Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/page_alloc.c
4 *
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
7 *
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
16 */
17
18#include <linux/stddef.h>
19#include <linux/mm.h>
20#include <linux/highmem.h>
21#include <linux/swap.h>
22#include <linux/interrupt.h>
23#include <linux/pagemap.h>
24#include <linux/jiffies.h>
25#include <linux/memblock.h>
26#include <linux/compiler.h>
27#include <linux/kernel.h>
28#include <linux/kasan.h>
29#include <linux/module.h>
30#include <linux/suspend.h>
31#include <linux/pagevec.h>
32#include <linux/blkdev.h>
33#include <linux/slab.h>
34#include <linux/ratelimit.h>
35#include <linux/oom.h>
36#include <linux/topology.h>
37#include <linux/sysctl.h>
38#include <linux/cpu.h>
39#include <linux/cpuset.h>
40#include <linux/memory_hotplug.h>
41#include <linux/nodemask.h>
42#include <linux/vmalloc.h>
43#include <linux/vmstat.h>
44#include <linux/mempolicy.h>
45#include <linux/memremap.h>
46#include <linux/stop_machine.h>
47#include <linux/random.h>
48#include <linux/sort.h>
49#include <linux/pfn.h>
50#include <linux/backing-dev.h>
51#include <linux/fault-inject.h>
52#include <linux/page-isolation.h>
53#include <linux/debugobjects.h>
54#include <linux/kmemleak.h>
55#include <linux/compaction.h>
56#include <trace/events/kmem.h>
57#include <trace/events/oom.h>
58#include <linux/prefetch.h>
59#include <linux/mm_inline.h>
60#include <linux/mmu_notifier.h>
61#include <linux/migrate.h>
62#include <linux/hugetlb.h>
63#include <linux/sched/rt.h>
64#include <linux/sched/mm.h>
65#include <linux/page_owner.h>
66#include <linux/kthread.h>
67#include <linux/memcontrol.h>
68#include <linux/ftrace.h>
69#include <linux/lockdep.h>
70#include <linux/nmi.h>
71#include <linux/psi.h>
72#include <linux/padata.h>
73#include <linux/khugepaged.h>
74#include <linux/buffer_head.h>
75#include <asm/sections.h>
76#include <asm/tlbflush.h>
77#include <asm/div64.h>
78#include "internal.h"
79#include "shuffle.h"
80#include "page_reporting.h"
81
82/* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
83typedef int __bitwise fpi_t;
84
85/* No special request */
86#define FPI_NONE ((__force fpi_t)0)
87
88/*
89 * Skip free page reporting notification for the (possibly merged) page.
90 * This does not hinder free page reporting from grabbing the page,
91 * reporting it and marking it "reported" - it only skips notifying
92 * the free page reporting infrastructure about a newly freed page. For
93 * example, used when temporarily pulling a page from a freelist and
94 * putting it back unmodified.
95 */
96#define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
97
98/*
99 * Place the (possibly merged) page to the tail of the freelist. Will ignore
100 * page shuffling (relevant code - e.g., memory onlining - is expected to
101 * shuffle the whole zone).
102 *
103 * Note: No code should rely on this flag for correctness - it's purely
104 * to allow for optimizations when handing back either fresh pages
105 * (memory onlining) or untouched pages (page isolation, free page
106 * reporting).
107 */
108#define FPI_TO_TAIL ((__force fpi_t)BIT(1))
109
110/*
111 * Don't poison memory with KASAN (only for the tag-based modes).
112 * During boot, all non-reserved memblock memory is exposed to page_alloc.
113 * Poisoning all that memory lengthens boot time, especially on systems with
114 * large amount of RAM. This flag is used to skip that poisoning.
115 * This is only done for the tag-based KASAN modes, as those are able to
116 * detect memory corruptions with the memory tags assigned by default.
117 * All memory allocated normally after boot gets poisoned as usual.
118 */
119#define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
120
121/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
122static DEFINE_MUTEX(pcp_batch_high_lock);
123#define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
124
125struct pagesets {
126 local_lock_t lock;
127};
128static DEFINE_PER_CPU(struct pagesets, pagesets) = {
129 .lock = INIT_LOCAL_LOCK(lock),
130};
131
132#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
133DEFINE_PER_CPU(int, numa_node);
134EXPORT_PER_CPU_SYMBOL(numa_node);
135#endif
136
137DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
138
139#ifdef CONFIG_HAVE_MEMORYLESS_NODES
140/*
141 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
142 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
143 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
144 * defined in <linux/topology.h>.
145 */
146DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
147EXPORT_PER_CPU_SYMBOL(_numa_mem_);
148#endif
149
150/* work_structs for global per-cpu drains */
151struct pcpu_drain {
152 struct zone *zone;
153 struct work_struct work;
154};
155static DEFINE_MUTEX(pcpu_drain_mutex);
156static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
157
158#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
159volatile unsigned long latent_entropy __latent_entropy;
160EXPORT_SYMBOL(latent_entropy);
161#endif
162
163/*
164 * Array of node states.
165 */
166nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
167 [N_POSSIBLE] = NODE_MASK_ALL,
168 [N_ONLINE] = { { [0] = 1UL } },
169#ifndef CONFIG_NUMA
170 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
171#ifdef CONFIG_HIGHMEM
172 [N_HIGH_MEMORY] = { { [0] = 1UL } },
173#endif
174 [N_MEMORY] = { { [0] = 1UL } },
175 [N_CPU] = { { [0] = 1UL } },
176#endif /* NUMA */
177};
178EXPORT_SYMBOL(node_states);
179
180atomic_long_t _totalram_pages __read_mostly;
181EXPORT_SYMBOL(_totalram_pages);
182unsigned long totalreserve_pages __read_mostly;
183unsigned long totalcma_pages __read_mostly;
184
185int percpu_pagelist_high_fraction;
186gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
187DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
188EXPORT_SYMBOL(init_on_alloc);
189
190DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
191EXPORT_SYMBOL(init_on_free);
192
193static bool _init_on_alloc_enabled_early __read_mostly
194 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
195static int __init early_init_on_alloc(char *buf)
196{
197
198 return kstrtobool(buf, &_init_on_alloc_enabled_early);
199}
200early_param("init_on_alloc", early_init_on_alloc);
201
202static bool _init_on_free_enabled_early __read_mostly
203 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
204static int __init early_init_on_free(char *buf)
205{
206 return kstrtobool(buf, &_init_on_free_enabled_early);
207}
208early_param("init_on_free", early_init_on_free);
209
210/*
211 * A cached value of the page's pageblock's migratetype, used when the page is
212 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
213 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
214 * Also the migratetype set in the page does not necessarily match the pcplist
215 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
216 * other index - this ensures that it will be put on the correct CMA freelist.
217 */
218static inline int get_pcppage_migratetype(struct page *page)
219{
220 return page->index;
221}
222
223static inline void set_pcppage_migratetype(struct page *page, int migratetype)
224{
225 page->index = migratetype;
226}
227
228#ifdef CONFIG_PM_SLEEP
229/*
230 * The following functions are used by the suspend/hibernate code to temporarily
231 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
232 * while devices are suspended. To avoid races with the suspend/hibernate code,
233 * they should always be called with system_transition_mutex held
234 * (gfp_allowed_mask also should only be modified with system_transition_mutex
235 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
236 * with that modification).
237 */
238
239static gfp_t saved_gfp_mask;
240
241void pm_restore_gfp_mask(void)
242{
243 WARN_ON(!mutex_is_locked(&system_transition_mutex));
244 if (saved_gfp_mask) {
245 gfp_allowed_mask = saved_gfp_mask;
246 saved_gfp_mask = 0;
247 }
248}
249
250void pm_restrict_gfp_mask(void)
251{
252 WARN_ON(!mutex_is_locked(&system_transition_mutex));
253 WARN_ON(saved_gfp_mask);
254 saved_gfp_mask = gfp_allowed_mask;
255 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
256}
257
258bool pm_suspended_storage(void)
259{
260 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
261 return false;
262 return true;
263}
264#endif /* CONFIG_PM_SLEEP */
265
266#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
267unsigned int pageblock_order __read_mostly;
268#endif
269
270static void __free_pages_ok(struct page *page, unsigned int order,
271 fpi_t fpi_flags);
272
273/*
274 * results with 256, 32 in the lowmem_reserve sysctl:
275 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
276 * 1G machine -> (16M dma, 784M normal, 224M high)
277 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
278 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
279 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
280 *
281 * TBD: should special case ZONE_DMA32 machines here - in those we normally
282 * don't need any ZONE_NORMAL reservation
283 */
284int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
285#ifdef CONFIG_ZONE_DMA
286 [ZONE_DMA] = 256,
287#endif
288#ifdef CONFIG_ZONE_DMA32
289 [ZONE_DMA32] = 256,
290#endif
291 [ZONE_NORMAL] = 32,
292#ifdef CONFIG_HIGHMEM
293 [ZONE_HIGHMEM] = 0,
294#endif
295 [ZONE_MOVABLE] = 0,
296};
297
298static char * const zone_names[MAX_NR_ZONES] = {
299#ifdef CONFIG_ZONE_DMA
300 "DMA",
301#endif
302#ifdef CONFIG_ZONE_DMA32
303 "DMA32",
304#endif
305 "Normal",
306#ifdef CONFIG_HIGHMEM
307 "HighMem",
308#endif
309 "Movable",
310#ifdef CONFIG_ZONE_DEVICE
311 "Device",
312#endif
313};
314
315const char * const migratetype_names[MIGRATE_TYPES] = {
316 "Unmovable",
317 "Movable",
318 "Reclaimable",
319 "HighAtomic",
320#ifdef CONFIG_CMA
321 "CMA",
322#endif
323#ifdef CONFIG_MEMORY_ISOLATION
324 "Isolate",
325#endif
326};
327
328compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
329 [NULL_COMPOUND_DTOR] = NULL,
330 [COMPOUND_PAGE_DTOR] = free_compound_page,
331#ifdef CONFIG_HUGETLB_PAGE
332 [HUGETLB_PAGE_DTOR] = free_huge_page,
333#endif
334#ifdef CONFIG_TRANSPARENT_HUGEPAGE
335 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
336#endif
337};
338
339int min_free_kbytes = 1024;
340int user_min_free_kbytes = -1;
341int watermark_boost_factor __read_mostly = 15000;
342int watermark_scale_factor = 10;
343
344static unsigned long nr_kernel_pages __initdata;
345static unsigned long nr_all_pages __initdata;
346static unsigned long dma_reserve __initdata;
347
348static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
349static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
350static unsigned long required_kernelcore __initdata;
351static unsigned long required_kernelcore_percent __initdata;
352static unsigned long required_movablecore __initdata;
353static unsigned long required_movablecore_percent __initdata;
354static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
355static bool mirrored_kernelcore __meminitdata;
356
357/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
358int movable_zone;
359EXPORT_SYMBOL(movable_zone);
360
361#if MAX_NUMNODES > 1
362unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
363unsigned int nr_online_nodes __read_mostly = 1;
364EXPORT_SYMBOL(nr_node_ids);
365EXPORT_SYMBOL(nr_online_nodes);
366#endif
367
368int page_group_by_mobility_disabled __read_mostly;
369
370#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
371/*
372 * During boot we initialize deferred pages on-demand, as needed, but once
373 * page_alloc_init_late() has finished, the deferred pages are all initialized,
374 * and we can permanently disable that path.
375 */
376static DEFINE_STATIC_KEY_TRUE(deferred_pages);
377
378/*
379 * Calling kasan_poison_pages() only after deferred memory initialization
380 * has completed. Poisoning pages during deferred memory init will greatly
381 * lengthen the process and cause problem in large memory systems as the
382 * deferred pages initialization is done with interrupt disabled.
383 *
384 * Assuming that there will be no reference to those newly initialized
385 * pages before they are ever allocated, this should have no effect on
386 * KASAN memory tracking as the poison will be properly inserted at page
387 * allocation time. The only corner case is when pages are allocated by
388 * on-demand allocation and then freed again before the deferred pages
389 * initialization is done, but this is not likely to happen.
390 */
391static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
392{
393 return static_branch_unlikely(&deferred_pages) ||
394 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
395 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
396 PageSkipKASanPoison(page);
397}
398
399/* Returns true if the struct page for the pfn is uninitialised */
400static inline bool __meminit early_page_uninitialised(unsigned long pfn)
401{
402 int nid = early_pfn_to_nid(pfn);
403
404 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
405 return true;
406
407 return false;
408}
409
410/*
411 * Returns true when the remaining initialisation should be deferred until
412 * later in the boot cycle when it can be parallelised.
413 */
414static bool __meminit
415defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
416{
417 static unsigned long prev_end_pfn, nr_initialised;
418
419 /*
420 * prev_end_pfn static that contains the end of previous zone
421 * No need to protect because called very early in boot before smp_init.
422 */
423 if (prev_end_pfn != end_pfn) {
424 prev_end_pfn = end_pfn;
425 nr_initialised = 0;
426 }
427
428 /* Always populate low zones for address-constrained allocations */
429 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
430 return false;
431
432 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
433 return true;
434 /*
435 * We start only with one section of pages, more pages are added as
436 * needed until the rest of deferred pages are initialized.
437 */
438 nr_initialised++;
439 if ((nr_initialised > PAGES_PER_SECTION) &&
440 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
441 NODE_DATA(nid)->first_deferred_pfn = pfn;
442 return true;
443 }
444 return false;
445}
446#else
447static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
448{
449 return (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
450 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
451 PageSkipKASanPoison(page);
452}
453
454static inline bool early_page_uninitialised(unsigned long pfn)
455{
456 return false;
457}
458
459static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
460{
461 return false;
462}
463#endif
464
465/* Return a pointer to the bitmap storing bits affecting a block of pages */
466static inline unsigned long *get_pageblock_bitmap(const struct page *page,
467 unsigned long pfn)
468{
469#ifdef CONFIG_SPARSEMEM
470 return section_to_usemap(__pfn_to_section(pfn));
471#else
472 return page_zone(page)->pageblock_flags;
473#endif /* CONFIG_SPARSEMEM */
474}
475
476static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
477{
478#ifdef CONFIG_SPARSEMEM
479 pfn &= (PAGES_PER_SECTION-1);
480#else
481 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
482#endif /* CONFIG_SPARSEMEM */
483 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
484}
485
486static __always_inline
487unsigned long __get_pfnblock_flags_mask(const struct page *page,
488 unsigned long pfn,
489 unsigned long mask)
490{
491 unsigned long *bitmap;
492 unsigned long bitidx, word_bitidx;
493 unsigned long word;
494
495 bitmap = get_pageblock_bitmap(page, pfn);
496 bitidx = pfn_to_bitidx(page, pfn);
497 word_bitidx = bitidx / BITS_PER_LONG;
498 bitidx &= (BITS_PER_LONG-1);
499
500 word = bitmap[word_bitidx];
501 return (word >> bitidx) & mask;
502}
503
504/**
505 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
506 * @page: The page within the block of interest
507 * @pfn: The target page frame number
508 * @mask: mask of bits that the caller is interested in
509 *
510 * Return: pageblock_bits flags
511 */
512unsigned long get_pfnblock_flags_mask(const struct page *page,
513 unsigned long pfn, unsigned long mask)
514{
515 return __get_pfnblock_flags_mask(page, pfn, mask);
516}
517
518static __always_inline int get_pfnblock_migratetype(const struct page *page,
519 unsigned long pfn)
520{
521 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
522}
523
524/**
525 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
526 * @page: The page within the block of interest
527 * @flags: The flags to set
528 * @pfn: The target page frame number
529 * @mask: mask of bits that the caller is interested in
530 */
531void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
532 unsigned long pfn,
533 unsigned long mask)
534{
535 unsigned long *bitmap;
536 unsigned long bitidx, word_bitidx;
537 unsigned long old_word, word;
538
539 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
540 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
541
542 bitmap = get_pageblock_bitmap(page, pfn);
543 bitidx = pfn_to_bitidx(page, pfn);
544 word_bitidx = bitidx / BITS_PER_LONG;
545 bitidx &= (BITS_PER_LONG-1);
546
547 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
548
549 mask <<= bitidx;
550 flags <<= bitidx;
551
552 word = READ_ONCE(bitmap[word_bitidx]);
553 for (;;) {
554 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
555 if (word == old_word)
556 break;
557 word = old_word;
558 }
559}
560
561void set_pageblock_migratetype(struct page *page, int migratetype)
562{
563 if (unlikely(page_group_by_mobility_disabled &&
564 migratetype < MIGRATE_PCPTYPES))
565 migratetype = MIGRATE_UNMOVABLE;
566
567 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
568 page_to_pfn(page), MIGRATETYPE_MASK);
569}
570
571#ifdef CONFIG_DEBUG_VM
572static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
573{
574 int ret = 0;
575 unsigned seq;
576 unsigned long pfn = page_to_pfn(page);
577 unsigned long sp, start_pfn;
578
579 do {
580 seq = zone_span_seqbegin(zone);
581 start_pfn = zone->zone_start_pfn;
582 sp = zone->spanned_pages;
583 if (!zone_spans_pfn(zone, pfn))
584 ret = 1;
585 } while (zone_span_seqretry(zone, seq));
586
587 if (ret)
588 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
589 pfn, zone_to_nid(zone), zone->name,
590 start_pfn, start_pfn + sp);
591
592 return ret;
593}
594
595static int page_is_consistent(struct zone *zone, struct page *page)
596{
597 if (!pfn_valid_within(page_to_pfn(page)))
598 return 0;
599 if (zone != page_zone(page))
600 return 0;
601
602 return 1;
603}
604/*
605 * Temporary debugging check for pages not lying within a given zone.
606 */
607static int __maybe_unused bad_range(struct zone *zone, struct page *page)
608{
609 if (page_outside_zone_boundaries(zone, page))
610 return 1;
611 if (!page_is_consistent(zone, page))
612 return 1;
613
614 return 0;
615}
616#else
617static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
618{
619 return 0;
620}
621#endif
622
623static void bad_page(struct page *page, const char *reason)
624{
625 static unsigned long resume;
626 static unsigned long nr_shown;
627 static unsigned long nr_unshown;
628
629 /*
630 * Allow a burst of 60 reports, then keep quiet for that minute;
631 * or allow a steady drip of one report per second.
632 */
633 if (nr_shown == 60) {
634 if (time_before(jiffies, resume)) {
635 nr_unshown++;
636 goto out;
637 }
638 if (nr_unshown) {
639 pr_alert(
640 "BUG: Bad page state: %lu messages suppressed\n",
641 nr_unshown);
642 nr_unshown = 0;
643 }
644 nr_shown = 0;
645 }
646 if (nr_shown++ == 0)
647 resume = jiffies + 60 * HZ;
648
649 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
650 current->comm, page_to_pfn(page));
651 dump_page(page, reason);
652
653 print_modules();
654 dump_stack();
655out:
656 /* Leave bad fields for debug, except PageBuddy could make trouble */
657 page_mapcount_reset(page); /* remove PageBuddy */
658 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
659}
660
661static inline unsigned int order_to_pindex(int migratetype, int order)
662{
663 int base = order;
664
665#ifdef CONFIG_TRANSPARENT_HUGEPAGE
666 if (order > PAGE_ALLOC_COSTLY_ORDER) {
667 VM_BUG_ON(order != pageblock_order);
668 base = PAGE_ALLOC_COSTLY_ORDER + 1;
669 }
670#else
671 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
672#endif
673
674 return (MIGRATE_PCPTYPES * base) + migratetype;
675}
676
677static inline int pindex_to_order(unsigned int pindex)
678{
679 int order = pindex / MIGRATE_PCPTYPES;
680
681#ifdef CONFIG_TRANSPARENT_HUGEPAGE
682 if (order > PAGE_ALLOC_COSTLY_ORDER) {
683 order = pageblock_order;
684 VM_BUG_ON(order != pageblock_order);
685 }
686#else
687 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
688#endif
689
690 return order;
691}
692
693static inline bool pcp_allowed_order(unsigned int order)
694{
695 if (order <= PAGE_ALLOC_COSTLY_ORDER)
696 return true;
697#ifdef CONFIG_TRANSPARENT_HUGEPAGE
698 if (order == pageblock_order)
699 return true;
700#endif
701 return false;
702}
703
704static inline void free_the_page(struct page *page, unsigned int order)
705{
706 if (pcp_allowed_order(order)) /* Via pcp? */
707 free_unref_page(page, order);
708 else
709 __free_pages_ok(page, order, FPI_NONE);
710}
711
712/*
713 * Higher-order pages are called "compound pages". They are structured thusly:
714 *
715 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
716 *
717 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
718 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
719 *
720 * The first tail page's ->compound_dtor holds the offset in array of compound
721 * page destructors. See compound_page_dtors.
722 *
723 * The first tail page's ->compound_order holds the order of allocation.
724 * This usage means that zero-order pages may not be compound.
725 */
726
727void free_compound_page(struct page *page)
728{
729 mem_cgroup_uncharge(page);
730 free_the_page(page, compound_order(page));
731}
732
733void prep_compound_page(struct page *page, unsigned int order)
734{
735 int i;
736 int nr_pages = 1 << order;
737
738 __SetPageHead(page);
739 for (i = 1; i < nr_pages; i++) {
740 struct page *p = page + i;
741 p->mapping = TAIL_MAPPING;
742 set_compound_head(p, page);
743 }
744
745 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
746 set_compound_order(page, order);
747 atomic_set(compound_mapcount_ptr(page), -1);
748 if (hpage_pincount_available(page))
749 atomic_set(compound_pincount_ptr(page), 0);
750}
751
752#ifdef CONFIG_DEBUG_PAGEALLOC
753unsigned int _debug_guardpage_minorder;
754
755bool _debug_pagealloc_enabled_early __read_mostly
756 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
757EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
758DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
759EXPORT_SYMBOL(_debug_pagealloc_enabled);
760
761DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
762
763static int __init early_debug_pagealloc(char *buf)
764{
765 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
766}
767early_param("debug_pagealloc", early_debug_pagealloc);
768
769static int __init debug_guardpage_minorder_setup(char *buf)
770{
771 unsigned long res;
772
773 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
774 pr_err("Bad debug_guardpage_minorder value\n");
775 return 0;
776 }
777 _debug_guardpage_minorder = res;
778 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
779 return 0;
780}
781early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
782
783static inline bool set_page_guard(struct zone *zone, struct page *page,
784 unsigned int order, int migratetype)
785{
786 if (!debug_guardpage_enabled())
787 return false;
788
789 if (order >= debug_guardpage_minorder())
790 return false;
791
792 __SetPageGuard(page);
793 INIT_LIST_HEAD(&page->lru);
794 set_page_private(page, order);
795 /* Guard pages are not available for any usage */
796 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
797
798 return true;
799}
800
801static inline void clear_page_guard(struct zone *zone, struct page *page,
802 unsigned int order, int migratetype)
803{
804 if (!debug_guardpage_enabled())
805 return;
806
807 __ClearPageGuard(page);
808
809 set_page_private(page, 0);
810 if (!is_migrate_isolate(migratetype))
811 __mod_zone_freepage_state(zone, (1 << order), migratetype);
812}
813#else
814static inline bool set_page_guard(struct zone *zone, struct page *page,
815 unsigned int order, int migratetype) { return false; }
816static inline void clear_page_guard(struct zone *zone, struct page *page,
817 unsigned int order, int migratetype) {}
818#endif
819
820/*
821 * Enable static keys related to various memory debugging and hardening options.
822 * Some override others, and depend on early params that are evaluated in the
823 * order of appearance. So we need to first gather the full picture of what was
824 * enabled, and then make decisions.
825 */
826void init_mem_debugging_and_hardening(void)
827{
828 bool page_poisoning_requested = false;
829
830#ifdef CONFIG_PAGE_POISONING
831 /*
832 * Page poisoning is debug page alloc for some arches. If
833 * either of those options are enabled, enable poisoning.
834 */
835 if (page_poisoning_enabled() ||
836 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
837 debug_pagealloc_enabled())) {
838 static_branch_enable(&_page_poisoning_enabled);
839 page_poisoning_requested = true;
840 }
841#endif
842
843 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
844 page_poisoning_requested) {
845 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
846 "will take precedence over init_on_alloc and init_on_free\n");
847 _init_on_alloc_enabled_early = false;
848 _init_on_free_enabled_early = false;
849 }
850
851 if (_init_on_alloc_enabled_early)
852 static_branch_enable(&init_on_alloc);
853 else
854 static_branch_disable(&init_on_alloc);
855
856 if (_init_on_free_enabled_early)
857 static_branch_enable(&init_on_free);
858 else
859 static_branch_disable(&init_on_free);
860
861#ifdef CONFIG_DEBUG_PAGEALLOC
862 if (!debug_pagealloc_enabled())
863 return;
864
865 static_branch_enable(&_debug_pagealloc_enabled);
866
867 if (!debug_guardpage_minorder())
868 return;
869
870 static_branch_enable(&_debug_guardpage_enabled);
871#endif
872}
873
874static inline void set_buddy_order(struct page *page, unsigned int order)
875{
876 set_page_private(page, order);
877 __SetPageBuddy(page);
878}
879
880/*
881 * This function checks whether a page is free && is the buddy
882 * we can coalesce a page and its buddy if
883 * (a) the buddy is not in a hole (check before calling!) &&
884 * (b) the buddy is in the buddy system &&
885 * (c) a page and its buddy have the same order &&
886 * (d) a page and its buddy are in the same zone.
887 *
888 * For recording whether a page is in the buddy system, we set PageBuddy.
889 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
890 *
891 * For recording page's order, we use page_private(page).
892 */
893static inline bool page_is_buddy(struct page *page, struct page *buddy,
894 unsigned int order)
895{
896 if (!page_is_guard(buddy) && !PageBuddy(buddy))
897 return false;
898
899 if (buddy_order(buddy) != order)
900 return false;
901
902 /*
903 * zone check is done late to avoid uselessly calculating
904 * zone/node ids for pages that could never merge.
905 */
906 if (page_zone_id(page) != page_zone_id(buddy))
907 return false;
908
909 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
910
911 return true;
912}
913
914#ifdef CONFIG_COMPACTION
915static inline struct capture_control *task_capc(struct zone *zone)
916{
917 struct capture_control *capc = current->capture_control;
918
919 return unlikely(capc) &&
920 !(current->flags & PF_KTHREAD) &&
921 !capc->page &&
922 capc->cc->zone == zone ? capc : NULL;
923}
924
925static inline bool
926compaction_capture(struct capture_control *capc, struct page *page,
927 int order, int migratetype)
928{
929 if (!capc || order != capc->cc->order)
930 return false;
931
932 /* Do not accidentally pollute CMA or isolated regions*/
933 if (is_migrate_cma(migratetype) ||
934 is_migrate_isolate(migratetype))
935 return false;
936
937 /*
938 * Do not let lower order allocations pollute a movable pageblock.
939 * This might let an unmovable request use a reclaimable pageblock
940 * and vice-versa but no more than normal fallback logic which can
941 * have trouble finding a high-order free page.
942 */
943 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
944 return false;
945
946 capc->page = page;
947 return true;
948}
949
950#else
951static inline struct capture_control *task_capc(struct zone *zone)
952{
953 return NULL;
954}
955
956static inline bool
957compaction_capture(struct capture_control *capc, struct page *page,
958 int order, int migratetype)
959{
960 return false;
961}
962#endif /* CONFIG_COMPACTION */
963
964/* Used for pages not on another list */
965static inline void add_to_free_list(struct page *page, struct zone *zone,
966 unsigned int order, int migratetype)
967{
968 struct free_area *area = &zone->free_area[order];
969
970 list_add(&page->lru, &area->free_list[migratetype]);
971 area->nr_free++;
972}
973
974/* Used for pages not on another list */
975static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
976 unsigned int order, int migratetype)
977{
978 struct free_area *area = &zone->free_area[order];
979
980 list_add_tail(&page->lru, &area->free_list[migratetype]);
981 area->nr_free++;
982}
983
984/*
985 * Used for pages which are on another list. Move the pages to the tail
986 * of the list - so the moved pages won't immediately be considered for
987 * allocation again (e.g., optimization for memory onlining).
988 */
989static inline void move_to_free_list(struct page *page, struct zone *zone,
990 unsigned int order, int migratetype)
991{
992 struct free_area *area = &zone->free_area[order];
993
994 list_move_tail(&page->lru, &area->free_list[migratetype]);
995}
996
997static inline void del_page_from_free_list(struct page *page, struct zone *zone,
998 unsigned int order)
999{
1000 /* clear reported state and update reported page count */
1001 if (page_reported(page))
1002 __ClearPageReported(page);
1003
1004 list_del(&page->lru);
1005 __ClearPageBuddy(page);
1006 set_page_private(page, 0);
1007 zone->free_area[order].nr_free--;
1008}
1009
1010/*
1011 * If this is not the largest possible page, check if the buddy
1012 * of the next-highest order is free. If it is, it's possible
1013 * that pages are being freed that will coalesce soon. In case,
1014 * that is happening, add the free page to the tail of the list
1015 * so it's less likely to be used soon and more likely to be merged
1016 * as a higher order page
1017 */
1018static inline bool
1019buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1020 struct page *page, unsigned int order)
1021{
1022 struct page *higher_page, *higher_buddy;
1023 unsigned long combined_pfn;
1024
1025 if (order >= MAX_ORDER - 2)
1026 return false;
1027
1028 if (!pfn_valid_within(buddy_pfn))
1029 return false;
1030
1031 combined_pfn = buddy_pfn & pfn;
1032 higher_page = page + (combined_pfn - pfn);
1033 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1034 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1035
1036 return pfn_valid_within(buddy_pfn) &&
1037 page_is_buddy(higher_page, higher_buddy, order + 1);
1038}
1039
1040/*
1041 * Freeing function for a buddy system allocator.
1042 *
1043 * The concept of a buddy system is to maintain direct-mapped table
1044 * (containing bit values) for memory blocks of various "orders".
1045 * The bottom level table contains the map for the smallest allocatable
1046 * units of memory (here, pages), and each level above it describes
1047 * pairs of units from the levels below, hence, "buddies".
1048 * At a high level, all that happens here is marking the table entry
1049 * at the bottom level available, and propagating the changes upward
1050 * as necessary, plus some accounting needed to play nicely with other
1051 * parts of the VM system.
1052 * At each level, we keep a list of pages, which are heads of continuous
1053 * free pages of length of (1 << order) and marked with PageBuddy.
1054 * Page's order is recorded in page_private(page) field.
1055 * So when we are allocating or freeing one, we can derive the state of the
1056 * other. That is, if we allocate a small block, and both were
1057 * free, the remainder of the region must be split into blocks.
1058 * If a block is freed, and its buddy is also free, then this
1059 * triggers coalescing into a block of larger size.
1060 *
1061 * -- nyc
1062 */
1063
1064static inline void __free_one_page(struct page *page,
1065 unsigned long pfn,
1066 struct zone *zone, unsigned int order,
1067 int migratetype, fpi_t fpi_flags)
1068{
1069 struct capture_control *capc = task_capc(zone);
1070 unsigned long buddy_pfn;
1071 unsigned long combined_pfn;
1072 unsigned int max_order;
1073 struct page *buddy;
1074 bool to_tail;
1075
1076 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1077
1078 VM_BUG_ON(!zone_is_initialized(zone));
1079 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1080
1081 VM_BUG_ON(migratetype == -1);
1082 if (likely(!is_migrate_isolate(migratetype)))
1083 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1084
1085 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1086 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1087
1088continue_merging:
1089 while (order < max_order) {
1090 if (compaction_capture(capc, page, order, migratetype)) {
1091 __mod_zone_freepage_state(zone, -(1 << order),
1092 migratetype);
1093 return;
1094 }
1095 buddy_pfn = __find_buddy_pfn(pfn, order);
1096 buddy = page + (buddy_pfn - pfn);
1097
1098 if (!pfn_valid_within(buddy_pfn))
1099 goto done_merging;
1100 if (!page_is_buddy(page, buddy, order))
1101 goto done_merging;
1102 /*
1103 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1104 * merge with it and move up one order.
1105 */
1106 if (page_is_guard(buddy))
1107 clear_page_guard(zone, buddy, order, migratetype);
1108 else
1109 del_page_from_free_list(buddy, zone, order);
1110 combined_pfn = buddy_pfn & pfn;
1111 page = page + (combined_pfn - pfn);
1112 pfn = combined_pfn;
1113 order++;
1114 }
1115 if (order < MAX_ORDER - 1) {
1116 /* If we are here, it means order is >= pageblock_order.
1117 * We want to prevent merge between freepages on isolate
1118 * pageblock and normal pageblock. Without this, pageblock
1119 * isolation could cause incorrect freepage or CMA accounting.
1120 *
1121 * We don't want to hit this code for the more frequent
1122 * low-order merging.
1123 */
1124 if (unlikely(has_isolate_pageblock(zone))) {
1125 int buddy_mt;
1126
1127 buddy_pfn = __find_buddy_pfn(pfn, order);
1128 buddy = page + (buddy_pfn - pfn);
1129 buddy_mt = get_pageblock_migratetype(buddy);
1130
1131 if (migratetype != buddy_mt
1132 && (is_migrate_isolate(migratetype) ||
1133 is_migrate_isolate(buddy_mt)))
1134 goto done_merging;
1135 }
1136 max_order = order + 1;
1137 goto continue_merging;
1138 }
1139
1140done_merging:
1141 set_buddy_order(page, order);
1142
1143 if (fpi_flags & FPI_TO_TAIL)
1144 to_tail = true;
1145 else if (is_shuffle_order(order))
1146 to_tail = shuffle_pick_tail();
1147 else
1148 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1149
1150 if (to_tail)
1151 add_to_free_list_tail(page, zone, order, migratetype);
1152 else
1153 add_to_free_list(page, zone, order, migratetype);
1154
1155 /* Notify page reporting subsystem of freed page */
1156 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1157 page_reporting_notify_free(order);
1158}
1159
1160/*
1161 * A bad page could be due to a number of fields. Instead of multiple branches,
1162 * try and check multiple fields with one check. The caller must do a detailed
1163 * check if necessary.
1164 */
1165static inline bool page_expected_state(struct page *page,
1166 unsigned long check_flags)
1167{
1168 if (unlikely(atomic_read(&page->_mapcount) != -1))
1169 return false;
1170
1171 if (unlikely((unsigned long)page->mapping |
1172 page_ref_count(page) |
1173#ifdef CONFIG_MEMCG
1174 page->memcg_data |
1175#endif
1176 (page->flags & check_flags)))
1177 return false;
1178
1179 return true;
1180}
1181
1182static const char *page_bad_reason(struct page *page, unsigned long flags)
1183{
1184 const char *bad_reason = NULL;
1185
1186 if (unlikely(atomic_read(&page->_mapcount) != -1))
1187 bad_reason = "nonzero mapcount";
1188 if (unlikely(page->mapping != NULL))
1189 bad_reason = "non-NULL mapping";
1190 if (unlikely(page_ref_count(page) != 0))
1191 bad_reason = "nonzero _refcount";
1192 if (unlikely(page->flags & flags)) {
1193 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1194 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1195 else
1196 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1197 }
1198#ifdef CONFIG_MEMCG
1199 if (unlikely(page->memcg_data))
1200 bad_reason = "page still charged to cgroup";
1201#endif
1202 return bad_reason;
1203}
1204
1205static void check_free_page_bad(struct page *page)
1206{
1207 bad_page(page,
1208 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1209}
1210
1211static inline int check_free_page(struct page *page)
1212{
1213 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1214 return 0;
1215
1216 /* Something has gone sideways, find it */
1217 check_free_page_bad(page);
1218 return 1;
1219}
1220
1221static int free_tail_pages_check(struct page *head_page, struct page *page)
1222{
1223 int ret = 1;
1224
1225 /*
1226 * We rely page->lru.next never has bit 0 set, unless the page
1227 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1228 */
1229 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1230
1231 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1232 ret = 0;
1233 goto out;
1234 }
1235 switch (page - head_page) {
1236 case 1:
1237 /* the first tail page: ->mapping may be compound_mapcount() */
1238 if (unlikely(compound_mapcount(page))) {
1239 bad_page(page, "nonzero compound_mapcount");
1240 goto out;
1241 }
1242 break;
1243 case 2:
1244 /*
1245 * the second tail page: ->mapping is
1246 * deferred_list.next -- ignore value.
1247 */
1248 break;
1249 default:
1250 if (page->mapping != TAIL_MAPPING) {
1251 bad_page(page, "corrupted mapping in tail page");
1252 goto out;
1253 }
1254 break;
1255 }
1256 if (unlikely(!PageTail(page))) {
1257 bad_page(page, "PageTail not set");
1258 goto out;
1259 }
1260 if (unlikely(compound_head(page) != head_page)) {
1261 bad_page(page, "compound_head not consistent");
1262 goto out;
1263 }
1264 ret = 0;
1265out:
1266 page->mapping = NULL;
1267 clear_compound_head(page);
1268 return ret;
1269}
1270
1271static void kernel_init_free_pages(struct page *page, int numpages, bool zero_tags)
1272{
1273 int i;
1274
1275 if (zero_tags) {
1276 for (i = 0; i < numpages; i++)
1277 tag_clear_highpage(page + i);
1278 return;
1279 }
1280
1281 /* s390's use of memset() could override KASAN redzones. */
1282 kasan_disable_current();
1283 for (i = 0; i < numpages; i++) {
1284 u8 tag = page_kasan_tag(page + i);
1285 page_kasan_tag_reset(page + i);
1286 clear_highpage(page + i);
1287 page_kasan_tag_set(page + i, tag);
1288 }
1289 kasan_enable_current();
1290}
1291
1292static __always_inline bool free_pages_prepare(struct page *page,
1293 unsigned int order, bool check_free, fpi_t fpi_flags)
1294{
1295 int bad = 0;
1296 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1297
1298 VM_BUG_ON_PAGE(PageTail(page), page);
1299
1300 trace_mm_page_free(page, order);
1301
1302 if (unlikely(PageHWPoison(page)) && !order) {
1303 /*
1304 * Do not let hwpoison pages hit pcplists/buddy
1305 * Untie memcg state and reset page's owner
1306 */
1307 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1308 __memcg_kmem_uncharge_page(page, order);
1309 reset_page_owner(page, order);
1310 return false;
1311 }
1312
1313 /*
1314 * Check tail pages before head page information is cleared to
1315 * avoid checking PageCompound for order-0 pages.
1316 */
1317 if (unlikely(order)) {
1318 bool compound = PageCompound(page);
1319 int i;
1320
1321 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1322
1323 if (compound)
1324 ClearPageDoubleMap(page);
1325 for (i = 1; i < (1 << order); i++) {
1326 if (compound)
1327 bad += free_tail_pages_check(page, page + i);
1328 if (unlikely(check_free_page(page + i))) {
1329 bad++;
1330 continue;
1331 }
1332 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1333 }
1334 }
1335 if (PageMappingFlags(page))
1336 page->mapping = NULL;
1337 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1338 __memcg_kmem_uncharge_page(page, order);
1339 if (check_free)
1340 bad += check_free_page(page);
1341 if (bad)
1342 return false;
1343
1344 page_cpupid_reset_last(page);
1345 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1346 reset_page_owner(page, order);
1347
1348 if (!PageHighMem(page)) {
1349 debug_check_no_locks_freed(page_address(page),
1350 PAGE_SIZE << order);
1351 debug_check_no_obj_freed(page_address(page),
1352 PAGE_SIZE << order);
1353 }
1354
1355 kernel_poison_pages(page, 1 << order);
1356
1357 /*
1358 * As memory initialization might be integrated into KASAN,
1359 * kasan_free_pages and kernel_init_free_pages must be
1360 * kept together to avoid discrepancies in behavior.
1361 *
1362 * With hardware tag-based KASAN, memory tags must be set before the
1363 * page becomes unavailable via debug_pagealloc or arch_free_page.
1364 */
1365 if (kasan_has_integrated_init()) {
1366 if (!skip_kasan_poison)
1367 kasan_free_pages(page, order);
1368 } else {
1369 bool init = want_init_on_free();
1370
1371 if (init)
1372 kernel_init_free_pages(page, 1 << order, false);
1373 if (!skip_kasan_poison)
1374 kasan_poison_pages(page, order, init);
1375 }
1376
1377 /*
1378 * arch_free_page() can make the page's contents inaccessible. s390
1379 * does this. So nothing which can access the page's contents should
1380 * happen after this.
1381 */
1382 arch_free_page(page, order);
1383
1384 debug_pagealloc_unmap_pages(page, 1 << order);
1385
1386 return true;
1387}
1388
1389#ifdef CONFIG_DEBUG_VM
1390/*
1391 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1392 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1393 * moved from pcp lists to free lists.
1394 */
1395static bool free_pcp_prepare(struct page *page, unsigned int order)
1396{
1397 return free_pages_prepare(page, order, true, FPI_NONE);
1398}
1399
1400static bool bulkfree_pcp_prepare(struct page *page)
1401{
1402 if (debug_pagealloc_enabled_static())
1403 return check_free_page(page);
1404 else
1405 return false;
1406}
1407#else
1408/*
1409 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1410 * moving from pcp lists to free list in order to reduce overhead. With
1411 * debug_pagealloc enabled, they are checked also immediately when being freed
1412 * to the pcp lists.
1413 */
1414static bool free_pcp_prepare(struct page *page, unsigned int order)
1415{
1416 if (debug_pagealloc_enabled_static())
1417 return free_pages_prepare(page, order, true, FPI_NONE);
1418 else
1419 return free_pages_prepare(page, order, false, FPI_NONE);
1420}
1421
1422static bool bulkfree_pcp_prepare(struct page *page)
1423{
1424 return check_free_page(page);
1425}
1426#endif /* CONFIG_DEBUG_VM */
1427
1428static inline void prefetch_buddy(struct page *page)
1429{
1430 unsigned long pfn = page_to_pfn(page);
1431 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1432 struct page *buddy = page + (buddy_pfn - pfn);
1433
1434 prefetch(buddy);
1435}
1436
1437/*
1438 * Frees a number of pages from the PCP lists
1439 * Assumes all pages on list are in same zone, and of same order.
1440 * count is the number of pages to free.
1441 *
1442 * If the zone was previously in an "all pages pinned" state then look to
1443 * see if this freeing clears that state.
1444 *
1445 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1446 * pinned" detection logic.
1447 */
1448static void free_pcppages_bulk(struct zone *zone, int count,
1449 struct per_cpu_pages *pcp)
1450{
1451 int pindex = 0;
1452 int batch_free = 0;
1453 int nr_freed = 0;
1454 unsigned int order;
1455 int prefetch_nr = READ_ONCE(pcp->batch);
1456 bool isolated_pageblocks;
1457 struct page *page, *tmp;
1458 LIST_HEAD(head);
1459
1460 /*
1461 * Ensure proper count is passed which otherwise would stuck in the
1462 * below while (list_empty(list)) loop.
1463 */
1464 count = min(pcp->count, count);
1465 while (count > 0) {
1466 struct list_head *list;
1467
1468 /*
1469 * Remove pages from lists in a round-robin fashion. A
1470 * batch_free count is maintained that is incremented when an
1471 * empty list is encountered. This is so more pages are freed
1472 * off fuller lists instead of spinning excessively around empty
1473 * lists
1474 */
1475 do {
1476 batch_free++;
1477 if (++pindex == NR_PCP_LISTS)
1478 pindex = 0;
1479 list = &pcp->lists[pindex];
1480 } while (list_empty(list));
1481
1482 /* This is the only non-empty list. Free them all. */
1483 if (batch_free == NR_PCP_LISTS)
1484 batch_free = count;
1485
1486 order = pindex_to_order(pindex);
1487 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1488 do {
1489 page = list_last_entry(list, struct page, lru);
1490 /* must delete to avoid corrupting pcp list */
1491 list_del(&page->lru);
1492 nr_freed += 1 << order;
1493 count -= 1 << order;
1494
1495 if (bulkfree_pcp_prepare(page))
1496 continue;
1497
1498 /* Encode order with the migratetype */
1499 page->index <<= NR_PCP_ORDER_WIDTH;
1500 page->index |= order;
1501
1502 list_add_tail(&page->lru, &head);
1503
1504 /*
1505 * We are going to put the page back to the global
1506 * pool, prefetch its buddy to speed up later access
1507 * under zone->lock. It is believed the overhead of
1508 * an additional test and calculating buddy_pfn here
1509 * can be offset by reduced memory latency later. To
1510 * avoid excessive prefetching due to large count, only
1511 * prefetch buddy for the first pcp->batch nr of pages.
1512 */
1513 if (prefetch_nr) {
1514 prefetch_buddy(page);
1515 prefetch_nr--;
1516 }
1517 } while (count > 0 && --batch_free && !list_empty(list));
1518 }
1519 pcp->count -= nr_freed;
1520
1521 /*
1522 * local_lock_irq held so equivalent to spin_lock_irqsave for
1523 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1524 */
1525 spin_lock(&zone->lock);
1526 isolated_pageblocks = has_isolate_pageblock(zone);
1527
1528 /*
1529 * Use safe version since after __free_one_page(),
1530 * page->lru.next will not point to original list.
1531 */
1532 list_for_each_entry_safe(page, tmp, &head, lru) {
1533 int mt = get_pcppage_migratetype(page);
1534
1535 /* mt has been encoded with the order (see above) */
1536 order = mt & NR_PCP_ORDER_MASK;
1537 mt >>= NR_PCP_ORDER_WIDTH;
1538
1539 /* MIGRATE_ISOLATE page should not go to pcplists */
1540 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1541 /* Pageblock could have been isolated meanwhile */
1542 if (unlikely(isolated_pageblocks))
1543 mt = get_pageblock_migratetype(page);
1544
1545 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1546 trace_mm_page_pcpu_drain(page, order, mt);
1547 }
1548 spin_unlock(&zone->lock);
1549}
1550
1551static void free_one_page(struct zone *zone,
1552 struct page *page, unsigned long pfn,
1553 unsigned int order,
1554 int migratetype, fpi_t fpi_flags)
1555{
1556 unsigned long flags;
1557
1558 spin_lock_irqsave(&zone->lock, flags);
1559 if (unlikely(has_isolate_pageblock(zone) ||
1560 is_migrate_isolate(migratetype))) {
1561 migratetype = get_pfnblock_migratetype(page, pfn);
1562 }
1563 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1564 spin_unlock_irqrestore(&zone->lock, flags);
1565}
1566
1567static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1568 unsigned long zone, int nid)
1569{
1570 mm_zero_struct_page(page);
1571 set_page_links(page, zone, nid, pfn);
1572 init_page_count(page);
1573 page_mapcount_reset(page);
1574 page_cpupid_reset_last(page);
1575 page_kasan_tag_reset(page);
1576
1577 INIT_LIST_HEAD(&page->lru);
1578#ifdef WANT_PAGE_VIRTUAL
1579 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1580 if (!is_highmem_idx(zone))
1581 set_page_address(page, __va(pfn << PAGE_SHIFT));
1582#endif
1583}
1584
1585#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1586static void __meminit init_reserved_page(unsigned long pfn)
1587{
1588 pg_data_t *pgdat;
1589 int nid, zid;
1590
1591 if (!early_page_uninitialised(pfn))
1592 return;
1593
1594 nid = early_pfn_to_nid(pfn);
1595 pgdat = NODE_DATA(nid);
1596
1597 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1598 struct zone *zone = &pgdat->node_zones[zid];
1599
1600 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1601 break;
1602 }
1603 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1604}
1605#else
1606static inline void init_reserved_page(unsigned long pfn)
1607{
1608}
1609#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1610
1611/*
1612 * Initialised pages do not have PageReserved set. This function is
1613 * called for each range allocated by the bootmem allocator and
1614 * marks the pages PageReserved. The remaining valid pages are later
1615 * sent to the buddy page allocator.
1616 */
1617void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1618{
1619 unsigned long start_pfn = PFN_DOWN(start);
1620 unsigned long end_pfn = PFN_UP(end);
1621
1622 for (; start_pfn < end_pfn; start_pfn++) {
1623 if (pfn_valid(start_pfn)) {
1624 struct page *page = pfn_to_page(start_pfn);
1625
1626 init_reserved_page(start_pfn);
1627
1628 /* Avoid false-positive PageTail() */
1629 INIT_LIST_HEAD(&page->lru);
1630
1631 /*
1632 * no need for atomic set_bit because the struct
1633 * page is not visible yet so nobody should
1634 * access it yet.
1635 */
1636 __SetPageReserved(page);
1637 }
1638 }
1639}
1640
1641static void __free_pages_ok(struct page *page, unsigned int order,
1642 fpi_t fpi_flags)
1643{
1644 unsigned long flags;
1645 int migratetype;
1646 unsigned long pfn = page_to_pfn(page);
1647 struct zone *zone = page_zone(page);
1648
1649 if (!free_pages_prepare(page, order, true, fpi_flags))
1650 return;
1651
1652 migratetype = get_pfnblock_migratetype(page, pfn);
1653
1654 spin_lock_irqsave(&zone->lock, flags);
1655 if (unlikely(has_isolate_pageblock(zone) ||
1656 is_migrate_isolate(migratetype))) {
1657 migratetype = get_pfnblock_migratetype(page, pfn);
1658 }
1659 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1660 spin_unlock_irqrestore(&zone->lock, flags);
1661
1662 __count_vm_events(PGFREE, 1 << order);
1663}
1664
1665void __free_pages_core(struct page *page, unsigned int order)
1666{
1667 unsigned int nr_pages = 1 << order;
1668 struct page *p = page;
1669 unsigned int loop;
1670
1671 /*
1672 * When initializing the memmap, __init_single_page() sets the refcount
1673 * of all pages to 1 ("allocated"/"not free"). We have to set the
1674 * refcount of all involved pages to 0.
1675 */
1676 prefetchw(p);
1677 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1678 prefetchw(p + 1);
1679 __ClearPageReserved(p);
1680 set_page_count(p, 0);
1681 }
1682 __ClearPageReserved(p);
1683 set_page_count(p, 0);
1684
1685 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1686
1687 /*
1688 * Bypass PCP and place fresh pages right to the tail, primarily
1689 * relevant for memory onlining.
1690 */
1691 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1692}
1693
1694#ifdef CONFIG_NUMA
1695
1696/*
1697 * During memory init memblocks map pfns to nids. The search is expensive and
1698 * this caches recent lookups. The implementation of __early_pfn_to_nid
1699 * treats start/end as pfns.
1700 */
1701struct mminit_pfnnid_cache {
1702 unsigned long last_start;
1703 unsigned long last_end;
1704 int last_nid;
1705};
1706
1707static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1708
1709/*
1710 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1711 */
1712static int __meminit __early_pfn_to_nid(unsigned long pfn,
1713 struct mminit_pfnnid_cache *state)
1714{
1715 unsigned long start_pfn, end_pfn;
1716 int nid;
1717
1718 if (state->last_start <= pfn && pfn < state->last_end)
1719 return state->last_nid;
1720
1721 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1722 if (nid != NUMA_NO_NODE) {
1723 state->last_start = start_pfn;
1724 state->last_end = end_pfn;
1725 state->last_nid = nid;
1726 }
1727
1728 return nid;
1729}
1730
1731int __meminit early_pfn_to_nid(unsigned long pfn)
1732{
1733 static DEFINE_SPINLOCK(early_pfn_lock);
1734 int nid;
1735
1736 spin_lock(&early_pfn_lock);
1737 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1738 if (nid < 0)
1739 nid = first_online_node;
1740 spin_unlock(&early_pfn_lock);
1741
1742 return nid;
1743}
1744#endif /* CONFIG_NUMA */
1745
1746void __init memblock_free_pages(struct page *page, unsigned long pfn,
1747 unsigned int order)
1748{
1749 if (early_page_uninitialised(pfn))
1750 return;
1751 __free_pages_core(page, order);
1752}
1753
1754/*
1755 * Check that the whole (or subset of) a pageblock given by the interval of
1756 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1757 * with the migration of free compaction scanner. The scanners then need to
1758 * use only pfn_valid_within() check for arches that allow holes within
1759 * pageblocks.
1760 *
1761 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1762 *
1763 * It's possible on some configurations to have a setup like node0 node1 node0
1764 * i.e. it's possible that all pages within a zones range of pages do not
1765 * belong to a single zone. We assume that a border between node0 and node1
1766 * can occur within a single pageblock, but not a node0 node1 node0
1767 * interleaving within a single pageblock. It is therefore sufficient to check
1768 * the first and last page of a pageblock and avoid checking each individual
1769 * page in a pageblock.
1770 */
1771struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1772 unsigned long end_pfn, struct zone *zone)
1773{
1774 struct page *start_page;
1775 struct page *end_page;
1776
1777 /* end_pfn is one past the range we are checking */
1778 end_pfn--;
1779
1780 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1781 return NULL;
1782
1783 start_page = pfn_to_online_page(start_pfn);
1784 if (!start_page)
1785 return NULL;
1786
1787 if (page_zone(start_page) != zone)
1788 return NULL;
1789
1790 end_page = pfn_to_page(end_pfn);
1791
1792 /* This gives a shorter code than deriving page_zone(end_page) */
1793 if (page_zone_id(start_page) != page_zone_id(end_page))
1794 return NULL;
1795
1796 return start_page;
1797}
1798
1799void set_zone_contiguous(struct zone *zone)
1800{
1801 unsigned long block_start_pfn = zone->zone_start_pfn;
1802 unsigned long block_end_pfn;
1803
1804 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1805 for (; block_start_pfn < zone_end_pfn(zone);
1806 block_start_pfn = block_end_pfn,
1807 block_end_pfn += pageblock_nr_pages) {
1808
1809 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1810
1811 if (!__pageblock_pfn_to_page(block_start_pfn,
1812 block_end_pfn, zone))
1813 return;
1814 cond_resched();
1815 }
1816
1817 /* We confirm that there is no hole */
1818 zone->contiguous = true;
1819}
1820
1821void clear_zone_contiguous(struct zone *zone)
1822{
1823 zone->contiguous = false;
1824}
1825
1826#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1827static void __init deferred_free_range(unsigned long pfn,
1828 unsigned long nr_pages)
1829{
1830 struct page *page;
1831 unsigned long i;
1832
1833 if (!nr_pages)
1834 return;
1835
1836 page = pfn_to_page(pfn);
1837
1838 /* Free a large naturally-aligned chunk if possible */
1839 if (nr_pages == pageblock_nr_pages &&
1840 (pfn & (pageblock_nr_pages - 1)) == 0) {
1841 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1842 __free_pages_core(page, pageblock_order);
1843 return;
1844 }
1845
1846 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1847 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1848 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1849 __free_pages_core(page, 0);
1850 }
1851}
1852
1853/* Completion tracking for deferred_init_memmap() threads */
1854static atomic_t pgdat_init_n_undone __initdata;
1855static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1856
1857static inline void __init pgdat_init_report_one_done(void)
1858{
1859 if (atomic_dec_and_test(&pgdat_init_n_undone))
1860 complete(&pgdat_init_all_done_comp);
1861}
1862
1863/*
1864 * Returns true if page needs to be initialized or freed to buddy allocator.
1865 *
1866 * First we check if pfn is valid on architectures where it is possible to have
1867 * holes within pageblock_nr_pages. On systems where it is not possible, this
1868 * function is optimized out.
1869 *
1870 * Then, we check if a current large page is valid by only checking the validity
1871 * of the head pfn.
1872 */
1873static inline bool __init deferred_pfn_valid(unsigned long pfn)
1874{
1875 if (!pfn_valid_within(pfn))
1876 return false;
1877 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1878 return false;
1879 return true;
1880}
1881
1882/*
1883 * Free pages to buddy allocator. Try to free aligned pages in
1884 * pageblock_nr_pages sizes.
1885 */
1886static void __init deferred_free_pages(unsigned long pfn,
1887 unsigned long end_pfn)
1888{
1889 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1890 unsigned long nr_free = 0;
1891
1892 for (; pfn < end_pfn; pfn++) {
1893 if (!deferred_pfn_valid(pfn)) {
1894 deferred_free_range(pfn - nr_free, nr_free);
1895 nr_free = 0;
1896 } else if (!(pfn & nr_pgmask)) {
1897 deferred_free_range(pfn - nr_free, nr_free);
1898 nr_free = 1;
1899 } else {
1900 nr_free++;
1901 }
1902 }
1903 /* Free the last block of pages to allocator */
1904 deferred_free_range(pfn - nr_free, nr_free);
1905}
1906
1907/*
1908 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1909 * by performing it only once every pageblock_nr_pages.
1910 * Return number of pages initialized.
1911 */
1912static unsigned long __init deferred_init_pages(struct zone *zone,
1913 unsigned long pfn,
1914 unsigned long end_pfn)
1915{
1916 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1917 int nid = zone_to_nid(zone);
1918 unsigned long nr_pages = 0;
1919 int zid = zone_idx(zone);
1920 struct page *page = NULL;
1921
1922 for (; pfn < end_pfn; pfn++) {
1923 if (!deferred_pfn_valid(pfn)) {
1924 page = NULL;
1925 continue;
1926 } else if (!page || !(pfn & nr_pgmask)) {
1927 page = pfn_to_page(pfn);
1928 } else {
1929 page++;
1930 }
1931 __init_single_page(page, pfn, zid, nid);
1932 nr_pages++;
1933 }
1934 return (nr_pages);
1935}
1936
1937/*
1938 * This function is meant to pre-load the iterator for the zone init.
1939 * Specifically it walks through the ranges until we are caught up to the
1940 * first_init_pfn value and exits there. If we never encounter the value we
1941 * return false indicating there are no valid ranges left.
1942 */
1943static bool __init
1944deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1945 unsigned long *spfn, unsigned long *epfn,
1946 unsigned long first_init_pfn)
1947{
1948 u64 j;
1949
1950 /*
1951 * Start out by walking through the ranges in this zone that have
1952 * already been initialized. We don't need to do anything with them
1953 * so we just need to flush them out of the system.
1954 */
1955 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1956 if (*epfn <= first_init_pfn)
1957 continue;
1958 if (*spfn < first_init_pfn)
1959 *spfn = first_init_pfn;
1960 *i = j;
1961 return true;
1962 }
1963
1964 return false;
1965}
1966
1967/*
1968 * Initialize and free pages. We do it in two loops: first we initialize
1969 * struct page, then free to buddy allocator, because while we are
1970 * freeing pages we can access pages that are ahead (computing buddy
1971 * page in __free_one_page()).
1972 *
1973 * In order to try and keep some memory in the cache we have the loop
1974 * broken along max page order boundaries. This way we will not cause
1975 * any issues with the buddy page computation.
1976 */
1977static unsigned long __init
1978deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1979 unsigned long *end_pfn)
1980{
1981 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1982 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1983 unsigned long nr_pages = 0;
1984 u64 j = *i;
1985
1986 /* First we loop through and initialize the page values */
1987 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1988 unsigned long t;
1989
1990 if (mo_pfn <= *start_pfn)
1991 break;
1992
1993 t = min(mo_pfn, *end_pfn);
1994 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1995
1996 if (mo_pfn < *end_pfn) {
1997 *start_pfn = mo_pfn;
1998 break;
1999 }
2000 }
2001
2002 /* Reset values and now loop through freeing pages as needed */
2003 swap(j, *i);
2004
2005 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2006 unsigned long t;
2007
2008 if (mo_pfn <= spfn)
2009 break;
2010
2011 t = min(mo_pfn, epfn);
2012 deferred_free_pages(spfn, t);
2013
2014 if (mo_pfn <= epfn)
2015 break;
2016 }
2017
2018 return nr_pages;
2019}
2020
2021static void __init
2022deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2023 void *arg)
2024{
2025 unsigned long spfn, epfn;
2026 struct zone *zone = arg;
2027 u64 i;
2028
2029 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2030
2031 /*
2032 * Initialize and free pages in MAX_ORDER sized increments so that we
2033 * can avoid introducing any issues with the buddy allocator.
2034 */
2035 while (spfn < end_pfn) {
2036 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2037 cond_resched();
2038 }
2039}
2040
2041/* An arch may override for more concurrency. */
2042__weak int __init
2043deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2044{
2045 return 1;
2046}
2047
2048/* Initialise remaining memory on a node */
2049static int __init deferred_init_memmap(void *data)
2050{
2051 pg_data_t *pgdat = data;
2052 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2053 unsigned long spfn = 0, epfn = 0;
2054 unsigned long first_init_pfn, flags;
2055 unsigned long start = jiffies;
2056 struct zone *zone;
2057 int zid, max_threads;
2058 u64 i;
2059
2060 /* Bind memory initialisation thread to a local node if possible */
2061 if (!cpumask_empty(cpumask))
2062 set_cpus_allowed_ptr(current, cpumask);
2063
2064 pgdat_resize_lock(pgdat, &flags);
2065 first_init_pfn = pgdat->first_deferred_pfn;
2066 if (first_init_pfn == ULONG_MAX) {
2067 pgdat_resize_unlock(pgdat, &flags);
2068 pgdat_init_report_one_done();
2069 return 0;
2070 }
2071
2072 /* Sanity check boundaries */
2073 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2074 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2075 pgdat->first_deferred_pfn = ULONG_MAX;
2076
2077 /*
2078 * Once we unlock here, the zone cannot be grown anymore, thus if an
2079 * interrupt thread must allocate this early in boot, zone must be
2080 * pre-grown prior to start of deferred page initialization.
2081 */
2082 pgdat_resize_unlock(pgdat, &flags);
2083
2084 /* Only the highest zone is deferred so find it */
2085 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2086 zone = pgdat->node_zones + zid;
2087 if (first_init_pfn < zone_end_pfn(zone))
2088 break;
2089 }
2090
2091 /* If the zone is empty somebody else may have cleared out the zone */
2092 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2093 first_init_pfn))
2094 goto zone_empty;
2095
2096 max_threads = deferred_page_init_max_threads(cpumask);
2097
2098 while (spfn < epfn) {
2099 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2100 struct padata_mt_job job = {
2101 .thread_fn = deferred_init_memmap_chunk,
2102 .fn_arg = zone,
2103 .start = spfn,
2104 .size = epfn_align - spfn,
2105 .align = PAGES_PER_SECTION,
2106 .min_chunk = PAGES_PER_SECTION,
2107 .max_threads = max_threads,
2108 };
2109
2110 padata_do_multithreaded(&job);
2111 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2112 epfn_align);
2113 }
2114zone_empty:
2115 /* Sanity check that the next zone really is unpopulated */
2116 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2117
2118 pr_info("node %d deferred pages initialised in %ums\n",
2119 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2120
2121 pgdat_init_report_one_done();
2122 return 0;
2123}
2124
2125/*
2126 * If this zone has deferred pages, try to grow it by initializing enough
2127 * deferred pages to satisfy the allocation specified by order, rounded up to
2128 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2129 * of SECTION_SIZE bytes by initializing struct pages in increments of
2130 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2131 *
2132 * Return true when zone was grown, otherwise return false. We return true even
2133 * when we grow less than requested, to let the caller decide if there are
2134 * enough pages to satisfy the allocation.
2135 *
2136 * Note: We use noinline because this function is needed only during boot, and
2137 * it is called from a __ref function _deferred_grow_zone. This way we are
2138 * making sure that it is not inlined into permanent text section.
2139 */
2140static noinline bool __init
2141deferred_grow_zone(struct zone *zone, unsigned int order)
2142{
2143 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2144 pg_data_t *pgdat = zone->zone_pgdat;
2145 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2146 unsigned long spfn, epfn, flags;
2147 unsigned long nr_pages = 0;
2148 u64 i;
2149
2150 /* Only the last zone may have deferred pages */
2151 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2152 return false;
2153
2154 pgdat_resize_lock(pgdat, &flags);
2155
2156 /*
2157 * If someone grew this zone while we were waiting for spinlock, return
2158 * true, as there might be enough pages already.
2159 */
2160 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2161 pgdat_resize_unlock(pgdat, &flags);
2162 return true;
2163 }
2164
2165 /* If the zone is empty somebody else may have cleared out the zone */
2166 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2167 first_deferred_pfn)) {
2168 pgdat->first_deferred_pfn = ULONG_MAX;
2169 pgdat_resize_unlock(pgdat, &flags);
2170 /* Retry only once. */
2171 return first_deferred_pfn != ULONG_MAX;
2172 }
2173
2174 /*
2175 * Initialize and free pages in MAX_ORDER sized increments so
2176 * that we can avoid introducing any issues with the buddy
2177 * allocator.
2178 */
2179 while (spfn < epfn) {
2180 /* update our first deferred PFN for this section */
2181 first_deferred_pfn = spfn;
2182
2183 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2184 touch_nmi_watchdog();
2185
2186 /* We should only stop along section boundaries */
2187 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2188 continue;
2189
2190 /* If our quota has been met we can stop here */
2191 if (nr_pages >= nr_pages_needed)
2192 break;
2193 }
2194
2195 pgdat->first_deferred_pfn = spfn;
2196 pgdat_resize_unlock(pgdat, &flags);
2197
2198 return nr_pages > 0;
2199}
2200
2201/*
2202 * deferred_grow_zone() is __init, but it is called from
2203 * get_page_from_freelist() during early boot until deferred_pages permanently
2204 * disables this call. This is why we have refdata wrapper to avoid warning,
2205 * and to ensure that the function body gets unloaded.
2206 */
2207static bool __ref
2208_deferred_grow_zone(struct zone *zone, unsigned int order)
2209{
2210 return deferred_grow_zone(zone, order);
2211}
2212
2213#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2214
2215void __init page_alloc_init_late(void)
2216{
2217 struct zone *zone;
2218 int nid;
2219
2220#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2221
2222 /* There will be num_node_state(N_MEMORY) threads */
2223 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2224 for_each_node_state(nid, N_MEMORY) {
2225 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2226 }
2227
2228 /* Block until all are initialised */
2229 wait_for_completion(&pgdat_init_all_done_comp);
2230
2231 /*
2232 * We initialized the rest of the deferred pages. Permanently disable
2233 * on-demand struct page initialization.
2234 */
2235 static_branch_disable(&deferred_pages);
2236
2237 /* Reinit limits that are based on free pages after the kernel is up */
2238 files_maxfiles_init();
2239#endif
2240
2241 buffer_init();
2242
2243 /* Discard memblock private memory */
2244 memblock_discard();
2245
2246 for_each_node_state(nid, N_MEMORY)
2247 shuffle_free_memory(NODE_DATA(nid));
2248
2249 for_each_populated_zone(zone)
2250 set_zone_contiguous(zone);
2251}
2252
2253#ifdef CONFIG_CMA
2254/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2255void __init init_cma_reserved_pageblock(struct page *page)
2256{
2257 unsigned i = pageblock_nr_pages;
2258 struct page *p = page;
2259
2260 do {
2261 __ClearPageReserved(p);
2262 set_page_count(p, 0);
2263 } while (++p, --i);
2264
2265 set_pageblock_migratetype(page, MIGRATE_CMA);
2266
2267 if (pageblock_order >= MAX_ORDER) {
2268 i = pageblock_nr_pages;
2269 p = page;
2270 do {
2271 set_page_refcounted(p);
2272 __free_pages(p, MAX_ORDER - 1);
2273 p += MAX_ORDER_NR_PAGES;
2274 } while (i -= MAX_ORDER_NR_PAGES);
2275 } else {
2276 set_page_refcounted(page);
2277 __free_pages(page, pageblock_order);
2278 }
2279
2280 adjust_managed_page_count(page, pageblock_nr_pages);
2281 page_zone(page)->cma_pages += pageblock_nr_pages;
2282}
2283#endif
2284
2285/*
2286 * The order of subdivision here is critical for the IO subsystem.
2287 * Please do not alter this order without good reasons and regression
2288 * testing. Specifically, as large blocks of memory are subdivided,
2289 * the order in which smaller blocks are delivered depends on the order
2290 * they're subdivided in this function. This is the primary factor
2291 * influencing the order in which pages are delivered to the IO
2292 * subsystem according to empirical testing, and this is also justified
2293 * by considering the behavior of a buddy system containing a single
2294 * large block of memory acted on by a series of small allocations.
2295 * This behavior is a critical factor in sglist merging's success.
2296 *
2297 * -- nyc
2298 */
2299static inline void expand(struct zone *zone, struct page *page,
2300 int low, int high, int migratetype)
2301{
2302 unsigned long size = 1 << high;
2303
2304 while (high > low) {
2305 high--;
2306 size >>= 1;
2307 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2308
2309 /*
2310 * Mark as guard pages (or page), that will allow to
2311 * merge back to allocator when buddy will be freed.
2312 * Corresponding page table entries will not be touched,
2313 * pages will stay not present in virtual address space
2314 */
2315 if (set_page_guard(zone, &page[size], high, migratetype))
2316 continue;
2317
2318 add_to_free_list(&page[size], zone, high, migratetype);
2319 set_buddy_order(&page[size], high);
2320 }
2321}
2322
2323static void check_new_page_bad(struct page *page)
2324{
2325 if (unlikely(page->flags & __PG_HWPOISON)) {
2326 /* Don't complain about hwpoisoned pages */
2327 page_mapcount_reset(page); /* remove PageBuddy */
2328 return;
2329 }
2330
2331 bad_page(page,
2332 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2333}
2334
2335/*
2336 * This page is about to be returned from the page allocator
2337 */
2338static inline int check_new_page(struct page *page)
2339{
2340 if (likely(page_expected_state(page,
2341 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2342 return 0;
2343
2344 check_new_page_bad(page);
2345 return 1;
2346}
2347
2348#ifdef CONFIG_DEBUG_VM
2349/*
2350 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2351 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2352 * also checked when pcp lists are refilled from the free lists.
2353 */
2354static inline bool check_pcp_refill(struct page *page)
2355{
2356 if (debug_pagealloc_enabled_static())
2357 return check_new_page(page);
2358 else
2359 return false;
2360}
2361
2362static inline bool check_new_pcp(struct page *page)
2363{
2364 return check_new_page(page);
2365}
2366#else
2367/*
2368 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2369 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2370 * enabled, they are also checked when being allocated from the pcp lists.
2371 */
2372static inline bool check_pcp_refill(struct page *page)
2373{
2374 return check_new_page(page);
2375}
2376static inline bool check_new_pcp(struct page *page)
2377{
2378 if (debug_pagealloc_enabled_static())
2379 return check_new_page(page);
2380 else
2381 return false;
2382}
2383#endif /* CONFIG_DEBUG_VM */
2384
2385static bool check_new_pages(struct page *page, unsigned int order)
2386{
2387 int i;
2388 for (i = 0; i < (1 << order); i++) {
2389 struct page *p = page + i;
2390
2391 if (unlikely(check_new_page(p)))
2392 return true;
2393 }
2394
2395 return false;
2396}
2397
2398inline void post_alloc_hook(struct page *page, unsigned int order,
2399 gfp_t gfp_flags)
2400{
2401 set_page_private(page, 0);
2402 set_page_refcounted(page);
2403
2404 arch_alloc_page(page, order);
2405 debug_pagealloc_map_pages(page, 1 << order);
2406
2407 /*
2408 * Page unpoisoning must happen before memory initialization.
2409 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2410 * allocations and the page unpoisoning code will complain.
2411 */
2412 kernel_unpoison_pages(page, 1 << order);
2413
2414 /*
2415 * As memory initialization might be integrated into KASAN,
2416 * kasan_alloc_pages and kernel_init_free_pages must be
2417 * kept together to avoid discrepancies in behavior.
2418 */
2419 if (kasan_has_integrated_init()) {
2420 kasan_alloc_pages(page, order, gfp_flags);
2421 } else {
2422 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2423
2424 kasan_unpoison_pages(page, order, init);
2425 if (init)
2426 kernel_init_free_pages(page, 1 << order,
2427 gfp_flags & __GFP_ZEROTAGS);
2428 }
2429
2430 set_page_owner(page, order, gfp_flags);
2431}
2432
2433static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2434 unsigned int alloc_flags)
2435{
2436 post_alloc_hook(page, order, gfp_flags);
2437
2438 if (order && (gfp_flags & __GFP_COMP))
2439 prep_compound_page(page, order);
2440
2441 /*
2442 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2443 * allocate the page. The expectation is that the caller is taking
2444 * steps that will free more memory. The caller should avoid the page
2445 * being used for !PFMEMALLOC purposes.
2446 */
2447 if (alloc_flags & ALLOC_NO_WATERMARKS)
2448 set_page_pfmemalloc(page);
2449 else
2450 clear_page_pfmemalloc(page);
2451}
2452
2453/*
2454 * Go through the free lists for the given migratetype and remove
2455 * the smallest available page from the freelists
2456 */
2457static __always_inline
2458struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2459 int migratetype)
2460{
2461 unsigned int current_order;
2462 struct free_area *area;
2463 struct page *page;
2464
2465 /* Find a page of the appropriate size in the preferred list */
2466 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2467 area = &(zone->free_area[current_order]);
2468 page = get_page_from_free_area(area, migratetype);
2469 if (!page)
2470 continue;
2471 del_page_from_free_list(page, zone, current_order);
2472 expand(zone, page, order, current_order, migratetype);
2473 set_pcppage_migratetype(page, migratetype);
2474 return page;
2475 }
2476
2477 return NULL;
2478}
2479
2480
2481/*
2482 * This array describes the order lists are fallen back to when
2483 * the free lists for the desirable migrate type are depleted
2484 */
2485static int fallbacks[MIGRATE_TYPES][3] = {
2486 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2487 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2488 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2489#ifdef CONFIG_CMA
2490 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2491#endif
2492#ifdef CONFIG_MEMORY_ISOLATION
2493 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2494#endif
2495};
2496
2497#ifdef CONFIG_CMA
2498static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2499 unsigned int order)
2500{
2501 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2502}
2503#else
2504static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2505 unsigned int order) { return NULL; }
2506#endif
2507
2508/*
2509 * Move the free pages in a range to the freelist tail of the requested type.
2510 * Note that start_page and end_pages are not aligned on a pageblock
2511 * boundary. If alignment is required, use move_freepages_block()
2512 */
2513static int move_freepages(struct zone *zone,
2514 unsigned long start_pfn, unsigned long end_pfn,
2515 int migratetype, int *num_movable)
2516{
2517 struct page *page;
2518 unsigned long pfn;
2519 unsigned int order;
2520 int pages_moved = 0;
2521
2522 for (pfn = start_pfn; pfn <= end_pfn;) {
2523 if (!pfn_valid_within(pfn)) {
2524 pfn++;
2525 continue;
2526 }
2527
2528 page = pfn_to_page(pfn);
2529 if (!PageBuddy(page)) {
2530 /*
2531 * We assume that pages that could be isolated for
2532 * migration are movable. But we don't actually try
2533 * isolating, as that would be expensive.
2534 */
2535 if (num_movable &&
2536 (PageLRU(page) || __PageMovable(page)))
2537 (*num_movable)++;
2538 pfn++;
2539 continue;
2540 }
2541
2542 /* Make sure we are not inadvertently changing nodes */
2543 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2544 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2545
2546 order = buddy_order(page);
2547 move_to_free_list(page, zone, order, migratetype);
2548 pfn += 1 << order;
2549 pages_moved += 1 << order;
2550 }
2551
2552 return pages_moved;
2553}
2554
2555int move_freepages_block(struct zone *zone, struct page *page,
2556 int migratetype, int *num_movable)
2557{
2558 unsigned long start_pfn, end_pfn, pfn;
2559
2560 if (num_movable)
2561 *num_movable = 0;
2562
2563 pfn = page_to_pfn(page);
2564 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2565 end_pfn = start_pfn + pageblock_nr_pages - 1;
2566
2567 /* Do not cross zone boundaries */
2568 if (!zone_spans_pfn(zone, start_pfn))
2569 start_pfn = pfn;
2570 if (!zone_spans_pfn(zone, end_pfn))
2571 return 0;
2572
2573 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2574 num_movable);
2575}
2576
2577static void change_pageblock_range(struct page *pageblock_page,
2578 int start_order, int migratetype)
2579{
2580 int nr_pageblocks = 1 << (start_order - pageblock_order);
2581
2582 while (nr_pageblocks--) {
2583 set_pageblock_migratetype(pageblock_page, migratetype);
2584 pageblock_page += pageblock_nr_pages;
2585 }
2586}
2587
2588/*
2589 * When we are falling back to another migratetype during allocation, try to
2590 * steal extra free pages from the same pageblocks to satisfy further
2591 * allocations, instead of polluting multiple pageblocks.
2592 *
2593 * If we are stealing a relatively large buddy page, it is likely there will
2594 * be more free pages in the pageblock, so try to steal them all. For
2595 * reclaimable and unmovable allocations, we steal regardless of page size,
2596 * as fragmentation caused by those allocations polluting movable pageblocks
2597 * is worse than movable allocations stealing from unmovable and reclaimable
2598 * pageblocks.
2599 */
2600static bool can_steal_fallback(unsigned int order, int start_mt)
2601{
2602 /*
2603 * Leaving this order check is intended, although there is
2604 * relaxed order check in next check. The reason is that
2605 * we can actually steal whole pageblock if this condition met,
2606 * but, below check doesn't guarantee it and that is just heuristic
2607 * so could be changed anytime.
2608 */
2609 if (order >= pageblock_order)
2610 return true;
2611
2612 if (order >= pageblock_order / 2 ||
2613 start_mt == MIGRATE_RECLAIMABLE ||
2614 start_mt == MIGRATE_UNMOVABLE ||
2615 page_group_by_mobility_disabled)
2616 return true;
2617
2618 return false;
2619}
2620
2621static inline bool boost_watermark(struct zone *zone)
2622{
2623 unsigned long max_boost;
2624
2625 if (!watermark_boost_factor)
2626 return false;
2627 /*
2628 * Don't bother in zones that are unlikely to produce results.
2629 * On small machines, including kdump capture kernels running
2630 * in a small area, boosting the watermark can cause an out of
2631 * memory situation immediately.
2632 */
2633 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2634 return false;
2635
2636 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2637 watermark_boost_factor, 10000);
2638
2639 /*
2640 * high watermark may be uninitialised if fragmentation occurs
2641 * very early in boot so do not boost. We do not fall
2642 * through and boost by pageblock_nr_pages as failing
2643 * allocations that early means that reclaim is not going
2644 * to help and it may even be impossible to reclaim the
2645 * boosted watermark resulting in a hang.
2646 */
2647 if (!max_boost)
2648 return false;
2649
2650 max_boost = max(pageblock_nr_pages, max_boost);
2651
2652 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2653 max_boost);
2654
2655 return true;
2656}
2657
2658/*
2659 * This function implements actual steal behaviour. If order is large enough,
2660 * we can steal whole pageblock. If not, we first move freepages in this
2661 * pageblock to our migratetype and determine how many already-allocated pages
2662 * are there in the pageblock with a compatible migratetype. If at least half
2663 * of pages are free or compatible, we can change migratetype of the pageblock
2664 * itself, so pages freed in the future will be put on the correct free list.
2665 */
2666static void steal_suitable_fallback(struct zone *zone, struct page *page,
2667 unsigned int alloc_flags, int start_type, bool whole_block)
2668{
2669 unsigned int current_order = buddy_order(page);
2670 int free_pages, movable_pages, alike_pages;
2671 int old_block_type;
2672
2673 old_block_type = get_pageblock_migratetype(page);
2674
2675 /*
2676 * This can happen due to races and we want to prevent broken
2677 * highatomic accounting.
2678 */
2679 if (is_migrate_highatomic(old_block_type))
2680 goto single_page;
2681
2682 /* Take ownership for orders >= pageblock_order */
2683 if (current_order >= pageblock_order) {
2684 change_pageblock_range(page, current_order, start_type);
2685 goto single_page;
2686 }
2687
2688 /*
2689 * Boost watermarks to increase reclaim pressure to reduce the
2690 * likelihood of future fallbacks. Wake kswapd now as the node
2691 * may be balanced overall and kswapd will not wake naturally.
2692 */
2693 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2694 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2695
2696 /* We are not allowed to try stealing from the whole block */
2697 if (!whole_block)
2698 goto single_page;
2699
2700 free_pages = move_freepages_block(zone, page, start_type,
2701 &movable_pages);
2702 /*
2703 * Determine how many pages are compatible with our allocation.
2704 * For movable allocation, it's the number of movable pages which
2705 * we just obtained. For other types it's a bit more tricky.
2706 */
2707 if (start_type == MIGRATE_MOVABLE) {
2708 alike_pages = movable_pages;
2709 } else {
2710 /*
2711 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2712 * to MOVABLE pageblock, consider all non-movable pages as
2713 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2714 * vice versa, be conservative since we can't distinguish the
2715 * exact migratetype of non-movable pages.
2716 */
2717 if (old_block_type == MIGRATE_MOVABLE)
2718 alike_pages = pageblock_nr_pages
2719 - (free_pages + movable_pages);
2720 else
2721 alike_pages = 0;
2722 }
2723
2724 /* moving whole block can fail due to zone boundary conditions */
2725 if (!free_pages)
2726 goto single_page;
2727
2728 /*
2729 * If a sufficient number of pages in the block are either free or of
2730 * comparable migratability as our allocation, claim the whole block.
2731 */
2732 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2733 page_group_by_mobility_disabled)
2734 set_pageblock_migratetype(page, start_type);
2735
2736 return;
2737
2738single_page:
2739 move_to_free_list(page, zone, current_order, start_type);
2740}
2741
2742/*
2743 * Check whether there is a suitable fallback freepage with requested order.
2744 * If only_stealable is true, this function returns fallback_mt only if
2745 * we can steal other freepages all together. This would help to reduce
2746 * fragmentation due to mixed migratetype pages in one pageblock.
2747 */
2748int find_suitable_fallback(struct free_area *area, unsigned int order,
2749 int migratetype, bool only_stealable, bool *can_steal)
2750{
2751 int i;
2752 int fallback_mt;
2753
2754 if (area->nr_free == 0)
2755 return -1;
2756
2757 *can_steal = false;
2758 for (i = 0;; i++) {
2759 fallback_mt = fallbacks[migratetype][i];
2760 if (fallback_mt == MIGRATE_TYPES)
2761 break;
2762
2763 if (free_area_empty(area, fallback_mt))
2764 continue;
2765
2766 if (can_steal_fallback(order, migratetype))
2767 *can_steal = true;
2768
2769 if (!only_stealable)
2770 return fallback_mt;
2771
2772 if (*can_steal)
2773 return fallback_mt;
2774 }
2775
2776 return -1;
2777}
2778
2779/*
2780 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2781 * there are no empty page blocks that contain a page with a suitable order
2782 */
2783static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2784 unsigned int alloc_order)
2785{
2786 int mt;
2787 unsigned long max_managed, flags;
2788
2789 /*
2790 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2791 * Check is race-prone but harmless.
2792 */
2793 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2794 if (zone->nr_reserved_highatomic >= max_managed)
2795 return;
2796
2797 spin_lock_irqsave(&zone->lock, flags);
2798
2799 /* Recheck the nr_reserved_highatomic limit under the lock */
2800 if (zone->nr_reserved_highatomic >= max_managed)
2801 goto out_unlock;
2802
2803 /* Yoink! */
2804 mt = get_pageblock_migratetype(page);
2805 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2806 && !is_migrate_cma(mt)) {
2807 zone->nr_reserved_highatomic += pageblock_nr_pages;
2808 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2809 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2810 }
2811
2812out_unlock:
2813 spin_unlock_irqrestore(&zone->lock, flags);
2814}
2815
2816/*
2817 * Used when an allocation is about to fail under memory pressure. This
2818 * potentially hurts the reliability of high-order allocations when under
2819 * intense memory pressure but failed atomic allocations should be easier
2820 * to recover from than an OOM.
2821 *
2822 * If @force is true, try to unreserve a pageblock even though highatomic
2823 * pageblock is exhausted.
2824 */
2825static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2826 bool force)
2827{
2828 struct zonelist *zonelist = ac->zonelist;
2829 unsigned long flags;
2830 struct zoneref *z;
2831 struct zone *zone;
2832 struct page *page;
2833 int order;
2834 bool ret;
2835
2836 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2837 ac->nodemask) {
2838 /*
2839 * Preserve at least one pageblock unless memory pressure
2840 * is really high.
2841 */
2842 if (!force && zone->nr_reserved_highatomic <=
2843 pageblock_nr_pages)
2844 continue;
2845
2846 spin_lock_irqsave(&zone->lock, flags);
2847 for (order = 0; order < MAX_ORDER; order++) {
2848 struct free_area *area = &(zone->free_area[order]);
2849
2850 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2851 if (!page)
2852 continue;
2853
2854 /*
2855 * In page freeing path, migratetype change is racy so
2856 * we can counter several free pages in a pageblock
2857 * in this loop although we changed the pageblock type
2858 * from highatomic to ac->migratetype. So we should
2859 * adjust the count once.
2860 */
2861 if (is_migrate_highatomic_page(page)) {
2862 /*
2863 * It should never happen but changes to
2864 * locking could inadvertently allow a per-cpu
2865 * drain to add pages to MIGRATE_HIGHATOMIC
2866 * while unreserving so be safe and watch for
2867 * underflows.
2868 */
2869 zone->nr_reserved_highatomic -= min(
2870 pageblock_nr_pages,
2871 zone->nr_reserved_highatomic);
2872 }
2873
2874 /*
2875 * Convert to ac->migratetype and avoid the normal
2876 * pageblock stealing heuristics. Minimally, the caller
2877 * is doing the work and needs the pages. More
2878 * importantly, if the block was always converted to
2879 * MIGRATE_UNMOVABLE or another type then the number
2880 * of pageblocks that cannot be completely freed
2881 * may increase.
2882 */
2883 set_pageblock_migratetype(page, ac->migratetype);
2884 ret = move_freepages_block(zone, page, ac->migratetype,
2885 NULL);
2886 if (ret) {
2887 spin_unlock_irqrestore(&zone->lock, flags);
2888 return ret;
2889 }
2890 }
2891 spin_unlock_irqrestore(&zone->lock, flags);
2892 }
2893
2894 return false;
2895}
2896
2897/*
2898 * Try finding a free buddy page on the fallback list and put it on the free
2899 * list of requested migratetype, possibly along with other pages from the same
2900 * block, depending on fragmentation avoidance heuristics. Returns true if
2901 * fallback was found so that __rmqueue_smallest() can grab it.
2902 *
2903 * The use of signed ints for order and current_order is a deliberate
2904 * deviation from the rest of this file, to make the for loop
2905 * condition simpler.
2906 */
2907static __always_inline bool
2908__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2909 unsigned int alloc_flags)
2910{
2911 struct free_area *area;
2912 int current_order;
2913 int min_order = order;
2914 struct page *page;
2915 int fallback_mt;
2916 bool can_steal;
2917
2918 /*
2919 * Do not steal pages from freelists belonging to other pageblocks
2920 * i.e. orders < pageblock_order. If there are no local zones free,
2921 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2922 */
2923 if (alloc_flags & ALLOC_NOFRAGMENT)
2924 min_order = pageblock_order;
2925
2926 /*
2927 * Find the largest available free page in the other list. This roughly
2928 * approximates finding the pageblock with the most free pages, which
2929 * would be too costly to do exactly.
2930 */
2931 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2932 --current_order) {
2933 area = &(zone->free_area[current_order]);
2934 fallback_mt = find_suitable_fallback(area, current_order,
2935 start_migratetype, false, &can_steal);
2936 if (fallback_mt == -1)
2937 continue;
2938
2939 /*
2940 * We cannot steal all free pages from the pageblock and the
2941 * requested migratetype is movable. In that case it's better to
2942 * steal and split the smallest available page instead of the
2943 * largest available page, because even if the next movable
2944 * allocation falls back into a different pageblock than this
2945 * one, it won't cause permanent fragmentation.
2946 */
2947 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2948 && current_order > order)
2949 goto find_smallest;
2950
2951 goto do_steal;
2952 }
2953
2954 return false;
2955
2956find_smallest:
2957 for (current_order = order; current_order < MAX_ORDER;
2958 current_order++) {
2959 area = &(zone->free_area[current_order]);
2960 fallback_mt = find_suitable_fallback(area, current_order,
2961 start_migratetype, false, &can_steal);
2962 if (fallback_mt != -1)
2963 break;
2964 }
2965
2966 /*
2967 * This should not happen - we already found a suitable fallback
2968 * when looking for the largest page.
2969 */
2970 VM_BUG_ON(current_order == MAX_ORDER);
2971
2972do_steal:
2973 page = get_page_from_free_area(area, fallback_mt);
2974
2975 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2976 can_steal);
2977
2978 trace_mm_page_alloc_extfrag(page, order, current_order,
2979 start_migratetype, fallback_mt);
2980
2981 return true;
2982
2983}
2984
2985/*
2986 * Do the hard work of removing an element from the buddy allocator.
2987 * Call me with the zone->lock already held.
2988 */
2989static __always_inline struct page *
2990__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2991 unsigned int alloc_flags)
2992{
2993 struct page *page;
2994
2995 if (IS_ENABLED(CONFIG_CMA)) {
2996 /*
2997 * Balance movable allocations between regular and CMA areas by
2998 * allocating from CMA when over half of the zone's free memory
2999 * is in the CMA area.
3000 */
3001 if (alloc_flags & ALLOC_CMA &&
3002 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3003 zone_page_state(zone, NR_FREE_PAGES) / 2) {
3004 page = __rmqueue_cma_fallback(zone, order);
3005 if (page)
3006 goto out;
3007 }
3008 }
3009retry:
3010 page = __rmqueue_smallest(zone, order, migratetype);
3011 if (unlikely(!page)) {
3012 if (alloc_flags & ALLOC_CMA)
3013 page = __rmqueue_cma_fallback(zone, order);
3014
3015 if (!page && __rmqueue_fallback(zone, order, migratetype,
3016 alloc_flags))
3017 goto retry;
3018 }
3019out:
3020 if (page)
3021 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3022 return page;
3023}
3024
3025/*
3026 * Obtain a specified number of elements from the buddy allocator, all under
3027 * a single hold of the lock, for efficiency. Add them to the supplied list.
3028 * Returns the number of new pages which were placed at *list.
3029 */
3030static int rmqueue_bulk(struct zone *zone, unsigned int order,
3031 unsigned long count, struct list_head *list,
3032 int migratetype, unsigned int alloc_flags)
3033{
3034 int i, allocated = 0;
3035
3036 /*
3037 * local_lock_irq held so equivalent to spin_lock_irqsave for
3038 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3039 */
3040 spin_lock(&zone->lock);
3041 for (i = 0; i < count; ++i) {
3042 struct page *page = __rmqueue(zone, order, migratetype,
3043 alloc_flags);
3044 if (unlikely(page == NULL))
3045 break;
3046
3047 if (unlikely(check_pcp_refill(page)))
3048 continue;
3049
3050 /*
3051 * Split buddy pages returned by expand() are received here in
3052 * physical page order. The page is added to the tail of
3053 * caller's list. From the callers perspective, the linked list
3054 * is ordered by page number under some conditions. This is
3055 * useful for IO devices that can forward direction from the
3056 * head, thus also in the physical page order. This is useful
3057 * for IO devices that can merge IO requests if the physical
3058 * pages are ordered properly.
3059 */
3060 list_add_tail(&page->lru, list);
3061 allocated++;
3062 if (is_migrate_cma(get_pcppage_migratetype(page)))
3063 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3064 -(1 << order));
3065 }
3066
3067 /*
3068 * i pages were removed from the buddy list even if some leak due
3069 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3070 * on i. Do not confuse with 'allocated' which is the number of
3071 * pages added to the pcp list.
3072 */
3073 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3074 spin_unlock(&zone->lock);
3075 return allocated;
3076}
3077
3078#ifdef CONFIG_NUMA
3079/*
3080 * Called from the vmstat counter updater to drain pagesets of this
3081 * currently executing processor on remote nodes after they have
3082 * expired.
3083 *
3084 * Note that this function must be called with the thread pinned to
3085 * a single processor.
3086 */
3087void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3088{
3089 unsigned long flags;
3090 int to_drain, batch;
3091
3092 local_lock_irqsave(&pagesets.lock, flags);
3093 batch = READ_ONCE(pcp->batch);
3094 to_drain = min(pcp->count, batch);
3095 if (to_drain > 0)
3096 free_pcppages_bulk(zone, to_drain, pcp);
3097 local_unlock_irqrestore(&pagesets.lock, flags);
3098}
3099#endif
3100
3101/*
3102 * Drain pcplists of the indicated processor and zone.
3103 *
3104 * The processor must either be the current processor and the
3105 * thread pinned to the current processor or a processor that
3106 * is not online.
3107 */
3108static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3109{
3110 unsigned long flags;
3111 struct per_cpu_pages *pcp;
3112
3113 local_lock_irqsave(&pagesets.lock, flags);
3114
3115 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3116 if (pcp->count)
3117 free_pcppages_bulk(zone, pcp->count, pcp);
3118
3119 local_unlock_irqrestore(&pagesets.lock, flags);
3120}
3121
3122/*
3123 * Drain pcplists of all zones on the indicated processor.
3124 *
3125 * The processor must either be the current processor and the
3126 * thread pinned to the current processor or a processor that
3127 * is not online.
3128 */
3129static void drain_pages(unsigned int cpu)
3130{
3131 struct zone *zone;
3132
3133 for_each_populated_zone(zone) {
3134 drain_pages_zone(cpu, zone);
3135 }
3136}
3137
3138/*
3139 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3140 *
3141 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3142 * the single zone's pages.
3143 */
3144void drain_local_pages(struct zone *zone)
3145{
3146 int cpu = smp_processor_id();
3147
3148 if (zone)
3149 drain_pages_zone(cpu, zone);
3150 else
3151 drain_pages(cpu);
3152}
3153
3154static void drain_local_pages_wq(struct work_struct *work)
3155{
3156 struct pcpu_drain *drain;
3157
3158 drain = container_of(work, struct pcpu_drain, work);
3159
3160 /*
3161 * drain_all_pages doesn't use proper cpu hotplug protection so
3162 * we can race with cpu offline when the WQ can move this from
3163 * a cpu pinned worker to an unbound one. We can operate on a different
3164 * cpu which is alright but we also have to make sure to not move to
3165 * a different one.
3166 */
3167 preempt_disable();
3168 drain_local_pages(drain->zone);
3169 preempt_enable();
3170}
3171
3172/*
3173 * The implementation of drain_all_pages(), exposing an extra parameter to
3174 * drain on all cpus.
3175 *
3176 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3177 * not empty. The check for non-emptiness can however race with a free to
3178 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3179 * that need the guarantee that every CPU has drained can disable the
3180 * optimizing racy check.
3181 */
3182static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3183{
3184 int cpu;
3185
3186 /*
3187 * Allocate in the BSS so we won't require allocation in
3188 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3189 */
3190 static cpumask_t cpus_with_pcps;
3191
3192 /*
3193 * Make sure nobody triggers this path before mm_percpu_wq is fully
3194 * initialized.
3195 */
3196 if (WARN_ON_ONCE(!mm_percpu_wq))
3197 return;
3198
3199 /*
3200 * Do not drain if one is already in progress unless it's specific to
3201 * a zone. Such callers are primarily CMA and memory hotplug and need
3202 * the drain to be complete when the call returns.
3203 */
3204 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3205 if (!zone)
3206 return;
3207 mutex_lock(&pcpu_drain_mutex);
3208 }
3209
3210 /*
3211 * We don't care about racing with CPU hotplug event
3212 * as offline notification will cause the notified
3213 * cpu to drain that CPU pcps and on_each_cpu_mask
3214 * disables preemption as part of its processing
3215 */
3216 for_each_online_cpu(cpu) {
3217 struct per_cpu_pages *pcp;
3218 struct zone *z;
3219 bool has_pcps = false;
3220
3221 if (force_all_cpus) {
3222 /*
3223 * The pcp.count check is racy, some callers need a
3224 * guarantee that no cpu is missed.
3225 */
3226 has_pcps = true;
3227 } else if (zone) {
3228 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3229 if (pcp->count)
3230 has_pcps = true;
3231 } else {
3232 for_each_populated_zone(z) {
3233 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3234 if (pcp->count) {
3235 has_pcps = true;
3236 break;
3237 }
3238 }
3239 }
3240
3241 if (has_pcps)
3242 cpumask_set_cpu(cpu, &cpus_with_pcps);
3243 else
3244 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3245 }
3246
3247 for_each_cpu(cpu, &cpus_with_pcps) {
3248 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3249
3250 drain->zone = zone;
3251 INIT_WORK(&drain->work, drain_local_pages_wq);
3252 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3253 }
3254 for_each_cpu(cpu, &cpus_with_pcps)
3255 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3256
3257 mutex_unlock(&pcpu_drain_mutex);
3258}
3259
3260/*
3261 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3262 *
3263 * When zone parameter is non-NULL, spill just the single zone's pages.
3264 *
3265 * Note that this can be extremely slow as the draining happens in a workqueue.
3266 */
3267void drain_all_pages(struct zone *zone)
3268{
3269 __drain_all_pages(zone, false);
3270}
3271
3272#ifdef CONFIG_HIBERNATION
3273
3274/*
3275 * Touch the watchdog for every WD_PAGE_COUNT pages.
3276 */
3277#define WD_PAGE_COUNT (128*1024)
3278
3279void mark_free_pages(struct zone *zone)
3280{
3281 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3282 unsigned long flags;
3283 unsigned int order, t;
3284 struct page *page;
3285
3286 if (zone_is_empty(zone))
3287 return;
3288
3289 spin_lock_irqsave(&zone->lock, flags);
3290
3291 max_zone_pfn = zone_end_pfn(zone);
3292 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3293 if (pfn_valid(pfn)) {
3294 page = pfn_to_page(pfn);
3295
3296 if (!--page_count) {
3297 touch_nmi_watchdog();
3298 page_count = WD_PAGE_COUNT;
3299 }
3300
3301 if (page_zone(page) != zone)
3302 continue;
3303
3304 if (!swsusp_page_is_forbidden(page))
3305 swsusp_unset_page_free(page);
3306 }
3307
3308 for_each_migratetype_order(order, t) {
3309 list_for_each_entry(page,
3310 &zone->free_area[order].free_list[t], lru) {
3311 unsigned long i;
3312
3313 pfn = page_to_pfn(page);
3314 for (i = 0; i < (1UL << order); i++) {
3315 if (!--page_count) {
3316 touch_nmi_watchdog();
3317 page_count = WD_PAGE_COUNT;
3318 }
3319 swsusp_set_page_free(pfn_to_page(pfn + i));
3320 }
3321 }
3322 }
3323 spin_unlock_irqrestore(&zone->lock, flags);
3324}
3325#endif /* CONFIG_PM */
3326
3327static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3328 unsigned int order)
3329{
3330 int migratetype;
3331
3332 if (!free_pcp_prepare(page, order))
3333 return false;
3334
3335 migratetype = get_pfnblock_migratetype(page, pfn);
3336 set_pcppage_migratetype(page, migratetype);
3337 return true;
3338}
3339
3340static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch)
3341{
3342 int min_nr_free, max_nr_free;
3343
3344 /* Check for PCP disabled or boot pageset */
3345 if (unlikely(high < batch))
3346 return 1;
3347
3348 /* Leave at least pcp->batch pages on the list */
3349 min_nr_free = batch;
3350 max_nr_free = high - batch;
3351
3352 /*
3353 * Double the number of pages freed each time there is subsequent
3354 * freeing of pages without any allocation.
3355 */
3356 batch <<= pcp->free_factor;
3357 if (batch < max_nr_free)
3358 pcp->free_factor++;
3359 batch = clamp(batch, min_nr_free, max_nr_free);
3360
3361 return batch;
3362}
3363
3364static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone)
3365{
3366 int high = READ_ONCE(pcp->high);
3367
3368 if (unlikely(!high))
3369 return 0;
3370
3371 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3372 return high;
3373
3374 /*
3375 * If reclaim is active, limit the number of pages that can be
3376 * stored on pcp lists
3377 */
3378 return min(READ_ONCE(pcp->batch) << 2, high);
3379}
3380
3381static void free_unref_page_commit(struct page *page, unsigned long pfn,
3382 int migratetype, unsigned int order)
3383{
3384 struct zone *zone = page_zone(page);
3385 struct per_cpu_pages *pcp;
3386 int high;
3387 int pindex;
3388
3389 __count_vm_event(PGFREE);
3390 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3391 pindex = order_to_pindex(migratetype, order);
3392 list_add(&page->lru, &pcp->lists[pindex]);
3393 pcp->count += 1 << order;
3394 high = nr_pcp_high(pcp, zone);
3395 if (pcp->count >= high) {
3396 int batch = READ_ONCE(pcp->batch);
3397
3398 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch), pcp);
3399 }
3400}
3401
3402/*
3403 * Free a pcp page
3404 */
3405void free_unref_page(struct page *page, unsigned int order)
3406{
3407 unsigned long flags;
3408 unsigned long pfn = page_to_pfn(page);
3409 int migratetype;
3410
3411 if (!free_unref_page_prepare(page, pfn, order))
3412 return;
3413
3414 /*
3415 * We only track unmovable, reclaimable and movable on pcp lists.
3416 * Place ISOLATE pages on the isolated list because they are being
3417 * offlined but treat HIGHATOMIC as movable pages so we can get those
3418 * areas back if necessary. Otherwise, we may have to free
3419 * excessively into the page allocator
3420 */
3421 migratetype = get_pcppage_migratetype(page);
3422 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3423 if (unlikely(is_migrate_isolate(migratetype))) {
3424 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3425 return;
3426 }
3427 migratetype = MIGRATE_MOVABLE;
3428 }
3429
3430 local_lock_irqsave(&pagesets.lock, flags);
3431 free_unref_page_commit(page, pfn, migratetype, order);
3432 local_unlock_irqrestore(&pagesets.lock, flags);
3433}
3434
3435/*
3436 * Free a list of 0-order pages
3437 */
3438void free_unref_page_list(struct list_head *list)
3439{
3440 struct page *page, *next;
3441 unsigned long flags, pfn;
3442 int batch_count = 0;
3443 int migratetype;
3444
3445 /* Prepare pages for freeing */
3446 list_for_each_entry_safe(page, next, list, lru) {
3447 pfn = page_to_pfn(page);
3448 if (!free_unref_page_prepare(page, pfn, 0)) {
3449 list_del(&page->lru);
3450 continue;
3451 }
3452
3453 /*
3454 * Free isolated pages directly to the allocator, see
3455 * comment in free_unref_page.
3456 */
3457 migratetype = get_pcppage_migratetype(page);
3458 if (unlikely(is_migrate_isolate(migratetype))) {
3459 list_del(&page->lru);
3460 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3461 continue;
3462 }
3463
3464 set_page_private(page, pfn);
3465 }
3466
3467 local_lock_irqsave(&pagesets.lock, flags);
3468 list_for_each_entry_safe(page, next, list, lru) {
3469 pfn = page_private(page);
3470 set_page_private(page, 0);
3471
3472 /*
3473 * Non-isolated types over MIGRATE_PCPTYPES get added
3474 * to the MIGRATE_MOVABLE pcp list.
3475 */
3476 migratetype = get_pcppage_migratetype(page);
3477 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3478 migratetype = MIGRATE_MOVABLE;
3479
3480 trace_mm_page_free_batched(page);
3481 free_unref_page_commit(page, pfn, migratetype, 0);
3482
3483 /*
3484 * Guard against excessive IRQ disabled times when we get
3485 * a large list of pages to free.
3486 */
3487 if (++batch_count == SWAP_CLUSTER_MAX) {
3488 local_unlock_irqrestore(&pagesets.lock, flags);
3489 batch_count = 0;
3490 local_lock_irqsave(&pagesets.lock, flags);
3491 }
3492 }
3493 local_unlock_irqrestore(&pagesets.lock, flags);
3494}
3495
3496/*
3497 * split_page takes a non-compound higher-order page, and splits it into
3498 * n (1<<order) sub-pages: page[0..n]
3499 * Each sub-page must be freed individually.
3500 *
3501 * Note: this is probably too low level an operation for use in drivers.
3502 * Please consult with lkml before using this in your driver.
3503 */
3504void split_page(struct page *page, unsigned int order)
3505{
3506 int i;
3507
3508 VM_BUG_ON_PAGE(PageCompound(page), page);
3509 VM_BUG_ON_PAGE(!page_count(page), page);
3510
3511 for (i = 1; i < (1 << order); i++)
3512 set_page_refcounted(page + i);
3513 split_page_owner(page, 1 << order);
3514 split_page_memcg(page, 1 << order);
3515}
3516EXPORT_SYMBOL_GPL(split_page);
3517
3518int __isolate_free_page(struct page *page, unsigned int order)
3519{
3520 unsigned long watermark;
3521 struct zone *zone;
3522 int mt;
3523
3524 BUG_ON(!PageBuddy(page));
3525
3526 zone = page_zone(page);
3527 mt = get_pageblock_migratetype(page);
3528
3529 if (!is_migrate_isolate(mt)) {
3530 /*
3531 * Obey watermarks as if the page was being allocated. We can
3532 * emulate a high-order watermark check with a raised order-0
3533 * watermark, because we already know our high-order page
3534 * exists.
3535 */
3536 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3537 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3538 return 0;
3539
3540 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3541 }
3542
3543 /* Remove page from free list */
3544
3545 del_page_from_free_list(page, zone, order);
3546
3547 /*
3548 * Set the pageblock if the isolated page is at least half of a
3549 * pageblock
3550 */
3551 if (order >= pageblock_order - 1) {
3552 struct page *endpage = page + (1 << order) - 1;
3553 for (; page < endpage; page += pageblock_nr_pages) {
3554 int mt = get_pageblock_migratetype(page);
3555 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3556 && !is_migrate_highatomic(mt))
3557 set_pageblock_migratetype(page,
3558 MIGRATE_MOVABLE);
3559 }
3560 }
3561
3562
3563 return 1UL << order;
3564}
3565
3566/**
3567 * __putback_isolated_page - Return a now-isolated page back where we got it
3568 * @page: Page that was isolated
3569 * @order: Order of the isolated page
3570 * @mt: The page's pageblock's migratetype
3571 *
3572 * This function is meant to return a page pulled from the free lists via
3573 * __isolate_free_page back to the free lists they were pulled from.
3574 */
3575void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3576{
3577 struct zone *zone = page_zone(page);
3578
3579 /* zone lock should be held when this function is called */
3580 lockdep_assert_held(&zone->lock);
3581
3582 /* Return isolated page to tail of freelist. */
3583 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3584 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3585}
3586
3587/*
3588 * Update NUMA hit/miss statistics
3589 *
3590 * Must be called with interrupts disabled.
3591 */
3592static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3593 long nr_account)
3594{
3595#ifdef CONFIG_NUMA
3596 enum numa_stat_item local_stat = NUMA_LOCAL;
3597
3598 /* skip numa counters update if numa stats is disabled */
3599 if (!static_branch_likely(&vm_numa_stat_key))
3600 return;
3601
3602 if (zone_to_nid(z) != numa_node_id())
3603 local_stat = NUMA_OTHER;
3604
3605 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3606 __count_numa_events(z, NUMA_HIT, nr_account);
3607 else {
3608 __count_numa_events(z, NUMA_MISS, nr_account);
3609 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3610 }
3611 __count_numa_events(z, local_stat, nr_account);
3612#endif
3613}
3614
3615/* Remove page from the per-cpu list, caller must protect the list */
3616static inline
3617struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3618 int migratetype,
3619 unsigned int alloc_flags,
3620 struct per_cpu_pages *pcp,
3621 struct list_head *list)
3622{
3623 struct page *page;
3624
3625 do {
3626 if (list_empty(list)) {
3627 int batch = READ_ONCE(pcp->batch);
3628 int alloced;
3629
3630 /*
3631 * Scale batch relative to order if batch implies
3632 * free pages can be stored on the PCP. Batch can
3633 * be 1 for small zones or for boot pagesets which
3634 * should never store free pages as the pages may
3635 * belong to arbitrary zones.
3636 */
3637 if (batch > 1)
3638 batch = max(batch >> order, 2);
3639 alloced = rmqueue_bulk(zone, order,
3640 batch, list,
3641 migratetype, alloc_flags);
3642
3643 pcp->count += alloced << order;
3644 if (unlikely(list_empty(list)))
3645 return NULL;
3646 }
3647
3648 page = list_first_entry(list, struct page, lru);
3649 list_del(&page->lru);
3650 pcp->count -= 1 << order;
3651 } while (check_new_pcp(page));
3652
3653 return page;
3654}
3655
3656/* Lock and remove page from the per-cpu list */
3657static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3658 struct zone *zone, unsigned int order,
3659 gfp_t gfp_flags, int migratetype,
3660 unsigned int alloc_flags)
3661{
3662 struct per_cpu_pages *pcp;
3663 struct list_head *list;
3664 struct page *page;
3665 unsigned long flags;
3666
3667 local_lock_irqsave(&pagesets.lock, flags);
3668
3669 /*
3670 * On allocation, reduce the number of pages that are batch freed.
3671 * See nr_pcp_free() where free_factor is increased for subsequent
3672 * frees.
3673 */
3674 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3675 pcp->free_factor >>= 1;
3676 list = &pcp->lists[order_to_pindex(migratetype, order)];
3677 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3678 local_unlock_irqrestore(&pagesets.lock, flags);
3679 if (page) {
3680 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3681 zone_statistics(preferred_zone, zone, 1);
3682 }
3683 return page;
3684}
3685
3686/*
3687 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3688 */
3689static inline
3690struct page *rmqueue(struct zone *preferred_zone,
3691 struct zone *zone, unsigned int order,
3692 gfp_t gfp_flags, unsigned int alloc_flags,
3693 int migratetype)
3694{
3695 unsigned long flags;
3696 struct page *page;
3697
3698 if (likely(pcp_allowed_order(order))) {
3699 /*
3700 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3701 * we need to skip it when CMA area isn't allowed.
3702 */
3703 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3704 migratetype != MIGRATE_MOVABLE) {
3705 page = rmqueue_pcplist(preferred_zone, zone, order,
3706 gfp_flags, migratetype, alloc_flags);
3707 goto out;
3708 }
3709 }
3710
3711 /*
3712 * We most definitely don't want callers attempting to
3713 * allocate greater than order-1 page units with __GFP_NOFAIL.
3714 */
3715 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3716 spin_lock_irqsave(&zone->lock, flags);
3717
3718 do {
3719 page = NULL;
3720 /*
3721 * order-0 request can reach here when the pcplist is skipped
3722 * due to non-CMA allocation context. HIGHATOMIC area is
3723 * reserved for high-order atomic allocation, so order-0
3724 * request should skip it.
3725 */
3726 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3727 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3728 if (page)
3729 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3730 }
3731 if (!page)
3732 page = __rmqueue(zone, order, migratetype, alloc_flags);
3733 } while (page && check_new_pages(page, order));
3734 if (!page)
3735 goto failed;
3736
3737 __mod_zone_freepage_state(zone, -(1 << order),
3738 get_pcppage_migratetype(page));
3739 spin_unlock_irqrestore(&zone->lock, flags);
3740
3741 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3742 zone_statistics(preferred_zone, zone, 1);
3743
3744out:
3745 /* Separate test+clear to avoid unnecessary atomics */
3746 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3747 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3748 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3749 }
3750
3751 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3752 return page;
3753
3754failed:
3755 spin_unlock_irqrestore(&zone->lock, flags);
3756 return NULL;
3757}
3758
3759#ifdef CONFIG_FAIL_PAGE_ALLOC
3760
3761static struct {
3762 struct fault_attr attr;
3763
3764 bool ignore_gfp_highmem;
3765 bool ignore_gfp_reclaim;
3766 u32 min_order;
3767} fail_page_alloc = {
3768 .attr = FAULT_ATTR_INITIALIZER,
3769 .ignore_gfp_reclaim = true,
3770 .ignore_gfp_highmem = true,
3771 .min_order = 1,
3772};
3773
3774static int __init setup_fail_page_alloc(char *str)
3775{
3776 return setup_fault_attr(&fail_page_alloc.attr, str);
3777}
3778__setup("fail_page_alloc=", setup_fail_page_alloc);
3779
3780static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3781{
3782 if (order < fail_page_alloc.min_order)
3783 return false;
3784 if (gfp_mask & __GFP_NOFAIL)
3785 return false;
3786 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3787 return false;
3788 if (fail_page_alloc.ignore_gfp_reclaim &&
3789 (gfp_mask & __GFP_DIRECT_RECLAIM))
3790 return false;
3791
3792 return should_fail(&fail_page_alloc.attr, 1 << order);
3793}
3794
3795#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3796
3797static int __init fail_page_alloc_debugfs(void)
3798{
3799 umode_t mode = S_IFREG | 0600;
3800 struct dentry *dir;
3801
3802 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3803 &fail_page_alloc.attr);
3804
3805 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3806 &fail_page_alloc.ignore_gfp_reclaim);
3807 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3808 &fail_page_alloc.ignore_gfp_highmem);
3809 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3810
3811 return 0;
3812}
3813
3814late_initcall(fail_page_alloc_debugfs);
3815
3816#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3817
3818#else /* CONFIG_FAIL_PAGE_ALLOC */
3819
3820static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3821{
3822 return false;
3823}
3824
3825#endif /* CONFIG_FAIL_PAGE_ALLOC */
3826
3827noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3828{
3829 return __should_fail_alloc_page(gfp_mask, order);
3830}
3831ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3832
3833static inline long __zone_watermark_unusable_free(struct zone *z,
3834 unsigned int order, unsigned int alloc_flags)
3835{
3836 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3837 long unusable_free = (1 << order) - 1;
3838
3839 /*
3840 * If the caller does not have rights to ALLOC_HARDER then subtract
3841 * the high-atomic reserves. This will over-estimate the size of the
3842 * atomic reserve but it avoids a search.
3843 */
3844 if (likely(!alloc_harder))
3845 unusable_free += z->nr_reserved_highatomic;
3846
3847#ifdef CONFIG_CMA
3848 /* If allocation can't use CMA areas don't use free CMA pages */
3849 if (!(alloc_flags & ALLOC_CMA))
3850 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3851#endif
3852
3853 return unusable_free;
3854}
3855
3856/*
3857 * Return true if free base pages are above 'mark'. For high-order checks it
3858 * will return true of the order-0 watermark is reached and there is at least
3859 * one free page of a suitable size. Checking now avoids taking the zone lock
3860 * to check in the allocation paths if no pages are free.
3861 */
3862bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3863 int highest_zoneidx, unsigned int alloc_flags,
3864 long free_pages)
3865{
3866 long min = mark;
3867 int o;
3868 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3869
3870 /* free_pages may go negative - that's OK */
3871 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3872
3873 if (alloc_flags & ALLOC_HIGH)
3874 min -= min / 2;
3875
3876 if (unlikely(alloc_harder)) {
3877 /*
3878 * OOM victims can try even harder than normal ALLOC_HARDER
3879 * users on the grounds that it's definitely going to be in
3880 * the exit path shortly and free memory. Any allocation it
3881 * makes during the free path will be small and short-lived.
3882 */
3883 if (alloc_flags & ALLOC_OOM)
3884 min -= min / 2;
3885 else
3886 min -= min / 4;
3887 }
3888
3889 /*
3890 * Check watermarks for an order-0 allocation request. If these
3891 * are not met, then a high-order request also cannot go ahead
3892 * even if a suitable page happened to be free.
3893 */
3894 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3895 return false;
3896
3897 /* If this is an order-0 request then the watermark is fine */
3898 if (!order)
3899 return true;
3900
3901 /* For a high-order request, check at least one suitable page is free */
3902 for (o = order; o < MAX_ORDER; o++) {
3903 struct free_area *area = &z->free_area[o];
3904 int mt;
3905
3906 if (!area->nr_free)
3907 continue;
3908
3909 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3910 if (!free_area_empty(area, mt))
3911 return true;
3912 }
3913
3914#ifdef CONFIG_CMA
3915 if ((alloc_flags & ALLOC_CMA) &&
3916 !free_area_empty(area, MIGRATE_CMA)) {
3917 return true;
3918 }
3919#endif
3920 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3921 return true;
3922 }
3923 return false;
3924}
3925
3926bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3927 int highest_zoneidx, unsigned int alloc_flags)
3928{
3929 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3930 zone_page_state(z, NR_FREE_PAGES));
3931}
3932
3933static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3934 unsigned long mark, int highest_zoneidx,
3935 unsigned int alloc_flags, gfp_t gfp_mask)
3936{
3937 long free_pages;
3938
3939 free_pages = zone_page_state(z, NR_FREE_PAGES);
3940
3941 /*
3942 * Fast check for order-0 only. If this fails then the reserves
3943 * need to be calculated.
3944 */
3945 if (!order) {
3946 long fast_free;
3947
3948 fast_free = free_pages;
3949 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3950 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3951 return true;
3952 }
3953
3954 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3955 free_pages))
3956 return true;
3957 /*
3958 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3959 * when checking the min watermark. The min watermark is the
3960 * point where boosting is ignored so that kswapd is woken up
3961 * when below the low watermark.
3962 */
3963 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3964 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3965 mark = z->_watermark[WMARK_MIN];
3966 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3967 alloc_flags, free_pages);
3968 }
3969
3970 return false;
3971}
3972
3973bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3974 unsigned long mark, int highest_zoneidx)
3975{
3976 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3977
3978 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3979 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3980
3981 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3982 free_pages);
3983}
3984
3985#ifdef CONFIG_NUMA
3986static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3987{
3988 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3989 node_reclaim_distance;
3990}
3991#else /* CONFIG_NUMA */
3992static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3993{
3994 return true;
3995}
3996#endif /* CONFIG_NUMA */
3997
3998/*
3999 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4000 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4001 * premature use of a lower zone may cause lowmem pressure problems that
4002 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4003 * probably too small. It only makes sense to spread allocations to avoid
4004 * fragmentation between the Normal and DMA32 zones.
4005 */
4006static inline unsigned int
4007alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4008{
4009 unsigned int alloc_flags;
4010
4011 /*
4012 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4013 * to save a branch.
4014 */
4015 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4016
4017#ifdef CONFIG_ZONE_DMA32
4018 if (!zone)
4019 return alloc_flags;
4020
4021 if (zone_idx(zone) != ZONE_NORMAL)
4022 return alloc_flags;
4023
4024 /*
4025 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4026 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4027 * on UMA that if Normal is populated then so is DMA32.
4028 */
4029 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4030 if (nr_online_nodes > 1 && !populated_zone(--zone))
4031 return alloc_flags;
4032
4033 alloc_flags |= ALLOC_NOFRAGMENT;
4034#endif /* CONFIG_ZONE_DMA32 */
4035 return alloc_flags;
4036}
4037
4038/* Must be called after current_gfp_context() which can change gfp_mask */
4039static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4040 unsigned int alloc_flags)
4041{
4042#ifdef CONFIG_CMA
4043 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4044 alloc_flags |= ALLOC_CMA;
4045#endif
4046 return alloc_flags;
4047}
4048
4049/*
4050 * get_page_from_freelist goes through the zonelist trying to allocate
4051 * a page.
4052 */
4053static struct page *
4054get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4055 const struct alloc_context *ac)
4056{
4057 struct zoneref *z;
4058 struct zone *zone;
4059 struct pglist_data *last_pgdat_dirty_limit = NULL;
4060 bool no_fallback;
4061
4062retry:
4063 /*
4064 * Scan zonelist, looking for a zone with enough free.
4065 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4066 */
4067 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4068 z = ac->preferred_zoneref;
4069 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4070 ac->nodemask) {
4071 struct page *page;
4072 unsigned long mark;
4073
4074 if (cpusets_enabled() &&
4075 (alloc_flags & ALLOC_CPUSET) &&
4076 !__cpuset_zone_allowed(zone, gfp_mask))
4077 continue;
4078 /*
4079 * When allocating a page cache page for writing, we
4080 * want to get it from a node that is within its dirty
4081 * limit, such that no single node holds more than its
4082 * proportional share of globally allowed dirty pages.
4083 * The dirty limits take into account the node's
4084 * lowmem reserves and high watermark so that kswapd
4085 * should be able to balance it without having to
4086 * write pages from its LRU list.
4087 *
4088 * XXX: For now, allow allocations to potentially
4089 * exceed the per-node dirty limit in the slowpath
4090 * (spread_dirty_pages unset) before going into reclaim,
4091 * which is important when on a NUMA setup the allowed
4092 * nodes are together not big enough to reach the
4093 * global limit. The proper fix for these situations
4094 * will require awareness of nodes in the
4095 * dirty-throttling and the flusher threads.
4096 */
4097 if (ac->spread_dirty_pages) {
4098 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4099 continue;
4100
4101 if (!node_dirty_ok(zone->zone_pgdat)) {
4102 last_pgdat_dirty_limit = zone->zone_pgdat;
4103 continue;
4104 }
4105 }
4106
4107 if (no_fallback && nr_online_nodes > 1 &&
4108 zone != ac->preferred_zoneref->zone) {
4109 int local_nid;
4110
4111 /*
4112 * If moving to a remote node, retry but allow
4113 * fragmenting fallbacks. Locality is more important
4114 * than fragmentation avoidance.
4115 */
4116 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4117 if (zone_to_nid(zone) != local_nid) {
4118 alloc_flags &= ~ALLOC_NOFRAGMENT;
4119 goto retry;
4120 }
4121 }
4122
4123 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4124 if (!zone_watermark_fast(zone, order, mark,
4125 ac->highest_zoneidx, alloc_flags,
4126 gfp_mask)) {
4127 int ret;
4128
4129#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4130 /*
4131 * Watermark failed for this zone, but see if we can
4132 * grow this zone if it contains deferred pages.
4133 */
4134 if (static_branch_unlikely(&deferred_pages)) {
4135 if (_deferred_grow_zone(zone, order))
4136 goto try_this_zone;
4137 }
4138#endif
4139 /* Checked here to keep the fast path fast */
4140 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4141 if (alloc_flags & ALLOC_NO_WATERMARKS)
4142 goto try_this_zone;
4143
4144 if (!node_reclaim_enabled() ||
4145 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4146 continue;
4147
4148 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4149 switch (ret) {
4150 case NODE_RECLAIM_NOSCAN:
4151 /* did not scan */
4152 continue;
4153 case NODE_RECLAIM_FULL:
4154 /* scanned but unreclaimable */
4155 continue;
4156 default:
4157 /* did we reclaim enough */
4158 if (zone_watermark_ok(zone, order, mark,
4159 ac->highest_zoneidx, alloc_flags))
4160 goto try_this_zone;
4161
4162 continue;
4163 }
4164 }
4165
4166try_this_zone:
4167 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4168 gfp_mask, alloc_flags, ac->migratetype);
4169 if (page) {
4170 prep_new_page(page, order, gfp_mask, alloc_flags);
4171
4172 /*
4173 * If this is a high-order atomic allocation then check
4174 * if the pageblock should be reserved for the future
4175 */
4176 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4177 reserve_highatomic_pageblock(page, zone, order);
4178
4179 return page;
4180 } else {
4181#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4182 /* Try again if zone has deferred pages */
4183 if (static_branch_unlikely(&deferred_pages)) {
4184 if (_deferred_grow_zone(zone, order))
4185 goto try_this_zone;
4186 }
4187#endif
4188 }
4189 }
4190
4191 /*
4192 * It's possible on a UMA machine to get through all zones that are
4193 * fragmented. If avoiding fragmentation, reset and try again.
4194 */
4195 if (no_fallback) {
4196 alloc_flags &= ~ALLOC_NOFRAGMENT;
4197 goto retry;
4198 }
4199
4200 return NULL;
4201}
4202
4203static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4204{
4205 unsigned int filter = SHOW_MEM_FILTER_NODES;
4206
4207 /*
4208 * This documents exceptions given to allocations in certain
4209 * contexts that are allowed to allocate outside current's set
4210 * of allowed nodes.
4211 */
4212 if (!(gfp_mask & __GFP_NOMEMALLOC))
4213 if (tsk_is_oom_victim(current) ||
4214 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4215 filter &= ~SHOW_MEM_FILTER_NODES;
4216 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4217 filter &= ~SHOW_MEM_FILTER_NODES;
4218
4219 show_mem(filter, nodemask);
4220}
4221
4222void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4223{
4224 struct va_format vaf;
4225 va_list args;
4226 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4227
4228 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4229 return;
4230
4231 va_start(args, fmt);
4232 vaf.fmt = fmt;
4233 vaf.va = &args;
4234 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4235 current->comm, &vaf, gfp_mask, &gfp_mask,
4236 nodemask_pr_args(nodemask));
4237 va_end(args);
4238
4239 cpuset_print_current_mems_allowed();
4240 pr_cont("\n");
4241 dump_stack();
4242 warn_alloc_show_mem(gfp_mask, nodemask);
4243}
4244
4245static inline struct page *
4246__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4247 unsigned int alloc_flags,
4248 const struct alloc_context *ac)
4249{
4250 struct page *page;
4251
4252 page = get_page_from_freelist(gfp_mask, order,
4253 alloc_flags|ALLOC_CPUSET, ac);
4254 /*
4255 * fallback to ignore cpuset restriction if our nodes
4256 * are depleted
4257 */
4258 if (!page)
4259 page = get_page_from_freelist(gfp_mask, order,
4260 alloc_flags, ac);
4261
4262 return page;
4263}
4264
4265static inline struct page *
4266__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4267 const struct alloc_context *ac, unsigned long *did_some_progress)
4268{
4269 struct oom_control oc = {
4270 .zonelist = ac->zonelist,
4271 .nodemask = ac->nodemask,
4272 .memcg = NULL,
4273 .gfp_mask = gfp_mask,
4274 .order = order,
4275 };
4276 struct page *page;
4277
4278 *did_some_progress = 0;
4279
4280 /*
4281 * Acquire the oom lock. If that fails, somebody else is
4282 * making progress for us.
4283 */
4284 if (!mutex_trylock(&oom_lock)) {
4285 *did_some_progress = 1;
4286 schedule_timeout_uninterruptible(1);
4287 return NULL;
4288 }
4289
4290 /*
4291 * Go through the zonelist yet one more time, keep very high watermark
4292 * here, this is only to catch a parallel oom killing, we must fail if
4293 * we're still under heavy pressure. But make sure that this reclaim
4294 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4295 * allocation which will never fail due to oom_lock already held.
4296 */
4297 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4298 ~__GFP_DIRECT_RECLAIM, order,
4299 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4300 if (page)
4301 goto out;
4302
4303 /* Coredumps can quickly deplete all memory reserves */
4304 if (current->flags & PF_DUMPCORE)
4305 goto out;
4306 /* The OOM killer will not help higher order allocs */
4307 if (order > PAGE_ALLOC_COSTLY_ORDER)
4308 goto out;
4309 /*
4310 * We have already exhausted all our reclaim opportunities without any
4311 * success so it is time to admit defeat. We will skip the OOM killer
4312 * because it is very likely that the caller has a more reasonable
4313 * fallback than shooting a random task.
4314 *
4315 * The OOM killer may not free memory on a specific node.
4316 */
4317 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4318 goto out;
4319 /* The OOM killer does not needlessly kill tasks for lowmem */
4320 if (ac->highest_zoneidx < ZONE_NORMAL)
4321 goto out;
4322 if (pm_suspended_storage())
4323 goto out;
4324 /*
4325 * XXX: GFP_NOFS allocations should rather fail than rely on
4326 * other request to make a forward progress.
4327 * We are in an unfortunate situation where out_of_memory cannot
4328 * do much for this context but let's try it to at least get
4329 * access to memory reserved if the current task is killed (see
4330 * out_of_memory). Once filesystems are ready to handle allocation
4331 * failures more gracefully we should just bail out here.
4332 */
4333
4334 /* Exhausted what can be done so it's blame time */
4335 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4336 *did_some_progress = 1;
4337
4338 /*
4339 * Help non-failing allocations by giving them access to memory
4340 * reserves
4341 */
4342 if (gfp_mask & __GFP_NOFAIL)
4343 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4344 ALLOC_NO_WATERMARKS, ac);
4345 }
4346out:
4347 mutex_unlock(&oom_lock);
4348 return page;
4349}
4350
4351/*
4352 * Maximum number of compaction retries with a progress before OOM
4353 * killer is consider as the only way to move forward.
4354 */
4355#define MAX_COMPACT_RETRIES 16
4356
4357#ifdef CONFIG_COMPACTION
4358/* Try memory compaction for high-order allocations before reclaim */
4359static struct page *
4360__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4361 unsigned int alloc_flags, const struct alloc_context *ac,
4362 enum compact_priority prio, enum compact_result *compact_result)
4363{
4364 struct page *page = NULL;
4365 unsigned long pflags;
4366 unsigned int noreclaim_flag;
4367
4368 if (!order)
4369 return NULL;
4370
4371 psi_memstall_enter(&pflags);
4372 noreclaim_flag = memalloc_noreclaim_save();
4373
4374 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4375 prio, &page);
4376
4377 memalloc_noreclaim_restore(noreclaim_flag);
4378 psi_memstall_leave(&pflags);
4379
4380 if (*compact_result == COMPACT_SKIPPED)
4381 return NULL;
4382 /*
4383 * At least in one zone compaction wasn't deferred or skipped, so let's
4384 * count a compaction stall
4385 */
4386 count_vm_event(COMPACTSTALL);
4387
4388 /* Prep a captured page if available */
4389 if (page)
4390 prep_new_page(page, order, gfp_mask, alloc_flags);
4391
4392 /* Try get a page from the freelist if available */
4393 if (!page)
4394 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4395
4396 if (page) {
4397 struct zone *zone = page_zone(page);
4398
4399 zone->compact_blockskip_flush = false;
4400 compaction_defer_reset(zone, order, true);
4401 count_vm_event(COMPACTSUCCESS);
4402 return page;
4403 }
4404
4405 /*
4406 * It's bad if compaction run occurs and fails. The most likely reason
4407 * is that pages exist, but not enough to satisfy watermarks.
4408 */
4409 count_vm_event(COMPACTFAIL);
4410
4411 cond_resched();
4412
4413 return NULL;
4414}
4415
4416static inline bool
4417should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4418 enum compact_result compact_result,
4419 enum compact_priority *compact_priority,
4420 int *compaction_retries)
4421{
4422 int max_retries = MAX_COMPACT_RETRIES;
4423 int min_priority;
4424 bool ret = false;
4425 int retries = *compaction_retries;
4426 enum compact_priority priority = *compact_priority;
4427
4428 if (!order)
4429 return false;
4430
4431 if (fatal_signal_pending(current))
4432 return false;
4433
4434 if (compaction_made_progress(compact_result))
4435 (*compaction_retries)++;
4436
4437 /*
4438 * compaction considers all the zone as desperately out of memory
4439 * so it doesn't really make much sense to retry except when the
4440 * failure could be caused by insufficient priority
4441 */
4442 if (compaction_failed(compact_result))
4443 goto check_priority;
4444
4445 /*
4446 * compaction was skipped because there are not enough order-0 pages
4447 * to work with, so we retry only if it looks like reclaim can help.
4448 */
4449 if (compaction_needs_reclaim(compact_result)) {
4450 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4451 goto out;
4452 }
4453
4454 /*
4455 * make sure the compaction wasn't deferred or didn't bail out early
4456 * due to locks contention before we declare that we should give up.
4457 * But the next retry should use a higher priority if allowed, so
4458 * we don't just keep bailing out endlessly.
4459 */
4460 if (compaction_withdrawn(compact_result)) {
4461 goto check_priority;
4462 }
4463
4464 /*
4465 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4466 * costly ones because they are de facto nofail and invoke OOM
4467 * killer to move on while costly can fail and users are ready
4468 * to cope with that. 1/4 retries is rather arbitrary but we
4469 * would need much more detailed feedback from compaction to
4470 * make a better decision.
4471 */
4472 if (order > PAGE_ALLOC_COSTLY_ORDER)
4473 max_retries /= 4;
4474 if (*compaction_retries <= max_retries) {
4475 ret = true;
4476 goto out;
4477 }
4478
4479 /*
4480 * Make sure there are attempts at the highest priority if we exhausted
4481 * all retries or failed at the lower priorities.
4482 */
4483check_priority:
4484 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4485 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4486
4487 if (*compact_priority > min_priority) {
4488 (*compact_priority)--;
4489 *compaction_retries = 0;
4490 ret = true;
4491 }
4492out:
4493 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4494 return ret;
4495}
4496#else
4497static inline struct page *
4498__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4499 unsigned int alloc_flags, const struct alloc_context *ac,
4500 enum compact_priority prio, enum compact_result *compact_result)
4501{
4502 *compact_result = COMPACT_SKIPPED;
4503 return NULL;
4504}
4505
4506static inline bool
4507should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4508 enum compact_result compact_result,
4509 enum compact_priority *compact_priority,
4510 int *compaction_retries)
4511{
4512 struct zone *zone;
4513 struct zoneref *z;
4514
4515 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4516 return false;
4517
4518 /*
4519 * There are setups with compaction disabled which would prefer to loop
4520 * inside the allocator rather than hit the oom killer prematurely.
4521 * Let's give them a good hope and keep retrying while the order-0
4522 * watermarks are OK.
4523 */
4524 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4525 ac->highest_zoneidx, ac->nodemask) {
4526 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4527 ac->highest_zoneidx, alloc_flags))
4528 return true;
4529 }
4530 return false;
4531}
4532#endif /* CONFIG_COMPACTION */
4533
4534#ifdef CONFIG_LOCKDEP
4535static struct lockdep_map __fs_reclaim_map =
4536 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4537
4538static bool __need_reclaim(gfp_t gfp_mask)
4539{
4540 /* no reclaim without waiting on it */
4541 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4542 return false;
4543
4544 /* this guy won't enter reclaim */
4545 if (current->flags & PF_MEMALLOC)
4546 return false;
4547
4548 if (gfp_mask & __GFP_NOLOCKDEP)
4549 return false;
4550
4551 return true;
4552}
4553
4554void __fs_reclaim_acquire(void)
4555{
4556 lock_map_acquire(&__fs_reclaim_map);
4557}
4558
4559void __fs_reclaim_release(void)
4560{
4561 lock_map_release(&__fs_reclaim_map);
4562}
4563
4564void fs_reclaim_acquire(gfp_t gfp_mask)
4565{
4566 gfp_mask = current_gfp_context(gfp_mask);
4567
4568 if (__need_reclaim(gfp_mask)) {
4569 if (gfp_mask & __GFP_FS)
4570 __fs_reclaim_acquire();
4571
4572#ifdef CONFIG_MMU_NOTIFIER
4573 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4574 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4575#endif
4576
4577 }
4578}
4579EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4580
4581void fs_reclaim_release(gfp_t gfp_mask)
4582{
4583 gfp_mask = current_gfp_context(gfp_mask);
4584
4585 if (__need_reclaim(gfp_mask)) {
4586 if (gfp_mask & __GFP_FS)
4587 __fs_reclaim_release();
4588 }
4589}
4590EXPORT_SYMBOL_GPL(fs_reclaim_release);
4591#endif
4592
4593/* Perform direct synchronous page reclaim */
4594static unsigned long
4595__perform_reclaim(gfp_t gfp_mask, unsigned int order,
4596 const struct alloc_context *ac)
4597{
4598 unsigned int noreclaim_flag;
4599 unsigned long pflags, progress;
4600
4601 cond_resched();
4602
4603 /* We now go into synchronous reclaim */
4604 cpuset_memory_pressure_bump();
4605 psi_memstall_enter(&pflags);
4606 fs_reclaim_acquire(gfp_mask);
4607 noreclaim_flag = memalloc_noreclaim_save();
4608
4609 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4610 ac->nodemask);
4611
4612 memalloc_noreclaim_restore(noreclaim_flag);
4613 fs_reclaim_release(gfp_mask);
4614 psi_memstall_leave(&pflags);
4615
4616 cond_resched();
4617
4618 return progress;
4619}
4620
4621/* The really slow allocator path where we enter direct reclaim */
4622static inline struct page *
4623__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4624 unsigned int alloc_flags, const struct alloc_context *ac,
4625 unsigned long *did_some_progress)
4626{
4627 struct page *page = NULL;
4628 bool drained = false;
4629
4630 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4631 if (unlikely(!(*did_some_progress)))
4632 return NULL;
4633
4634retry:
4635 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4636
4637 /*
4638 * If an allocation failed after direct reclaim, it could be because
4639 * pages are pinned on the per-cpu lists or in high alloc reserves.
4640 * Shrink them and try again
4641 */
4642 if (!page && !drained) {
4643 unreserve_highatomic_pageblock(ac, false);
4644 drain_all_pages(NULL);
4645 drained = true;
4646 goto retry;
4647 }
4648
4649 return page;
4650}
4651
4652static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4653 const struct alloc_context *ac)
4654{
4655 struct zoneref *z;
4656 struct zone *zone;
4657 pg_data_t *last_pgdat = NULL;
4658 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4659
4660 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4661 ac->nodemask) {
4662 if (last_pgdat != zone->zone_pgdat)
4663 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4664 last_pgdat = zone->zone_pgdat;
4665 }
4666}
4667
4668static inline unsigned int
4669gfp_to_alloc_flags(gfp_t gfp_mask)
4670{
4671 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4672
4673 /*
4674 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4675 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4676 * to save two branches.
4677 */
4678 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4679 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4680
4681 /*
4682 * The caller may dip into page reserves a bit more if the caller
4683 * cannot run direct reclaim, or if the caller has realtime scheduling
4684 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4685 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4686 */
4687 alloc_flags |= (__force int)
4688 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4689
4690 if (gfp_mask & __GFP_ATOMIC) {
4691 /*
4692 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4693 * if it can't schedule.
4694 */
4695 if (!(gfp_mask & __GFP_NOMEMALLOC))
4696 alloc_flags |= ALLOC_HARDER;
4697 /*
4698 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4699 * comment for __cpuset_node_allowed().
4700 */
4701 alloc_flags &= ~ALLOC_CPUSET;
4702 } else if (unlikely(rt_task(current)) && !in_interrupt())
4703 alloc_flags |= ALLOC_HARDER;
4704
4705 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4706
4707 return alloc_flags;
4708}
4709
4710static bool oom_reserves_allowed(struct task_struct *tsk)
4711{
4712 if (!tsk_is_oom_victim(tsk))
4713 return false;
4714
4715 /*
4716 * !MMU doesn't have oom reaper so give access to memory reserves
4717 * only to the thread with TIF_MEMDIE set
4718 */
4719 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4720 return false;
4721
4722 return true;
4723}
4724
4725/*
4726 * Distinguish requests which really need access to full memory
4727 * reserves from oom victims which can live with a portion of it
4728 */
4729static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4730{
4731 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4732 return 0;
4733 if (gfp_mask & __GFP_MEMALLOC)
4734 return ALLOC_NO_WATERMARKS;
4735 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4736 return ALLOC_NO_WATERMARKS;
4737 if (!in_interrupt()) {
4738 if (current->flags & PF_MEMALLOC)
4739 return ALLOC_NO_WATERMARKS;
4740 else if (oom_reserves_allowed(current))
4741 return ALLOC_OOM;
4742 }
4743
4744 return 0;
4745}
4746
4747bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4748{
4749 return !!__gfp_pfmemalloc_flags(gfp_mask);
4750}
4751
4752/*
4753 * Checks whether it makes sense to retry the reclaim to make a forward progress
4754 * for the given allocation request.
4755 *
4756 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4757 * without success, or when we couldn't even meet the watermark if we
4758 * reclaimed all remaining pages on the LRU lists.
4759 *
4760 * Returns true if a retry is viable or false to enter the oom path.
4761 */
4762static inline bool
4763should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4764 struct alloc_context *ac, int alloc_flags,
4765 bool did_some_progress, int *no_progress_loops)
4766{
4767 struct zone *zone;
4768 struct zoneref *z;
4769 bool ret = false;
4770
4771 /*
4772 * Costly allocations might have made a progress but this doesn't mean
4773 * their order will become available due to high fragmentation so
4774 * always increment the no progress counter for them
4775 */
4776 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4777 *no_progress_loops = 0;
4778 else
4779 (*no_progress_loops)++;
4780
4781 /*
4782 * Make sure we converge to OOM if we cannot make any progress
4783 * several times in the row.
4784 */
4785 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4786 /* Before OOM, exhaust highatomic_reserve */
4787 return unreserve_highatomic_pageblock(ac, true);
4788 }
4789
4790 /*
4791 * Keep reclaiming pages while there is a chance this will lead
4792 * somewhere. If none of the target zones can satisfy our allocation
4793 * request even if all reclaimable pages are considered then we are
4794 * screwed and have to go OOM.
4795 */
4796 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4797 ac->highest_zoneidx, ac->nodemask) {
4798 unsigned long available;
4799 unsigned long reclaimable;
4800 unsigned long min_wmark = min_wmark_pages(zone);
4801 bool wmark;
4802
4803 available = reclaimable = zone_reclaimable_pages(zone);
4804 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4805
4806 /*
4807 * Would the allocation succeed if we reclaimed all
4808 * reclaimable pages?
4809 */
4810 wmark = __zone_watermark_ok(zone, order, min_wmark,
4811 ac->highest_zoneidx, alloc_flags, available);
4812 trace_reclaim_retry_zone(z, order, reclaimable,
4813 available, min_wmark, *no_progress_loops, wmark);
4814 if (wmark) {
4815 /*
4816 * If we didn't make any progress and have a lot of
4817 * dirty + writeback pages then we should wait for
4818 * an IO to complete to slow down the reclaim and
4819 * prevent from pre mature OOM
4820 */
4821 if (!did_some_progress) {
4822 unsigned long write_pending;
4823
4824 write_pending = zone_page_state_snapshot(zone,
4825 NR_ZONE_WRITE_PENDING);
4826
4827 if (2 * write_pending > reclaimable) {
4828 congestion_wait(BLK_RW_ASYNC, HZ/10);
4829 return true;
4830 }
4831 }
4832
4833 ret = true;
4834 goto out;
4835 }
4836 }
4837
4838out:
4839 /*
4840 * Memory allocation/reclaim might be called from a WQ context and the
4841 * current implementation of the WQ concurrency control doesn't
4842 * recognize that a particular WQ is congested if the worker thread is
4843 * looping without ever sleeping. Therefore we have to do a short sleep
4844 * here rather than calling cond_resched().
4845 */
4846 if (current->flags & PF_WQ_WORKER)
4847 schedule_timeout_uninterruptible(1);
4848 else
4849 cond_resched();
4850 return ret;
4851}
4852
4853static inline bool
4854check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4855{
4856 /*
4857 * It's possible that cpuset's mems_allowed and the nodemask from
4858 * mempolicy don't intersect. This should be normally dealt with by
4859 * policy_nodemask(), but it's possible to race with cpuset update in
4860 * such a way the check therein was true, and then it became false
4861 * before we got our cpuset_mems_cookie here.
4862 * This assumes that for all allocations, ac->nodemask can come only
4863 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4864 * when it does not intersect with the cpuset restrictions) or the
4865 * caller can deal with a violated nodemask.
4866 */
4867 if (cpusets_enabled() && ac->nodemask &&
4868 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4869 ac->nodemask = NULL;
4870 return true;
4871 }
4872
4873 /*
4874 * When updating a task's mems_allowed or mempolicy nodemask, it is
4875 * possible to race with parallel threads in such a way that our
4876 * allocation can fail while the mask is being updated. If we are about
4877 * to fail, check if the cpuset changed during allocation and if so,
4878 * retry.
4879 */
4880 if (read_mems_allowed_retry(cpuset_mems_cookie))
4881 return true;
4882
4883 return false;
4884}
4885
4886static inline struct page *
4887__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4888 struct alloc_context *ac)
4889{
4890 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4891 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4892 struct page *page = NULL;
4893 unsigned int alloc_flags;
4894 unsigned long did_some_progress;
4895 enum compact_priority compact_priority;
4896 enum compact_result compact_result;
4897 int compaction_retries;
4898 int no_progress_loops;
4899 unsigned int cpuset_mems_cookie;
4900 int reserve_flags;
4901
4902 /*
4903 * We also sanity check to catch abuse of atomic reserves being used by
4904 * callers that are not in atomic context.
4905 */
4906 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4907 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4908 gfp_mask &= ~__GFP_ATOMIC;
4909
4910retry_cpuset:
4911 compaction_retries = 0;
4912 no_progress_loops = 0;
4913 compact_priority = DEF_COMPACT_PRIORITY;
4914 cpuset_mems_cookie = read_mems_allowed_begin();
4915
4916 /*
4917 * The fast path uses conservative alloc_flags to succeed only until
4918 * kswapd needs to be woken up, and to avoid the cost of setting up
4919 * alloc_flags precisely. So we do that now.
4920 */
4921 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4922
4923 /*
4924 * We need to recalculate the starting point for the zonelist iterator
4925 * because we might have used different nodemask in the fast path, or
4926 * there was a cpuset modification and we are retrying - otherwise we
4927 * could end up iterating over non-eligible zones endlessly.
4928 */
4929 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4930 ac->highest_zoneidx, ac->nodemask);
4931 if (!ac->preferred_zoneref->zone)
4932 goto nopage;
4933
4934 if (alloc_flags & ALLOC_KSWAPD)
4935 wake_all_kswapds(order, gfp_mask, ac);
4936
4937 /*
4938 * The adjusted alloc_flags might result in immediate success, so try
4939 * that first
4940 */
4941 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4942 if (page)
4943 goto got_pg;
4944
4945 /*
4946 * For costly allocations, try direct compaction first, as it's likely
4947 * that we have enough base pages and don't need to reclaim. For non-
4948 * movable high-order allocations, do that as well, as compaction will
4949 * try prevent permanent fragmentation by migrating from blocks of the
4950 * same migratetype.
4951 * Don't try this for allocations that are allowed to ignore
4952 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4953 */
4954 if (can_direct_reclaim &&
4955 (costly_order ||
4956 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4957 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4958 page = __alloc_pages_direct_compact(gfp_mask, order,
4959 alloc_flags, ac,
4960 INIT_COMPACT_PRIORITY,
4961 &compact_result);
4962 if (page)
4963 goto got_pg;
4964
4965 /*
4966 * Checks for costly allocations with __GFP_NORETRY, which
4967 * includes some THP page fault allocations
4968 */
4969 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4970 /*
4971 * If allocating entire pageblock(s) and compaction
4972 * failed because all zones are below low watermarks
4973 * or is prohibited because it recently failed at this
4974 * order, fail immediately unless the allocator has
4975 * requested compaction and reclaim retry.
4976 *
4977 * Reclaim is
4978 * - potentially very expensive because zones are far
4979 * below their low watermarks or this is part of very
4980 * bursty high order allocations,
4981 * - not guaranteed to help because isolate_freepages()
4982 * may not iterate over freed pages as part of its
4983 * linear scan, and
4984 * - unlikely to make entire pageblocks free on its
4985 * own.
4986 */
4987 if (compact_result == COMPACT_SKIPPED ||
4988 compact_result == COMPACT_DEFERRED)
4989 goto nopage;
4990
4991 /*
4992 * Looks like reclaim/compaction is worth trying, but
4993 * sync compaction could be very expensive, so keep
4994 * using async compaction.
4995 */
4996 compact_priority = INIT_COMPACT_PRIORITY;
4997 }
4998 }
4999
5000retry:
5001 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5002 if (alloc_flags & ALLOC_KSWAPD)
5003 wake_all_kswapds(order, gfp_mask, ac);
5004
5005 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5006 if (reserve_flags)
5007 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
5008
5009 /*
5010 * Reset the nodemask and zonelist iterators if memory policies can be
5011 * ignored. These allocations are high priority and system rather than
5012 * user oriented.
5013 */
5014 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5015 ac->nodemask = NULL;
5016 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5017 ac->highest_zoneidx, ac->nodemask);
5018 }
5019
5020 /* Attempt with potentially adjusted zonelist and alloc_flags */
5021 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5022 if (page)
5023 goto got_pg;
5024
5025 /* Caller is not willing to reclaim, we can't balance anything */
5026 if (!can_direct_reclaim)
5027 goto nopage;
5028
5029 /* Avoid recursion of direct reclaim */
5030 if (current->flags & PF_MEMALLOC)
5031 goto nopage;
5032
5033 /* Try direct reclaim and then allocating */
5034 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5035 &did_some_progress);
5036 if (page)
5037 goto got_pg;
5038
5039 /* Try direct compaction and then allocating */
5040 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5041 compact_priority, &compact_result);
5042 if (page)
5043 goto got_pg;
5044
5045 /* Do not loop if specifically requested */
5046 if (gfp_mask & __GFP_NORETRY)
5047 goto nopage;
5048
5049 /*
5050 * Do not retry costly high order allocations unless they are
5051 * __GFP_RETRY_MAYFAIL
5052 */
5053 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5054 goto nopage;
5055
5056 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5057 did_some_progress > 0, &no_progress_loops))
5058 goto retry;
5059
5060 /*
5061 * It doesn't make any sense to retry for the compaction if the order-0
5062 * reclaim is not able to make any progress because the current
5063 * implementation of the compaction depends on the sufficient amount
5064 * of free memory (see __compaction_suitable)
5065 */
5066 if (did_some_progress > 0 &&
5067 should_compact_retry(ac, order, alloc_flags,
5068 compact_result, &compact_priority,
5069 &compaction_retries))
5070 goto retry;
5071
5072
5073 /* Deal with possible cpuset update races before we start OOM killing */
5074 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5075 goto retry_cpuset;
5076
5077 /* Reclaim has failed us, start killing things */
5078 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5079 if (page)
5080 goto got_pg;
5081
5082 /* Avoid allocations with no watermarks from looping endlessly */
5083 if (tsk_is_oom_victim(current) &&
5084 (alloc_flags & ALLOC_OOM ||
5085 (gfp_mask & __GFP_NOMEMALLOC)))
5086 goto nopage;
5087
5088 /* Retry as long as the OOM killer is making progress */
5089 if (did_some_progress) {
5090 no_progress_loops = 0;
5091 goto retry;
5092 }
5093
5094nopage:
5095 /* Deal with possible cpuset update races before we fail */
5096 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5097 goto retry_cpuset;
5098
5099 /*
5100 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5101 * we always retry
5102 */
5103 if (gfp_mask & __GFP_NOFAIL) {
5104 /*
5105 * All existing users of the __GFP_NOFAIL are blockable, so warn
5106 * of any new users that actually require GFP_NOWAIT
5107 */
5108 if (WARN_ON_ONCE(!can_direct_reclaim))
5109 goto fail;
5110
5111 /*
5112 * PF_MEMALLOC request from this context is rather bizarre
5113 * because we cannot reclaim anything and only can loop waiting
5114 * for somebody to do a work for us
5115 */
5116 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5117
5118 /*
5119 * non failing costly orders are a hard requirement which we
5120 * are not prepared for much so let's warn about these users
5121 * so that we can identify them and convert them to something
5122 * else.
5123 */
5124 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5125
5126 /*
5127 * Help non-failing allocations by giving them access to memory
5128 * reserves but do not use ALLOC_NO_WATERMARKS because this
5129 * could deplete whole memory reserves which would just make
5130 * the situation worse
5131 */
5132 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5133 if (page)
5134 goto got_pg;
5135
5136 cond_resched();
5137 goto retry;
5138 }
5139fail:
5140 warn_alloc(gfp_mask, ac->nodemask,
5141 "page allocation failure: order:%u", order);
5142got_pg:
5143 return page;
5144}
5145
5146static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5147 int preferred_nid, nodemask_t *nodemask,
5148 struct alloc_context *ac, gfp_t *alloc_gfp,
5149 unsigned int *alloc_flags)
5150{
5151 ac->highest_zoneidx = gfp_zone(gfp_mask);
5152 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5153 ac->nodemask = nodemask;
5154 ac->migratetype = gfp_migratetype(gfp_mask);
5155
5156 if (cpusets_enabled()) {
5157 *alloc_gfp |= __GFP_HARDWALL;
5158 /*
5159 * When we are in the interrupt context, it is irrelevant
5160 * to the current task context. It means that any node ok.
5161 */
5162 if (!in_interrupt() && !ac->nodemask)
5163 ac->nodemask = &cpuset_current_mems_allowed;
5164 else
5165 *alloc_flags |= ALLOC_CPUSET;
5166 }
5167
5168 fs_reclaim_acquire(gfp_mask);
5169 fs_reclaim_release(gfp_mask);
5170
5171 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5172
5173 if (should_fail_alloc_page(gfp_mask, order))
5174 return false;
5175
5176 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5177
5178 /* Dirty zone balancing only done in the fast path */
5179 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5180
5181 /*
5182 * The preferred zone is used for statistics but crucially it is
5183 * also used as the starting point for the zonelist iterator. It
5184 * may get reset for allocations that ignore memory policies.
5185 */
5186 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5187 ac->highest_zoneidx, ac->nodemask);
5188
5189 return true;
5190}
5191
5192/*
5193 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5194 * @gfp: GFP flags for the allocation
5195 * @preferred_nid: The preferred NUMA node ID to allocate from
5196 * @nodemask: Set of nodes to allocate from, may be NULL
5197 * @nr_pages: The number of pages desired on the list or array
5198 * @page_list: Optional list to store the allocated pages
5199 * @page_array: Optional array to store the pages
5200 *
5201 * This is a batched version of the page allocator that attempts to
5202 * allocate nr_pages quickly. Pages are added to page_list if page_list
5203 * is not NULL, otherwise it is assumed that the page_array is valid.
5204 *
5205 * For lists, nr_pages is the number of pages that should be allocated.
5206 *
5207 * For arrays, only NULL elements are populated with pages and nr_pages
5208 * is the maximum number of pages that will be stored in the array.
5209 *
5210 * Returns the number of pages on the list or array.
5211 */
5212unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5213 nodemask_t *nodemask, int nr_pages,
5214 struct list_head *page_list,
5215 struct page **page_array)
5216{
5217 struct page *page;
5218 unsigned long flags;
5219 struct zone *zone;
5220 struct zoneref *z;
5221 struct per_cpu_pages *pcp;
5222 struct list_head *pcp_list;
5223 struct alloc_context ac;
5224 gfp_t alloc_gfp;
5225 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5226 int nr_populated = 0, nr_account = 0;
5227
5228 /*
5229 * Skip populated array elements to determine if any pages need
5230 * to be allocated before disabling IRQs.
5231 */
5232 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5233 nr_populated++;
5234
5235 /* No pages requested? */
5236 if (unlikely(nr_pages <= 0))
5237 goto out;
5238
5239 /* Already populated array? */
5240 if (unlikely(page_array && nr_pages - nr_populated == 0))
5241 goto out;
5242
5243 /* Use the single page allocator for one page. */
5244 if (nr_pages - nr_populated == 1)
5245 goto failed;
5246
5247#ifdef CONFIG_PAGE_OWNER
5248 /*
5249 * PAGE_OWNER may recurse into the allocator to allocate space to
5250 * save the stack with pagesets.lock held. Releasing/reacquiring
5251 * removes much of the performance benefit of bulk allocation so
5252 * force the caller to allocate one page at a time as it'll have
5253 * similar performance to added complexity to the bulk allocator.
5254 */
5255 if (static_branch_unlikely(&page_owner_inited))
5256 goto failed;
5257#endif
5258
5259 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5260 gfp &= gfp_allowed_mask;
5261 alloc_gfp = gfp;
5262 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5263 goto out;
5264 gfp = alloc_gfp;
5265
5266 /* Find an allowed local zone that meets the low watermark. */
5267 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5268 unsigned long mark;
5269
5270 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5271 !__cpuset_zone_allowed(zone, gfp)) {
5272 continue;
5273 }
5274
5275 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5276 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5277 goto failed;
5278 }
5279
5280 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5281 if (zone_watermark_fast(zone, 0, mark,
5282 zonelist_zone_idx(ac.preferred_zoneref),
5283 alloc_flags, gfp)) {
5284 break;
5285 }
5286 }
5287
5288 /*
5289 * If there are no allowed local zones that meets the watermarks then
5290 * try to allocate a single page and reclaim if necessary.
5291 */
5292 if (unlikely(!zone))
5293 goto failed;
5294
5295 /* Attempt the batch allocation */
5296 local_lock_irqsave(&pagesets.lock, flags);
5297 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5298 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5299
5300 while (nr_populated < nr_pages) {
5301
5302 /* Skip existing pages */
5303 if (page_array && page_array[nr_populated]) {
5304 nr_populated++;
5305 continue;
5306 }
5307
5308 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5309 pcp, pcp_list);
5310 if (unlikely(!page)) {
5311 /* Try and get at least one page */
5312 if (!nr_populated)
5313 goto failed_irq;
5314 break;
5315 }
5316 nr_account++;
5317
5318 prep_new_page(page, 0, gfp, 0);
5319 if (page_list)
5320 list_add(&page->lru, page_list);
5321 else
5322 page_array[nr_populated] = page;
5323 nr_populated++;
5324 }
5325
5326 local_unlock_irqrestore(&pagesets.lock, flags);
5327
5328 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5329 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5330
5331out:
5332 return nr_populated;
5333
5334failed_irq:
5335 local_unlock_irqrestore(&pagesets.lock, flags);
5336
5337failed:
5338 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5339 if (page) {
5340 if (page_list)
5341 list_add(&page->lru, page_list);
5342 else
5343 page_array[nr_populated] = page;
5344 nr_populated++;
5345 }
5346
5347 goto out;
5348}
5349EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5350
5351/*
5352 * This is the 'heart' of the zoned buddy allocator.
5353 */
5354struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5355 nodemask_t *nodemask)
5356{
5357 struct page *page;
5358 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5359 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5360 struct alloc_context ac = { };
5361
5362 /*
5363 * There are several places where we assume that the order value is sane
5364 * so bail out early if the request is out of bound.
5365 */
5366 if (unlikely(order >= MAX_ORDER)) {
5367 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5368 return NULL;
5369 }
5370
5371 gfp &= gfp_allowed_mask;
5372 /*
5373 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5374 * resp. GFP_NOIO which has to be inherited for all allocation requests
5375 * from a particular context which has been marked by
5376 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5377 * movable zones are not used during allocation.
5378 */
5379 gfp = current_gfp_context(gfp);
5380 alloc_gfp = gfp;
5381 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5382 &alloc_gfp, &alloc_flags))
5383 return NULL;
5384
5385 /*
5386 * Forbid the first pass from falling back to types that fragment
5387 * memory until all local zones are considered.
5388 */
5389 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5390
5391 /* First allocation attempt */
5392 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5393 if (likely(page))
5394 goto out;
5395
5396 alloc_gfp = gfp;
5397 ac.spread_dirty_pages = false;
5398
5399 /*
5400 * Restore the original nodemask if it was potentially replaced with
5401 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5402 */
5403 ac.nodemask = nodemask;
5404
5405 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5406
5407out:
5408 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5409 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5410 __free_pages(page, order);
5411 page = NULL;
5412 }
5413
5414 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5415
5416 return page;
5417}
5418EXPORT_SYMBOL(__alloc_pages);
5419
5420/*
5421 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5422 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5423 * you need to access high mem.
5424 */
5425unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5426{
5427 struct page *page;
5428
5429 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5430 if (!page)
5431 return 0;
5432 return (unsigned long) page_address(page);
5433}
5434EXPORT_SYMBOL(__get_free_pages);
5435
5436unsigned long get_zeroed_page(gfp_t gfp_mask)
5437{
5438 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5439}
5440EXPORT_SYMBOL(get_zeroed_page);
5441
5442/**
5443 * __free_pages - Free pages allocated with alloc_pages().
5444 * @page: The page pointer returned from alloc_pages().
5445 * @order: The order of the allocation.
5446 *
5447 * This function can free multi-page allocations that are not compound
5448 * pages. It does not check that the @order passed in matches that of
5449 * the allocation, so it is easy to leak memory. Freeing more memory
5450 * than was allocated will probably emit a warning.
5451 *
5452 * If the last reference to this page is speculative, it will be released
5453 * by put_page() which only frees the first page of a non-compound
5454 * allocation. To prevent the remaining pages from being leaked, we free
5455 * the subsequent pages here. If you want to use the page's reference
5456 * count to decide when to free the allocation, you should allocate a
5457 * compound page, and use put_page() instead of __free_pages().
5458 *
5459 * Context: May be called in interrupt context or while holding a normal
5460 * spinlock, but not in NMI context or while holding a raw spinlock.
5461 */
5462void __free_pages(struct page *page, unsigned int order)
5463{
5464 if (put_page_testzero(page))
5465 free_the_page(page, order);
5466 else if (!PageHead(page))
5467 while (order-- > 0)
5468 free_the_page(page + (1 << order), order);
5469}
5470EXPORT_SYMBOL(__free_pages);
5471
5472void free_pages(unsigned long addr, unsigned int order)
5473{
5474 if (addr != 0) {
5475 VM_BUG_ON(!virt_addr_valid((void *)addr));
5476 __free_pages(virt_to_page((void *)addr), order);
5477 }
5478}
5479
5480EXPORT_SYMBOL(free_pages);
5481
5482/*
5483 * Page Fragment:
5484 * An arbitrary-length arbitrary-offset area of memory which resides
5485 * within a 0 or higher order page. Multiple fragments within that page
5486 * are individually refcounted, in the page's reference counter.
5487 *
5488 * The page_frag functions below provide a simple allocation framework for
5489 * page fragments. This is used by the network stack and network device
5490 * drivers to provide a backing region of memory for use as either an
5491 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5492 */
5493static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5494 gfp_t gfp_mask)
5495{
5496 struct page *page = NULL;
5497 gfp_t gfp = gfp_mask;
5498
5499#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5500 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5501 __GFP_NOMEMALLOC;
5502 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5503 PAGE_FRAG_CACHE_MAX_ORDER);
5504 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5505#endif
5506 if (unlikely(!page))
5507 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5508
5509 nc->va = page ? page_address(page) : NULL;
5510
5511 return page;
5512}
5513
5514void __page_frag_cache_drain(struct page *page, unsigned int count)
5515{
5516 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5517
5518 if (page_ref_sub_and_test(page, count))
5519 free_the_page(page, compound_order(page));
5520}
5521EXPORT_SYMBOL(__page_frag_cache_drain);
5522
5523void *page_frag_alloc_align(struct page_frag_cache *nc,
5524 unsigned int fragsz, gfp_t gfp_mask,
5525 unsigned int align_mask)
5526{
5527 unsigned int size = PAGE_SIZE;
5528 struct page *page;
5529 int offset;
5530
5531 if (unlikely(!nc->va)) {
5532refill:
5533 page = __page_frag_cache_refill(nc, gfp_mask);
5534 if (!page)
5535 return NULL;
5536
5537#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5538 /* if size can vary use size else just use PAGE_SIZE */
5539 size = nc->size;
5540#endif
5541 /* Even if we own the page, we do not use atomic_set().
5542 * This would break get_page_unless_zero() users.
5543 */
5544 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5545
5546 /* reset page count bias and offset to start of new frag */
5547 nc->pfmemalloc = page_is_pfmemalloc(page);
5548 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5549 nc->offset = size;
5550 }
5551
5552 offset = nc->offset - fragsz;
5553 if (unlikely(offset < 0)) {
5554 page = virt_to_page(nc->va);
5555
5556 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5557 goto refill;
5558
5559 if (unlikely(nc->pfmemalloc)) {
5560 free_the_page(page, compound_order(page));
5561 goto refill;
5562 }
5563
5564#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5565 /* if size can vary use size else just use PAGE_SIZE */
5566 size = nc->size;
5567#endif
5568 /* OK, page count is 0, we can safely set it */
5569 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5570
5571 /* reset page count bias and offset to start of new frag */
5572 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5573 offset = size - fragsz;
5574 }
5575
5576 nc->pagecnt_bias--;
5577 offset &= align_mask;
5578 nc->offset = offset;
5579
5580 return nc->va + offset;
5581}
5582EXPORT_SYMBOL(page_frag_alloc_align);
5583
5584/*
5585 * Frees a page fragment allocated out of either a compound or order 0 page.
5586 */
5587void page_frag_free(void *addr)
5588{
5589 struct page *page = virt_to_head_page(addr);
5590
5591 if (unlikely(put_page_testzero(page)))
5592 free_the_page(page, compound_order(page));
5593}
5594EXPORT_SYMBOL(page_frag_free);
5595
5596static void *make_alloc_exact(unsigned long addr, unsigned int order,
5597 size_t size)
5598{
5599 if (addr) {
5600 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5601 unsigned long used = addr + PAGE_ALIGN(size);
5602
5603 split_page(virt_to_page((void *)addr), order);
5604 while (used < alloc_end) {
5605 free_page(used);
5606 used += PAGE_SIZE;
5607 }
5608 }
5609 return (void *)addr;
5610}
5611
5612/**
5613 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5614 * @size: the number of bytes to allocate
5615 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5616 *
5617 * This function is similar to alloc_pages(), except that it allocates the
5618 * minimum number of pages to satisfy the request. alloc_pages() can only
5619 * allocate memory in power-of-two pages.
5620 *
5621 * This function is also limited by MAX_ORDER.
5622 *
5623 * Memory allocated by this function must be released by free_pages_exact().
5624 *
5625 * Return: pointer to the allocated area or %NULL in case of error.
5626 */
5627void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5628{
5629 unsigned int order = get_order(size);
5630 unsigned long addr;
5631
5632 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5633 gfp_mask &= ~__GFP_COMP;
5634
5635 addr = __get_free_pages(gfp_mask, order);
5636 return make_alloc_exact(addr, order, size);
5637}
5638EXPORT_SYMBOL(alloc_pages_exact);
5639
5640/**
5641 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5642 * pages on a node.
5643 * @nid: the preferred node ID where memory should be allocated
5644 * @size: the number of bytes to allocate
5645 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5646 *
5647 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5648 * back.
5649 *
5650 * Return: pointer to the allocated area or %NULL in case of error.
5651 */
5652void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5653{
5654 unsigned int order = get_order(size);
5655 struct page *p;
5656
5657 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5658 gfp_mask &= ~__GFP_COMP;
5659
5660 p = alloc_pages_node(nid, gfp_mask, order);
5661 if (!p)
5662 return NULL;
5663 return make_alloc_exact((unsigned long)page_address(p), order, size);
5664}
5665
5666/**
5667 * free_pages_exact - release memory allocated via alloc_pages_exact()
5668 * @virt: the value returned by alloc_pages_exact.
5669 * @size: size of allocation, same value as passed to alloc_pages_exact().
5670 *
5671 * Release the memory allocated by a previous call to alloc_pages_exact.
5672 */
5673void free_pages_exact(void *virt, size_t size)
5674{
5675 unsigned long addr = (unsigned long)virt;
5676 unsigned long end = addr + PAGE_ALIGN(size);
5677
5678 while (addr < end) {
5679 free_page(addr);
5680 addr += PAGE_SIZE;
5681 }
5682}
5683EXPORT_SYMBOL(free_pages_exact);
5684
5685/**
5686 * nr_free_zone_pages - count number of pages beyond high watermark
5687 * @offset: The zone index of the highest zone
5688 *
5689 * nr_free_zone_pages() counts the number of pages which are beyond the
5690 * high watermark within all zones at or below a given zone index. For each
5691 * zone, the number of pages is calculated as:
5692 *
5693 * nr_free_zone_pages = managed_pages - high_pages
5694 *
5695 * Return: number of pages beyond high watermark.
5696 */
5697static unsigned long nr_free_zone_pages(int offset)
5698{
5699 struct zoneref *z;
5700 struct zone *zone;
5701
5702 /* Just pick one node, since fallback list is circular */
5703 unsigned long sum = 0;
5704
5705 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5706
5707 for_each_zone_zonelist(zone, z, zonelist, offset) {
5708 unsigned long size = zone_managed_pages(zone);
5709 unsigned long high = high_wmark_pages(zone);
5710 if (size > high)
5711 sum += size - high;
5712 }
5713
5714 return sum;
5715}
5716
5717/**
5718 * nr_free_buffer_pages - count number of pages beyond high watermark
5719 *
5720 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5721 * watermark within ZONE_DMA and ZONE_NORMAL.
5722 *
5723 * Return: number of pages beyond high watermark within ZONE_DMA and
5724 * ZONE_NORMAL.
5725 */
5726unsigned long nr_free_buffer_pages(void)
5727{
5728 return nr_free_zone_pages(gfp_zone(GFP_USER));
5729}
5730EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5731
5732static inline void show_node(struct zone *zone)
5733{
5734 if (IS_ENABLED(CONFIG_NUMA))
5735 printk("Node %d ", zone_to_nid(zone));
5736}
5737
5738long si_mem_available(void)
5739{
5740 long available;
5741 unsigned long pagecache;
5742 unsigned long wmark_low = 0;
5743 unsigned long pages[NR_LRU_LISTS];
5744 unsigned long reclaimable;
5745 struct zone *zone;
5746 int lru;
5747
5748 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5749 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5750
5751 for_each_zone(zone)
5752 wmark_low += low_wmark_pages(zone);
5753
5754 /*
5755 * Estimate the amount of memory available for userspace allocations,
5756 * without causing swapping.
5757 */
5758 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5759
5760 /*
5761 * Not all the page cache can be freed, otherwise the system will
5762 * start swapping. Assume at least half of the page cache, or the
5763 * low watermark worth of cache, needs to stay.
5764 */
5765 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5766 pagecache -= min(pagecache / 2, wmark_low);
5767 available += pagecache;
5768
5769 /*
5770 * Part of the reclaimable slab and other kernel memory consists of
5771 * items that are in use, and cannot be freed. Cap this estimate at the
5772 * low watermark.
5773 */
5774 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5775 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5776 available += reclaimable - min(reclaimable / 2, wmark_low);
5777
5778 if (available < 0)
5779 available = 0;
5780 return available;
5781}
5782EXPORT_SYMBOL_GPL(si_mem_available);
5783
5784void si_meminfo(struct sysinfo *val)
5785{
5786 val->totalram = totalram_pages();
5787 val->sharedram = global_node_page_state(NR_SHMEM);
5788 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5789 val->bufferram = nr_blockdev_pages();
5790 val->totalhigh = totalhigh_pages();
5791 val->freehigh = nr_free_highpages();
5792 val->mem_unit = PAGE_SIZE;
5793}
5794
5795EXPORT_SYMBOL(si_meminfo);
5796
5797#ifdef CONFIG_NUMA
5798void si_meminfo_node(struct sysinfo *val, int nid)
5799{
5800 int zone_type; /* needs to be signed */
5801 unsigned long managed_pages = 0;
5802 unsigned long managed_highpages = 0;
5803 unsigned long free_highpages = 0;
5804 pg_data_t *pgdat = NODE_DATA(nid);
5805
5806 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5807 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5808 val->totalram = managed_pages;
5809 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5810 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5811#ifdef CONFIG_HIGHMEM
5812 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5813 struct zone *zone = &pgdat->node_zones[zone_type];
5814
5815 if (is_highmem(zone)) {
5816 managed_highpages += zone_managed_pages(zone);
5817 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5818 }
5819 }
5820 val->totalhigh = managed_highpages;
5821 val->freehigh = free_highpages;
5822#else
5823 val->totalhigh = managed_highpages;
5824 val->freehigh = free_highpages;
5825#endif
5826 val->mem_unit = PAGE_SIZE;
5827}
5828#endif
5829
5830/*
5831 * Determine whether the node should be displayed or not, depending on whether
5832 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5833 */
5834static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5835{
5836 if (!(flags & SHOW_MEM_FILTER_NODES))
5837 return false;
5838
5839 /*
5840 * no node mask - aka implicit memory numa policy. Do not bother with
5841 * the synchronization - read_mems_allowed_begin - because we do not
5842 * have to be precise here.
5843 */
5844 if (!nodemask)
5845 nodemask = &cpuset_current_mems_allowed;
5846
5847 return !node_isset(nid, *nodemask);
5848}
5849
5850#define K(x) ((x) << (PAGE_SHIFT-10))
5851
5852static void show_migration_types(unsigned char type)
5853{
5854 static const char types[MIGRATE_TYPES] = {
5855 [MIGRATE_UNMOVABLE] = 'U',
5856 [MIGRATE_MOVABLE] = 'M',
5857 [MIGRATE_RECLAIMABLE] = 'E',
5858 [MIGRATE_HIGHATOMIC] = 'H',
5859#ifdef CONFIG_CMA
5860 [MIGRATE_CMA] = 'C',
5861#endif
5862#ifdef CONFIG_MEMORY_ISOLATION
5863 [MIGRATE_ISOLATE] = 'I',
5864#endif
5865 };
5866 char tmp[MIGRATE_TYPES + 1];
5867 char *p = tmp;
5868 int i;
5869
5870 for (i = 0; i < MIGRATE_TYPES; i++) {
5871 if (type & (1 << i))
5872 *p++ = types[i];
5873 }
5874
5875 *p = '\0';
5876 printk(KERN_CONT "(%s) ", tmp);
5877}
5878
5879/*
5880 * Show free area list (used inside shift_scroll-lock stuff)
5881 * We also calculate the percentage fragmentation. We do this by counting the
5882 * memory on each free list with the exception of the first item on the list.
5883 *
5884 * Bits in @filter:
5885 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5886 * cpuset.
5887 */
5888void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5889{
5890 unsigned long free_pcp = 0;
5891 int cpu;
5892 struct zone *zone;
5893 pg_data_t *pgdat;
5894
5895 for_each_populated_zone(zone) {
5896 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5897 continue;
5898
5899 for_each_online_cpu(cpu)
5900 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5901 }
5902
5903 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5904 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5905 " unevictable:%lu dirty:%lu writeback:%lu\n"
5906 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5907 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5908 " free:%lu free_pcp:%lu free_cma:%lu\n",
5909 global_node_page_state(NR_ACTIVE_ANON),
5910 global_node_page_state(NR_INACTIVE_ANON),
5911 global_node_page_state(NR_ISOLATED_ANON),
5912 global_node_page_state(NR_ACTIVE_FILE),
5913 global_node_page_state(NR_INACTIVE_FILE),
5914 global_node_page_state(NR_ISOLATED_FILE),
5915 global_node_page_state(NR_UNEVICTABLE),
5916 global_node_page_state(NR_FILE_DIRTY),
5917 global_node_page_state(NR_WRITEBACK),
5918 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5919 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5920 global_node_page_state(NR_FILE_MAPPED),
5921 global_node_page_state(NR_SHMEM),
5922 global_node_page_state(NR_PAGETABLE),
5923 global_zone_page_state(NR_BOUNCE),
5924 global_zone_page_state(NR_FREE_PAGES),
5925 free_pcp,
5926 global_zone_page_state(NR_FREE_CMA_PAGES));
5927
5928 for_each_online_pgdat(pgdat) {
5929 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5930 continue;
5931
5932 printk("Node %d"
5933 " active_anon:%lukB"
5934 " inactive_anon:%lukB"
5935 " active_file:%lukB"
5936 " inactive_file:%lukB"
5937 " unevictable:%lukB"
5938 " isolated(anon):%lukB"
5939 " isolated(file):%lukB"
5940 " mapped:%lukB"
5941 " dirty:%lukB"
5942 " writeback:%lukB"
5943 " shmem:%lukB"
5944#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5945 " shmem_thp: %lukB"
5946 " shmem_pmdmapped: %lukB"
5947 " anon_thp: %lukB"
5948#endif
5949 " writeback_tmp:%lukB"
5950 " kernel_stack:%lukB"
5951#ifdef CONFIG_SHADOW_CALL_STACK
5952 " shadow_call_stack:%lukB"
5953#endif
5954 " pagetables:%lukB"
5955 " all_unreclaimable? %s"
5956 "\n",
5957 pgdat->node_id,
5958 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5959 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5960 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5961 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5962 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5963 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5964 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5965 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5966 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5967 K(node_page_state(pgdat, NR_WRITEBACK)),
5968 K(node_page_state(pgdat, NR_SHMEM)),
5969#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5970 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5971 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5972 K(node_page_state(pgdat, NR_ANON_THPS)),
5973#endif
5974 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5975 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5976#ifdef CONFIG_SHADOW_CALL_STACK
5977 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5978#endif
5979 K(node_page_state(pgdat, NR_PAGETABLE)),
5980 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5981 "yes" : "no");
5982 }
5983
5984 for_each_populated_zone(zone) {
5985 int i;
5986
5987 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5988 continue;
5989
5990 free_pcp = 0;
5991 for_each_online_cpu(cpu)
5992 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5993
5994 show_node(zone);
5995 printk(KERN_CONT
5996 "%s"
5997 " free:%lukB"
5998 " min:%lukB"
5999 " low:%lukB"
6000 " high:%lukB"
6001 " reserved_highatomic:%luKB"
6002 " active_anon:%lukB"
6003 " inactive_anon:%lukB"
6004 " active_file:%lukB"
6005 " inactive_file:%lukB"
6006 " unevictable:%lukB"
6007 " writepending:%lukB"
6008 " present:%lukB"
6009 " managed:%lukB"
6010 " mlocked:%lukB"
6011 " bounce:%lukB"
6012 " free_pcp:%lukB"
6013 " local_pcp:%ukB"
6014 " free_cma:%lukB"
6015 "\n",
6016 zone->name,
6017 K(zone_page_state(zone, NR_FREE_PAGES)),
6018 K(min_wmark_pages(zone)),
6019 K(low_wmark_pages(zone)),
6020 K(high_wmark_pages(zone)),
6021 K(zone->nr_reserved_highatomic),
6022 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6023 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6024 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6025 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6026 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6027 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6028 K(zone->present_pages),
6029 K(zone_managed_pages(zone)),
6030 K(zone_page_state(zone, NR_MLOCK)),
6031 K(zone_page_state(zone, NR_BOUNCE)),
6032 K(free_pcp),
6033 K(this_cpu_read(zone->per_cpu_pageset->count)),
6034 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6035 printk("lowmem_reserve[]:");
6036 for (i = 0; i < MAX_NR_ZONES; i++)
6037 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6038 printk(KERN_CONT "\n");
6039 }
6040
6041 for_each_populated_zone(zone) {
6042 unsigned int order;
6043 unsigned long nr[MAX_ORDER], flags, total = 0;
6044 unsigned char types[MAX_ORDER];
6045
6046 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6047 continue;
6048 show_node(zone);
6049 printk(KERN_CONT "%s: ", zone->name);
6050
6051 spin_lock_irqsave(&zone->lock, flags);
6052 for (order = 0; order < MAX_ORDER; order++) {
6053 struct free_area *area = &zone->free_area[order];
6054 int type;
6055
6056 nr[order] = area->nr_free;
6057 total += nr[order] << order;
6058
6059 types[order] = 0;
6060 for (type = 0; type < MIGRATE_TYPES; type++) {
6061 if (!free_area_empty(area, type))
6062 types[order] |= 1 << type;
6063 }
6064 }
6065 spin_unlock_irqrestore(&zone->lock, flags);
6066 for (order = 0; order < MAX_ORDER; order++) {
6067 printk(KERN_CONT "%lu*%lukB ",
6068 nr[order], K(1UL) << order);
6069 if (nr[order])
6070 show_migration_types(types[order]);
6071 }
6072 printk(KERN_CONT "= %lukB\n", K(total));
6073 }
6074
6075 hugetlb_show_meminfo();
6076
6077 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6078
6079 show_swap_cache_info();
6080}
6081
6082static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6083{
6084 zoneref->zone = zone;
6085 zoneref->zone_idx = zone_idx(zone);
6086}
6087
6088/*
6089 * Builds allocation fallback zone lists.
6090 *
6091 * Add all populated zones of a node to the zonelist.
6092 */
6093static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6094{
6095 struct zone *zone;
6096 enum zone_type zone_type = MAX_NR_ZONES;
6097 int nr_zones = 0;
6098
6099 do {
6100 zone_type--;
6101 zone = pgdat->node_zones + zone_type;
6102 if (managed_zone(zone)) {
6103 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6104 check_highest_zone(zone_type);
6105 }
6106 } while (zone_type);
6107
6108 return nr_zones;
6109}
6110
6111#ifdef CONFIG_NUMA
6112
6113static int __parse_numa_zonelist_order(char *s)
6114{
6115 /*
6116 * We used to support different zonelists modes but they turned
6117 * out to be just not useful. Let's keep the warning in place
6118 * if somebody still use the cmd line parameter so that we do
6119 * not fail it silently
6120 */
6121 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6122 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6123 return -EINVAL;
6124 }
6125 return 0;
6126}
6127
6128char numa_zonelist_order[] = "Node";
6129
6130/*
6131 * sysctl handler for numa_zonelist_order
6132 */
6133int numa_zonelist_order_handler(struct ctl_table *table, int write,
6134 void *buffer, size_t *length, loff_t *ppos)
6135{
6136 if (write)
6137 return __parse_numa_zonelist_order(buffer);
6138 return proc_dostring(table, write, buffer, length, ppos);
6139}
6140
6141
6142#define MAX_NODE_LOAD (nr_online_nodes)
6143static int node_load[MAX_NUMNODES];
6144
6145/**
6146 * find_next_best_node - find the next node that should appear in a given node's fallback list
6147 * @node: node whose fallback list we're appending
6148 * @used_node_mask: nodemask_t of already used nodes
6149 *
6150 * We use a number of factors to determine which is the next node that should
6151 * appear on a given node's fallback list. The node should not have appeared
6152 * already in @node's fallback list, and it should be the next closest node
6153 * according to the distance array (which contains arbitrary distance values
6154 * from each node to each node in the system), and should also prefer nodes
6155 * with no CPUs, since presumably they'll have very little allocation pressure
6156 * on them otherwise.
6157 *
6158 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6159 */
6160static int find_next_best_node(int node, nodemask_t *used_node_mask)
6161{
6162 int n, val;
6163 int min_val = INT_MAX;
6164 int best_node = NUMA_NO_NODE;
6165
6166 /* Use the local node if we haven't already */
6167 if (!node_isset(node, *used_node_mask)) {
6168 node_set(node, *used_node_mask);
6169 return node;
6170 }
6171
6172 for_each_node_state(n, N_MEMORY) {
6173
6174 /* Don't want a node to appear more than once */
6175 if (node_isset(n, *used_node_mask))
6176 continue;
6177
6178 /* Use the distance array to find the distance */
6179 val = node_distance(node, n);
6180
6181 /* Penalize nodes under us ("prefer the next node") */
6182 val += (n < node);
6183
6184 /* Give preference to headless and unused nodes */
6185 if (!cpumask_empty(cpumask_of_node(n)))
6186 val += PENALTY_FOR_NODE_WITH_CPUS;
6187
6188 /* Slight preference for less loaded node */
6189 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6190 val += node_load[n];
6191
6192 if (val < min_val) {
6193 min_val = val;
6194 best_node = n;
6195 }
6196 }
6197
6198 if (best_node >= 0)
6199 node_set(best_node, *used_node_mask);
6200
6201 return best_node;
6202}
6203
6204
6205/*
6206 * Build zonelists ordered by node and zones within node.
6207 * This results in maximum locality--normal zone overflows into local
6208 * DMA zone, if any--but risks exhausting DMA zone.
6209 */
6210static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6211 unsigned nr_nodes)
6212{
6213 struct zoneref *zonerefs;
6214 int i;
6215
6216 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6217
6218 for (i = 0; i < nr_nodes; i++) {
6219 int nr_zones;
6220
6221 pg_data_t *node = NODE_DATA(node_order[i]);
6222
6223 nr_zones = build_zonerefs_node(node, zonerefs);
6224 zonerefs += nr_zones;
6225 }
6226 zonerefs->zone = NULL;
6227 zonerefs->zone_idx = 0;
6228}
6229
6230/*
6231 * Build gfp_thisnode zonelists
6232 */
6233static void build_thisnode_zonelists(pg_data_t *pgdat)
6234{
6235 struct zoneref *zonerefs;
6236 int nr_zones;
6237
6238 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6239 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6240 zonerefs += nr_zones;
6241 zonerefs->zone = NULL;
6242 zonerefs->zone_idx = 0;
6243}
6244
6245/*
6246 * Build zonelists ordered by zone and nodes within zones.
6247 * This results in conserving DMA zone[s] until all Normal memory is
6248 * exhausted, but results in overflowing to remote node while memory
6249 * may still exist in local DMA zone.
6250 */
6251
6252static void build_zonelists(pg_data_t *pgdat)
6253{
6254 static int node_order[MAX_NUMNODES];
6255 int node, load, nr_nodes = 0;
6256 nodemask_t used_mask = NODE_MASK_NONE;
6257 int local_node, prev_node;
6258
6259 /* NUMA-aware ordering of nodes */
6260 local_node = pgdat->node_id;
6261 load = nr_online_nodes;
6262 prev_node = local_node;
6263
6264 memset(node_order, 0, sizeof(node_order));
6265 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6266 /*
6267 * We don't want to pressure a particular node.
6268 * So adding penalty to the first node in same
6269 * distance group to make it round-robin.
6270 */
6271 if (node_distance(local_node, node) !=
6272 node_distance(local_node, prev_node))
6273 node_load[node] = load;
6274
6275 node_order[nr_nodes++] = node;
6276 prev_node = node;
6277 load--;
6278 }
6279
6280 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6281 build_thisnode_zonelists(pgdat);
6282}
6283
6284#ifdef CONFIG_HAVE_MEMORYLESS_NODES
6285/*
6286 * Return node id of node used for "local" allocations.
6287 * I.e., first node id of first zone in arg node's generic zonelist.
6288 * Used for initializing percpu 'numa_mem', which is used primarily
6289 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6290 */
6291int local_memory_node(int node)
6292{
6293 struct zoneref *z;
6294
6295 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6296 gfp_zone(GFP_KERNEL),
6297 NULL);
6298 return zone_to_nid(z->zone);
6299}
6300#endif
6301
6302static void setup_min_unmapped_ratio(void);
6303static void setup_min_slab_ratio(void);
6304#else /* CONFIG_NUMA */
6305
6306static void build_zonelists(pg_data_t *pgdat)
6307{
6308 int node, local_node;
6309 struct zoneref *zonerefs;
6310 int nr_zones;
6311
6312 local_node = pgdat->node_id;
6313
6314 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6315 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6316 zonerefs += nr_zones;
6317
6318 /*
6319 * Now we build the zonelist so that it contains the zones
6320 * of all the other nodes.
6321 * We don't want to pressure a particular node, so when
6322 * building the zones for node N, we make sure that the
6323 * zones coming right after the local ones are those from
6324 * node N+1 (modulo N)
6325 */
6326 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6327 if (!node_online(node))
6328 continue;
6329 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6330 zonerefs += nr_zones;
6331 }
6332 for (node = 0; node < local_node; node++) {
6333 if (!node_online(node))
6334 continue;
6335 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6336 zonerefs += nr_zones;
6337 }
6338
6339 zonerefs->zone = NULL;
6340 zonerefs->zone_idx = 0;
6341}
6342
6343#endif /* CONFIG_NUMA */
6344
6345/*
6346 * Boot pageset table. One per cpu which is going to be used for all
6347 * zones and all nodes. The parameters will be set in such a way
6348 * that an item put on a list will immediately be handed over to
6349 * the buddy list. This is safe since pageset manipulation is done
6350 * with interrupts disabled.
6351 *
6352 * The boot_pagesets must be kept even after bootup is complete for
6353 * unused processors and/or zones. They do play a role for bootstrapping
6354 * hotplugged processors.
6355 *
6356 * zoneinfo_show() and maybe other functions do
6357 * not check if the processor is online before following the pageset pointer.
6358 * Other parts of the kernel may not check if the zone is available.
6359 */
6360static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6361/* These effectively disable the pcplists in the boot pageset completely */
6362#define BOOT_PAGESET_HIGH 0
6363#define BOOT_PAGESET_BATCH 1
6364static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6365static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6366static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6367
6368static void __build_all_zonelists(void *data)
6369{
6370 int nid;
6371 int __maybe_unused cpu;
6372 pg_data_t *self = data;
6373 static DEFINE_SPINLOCK(lock);
6374
6375 spin_lock(&lock);
6376
6377#ifdef CONFIG_NUMA
6378 memset(node_load, 0, sizeof(node_load));
6379#endif
6380
6381 /*
6382 * This node is hotadded and no memory is yet present. So just
6383 * building zonelists is fine - no need to touch other nodes.
6384 */
6385 if (self && !node_online(self->node_id)) {
6386 build_zonelists(self);
6387 } else {
6388 for_each_online_node(nid) {
6389 pg_data_t *pgdat = NODE_DATA(nid);
6390
6391 build_zonelists(pgdat);
6392 }
6393
6394#ifdef CONFIG_HAVE_MEMORYLESS_NODES
6395 /*
6396 * We now know the "local memory node" for each node--
6397 * i.e., the node of the first zone in the generic zonelist.
6398 * Set up numa_mem percpu variable for on-line cpus. During
6399 * boot, only the boot cpu should be on-line; we'll init the
6400 * secondary cpus' numa_mem as they come on-line. During
6401 * node/memory hotplug, we'll fixup all on-line cpus.
6402 */
6403 for_each_online_cpu(cpu)
6404 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6405#endif
6406 }
6407
6408 spin_unlock(&lock);
6409}
6410
6411static noinline void __init
6412build_all_zonelists_init(void)
6413{
6414 int cpu;
6415
6416 __build_all_zonelists(NULL);
6417
6418 /*
6419 * Initialize the boot_pagesets that are going to be used
6420 * for bootstrapping processors. The real pagesets for
6421 * each zone will be allocated later when the per cpu
6422 * allocator is available.
6423 *
6424 * boot_pagesets are used also for bootstrapping offline
6425 * cpus if the system is already booted because the pagesets
6426 * are needed to initialize allocators on a specific cpu too.
6427 * F.e. the percpu allocator needs the page allocator which
6428 * needs the percpu allocator in order to allocate its pagesets
6429 * (a chicken-egg dilemma).
6430 */
6431 for_each_possible_cpu(cpu)
6432 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6433
6434 mminit_verify_zonelist();
6435 cpuset_init_current_mems_allowed();
6436}
6437
6438/*
6439 * unless system_state == SYSTEM_BOOTING.
6440 *
6441 * __ref due to call of __init annotated helper build_all_zonelists_init
6442 * [protected by SYSTEM_BOOTING].
6443 */
6444void __ref build_all_zonelists(pg_data_t *pgdat)
6445{
6446 unsigned long vm_total_pages;
6447
6448 if (system_state == SYSTEM_BOOTING) {
6449 build_all_zonelists_init();
6450 } else {
6451 __build_all_zonelists(pgdat);
6452 /* cpuset refresh routine should be here */
6453 }
6454 /* Get the number of free pages beyond high watermark in all zones. */
6455 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6456 /*
6457 * Disable grouping by mobility if the number of pages in the
6458 * system is too low to allow the mechanism to work. It would be
6459 * more accurate, but expensive to check per-zone. This check is
6460 * made on memory-hotadd so a system can start with mobility
6461 * disabled and enable it later
6462 */
6463 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6464 page_group_by_mobility_disabled = 1;
6465 else
6466 page_group_by_mobility_disabled = 0;
6467
6468 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6469 nr_online_nodes,
6470 page_group_by_mobility_disabled ? "off" : "on",
6471 vm_total_pages);
6472#ifdef CONFIG_NUMA
6473 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6474#endif
6475}
6476
6477/* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6478static bool __meminit
6479overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6480{
6481 static struct memblock_region *r;
6482
6483 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6484 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6485 for_each_mem_region(r) {
6486 if (*pfn < memblock_region_memory_end_pfn(r))
6487 break;
6488 }
6489 }
6490 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6491 memblock_is_mirror(r)) {
6492 *pfn = memblock_region_memory_end_pfn(r);
6493 return true;
6494 }
6495 }
6496 return false;
6497}
6498
6499/*
6500 * Initially all pages are reserved - free ones are freed
6501 * up by memblock_free_all() once the early boot process is
6502 * done. Non-atomic initialization, single-pass.
6503 *
6504 * All aligned pageblocks are initialized to the specified migratetype
6505 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6506 * zone stats (e.g., nr_isolate_pageblock) are touched.
6507 */
6508void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6509 unsigned long start_pfn, unsigned long zone_end_pfn,
6510 enum meminit_context context,
6511 struct vmem_altmap *altmap, int migratetype)
6512{
6513 unsigned long pfn, end_pfn = start_pfn + size;
6514 struct page *page;
6515
6516 if (highest_memmap_pfn < end_pfn - 1)
6517 highest_memmap_pfn = end_pfn - 1;
6518
6519#ifdef CONFIG_ZONE_DEVICE
6520 /*
6521 * Honor reservation requested by the driver for this ZONE_DEVICE
6522 * memory. We limit the total number of pages to initialize to just
6523 * those that might contain the memory mapping. We will defer the
6524 * ZONE_DEVICE page initialization until after we have released
6525 * the hotplug lock.
6526 */
6527 if (zone == ZONE_DEVICE) {
6528 if (!altmap)
6529 return;
6530
6531 if (start_pfn == altmap->base_pfn)
6532 start_pfn += altmap->reserve;
6533 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6534 }
6535#endif
6536
6537 for (pfn = start_pfn; pfn < end_pfn; ) {
6538 /*
6539 * There can be holes in boot-time mem_map[]s handed to this
6540 * function. They do not exist on hotplugged memory.
6541 */
6542 if (context == MEMINIT_EARLY) {
6543 if (overlap_memmap_init(zone, &pfn))
6544 continue;
6545 if (defer_init(nid, pfn, zone_end_pfn))
6546 break;
6547 }
6548
6549 page = pfn_to_page(pfn);
6550 __init_single_page(page, pfn, zone, nid);
6551 if (context == MEMINIT_HOTPLUG)
6552 __SetPageReserved(page);
6553
6554 /*
6555 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6556 * such that unmovable allocations won't be scattered all
6557 * over the place during system boot.
6558 */
6559 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6560 set_pageblock_migratetype(page, migratetype);
6561 cond_resched();
6562 }
6563 pfn++;
6564 }
6565}
6566
6567#ifdef CONFIG_ZONE_DEVICE
6568void __ref memmap_init_zone_device(struct zone *zone,
6569 unsigned long start_pfn,
6570 unsigned long nr_pages,
6571 struct dev_pagemap *pgmap)
6572{
6573 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6574 struct pglist_data *pgdat = zone->zone_pgdat;
6575 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6576 unsigned long zone_idx = zone_idx(zone);
6577 unsigned long start = jiffies;
6578 int nid = pgdat->node_id;
6579
6580 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6581 return;
6582
6583 /*
6584 * The call to memmap_init should have already taken care
6585 * of the pages reserved for the memmap, so we can just jump to
6586 * the end of that region and start processing the device pages.
6587 */
6588 if (altmap) {
6589 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6590 nr_pages = end_pfn - start_pfn;
6591 }
6592
6593 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6594 struct page *page = pfn_to_page(pfn);
6595
6596 __init_single_page(page, pfn, zone_idx, nid);
6597
6598 /*
6599 * Mark page reserved as it will need to wait for onlining
6600 * phase for it to be fully associated with a zone.
6601 *
6602 * We can use the non-atomic __set_bit operation for setting
6603 * the flag as we are still initializing the pages.
6604 */
6605 __SetPageReserved(page);
6606
6607 /*
6608 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6609 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6610 * ever freed or placed on a driver-private list.
6611 */
6612 page->pgmap = pgmap;
6613 page->zone_device_data = NULL;
6614
6615 /*
6616 * Mark the block movable so that blocks are reserved for
6617 * movable at startup. This will force kernel allocations
6618 * to reserve their blocks rather than leaking throughout
6619 * the address space during boot when many long-lived
6620 * kernel allocations are made.
6621 *
6622 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6623 * because this is done early in section_activate()
6624 */
6625 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6626 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6627 cond_resched();
6628 }
6629 }
6630
6631 pr_info("%s initialised %lu pages in %ums\n", __func__,
6632 nr_pages, jiffies_to_msecs(jiffies - start));
6633}
6634
6635#endif
6636static void __meminit zone_init_free_lists(struct zone *zone)
6637{
6638 unsigned int order, t;
6639 for_each_migratetype_order(order, t) {
6640 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6641 zone->free_area[order].nr_free = 0;
6642 }
6643}
6644
6645#if !defined(CONFIG_FLATMEM)
6646/*
6647 * Only struct pages that correspond to ranges defined by memblock.memory
6648 * are zeroed and initialized by going through __init_single_page() during
6649 * memmap_init_zone_range().
6650 *
6651 * But, there could be struct pages that correspond to holes in
6652 * memblock.memory. This can happen because of the following reasons:
6653 * - physical memory bank size is not necessarily the exact multiple of the
6654 * arbitrary section size
6655 * - early reserved memory may not be listed in memblock.memory
6656 * - memory layouts defined with memmap= kernel parameter may not align
6657 * nicely with memmap sections
6658 *
6659 * Explicitly initialize those struct pages so that:
6660 * - PG_Reserved is set
6661 * - zone and node links point to zone and node that span the page if the
6662 * hole is in the middle of a zone
6663 * - zone and node links point to adjacent zone/node if the hole falls on
6664 * the zone boundary; the pages in such holes will be prepended to the
6665 * zone/node above the hole except for the trailing pages in the last
6666 * section that will be appended to the zone/node below.
6667 */
6668static void __init init_unavailable_range(unsigned long spfn,
6669 unsigned long epfn,
6670 int zone, int node)
6671{
6672 unsigned long pfn;
6673 u64 pgcnt = 0;
6674
6675 for (pfn = spfn; pfn < epfn; pfn++) {
6676 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6677 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6678 + pageblock_nr_pages - 1;
6679 continue;
6680 }
6681 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6682 __SetPageReserved(pfn_to_page(pfn));
6683 pgcnt++;
6684 }
6685
6686 if (pgcnt)
6687 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6688 node, zone_names[zone], pgcnt);
6689}
6690#else
6691static inline void init_unavailable_range(unsigned long spfn,
6692 unsigned long epfn,
6693 int zone, int node)
6694{
6695}
6696#endif
6697
6698static void __init memmap_init_zone_range(struct zone *zone,
6699 unsigned long start_pfn,
6700 unsigned long end_pfn,
6701 unsigned long *hole_pfn)
6702{
6703 unsigned long zone_start_pfn = zone->zone_start_pfn;
6704 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6705 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6706
6707 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6708 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6709
6710 if (start_pfn >= end_pfn)
6711 return;
6712
6713 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6714 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6715
6716 if (*hole_pfn < start_pfn)
6717 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6718
6719 *hole_pfn = end_pfn;
6720}
6721
6722static void __init memmap_init(void)
6723{
6724 unsigned long start_pfn, end_pfn;
6725 unsigned long hole_pfn = 0;
6726 int i, j, zone_id, nid;
6727
6728 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6729 struct pglist_data *node = NODE_DATA(nid);
6730
6731 for (j = 0; j < MAX_NR_ZONES; j++) {
6732 struct zone *zone = node->node_zones + j;
6733
6734 if (!populated_zone(zone))
6735 continue;
6736
6737 memmap_init_zone_range(zone, start_pfn, end_pfn,
6738 &hole_pfn);
6739 zone_id = j;
6740 }
6741 }
6742
6743#ifdef CONFIG_SPARSEMEM
6744 /*
6745 * Initialize the memory map for hole in the range [memory_end,
6746 * section_end].
6747 * Append the pages in this hole to the highest zone in the last
6748 * node.
6749 * The call to init_unavailable_range() is outside the ifdef to
6750 * silence the compiler warining about zone_id set but not used;
6751 * for FLATMEM it is a nop anyway
6752 */
6753 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6754 if (hole_pfn < end_pfn)
6755#endif
6756 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6757}
6758
6759static int zone_batchsize(struct zone *zone)
6760{
6761#ifdef CONFIG_MMU
6762 int batch;
6763
6764 /*
6765 * The number of pages to batch allocate is either ~0.1%
6766 * of the zone or 1MB, whichever is smaller. The batch
6767 * size is striking a balance between allocation latency
6768 * and zone lock contention.
6769 */
6770 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6771 batch /= 4; /* We effectively *= 4 below */
6772 if (batch < 1)
6773 batch = 1;
6774
6775 /*
6776 * Clamp the batch to a 2^n - 1 value. Having a power
6777 * of 2 value was found to be more likely to have
6778 * suboptimal cache aliasing properties in some cases.
6779 *
6780 * For example if 2 tasks are alternately allocating
6781 * batches of pages, one task can end up with a lot
6782 * of pages of one half of the possible page colors
6783 * and the other with pages of the other colors.
6784 */
6785 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6786
6787 return batch;
6788
6789#else
6790 /* The deferral and batching of frees should be suppressed under NOMMU
6791 * conditions.
6792 *
6793 * The problem is that NOMMU needs to be able to allocate large chunks
6794 * of contiguous memory as there's no hardware page translation to
6795 * assemble apparent contiguous memory from discontiguous pages.
6796 *
6797 * Queueing large contiguous runs of pages for batching, however,
6798 * causes the pages to actually be freed in smaller chunks. As there
6799 * can be a significant delay between the individual batches being
6800 * recycled, this leads to the once large chunks of space being
6801 * fragmented and becoming unavailable for high-order allocations.
6802 */
6803 return 0;
6804#endif
6805}
6806
6807static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6808{
6809#ifdef CONFIG_MMU
6810 int high;
6811 int nr_split_cpus;
6812 unsigned long total_pages;
6813
6814 if (!percpu_pagelist_high_fraction) {
6815 /*
6816 * By default, the high value of the pcp is based on the zone
6817 * low watermark so that if they are full then background
6818 * reclaim will not be started prematurely.
6819 */
6820 total_pages = low_wmark_pages(zone);
6821 } else {
6822 /*
6823 * If percpu_pagelist_high_fraction is configured, the high
6824 * value is based on a fraction of the managed pages in the
6825 * zone.
6826 */
6827 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6828 }
6829
6830 /*
6831 * Split the high value across all online CPUs local to the zone. Note
6832 * that early in boot that CPUs may not be online yet and that during
6833 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6834 * onlined. For memory nodes that have no CPUs, split pcp->high across
6835 * all online CPUs to mitigate the risk that reclaim is triggered
6836 * prematurely due to pages stored on pcp lists.
6837 */
6838 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6839 if (!nr_split_cpus)
6840 nr_split_cpus = num_online_cpus();
6841 high = total_pages / nr_split_cpus;
6842
6843 /*
6844 * Ensure high is at least batch*4. The multiple is based on the
6845 * historical relationship between high and batch.
6846 */
6847 high = max(high, batch << 2);
6848
6849 return high;
6850#else
6851 return 0;
6852#endif
6853}
6854
6855/*
6856 * pcp->high and pcp->batch values are related and generally batch is lower
6857 * than high. They are also related to pcp->count such that count is lower
6858 * than high, and as soon as it reaches high, the pcplist is flushed.
6859 *
6860 * However, guaranteeing these relations at all times would require e.g. write
6861 * barriers here but also careful usage of read barriers at the read side, and
6862 * thus be prone to error and bad for performance. Thus the update only prevents
6863 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6864 * can cope with those fields changing asynchronously, and fully trust only the
6865 * pcp->count field on the local CPU with interrupts disabled.
6866 *
6867 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6868 * outside of boot time (or some other assurance that no concurrent updaters
6869 * exist).
6870 */
6871static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6872 unsigned long batch)
6873{
6874 WRITE_ONCE(pcp->batch, batch);
6875 WRITE_ONCE(pcp->high, high);
6876}
6877
6878static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6879{
6880 int pindex;
6881
6882 memset(pcp, 0, sizeof(*pcp));
6883 memset(pzstats, 0, sizeof(*pzstats));
6884
6885 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6886 INIT_LIST_HEAD(&pcp->lists[pindex]);
6887
6888 /*
6889 * Set batch and high values safe for a boot pageset. A true percpu
6890 * pageset's initialization will update them subsequently. Here we don't
6891 * need to be as careful as pageset_update() as nobody can access the
6892 * pageset yet.
6893 */
6894 pcp->high = BOOT_PAGESET_HIGH;
6895 pcp->batch = BOOT_PAGESET_BATCH;
6896 pcp->free_factor = 0;
6897}
6898
6899static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6900 unsigned long batch)
6901{
6902 struct per_cpu_pages *pcp;
6903 int cpu;
6904
6905 for_each_possible_cpu(cpu) {
6906 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6907 pageset_update(pcp, high, batch);
6908 }
6909}
6910
6911/*
6912 * Calculate and set new high and batch values for all per-cpu pagesets of a
6913 * zone based on the zone's size.
6914 */
6915static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6916{
6917 int new_high, new_batch;
6918
6919 new_batch = max(1, zone_batchsize(zone));
6920 new_high = zone_highsize(zone, new_batch, cpu_online);
6921
6922 if (zone->pageset_high == new_high &&
6923 zone->pageset_batch == new_batch)
6924 return;
6925
6926 zone->pageset_high = new_high;
6927 zone->pageset_batch = new_batch;
6928
6929 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6930}
6931
6932void __meminit setup_zone_pageset(struct zone *zone)
6933{
6934 int cpu;
6935
6936 /* Size may be 0 on !SMP && !NUMA */
6937 if (sizeof(struct per_cpu_zonestat) > 0)
6938 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6939
6940 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6941 for_each_possible_cpu(cpu) {
6942 struct per_cpu_pages *pcp;
6943 struct per_cpu_zonestat *pzstats;
6944
6945 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6946 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6947 per_cpu_pages_init(pcp, pzstats);
6948 }
6949
6950 zone_set_pageset_high_and_batch(zone, 0);
6951}
6952
6953/*
6954 * Allocate per cpu pagesets and initialize them.
6955 * Before this call only boot pagesets were available.
6956 */
6957void __init setup_per_cpu_pageset(void)
6958{
6959 struct pglist_data *pgdat;
6960 struct zone *zone;
6961 int __maybe_unused cpu;
6962
6963 for_each_populated_zone(zone)
6964 setup_zone_pageset(zone);
6965
6966#ifdef CONFIG_NUMA
6967 /*
6968 * Unpopulated zones continue using the boot pagesets.
6969 * The numa stats for these pagesets need to be reset.
6970 * Otherwise, they will end up skewing the stats of
6971 * the nodes these zones are associated with.
6972 */
6973 for_each_possible_cpu(cpu) {
6974 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
6975 memset(pzstats->vm_numa_event, 0,
6976 sizeof(pzstats->vm_numa_event));
6977 }
6978#endif
6979
6980 for_each_online_pgdat(pgdat)
6981 pgdat->per_cpu_nodestats =
6982 alloc_percpu(struct per_cpu_nodestat);
6983}
6984
6985static __meminit void zone_pcp_init(struct zone *zone)
6986{
6987 /*
6988 * per cpu subsystem is not up at this point. The following code
6989 * relies on the ability of the linker to provide the
6990 * offset of a (static) per cpu variable into the per cpu area.
6991 */
6992 zone->per_cpu_pageset = &boot_pageset;
6993 zone->per_cpu_zonestats = &boot_zonestats;
6994 zone->pageset_high = BOOT_PAGESET_HIGH;
6995 zone->pageset_batch = BOOT_PAGESET_BATCH;
6996
6997 if (populated_zone(zone))
6998 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
6999 zone->present_pages, zone_batchsize(zone));
7000}
7001
7002void __meminit init_currently_empty_zone(struct zone *zone,
7003 unsigned long zone_start_pfn,
7004 unsigned long size)
7005{
7006 struct pglist_data *pgdat = zone->zone_pgdat;
7007 int zone_idx = zone_idx(zone) + 1;
7008
7009 if (zone_idx > pgdat->nr_zones)
7010 pgdat->nr_zones = zone_idx;
7011
7012 zone->zone_start_pfn = zone_start_pfn;
7013
7014 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7015 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7016 pgdat->node_id,
7017 (unsigned long)zone_idx(zone),
7018 zone_start_pfn, (zone_start_pfn + size));
7019
7020 zone_init_free_lists(zone);
7021 zone->initialized = 1;
7022}
7023
7024/**
7025 * get_pfn_range_for_nid - Return the start and end page frames for a node
7026 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7027 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7028 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7029 *
7030 * It returns the start and end page frame of a node based on information
7031 * provided by memblock_set_node(). If called for a node
7032 * with no available memory, a warning is printed and the start and end
7033 * PFNs will be 0.
7034 */
7035void __init get_pfn_range_for_nid(unsigned int nid,
7036 unsigned long *start_pfn, unsigned long *end_pfn)
7037{
7038 unsigned long this_start_pfn, this_end_pfn;
7039 int i;
7040
7041 *start_pfn = -1UL;
7042 *end_pfn = 0;
7043
7044 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7045 *start_pfn = min(*start_pfn, this_start_pfn);
7046 *end_pfn = max(*end_pfn, this_end_pfn);
7047 }
7048
7049 if (*start_pfn == -1UL)
7050 *start_pfn = 0;
7051}
7052
7053/*
7054 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7055 * assumption is made that zones within a node are ordered in monotonic
7056 * increasing memory addresses so that the "highest" populated zone is used
7057 */
7058static void __init find_usable_zone_for_movable(void)
7059{
7060 int zone_index;
7061 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7062 if (zone_index == ZONE_MOVABLE)
7063 continue;
7064
7065 if (arch_zone_highest_possible_pfn[zone_index] >
7066 arch_zone_lowest_possible_pfn[zone_index])
7067 break;
7068 }
7069
7070 VM_BUG_ON(zone_index == -1);
7071 movable_zone = zone_index;
7072}
7073
7074/*
7075 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7076 * because it is sized independent of architecture. Unlike the other zones,
7077 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7078 * in each node depending on the size of each node and how evenly kernelcore
7079 * is distributed. This helper function adjusts the zone ranges
7080 * provided by the architecture for a given node by using the end of the
7081 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7082 * zones within a node are in order of monotonic increases memory addresses
7083 */
7084static void __init adjust_zone_range_for_zone_movable(int nid,
7085 unsigned long zone_type,
7086 unsigned long node_start_pfn,
7087 unsigned long node_end_pfn,
7088 unsigned long *zone_start_pfn,
7089 unsigned long *zone_end_pfn)
7090{
7091 /* Only adjust if ZONE_MOVABLE is on this node */
7092 if (zone_movable_pfn[nid]) {
7093 /* Size ZONE_MOVABLE */
7094 if (zone_type == ZONE_MOVABLE) {
7095 *zone_start_pfn = zone_movable_pfn[nid];
7096 *zone_end_pfn = min(node_end_pfn,
7097 arch_zone_highest_possible_pfn[movable_zone]);
7098
7099 /* Adjust for ZONE_MOVABLE starting within this range */
7100 } else if (!mirrored_kernelcore &&
7101 *zone_start_pfn < zone_movable_pfn[nid] &&
7102 *zone_end_pfn > zone_movable_pfn[nid]) {
7103 *zone_end_pfn = zone_movable_pfn[nid];
7104
7105 /* Check if this whole range is within ZONE_MOVABLE */
7106 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7107 *zone_start_pfn = *zone_end_pfn;
7108 }
7109}
7110
7111/*
7112 * Return the number of pages a zone spans in a node, including holes
7113 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7114 */
7115static unsigned long __init zone_spanned_pages_in_node(int nid,
7116 unsigned long zone_type,
7117 unsigned long node_start_pfn,
7118 unsigned long node_end_pfn,
7119 unsigned long *zone_start_pfn,
7120 unsigned long *zone_end_pfn)
7121{
7122 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7123 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7124 /* When hotadd a new node from cpu_up(), the node should be empty */
7125 if (!node_start_pfn && !node_end_pfn)
7126 return 0;
7127
7128 /* Get the start and end of the zone */
7129 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7130 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7131 adjust_zone_range_for_zone_movable(nid, zone_type,
7132 node_start_pfn, node_end_pfn,
7133 zone_start_pfn, zone_end_pfn);
7134
7135 /* Check that this node has pages within the zone's required range */
7136 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7137 return 0;
7138
7139 /* Move the zone boundaries inside the node if necessary */
7140 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7141 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7142
7143 /* Return the spanned pages */
7144 return *zone_end_pfn - *zone_start_pfn;
7145}
7146
7147/*
7148 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7149 * then all holes in the requested range will be accounted for.
7150 */
7151unsigned long __init __absent_pages_in_range(int nid,
7152 unsigned long range_start_pfn,
7153 unsigned long range_end_pfn)
7154{
7155 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7156 unsigned long start_pfn, end_pfn;
7157 int i;
7158
7159 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7160 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7161 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7162 nr_absent -= end_pfn - start_pfn;
7163 }
7164 return nr_absent;
7165}
7166
7167/**
7168 * absent_pages_in_range - Return number of page frames in holes within a range
7169 * @start_pfn: The start PFN to start searching for holes
7170 * @end_pfn: The end PFN to stop searching for holes
7171 *
7172 * Return: the number of pages frames in memory holes within a range.
7173 */
7174unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7175 unsigned long end_pfn)
7176{
7177 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7178}
7179
7180/* Return the number of page frames in holes in a zone on a node */
7181static unsigned long __init zone_absent_pages_in_node(int nid,
7182 unsigned long zone_type,
7183 unsigned long node_start_pfn,
7184 unsigned long node_end_pfn)
7185{
7186 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7187 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7188 unsigned long zone_start_pfn, zone_end_pfn;
7189 unsigned long nr_absent;
7190
7191 /* When hotadd a new node from cpu_up(), the node should be empty */
7192 if (!node_start_pfn && !node_end_pfn)
7193 return 0;
7194
7195 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7196 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7197
7198 adjust_zone_range_for_zone_movable(nid, zone_type,
7199 node_start_pfn, node_end_pfn,
7200 &zone_start_pfn, &zone_end_pfn);
7201 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7202
7203 /*
7204 * ZONE_MOVABLE handling.
7205 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7206 * and vice versa.
7207 */
7208 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7209 unsigned long start_pfn, end_pfn;
7210 struct memblock_region *r;
7211
7212 for_each_mem_region(r) {
7213 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7214 zone_start_pfn, zone_end_pfn);
7215 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7216 zone_start_pfn, zone_end_pfn);
7217
7218 if (zone_type == ZONE_MOVABLE &&
7219 memblock_is_mirror(r))
7220 nr_absent += end_pfn - start_pfn;
7221
7222 if (zone_type == ZONE_NORMAL &&
7223 !memblock_is_mirror(r))
7224 nr_absent += end_pfn - start_pfn;
7225 }
7226 }
7227
7228 return nr_absent;
7229}
7230
7231static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7232 unsigned long node_start_pfn,
7233 unsigned long node_end_pfn)
7234{
7235 unsigned long realtotalpages = 0, totalpages = 0;
7236 enum zone_type i;
7237
7238 for (i = 0; i < MAX_NR_ZONES; i++) {
7239 struct zone *zone = pgdat->node_zones + i;
7240 unsigned long zone_start_pfn, zone_end_pfn;
7241 unsigned long spanned, absent;
7242 unsigned long size, real_size;
7243
7244 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7245 node_start_pfn,
7246 node_end_pfn,
7247 &zone_start_pfn,
7248 &zone_end_pfn);
7249 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7250 node_start_pfn,
7251 node_end_pfn);
7252
7253 size = spanned;
7254 real_size = size - absent;
7255
7256 if (size)
7257 zone->zone_start_pfn = zone_start_pfn;
7258 else
7259 zone->zone_start_pfn = 0;
7260 zone->spanned_pages = size;
7261 zone->present_pages = real_size;
7262
7263 totalpages += size;
7264 realtotalpages += real_size;
7265 }
7266
7267 pgdat->node_spanned_pages = totalpages;
7268 pgdat->node_present_pages = realtotalpages;
7269 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7270}
7271
7272#ifndef CONFIG_SPARSEMEM
7273/*
7274 * Calculate the size of the zone->blockflags rounded to an unsigned long
7275 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7276 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7277 * round what is now in bits to nearest long in bits, then return it in
7278 * bytes.
7279 */
7280static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7281{
7282 unsigned long usemapsize;
7283
7284 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7285 usemapsize = roundup(zonesize, pageblock_nr_pages);
7286 usemapsize = usemapsize >> pageblock_order;
7287 usemapsize *= NR_PAGEBLOCK_BITS;
7288 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7289
7290 return usemapsize / 8;
7291}
7292
7293static void __ref setup_usemap(struct zone *zone)
7294{
7295 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7296 zone->spanned_pages);
7297 zone->pageblock_flags = NULL;
7298 if (usemapsize) {
7299 zone->pageblock_flags =
7300 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7301 zone_to_nid(zone));
7302 if (!zone->pageblock_flags)
7303 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7304 usemapsize, zone->name, zone_to_nid(zone));
7305 }
7306}
7307#else
7308static inline void setup_usemap(struct zone *zone) {}
7309#endif /* CONFIG_SPARSEMEM */
7310
7311#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7312
7313/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7314void __init set_pageblock_order(void)
7315{
7316 unsigned int order;
7317
7318 /* Check that pageblock_nr_pages has not already been setup */
7319 if (pageblock_order)
7320 return;
7321
7322 if (HPAGE_SHIFT > PAGE_SHIFT)
7323 order = HUGETLB_PAGE_ORDER;
7324 else
7325 order = MAX_ORDER - 1;
7326
7327 /*
7328 * Assume the largest contiguous order of interest is a huge page.
7329 * This value may be variable depending on boot parameters on IA64 and
7330 * powerpc.
7331 */
7332 pageblock_order = order;
7333}
7334#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7335
7336/*
7337 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7338 * is unused as pageblock_order is set at compile-time. See
7339 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7340 * the kernel config
7341 */
7342void __init set_pageblock_order(void)
7343{
7344}
7345
7346#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7347
7348static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7349 unsigned long present_pages)
7350{
7351 unsigned long pages = spanned_pages;
7352
7353 /*
7354 * Provide a more accurate estimation if there are holes within
7355 * the zone and SPARSEMEM is in use. If there are holes within the
7356 * zone, each populated memory region may cost us one or two extra
7357 * memmap pages due to alignment because memmap pages for each
7358 * populated regions may not be naturally aligned on page boundary.
7359 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7360 */
7361 if (spanned_pages > present_pages + (present_pages >> 4) &&
7362 IS_ENABLED(CONFIG_SPARSEMEM))
7363 pages = present_pages;
7364
7365 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7366}
7367
7368#ifdef CONFIG_TRANSPARENT_HUGEPAGE
7369static void pgdat_init_split_queue(struct pglist_data *pgdat)
7370{
7371 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7372
7373 spin_lock_init(&ds_queue->split_queue_lock);
7374 INIT_LIST_HEAD(&ds_queue->split_queue);
7375 ds_queue->split_queue_len = 0;
7376}
7377#else
7378static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7379#endif
7380
7381#ifdef CONFIG_COMPACTION
7382static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7383{
7384 init_waitqueue_head(&pgdat->kcompactd_wait);
7385}
7386#else
7387static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7388#endif
7389
7390static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7391{
7392 pgdat_resize_init(pgdat);
7393
7394 pgdat_init_split_queue(pgdat);
7395 pgdat_init_kcompactd(pgdat);
7396
7397 init_waitqueue_head(&pgdat->kswapd_wait);
7398 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7399
7400 pgdat_page_ext_init(pgdat);
7401 lruvec_init(&pgdat->__lruvec);
7402}
7403
7404static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7405 unsigned long remaining_pages)
7406{
7407 atomic_long_set(&zone->managed_pages, remaining_pages);
7408 zone_set_nid(zone, nid);
7409 zone->name = zone_names[idx];
7410 zone->zone_pgdat = NODE_DATA(nid);
7411 spin_lock_init(&zone->lock);
7412 zone_seqlock_init(zone);
7413 zone_pcp_init(zone);
7414}
7415
7416/*
7417 * Set up the zone data structures
7418 * - init pgdat internals
7419 * - init all zones belonging to this node
7420 *
7421 * NOTE: this function is only called during memory hotplug
7422 */
7423#ifdef CONFIG_MEMORY_HOTPLUG
7424void __ref free_area_init_core_hotplug(int nid)
7425{
7426 enum zone_type z;
7427 pg_data_t *pgdat = NODE_DATA(nid);
7428
7429 pgdat_init_internals(pgdat);
7430 for (z = 0; z < MAX_NR_ZONES; z++)
7431 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7432}
7433#endif
7434
7435/*
7436 * Set up the zone data structures:
7437 * - mark all pages reserved
7438 * - mark all memory queues empty
7439 * - clear the memory bitmaps
7440 *
7441 * NOTE: pgdat should get zeroed by caller.
7442 * NOTE: this function is only called during early init.
7443 */
7444static void __init free_area_init_core(struct pglist_data *pgdat)
7445{
7446 enum zone_type j;
7447 int nid = pgdat->node_id;
7448
7449 pgdat_init_internals(pgdat);
7450 pgdat->per_cpu_nodestats = &boot_nodestats;
7451
7452 for (j = 0; j < MAX_NR_ZONES; j++) {
7453 struct zone *zone = pgdat->node_zones + j;
7454 unsigned long size, freesize, memmap_pages;
7455
7456 size = zone->spanned_pages;
7457 freesize = zone->present_pages;
7458
7459 /*
7460 * Adjust freesize so that it accounts for how much memory
7461 * is used by this zone for memmap. This affects the watermark
7462 * and per-cpu initialisations
7463 */
7464 memmap_pages = calc_memmap_size(size, freesize);
7465 if (!is_highmem_idx(j)) {
7466 if (freesize >= memmap_pages) {
7467 freesize -= memmap_pages;
7468 if (memmap_pages)
7469 pr_debug(" %s zone: %lu pages used for memmap\n",
7470 zone_names[j], memmap_pages);
7471 } else
7472 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7473 zone_names[j], memmap_pages, freesize);
7474 }
7475
7476 /* Account for reserved pages */
7477 if (j == 0 && freesize > dma_reserve) {
7478 freesize -= dma_reserve;
7479 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7480 }
7481
7482 if (!is_highmem_idx(j))
7483 nr_kernel_pages += freesize;
7484 /* Charge for highmem memmap if there are enough kernel pages */
7485 else if (nr_kernel_pages > memmap_pages * 2)
7486 nr_kernel_pages -= memmap_pages;
7487 nr_all_pages += freesize;
7488
7489 /*
7490 * Set an approximate value for lowmem here, it will be adjusted
7491 * when the bootmem allocator frees pages into the buddy system.
7492 * And all highmem pages will be managed by the buddy system.
7493 */
7494 zone_init_internals(zone, j, nid, freesize);
7495
7496 if (!size)
7497 continue;
7498
7499 set_pageblock_order();
7500 setup_usemap(zone);
7501 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7502 }
7503}
7504
7505#ifdef CONFIG_FLATMEM
7506static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7507{
7508 unsigned long __maybe_unused start = 0;
7509 unsigned long __maybe_unused offset = 0;
7510
7511 /* Skip empty nodes */
7512 if (!pgdat->node_spanned_pages)
7513 return;
7514
7515 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7516 offset = pgdat->node_start_pfn - start;
7517 /* ia64 gets its own node_mem_map, before this, without bootmem */
7518 if (!pgdat->node_mem_map) {
7519 unsigned long size, end;
7520 struct page *map;
7521
7522 /*
7523 * The zone's endpoints aren't required to be MAX_ORDER
7524 * aligned but the node_mem_map endpoints must be in order
7525 * for the buddy allocator to function correctly.
7526 */
7527 end = pgdat_end_pfn(pgdat);
7528 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7529 size = (end - start) * sizeof(struct page);
7530 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7531 pgdat->node_id);
7532 if (!map)
7533 panic("Failed to allocate %ld bytes for node %d memory map\n",
7534 size, pgdat->node_id);
7535 pgdat->node_mem_map = map + offset;
7536 }
7537 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7538 __func__, pgdat->node_id, (unsigned long)pgdat,
7539 (unsigned long)pgdat->node_mem_map);
7540#ifndef CONFIG_NUMA
7541 /*
7542 * With no DISCONTIG, the global mem_map is just set as node 0's
7543 */
7544 if (pgdat == NODE_DATA(0)) {
7545 mem_map = NODE_DATA(0)->node_mem_map;
7546 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7547 mem_map -= offset;
7548 }
7549#endif
7550}
7551#else
7552static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7553#endif /* CONFIG_FLATMEM */
7554
7555#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7556static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7557{
7558 pgdat->first_deferred_pfn = ULONG_MAX;
7559}
7560#else
7561static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7562#endif
7563
7564static void __init free_area_init_node(int nid)
7565{
7566 pg_data_t *pgdat = NODE_DATA(nid);
7567 unsigned long start_pfn = 0;
7568 unsigned long end_pfn = 0;
7569
7570 /* pg_data_t should be reset to zero when it's allocated */
7571 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7572
7573 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7574
7575 pgdat->node_id = nid;
7576 pgdat->node_start_pfn = start_pfn;
7577 pgdat->per_cpu_nodestats = NULL;
7578
7579 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7580 (u64)start_pfn << PAGE_SHIFT,
7581 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7582 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7583
7584 alloc_node_mem_map(pgdat);
7585 pgdat_set_deferred_range(pgdat);
7586
7587 free_area_init_core(pgdat);
7588}
7589
7590void __init free_area_init_memoryless_node(int nid)
7591{
7592 free_area_init_node(nid);
7593}
7594
7595#if MAX_NUMNODES > 1
7596/*
7597 * Figure out the number of possible node ids.
7598 */
7599void __init setup_nr_node_ids(void)
7600{
7601 unsigned int highest;
7602
7603 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7604 nr_node_ids = highest + 1;
7605}
7606#endif
7607
7608/**
7609 * node_map_pfn_alignment - determine the maximum internode alignment
7610 *
7611 * This function should be called after node map is populated and sorted.
7612 * It calculates the maximum power of two alignment which can distinguish
7613 * all the nodes.
7614 *
7615 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7616 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7617 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7618 * shifted, 1GiB is enough and this function will indicate so.
7619 *
7620 * This is used to test whether pfn -> nid mapping of the chosen memory
7621 * model has fine enough granularity to avoid incorrect mapping for the
7622 * populated node map.
7623 *
7624 * Return: the determined alignment in pfn's. 0 if there is no alignment
7625 * requirement (single node).
7626 */
7627unsigned long __init node_map_pfn_alignment(void)
7628{
7629 unsigned long accl_mask = 0, last_end = 0;
7630 unsigned long start, end, mask;
7631 int last_nid = NUMA_NO_NODE;
7632 int i, nid;
7633
7634 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7635 if (!start || last_nid < 0 || last_nid == nid) {
7636 last_nid = nid;
7637 last_end = end;
7638 continue;
7639 }
7640
7641 /*
7642 * Start with a mask granular enough to pin-point to the
7643 * start pfn and tick off bits one-by-one until it becomes
7644 * too coarse to separate the current node from the last.
7645 */
7646 mask = ~((1 << __ffs(start)) - 1);
7647 while (mask && last_end <= (start & (mask << 1)))
7648 mask <<= 1;
7649
7650 /* accumulate all internode masks */
7651 accl_mask |= mask;
7652 }
7653
7654 /* convert mask to number of pages */
7655 return ~accl_mask + 1;
7656}
7657
7658/**
7659 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7660 *
7661 * Return: the minimum PFN based on information provided via
7662 * memblock_set_node().
7663 */
7664unsigned long __init find_min_pfn_with_active_regions(void)
7665{
7666 return PHYS_PFN(memblock_start_of_DRAM());
7667}
7668
7669/*
7670 * early_calculate_totalpages()
7671 * Sum pages in active regions for movable zone.
7672 * Populate N_MEMORY for calculating usable_nodes.
7673 */
7674static unsigned long __init early_calculate_totalpages(void)
7675{
7676 unsigned long totalpages = 0;
7677 unsigned long start_pfn, end_pfn;
7678 int i, nid;
7679
7680 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7681 unsigned long pages = end_pfn - start_pfn;
7682
7683 totalpages += pages;
7684 if (pages)
7685 node_set_state(nid, N_MEMORY);
7686 }
7687 return totalpages;
7688}
7689
7690/*
7691 * Find the PFN the Movable zone begins in each node. Kernel memory
7692 * is spread evenly between nodes as long as the nodes have enough
7693 * memory. When they don't, some nodes will have more kernelcore than
7694 * others
7695 */
7696static void __init find_zone_movable_pfns_for_nodes(void)
7697{
7698 int i, nid;
7699 unsigned long usable_startpfn;
7700 unsigned long kernelcore_node, kernelcore_remaining;
7701 /* save the state before borrow the nodemask */
7702 nodemask_t saved_node_state = node_states[N_MEMORY];
7703 unsigned long totalpages = early_calculate_totalpages();
7704 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7705 struct memblock_region *r;
7706
7707 /* Need to find movable_zone earlier when movable_node is specified. */
7708 find_usable_zone_for_movable();
7709
7710 /*
7711 * If movable_node is specified, ignore kernelcore and movablecore
7712 * options.
7713 */
7714 if (movable_node_is_enabled()) {
7715 for_each_mem_region(r) {
7716 if (!memblock_is_hotpluggable(r))
7717 continue;
7718
7719 nid = memblock_get_region_node(r);
7720
7721 usable_startpfn = PFN_DOWN(r->base);
7722 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7723 min(usable_startpfn, zone_movable_pfn[nid]) :
7724 usable_startpfn;
7725 }
7726
7727 goto out2;
7728 }
7729
7730 /*
7731 * If kernelcore=mirror is specified, ignore movablecore option
7732 */
7733 if (mirrored_kernelcore) {
7734 bool mem_below_4gb_not_mirrored = false;
7735
7736 for_each_mem_region(r) {
7737 if (memblock_is_mirror(r))
7738 continue;
7739
7740 nid = memblock_get_region_node(r);
7741
7742 usable_startpfn = memblock_region_memory_base_pfn(r);
7743
7744 if (usable_startpfn < 0x100000) {
7745 mem_below_4gb_not_mirrored = true;
7746 continue;
7747 }
7748
7749 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7750 min(usable_startpfn, zone_movable_pfn[nid]) :
7751 usable_startpfn;
7752 }
7753
7754 if (mem_below_4gb_not_mirrored)
7755 pr_warn("This configuration results in unmirrored kernel memory.\n");
7756
7757 goto out2;
7758 }
7759
7760 /*
7761 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7762 * amount of necessary memory.
7763 */
7764 if (required_kernelcore_percent)
7765 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7766 10000UL;
7767 if (required_movablecore_percent)
7768 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7769 10000UL;
7770
7771 /*
7772 * If movablecore= was specified, calculate what size of
7773 * kernelcore that corresponds so that memory usable for
7774 * any allocation type is evenly spread. If both kernelcore
7775 * and movablecore are specified, then the value of kernelcore
7776 * will be used for required_kernelcore if it's greater than
7777 * what movablecore would have allowed.
7778 */
7779 if (required_movablecore) {
7780 unsigned long corepages;
7781
7782 /*
7783 * Round-up so that ZONE_MOVABLE is at least as large as what
7784 * was requested by the user
7785 */
7786 required_movablecore =
7787 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7788 required_movablecore = min(totalpages, required_movablecore);
7789 corepages = totalpages - required_movablecore;
7790
7791 required_kernelcore = max(required_kernelcore, corepages);
7792 }
7793
7794 /*
7795 * If kernelcore was not specified or kernelcore size is larger
7796 * than totalpages, there is no ZONE_MOVABLE.
7797 */
7798 if (!required_kernelcore || required_kernelcore >= totalpages)
7799 goto out;
7800
7801 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7802 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7803
7804restart:
7805 /* Spread kernelcore memory as evenly as possible throughout nodes */
7806 kernelcore_node = required_kernelcore / usable_nodes;
7807 for_each_node_state(nid, N_MEMORY) {
7808 unsigned long start_pfn, end_pfn;
7809
7810 /*
7811 * Recalculate kernelcore_node if the division per node
7812 * now exceeds what is necessary to satisfy the requested
7813 * amount of memory for the kernel
7814 */
7815 if (required_kernelcore < kernelcore_node)
7816 kernelcore_node = required_kernelcore / usable_nodes;
7817
7818 /*
7819 * As the map is walked, we track how much memory is usable
7820 * by the kernel using kernelcore_remaining. When it is
7821 * 0, the rest of the node is usable by ZONE_MOVABLE
7822 */
7823 kernelcore_remaining = kernelcore_node;
7824
7825 /* Go through each range of PFNs within this node */
7826 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7827 unsigned long size_pages;
7828
7829 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7830 if (start_pfn >= end_pfn)
7831 continue;
7832
7833 /* Account for what is only usable for kernelcore */
7834 if (start_pfn < usable_startpfn) {
7835 unsigned long kernel_pages;
7836 kernel_pages = min(end_pfn, usable_startpfn)
7837 - start_pfn;
7838
7839 kernelcore_remaining -= min(kernel_pages,
7840 kernelcore_remaining);
7841 required_kernelcore -= min(kernel_pages,
7842 required_kernelcore);
7843
7844 /* Continue if range is now fully accounted */
7845 if (end_pfn <= usable_startpfn) {
7846
7847 /*
7848 * Push zone_movable_pfn to the end so
7849 * that if we have to rebalance
7850 * kernelcore across nodes, we will
7851 * not double account here
7852 */
7853 zone_movable_pfn[nid] = end_pfn;
7854 continue;
7855 }
7856 start_pfn = usable_startpfn;
7857 }
7858
7859 /*
7860 * The usable PFN range for ZONE_MOVABLE is from
7861 * start_pfn->end_pfn. Calculate size_pages as the
7862 * number of pages used as kernelcore
7863 */
7864 size_pages = end_pfn - start_pfn;
7865 if (size_pages > kernelcore_remaining)
7866 size_pages = kernelcore_remaining;
7867 zone_movable_pfn[nid] = start_pfn + size_pages;
7868
7869 /*
7870 * Some kernelcore has been met, update counts and
7871 * break if the kernelcore for this node has been
7872 * satisfied
7873 */
7874 required_kernelcore -= min(required_kernelcore,
7875 size_pages);
7876 kernelcore_remaining -= size_pages;
7877 if (!kernelcore_remaining)
7878 break;
7879 }
7880 }
7881
7882 /*
7883 * If there is still required_kernelcore, we do another pass with one
7884 * less node in the count. This will push zone_movable_pfn[nid] further
7885 * along on the nodes that still have memory until kernelcore is
7886 * satisfied
7887 */
7888 usable_nodes--;
7889 if (usable_nodes && required_kernelcore > usable_nodes)
7890 goto restart;
7891
7892out2:
7893 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7894 for (nid = 0; nid < MAX_NUMNODES; nid++)
7895 zone_movable_pfn[nid] =
7896 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7897
7898out:
7899 /* restore the node_state */
7900 node_states[N_MEMORY] = saved_node_state;
7901}
7902
7903/* Any regular or high memory on that node ? */
7904static void check_for_memory(pg_data_t *pgdat, int nid)
7905{
7906 enum zone_type zone_type;
7907
7908 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7909 struct zone *zone = &pgdat->node_zones[zone_type];
7910 if (populated_zone(zone)) {
7911 if (IS_ENABLED(CONFIG_HIGHMEM))
7912 node_set_state(nid, N_HIGH_MEMORY);
7913 if (zone_type <= ZONE_NORMAL)
7914 node_set_state(nid, N_NORMAL_MEMORY);
7915 break;
7916 }
7917 }
7918}
7919
7920/*
7921 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7922 * such cases we allow max_zone_pfn sorted in the descending order
7923 */
7924bool __weak arch_has_descending_max_zone_pfns(void)
7925{
7926 return false;
7927}
7928
7929/**
7930 * free_area_init - Initialise all pg_data_t and zone data
7931 * @max_zone_pfn: an array of max PFNs for each zone
7932 *
7933 * This will call free_area_init_node() for each active node in the system.
7934 * Using the page ranges provided by memblock_set_node(), the size of each
7935 * zone in each node and their holes is calculated. If the maximum PFN
7936 * between two adjacent zones match, it is assumed that the zone is empty.
7937 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7938 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7939 * starts where the previous one ended. For example, ZONE_DMA32 starts
7940 * at arch_max_dma_pfn.
7941 */
7942void __init free_area_init(unsigned long *max_zone_pfn)
7943{
7944 unsigned long start_pfn, end_pfn;
7945 int i, nid, zone;
7946 bool descending;
7947
7948 /* Record where the zone boundaries are */
7949 memset(arch_zone_lowest_possible_pfn, 0,
7950 sizeof(arch_zone_lowest_possible_pfn));
7951 memset(arch_zone_highest_possible_pfn, 0,
7952 sizeof(arch_zone_highest_possible_pfn));
7953
7954 start_pfn = find_min_pfn_with_active_regions();
7955 descending = arch_has_descending_max_zone_pfns();
7956
7957 for (i = 0; i < MAX_NR_ZONES; i++) {
7958 if (descending)
7959 zone = MAX_NR_ZONES - i - 1;
7960 else
7961 zone = i;
7962
7963 if (zone == ZONE_MOVABLE)
7964 continue;
7965
7966 end_pfn = max(max_zone_pfn[zone], start_pfn);
7967 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7968 arch_zone_highest_possible_pfn[zone] = end_pfn;
7969
7970 start_pfn = end_pfn;
7971 }
7972
7973 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7974 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7975 find_zone_movable_pfns_for_nodes();
7976
7977 /* Print out the zone ranges */
7978 pr_info("Zone ranges:\n");
7979 for (i = 0; i < MAX_NR_ZONES; i++) {
7980 if (i == ZONE_MOVABLE)
7981 continue;
7982 pr_info(" %-8s ", zone_names[i]);
7983 if (arch_zone_lowest_possible_pfn[i] ==
7984 arch_zone_highest_possible_pfn[i])
7985 pr_cont("empty\n");
7986 else
7987 pr_cont("[mem %#018Lx-%#018Lx]\n",
7988 (u64)arch_zone_lowest_possible_pfn[i]
7989 << PAGE_SHIFT,
7990 ((u64)arch_zone_highest_possible_pfn[i]
7991 << PAGE_SHIFT) - 1);
7992 }
7993
7994 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7995 pr_info("Movable zone start for each node\n");
7996 for (i = 0; i < MAX_NUMNODES; i++) {
7997 if (zone_movable_pfn[i])
7998 pr_info(" Node %d: %#018Lx\n", i,
7999 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8000 }
8001
8002 /*
8003 * Print out the early node map, and initialize the
8004 * subsection-map relative to active online memory ranges to
8005 * enable future "sub-section" extensions of the memory map.
8006 */
8007 pr_info("Early memory node ranges\n");
8008 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8009 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8010 (u64)start_pfn << PAGE_SHIFT,
8011 ((u64)end_pfn << PAGE_SHIFT) - 1);
8012 subsection_map_init(start_pfn, end_pfn - start_pfn);
8013 }
8014
8015 /* Initialise every node */
8016 mminit_verify_pageflags_layout();
8017 setup_nr_node_ids();
8018 for_each_online_node(nid) {
8019 pg_data_t *pgdat = NODE_DATA(nid);
8020 free_area_init_node(nid);
8021
8022 /* Any memory on that node */
8023 if (pgdat->node_present_pages)
8024 node_set_state(nid, N_MEMORY);
8025 check_for_memory(pgdat, nid);
8026 }
8027
8028 memmap_init();
8029}
8030
8031static int __init cmdline_parse_core(char *p, unsigned long *core,
8032 unsigned long *percent)
8033{
8034 unsigned long long coremem;
8035 char *endptr;
8036
8037 if (!p)
8038 return -EINVAL;
8039
8040 /* Value may be a percentage of total memory, otherwise bytes */
8041 coremem = simple_strtoull(p, &endptr, 0);
8042 if (*endptr == '%') {
8043 /* Paranoid check for percent values greater than 100 */
8044 WARN_ON(coremem > 100);
8045
8046 *percent = coremem;
8047 } else {
8048 coremem = memparse(p, &p);
8049 /* Paranoid check that UL is enough for the coremem value */
8050 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8051
8052 *core = coremem >> PAGE_SHIFT;
8053 *percent = 0UL;
8054 }
8055 return 0;
8056}
8057
8058/*
8059 * kernelcore=size sets the amount of memory for use for allocations that
8060 * cannot be reclaimed or migrated.
8061 */
8062static int __init cmdline_parse_kernelcore(char *p)
8063{
8064 /* parse kernelcore=mirror */
8065 if (parse_option_str(p, "mirror")) {
8066 mirrored_kernelcore = true;
8067 return 0;
8068 }
8069
8070 return cmdline_parse_core(p, &required_kernelcore,
8071 &required_kernelcore_percent);
8072}
8073
8074/*
8075 * movablecore=size sets the amount of memory for use for allocations that
8076 * can be reclaimed or migrated.
8077 */
8078static int __init cmdline_parse_movablecore(char *p)
8079{
8080 return cmdline_parse_core(p, &required_movablecore,
8081 &required_movablecore_percent);
8082}
8083
8084early_param("kernelcore", cmdline_parse_kernelcore);
8085early_param("movablecore", cmdline_parse_movablecore);
8086
8087void adjust_managed_page_count(struct page *page, long count)
8088{
8089 atomic_long_add(count, &page_zone(page)->managed_pages);
8090 totalram_pages_add(count);
8091#ifdef CONFIG_HIGHMEM
8092 if (PageHighMem(page))
8093 totalhigh_pages_add(count);
8094#endif
8095}
8096EXPORT_SYMBOL(adjust_managed_page_count);
8097
8098unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8099{
8100 void *pos;
8101 unsigned long pages = 0;
8102
8103 start = (void *)PAGE_ALIGN((unsigned long)start);
8104 end = (void *)((unsigned long)end & PAGE_MASK);
8105 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8106 struct page *page = virt_to_page(pos);
8107 void *direct_map_addr;
8108
8109 /*
8110 * 'direct_map_addr' might be different from 'pos'
8111 * because some architectures' virt_to_page()
8112 * work with aliases. Getting the direct map
8113 * address ensures that we get a _writeable_
8114 * alias for the memset().
8115 */
8116 direct_map_addr = page_address(page);
8117 /*
8118 * Perform a kasan-unchecked memset() since this memory
8119 * has not been initialized.
8120 */
8121 direct_map_addr = kasan_reset_tag(direct_map_addr);
8122 if ((unsigned int)poison <= 0xFF)
8123 memset(direct_map_addr, poison, PAGE_SIZE);
8124
8125 free_reserved_page(page);
8126 }
8127
8128 if (pages && s)
8129 pr_info("Freeing %s memory: %ldK\n",
8130 s, pages << (PAGE_SHIFT - 10));
8131
8132 return pages;
8133}
8134
8135void __init mem_init_print_info(void)
8136{
8137 unsigned long physpages, codesize, datasize, rosize, bss_size;
8138 unsigned long init_code_size, init_data_size;
8139
8140 physpages = get_num_physpages();
8141 codesize = _etext - _stext;
8142 datasize = _edata - _sdata;
8143 rosize = __end_rodata - __start_rodata;
8144 bss_size = __bss_stop - __bss_start;
8145 init_data_size = __init_end - __init_begin;
8146 init_code_size = _einittext - _sinittext;
8147
8148 /*
8149 * Detect special cases and adjust section sizes accordingly:
8150 * 1) .init.* may be embedded into .data sections
8151 * 2) .init.text.* may be out of [__init_begin, __init_end],
8152 * please refer to arch/tile/kernel/vmlinux.lds.S.
8153 * 3) .rodata.* may be embedded into .text or .data sections.
8154 */
8155#define adj_init_size(start, end, size, pos, adj) \
8156 do { \
8157 if (start <= pos && pos < end && size > adj) \
8158 size -= adj; \
8159 } while (0)
8160
8161 adj_init_size(__init_begin, __init_end, init_data_size,
8162 _sinittext, init_code_size);
8163 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8164 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8165 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8166 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8167
8168#undef adj_init_size
8169
8170 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8171#ifdef CONFIG_HIGHMEM
8172 ", %luK highmem"
8173#endif
8174 ")\n",
8175 nr_free_pages() << (PAGE_SHIFT - 10),
8176 physpages << (PAGE_SHIFT - 10),
8177 codesize >> 10, datasize >> 10, rosize >> 10,
8178 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8179 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
8180 totalcma_pages << (PAGE_SHIFT - 10)
8181#ifdef CONFIG_HIGHMEM
8182 , totalhigh_pages() << (PAGE_SHIFT - 10)
8183#endif
8184 );
8185}
8186
8187/**
8188 * set_dma_reserve - set the specified number of pages reserved in the first zone
8189 * @new_dma_reserve: The number of pages to mark reserved
8190 *
8191 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8192 * In the DMA zone, a significant percentage may be consumed by kernel image
8193 * and other unfreeable allocations which can skew the watermarks badly. This
8194 * function may optionally be used to account for unfreeable pages in the
8195 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8196 * smaller per-cpu batchsize.
8197 */
8198void __init set_dma_reserve(unsigned long new_dma_reserve)
8199{
8200 dma_reserve = new_dma_reserve;
8201}
8202
8203static int page_alloc_cpu_dead(unsigned int cpu)
8204{
8205 struct zone *zone;
8206
8207 lru_add_drain_cpu(cpu);
8208 drain_pages(cpu);
8209
8210 /*
8211 * Spill the event counters of the dead processor
8212 * into the current processors event counters.
8213 * This artificially elevates the count of the current
8214 * processor.
8215 */
8216 vm_events_fold_cpu(cpu);
8217
8218 /*
8219 * Zero the differential counters of the dead processor
8220 * so that the vm statistics are consistent.
8221 *
8222 * This is only okay since the processor is dead and cannot
8223 * race with what we are doing.
8224 */
8225 cpu_vm_stats_fold(cpu);
8226
8227 for_each_populated_zone(zone)
8228 zone_pcp_update(zone, 0);
8229
8230 return 0;
8231}
8232
8233static int page_alloc_cpu_online(unsigned int cpu)
8234{
8235 struct zone *zone;
8236
8237 for_each_populated_zone(zone)
8238 zone_pcp_update(zone, 1);
8239 return 0;
8240}
8241
8242#ifdef CONFIG_NUMA
8243int hashdist = HASHDIST_DEFAULT;
8244
8245static int __init set_hashdist(char *str)
8246{
8247 if (!str)
8248 return 0;
8249 hashdist = simple_strtoul(str, &str, 0);
8250 return 1;
8251}
8252__setup("hashdist=", set_hashdist);
8253#endif
8254
8255void __init page_alloc_init(void)
8256{
8257 int ret;
8258
8259#ifdef CONFIG_NUMA
8260 if (num_node_state(N_MEMORY) == 1)
8261 hashdist = 0;
8262#endif
8263
8264 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8265 "mm/page_alloc:pcp",
8266 page_alloc_cpu_online,
8267 page_alloc_cpu_dead);
8268 WARN_ON(ret < 0);
8269}
8270
8271/*
8272 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8273 * or min_free_kbytes changes.
8274 */
8275static void calculate_totalreserve_pages(void)
8276{
8277 struct pglist_data *pgdat;
8278 unsigned long reserve_pages = 0;
8279 enum zone_type i, j;
8280
8281 for_each_online_pgdat(pgdat) {
8282
8283 pgdat->totalreserve_pages = 0;
8284
8285 for (i = 0; i < MAX_NR_ZONES; i++) {
8286 struct zone *zone = pgdat->node_zones + i;
8287 long max = 0;
8288 unsigned long managed_pages = zone_managed_pages(zone);
8289
8290 /* Find valid and maximum lowmem_reserve in the zone */
8291 for (j = i; j < MAX_NR_ZONES; j++) {
8292 if (zone->lowmem_reserve[j] > max)
8293 max = zone->lowmem_reserve[j];
8294 }
8295
8296 /* we treat the high watermark as reserved pages. */
8297 max += high_wmark_pages(zone);
8298
8299 if (max > managed_pages)
8300 max = managed_pages;
8301
8302 pgdat->totalreserve_pages += max;
8303
8304 reserve_pages += max;
8305 }
8306 }
8307 totalreserve_pages = reserve_pages;
8308}
8309
8310/*
8311 * setup_per_zone_lowmem_reserve - called whenever
8312 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8313 * has a correct pages reserved value, so an adequate number of
8314 * pages are left in the zone after a successful __alloc_pages().
8315 */
8316static void setup_per_zone_lowmem_reserve(void)
8317{
8318 struct pglist_data *pgdat;
8319 enum zone_type i, j;
8320
8321 for_each_online_pgdat(pgdat) {
8322 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8323 struct zone *zone = &pgdat->node_zones[i];
8324 int ratio = sysctl_lowmem_reserve_ratio[i];
8325 bool clear = !ratio || !zone_managed_pages(zone);
8326 unsigned long managed_pages = 0;
8327
8328 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8329 struct zone *upper_zone = &pgdat->node_zones[j];
8330
8331 managed_pages += zone_managed_pages(upper_zone);
8332
8333 if (clear)
8334 zone->lowmem_reserve[j] = 0;
8335 else
8336 zone->lowmem_reserve[j] = managed_pages / ratio;
8337 }
8338 }
8339 }
8340
8341 /* update totalreserve_pages */
8342 calculate_totalreserve_pages();
8343}
8344
8345static void __setup_per_zone_wmarks(void)
8346{
8347 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8348 unsigned long lowmem_pages = 0;
8349 struct zone *zone;
8350 unsigned long flags;
8351
8352 /* Calculate total number of !ZONE_HIGHMEM pages */
8353 for_each_zone(zone) {
8354 if (!is_highmem(zone))
8355 lowmem_pages += zone_managed_pages(zone);
8356 }
8357
8358 for_each_zone(zone) {
8359 u64 tmp;
8360
8361 spin_lock_irqsave(&zone->lock, flags);
8362 tmp = (u64)pages_min * zone_managed_pages(zone);
8363 do_div(tmp, lowmem_pages);
8364 if (is_highmem(zone)) {
8365 /*
8366 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8367 * need highmem pages, so cap pages_min to a small
8368 * value here.
8369 *
8370 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8371 * deltas control async page reclaim, and so should
8372 * not be capped for highmem.
8373 */
8374 unsigned long min_pages;
8375
8376 min_pages = zone_managed_pages(zone) / 1024;
8377 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8378 zone->_watermark[WMARK_MIN] = min_pages;
8379 } else {
8380 /*
8381 * If it's a lowmem zone, reserve a number of pages
8382 * proportionate to the zone's size.
8383 */
8384 zone->_watermark[WMARK_MIN] = tmp;
8385 }
8386
8387 /*
8388 * Set the kswapd watermarks distance according to the
8389 * scale factor in proportion to available memory, but
8390 * ensure a minimum size on small systems.
8391 */
8392 tmp = max_t(u64, tmp >> 2,
8393 mult_frac(zone_managed_pages(zone),
8394 watermark_scale_factor, 10000));
8395
8396 zone->watermark_boost = 0;
8397 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8398 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8399
8400 spin_unlock_irqrestore(&zone->lock, flags);
8401 }
8402
8403 /* update totalreserve_pages */
8404 calculate_totalreserve_pages();
8405}
8406
8407/**
8408 * setup_per_zone_wmarks - called when min_free_kbytes changes
8409 * or when memory is hot-{added|removed}
8410 *
8411 * Ensures that the watermark[min,low,high] values for each zone are set
8412 * correctly with respect to min_free_kbytes.
8413 */
8414void setup_per_zone_wmarks(void)
8415{
8416 struct zone *zone;
8417 static DEFINE_SPINLOCK(lock);
8418
8419 spin_lock(&lock);
8420 __setup_per_zone_wmarks();
8421 spin_unlock(&lock);
8422
8423 /*
8424 * The watermark size have changed so update the pcpu batch
8425 * and high limits or the limits may be inappropriate.
8426 */
8427 for_each_zone(zone)
8428 zone_pcp_update(zone, 0);
8429}
8430
8431/*
8432 * Initialise min_free_kbytes.
8433 *
8434 * For small machines we want it small (128k min). For large machines
8435 * we want it large (256MB max). But it is not linear, because network
8436 * bandwidth does not increase linearly with machine size. We use
8437 *
8438 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8439 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8440 *
8441 * which yields
8442 *
8443 * 16MB: 512k
8444 * 32MB: 724k
8445 * 64MB: 1024k
8446 * 128MB: 1448k
8447 * 256MB: 2048k
8448 * 512MB: 2896k
8449 * 1024MB: 4096k
8450 * 2048MB: 5792k
8451 * 4096MB: 8192k
8452 * 8192MB: 11584k
8453 * 16384MB: 16384k
8454 */
8455int __meminit init_per_zone_wmark_min(void)
8456{
8457 unsigned long lowmem_kbytes;
8458 int new_min_free_kbytes;
8459
8460 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8461 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8462
8463 if (new_min_free_kbytes > user_min_free_kbytes) {
8464 min_free_kbytes = new_min_free_kbytes;
8465 if (min_free_kbytes < 128)
8466 min_free_kbytes = 128;
8467 if (min_free_kbytes > 262144)
8468 min_free_kbytes = 262144;
8469 } else {
8470 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8471 new_min_free_kbytes, user_min_free_kbytes);
8472 }
8473 setup_per_zone_wmarks();
8474 refresh_zone_stat_thresholds();
8475 setup_per_zone_lowmem_reserve();
8476
8477#ifdef CONFIG_NUMA
8478 setup_min_unmapped_ratio();
8479 setup_min_slab_ratio();
8480#endif
8481
8482 khugepaged_min_free_kbytes_update();
8483
8484 return 0;
8485}
8486postcore_initcall(init_per_zone_wmark_min)
8487
8488/*
8489 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8490 * that we can call two helper functions whenever min_free_kbytes
8491 * changes.
8492 */
8493int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8494 void *buffer, size_t *length, loff_t *ppos)
8495{
8496 int rc;
8497
8498 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8499 if (rc)
8500 return rc;
8501
8502 if (write) {
8503 user_min_free_kbytes = min_free_kbytes;
8504 setup_per_zone_wmarks();
8505 }
8506 return 0;
8507}
8508
8509int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8510 void *buffer, size_t *length, loff_t *ppos)
8511{
8512 int rc;
8513
8514 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8515 if (rc)
8516 return rc;
8517
8518 if (write)
8519 setup_per_zone_wmarks();
8520
8521 return 0;
8522}
8523
8524#ifdef CONFIG_NUMA
8525static void setup_min_unmapped_ratio(void)
8526{
8527 pg_data_t *pgdat;
8528 struct zone *zone;
8529
8530 for_each_online_pgdat(pgdat)
8531 pgdat->min_unmapped_pages = 0;
8532
8533 for_each_zone(zone)
8534 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8535 sysctl_min_unmapped_ratio) / 100;
8536}
8537
8538
8539int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8540 void *buffer, size_t *length, loff_t *ppos)
8541{
8542 int rc;
8543
8544 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8545 if (rc)
8546 return rc;
8547
8548 setup_min_unmapped_ratio();
8549
8550 return 0;
8551}
8552
8553static void setup_min_slab_ratio(void)
8554{
8555 pg_data_t *pgdat;
8556 struct zone *zone;
8557
8558 for_each_online_pgdat(pgdat)
8559 pgdat->min_slab_pages = 0;
8560
8561 for_each_zone(zone)
8562 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8563 sysctl_min_slab_ratio) / 100;
8564}
8565
8566int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8567 void *buffer, size_t *length, loff_t *ppos)
8568{
8569 int rc;
8570
8571 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8572 if (rc)
8573 return rc;
8574
8575 setup_min_slab_ratio();
8576
8577 return 0;
8578}
8579#endif
8580
8581/*
8582 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8583 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8584 * whenever sysctl_lowmem_reserve_ratio changes.
8585 *
8586 * The reserve ratio obviously has absolutely no relation with the
8587 * minimum watermarks. The lowmem reserve ratio can only make sense
8588 * if in function of the boot time zone sizes.
8589 */
8590int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8591 void *buffer, size_t *length, loff_t *ppos)
8592{
8593 int i;
8594
8595 proc_dointvec_minmax(table, write, buffer, length, ppos);
8596
8597 for (i = 0; i < MAX_NR_ZONES; i++) {
8598 if (sysctl_lowmem_reserve_ratio[i] < 1)
8599 sysctl_lowmem_reserve_ratio[i] = 0;
8600 }
8601
8602 setup_per_zone_lowmem_reserve();
8603 return 0;
8604}
8605
8606/*
8607 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8608 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8609 * pagelist can have before it gets flushed back to buddy allocator.
8610 */
8611int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8612 int write, void *buffer, size_t *length, loff_t *ppos)
8613{
8614 struct zone *zone;
8615 int old_percpu_pagelist_high_fraction;
8616 int ret;
8617
8618 mutex_lock(&pcp_batch_high_lock);
8619 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8620
8621 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8622 if (!write || ret < 0)
8623 goto out;
8624
8625 /* Sanity checking to avoid pcp imbalance */
8626 if (percpu_pagelist_high_fraction &&
8627 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8628 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8629 ret = -EINVAL;
8630 goto out;
8631 }
8632
8633 /* No change? */
8634 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8635 goto out;
8636
8637 for_each_populated_zone(zone)
8638 zone_set_pageset_high_and_batch(zone, 0);
8639out:
8640 mutex_unlock(&pcp_batch_high_lock);
8641 return ret;
8642}
8643
8644#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8645/*
8646 * Returns the number of pages that arch has reserved but
8647 * is not known to alloc_large_system_hash().
8648 */
8649static unsigned long __init arch_reserved_kernel_pages(void)
8650{
8651 return 0;
8652}
8653#endif
8654
8655/*
8656 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8657 * machines. As memory size is increased the scale is also increased but at
8658 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8659 * quadruples the scale is increased by one, which means the size of hash table
8660 * only doubles, instead of quadrupling as well.
8661 * Because 32-bit systems cannot have large physical memory, where this scaling
8662 * makes sense, it is disabled on such platforms.
8663 */
8664#if __BITS_PER_LONG > 32
8665#define ADAPT_SCALE_BASE (64ul << 30)
8666#define ADAPT_SCALE_SHIFT 2
8667#define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8668#endif
8669
8670/*
8671 * allocate a large system hash table from bootmem
8672 * - it is assumed that the hash table must contain an exact power-of-2
8673 * quantity of entries
8674 * - limit is the number of hash buckets, not the total allocation size
8675 */
8676void *__init alloc_large_system_hash(const char *tablename,
8677 unsigned long bucketsize,
8678 unsigned long numentries,
8679 int scale,
8680 int flags,
8681 unsigned int *_hash_shift,
8682 unsigned int *_hash_mask,
8683 unsigned long low_limit,
8684 unsigned long high_limit)
8685{
8686 unsigned long long max = high_limit;
8687 unsigned long log2qty, size;
8688 void *table = NULL;
8689 gfp_t gfp_flags;
8690 bool virt;
8691 bool huge;
8692
8693 /* allow the kernel cmdline to have a say */
8694 if (!numentries) {
8695 /* round applicable memory size up to nearest megabyte */
8696 numentries = nr_kernel_pages;
8697 numentries -= arch_reserved_kernel_pages();
8698
8699 /* It isn't necessary when PAGE_SIZE >= 1MB */
8700 if (PAGE_SHIFT < 20)
8701 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8702
8703#if __BITS_PER_LONG > 32
8704 if (!high_limit) {
8705 unsigned long adapt;
8706
8707 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8708 adapt <<= ADAPT_SCALE_SHIFT)
8709 scale++;
8710 }
8711#endif
8712
8713 /* limit to 1 bucket per 2^scale bytes of low memory */
8714 if (scale > PAGE_SHIFT)
8715 numentries >>= (scale - PAGE_SHIFT);
8716 else
8717 numentries <<= (PAGE_SHIFT - scale);
8718
8719 /* Make sure we've got at least a 0-order allocation.. */
8720 if (unlikely(flags & HASH_SMALL)) {
8721 /* Makes no sense without HASH_EARLY */
8722 WARN_ON(!(flags & HASH_EARLY));
8723 if (!(numentries >> *_hash_shift)) {
8724 numentries = 1UL << *_hash_shift;
8725 BUG_ON(!numentries);
8726 }
8727 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8728 numentries = PAGE_SIZE / bucketsize;
8729 }
8730 numentries = roundup_pow_of_two(numentries);
8731
8732 /* limit allocation size to 1/16 total memory by default */
8733 if (max == 0) {
8734 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8735 do_div(max, bucketsize);
8736 }
8737 max = min(max, 0x80000000ULL);
8738
8739 if (numentries < low_limit)
8740 numentries = low_limit;
8741 if (numentries > max)
8742 numentries = max;
8743
8744 log2qty = ilog2(numentries);
8745
8746 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8747 do {
8748 virt = false;
8749 size = bucketsize << log2qty;
8750 if (flags & HASH_EARLY) {
8751 if (flags & HASH_ZERO)
8752 table = memblock_alloc(size, SMP_CACHE_BYTES);
8753 else
8754 table = memblock_alloc_raw(size,
8755 SMP_CACHE_BYTES);
8756 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8757 table = __vmalloc(size, gfp_flags);
8758 virt = true;
8759 huge = is_vm_area_hugepages(table);
8760 } else {
8761 /*
8762 * If bucketsize is not a power-of-two, we may free
8763 * some pages at the end of hash table which
8764 * alloc_pages_exact() automatically does
8765 */
8766 table = alloc_pages_exact(size, gfp_flags);
8767 kmemleak_alloc(table, size, 1, gfp_flags);
8768 }
8769 } while (!table && size > PAGE_SIZE && --log2qty);
8770
8771 if (!table)
8772 panic("Failed to allocate %s hash table\n", tablename);
8773
8774 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8775 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8776 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8777
8778 if (_hash_shift)
8779 *_hash_shift = log2qty;
8780 if (_hash_mask)
8781 *_hash_mask = (1 << log2qty) - 1;
8782
8783 return table;
8784}
8785
8786/*
8787 * This function checks whether pageblock includes unmovable pages or not.
8788 *
8789 * PageLRU check without isolation or lru_lock could race so that
8790 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8791 * check without lock_page also may miss some movable non-lru pages at
8792 * race condition. So you can't expect this function should be exact.
8793 *
8794 * Returns a page without holding a reference. If the caller wants to
8795 * dereference that page (e.g., dumping), it has to make sure that it
8796 * cannot get removed (e.g., via memory unplug) concurrently.
8797 *
8798 */
8799struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8800 int migratetype, int flags)
8801{
8802 unsigned long iter = 0;
8803 unsigned long pfn = page_to_pfn(page);
8804 unsigned long offset = pfn % pageblock_nr_pages;
8805
8806 if (is_migrate_cma_page(page)) {
8807 /*
8808 * CMA allocations (alloc_contig_range) really need to mark
8809 * isolate CMA pageblocks even when they are not movable in fact
8810 * so consider them movable here.
8811 */
8812 if (is_migrate_cma(migratetype))
8813 return NULL;
8814
8815 return page;
8816 }
8817
8818 for (; iter < pageblock_nr_pages - offset; iter++) {
8819 if (!pfn_valid_within(pfn + iter))
8820 continue;
8821
8822 page = pfn_to_page(pfn + iter);
8823
8824 /*
8825 * Both, bootmem allocations and memory holes are marked
8826 * PG_reserved and are unmovable. We can even have unmovable
8827 * allocations inside ZONE_MOVABLE, for example when
8828 * specifying "movablecore".
8829 */
8830 if (PageReserved(page))
8831 return page;
8832
8833 /*
8834 * If the zone is movable and we have ruled out all reserved
8835 * pages then it should be reasonably safe to assume the rest
8836 * is movable.
8837 */
8838 if (zone_idx(zone) == ZONE_MOVABLE)
8839 continue;
8840
8841 /*
8842 * Hugepages are not in LRU lists, but they're movable.
8843 * THPs are on the LRU, but need to be counted as #small pages.
8844 * We need not scan over tail pages because we don't
8845 * handle each tail page individually in migration.
8846 */
8847 if (PageHuge(page) || PageTransCompound(page)) {
8848 struct page *head = compound_head(page);
8849 unsigned int skip_pages;
8850
8851 if (PageHuge(page)) {
8852 if (!hugepage_migration_supported(page_hstate(head)))
8853 return page;
8854 } else if (!PageLRU(head) && !__PageMovable(head)) {
8855 return page;
8856 }
8857
8858 skip_pages = compound_nr(head) - (page - head);
8859 iter += skip_pages - 1;
8860 continue;
8861 }
8862
8863 /*
8864 * We can't use page_count without pin a page
8865 * because another CPU can free compound page.
8866 * This check already skips compound tails of THP
8867 * because their page->_refcount is zero at all time.
8868 */
8869 if (!page_ref_count(page)) {
8870 if (PageBuddy(page))
8871 iter += (1 << buddy_order(page)) - 1;
8872 continue;
8873 }
8874
8875 /*
8876 * The HWPoisoned page may be not in buddy system, and
8877 * page_count() is not 0.
8878 */
8879 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8880 continue;
8881
8882 /*
8883 * We treat all PageOffline() pages as movable when offlining
8884 * to give drivers a chance to decrement their reference count
8885 * in MEM_GOING_OFFLINE in order to indicate that these pages
8886 * can be offlined as there are no direct references anymore.
8887 * For actually unmovable PageOffline() where the driver does
8888 * not support this, we will fail later when trying to actually
8889 * move these pages that still have a reference count > 0.
8890 * (false negatives in this function only)
8891 */
8892 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8893 continue;
8894
8895 if (__PageMovable(page) || PageLRU(page))
8896 continue;
8897
8898 /*
8899 * If there are RECLAIMABLE pages, we need to check
8900 * it. But now, memory offline itself doesn't call
8901 * shrink_node_slabs() and it still to be fixed.
8902 */
8903 return page;
8904 }
8905 return NULL;
8906}
8907
8908#ifdef CONFIG_CONTIG_ALLOC
8909static unsigned long pfn_max_align_down(unsigned long pfn)
8910{
8911 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8912 pageblock_nr_pages) - 1);
8913}
8914
8915static unsigned long pfn_max_align_up(unsigned long pfn)
8916{
8917 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8918 pageblock_nr_pages));
8919}
8920
8921#if defined(CONFIG_DYNAMIC_DEBUG) || \
8922 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8923/* Usage: See admin-guide/dynamic-debug-howto.rst */
8924static void alloc_contig_dump_pages(struct list_head *page_list)
8925{
8926 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8927
8928 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8929 struct page *page;
8930
8931 dump_stack();
8932 list_for_each_entry(page, page_list, lru)
8933 dump_page(page, "migration failure");
8934 }
8935}
8936#else
8937static inline void alloc_contig_dump_pages(struct list_head *page_list)
8938{
8939}
8940#endif
8941
8942/* [start, end) must belong to a single zone. */
8943static int __alloc_contig_migrate_range(struct compact_control *cc,
8944 unsigned long start, unsigned long end)
8945{
8946 /* This function is based on compact_zone() from compaction.c. */
8947 unsigned int nr_reclaimed;
8948 unsigned long pfn = start;
8949 unsigned int tries = 0;
8950 int ret = 0;
8951 struct migration_target_control mtc = {
8952 .nid = zone_to_nid(cc->zone),
8953 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8954 };
8955
8956 lru_cache_disable();
8957
8958 while (pfn < end || !list_empty(&cc->migratepages)) {
8959 if (fatal_signal_pending(current)) {
8960 ret = -EINTR;
8961 break;
8962 }
8963
8964 if (list_empty(&cc->migratepages)) {
8965 cc->nr_migratepages = 0;
8966 ret = isolate_migratepages_range(cc, pfn, end);
8967 if (ret && ret != -EAGAIN)
8968 break;
8969 pfn = cc->migrate_pfn;
8970 tries = 0;
8971 } else if (++tries == 5) {
8972 ret = -EBUSY;
8973 break;
8974 }
8975
8976 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8977 &cc->migratepages);
8978 cc->nr_migratepages -= nr_reclaimed;
8979
8980 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8981 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8982
8983 /*
8984 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
8985 * to retry again over this error, so do the same here.
8986 */
8987 if (ret == -ENOMEM)
8988 break;
8989 }
8990
8991 lru_cache_enable();
8992 if (ret < 0) {
8993 if (ret == -EBUSY)
8994 alloc_contig_dump_pages(&cc->migratepages);
8995 putback_movable_pages(&cc->migratepages);
8996 return ret;
8997 }
8998 return 0;
8999}
9000
9001/**
9002 * alloc_contig_range() -- tries to allocate given range of pages
9003 * @start: start PFN to allocate
9004 * @end: one-past-the-last PFN to allocate
9005 * @migratetype: migratetype of the underlying pageblocks (either
9006 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9007 * in range must have the same migratetype and it must
9008 * be either of the two.
9009 * @gfp_mask: GFP mask to use during compaction
9010 *
9011 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9012 * aligned. The PFN range must belong to a single zone.
9013 *
9014 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9015 * pageblocks in the range. Once isolated, the pageblocks should not
9016 * be modified by others.
9017 *
9018 * Return: zero on success or negative error code. On success all
9019 * pages which PFN is in [start, end) are allocated for the caller and
9020 * need to be freed with free_contig_range().
9021 */
9022int alloc_contig_range(unsigned long start, unsigned long end,
9023 unsigned migratetype, gfp_t gfp_mask)
9024{
9025 unsigned long outer_start, outer_end;
9026 unsigned int order;
9027 int ret = 0;
9028
9029 struct compact_control cc = {
9030 .nr_migratepages = 0,
9031 .order = -1,
9032 .zone = page_zone(pfn_to_page(start)),
9033 .mode = MIGRATE_SYNC,
9034 .ignore_skip_hint = true,
9035 .no_set_skip_hint = true,
9036 .gfp_mask = current_gfp_context(gfp_mask),
9037 .alloc_contig = true,
9038 };
9039 INIT_LIST_HEAD(&cc.migratepages);
9040
9041 /*
9042 * What we do here is we mark all pageblocks in range as
9043 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9044 * have different sizes, and due to the way page allocator
9045 * work, we align the range to biggest of the two pages so
9046 * that page allocator won't try to merge buddies from
9047 * different pageblocks and change MIGRATE_ISOLATE to some
9048 * other migration type.
9049 *
9050 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9051 * migrate the pages from an unaligned range (ie. pages that
9052 * we are interested in). This will put all the pages in
9053 * range back to page allocator as MIGRATE_ISOLATE.
9054 *
9055 * When this is done, we take the pages in range from page
9056 * allocator removing them from the buddy system. This way
9057 * page allocator will never consider using them.
9058 *
9059 * This lets us mark the pageblocks back as
9060 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9061 * aligned range but not in the unaligned, original range are
9062 * put back to page allocator so that buddy can use them.
9063 */
9064
9065 ret = start_isolate_page_range(pfn_max_align_down(start),
9066 pfn_max_align_up(end), migratetype, 0);
9067 if (ret)
9068 return ret;
9069
9070 drain_all_pages(cc.zone);
9071
9072 /*
9073 * In case of -EBUSY, we'd like to know which page causes problem.
9074 * So, just fall through. test_pages_isolated() has a tracepoint
9075 * which will report the busy page.
9076 *
9077 * It is possible that busy pages could become available before
9078 * the call to test_pages_isolated, and the range will actually be
9079 * allocated. So, if we fall through be sure to clear ret so that
9080 * -EBUSY is not accidentally used or returned to caller.
9081 */
9082 ret = __alloc_contig_migrate_range(&cc, start, end);
9083 if (ret && ret != -EBUSY)
9084 goto done;
9085 ret = 0;
9086
9087 /*
9088 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9089 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9090 * more, all pages in [start, end) are free in page allocator.
9091 * What we are going to do is to allocate all pages from
9092 * [start, end) (that is remove them from page allocator).
9093 *
9094 * The only problem is that pages at the beginning and at the
9095 * end of interesting range may be not aligned with pages that
9096 * page allocator holds, ie. they can be part of higher order
9097 * pages. Because of this, we reserve the bigger range and
9098 * once this is done free the pages we are not interested in.
9099 *
9100 * We don't have to hold zone->lock here because the pages are
9101 * isolated thus they won't get removed from buddy.
9102 */
9103
9104 order = 0;
9105 outer_start = start;
9106 while (!PageBuddy(pfn_to_page(outer_start))) {
9107 if (++order >= MAX_ORDER) {
9108 outer_start = start;
9109 break;
9110 }
9111 outer_start &= ~0UL << order;
9112 }
9113
9114 if (outer_start != start) {
9115 order = buddy_order(pfn_to_page(outer_start));
9116
9117 /*
9118 * outer_start page could be small order buddy page and
9119 * it doesn't include start page. Adjust outer_start
9120 * in this case to report failed page properly
9121 * on tracepoint in test_pages_isolated()
9122 */
9123 if (outer_start + (1UL << order) <= start)
9124 outer_start = start;
9125 }
9126
9127 /* Make sure the range is really isolated. */
9128 if (test_pages_isolated(outer_start, end, 0)) {
9129 ret = -EBUSY;
9130 goto done;
9131 }
9132
9133 /* Grab isolated pages from freelists. */
9134 outer_end = isolate_freepages_range(&cc, outer_start, end);
9135 if (!outer_end) {
9136 ret = -EBUSY;
9137 goto done;
9138 }
9139
9140 /* Free head and tail (if any) */
9141 if (start != outer_start)
9142 free_contig_range(outer_start, start - outer_start);
9143 if (end != outer_end)
9144 free_contig_range(end, outer_end - end);
9145
9146done:
9147 undo_isolate_page_range(pfn_max_align_down(start),
9148 pfn_max_align_up(end), migratetype);
9149 return ret;
9150}
9151EXPORT_SYMBOL(alloc_contig_range);
9152
9153static int __alloc_contig_pages(unsigned long start_pfn,
9154 unsigned long nr_pages, gfp_t gfp_mask)
9155{
9156 unsigned long end_pfn = start_pfn + nr_pages;
9157
9158 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9159 gfp_mask);
9160}
9161
9162static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9163 unsigned long nr_pages)
9164{
9165 unsigned long i, end_pfn = start_pfn + nr_pages;
9166 struct page *page;
9167
9168 for (i = start_pfn; i < end_pfn; i++) {
9169 page = pfn_to_online_page(i);
9170 if (!page)
9171 return false;
9172
9173 if (page_zone(page) != z)
9174 return false;
9175
9176 if (PageReserved(page))
9177 return false;
9178 }
9179 return true;
9180}
9181
9182static bool zone_spans_last_pfn(const struct zone *zone,
9183 unsigned long start_pfn, unsigned long nr_pages)
9184{
9185 unsigned long last_pfn = start_pfn + nr_pages - 1;
9186
9187 return zone_spans_pfn(zone, last_pfn);
9188}
9189
9190/**
9191 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9192 * @nr_pages: Number of contiguous pages to allocate
9193 * @gfp_mask: GFP mask to limit search and used during compaction
9194 * @nid: Target node
9195 * @nodemask: Mask for other possible nodes
9196 *
9197 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9198 * on an applicable zonelist to find a contiguous pfn range which can then be
9199 * tried for allocation with alloc_contig_range(). This routine is intended
9200 * for allocation requests which can not be fulfilled with the buddy allocator.
9201 *
9202 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9203 * power of two then the alignment is guaranteed to be to the given nr_pages
9204 * (e.g. 1GB request would be aligned to 1GB).
9205 *
9206 * Allocated pages can be freed with free_contig_range() or by manually calling
9207 * __free_page() on each allocated page.
9208 *
9209 * Return: pointer to contiguous pages on success, or NULL if not successful.
9210 */
9211struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9212 int nid, nodemask_t *nodemask)
9213{
9214 unsigned long ret, pfn, flags;
9215 struct zonelist *zonelist;
9216 struct zone *zone;
9217 struct zoneref *z;
9218
9219 zonelist = node_zonelist(nid, gfp_mask);
9220 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9221 gfp_zone(gfp_mask), nodemask) {
9222 spin_lock_irqsave(&zone->lock, flags);
9223
9224 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9225 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9226 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9227 /*
9228 * We release the zone lock here because
9229 * alloc_contig_range() will also lock the zone
9230 * at some point. If there's an allocation
9231 * spinning on this lock, it may win the race
9232 * and cause alloc_contig_range() to fail...
9233 */
9234 spin_unlock_irqrestore(&zone->lock, flags);
9235 ret = __alloc_contig_pages(pfn, nr_pages,
9236 gfp_mask);
9237 if (!ret)
9238 return pfn_to_page(pfn);
9239 spin_lock_irqsave(&zone->lock, flags);
9240 }
9241 pfn += nr_pages;
9242 }
9243 spin_unlock_irqrestore(&zone->lock, flags);
9244 }
9245 return NULL;
9246}
9247#endif /* CONFIG_CONTIG_ALLOC */
9248
9249void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9250{
9251 unsigned long count = 0;
9252
9253 for (; nr_pages--; pfn++) {
9254 struct page *page = pfn_to_page(pfn);
9255
9256 count += page_count(page) != 1;
9257 __free_page(page);
9258 }
9259 WARN(count != 0, "%lu pages are still in use!\n", count);
9260}
9261EXPORT_SYMBOL(free_contig_range);
9262
9263/*
9264 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9265 * page high values need to be recalculated.
9266 */
9267void zone_pcp_update(struct zone *zone, int cpu_online)
9268{
9269 mutex_lock(&pcp_batch_high_lock);
9270 zone_set_pageset_high_and_batch(zone, cpu_online);
9271 mutex_unlock(&pcp_batch_high_lock);
9272}
9273
9274/*
9275 * Effectively disable pcplists for the zone by setting the high limit to 0
9276 * and draining all cpus. A concurrent page freeing on another CPU that's about
9277 * to put the page on pcplist will either finish before the drain and the page
9278 * will be drained, or observe the new high limit and skip the pcplist.
9279 *
9280 * Must be paired with a call to zone_pcp_enable().
9281 */
9282void zone_pcp_disable(struct zone *zone)
9283{
9284 mutex_lock(&pcp_batch_high_lock);
9285 __zone_set_pageset_high_and_batch(zone, 0, 1);
9286 __drain_all_pages(zone, true);
9287}
9288
9289void zone_pcp_enable(struct zone *zone)
9290{
9291 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9292 mutex_unlock(&pcp_batch_high_lock);
9293}
9294
9295void zone_pcp_reset(struct zone *zone)
9296{
9297 int cpu;
9298 struct per_cpu_zonestat *pzstats;
9299
9300 if (zone->per_cpu_pageset != &boot_pageset) {
9301 for_each_online_cpu(cpu) {
9302 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9303 drain_zonestat(zone, pzstats);
9304 }
9305 free_percpu(zone->per_cpu_pageset);
9306 free_percpu(zone->per_cpu_zonestats);
9307 zone->per_cpu_pageset = &boot_pageset;
9308 zone->per_cpu_zonestats = &boot_zonestats;
9309 }
9310}
9311
9312#ifdef CONFIG_MEMORY_HOTREMOVE
9313/*
9314 * All pages in the range must be in a single zone, must not contain holes,
9315 * must span full sections, and must be isolated before calling this function.
9316 */
9317void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9318{
9319 unsigned long pfn = start_pfn;
9320 struct page *page;
9321 struct zone *zone;
9322 unsigned int order;
9323 unsigned long flags;
9324
9325 offline_mem_sections(pfn, end_pfn);
9326 zone = page_zone(pfn_to_page(pfn));
9327 spin_lock_irqsave(&zone->lock, flags);
9328 while (pfn < end_pfn) {
9329 page = pfn_to_page(pfn);
9330 /*
9331 * The HWPoisoned page may be not in buddy system, and
9332 * page_count() is not 0.
9333 */
9334 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9335 pfn++;
9336 continue;
9337 }
9338 /*
9339 * At this point all remaining PageOffline() pages have a
9340 * reference count of 0 and can simply be skipped.
9341 */
9342 if (PageOffline(page)) {
9343 BUG_ON(page_count(page));
9344 BUG_ON(PageBuddy(page));
9345 pfn++;
9346 continue;
9347 }
9348
9349 BUG_ON(page_count(page));
9350 BUG_ON(!PageBuddy(page));
9351 order = buddy_order(page);
9352 del_page_from_free_list(page, zone, order);
9353 pfn += (1 << order);
9354 }
9355 spin_unlock_irqrestore(&zone->lock, flags);
9356}
9357#endif
9358
9359bool is_free_buddy_page(struct page *page)
9360{
9361 struct zone *zone = page_zone(page);
9362 unsigned long pfn = page_to_pfn(page);
9363 unsigned long flags;
9364 unsigned int order;
9365
9366 spin_lock_irqsave(&zone->lock, flags);
9367 for (order = 0; order < MAX_ORDER; order++) {
9368 struct page *page_head = page - (pfn & ((1 << order) - 1));
9369
9370 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
9371 break;
9372 }
9373 spin_unlock_irqrestore(&zone->lock, flags);
9374
9375 return order < MAX_ORDER;
9376}
9377
9378#ifdef CONFIG_MEMORY_FAILURE
9379/*
9380 * Break down a higher-order page in sub-pages, and keep our target out of
9381 * buddy allocator.
9382 */
9383static void break_down_buddy_pages(struct zone *zone, struct page *page,
9384 struct page *target, int low, int high,
9385 int migratetype)
9386{
9387 unsigned long size = 1 << high;
9388 struct page *current_buddy, *next_page;
9389
9390 while (high > low) {
9391 high--;
9392 size >>= 1;
9393
9394 if (target >= &page[size]) {
9395 next_page = page + size;
9396 current_buddy = page;
9397 } else {
9398 next_page = page;
9399 current_buddy = page + size;
9400 }
9401
9402 if (set_page_guard(zone, current_buddy, high, migratetype))
9403 continue;
9404
9405 if (current_buddy != target) {
9406 add_to_free_list(current_buddy, zone, high, migratetype);
9407 set_buddy_order(current_buddy, high);
9408 page = next_page;
9409 }
9410 }
9411}
9412
9413/*
9414 * Take a page that will be marked as poisoned off the buddy allocator.
9415 */
9416bool take_page_off_buddy(struct page *page)
9417{
9418 struct zone *zone = page_zone(page);
9419 unsigned long pfn = page_to_pfn(page);
9420 unsigned long flags;
9421 unsigned int order;
9422 bool ret = false;
9423
9424 spin_lock_irqsave(&zone->lock, flags);
9425 for (order = 0; order < MAX_ORDER; order++) {
9426 struct page *page_head = page - (pfn & ((1 << order) - 1));
9427 int page_order = buddy_order(page_head);
9428
9429 if (PageBuddy(page_head) && page_order >= order) {
9430 unsigned long pfn_head = page_to_pfn(page_head);
9431 int migratetype = get_pfnblock_migratetype(page_head,
9432 pfn_head);
9433
9434 del_page_from_free_list(page_head, zone, page_order);
9435 break_down_buddy_pages(zone, page_head, page, 0,
9436 page_order, migratetype);
9437 if (!is_migrate_isolate(migratetype))
9438 __mod_zone_freepage_state(zone, -1, migratetype);
9439 ret = true;
9440 break;
9441 }
9442 if (page_count(page_head) > 0)
9443 break;
9444 }
9445 spin_unlock_irqrestore(&zone->lock, flags);
9446 return ret;
9447}
9448#endif
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/page_alloc.c
4 *
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
7 *
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
16 */
17
18#include <linux/stddef.h>
19#include <linux/mm.h>
20#include <linux/highmem.h>
21#include <linux/swap.h>
22#include <linux/swapops.h>
23#include <linux/interrupt.h>
24#include <linux/pagemap.h>
25#include <linux/jiffies.h>
26#include <linux/memblock.h>
27#include <linux/compiler.h>
28#include <linux/kernel.h>
29#include <linux/kasan.h>
30#include <linux/kmsan.h>
31#include <linux/module.h>
32#include <linux/suspend.h>
33#include <linux/pagevec.h>
34#include <linux/blkdev.h>
35#include <linux/slab.h>
36#include <linux/ratelimit.h>
37#include <linux/oom.h>
38#include <linux/topology.h>
39#include <linux/sysctl.h>
40#include <linux/cpu.h>
41#include <linux/cpuset.h>
42#include <linux/memory_hotplug.h>
43#include <linux/nodemask.h>
44#include <linux/vmalloc.h>
45#include <linux/vmstat.h>
46#include <linux/mempolicy.h>
47#include <linux/memremap.h>
48#include <linux/stop_machine.h>
49#include <linux/random.h>
50#include <linux/sort.h>
51#include <linux/pfn.h>
52#include <linux/backing-dev.h>
53#include <linux/fault-inject.h>
54#include <linux/page-isolation.h>
55#include <linux/debugobjects.h>
56#include <linux/kmemleak.h>
57#include <linux/compaction.h>
58#include <trace/events/kmem.h>
59#include <trace/events/oom.h>
60#include <linux/prefetch.h>
61#include <linux/mm_inline.h>
62#include <linux/mmu_notifier.h>
63#include <linux/migrate.h>
64#include <linux/hugetlb.h>
65#include <linux/sched/rt.h>
66#include <linux/sched/mm.h>
67#include <linux/page_owner.h>
68#include <linux/page_table_check.h>
69#include <linux/kthread.h>
70#include <linux/memcontrol.h>
71#include <linux/ftrace.h>
72#include <linux/lockdep.h>
73#include <linux/nmi.h>
74#include <linux/psi.h>
75#include <linux/padata.h>
76#include <linux/khugepaged.h>
77#include <linux/buffer_head.h>
78#include <linux/delayacct.h>
79#include <asm/sections.h>
80#include <asm/tlbflush.h>
81#include <asm/div64.h>
82#include "internal.h"
83#include "shuffle.h"
84#include "page_reporting.h"
85#include "swap.h"
86
87/* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
88typedef int __bitwise fpi_t;
89
90/* No special request */
91#define FPI_NONE ((__force fpi_t)0)
92
93/*
94 * Skip free page reporting notification for the (possibly merged) page.
95 * This does not hinder free page reporting from grabbing the page,
96 * reporting it and marking it "reported" - it only skips notifying
97 * the free page reporting infrastructure about a newly freed page. For
98 * example, used when temporarily pulling a page from a freelist and
99 * putting it back unmodified.
100 */
101#define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
102
103/*
104 * Place the (possibly merged) page to the tail of the freelist. Will ignore
105 * page shuffling (relevant code - e.g., memory onlining - is expected to
106 * shuffle the whole zone).
107 *
108 * Note: No code should rely on this flag for correctness - it's purely
109 * to allow for optimizations when handing back either fresh pages
110 * (memory onlining) or untouched pages (page isolation, free page
111 * reporting).
112 */
113#define FPI_TO_TAIL ((__force fpi_t)BIT(1))
114
115/*
116 * Don't poison memory with KASAN (only for the tag-based modes).
117 * During boot, all non-reserved memblock memory is exposed to page_alloc.
118 * Poisoning all that memory lengthens boot time, especially on systems with
119 * large amount of RAM. This flag is used to skip that poisoning.
120 * This is only done for the tag-based KASAN modes, as those are able to
121 * detect memory corruptions with the memory tags assigned by default.
122 * All memory allocated normally after boot gets poisoned as usual.
123 */
124#define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
125
126/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
127static DEFINE_MUTEX(pcp_batch_high_lock);
128#define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
129
130#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
131/*
132 * On SMP, spin_trylock is sufficient protection.
133 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
134 */
135#define pcp_trylock_prepare(flags) do { } while (0)
136#define pcp_trylock_finish(flag) do { } while (0)
137#else
138
139/* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
140#define pcp_trylock_prepare(flags) local_irq_save(flags)
141#define pcp_trylock_finish(flags) local_irq_restore(flags)
142#endif
143
144/*
145 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
146 * a migration causing the wrong PCP to be locked and remote memory being
147 * potentially allocated, pin the task to the CPU for the lookup+lock.
148 * preempt_disable is used on !RT because it is faster than migrate_disable.
149 * migrate_disable is used on RT because otherwise RT spinlock usage is
150 * interfered with and a high priority task cannot preempt the allocator.
151 */
152#ifndef CONFIG_PREEMPT_RT
153#define pcpu_task_pin() preempt_disable()
154#define pcpu_task_unpin() preempt_enable()
155#else
156#define pcpu_task_pin() migrate_disable()
157#define pcpu_task_unpin() migrate_enable()
158#endif
159
160/*
161 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
162 * Return value should be used with equivalent unlock helper.
163 */
164#define pcpu_spin_lock(type, member, ptr) \
165({ \
166 type *_ret; \
167 pcpu_task_pin(); \
168 _ret = this_cpu_ptr(ptr); \
169 spin_lock(&_ret->member); \
170 _ret; \
171})
172
173#define pcpu_spin_trylock(type, member, ptr) \
174({ \
175 type *_ret; \
176 pcpu_task_pin(); \
177 _ret = this_cpu_ptr(ptr); \
178 if (!spin_trylock(&_ret->member)) { \
179 pcpu_task_unpin(); \
180 _ret = NULL; \
181 } \
182 _ret; \
183})
184
185#define pcpu_spin_unlock(member, ptr) \
186({ \
187 spin_unlock(&ptr->member); \
188 pcpu_task_unpin(); \
189})
190
191/* struct per_cpu_pages specific helpers. */
192#define pcp_spin_lock(ptr) \
193 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
194
195#define pcp_spin_trylock(ptr) \
196 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
197
198#define pcp_spin_unlock(ptr) \
199 pcpu_spin_unlock(lock, ptr)
200
201#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
202DEFINE_PER_CPU(int, numa_node);
203EXPORT_PER_CPU_SYMBOL(numa_node);
204#endif
205
206DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
207
208#ifdef CONFIG_HAVE_MEMORYLESS_NODES
209/*
210 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
211 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
212 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
213 * defined in <linux/topology.h>.
214 */
215DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
216EXPORT_PER_CPU_SYMBOL(_numa_mem_);
217#endif
218
219static DEFINE_MUTEX(pcpu_drain_mutex);
220
221#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
222volatile unsigned long latent_entropy __latent_entropy;
223EXPORT_SYMBOL(latent_entropy);
224#endif
225
226/*
227 * Array of node states.
228 */
229nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
230 [N_POSSIBLE] = NODE_MASK_ALL,
231 [N_ONLINE] = { { [0] = 1UL } },
232#ifndef CONFIG_NUMA
233 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
234#ifdef CONFIG_HIGHMEM
235 [N_HIGH_MEMORY] = { { [0] = 1UL } },
236#endif
237 [N_MEMORY] = { { [0] = 1UL } },
238 [N_CPU] = { { [0] = 1UL } },
239#endif /* NUMA */
240};
241EXPORT_SYMBOL(node_states);
242
243atomic_long_t _totalram_pages __read_mostly;
244EXPORT_SYMBOL(_totalram_pages);
245unsigned long totalreserve_pages __read_mostly;
246unsigned long totalcma_pages __read_mostly;
247
248int percpu_pagelist_high_fraction;
249gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
250DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
251EXPORT_SYMBOL(init_on_alloc);
252
253DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
254EXPORT_SYMBOL(init_on_free);
255
256static bool _init_on_alloc_enabled_early __read_mostly
257 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
258static int __init early_init_on_alloc(char *buf)
259{
260
261 return kstrtobool(buf, &_init_on_alloc_enabled_early);
262}
263early_param("init_on_alloc", early_init_on_alloc);
264
265static bool _init_on_free_enabled_early __read_mostly
266 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
267static int __init early_init_on_free(char *buf)
268{
269 return kstrtobool(buf, &_init_on_free_enabled_early);
270}
271early_param("init_on_free", early_init_on_free);
272
273/*
274 * A cached value of the page's pageblock's migratetype, used when the page is
275 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
276 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
277 * Also the migratetype set in the page does not necessarily match the pcplist
278 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
279 * other index - this ensures that it will be put on the correct CMA freelist.
280 */
281static inline int get_pcppage_migratetype(struct page *page)
282{
283 return page->index;
284}
285
286static inline void set_pcppage_migratetype(struct page *page, int migratetype)
287{
288 page->index = migratetype;
289}
290
291#ifdef CONFIG_PM_SLEEP
292/*
293 * The following functions are used by the suspend/hibernate code to temporarily
294 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
295 * while devices are suspended. To avoid races with the suspend/hibernate code,
296 * they should always be called with system_transition_mutex held
297 * (gfp_allowed_mask also should only be modified with system_transition_mutex
298 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
299 * with that modification).
300 */
301
302static gfp_t saved_gfp_mask;
303
304void pm_restore_gfp_mask(void)
305{
306 WARN_ON(!mutex_is_locked(&system_transition_mutex));
307 if (saved_gfp_mask) {
308 gfp_allowed_mask = saved_gfp_mask;
309 saved_gfp_mask = 0;
310 }
311}
312
313void pm_restrict_gfp_mask(void)
314{
315 WARN_ON(!mutex_is_locked(&system_transition_mutex));
316 WARN_ON(saved_gfp_mask);
317 saved_gfp_mask = gfp_allowed_mask;
318 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
319}
320
321bool pm_suspended_storage(void)
322{
323 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
324 return false;
325 return true;
326}
327#endif /* CONFIG_PM_SLEEP */
328
329#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
330unsigned int pageblock_order __read_mostly;
331#endif
332
333static void __free_pages_ok(struct page *page, unsigned int order,
334 fpi_t fpi_flags);
335
336/*
337 * results with 256, 32 in the lowmem_reserve sysctl:
338 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
339 * 1G machine -> (16M dma, 784M normal, 224M high)
340 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
341 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
342 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
343 *
344 * TBD: should special case ZONE_DMA32 machines here - in those we normally
345 * don't need any ZONE_NORMAL reservation
346 */
347int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
348#ifdef CONFIG_ZONE_DMA
349 [ZONE_DMA] = 256,
350#endif
351#ifdef CONFIG_ZONE_DMA32
352 [ZONE_DMA32] = 256,
353#endif
354 [ZONE_NORMAL] = 32,
355#ifdef CONFIG_HIGHMEM
356 [ZONE_HIGHMEM] = 0,
357#endif
358 [ZONE_MOVABLE] = 0,
359};
360
361static char * const zone_names[MAX_NR_ZONES] = {
362#ifdef CONFIG_ZONE_DMA
363 "DMA",
364#endif
365#ifdef CONFIG_ZONE_DMA32
366 "DMA32",
367#endif
368 "Normal",
369#ifdef CONFIG_HIGHMEM
370 "HighMem",
371#endif
372 "Movable",
373#ifdef CONFIG_ZONE_DEVICE
374 "Device",
375#endif
376};
377
378const char * const migratetype_names[MIGRATE_TYPES] = {
379 "Unmovable",
380 "Movable",
381 "Reclaimable",
382 "HighAtomic",
383#ifdef CONFIG_CMA
384 "CMA",
385#endif
386#ifdef CONFIG_MEMORY_ISOLATION
387 "Isolate",
388#endif
389};
390
391compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
392 [NULL_COMPOUND_DTOR] = NULL,
393 [COMPOUND_PAGE_DTOR] = free_compound_page,
394#ifdef CONFIG_HUGETLB_PAGE
395 [HUGETLB_PAGE_DTOR] = free_huge_page,
396#endif
397#ifdef CONFIG_TRANSPARENT_HUGEPAGE
398 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
399#endif
400};
401
402int min_free_kbytes = 1024;
403int user_min_free_kbytes = -1;
404int watermark_boost_factor __read_mostly = 15000;
405int watermark_scale_factor = 10;
406
407static unsigned long nr_kernel_pages __initdata;
408static unsigned long nr_all_pages __initdata;
409static unsigned long dma_reserve __initdata;
410
411static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
412static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
413static unsigned long required_kernelcore __initdata;
414static unsigned long required_kernelcore_percent __initdata;
415static unsigned long required_movablecore __initdata;
416static unsigned long required_movablecore_percent __initdata;
417static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
418bool mirrored_kernelcore __initdata_memblock;
419
420/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
421int movable_zone;
422EXPORT_SYMBOL(movable_zone);
423
424#if MAX_NUMNODES > 1
425unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
426unsigned int nr_online_nodes __read_mostly = 1;
427EXPORT_SYMBOL(nr_node_ids);
428EXPORT_SYMBOL(nr_online_nodes);
429#endif
430
431int page_group_by_mobility_disabled __read_mostly;
432
433#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
434/*
435 * During boot we initialize deferred pages on-demand, as needed, but once
436 * page_alloc_init_late() has finished, the deferred pages are all initialized,
437 * and we can permanently disable that path.
438 */
439static DEFINE_STATIC_KEY_TRUE(deferred_pages);
440
441static inline bool deferred_pages_enabled(void)
442{
443 return static_branch_unlikely(&deferred_pages);
444}
445
446/* Returns true if the struct page for the pfn is uninitialised */
447static inline bool __meminit early_page_uninitialised(unsigned long pfn)
448{
449 int nid = early_pfn_to_nid(pfn);
450
451 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
452 return true;
453
454 return false;
455}
456
457/*
458 * Returns true when the remaining initialisation should be deferred until
459 * later in the boot cycle when it can be parallelised.
460 */
461static bool __meminit
462defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
463{
464 static unsigned long prev_end_pfn, nr_initialised;
465
466 if (early_page_ext_enabled())
467 return false;
468 /*
469 * prev_end_pfn static that contains the end of previous zone
470 * No need to protect because called very early in boot before smp_init.
471 */
472 if (prev_end_pfn != end_pfn) {
473 prev_end_pfn = end_pfn;
474 nr_initialised = 0;
475 }
476
477 /* Always populate low zones for address-constrained allocations */
478 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
479 return false;
480
481 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
482 return true;
483 /*
484 * We start only with one section of pages, more pages are added as
485 * needed until the rest of deferred pages are initialized.
486 */
487 nr_initialised++;
488 if ((nr_initialised > PAGES_PER_SECTION) &&
489 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
490 NODE_DATA(nid)->first_deferred_pfn = pfn;
491 return true;
492 }
493 return false;
494}
495#else
496static inline bool deferred_pages_enabled(void)
497{
498 return false;
499}
500
501static inline bool early_page_uninitialised(unsigned long pfn)
502{
503 return false;
504}
505
506static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
507{
508 return false;
509}
510#endif
511
512/* Return a pointer to the bitmap storing bits affecting a block of pages */
513static inline unsigned long *get_pageblock_bitmap(const struct page *page,
514 unsigned long pfn)
515{
516#ifdef CONFIG_SPARSEMEM
517 return section_to_usemap(__pfn_to_section(pfn));
518#else
519 return page_zone(page)->pageblock_flags;
520#endif /* CONFIG_SPARSEMEM */
521}
522
523static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
524{
525#ifdef CONFIG_SPARSEMEM
526 pfn &= (PAGES_PER_SECTION-1);
527#else
528 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
529#endif /* CONFIG_SPARSEMEM */
530 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
531}
532
533static __always_inline
534unsigned long __get_pfnblock_flags_mask(const struct page *page,
535 unsigned long pfn,
536 unsigned long mask)
537{
538 unsigned long *bitmap;
539 unsigned long bitidx, word_bitidx;
540 unsigned long word;
541
542 bitmap = get_pageblock_bitmap(page, pfn);
543 bitidx = pfn_to_bitidx(page, pfn);
544 word_bitidx = bitidx / BITS_PER_LONG;
545 bitidx &= (BITS_PER_LONG-1);
546 /*
547 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
548 * a consistent read of the memory array, so that results, even though
549 * racy, are not corrupted.
550 */
551 word = READ_ONCE(bitmap[word_bitidx]);
552 return (word >> bitidx) & mask;
553}
554
555/**
556 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
557 * @page: The page within the block of interest
558 * @pfn: The target page frame number
559 * @mask: mask of bits that the caller is interested in
560 *
561 * Return: pageblock_bits flags
562 */
563unsigned long get_pfnblock_flags_mask(const struct page *page,
564 unsigned long pfn, unsigned long mask)
565{
566 return __get_pfnblock_flags_mask(page, pfn, mask);
567}
568
569static __always_inline int get_pfnblock_migratetype(const struct page *page,
570 unsigned long pfn)
571{
572 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
573}
574
575/**
576 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
577 * @page: The page within the block of interest
578 * @flags: The flags to set
579 * @pfn: The target page frame number
580 * @mask: mask of bits that the caller is interested in
581 */
582void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
583 unsigned long pfn,
584 unsigned long mask)
585{
586 unsigned long *bitmap;
587 unsigned long bitidx, word_bitidx;
588 unsigned long word;
589
590 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
591 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
592
593 bitmap = get_pageblock_bitmap(page, pfn);
594 bitidx = pfn_to_bitidx(page, pfn);
595 word_bitidx = bitidx / BITS_PER_LONG;
596 bitidx &= (BITS_PER_LONG-1);
597
598 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
599
600 mask <<= bitidx;
601 flags <<= bitidx;
602
603 word = READ_ONCE(bitmap[word_bitidx]);
604 do {
605 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
606}
607
608void set_pageblock_migratetype(struct page *page, int migratetype)
609{
610 if (unlikely(page_group_by_mobility_disabled &&
611 migratetype < MIGRATE_PCPTYPES))
612 migratetype = MIGRATE_UNMOVABLE;
613
614 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
615 page_to_pfn(page), MIGRATETYPE_MASK);
616}
617
618#ifdef CONFIG_DEBUG_VM
619static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
620{
621 int ret = 0;
622 unsigned seq;
623 unsigned long pfn = page_to_pfn(page);
624 unsigned long sp, start_pfn;
625
626 do {
627 seq = zone_span_seqbegin(zone);
628 start_pfn = zone->zone_start_pfn;
629 sp = zone->spanned_pages;
630 if (!zone_spans_pfn(zone, pfn))
631 ret = 1;
632 } while (zone_span_seqretry(zone, seq));
633
634 if (ret)
635 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
636 pfn, zone_to_nid(zone), zone->name,
637 start_pfn, start_pfn + sp);
638
639 return ret;
640}
641
642static int page_is_consistent(struct zone *zone, struct page *page)
643{
644 if (zone != page_zone(page))
645 return 0;
646
647 return 1;
648}
649/*
650 * Temporary debugging check for pages not lying within a given zone.
651 */
652static int __maybe_unused bad_range(struct zone *zone, struct page *page)
653{
654 if (page_outside_zone_boundaries(zone, page))
655 return 1;
656 if (!page_is_consistent(zone, page))
657 return 1;
658
659 return 0;
660}
661#else
662static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
663{
664 return 0;
665}
666#endif
667
668static void bad_page(struct page *page, const char *reason)
669{
670 static unsigned long resume;
671 static unsigned long nr_shown;
672 static unsigned long nr_unshown;
673
674 /*
675 * Allow a burst of 60 reports, then keep quiet for that minute;
676 * or allow a steady drip of one report per second.
677 */
678 if (nr_shown == 60) {
679 if (time_before(jiffies, resume)) {
680 nr_unshown++;
681 goto out;
682 }
683 if (nr_unshown) {
684 pr_alert(
685 "BUG: Bad page state: %lu messages suppressed\n",
686 nr_unshown);
687 nr_unshown = 0;
688 }
689 nr_shown = 0;
690 }
691 if (nr_shown++ == 0)
692 resume = jiffies + 60 * HZ;
693
694 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
695 current->comm, page_to_pfn(page));
696 dump_page(page, reason);
697
698 print_modules();
699 dump_stack();
700out:
701 /* Leave bad fields for debug, except PageBuddy could make trouble */
702 page_mapcount_reset(page); /* remove PageBuddy */
703 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
704}
705
706static inline unsigned int order_to_pindex(int migratetype, int order)
707{
708 int base = order;
709
710#ifdef CONFIG_TRANSPARENT_HUGEPAGE
711 if (order > PAGE_ALLOC_COSTLY_ORDER) {
712 VM_BUG_ON(order != pageblock_order);
713 return NR_LOWORDER_PCP_LISTS;
714 }
715#else
716 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
717#endif
718
719 return (MIGRATE_PCPTYPES * base) + migratetype;
720}
721
722static inline int pindex_to_order(unsigned int pindex)
723{
724 int order = pindex / MIGRATE_PCPTYPES;
725
726#ifdef CONFIG_TRANSPARENT_HUGEPAGE
727 if (pindex == NR_LOWORDER_PCP_LISTS)
728 order = pageblock_order;
729#else
730 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
731#endif
732
733 return order;
734}
735
736static inline bool pcp_allowed_order(unsigned int order)
737{
738 if (order <= PAGE_ALLOC_COSTLY_ORDER)
739 return true;
740#ifdef CONFIG_TRANSPARENT_HUGEPAGE
741 if (order == pageblock_order)
742 return true;
743#endif
744 return false;
745}
746
747static inline void free_the_page(struct page *page, unsigned int order)
748{
749 if (pcp_allowed_order(order)) /* Via pcp? */
750 free_unref_page(page, order);
751 else
752 __free_pages_ok(page, order, FPI_NONE);
753}
754
755/*
756 * Higher-order pages are called "compound pages". They are structured thusly:
757 *
758 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
759 *
760 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
761 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
762 *
763 * The first tail page's ->compound_dtor holds the offset in array of compound
764 * page destructors. See compound_page_dtors.
765 *
766 * The first tail page's ->compound_order holds the order of allocation.
767 * This usage means that zero-order pages may not be compound.
768 */
769
770void free_compound_page(struct page *page)
771{
772 mem_cgroup_uncharge(page_folio(page));
773 free_the_page(page, compound_order(page));
774}
775
776static void prep_compound_head(struct page *page, unsigned int order)
777{
778 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
779 set_compound_order(page, order);
780 atomic_set(compound_mapcount_ptr(page), -1);
781 atomic_set(subpages_mapcount_ptr(page), 0);
782 atomic_set(compound_pincount_ptr(page), 0);
783}
784
785static void prep_compound_tail(struct page *head, int tail_idx)
786{
787 struct page *p = head + tail_idx;
788
789 p->mapping = TAIL_MAPPING;
790 set_compound_head(p, head);
791 set_page_private(p, 0);
792}
793
794void prep_compound_page(struct page *page, unsigned int order)
795{
796 int i;
797 int nr_pages = 1 << order;
798
799 __SetPageHead(page);
800 for (i = 1; i < nr_pages; i++)
801 prep_compound_tail(page, i);
802
803 prep_compound_head(page, order);
804}
805
806void destroy_large_folio(struct folio *folio)
807{
808 enum compound_dtor_id dtor = folio_page(folio, 1)->compound_dtor;
809
810 VM_BUG_ON_FOLIO(dtor >= NR_COMPOUND_DTORS, folio);
811 compound_page_dtors[dtor](&folio->page);
812}
813
814#ifdef CONFIG_DEBUG_PAGEALLOC
815unsigned int _debug_guardpage_minorder;
816
817bool _debug_pagealloc_enabled_early __read_mostly
818 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
819EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
820DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
821EXPORT_SYMBOL(_debug_pagealloc_enabled);
822
823DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
824
825static int __init early_debug_pagealloc(char *buf)
826{
827 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
828}
829early_param("debug_pagealloc", early_debug_pagealloc);
830
831static int __init debug_guardpage_minorder_setup(char *buf)
832{
833 unsigned long res;
834
835 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
836 pr_err("Bad debug_guardpage_minorder value\n");
837 return 0;
838 }
839 _debug_guardpage_minorder = res;
840 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
841 return 0;
842}
843early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
844
845static inline bool set_page_guard(struct zone *zone, struct page *page,
846 unsigned int order, int migratetype)
847{
848 if (!debug_guardpage_enabled())
849 return false;
850
851 if (order >= debug_guardpage_minorder())
852 return false;
853
854 __SetPageGuard(page);
855 INIT_LIST_HEAD(&page->buddy_list);
856 set_page_private(page, order);
857 /* Guard pages are not available for any usage */
858 if (!is_migrate_isolate(migratetype))
859 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
860
861 return true;
862}
863
864static inline void clear_page_guard(struct zone *zone, struct page *page,
865 unsigned int order, int migratetype)
866{
867 if (!debug_guardpage_enabled())
868 return;
869
870 __ClearPageGuard(page);
871
872 set_page_private(page, 0);
873 if (!is_migrate_isolate(migratetype))
874 __mod_zone_freepage_state(zone, (1 << order), migratetype);
875}
876#else
877static inline bool set_page_guard(struct zone *zone, struct page *page,
878 unsigned int order, int migratetype) { return false; }
879static inline void clear_page_guard(struct zone *zone, struct page *page,
880 unsigned int order, int migratetype) {}
881#endif
882
883/*
884 * Enable static keys related to various memory debugging and hardening options.
885 * Some override others, and depend on early params that are evaluated in the
886 * order of appearance. So we need to first gather the full picture of what was
887 * enabled, and then make decisions.
888 */
889void __init init_mem_debugging_and_hardening(void)
890{
891 bool page_poisoning_requested = false;
892
893#ifdef CONFIG_PAGE_POISONING
894 /*
895 * Page poisoning is debug page alloc for some arches. If
896 * either of those options are enabled, enable poisoning.
897 */
898 if (page_poisoning_enabled() ||
899 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
900 debug_pagealloc_enabled())) {
901 static_branch_enable(&_page_poisoning_enabled);
902 page_poisoning_requested = true;
903 }
904#endif
905
906 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
907 page_poisoning_requested) {
908 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
909 "will take precedence over init_on_alloc and init_on_free\n");
910 _init_on_alloc_enabled_early = false;
911 _init_on_free_enabled_early = false;
912 }
913
914 if (_init_on_alloc_enabled_early)
915 static_branch_enable(&init_on_alloc);
916 else
917 static_branch_disable(&init_on_alloc);
918
919 if (_init_on_free_enabled_early)
920 static_branch_enable(&init_on_free);
921 else
922 static_branch_disable(&init_on_free);
923
924 if (IS_ENABLED(CONFIG_KMSAN) &&
925 (_init_on_alloc_enabled_early || _init_on_free_enabled_early))
926 pr_info("mem auto-init: please make sure init_on_alloc and init_on_free are disabled when running KMSAN\n");
927
928#ifdef CONFIG_DEBUG_PAGEALLOC
929 if (!debug_pagealloc_enabled())
930 return;
931
932 static_branch_enable(&_debug_pagealloc_enabled);
933
934 if (!debug_guardpage_minorder())
935 return;
936
937 static_branch_enable(&_debug_guardpage_enabled);
938#endif
939}
940
941static inline void set_buddy_order(struct page *page, unsigned int order)
942{
943 set_page_private(page, order);
944 __SetPageBuddy(page);
945}
946
947#ifdef CONFIG_COMPACTION
948static inline struct capture_control *task_capc(struct zone *zone)
949{
950 struct capture_control *capc = current->capture_control;
951
952 return unlikely(capc) &&
953 !(current->flags & PF_KTHREAD) &&
954 !capc->page &&
955 capc->cc->zone == zone ? capc : NULL;
956}
957
958static inline bool
959compaction_capture(struct capture_control *capc, struct page *page,
960 int order, int migratetype)
961{
962 if (!capc || order != capc->cc->order)
963 return false;
964
965 /* Do not accidentally pollute CMA or isolated regions*/
966 if (is_migrate_cma(migratetype) ||
967 is_migrate_isolate(migratetype))
968 return false;
969
970 /*
971 * Do not let lower order allocations pollute a movable pageblock.
972 * This might let an unmovable request use a reclaimable pageblock
973 * and vice-versa but no more than normal fallback logic which can
974 * have trouble finding a high-order free page.
975 */
976 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
977 return false;
978
979 capc->page = page;
980 return true;
981}
982
983#else
984static inline struct capture_control *task_capc(struct zone *zone)
985{
986 return NULL;
987}
988
989static inline bool
990compaction_capture(struct capture_control *capc, struct page *page,
991 int order, int migratetype)
992{
993 return false;
994}
995#endif /* CONFIG_COMPACTION */
996
997/* Used for pages not on another list */
998static inline void add_to_free_list(struct page *page, struct zone *zone,
999 unsigned int order, int migratetype)
1000{
1001 struct free_area *area = &zone->free_area[order];
1002
1003 list_add(&page->buddy_list, &area->free_list[migratetype]);
1004 area->nr_free++;
1005}
1006
1007/* Used for pages not on another list */
1008static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
1009 unsigned int order, int migratetype)
1010{
1011 struct free_area *area = &zone->free_area[order];
1012
1013 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
1014 area->nr_free++;
1015}
1016
1017/*
1018 * Used for pages which are on another list. Move the pages to the tail
1019 * of the list - so the moved pages won't immediately be considered for
1020 * allocation again (e.g., optimization for memory onlining).
1021 */
1022static inline void move_to_free_list(struct page *page, struct zone *zone,
1023 unsigned int order, int migratetype)
1024{
1025 struct free_area *area = &zone->free_area[order];
1026
1027 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
1028}
1029
1030static inline void del_page_from_free_list(struct page *page, struct zone *zone,
1031 unsigned int order)
1032{
1033 /* clear reported state and update reported page count */
1034 if (page_reported(page))
1035 __ClearPageReported(page);
1036
1037 list_del(&page->buddy_list);
1038 __ClearPageBuddy(page);
1039 set_page_private(page, 0);
1040 zone->free_area[order].nr_free--;
1041}
1042
1043/*
1044 * If this is not the largest possible page, check if the buddy
1045 * of the next-highest order is free. If it is, it's possible
1046 * that pages are being freed that will coalesce soon. In case,
1047 * that is happening, add the free page to the tail of the list
1048 * so it's less likely to be used soon and more likely to be merged
1049 * as a higher order page
1050 */
1051static inline bool
1052buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1053 struct page *page, unsigned int order)
1054{
1055 unsigned long higher_page_pfn;
1056 struct page *higher_page;
1057
1058 if (order >= MAX_ORDER - 2)
1059 return false;
1060
1061 higher_page_pfn = buddy_pfn & pfn;
1062 higher_page = page + (higher_page_pfn - pfn);
1063
1064 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
1065 NULL) != NULL;
1066}
1067
1068/*
1069 * Freeing function for a buddy system allocator.
1070 *
1071 * The concept of a buddy system is to maintain direct-mapped table
1072 * (containing bit values) for memory blocks of various "orders".
1073 * The bottom level table contains the map for the smallest allocatable
1074 * units of memory (here, pages), and each level above it describes
1075 * pairs of units from the levels below, hence, "buddies".
1076 * At a high level, all that happens here is marking the table entry
1077 * at the bottom level available, and propagating the changes upward
1078 * as necessary, plus some accounting needed to play nicely with other
1079 * parts of the VM system.
1080 * At each level, we keep a list of pages, which are heads of continuous
1081 * free pages of length of (1 << order) and marked with PageBuddy.
1082 * Page's order is recorded in page_private(page) field.
1083 * So when we are allocating or freeing one, we can derive the state of the
1084 * other. That is, if we allocate a small block, and both were
1085 * free, the remainder of the region must be split into blocks.
1086 * If a block is freed, and its buddy is also free, then this
1087 * triggers coalescing into a block of larger size.
1088 *
1089 * -- nyc
1090 */
1091
1092static inline void __free_one_page(struct page *page,
1093 unsigned long pfn,
1094 struct zone *zone, unsigned int order,
1095 int migratetype, fpi_t fpi_flags)
1096{
1097 struct capture_control *capc = task_capc(zone);
1098 unsigned long buddy_pfn = 0;
1099 unsigned long combined_pfn;
1100 struct page *buddy;
1101 bool to_tail;
1102
1103 VM_BUG_ON(!zone_is_initialized(zone));
1104 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1105
1106 VM_BUG_ON(migratetype == -1);
1107 if (likely(!is_migrate_isolate(migratetype)))
1108 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1109
1110 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1111 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1112
1113 while (order < MAX_ORDER - 1) {
1114 if (compaction_capture(capc, page, order, migratetype)) {
1115 __mod_zone_freepage_state(zone, -(1 << order),
1116 migratetype);
1117 return;
1118 }
1119
1120 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
1121 if (!buddy)
1122 goto done_merging;
1123
1124 if (unlikely(order >= pageblock_order)) {
1125 /*
1126 * We want to prevent merge between freepages on pageblock
1127 * without fallbacks and normal pageblock. Without this,
1128 * pageblock isolation could cause incorrect freepage or CMA
1129 * accounting or HIGHATOMIC accounting.
1130 */
1131 int buddy_mt = get_pageblock_migratetype(buddy);
1132
1133 if (migratetype != buddy_mt
1134 && (!migratetype_is_mergeable(migratetype) ||
1135 !migratetype_is_mergeable(buddy_mt)))
1136 goto done_merging;
1137 }
1138
1139 /*
1140 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1141 * merge with it and move up one order.
1142 */
1143 if (page_is_guard(buddy))
1144 clear_page_guard(zone, buddy, order, migratetype);
1145 else
1146 del_page_from_free_list(buddy, zone, order);
1147 combined_pfn = buddy_pfn & pfn;
1148 page = page + (combined_pfn - pfn);
1149 pfn = combined_pfn;
1150 order++;
1151 }
1152
1153done_merging:
1154 set_buddy_order(page, order);
1155
1156 if (fpi_flags & FPI_TO_TAIL)
1157 to_tail = true;
1158 else if (is_shuffle_order(order))
1159 to_tail = shuffle_pick_tail();
1160 else
1161 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1162
1163 if (to_tail)
1164 add_to_free_list_tail(page, zone, order, migratetype);
1165 else
1166 add_to_free_list(page, zone, order, migratetype);
1167
1168 /* Notify page reporting subsystem of freed page */
1169 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1170 page_reporting_notify_free(order);
1171}
1172
1173/**
1174 * split_free_page() -- split a free page at split_pfn_offset
1175 * @free_page: the original free page
1176 * @order: the order of the page
1177 * @split_pfn_offset: split offset within the page
1178 *
1179 * Return -ENOENT if the free page is changed, otherwise 0
1180 *
1181 * It is used when the free page crosses two pageblocks with different migratetypes
1182 * at split_pfn_offset within the page. The split free page will be put into
1183 * separate migratetype lists afterwards. Otherwise, the function achieves
1184 * nothing.
1185 */
1186int split_free_page(struct page *free_page,
1187 unsigned int order, unsigned long split_pfn_offset)
1188{
1189 struct zone *zone = page_zone(free_page);
1190 unsigned long free_page_pfn = page_to_pfn(free_page);
1191 unsigned long pfn;
1192 unsigned long flags;
1193 int free_page_order;
1194 int mt;
1195 int ret = 0;
1196
1197 if (split_pfn_offset == 0)
1198 return ret;
1199
1200 spin_lock_irqsave(&zone->lock, flags);
1201
1202 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
1203 ret = -ENOENT;
1204 goto out;
1205 }
1206
1207 mt = get_pageblock_migratetype(free_page);
1208 if (likely(!is_migrate_isolate(mt)))
1209 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1210
1211 del_page_from_free_list(free_page, zone, order);
1212 for (pfn = free_page_pfn;
1213 pfn < free_page_pfn + (1UL << order);) {
1214 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
1215
1216 free_page_order = min_t(unsigned int,
1217 pfn ? __ffs(pfn) : order,
1218 __fls(split_pfn_offset));
1219 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
1220 mt, FPI_NONE);
1221 pfn += 1UL << free_page_order;
1222 split_pfn_offset -= (1UL << free_page_order);
1223 /* we have done the first part, now switch to second part */
1224 if (split_pfn_offset == 0)
1225 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
1226 }
1227out:
1228 spin_unlock_irqrestore(&zone->lock, flags);
1229 return ret;
1230}
1231/*
1232 * A bad page could be due to a number of fields. Instead of multiple branches,
1233 * try and check multiple fields with one check. The caller must do a detailed
1234 * check if necessary.
1235 */
1236static inline bool page_expected_state(struct page *page,
1237 unsigned long check_flags)
1238{
1239 if (unlikely(atomic_read(&page->_mapcount) != -1))
1240 return false;
1241
1242 if (unlikely((unsigned long)page->mapping |
1243 page_ref_count(page) |
1244#ifdef CONFIG_MEMCG
1245 page->memcg_data |
1246#endif
1247 (page->flags & check_flags)))
1248 return false;
1249
1250 return true;
1251}
1252
1253static const char *page_bad_reason(struct page *page, unsigned long flags)
1254{
1255 const char *bad_reason = NULL;
1256
1257 if (unlikely(atomic_read(&page->_mapcount) != -1))
1258 bad_reason = "nonzero mapcount";
1259 if (unlikely(page->mapping != NULL))
1260 bad_reason = "non-NULL mapping";
1261 if (unlikely(page_ref_count(page) != 0))
1262 bad_reason = "nonzero _refcount";
1263 if (unlikely(page->flags & flags)) {
1264 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1265 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1266 else
1267 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1268 }
1269#ifdef CONFIG_MEMCG
1270 if (unlikely(page->memcg_data))
1271 bad_reason = "page still charged to cgroup";
1272#endif
1273 return bad_reason;
1274}
1275
1276static void free_page_is_bad_report(struct page *page)
1277{
1278 bad_page(page,
1279 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1280}
1281
1282static inline bool free_page_is_bad(struct page *page)
1283{
1284 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1285 return false;
1286
1287 /* Something has gone sideways, find it */
1288 free_page_is_bad_report(page);
1289 return true;
1290}
1291
1292static int free_tail_pages_check(struct page *head_page, struct page *page)
1293{
1294 int ret = 1;
1295
1296 /*
1297 * We rely page->lru.next never has bit 0 set, unless the page
1298 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1299 */
1300 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1301
1302 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1303 ret = 0;
1304 goto out;
1305 }
1306 switch (page - head_page) {
1307 case 1:
1308 /* the first tail page: these may be in place of ->mapping */
1309 if (unlikely(head_compound_mapcount(head_page))) {
1310 bad_page(page, "nonzero compound_mapcount");
1311 goto out;
1312 }
1313 if (unlikely(atomic_read(subpages_mapcount_ptr(head_page)))) {
1314 bad_page(page, "nonzero subpages_mapcount");
1315 goto out;
1316 }
1317 if (unlikely(head_compound_pincount(head_page))) {
1318 bad_page(page, "nonzero compound_pincount");
1319 goto out;
1320 }
1321 break;
1322 case 2:
1323 /*
1324 * the second tail page: ->mapping is
1325 * deferred_list.next -- ignore value.
1326 */
1327 break;
1328 default:
1329 if (page->mapping != TAIL_MAPPING) {
1330 bad_page(page, "corrupted mapping in tail page");
1331 goto out;
1332 }
1333 break;
1334 }
1335 if (unlikely(!PageTail(page))) {
1336 bad_page(page, "PageTail not set");
1337 goto out;
1338 }
1339 if (unlikely(compound_head(page) != head_page)) {
1340 bad_page(page, "compound_head not consistent");
1341 goto out;
1342 }
1343 ret = 0;
1344out:
1345 page->mapping = NULL;
1346 clear_compound_head(page);
1347 return ret;
1348}
1349
1350/*
1351 * Skip KASAN memory poisoning when either:
1352 *
1353 * 1. Deferred memory initialization has not yet completed,
1354 * see the explanation below.
1355 * 2. Skipping poisoning is requested via FPI_SKIP_KASAN_POISON,
1356 * see the comment next to it.
1357 * 3. Skipping poisoning is requested via __GFP_SKIP_KASAN_POISON,
1358 * see the comment next to it.
1359 *
1360 * Poisoning pages during deferred memory init will greatly lengthen the
1361 * process and cause problem in large memory systems as the deferred pages
1362 * initialization is done with interrupt disabled.
1363 *
1364 * Assuming that there will be no reference to those newly initialized
1365 * pages before they are ever allocated, this should have no effect on
1366 * KASAN memory tracking as the poison will be properly inserted at page
1367 * allocation time. The only corner case is when pages are allocated by
1368 * on-demand allocation and then freed again before the deferred pages
1369 * initialization is done, but this is not likely to happen.
1370 */
1371static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1372{
1373 return deferred_pages_enabled() ||
1374 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
1375 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
1376 PageSkipKASanPoison(page);
1377}
1378
1379static void kernel_init_pages(struct page *page, int numpages)
1380{
1381 int i;
1382
1383 /* s390's use of memset() could override KASAN redzones. */
1384 kasan_disable_current();
1385 for (i = 0; i < numpages; i++)
1386 clear_highpage_kasan_tagged(page + i);
1387 kasan_enable_current();
1388}
1389
1390static __always_inline bool free_pages_prepare(struct page *page,
1391 unsigned int order, bool check_free, fpi_t fpi_flags)
1392{
1393 int bad = 0;
1394 bool init = want_init_on_free();
1395
1396 VM_BUG_ON_PAGE(PageTail(page), page);
1397
1398 trace_mm_page_free(page, order);
1399 kmsan_free_page(page, order);
1400
1401 if (unlikely(PageHWPoison(page)) && !order) {
1402 /*
1403 * Do not let hwpoison pages hit pcplists/buddy
1404 * Untie memcg state and reset page's owner
1405 */
1406 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1407 __memcg_kmem_uncharge_page(page, order);
1408 reset_page_owner(page, order);
1409 page_table_check_free(page, order);
1410 return false;
1411 }
1412
1413 /*
1414 * Check tail pages before head page information is cleared to
1415 * avoid checking PageCompound for order-0 pages.
1416 */
1417 if (unlikely(order)) {
1418 bool compound = PageCompound(page);
1419 int i;
1420
1421 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1422
1423 if (compound)
1424 ClearPageHasHWPoisoned(page);
1425 for (i = 1; i < (1 << order); i++) {
1426 if (compound)
1427 bad += free_tail_pages_check(page, page + i);
1428 if (unlikely(free_page_is_bad(page + i))) {
1429 bad++;
1430 continue;
1431 }
1432 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1433 }
1434 }
1435 if (PageMappingFlags(page))
1436 page->mapping = NULL;
1437 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1438 __memcg_kmem_uncharge_page(page, order);
1439 if (check_free && free_page_is_bad(page))
1440 bad++;
1441 if (bad)
1442 return false;
1443
1444 page_cpupid_reset_last(page);
1445 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1446 reset_page_owner(page, order);
1447 page_table_check_free(page, order);
1448
1449 if (!PageHighMem(page)) {
1450 debug_check_no_locks_freed(page_address(page),
1451 PAGE_SIZE << order);
1452 debug_check_no_obj_freed(page_address(page),
1453 PAGE_SIZE << order);
1454 }
1455
1456 kernel_poison_pages(page, 1 << order);
1457
1458 /*
1459 * As memory initialization might be integrated into KASAN,
1460 * KASAN poisoning and memory initialization code must be
1461 * kept together to avoid discrepancies in behavior.
1462 *
1463 * With hardware tag-based KASAN, memory tags must be set before the
1464 * page becomes unavailable via debug_pagealloc or arch_free_page.
1465 */
1466 if (!should_skip_kasan_poison(page, fpi_flags)) {
1467 kasan_poison_pages(page, order, init);
1468
1469 /* Memory is already initialized if KASAN did it internally. */
1470 if (kasan_has_integrated_init())
1471 init = false;
1472 }
1473 if (init)
1474 kernel_init_pages(page, 1 << order);
1475
1476 /*
1477 * arch_free_page() can make the page's contents inaccessible. s390
1478 * does this. So nothing which can access the page's contents should
1479 * happen after this.
1480 */
1481 arch_free_page(page, order);
1482
1483 debug_pagealloc_unmap_pages(page, 1 << order);
1484
1485 return true;
1486}
1487
1488#ifdef CONFIG_DEBUG_VM
1489/*
1490 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1491 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1492 * moved from pcp lists to free lists.
1493 */
1494static bool free_pcp_prepare(struct page *page, unsigned int order)
1495{
1496 return free_pages_prepare(page, order, true, FPI_NONE);
1497}
1498
1499/* return true if this page has an inappropriate state */
1500static bool bulkfree_pcp_prepare(struct page *page)
1501{
1502 if (debug_pagealloc_enabled_static())
1503 return free_page_is_bad(page);
1504 else
1505 return false;
1506}
1507#else
1508/*
1509 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1510 * moving from pcp lists to free list in order to reduce overhead. With
1511 * debug_pagealloc enabled, they are checked also immediately when being freed
1512 * to the pcp lists.
1513 */
1514static bool free_pcp_prepare(struct page *page, unsigned int order)
1515{
1516 if (debug_pagealloc_enabled_static())
1517 return free_pages_prepare(page, order, true, FPI_NONE);
1518 else
1519 return free_pages_prepare(page, order, false, FPI_NONE);
1520}
1521
1522static bool bulkfree_pcp_prepare(struct page *page)
1523{
1524 return free_page_is_bad(page);
1525}
1526#endif /* CONFIG_DEBUG_VM */
1527
1528/*
1529 * Frees a number of pages from the PCP lists
1530 * Assumes all pages on list are in same zone.
1531 * count is the number of pages to free.
1532 */
1533static void free_pcppages_bulk(struct zone *zone, int count,
1534 struct per_cpu_pages *pcp,
1535 int pindex)
1536{
1537 unsigned long flags;
1538 int min_pindex = 0;
1539 int max_pindex = NR_PCP_LISTS - 1;
1540 unsigned int order;
1541 bool isolated_pageblocks;
1542 struct page *page;
1543
1544 /*
1545 * Ensure proper count is passed which otherwise would stuck in the
1546 * below while (list_empty(list)) loop.
1547 */
1548 count = min(pcp->count, count);
1549
1550 /* Ensure requested pindex is drained first. */
1551 pindex = pindex - 1;
1552
1553 spin_lock_irqsave(&zone->lock, flags);
1554 isolated_pageblocks = has_isolate_pageblock(zone);
1555
1556 while (count > 0) {
1557 struct list_head *list;
1558 int nr_pages;
1559
1560 /* Remove pages from lists in a round-robin fashion. */
1561 do {
1562 if (++pindex > max_pindex)
1563 pindex = min_pindex;
1564 list = &pcp->lists[pindex];
1565 if (!list_empty(list))
1566 break;
1567
1568 if (pindex == max_pindex)
1569 max_pindex--;
1570 if (pindex == min_pindex)
1571 min_pindex++;
1572 } while (1);
1573
1574 order = pindex_to_order(pindex);
1575 nr_pages = 1 << order;
1576 do {
1577 int mt;
1578
1579 page = list_last_entry(list, struct page, pcp_list);
1580 mt = get_pcppage_migratetype(page);
1581
1582 /* must delete to avoid corrupting pcp list */
1583 list_del(&page->pcp_list);
1584 count -= nr_pages;
1585 pcp->count -= nr_pages;
1586
1587 if (bulkfree_pcp_prepare(page))
1588 continue;
1589
1590 /* MIGRATE_ISOLATE page should not go to pcplists */
1591 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1592 /* Pageblock could have been isolated meanwhile */
1593 if (unlikely(isolated_pageblocks))
1594 mt = get_pageblock_migratetype(page);
1595
1596 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1597 trace_mm_page_pcpu_drain(page, order, mt);
1598 } while (count > 0 && !list_empty(list));
1599 }
1600
1601 spin_unlock_irqrestore(&zone->lock, flags);
1602}
1603
1604static void free_one_page(struct zone *zone,
1605 struct page *page, unsigned long pfn,
1606 unsigned int order,
1607 int migratetype, fpi_t fpi_flags)
1608{
1609 unsigned long flags;
1610
1611 spin_lock_irqsave(&zone->lock, flags);
1612 if (unlikely(has_isolate_pageblock(zone) ||
1613 is_migrate_isolate(migratetype))) {
1614 migratetype = get_pfnblock_migratetype(page, pfn);
1615 }
1616 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1617 spin_unlock_irqrestore(&zone->lock, flags);
1618}
1619
1620static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1621 unsigned long zone, int nid)
1622{
1623 mm_zero_struct_page(page);
1624 set_page_links(page, zone, nid, pfn);
1625 init_page_count(page);
1626 page_mapcount_reset(page);
1627 page_cpupid_reset_last(page);
1628 page_kasan_tag_reset(page);
1629
1630 INIT_LIST_HEAD(&page->lru);
1631#ifdef WANT_PAGE_VIRTUAL
1632 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1633 if (!is_highmem_idx(zone))
1634 set_page_address(page, __va(pfn << PAGE_SHIFT));
1635#endif
1636}
1637
1638#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1639static void __meminit init_reserved_page(unsigned long pfn)
1640{
1641 pg_data_t *pgdat;
1642 int nid, zid;
1643
1644 if (!early_page_uninitialised(pfn))
1645 return;
1646
1647 nid = early_pfn_to_nid(pfn);
1648 pgdat = NODE_DATA(nid);
1649
1650 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1651 struct zone *zone = &pgdat->node_zones[zid];
1652
1653 if (zone_spans_pfn(zone, pfn))
1654 break;
1655 }
1656 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1657}
1658#else
1659static inline void init_reserved_page(unsigned long pfn)
1660{
1661}
1662#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1663
1664/*
1665 * Initialised pages do not have PageReserved set. This function is
1666 * called for each range allocated by the bootmem allocator and
1667 * marks the pages PageReserved. The remaining valid pages are later
1668 * sent to the buddy page allocator.
1669 */
1670void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1671{
1672 unsigned long start_pfn = PFN_DOWN(start);
1673 unsigned long end_pfn = PFN_UP(end);
1674
1675 for (; start_pfn < end_pfn; start_pfn++) {
1676 if (pfn_valid(start_pfn)) {
1677 struct page *page = pfn_to_page(start_pfn);
1678
1679 init_reserved_page(start_pfn);
1680
1681 /* Avoid false-positive PageTail() */
1682 INIT_LIST_HEAD(&page->lru);
1683
1684 /*
1685 * no need for atomic set_bit because the struct
1686 * page is not visible yet so nobody should
1687 * access it yet.
1688 */
1689 __SetPageReserved(page);
1690 }
1691 }
1692}
1693
1694static void __free_pages_ok(struct page *page, unsigned int order,
1695 fpi_t fpi_flags)
1696{
1697 unsigned long flags;
1698 int migratetype;
1699 unsigned long pfn = page_to_pfn(page);
1700 struct zone *zone = page_zone(page);
1701
1702 if (!free_pages_prepare(page, order, true, fpi_flags))
1703 return;
1704
1705 /*
1706 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1707 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1708 * This will reduce the lock holding time.
1709 */
1710 migratetype = get_pfnblock_migratetype(page, pfn);
1711
1712 spin_lock_irqsave(&zone->lock, flags);
1713 if (unlikely(has_isolate_pageblock(zone) ||
1714 is_migrate_isolate(migratetype))) {
1715 migratetype = get_pfnblock_migratetype(page, pfn);
1716 }
1717 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1718 spin_unlock_irqrestore(&zone->lock, flags);
1719
1720 __count_vm_events(PGFREE, 1 << order);
1721}
1722
1723void __free_pages_core(struct page *page, unsigned int order)
1724{
1725 unsigned int nr_pages = 1 << order;
1726 struct page *p = page;
1727 unsigned int loop;
1728
1729 /*
1730 * When initializing the memmap, __init_single_page() sets the refcount
1731 * of all pages to 1 ("allocated"/"not free"). We have to set the
1732 * refcount of all involved pages to 0.
1733 */
1734 prefetchw(p);
1735 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1736 prefetchw(p + 1);
1737 __ClearPageReserved(p);
1738 set_page_count(p, 0);
1739 }
1740 __ClearPageReserved(p);
1741 set_page_count(p, 0);
1742
1743 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1744
1745 /*
1746 * Bypass PCP and place fresh pages right to the tail, primarily
1747 * relevant for memory onlining.
1748 */
1749 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1750}
1751
1752#ifdef CONFIG_NUMA
1753
1754/*
1755 * During memory init memblocks map pfns to nids. The search is expensive and
1756 * this caches recent lookups. The implementation of __early_pfn_to_nid
1757 * treats start/end as pfns.
1758 */
1759struct mminit_pfnnid_cache {
1760 unsigned long last_start;
1761 unsigned long last_end;
1762 int last_nid;
1763};
1764
1765static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1766
1767/*
1768 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1769 */
1770static int __meminit __early_pfn_to_nid(unsigned long pfn,
1771 struct mminit_pfnnid_cache *state)
1772{
1773 unsigned long start_pfn, end_pfn;
1774 int nid;
1775
1776 if (state->last_start <= pfn && pfn < state->last_end)
1777 return state->last_nid;
1778
1779 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1780 if (nid != NUMA_NO_NODE) {
1781 state->last_start = start_pfn;
1782 state->last_end = end_pfn;
1783 state->last_nid = nid;
1784 }
1785
1786 return nid;
1787}
1788
1789int __meminit early_pfn_to_nid(unsigned long pfn)
1790{
1791 static DEFINE_SPINLOCK(early_pfn_lock);
1792 int nid;
1793
1794 spin_lock(&early_pfn_lock);
1795 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1796 if (nid < 0)
1797 nid = first_online_node;
1798 spin_unlock(&early_pfn_lock);
1799
1800 return nid;
1801}
1802#endif /* CONFIG_NUMA */
1803
1804void __init memblock_free_pages(struct page *page, unsigned long pfn,
1805 unsigned int order)
1806{
1807 if (early_page_uninitialised(pfn))
1808 return;
1809 if (!kmsan_memblock_free_pages(page, order)) {
1810 /* KMSAN will take care of these pages. */
1811 return;
1812 }
1813 __free_pages_core(page, order);
1814}
1815
1816/*
1817 * Check that the whole (or subset of) a pageblock given by the interval of
1818 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1819 * with the migration of free compaction scanner.
1820 *
1821 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1822 *
1823 * It's possible on some configurations to have a setup like node0 node1 node0
1824 * i.e. it's possible that all pages within a zones range of pages do not
1825 * belong to a single zone. We assume that a border between node0 and node1
1826 * can occur within a single pageblock, but not a node0 node1 node0
1827 * interleaving within a single pageblock. It is therefore sufficient to check
1828 * the first and last page of a pageblock and avoid checking each individual
1829 * page in a pageblock.
1830 */
1831struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1832 unsigned long end_pfn, struct zone *zone)
1833{
1834 struct page *start_page;
1835 struct page *end_page;
1836
1837 /* end_pfn is one past the range we are checking */
1838 end_pfn--;
1839
1840 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1841 return NULL;
1842
1843 start_page = pfn_to_online_page(start_pfn);
1844 if (!start_page)
1845 return NULL;
1846
1847 if (page_zone(start_page) != zone)
1848 return NULL;
1849
1850 end_page = pfn_to_page(end_pfn);
1851
1852 /* This gives a shorter code than deriving page_zone(end_page) */
1853 if (page_zone_id(start_page) != page_zone_id(end_page))
1854 return NULL;
1855
1856 return start_page;
1857}
1858
1859void set_zone_contiguous(struct zone *zone)
1860{
1861 unsigned long block_start_pfn = zone->zone_start_pfn;
1862 unsigned long block_end_pfn;
1863
1864 block_end_pfn = pageblock_end_pfn(block_start_pfn);
1865 for (; block_start_pfn < zone_end_pfn(zone);
1866 block_start_pfn = block_end_pfn,
1867 block_end_pfn += pageblock_nr_pages) {
1868
1869 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1870
1871 if (!__pageblock_pfn_to_page(block_start_pfn,
1872 block_end_pfn, zone))
1873 return;
1874 cond_resched();
1875 }
1876
1877 /* We confirm that there is no hole */
1878 zone->contiguous = true;
1879}
1880
1881void clear_zone_contiguous(struct zone *zone)
1882{
1883 zone->contiguous = false;
1884}
1885
1886#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1887static void __init deferred_free_range(unsigned long pfn,
1888 unsigned long nr_pages)
1889{
1890 struct page *page;
1891 unsigned long i;
1892
1893 if (!nr_pages)
1894 return;
1895
1896 page = pfn_to_page(pfn);
1897
1898 /* Free a large naturally-aligned chunk if possible */
1899 if (nr_pages == pageblock_nr_pages && pageblock_aligned(pfn)) {
1900 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1901 __free_pages_core(page, pageblock_order);
1902 return;
1903 }
1904
1905 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1906 if (pageblock_aligned(pfn))
1907 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1908 __free_pages_core(page, 0);
1909 }
1910}
1911
1912/* Completion tracking for deferred_init_memmap() threads */
1913static atomic_t pgdat_init_n_undone __initdata;
1914static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1915
1916static inline void __init pgdat_init_report_one_done(void)
1917{
1918 if (atomic_dec_and_test(&pgdat_init_n_undone))
1919 complete(&pgdat_init_all_done_comp);
1920}
1921
1922/*
1923 * Returns true if page needs to be initialized or freed to buddy allocator.
1924 *
1925 * We check if a current large page is valid by only checking the validity
1926 * of the head pfn.
1927 */
1928static inline bool __init deferred_pfn_valid(unsigned long pfn)
1929{
1930 if (pageblock_aligned(pfn) && !pfn_valid(pfn))
1931 return false;
1932 return true;
1933}
1934
1935/*
1936 * Free pages to buddy allocator. Try to free aligned pages in
1937 * pageblock_nr_pages sizes.
1938 */
1939static void __init deferred_free_pages(unsigned long pfn,
1940 unsigned long end_pfn)
1941{
1942 unsigned long nr_free = 0;
1943
1944 for (; pfn < end_pfn; pfn++) {
1945 if (!deferred_pfn_valid(pfn)) {
1946 deferred_free_range(pfn - nr_free, nr_free);
1947 nr_free = 0;
1948 } else if (pageblock_aligned(pfn)) {
1949 deferred_free_range(pfn - nr_free, nr_free);
1950 nr_free = 1;
1951 } else {
1952 nr_free++;
1953 }
1954 }
1955 /* Free the last block of pages to allocator */
1956 deferred_free_range(pfn - nr_free, nr_free);
1957}
1958
1959/*
1960 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1961 * by performing it only once every pageblock_nr_pages.
1962 * Return number of pages initialized.
1963 */
1964static unsigned long __init deferred_init_pages(struct zone *zone,
1965 unsigned long pfn,
1966 unsigned long end_pfn)
1967{
1968 int nid = zone_to_nid(zone);
1969 unsigned long nr_pages = 0;
1970 int zid = zone_idx(zone);
1971 struct page *page = NULL;
1972
1973 for (; pfn < end_pfn; pfn++) {
1974 if (!deferred_pfn_valid(pfn)) {
1975 page = NULL;
1976 continue;
1977 } else if (!page || pageblock_aligned(pfn)) {
1978 page = pfn_to_page(pfn);
1979 } else {
1980 page++;
1981 }
1982 __init_single_page(page, pfn, zid, nid);
1983 nr_pages++;
1984 }
1985 return (nr_pages);
1986}
1987
1988/*
1989 * This function is meant to pre-load the iterator for the zone init.
1990 * Specifically it walks through the ranges until we are caught up to the
1991 * first_init_pfn value and exits there. If we never encounter the value we
1992 * return false indicating there are no valid ranges left.
1993 */
1994static bool __init
1995deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1996 unsigned long *spfn, unsigned long *epfn,
1997 unsigned long first_init_pfn)
1998{
1999 u64 j;
2000
2001 /*
2002 * Start out by walking through the ranges in this zone that have
2003 * already been initialized. We don't need to do anything with them
2004 * so we just need to flush them out of the system.
2005 */
2006 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
2007 if (*epfn <= first_init_pfn)
2008 continue;
2009 if (*spfn < first_init_pfn)
2010 *spfn = first_init_pfn;
2011 *i = j;
2012 return true;
2013 }
2014
2015 return false;
2016}
2017
2018/*
2019 * Initialize and free pages. We do it in two loops: first we initialize
2020 * struct page, then free to buddy allocator, because while we are
2021 * freeing pages we can access pages that are ahead (computing buddy
2022 * page in __free_one_page()).
2023 *
2024 * In order to try and keep some memory in the cache we have the loop
2025 * broken along max page order boundaries. This way we will not cause
2026 * any issues with the buddy page computation.
2027 */
2028static unsigned long __init
2029deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
2030 unsigned long *end_pfn)
2031{
2032 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
2033 unsigned long spfn = *start_pfn, epfn = *end_pfn;
2034 unsigned long nr_pages = 0;
2035 u64 j = *i;
2036
2037 /* First we loop through and initialize the page values */
2038 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
2039 unsigned long t;
2040
2041 if (mo_pfn <= *start_pfn)
2042 break;
2043
2044 t = min(mo_pfn, *end_pfn);
2045 nr_pages += deferred_init_pages(zone, *start_pfn, t);
2046
2047 if (mo_pfn < *end_pfn) {
2048 *start_pfn = mo_pfn;
2049 break;
2050 }
2051 }
2052
2053 /* Reset values and now loop through freeing pages as needed */
2054 swap(j, *i);
2055
2056 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2057 unsigned long t;
2058
2059 if (mo_pfn <= spfn)
2060 break;
2061
2062 t = min(mo_pfn, epfn);
2063 deferred_free_pages(spfn, t);
2064
2065 if (mo_pfn <= epfn)
2066 break;
2067 }
2068
2069 return nr_pages;
2070}
2071
2072static void __init
2073deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2074 void *arg)
2075{
2076 unsigned long spfn, epfn;
2077 struct zone *zone = arg;
2078 u64 i;
2079
2080 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2081
2082 /*
2083 * Initialize and free pages in MAX_ORDER sized increments so that we
2084 * can avoid introducing any issues with the buddy allocator.
2085 */
2086 while (spfn < end_pfn) {
2087 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2088 cond_resched();
2089 }
2090}
2091
2092/* An arch may override for more concurrency. */
2093__weak int __init
2094deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2095{
2096 return 1;
2097}
2098
2099/* Initialise remaining memory on a node */
2100static int __init deferred_init_memmap(void *data)
2101{
2102 pg_data_t *pgdat = data;
2103 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2104 unsigned long spfn = 0, epfn = 0;
2105 unsigned long first_init_pfn, flags;
2106 unsigned long start = jiffies;
2107 struct zone *zone;
2108 int zid, max_threads;
2109 u64 i;
2110
2111 /* Bind memory initialisation thread to a local node if possible */
2112 if (!cpumask_empty(cpumask))
2113 set_cpus_allowed_ptr(current, cpumask);
2114
2115 pgdat_resize_lock(pgdat, &flags);
2116 first_init_pfn = pgdat->first_deferred_pfn;
2117 if (first_init_pfn == ULONG_MAX) {
2118 pgdat_resize_unlock(pgdat, &flags);
2119 pgdat_init_report_one_done();
2120 return 0;
2121 }
2122
2123 /* Sanity check boundaries */
2124 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2125 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2126 pgdat->first_deferred_pfn = ULONG_MAX;
2127
2128 /*
2129 * Once we unlock here, the zone cannot be grown anymore, thus if an
2130 * interrupt thread must allocate this early in boot, zone must be
2131 * pre-grown prior to start of deferred page initialization.
2132 */
2133 pgdat_resize_unlock(pgdat, &flags);
2134
2135 /* Only the highest zone is deferred so find it */
2136 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2137 zone = pgdat->node_zones + zid;
2138 if (first_init_pfn < zone_end_pfn(zone))
2139 break;
2140 }
2141
2142 /* If the zone is empty somebody else may have cleared out the zone */
2143 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2144 first_init_pfn))
2145 goto zone_empty;
2146
2147 max_threads = deferred_page_init_max_threads(cpumask);
2148
2149 while (spfn < epfn) {
2150 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2151 struct padata_mt_job job = {
2152 .thread_fn = deferred_init_memmap_chunk,
2153 .fn_arg = zone,
2154 .start = spfn,
2155 .size = epfn_align - spfn,
2156 .align = PAGES_PER_SECTION,
2157 .min_chunk = PAGES_PER_SECTION,
2158 .max_threads = max_threads,
2159 };
2160
2161 padata_do_multithreaded(&job);
2162 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2163 epfn_align);
2164 }
2165zone_empty:
2166 /* Sanity check that the next zone really is unpopulated */
2167 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2168
2169 pr_info("node %d deferred pages initialised in %ums\n",
2170 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2171
2172 pgdat_init_report_one_done();
2173 return 0;
2174}
2175
2176/*
2177 * If this zone has deferred pages, try to grow it by initializing enough
2178 * deferred pages to satisfy the allocation specified by order, rounded up to
2179 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2180 * of SECTION_SIZE bytes by initializing struct pages in increments of
2181 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2182 *
2183 * Return true when zone was grown, otherwise return false. We return true even
2184 * when we grow less than requested, to let the caller decide if there are
2185 * enough pages to satisfy the allocation.
2186 *
2187 * Note: We use noinline because this function is needed only during boot, and
2188 * it is called from a __ref function _deferred_grow_zone. This way we are
2189 * making sure that it is not inlined into permanent text section.
2190 */
2191static noinline bool __init
2192deferred_grow_zone(struct zone *zone, unsigned int order)
2193{
2194 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2195 pg_data_t *pgdat = zone->zone_pgdat;
2196 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2197 unsigned long spfn, epfn, flags;
2198 unsigned long nr_pages = 0;
2199 u64 i;
2200
2201 /* Only the last zone may have deferred pages */
2202 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2203 return false;
2204
2205 pgdat_resize_lock(pgdat, &flags);
2206
2207 /*
2208 * If someone grew this zone while we were waiting for spinlock, return
2209 * true, as there might be enough pages already.
2210 */
2211 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2212 pgdat_resize_unlock(pgdat, &flags);
2213 return true;
2214 }
2215
2216 /* If the zone is empty somebody else may have cleared out the zone */
2217 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2218 first_deferred_pfn)) {
2219 pgdat->first_deferred_pfn = ULONG_MAX;
2220 pgdat_resize_unlock(pgdat, &flags);
2221 /* Retry only once. */
2222 return first_deferred_pfn != ULONG_MAX;
2223 }
2224
2225 /*
2226 * Initialize and free pages in MAX_ORDER sized increments so
2227 * that we can avoid introducing any issues with the buddy
2228 * allocator.
2229 */
2230 while (spfn < epfn) {
2231 /* update our first deferred PFN for this section */
2232 first_deferred_pfn = spfn;
2233
2234 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2235 touch_nmi_watchdog();
2236
2237 /* We should only stop along section boundaries */
2238 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2239 continue;
2240
2241 /* If our quota has been met we can stop here */
2242 if (nr_pages >= nr_pages_needed)
2243 break;
2244 }
2245
2246 pgdat->first_deferred_pfn = spfn;
2247 pgdat_resize_unlock(pgdat, &flags);
2248
2249 return nr_pages > 0;
2250}
2251
2252/*
2253 * deferred_grow_zone() is __init, but it is called from
2254 * get_page_from_freelist() during early boot until deferred_pages permanently
2255 * disables this call. This is why we have refdata wrapper to avoid warning,
2256 * and to ensure that the function body gets unloaded.
2257 */
2258static bool __ref
2259_deferred_grow_zone(struct zone *zone, unsigned int order)
2260{
2261 return deferred_grow_zone(zone, order);
2262}
2263
2264#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2265
2266void __init page_alloc_init_late(void)
2267{
2268 struct zone *zone;
2269 int nid;
2270
2271#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2272
2273 /* There will be num_node_state(N_MEMORY) threads */
2274 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2275 for_each_node_state(nid, N_MEMORY) {
2276 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2277 }
2278
2279 /* Block until all are initialised */
2280 wait_for_completion(&pgdat_init_all_done_comp);
2281
2282 /*
2283 * We initialized the rest of the deferred pages. Permanently disable
2284 * on-demand struct page initialization.
2285 */
2286 static_branch_disable(&deferred_pages);
2287
2288 /* Reinit limits that are based on free pages after the kernel is up */
2289 files_maxfiles_init();
2290#endif
2291
2292 buffer_init();
2293
2294 /* Discard memblock private memory */
2295 memblock_discard();
2296
2297 for_each_node_state(nid, N_MEMORY)
2298 shuffle_free_memory(NODE_DATA(nid));
2299
2300 for_each_populated_zone(zone)
2301 set_zone_contiguous(zone);
2302}
2303
2304#ifdef CONFIG_CMA
2305/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2306void __init init_cma_reserved_pageblock(struct page *page)
2307{
2308 unsigned i = pageblock_nr_pages;
2309 struct page *p = page;
2310
2311 do {
2312 __ClearPageReserved(p);
2313 set_page_count(p, 0);
2314 } while (++p, --i);
2315
2316 set_pageblock_migratetype(page, MIGRATE_CMA);
2317 set_page_refcounted(page);
2318 __free_pages(page, pageblock_order);
2319
2320 adjust_managed_page_count(page, pageblock_nr_pages);
2321 page_zone(page)->cma_pages += pageblock_nr_pages;
2322}
2323#endif
2324
2325/*
2326 * The order of subdivision here is critical for the IO subsystem.
2327 * Please do not alter this order without good reasons and regression
2328 * testing. Specifically, as large blocks of memory are subdivided,
2329 * the order in which smaller blocks are delivered depends on the order
2330 * they're subdivided in this function. This is the primary factor
2331 * influencing the order in which pages are delivered to the IO
2332 * subsystem according to empirical testing, and this is also justified
2333 * by considering the behavior of a buddy system containing a single
2334 * large block of memory acted on by a series of small allocations.
2335 * This behavior is a critical factor in sglist merging's success.
2336 *
2337 * -- nyc
2338 */
2339static inline void expand(struct zone *zone, struct page *page,
2340 int low, int high, int migratetype)
2341{
2342 unsigned long size = 1 << high;
2343
2344 while (high > low) {
2345 high--;
2346 size >>= 1;
2347 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2348
2349 /*
2350 * Mark as guard pages (or page), that will allow to
2351 * merge back to allocator when buddy will be freed.
2352 * Corresponding page table entries will not be touched,
2353 * pages will stay not present in virtual address space
2354 */
2355 if (set_page_guard(zone, &page[size], high, migratetype))
2356 continue;
2357
2358 add_to_free_list(&page[size], zone, high, migratetype);
2359 set_buddy_order(&page[size], high);
2360 }
2361}
2362
2363static void check_new_page_bad(struct page *page)
2364{
2365 if (unlikely(page->flags & __PG_HWPOISON)) {
2366 /* Don't complain about hwpoisoned pages */
2367 page_mapcount_reset(page); /* remove PageBuddy */
2368 return;
2369 }
2370
2371 bad_page(page,
2372 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2373}
2374
2375/*
2376 * This page is about to be returned from the page allocator
2377 */
2378static inline int check_new_page(struct page *page)
2379{
2380 if (likely(page_expected_state(page,
2381 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2382 return 0;
2383
2384 check_new_page_bad(page);
2385 return 1;
2386}
2387
2388static bool check_new_pages(struct page *page, unsigned int order)
2389{
2390 int i;
2391 for (i = 0; i < (1 << order); i++) {
2392 struct page *p = page + i;
2393
2394 if (unlikely(check_new_page(p)))
2395 return true;
2396 }
2397
2398 return false;
2399}
2400
2401#ifdef CONFIG_DEBUG_VM
2402/*
2403 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2404 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2405 * also checked when pcp lists are refilled from the free lists.
2406 */
2407static inline bool check_pcp_refill(struct page *page, unsigned int order)
2408{
2409 if (debug_pagealloc_enabled_static())
2410 return check_new_pages(page, order);
2411 else
2412 return false;
2413}
2414
2415static inline bool check_new_pcp(struct page *page, unsigned int order)
2416{
2417 return check_new_pages(page, order);
2418}
2419#else
2420/*
2421 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2422 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2423 * enabled, they are also checked when being allocated from the pcp lists.
2424 */
2425static inline bool check_pcp_refill(struct page *page, unsigned int order)
2426{
2427 return check_new_pages(page, order);
2428}
2429static inline bool check_new_pcp(struct page *page, unsigned int order)
2430{
2431 if (debug_pagealloc_enabled_static())
2432 return check_new_pages(page, order);
2433 else
2434 return false;
2435}
2436#endif /* CONFIG_DEBUG_VM */
2437
2438static inline bool should_skip_kasan_unpoison(gfp_t flags)
2439{
2440 /* Don't skip if a software KASAN mode is enabled. */
2441 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
2442 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
2443 return false;
2444
2445 /* Skip, if hardware tag-based KASAN is not enabled. */
2446 if (!kasan_hw_tags_enabled())
2447 return true;
2448
2449 /*
2450 * With hardware tag-based KASAN enabled, skip if this has been
2451 * requested via __GFP_SKIP_KASAN_UNPOISON.
2452 */
2453 return flags & __GFP_SKIP_KASAN_UNPOISON;
2454}
2455
2456static inline bool should_skip_init(gfp_t flags)
2457{
2458 /* Don't skip, if hardware tag-based KASAN is not enabled. */
2459 if (!kasan_hw_tags_enabled())
2460 return false;
2461
2462 /* For hardware tag-based KASAN, skip if requested. */
2463 return (flags & __GFP_SKIP_ZERO);
2464}
2465
2466inline void post_alloc_hook(struct page *page, unsigned int order,
2467 gfp_t gfp_flags)
2468{
2469 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
2470 !should_skip_init(gfp_flags);
2471 bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
2472 int i;
2473
2474 set_page_private(page, 0);
2475 set_page_refcounted(page);
2476
2477 arch_alloc_page(page, order);
2478 debug_pagealloc_map_pages(page, 1 << order);
2479
2480 /*
2481 * Page unpoisoning must happen before memory initialization.
2482 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2483 * allocations and the page unpoisoning code will complain.
2484 */
2485 kernel_unpoison_pages(page, 1 << order);
2486
2487 /*
2488 * As memory initialization might be integrated into KASAN,
2489 * KASAN unpoisoning and memory initializion code must be
2490 * kept together to avoid discrepancies in behavior.
2491 */
2492
2493 /*
2494 * If memory tags should be zeroed (which happens only when memory
2495 * should be initialized as well).
2496 */
2497 if (init_tags) {
2498 /* Initialize both memory and tags. */
2499 for (i = 0; i != 1 << order; ++i)
2500 tag_clear_highpage(page + i);
2501
2502 /* Note that memory is already initialized by the loop above. */
2503 init = false;
2504 }
2505 if (!should_skip_kasan_unpoison(gfp_flags)) {
2506 /* Unpoison shadow memory or set memory tags. */
2507 kasan_unpoison_pages(page, order, init);
2508
2509 /* Note that memory is already initialized by KASAN. */
2510 if (kasan_has_integrated_init())
2511 init = false;
2512 } else {
2513 /* Ensure page_address() dereferencing does not fault. */
2514 for (i = 0; i != 1 << order; ++i)
2515 page_kasan_tag_reset(page + i);
2516 }
2517 /* If memory is still not initialized, do it now. */
2518 if (init)
2519 kernel_init_pages(page, 1 << order);
2520 /* Propagate __GFP_SKIP_KASAN_POISON to page flags. */
2521 if (kasan_hw_tags_enabled() && (gfp_flags & __GFP_SKIP_KASAN_POISON))
2522 SetPageSkipKASanPoison(page);
2523
2524 set_page_owner(page, order, gfp_flags);
2525 page_table_check_alloc(page, order);
2526}
2527
2528static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2529 unsigned int alloc_flags)
2530{
2531 post_alloc_hook(page, order, gfp_flags);
2532
2533 if (order && (gfp_flags & __GFP_COMP))
2534 prep_compound_page(page, order);
2535
2536 /*
2537 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2538 * allocate the page. The expectation is that the caller is taking
2539 * steps that will free more memory. The caller should avoid the page
2540 * being used for !PFMEMALLOC purposes.
2541 */
2542 if (alloc_flags & ALLOC_NO_WATERMARKS)
2543 set_page_pfmemalloc(page);
2544 else
2545 clear_page_pfmemalloc(page);
2546}
2547
2548/*
2549 * Go through the free lists for the given migratetype and remove
2550 * the smallest available page from the freelists
2551 */
2552static __always_inline
2553struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2554 int migratetype)
2555{
2556 unsigned int current_order;
2557 struct free_area *area;
2558 struct page *page;
2559
2560 /* Find a page of the appropriate size in the preferred list */
2561 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2562 area = &(zone->free_area[current_order]);
2563 page = get_page_from_free_area(area, migratetype);
2564 if (!page)
2565 continue;
2566 del_page_from_free_list(page, zone, current_order);
2567 expand(zone, page, order, current_order, migratetype);
2568 set_pcppage_migratetype(page, migratetype);
2569 trace_mm_page_alloc_zone_locked(page, order, migratetype,
2570 pcp_allowed_order(order) &&
2571 migratetype < MIGRATE_PCPTYPES);
2572 return page;
2573 }
2574
2575 return NULL;
2576}
2577
2578
2579/*
2580 * This array describes the order lists are fallen back to when
2581 * the free lists for the desirable migrate type are depleted
2582 *
2583 * The other migratetypes do not have fallbacks.
2584 */
2585static int fallbacks[MIGRATE_TYPES][3] = {
2586 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2587 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2588 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2589};
2590
2591#ifdef CONFIG_CMA
2592static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2593 unsigned int order)
2594{
2595 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2596}
2597#else
2598static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2599 unsigned int order) { return NULL; }
2600#endif
2601
2602/*
2603 * Move the free pages in a range to the freelist tail of the requested type.
2604 * Note that start_page and end_pages are not aligned on a pageblock
2605 * boundary. If alignment is required, use move_freepages_block()
2606 */
2607static int move_freepages(struct zone *zone,
2608 unsigned long start_pfn, unsigned long end_pfn,
2609 int migratetype, int *num_movable)
2610{
2611 struct page *page;
2612 unsigned long pfn;
2613 unsigned int order;
2614 int pages_moved = 0;
2615
2616 for (pfn = start_pfn; pfn <= end_pfn;) {
2617 page = pfn_to_page(pfn);
2618 if (!PageBuddy(page)) {
2619 /*
2620 * We assume that pages that could be isolated for
2621 * migration are movable. But we don't actually try
2622 * isolating, as that would be expensive.
2623 */
2624 if (num_movable &&
2625 (PageLRU(page) || __PageMovable(page)))
2626 (*num_movable)++;
2627 pfn++;
2628 continue;
2629 }
2630
2631 /* Make sure we are not inadvertently changing nodes */
2632 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2633 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2634
2635 order = buddy_order(page);
2636 move_to_free_list(page, zone, order, migratetype);
2637 pfn += 1 << order;
2638 pages_moved += 1 << order;
2639 }
2640
2641 return pages_moved;
2642}
2643
2644int move_freepages_block(struct zone *zone, struct page *page,
2645 int migratetype, int *num_movable)
2646{
2647 unsigned long start_pfn, end_pfn, pfn;
2648
2649 if (num_movable)
2650 *num_movable = 0;
2651
2652 pfn = page_to_pfn(page);
2653 start_pfn = pageblock_start_pfn(pfn);
2654 end_pfn = pageblock_end_pfn(pfn) - 1;
2655
2656 /* Do not cross zone boundaries */
2657 if (!zone_spans_pfn(zone, start_pfn))
2658 start_pfn = pfn;
2659 if (!zone_spans_pfn(zone, end_pfn))
2660 return 0;
2661
2662 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2663 num_movable);
2664}
2665
2666static void change_pageblock_range(struct page *pageblock_page,
2667 int start_order, int migratetype)
2668{
2669 int nr_pageblocks = 1 << (start_order - pageblock_order);
2670
2671 while (nr_pageblocks--) {
2672 set_pageblock_migratetype(pageblock_page, migratetype);
2673 pageblock_page += pageblock_nr_pages;
2674 }
2675}
2676
2677/*
2678 * When we are falling back to another migratetype during allocation, try to
2679 * steal extra free pages from the same pageblocks to satisfy further
2680 * allocations, instead of polluting multiple pageblocks.
2681 *
2682 * If we are stealing a relatively large buddy page, it is likely there will
2683 * be more free pages in the pageblock, so try to steal them all. For
2684 * reclaimable and unmovable allocations, we steal regardless of page size,
2685 * as fragmentation caused by those allocations polluting movable pageblocks
2686 * is worse than movable allocations stealing from unmovable and reclaimable
2687 * pageblocks.
2688 */
2689static bool can_steal_fallback(unsigned int order, int start_mt)
2690{
2691 /*
2692 * Leaving this order check is intended, although there is
2693 * relaxed order check in next check. The reason is that
2694 * we can actually steal whole pageblock if this condition met,
2695 * but, below check doesn't guarantee it and that is just heuristic
2696 * so could be changed anytime.
2697 */
2698 if (order >= pageblock_order)
2699 return true;
2700
2701 if (order >= pageblock_order / 2 ||
2702 start_mt == MIGRATE_RECLAIMABLE ||
2703 start_mt == MIGRATE_UNMOVABLE ||
2704 page_group_by_mobility_disabled)
2705 return true;
2706
2707 return false;
2708}
2709
2710static inline bool boost_watermark(struct zone *zone)
2711{
2712 unsigned long max_boost;
2713
2714 if (!watermark_boost_factor)
2715 return false;
2716 /*
2717 * Don't bother in zones that are unlikely to produce results.
2718 * On small machines, including kdump capture kernels running
2719 * in a small area, boosting the watermark can cause an out of
2720 * memory situation immediately.
2721 */
2722 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2723 return false;
2724
2725 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2726 watermark_boost_factor, 10000);
2727
2728 /*
2729 * high watermark may be uninitialised if fragmentation occurs
2730 * very early in boot so do not boost. We do not fall
2731 * through and boost by pageblock_nr_pages as failing
2732 * allocations that early means that reclaim is not going
2733 * to help and it may even be impossible to reclaim the
2734 * boosted watermark resulting in a hang.
2735 */
2736 if (!max_boost)
2737 return false;
2738
2739 max_boost = max(pageblock_nr_pages, max_boost);
2740
2741 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2742 max_boost);
2743
2744 return true;
2745}
2746
2747/*
2748 * This function implements actual steal behaviour. If order is large enough,
2749 * we can steal whole pageblock. If not, we first move freepages in this
2750 * pageblock to our migratetype and determine how many already-allocated pages
2751 * are there in the pageblock with a compatible migratetype. If at least half
2752 * of pages are free or compatible, we can change migratetype of the pageblock
2753 * itself, so pages freed in the future will be put on the correct free list.
2754 */
2755static void steal_suitable_fallback(struct zone *zone, struct page *page,
2756 unsigned int alloc_flags, int start_type, bool whole_block)
2757{
2758 unsigned int current_order = buddy_order(page);
2759 int free_pages, movable_pages, alike_pages;
2760 int old_block_type;
2761
2762 old_block_type = get_pageblock_migratetype(page);
2763
2764 /*
2765 * This can happen due to races and we want to prevent broken
2766 * highatomic accounting.
2767 */
2768 if (is_migrate_highatomic(old_block_type))
2769 goto single_page;
2770
2771 /* Take ownership for orders >= pageblock_order */
2772 if (current_order >= pageblock_order) {
2773 change_pageblock_range(page, current_order, start_type);
2774 goto single_page;
2775 }
2776
2777 /*
2778 * Boost watermarks to increase reclaim pressure to reduce the
2779 * likelihood of future fallbacks. Wake kswapd now as the node
2780 * may be balanced overall and kswapd will not wake naturally.
2781 */
2782 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2783 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2784
2785 /* We are not allowed to try stealing from the whole block */
2786 if (!whole_block)
2787 goto single_page;
2788
2789 free_pages = move_freepages_block(zone, page, start_type,
2790 &movable_pages);
2791 /*
2792 * Determine how many pages are compatible with our allocation.
2793 * For movable allocation, it's the number of movable pages which
2794 * we just obtained. For other types it's a bit more tricky.
2795 */
2796 if (start_type == MIGRATE_MOVABLE) {
2797 alike_pages = movable_pages;
2798 } else {
2799 /*
2800 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2801 * to MOVABLE pageblock, consider all non-movable pages as
2802 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2803 * vice versa, be conservative since we can't distinguish the
2804 * exact migratetype of non-movable pages.
2805 */
2806 if (old_block_type == MIGRATE_MOVABLE)
2807 alike_pages = pageblock_nr_pages
2808 - (free_pages + movable_pages);
2809 else
2810 alike_pages = 0;
2811 }
2812
2813 /* moving whole block can fail due to zone boundary conditions */
2814 if (!free_pages)
2815 goto single_page;
2816
2817 /*
2818 * If a sufficient number of pages in the block are either free or of
2819 * comparable migratability as our allocation, claim the whole block.
2820 */
2821 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2822 page_group_by_mobility_disabled)
2823 set_pageblock_migratetype(page, start_type);
2824
2825 return;
2826
2827single_page:
2828 move_to_free_list(page, zone, current_order, start_type);
2829}
2830
2831/*
2832 * Check whether there is a suitable fallback freepage with requested order.
2833 * If only_stealable is true, this function returns fallback_mt only if
2834 * we can steal other freepages all together. This would help to reduce
2835 * fragmentation due to mixed migratetype pages in one pageblock.
2836 */
2837int find_suitable_fallback(struct free_area *area, unsigned int order,
2838 int migratetype, bool only_stealable, bool *can_steal)
2839{
2840 int i;
2841 int fallback_mt;
2842
2843 if (area->nr_free == 0)
2844 return -1;
2845
2846 *can_steal = false;
2847 for (i = 0;; i++) {
2848 fallback_mt = fallbacks[migratetype][i];
2849 if (fallback_mt == MIGRATE_TYPES)
2850 break;
2851
2852 if (free_area_empty(area, fallback_mt))
2853 continue;
2854
2855 if (can_steal_fallback(order, migratetype))
2856 *can_steal = true;
2857
2858 if (!only_stealable)
2859 return fallback_mt;
2860
2861 if (*can_steal)
2862 return fallback_mt;
2863 }
2864
2865 return -1;
2866}
2867
2868/*
2869 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2870 * there are no empty page blocks that contain a page with a suitable order
2871 */
2872static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2873 unsigned int alloc_order)
2874{
2875 int mt;
2876 unsigned long max_managed, flags;
2877
2878 /*
2879 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2880 * Check is race-prone but harmless.
2881 */
2882 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2883 if (zone->nr_reserved_highatomic >= max_managed)
2884 return;
2885
2886 spin_lock_irqsave(&zone->lock, flags);
2887
2888 /* Recheck the nr_reserved_highatomic limit under the lock */
2889 if (zone->nr_reserved_highatomic >= max_managed)
2890 goto out_unlock;
2891
2892 /* Yoink! */
2893 mt = get_pageblock_migratetype(page);
2894 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2895 if (migratetype_is_mergeable(mt)) {
2896 zone->nr_reserved_highatomic += pageblock_nr_pages;
2897 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2898 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2899 }
2900
2901out_unlock:
2902 spin_unlock_irqrestore(&zone->lock, flags);
2903}
2904
2905/*
2906 * Used when an allocation is about to fail under memory pressure. This
2907 * potentially hurts the reliability of high-order allocations when under
2908 * intense memory pressure but failed atomic allocations should be easier
2909 * to recover from than an OOM.
2910 *
2911 * If @force is true, try to unreserve a pageblock even though highatomic
2912 * pageblock is exhausted.
2913 */
2914static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2915 bool force)
2916{
2917 struct zonelist *zonelist = ac->zonelist;
2918 unsigned long flags;
2919 struct zoneref *z;
2920 struct zone *zone;
2921 struct page *page;
2922 int order;
2923 bool ret;
2924
2925 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2926 ac->nodemask) {
2927 /*
2928 * Preserve at least one pageblock unless memory pressure
2929 * is really high.
2930 */
2931 if (!force && zone->nr_reserved_highatomic <=
2932 pageblock_nr_pages)
2933 continue;
2934
2935 spin_lock_irqsave(&zone->lock, flags);
2936 for (order = 0; order < MAX_ORDER; order++) {
2937 struct free_area *area = &(zone->free_area[order]);
2938
2939 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2940 if (!page)
2941 continue;
2942
2943 /*
2944 * In page freeing path, migratetype change is racy so
2945 * we can counter several free pages in a pageblock
2946 * in this loop although we changed the pageblock type
2947 * from highatomic to ac->migratetype. So we should
2948 * adjust the count once.
2949 */
2950 if (is_migrate_highatomic_page(page)) {
2951 /*
2952 * It should never happen but changes to
2953 * locking could inadvertently allow a per-cpu
2954 * drain to add pages to MIGRATE_HIGHATOMIC
2955 * while unreserving so be safe and watch for
2956 * underflows.
2957 */
2958 zone->nr_reserved_highatomic -= min(
2959 pageblock_nr_pages,
2960 zone->nr_reserved_highatomic);
2961 }
2962
2963 /*
2964 * Convert to ac->migratetype and avoid the normal
2965 * pageblock stealing heuristics. Minimally, the caller
2966 * is doing the work and needs the pages. More
2967 * importantly, if the block was always converted to
2968 * MIGRATE_UNMOVABLE or another type then the number
2969 * of pageblocks that cannot be completely freed
2970 * may increase.
2971 */
2972 set_pageblock_migratetype(page, ac->migratetype);
2973 ret = move_freepages_block(zone, page, ac->migratetype,
2974 NULL);
2975 if (ret) {
2976 spin_unlock_irqrestore(&zone->lock, flags);
2977 return ret;
2978 }
2979 }
2980 spin_unlock_irqrestore(&zone->lock, flags);
2981 }
2982
2983 return false;
2984}
2985
2986/*
2987 * Try finding a free buddy page on the fallback list and put it on the free
2988 * list of requested migratetype, possibly along with other pages from the same
2989 * block, depending on fragmentation avoidance heuristics. Returns true if
2990 * fallback was found so that __rmqueue_smallest() can grab it.
2991 *
2992 * The use of signed ints for order and current_order is a deliberate
2993 * deviation from the rest of this file, to make the for loop
2994 * condition simpler.
2995 */
2996static __always_inline bool
2997__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2998 unsigned int alloc_flags)
2999{
3000 struct free_area *area;
3001 int current_order;
3002 int min_order = order;
3003 struct page *page;
3004 int fallback_mt;
3005 bool can_steal;
3006
3007 /*
3008 * Do not steal pages from freelists belonging to other pageblocks
3009 * i.e. orders < pageblock_order. If there are no local zones free,
3010 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
3011 */
3012 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
3013 min_order = pageblock_order;
3014
3015 /*
3016 * Find the largest available free page in the other list. This roughly
3017 * approximates finding the pageblock with the most free pages, which
3018 * would be too costly to do exactly.
3019 */
3020 for (current_order = MAX_ORDER - 1; current_order >= min_order;
3021 --current_order) {
3022 area = &(zone->free_area[current_order]);
3023 fallback_mt = find_suitable_fallback(area, current_order,
3024 start_migratetype, false, &can_steal);
3025 if (fallback_mt == -1)
3026 continue;
3027
3028 /*
3029 * We cannot steal all free pages from the pageblock and the
3030 * requested migratetype is movable. In that case it's better to
3031 * steal and split the smallest available page instead of the
3032 * largest available page, because even if the next movable
3033 * allocation falls back into a different pageblock than this
3034 * one, it won't cause permanent fragmentation.
3035 */
3036 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
3037 && current_order > order)
3038 goto find_smallest;
3039
3040 goto do_steal;
3041 }
3042
3043 return false;
3044
3045find_smallest:
3046 for (current_order = order; current_order < MAX_ORDER;
3047 current_order++) {
3048 area = &(zone->free_area[current_order]);
3049 fallback_mt = find_suitable_fallback(area, current_order,
3050 start_migratetype, false, &can_steal);
3051 if (fallback_mt != -1)
3052 break;
3053 }
3054
3055 /*
3056 * This should not happen - we already found a suitable fallback
3057 * when looking for the largest page.
3058 */
3059 VM_BUG_ON(current_order == MAX_ORDER);
3060
3061do_steal:
3062 page = get_page_from_free_area(area, fallback_mt);
3063
3064 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
3065 can_steal);
3066
3067 trace_mm_page_alloc_extfrag(page, order, current_order,
3068 start_migratetype, fallback_mt);
3069
3070 return true;
3071
3072}
3073
3074/*
3075 * Do the hard work of removing an element from the buddy allocator.
3076 * Call me with the zone->lock already held.
3077 */
3078static __always_inline struct page *
3079__rmqueue(struct zone *zone, unsigned int order, int migratetype,
3080 unsigned int alloc_flags)
3081{
3082 struct page *page;
3083
3084 if (IS_ENABLED(CONFIG_CMA)) {
3085 /*
3086 * Balance movable allocations between regular and CMA areas by
3087 * allocating from CMA when over half of the zone's free memory
3088 * is in the CMA area.
3089 */
3090 if (alloc_flags & ALLOC_CMA &&
3091 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3092 zone_page_state(zone, NR_FREE_PAGES) / 2) {
3093 page = __rmqueue_cma_fallback(zone, order);
3094 if (page)
3095 return page;
3096 }
3097 }
3098retry:
3099 page = __rmqueue_smallest(zone, order, migratetype);
3100 if (unlikely(!page)) {
3101 if (alloc_flags & ALLOC_CMA)
3102 page = __rmqueue_cma_fallback(zone, order);
3103
3104 if (!page && __rmqueue_fallback(zone, order, migratetype,
3105 alloc_flags))
3106 goto retry;
3107 }
3108 return page;
3109}
3110
3111/*
3112 * Obtain a specified number of elements from the buddy allocator, all under
3113 * a single hold of the lock, for efficiency. Add them to the supplied list.
3114 * Returns the number of new pages which were placed at *list.
3115 */
3116static int rmqueue_bulk(struct zone *zone, unsigned int order,
3117 unsigned long count, struct list_head *list,
3118 int migratetype, unsigned int alloc_flags)
3119{
3120 unsigned long flags;
3121 int i, allocated = 0;
3122
3123 spin_lock_irqsave(&zone->lock, flags);
3124 for (i = 0; i < count; ++i) {
3125 struct page *page = __rmqueue(zone, order, migratetype,
3126 alloc_flags);
3127 if (unlikely(page == NULL))
3128 break;
3129
3130 if (unlikely(check_pcp_refill(page, order)))
3131 continue;
3132
3133 /*
3134 * Split buddy pages returned by expand() are received here in
3135 * physical page order. The page is added to the tail of
3136 * caller's list. From the callers perspective, the linked list
3137 * is ordered by page number under some conditions. This is
3138 * useful for IO devices that can forward direction from the
3139 * head, thus also in the physical page order. This is useful
3140 * for IO devices that can merge IO requests if the physical
3141 * pages are ordered properly.
3142 */
3143 list_add_tail(&page->pcp_list, list);
3144 allocated++;
3145 if (is_migrate_cma(get_pcppage_migratetype(page)))
3146 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3147 -(1 << order));
3148 }
3149
3150 /*
3151 * i pages were removed from the buddy list even if some leak due
3152 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3153 * on i. Do not confuse with 'allocated' which is the number of
3154 * pages added to the pcp list.
3155 */
3156 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3157 spin_unlock_irqrestore(&zone->lock, flags);
3158 return allocated;
3159}
3160
3161#ifdef CONFIG_NUMA
3162/*
3163 * Called from the vmstat counter updater to drain pagesets of this
3164 * currently executing processor on remote nodes after they have
3165 * expired.
3166 */
3167void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3168{
3169 int to_drain, batch;
3170
3171 batch = READ_ONCE(pcp->batch);
3172 to_drain = min(pcp->count, batch);
3173 if (to_drain > 0) {
3174 spin_lock(&pcp->lock);
3175 free_pcppages_bulk(zone, to_drain, pcp, 0);
3176 spin_unlock(&pcp->lock);
3177 }
3178}
3179#endif
3180
3181/*
3182 * Drain pcplists of the indicated processor and zone.
3183 */
3184static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3185{
3186 struct per_cpu_pages *pcp;
3187
3188 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3189 if (pcp->count) {
3190 spin_lock(&pcp->lock);
3191 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3192 spin_unlock(&pcp->lock);
3193 }
3194}
3195
3196/*
3197 * Drain pcplists of all zones on the indicated processor.
3198 */
3199static void drain_pages(unsigned int cpu)
3200{
3201 struct zone *zone;
3202
3203 for_each_populated_zone(zone) {
3204 drain_pages_zone(cpu, zone);
3205 }
3206}
3207
3208/*
3209 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3210 */
3211void drain_local_pages(struct zone *zone)
3212{
3213 int cpu = smp_processor_id();
3214
3215 if (zone)
3216 drain_pages_zone(cpu, zone);
3217 else
3218 drain_pages(cpu);
3219}
3220
3221/*
3222 * The implementation of drain_all_pages(), exposing an extra parameter to
3223 * drain on all cpus.
3224 *
3225 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3226 * not empty. The check for non-emptiness can however race with a free to
3227 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3228 * that need the guarantee that every CPU has drained can disable the
3229 * optimizing racy check.
3230 */
3231static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3232{
3233 int cpu;
3234
3235 /*
3236 * Allocate in the BSS so we won't require allocation in
3237 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3238 */
3239 static cpumask_t cpus_with_pcps;
3240
3241 /*
3242 * Do not drain if one is already in progress unless it's specific to
3243 * a zone. Such callers are primarily CMA and memory hotplug and need
3244 * the drain to be complete when the call returns.
3245 */
3246 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3247 if (!zone)
3248 return;
3249 mutex_lock(&pcpu_drain_mutex);
3250 }
3251
3252 /*
3253 * We don't care about racing with CPU hotplug event
3254 * as offline notification will cause the notified
3255 * cpu to drain that CPU pcps and on_each_cpu_mask
3256 * disables preemption as part of its processing
3257 */
3258 for_each_online_cpu(cpu) {
3259 struct per_cpu_pages *pcp;
3260 struct zone *z;
3261 bool has_pcps = false;
3262
3263 if (force_all_cpus) {
3264 /*
3265 * The pcp.count check is racy, some callers need a
3266 * guarantee that no cpu is missed.
3267 */
3268 has_pcps = true;
3269 } else if (zone) {
3270 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3271 if (pcp->count)
3272 has_pcps = true;
3273 } else {
3274 for_each_populated_zone(z) {
3275 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3276 if (pcp->count) {
3277 has_pcps = true;
3278 break;
3279 }
3280 }
3281 }
3282
3283 if (has_pcps)
3284 cpumask_set_cpu(cpu, &cpus_with_pcps);
3285 else
3286 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3287 }
3288
3289 for_each_cpu(cpu, &cpus_with_pcps) {
3290 if (zone)
3291 drain_pages_zone(cpu, zone);
3292 else
3293 drain_pages(cpu);
3294 }
3295
3296 mutex_unlock(&pcpu_drain_mutex);
3297}
3298
3299/*
3300 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3301 *
3302 * When zone parameter is non-NULL, spill just the single zone's pages.
3303 */
3304void drain_all_pages(struct zone *zone)
3305{
3306 __drain_all_pages(zone, false);
3307}
3308
3309#ifdef CONFIG_HIBERNATION
3310
3311/*
3312 * Touch the watchdog for every WD_PAGE_COUNT pages.
3313 */
3314#define WD_PAGE_COUNT (128*1024)
3315
3316void mark_free_pages(struct zone *zone)
3317{
3318 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3319 unsigned long flags;
3320 unsigned int order, t;
3321 struct page *page;
3322
3323 if (zone_is_empty(zone))
3324 return;
3325
3326 spin_lock_irqsave(&zone->lock, flags);
3327
3328 max_zone_pfn = zone_end_pfn(zone);
3329 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3330 if (pfn_valid(pfn)) {
3331 page = pfn_to_page(pfn);
3332
3333 if (!--page_count) {
3334 touch_nmi_watchdog();
3335 page_count = WD_PAGE_COUNT;
3336 }
3337
3338 if (page_zone(page) != zone)
3339 continue;
3340
3341 if (!swsusp_page_is_forbidden(page))
3342 swsusp_unset_page_free(page);
3343 }
3344
3345 for_each_migratetype_order(order, t) {
3346 list_for_each_entry(page,
3347 &zone->free_area[order].free_list[t], buddy_list) {
3348 unsigned long i;
3349
3350 pfn = page_to_pfn(page);
3351 for (i = 0; i < (1UL << order); i++) {
3352 if (!--page_count) {
3353 touch_nmi_watchdog();
3354 page_count = WD_PAGE_COUNT;
3355 }
3356 swsusp_set_page_free(pfn_to_page(pfn + i));
3357 }
3358 }
3359 }
3360 spin_unlock_irqrestore(&zone->lock, flags);
3361}
3362#endif /* CONFIG_PM */
3363
3364static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3365 unsigned int order)
3366{
3367 int migratetype;
3368
3369 if (!free_pcp_prepare(page, order))
3370 return false;
3371
3372 migratetype = get_pfnblock_migratetype(page, pfn);
3373 set_pcppage_migratetype(page, migratetype);
3374 return true;
3375}
3376
3377static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3378 bool free_high)
3379{
3380 int min_nr_free, max_nr_free;
3381
3382 /* Free everything if batch freeing high-order pages. */
3383 if (unlikely(free_high))
3384 return pcp->count;
3385
3386 /* Check for PCP disabled or boot pageset */
3387 if (unlikely(high < batch))
3388 return 1;
3389
3390 /* Leave at least pcp->batch pages on the list */
3391 min_nr_free = batch;
3392 max_nr_free = high - batch;
3393
3394 /*
3395 * Double the number of pages freed each time there is subsequent
3396 * freeing of pages without any allocation.
3397 */
3398 batch <<= pcp->free_factor;
3399 if (batch < max_nr_free)
3400 pcp->free_factor++;
3401 batch = clamp(batch, min_nr_free, max_nr_free);
3402
3403 return batch;
3404}
3405
3406static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3407 bool free_high)
3408{
3409 int high = READ_ONCE(pcp->high);
3410
3411 if (unlikely(!high || free_high))
3412 return 0;
3413
3414 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3415 return high;
3416
3417 /*
3418 * If reclaim is active, limit the number of pages that can be
3419 * stored on pcp lists
3420 */
3421 return min(READ_ONCE(pcp->batch) << 2, high);
3422}
3423
3424static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
3425 struct page *page, int migratetype,
3426 unsigned int order)
3427{
3428 int high;
3429 int pindex;
3430 bool free_high;
3431
3432 __count_vm_events(PGFREE, 1 << order);
3433 pindex = order_to_pindex(migratetype, order);
3434 list_add(&page->pcp_list, &pcp->lists[pindex]);
3435 pcp->count += 1 << order;
3436
3437 /*
3438 * As high-order pages other than THP's stored on PCP can contribute
3439 * to fragmentation, limit the number stored when PCP is heavily
3440 * freeing without allocation. The remainder after bulk freeing
3441 * stops will be drained from vmstat refresh context.
3442 */
3443 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3444
3445 high = nr_pcp_high(pcp, zone, free_high);
3446 if (pcp->count >= high) {
3447 int batch = READ_ONCE(pcp->batch);
3448
3449 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3450 }
3451}
3452
3453/*
3454 * Free a pcp page
3455 */
3456void free_unref_page(struct page *page, unsigned int order)
3457{
3458 unsigned long __maybe_unused UP_flags;
3459 struct per_cpu_pages *pcp;
3460 struct zone *zone;
3461 unsigned long pfn = page_to_pfn(page);
3462 int migratetype;
3463
3464 if (!free_unref_page_prepare(page, pfn, order))
3465 return;
3466
3467 /*
3468 * We only track unmovable, reclaimable and movable on pcp lists.
3469 * Place ISOLATE pages on the isolated list because they are being
3470 * offlined but treat HIGHATOMIC as movable pages so we can get those
3471 * areas back if necessary. Otherwise, we may have to free
3472 * excessively into the page allocator
3473 */
3474 migratetype = get_pcppage_migratetype(page);
3475 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3476 if (unlikely(is_migrate_isolate(migratetype))) {
3477 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3478 return;
3479 }
3480 migratetype = MIGRATE_MOVABLE;
3481 }
3482
3483 zone = page_zone(page);
3484 pcp_trylock_prepare(UP_flags);
3485 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3486 if (pcp) {
3487 free_unref_page_commit(zone, pcp, page, migratetype, order);
3488 pcp_spin_unlock(pcp);
3489 } else {
3490 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
3491 }
3492 pcp_trylock_finish(UP_flags);
3493}
3494
3495/*
3496 * Free a list of 0-order pages
3497 */
3498void free_unref_page_list(struct list_head *list)
3499{
3500 unsigned long __maybe_unused UP_flags;
3501 struct page *page, *next;
3502 struct per_cpu_pages *pcp = NULL;
3503 struct zone *locked_zone = NULL;
3504 int batch_count = 0;
3505 int migratetype;
3506
3507 /* Prepare pages for freeing */
3508 list_for_each_entry_safe(page, next, list, lru) {
3509 unsigned long pfn = page_to_pfn(page);
3510 if (!free_unref_page_prepare(page, pfn, 0)) {
3511 list_del(&page->lru);
3512 continue;
3513 }
3514
3515 /*
3516 * Free isolated pages directly to the allocator, see
3517 * comment in free_unref_page.
3518 */
3519 migratetype = get_pcppage_migratetype(page);
3520 if (unlikely(is_migrate_isolate(migratetype))) {
3521 list_del(&page->lru);
3522 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3523 continue;
3524 }
3525 }
3526
3527 list_for_each_entry_safe(page, next, list, lru) {
3528 struct zone *zone = page_zone(page);
3529
3530 list_del(&page->lru);
3531 migratetype = get_pcppage_migratetype(page);
3532
3533 /*
3534 * Either different zone requiring a different pcp lock or
3535 * excessive lock hold times when freeing a large list of
3536 * pages.
3537 */
3538 if (zone != locked_zone || batch_count == SWAP_CLUSTER_MAX) {
3539 if (pcp) {
3540 pcp_spin_unlock(pcp);
3541 pcp_trylock_finish(UP_flags);
3542 }
3543
3544 batch_count = 0;
3545
3546 /*
3547 * trylock is necessary as pages may be getting freed
3548 * from IRQ or SoftIRQ context after an IO completion.
3549 */
3550 pcp_trylock_prepare(UP_flags);
3551 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3552 if (unlikely(!pcp)) {
3553 pcp_trylock_finish(UP_flags);
3554 free_one_page(zone, page, page_to_pfn(page),
3555 0, migratetype, FPI_NONE);
3556 locked_zone = NULL;
3557 continue;
3558 }
3559 locked_zone = zone;
3560 }
3561
3562 /*
3563 * Non-isolated types over MIGRATE_PCPTYPES get added
3564 * to the MIGRATE_MOVABLE pcp list.
3565 */
3566 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3567 migratetype = MIGRATE_MOVABLE;
3568
3569 trace_mm_page_free_batched(page);
3570 free_unref_page_commit(zone, pcp, page, migratetype, 0);
3571 batch_count++;
3572 }
3573
3574 if (pcp) {
3575 pcp_spin_unlock(pcp);
3576 pcp_trylock_finish(UP_flags);
3577 }
3578}
3579
3580/*
3581 * split_page takes a non-compound higher-order page, and splits it into
3582 * n (1<<order) sub-pages: page[0..n]
3583 * Each sub-page must be freed individually.
3584 *
3585 * Note: this is probably too low level an operation for use in drivers.
3586 * Please consult with lkml before using this in your driver.
3587 */
3588void split_page(struct page *page, unsigned int order)
3589{
3590 int i;
3591
3592 VM_BUG_ON_PAGE(PageCompound(page), page);
3593 VM_BUG_ON_PAGE(!page_count(page), page);
3594
3595 for (i = 1; i < (1 << order); i++)
3596 set_page_refcounted(page + i);
3597 split_page_owner(page, 1 << order);
3598 split_page_memcg(page, 1 << order);
3599}
3600EXPORT_SYMBOL_GPL(split_page);
3601
3602int __isolate_free_page(struct page *page, unsigned int order)
3603{
3604 struct zone *zone = page_zone(page);
3605 int mt = get_pageblock_migratetype(page);
3606
3607 if (!is_migrate_isolate(mt)) {
3608 unsigned long watermark;
3609 /*
3610 * Obey watermarks as if the page was being allocated. We can
3611 * emulate a high-order watermark check with a raised order-0
3612 * watermark, because we already know our high-order page
3613 * exists.
3614 */
3615 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3616 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3617 return 0;
3618
3619 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3620 }
3621
3622 del_page_from_free_list(page, zone, order);
3623
3624 /*
3625 * Set the pageblock if the isolated page is at least half of a
3626 * pageblock
3627 */
3628 if (order >= pageblock_order - 1) {
3629 struct page *endpage = page + (1 << order) - 1;
3630 for (; page < endpage; page += pageblock_nr_pages) {
3631 int mt = get_pageblock_migratetype(page);
3632 /*
3633 * Only change normal pageblocks (i.e., they can merge
3634 * with others)
3635 */
3636 if (migratetype_is_mergeable(mt))
3637 set_pageblock_migratetype(page,
3638 MIGRATE_MOVABLE);
3639 }
3640 }
3641
3642 return 1UL << order;
3643}
3644
3645/**
3646 * __putback_isolated_page - Return a now-isolated page back where we got it
3647 * @page: Page that was isolated
3648 * @order: Order of the isolated page
3649 * @mt: The page's pageblock's migratetype
3650 *
3651 * This function is meant to return a page pulled from the free lists via
3652 * __isolate_free_page back to the free lists they were pulled from.
3653 */
3654void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3655{
3656 struct zone *zone = page_zone(page);
3657
3658 /* zone lock should be held when this function is called */
3659 lockdep_assert_held(&zone->lock);
3660
3661 /* Return isolated page to tail of freelist. */
3662 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3663 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3664}
3665
3666/*
3667 * Update NUMA hit/miss statistics
3668 */
3669static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3670 long nr_account)
3671{
3672#ifdef CONFIG_NUMA
3673 enum numa_stat_item local_stat = NUMA_LOCAL;
3674
3675 /* skip numa counters update if numa stats is disabled */
3676 if (!static_branch_likely(&vm_numa_stat_key))
3677 return;
3678
3679 if (zone_to_nid(z) != numa_node_id())
3680 local_stat = NUMA_OTHER;
3681
3682 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3683 __count_numa_events(z, NUMA_HIT, nr_account);
3684 else {
3685 __count_numa_events(z, NUMA_MISS, nr_account);
3686 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3687 }
3688 __count_numa_events(z, local_stat, nr_account);
3689#endif
3690}
3691
3692static __always_inline
3693struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
3694 unsigned int order, unsigned int alloc_flags,
3695 int migratetype)
3696{
3697 struct page *page;
3698 unsigned long flags;
3699
3700 do {
3701 page = NULL;
3702 spin_lock_irqsave(&zone->lock, flags);
3703 /*
3704 * order-0 request can reach here when the pcplist is skipped
3705 * due to non-CMA allocation context. HIGHATOMIC area is
3706 * reserved for high-order atomic allocation, so order-0
3707 * request should skip it.
3708 */
3709 if (order > 0 && alloc_flags & ALLOC_HARDER)
3710 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3711 if (!page) {
3712 page = __rmqueue(zone, order, migratetype, alloc_flags);
3713 if (!page) {
3714 spin_unlock_irqrestore(&zone->lock, flags);
3715 return NULL;
3716 }
3717 }
3718 __mod_zone_freepage_state(zone, -(1 << order),
3719 get_pcppage_migratetype(page));
3720 spin_unlock_irqrestore(&zone->lock, flags);
3721 } while (check_new_pages(page, order));
3722
3723 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3724 zone_statistics(preferred_zone, zone, 1);
3725
3726 return page;
3727}
3728
3729/* Remove page from the per-cpu list, caller must protect the list */
3730static inline
3731struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3732 int migratetype,
3733 unsigned int alloc_flags,
3734 struct per_cpu_pages *pcp,
3735 struct list_head *list)
3736{
3737 struct page *page;
3738
3739 do {
3740 if (list_empty(list)) {
3741 int batch = READ_ONCE(pcp->batch);
3742 int alloced;
3743
3744 /*
3745 * Scale batch relative to order if batch implies
3746 * free pages can be stored on the PCP. Batch can
3747 * be 1 for small zones or for boot pagesets which
3748 * should never store free pages as the pages may
3749 * belong to arbitrary zones.
3750 */
3751 if (batch > 1)
3752 batch = max(batch >> order, 2);
3753 alloced = rmqueue_bulk(zone, order,
3754 batch, list,
3755 migratetype, alloc_flags);
3756
3757 pcp->count += alloced << order;
3758 if (unlikely(list_empty(list)))
3759 return NULL;
3760 }
3761
3762 page = list_first_entry(list, struct page, pcp_list);
3763 list_del(&page->pcp_list);
3764 pcp->count -= 1 << order;
3765 } while (check_new_pcp(page, order));
3766
3767 return page;
3768}
3769
3770/* Lock and remove page from the per-cpu list */
3771static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3772 struct zone *zone, unsigned int order,
3773 int migratetype, unsigned int alloc_flags)
3774{
3775 struct per_cpu_pages *pcp;
3776 struct list_head *list;
3777 struct page *page;
3778 unsigned long __maybe_unused UP_flags;
3779
3780 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
3781 pcp_trylock_prepare(UP_flags);
3782 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3783 if (!pcp) {
3784 pcp_trylock_finish(UP_flags);
3785 return NULL;
3786 }
3787
3788 /*
3789 * On allocation, reduce the number of pages that are batch freed.
3790 * See nr_pcp_free() where free_factor is increased for subsequent
3791 * frees.
3792 */
3793 pcp->free_factor >>= 1;
3794 list = &pcp->lists[order_to_pindex(migratetype, order)];
3795 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3796 pcp_spin_unlock(pcp);
3797 pcp_trylock_finish(UP_flags);
3798 if (page) {
3799 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3800 zone_statistics(preferred_zone, zone, 1);
3801 }
3802 return page;
3803}
3804
3805/*
3806 * Allocate a page from the given zone.
3807 * Use pcplists for THP or "cheap" high-order allocations.
3808 */
3809
3810/*
3811 * Do not instrument rmqueue() with KMSAN. This function may call
3812 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
3813 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3814 * may call rmqueue() again, which will result in a deadlock.
3815 */
3816__no_sanitize_memory
3817static inline
3818struct page *rmqueue(struct zone *preferred_zone,
3819 struct zone *zone, unsigned int order,
3820 gfp_t gfp_flags, unsigned int alloc_flags,
3821 int migratetype)
3822{
3823 struct page *page;
3824
3825 /*
3826 * We most definitely don't want callers attempting to
3827 * allocate greater than order-1 page units with __GFP_NOFAIL.
3828 */
3829 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3830
3831 if (likely(pcp_allowed_order(order))) {
3832 /*
3833 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3834 * we need to skip it when CMA area isn't allowed.
3835 */
3836 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3837 migratetype != MIGRATE_MOVABLE) {
3838 page = rmqueue_pcplist(preferred_zone, zone, order,
3839 migratetype, alloc_flags);
3840 if (likely(page))
3841 goto out;
3842 }
3843 }
3844
3845 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3846 migratetype);
3847
3848out:
3849 /* Separate test+clear to avoid unnecessary atomics */
3850 if (unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3851 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3852 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3853 }
3854
3855 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3856 return page;
3857}
3858
3859#ifdef CONFIG_FAIL_PAGE_ALLOC
3860
3861static struct {
3862 struct fault_attr attr;
3863
3864 bool ignore_gfp_highmem;
3865 bool ignore_gfp_reclaim;
3866 u32 min_order;
3867} fail_page_alloc = {
3868 .attr = FAULT_ATTR_INITIALIZER,
3869 .ignore_gfp_reclaim = true,
3870 .ignore_gfp_highmem = true,
3871 .min_order = 1,
3872};
3873
3874static int __init setup_fail_page_alloc(char *str)
3875{
3876 return setup_fault_attr(&fail_page_alloc.attr, str);
3877}
3878__setup("fail_page_alloc=", setup_fail_page_alloc);
3879
3880static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3881{
3882 int flags = 0;
3883
3884 if (order < fail_page_alloc.min_order)
3885 return false;
3886 if (gfp_mask & __GFP_NOFAIL)
3887 return false;
3888 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3889 return false;
3890 if (fail_page_alloc.ignore_gfp_reclaim &&
3891 (gfp_mask & __GFP_DIRECT_RECLAIM))
3892 return false;
3893
3894 /* See comment in __should_failslab() */
3895 if (gfp_mask & __GFP_NOWARN)
3896 flags |= FAULT_NOWARN;
3897
3898 return should_fail_ex(&fail_page_alloc.attr, 1 << order, flags);
3899}
3900
3901#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3902
3903static int __init fail_page_alloc_debugfs(void)
3904{
3905 umode_t mode = S_IFREG | 0600;
3906 struct dentry *dir;
3907
3908 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3909 &fail_page_alloc.attr);
3910
3911 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3912 &fail_page_alloc.ignore_gfp_reclaim);
3913 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3914 &fail_page_alloc.ignore_gfp_highmem);
3915 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3916
3917 return 0;
3918}
3919
3920late_initcall(fail_page_alloc_debugfs);
3921
3922#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3923
3924#else /* CONFIG_FAIL_PAGE_ALLOC */
3925
3926static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3927{
3928 return false;
3929}
3930
3931#endif /* CONFIG_FAIL_PAGE_ALLOC */
3932
3933noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3934{
3935 return __should_fail_alloc_page(gfp_mask, order);
3936}
3937ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3938
3939static inline long __zone_watermark_unusable_free(struct zone *z,
3940 unsigned int order, unsigned int alloc_flags)
3941{
3942 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3943 long unusable_free = (1 << order) - 1;
3944
3945 /*
3946 * If the caller does not have rights to ALLOC_HARDER then subtract
3947 * the high-atomic reserves. This will over-estimate the size of the
3948 * atomic reserve but it avoids a search.
3949 */
3950 if (likely(!alloc_harder))
3951 unusable_free += z->nr_reserved_highatomic;
3952
3953#ifdef CONFIG_CMA
3954 /* If allocation can't use CMA areas don't use free CMA pages */
3955 if (!(alloc_flags & ALLOC_CMA))
3956 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3957#endif
3958
3959 return unusable_free;
3960}
3961
3962/*
3963 * Return true if free base pages are above 'mark'. For high-order checks it
3964 * will return true of the order-0 watermark is reached and there is at least
3965 * one free page of a suitable size. Checking now avoids taking the zone lock
3966 * to check in the allocation paths if no pages are free.
3967 */
3968bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3969 int highest_zoneidx, unsigned int alloc_flags,
3970 long free_pages)
3971{
3972 long min = mark;
3973 int o;
3974 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3975
3976 /* free_pages may go negative - that's OK */
3977 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3978
3979 if (alloc_flags & ALLOC_HIGH)
3980 min -= min / 2;
3981
3982 if (unlikely(alloc_harder)) {
3983 /*
3984 * OOM victims can try even harder than normal ALLOC_HARDER
3985 * users on the grounds that it's definitely going to be in
3986 * the exit path shortly and free memory. Any allocation it
3987 * makes during the free path will be small and short-lived.
3988 */
3989 if (alloc_flags & ALLOC_OOM)
3990 min -= min / 2;
3991 else
3992 min -= min / 4;
3993 }
3994
3995 /*
3996 * Check watermarks for an order-0 allocation request. If these
3997 * are not met, then a high-order request also cannot go ahead
3998 * even if a suitable page happened to be free.
3999 */
4000 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
4001 return false;
4002
4003 /* If this is an order-0 request then the watermark is fine */
4004 if (!order)
4005 return true;
4006
4007 /* For a high-order request, check at least one suitable page is free */
4008 for (o = order; o < MAX_ORDER; o++) {
4009 struct free_area *area = &z->free_area[o];
4010 int mt;
4011
4012 if (!area->nr_free)
4013 continue;
4014
4015 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
4016 if (!free_area_empty(area, mt))
4017 return true;
4018 }
4019
4020#ifdef CONFIG_CMA
4021 if ((alloc_flags & ALLOC_CMA) &&
4022 !free_area_empty(area, MIGRATE_CMA)) {
4023 return true;
4024 }
4025#endif
4026 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
4027 return true;
4028 }
4029 return false;
4030}
4031
4032bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
4033 int highest_zoneidx, unsigned int alloc_flags)
4034{
4035 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4036 zone_page_state(z, NR_FREE_PAGES));
4037}
4038
4039static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
4040 unsigned long mark, int highest_zoneidx,
4041 unsigned int alloc_flags, gfp_t gfp_mask)
4042{
4043 long free_pages;
4044
4045 free_pages = zone_page_state(z, NR_FREE_PAGES);
4046
4047 /*
4048 * Fast check for order-0 only. If this fails then the reserves
4049 * need to be calculated.
4050 */
4051 if (!order) {
4052 long usable_free;
4053 long reserved;
4054
4055 usable_free = free_pages;
4056 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
4057
4058 /* reserved may over estimate high-atomic reserves. */
4059 usable_free -= min(usable_free, reserved);
4060 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
4061 return true;
4062 }
4063
4064 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4065 free_pages))
4066 return true;
4067 /*
4068 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
4069 * when checking the min watermark. The min watermark is the
4070 * point where boosting is ignored so that kswapd is woken up
4071 * when below the low watermark.
4072 */
4073 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
4074 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
4075 mark = z->_watermark[WMARK_MIN];
4076 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
4077 alloc_flags, free_pages);
4078 }
4079
4080 return false;
4081}
4082
4083bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
4084 unsigned long mark, int highest_zoneidx)
4085{
4086 long free_pages = zone_page_state(z, NR_FREE_PAGES);
4087
4088 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
4089 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
4090
4091 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
4092 free_pages);
4093}
4094
4095#ifdef CONFIG_NUMA
4096int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
4097
4098static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4099{
4100 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
4101 node_reclaim_distance;
4102}
4103#else /* CONFIG_NUMA */
4104static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4105{
4106 return true;
4107}
4108#endif /* CONFIG_NUMA */
4109
4110/*
4111 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4112 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4113 * premature use of a lower zone may cause lowmem pressure problems that
4114 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4115 * probably too small. It only makes sense to spread allocations to avoid
4116 * fragmentation between the Normal and DMA32 zones.
4117 */
4118static inline unsigned int
4119alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4120{
4121 unsigned int alloc_flags;
4122
4123 /*
4124 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4125 * to save a branch.
4126 */
4127 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4128
4129#ifdef CONFIG_ZONE_DMA32
4130 if (!zone)
4131 return alloc_flags;
4132
4133 if (zone_idx(zone) != ZONE_NORMAL)
4134 return alloc_flags;
4135
4136 /*
4137 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4138 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4139 * on UMA that if Normal is populated then so is DMA32.
4140 */
4141 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4142 if (nr_online_nodes > 1 && !populated_zone(--zone))
4143 return alloc_flags;
4144
4145 alloc_flags |= ALLOC_NOFRAGMENT;
4146#endif /* CONFIG_ZONE_DMA32 */
4147 return alloc_flags;
4148}
4149
4150/* Must be called after current_gfp_context() which can change gfp_mask */
4151static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4152 unsigned int alloc_flags)
4153{
4154#ifdef CONFIG_CMA
4155 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4156 alloc_flags |= ALLOC_CMA;
4157#endif
4158 return alloc_flags;
4159}
4160
4161/*
4162 * get_page_from_freelist goes through the zonelist trying to allocate
4163 * a page.
4164 */
4165static struct page *
4166get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4167 const struct alloc_context *ac)
4168{
4169 struct zoneref *z;
4170 struct zone *zone;
4171 struct pglist_data *last_pgdat = NULL;
4172 bool last_pgdat_dirty_ok = false;
4173 bool no_fallback;
4174
4175retry:
4176 /*
4177 * Scan zonelist, looking for a zone with enough free.
4178 * See also __cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
4179 */
4180 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4181 z = ac->preferred_zoneref;
4182 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4183 ac->nodemask) {
4184 struct page *page;
4185 unsigned long mark;
4186
4187 if (cpusets_enabled() &&
4188 (alloc_flags & ALLOC_CPUSET) &&
4189 !__cpuset_zone_allowed(zone, gfp_mask))
4190 continue;
4191 /*
4192 * When allocating a page cache page for writing, we
4193 * want to get it from a node that is within its dirty
4194 * limit, such that no single node holds more than its
4195 * proportional share of globally allowed dirty pages.
4196 * The dirty limits take into account the node's
4197 * lowmem reserves and high watermark so that kswapd
4198 * should be able to balance it without having to
4199 * write pages from its LRU list.
4200 *
4201 * XXX: For now, allow allocations to potentially
4202 * exceed the per-node dirty limit in the slowpath
4203 * (spread_dirty_pages unset) before going into reclaim,
4204 * which is important when on a NUMA setup the allowed
4205 * nodes are together not big enough to reach the
4206 * global limit. The proper fix for these situations
4207 * will require awareness of nodes in the
4208 * dirty-throttling and the flusher threads.
4209 */
4210 if (ac->spread_dirty_pages) {
4211 if (last_pgdat != zone->zone_pgdat) {
4212 last_pgdat = zone->zone_pgdat;
4213 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
4214 }
4215
4216 if (!last_pgdat_dirty_ok)
4217 continue;
4218 }
4219
4220 if (no_fallback && nr_online_nodes > 1 &&
4221 zone != ac->preferred_zoneref->zone) {
4222 int local_nid;
4223
4224 /*
4225 * If moving to a remote node, retry but allow
4226 * fragmenting fallbacks. Locality is more important
4227 * than fragmentation avoidance.
4228 */
4229 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4230 if (zone_to_nid(zone) != local_nid) {
4231 alloc_flags &= ~ALLOC_NOFRAGMENT;
4232 goto retry;
4233 }
4234 }
4235
4236 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4237 if (!zone_watermark_fast(zone, order, mark,
4238 ac->highest_zoneidx, alloc_flags,
4239 gfp_mask)) {
4240 int ret;
4241
4242#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4243 /*
4244 * Watermark failed for this zone, but see if we can
4245 * grow this zone if it contains deferred pages.
4246 */
4247 if (static_branch_unlikely(&deferred_pages)) {
4248 if (_deferred_grow_zone(zone, order))
4249 goto try_this_zone;
4250 }
4251#endif
4252 /* Checked here to keep the fast path fast */
4253 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4254 if (alloc_flags & ALLOC_NO_WATERMARKS)
4255 goto try_this_zone;
4256
4257 if (!node_reclaim_enabled() ||
4258 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4259 continue;
4260
4261 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4262 switch (ret) {
4263 case NODE_RECLAIM_NOSCAN:
4264 /* did not scan */
4265 continue;
4266 case NODE_RECLAIM_FULL:
4267 /* scanned but unreclaimable */
4268 continue;
4269 default:
4270 /* did we reclaim enough */
4271 if (zone_watermark_ok(zone, order, mark,
4272 ac->highest_zoneidx, alloc_flags))
4273 goto try_this_zone;
4274
4275 continue;
4276 }
4277 }
4278
4279try_this_zone:
4280 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4281 gfp_mask, alloc_flags, ac->migratetype);
4282 if (page) {
4283 prep_new_page(page, order, gfp_mask, alloc_flags);
4284
4285 /*
4286 * If this is a high-order atomic allocation then check
4287 * if the pageblock should be reserved for the future
4288 */
4289 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4290 reserve_highatomic_pageblock(page, zone, order);
4291
4292 return page;
4293 } else {
4294#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4295 /* Try again if zone has deferred pages */
4296 if (static_branch_unlikely(&deferred_pages)) {
4297 if (_deferred_grow_zone(zone, order))
4298 goto try_this_zone;
4299 }
4300#endif
4301 }
4302 }
4303
4304 /*
4305 * It's possible on a UMA machine to get through all zones that are
4306 * fragmented. If avoiding fragmentation, reset and try again.
4307 */
4308 if (no_fallback) {
4309 alloc_flags &= ~ALLOC_NOFRAGMENT;
4310 goto retry;
4311 }
4312
4313 return NULL;
4314}
4315
4316static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4317{
4318 unsigned int filter = SHOW_MEM_FILTER_NODES;
4319
4320 /*
4321 * This documents exceptions given to allocations in certain
4322 * contexts that are allowed to allocate outside current's set
4323 * of allowed nodes.
4324 */
4325 if (!(gfp_mask & __GFP_NOMEMALLOC))
4326 if (tsk_is_oom_victim(current) ||
4327 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4328 filter &= ~SHOW_MEM_FILTER_NODES;
4329 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4330 filter &= ~SHOW_MEM_FILTER_NODES;
4331
4332 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
4333}
4334
4335void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4336{
4337 struct va_format vaf;
4338 va_list args;
4339 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4340
4341 if ((gfp_mask & __GFP_NOWARN) ||
4342 !__ratelimit(&nopage_rs) ||
4343 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4344 return;
4345
4346 va_start(args, fmt);
4347 vaf.fmt = fmt;
4348 vaf.va = &args;
4349 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4350 current->comm, &vaf, gfp_mask, &gfp_mask,
4351 nodemask_pr_args(nodemask));
4352 va_end(args);
4353
4354 cpuset_print_current_mems_allowed();
4355 pr_cont("\n");
4356 dump_stack();
4357 warn_alloc_show_mem(gfp_mask, nodemask);
4358}
4359
4360static inline struct page *
4361__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4362 unsigned int alloc_flags,
4363 const struct alloc_context *ac)
4364{
4365 struct page *page;
4366
4367 page = get_page_from_freelist(gfp_mask, order,
4368 alloc_flags|ALLOC_CPUSET, ac);
4369 /*
4370 * fallback to ignore cpuset restriction if our nodes
4371 * are depleted
4372 */
4373 if (!page)
4374 page = get_page_from_freelist(gfp_mask, order,
4375 alloc_flags, ac);
4376
4377 return page;
4378}
4379
4380static inline struct page *
4381__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4382 const struct alloc_context *ac, unsigned long *did_some_progress)
4383{
4384 struct oom_control oc = {
4385 .zonelist = ac->zonelist,
4386 .nodemask = ac->nodemask,
4387 .memcg = NULL,
4388 .gfp_mask = gfp_mask,
4389 .order = order,
4390 };
4391 struct page *page;
4392
4393 *did_some_progress = 0;
4394
4395 /*
4396 * Acquire the oom lock. If that fails, somebody else is
4397 * making progress for us.
4398 */
4399 if (!mutex_trylock(&oom_lock)) {
4400 *did_some_progress = 1;
4401 schedule_timeout_uninterruptible(1);
4402 return NULL;
4403 }
4404
4405 /*
4406 * Go through the zonelist yet one more time, keep very high watermark
4407 * here, this is only to catch a parallel oom killing, we must fail if
4408 * we're still under heavy pressure. But make sure that this reclaim
4409 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4410 * allocation which will never fail due to oom_lock already held.
4411 */
4412 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4413 ~__GFP_DIRECT_RECLAIM, order,
4414 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4415 if (page)
4416 goto out;
4417
4418 /* Coredumps can quickly deplete all memory reserves */
4419 if (current->flags & PF_DUMPCORE)
4420 goto out;
4421 /* The OOM killer will not help higher order allocs */
4422 if (order > PAGE_ALLOC_COSTLY_ORDER)
4423 goto out;
4424 /*
4425 * We have already exhausted all our reclaim opportunities without any
4426 * success so it is time to admit defeat. We will skip the OOM killer
4427 * because it is very likely that the caller has a more reasonable
4428 * fallback than shooting a random task.
4429 *
4430 * The OOM killer may not free memory on a specific node.
4431 */
4432 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4433 goto out;
4434 /* The OOM killer does not needlessly kill tasks for lowmem */
4435 if (ac->highest_zoneidx < ZONE_NORMAL)
4436 goto out;
4437 if (pm_suspended_storage())
4438 goto out;
4439 /*
4440 * XXX: GFP_NOFS allocations should rather fail than rely on
4441 * other request to make a forward progress.
4442 * We are in an unfortunate situation where out_of_memory cannot
4443 * do much for this context but let's try it to at least get
4444 * access to memory reserved if the current task is killed (see
4445 * out_of_memory). Once filesystems are ready to handle allocation
4446 * failures more gracefully we should just bail out here.
4447 */
4448
4449 /* Exhausted what can be done so it's blame time */
4450 if (out_of_memory(&oc) ||
4451 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4452 *did_some_progress = 1;
4453
4454 /*
4455 * Help non-failing allocations by giving them access to memory
4456 * reserves
4457 */
4458 if (gfp_mask & __GFP_NOFAIL)
4459 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4460 ALLOC_NO_WATERMARKS, ac);
4461 }
4462out:
4463 mutex_unlock(&oom_lock);
4464 return page;
4465}
4466
4467/*
4468 * Maximum number of compaction retries with a progress before OOM
4469 * killer is consider as the only way to move forward.
4470 */
4471#define MAX_COMPACT_RETRIES 16
4472
4473#ifdef CONFIG_COMPACTION
4474/* Try memory compaction for high-order allocations before reclaim */
4475static struct page *
4476__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4477 unsigned int alloc_flags, const struct alloc_context *ac,
4478 enum compact_priority prio, enum compact_result *compact_result)
4479{
4480 struct page *page = NULL;
4481 unsigned long pflags;
4482 unsigned int noreclaim_flag;
4483
4484 if (!order)
4485 return NULL;
4486
4487 psi_memstall_enter(&pflags);
4488 delayacct_compact_start();
4489 noreclaim_flag = memalloc_noreclaim_save();
4490
4491 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4492 prio, &page);
4493
4494 memalloc_noreclaim_restore(noreclaim_flag);
4495 psi_memstall_leave(&pflags);
4496 delayacct_compact_end();
4497
4498 if (*compact_result == COMPACT_SKIPPED)
4499 return NULL;
4500 /*
4501 * At least in one zone compaction wasn't deferred or skipped, so let's
4502 * count a compaction stall
4503 */
4504 count_vm_event(COMPACTSTALL);
4505
4506 /* Prep a captured page if available */
4507 if (page)
4508 prep_new_page(page, order, gfp_mask, alloc_flags);
4509
4510 /* Try get a page from the freelist if available */
4511 if (!page)
4512 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4513
4514 if (page) {
4515 struct zone *zone = page_zone(page);
4516
4517 zone->compact_blockskip_flush = false;
4518 compaction_defer_reset(zone, order, true);
4519 count_vm_event(COMPACTSUCCESS);
4520 return page;
4521 }
4522
4523 /*
4524 * It's bad if compaction run occurs and fails. The most likely reason
4525 * is that pages exist, but not enough to satisfy watermarks.
4526 */
4527 count_vm_event(COMPACTFAIL);
4528
4529 cond_resched();
4530
4531 return NULL;
4532}
4533
4534static inline bool
4535should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4536 enum compact_result compact_result,
4537 enum compact_priority *compact_priority,
4538 int *compaction_retries)
4539{
4540 int max_retries = MAX_COMPACT_RETRIES;
4541 int min_priority;
4542 bool ret = false;
4543 int retries = *compaction_retries;
4544 enum compact_priority priority = *compact_priority;
4545
4546 if (!order)
4547 return false;
4548
4549 if (fatal_signal_pending(current))
4550 return false;
4551
4552 if (compaction_made_progress(compact_result))
4553 (*compaction_retries)++;
4554
4555 /*
4556 * compaction considers all the zone as desperately out of memory
4557 * so it doesn't really make much sense to retry except when the
4558 * failure could be caused by insufficient priority
4559 */
4560 if (compaction_failed(compact_result))
4561 goto check_priority;
4562
4563 /*
4564 * compaction was skipped because there are not enough order-0 pages
4565 * to work with, so we retry only if it looks like reclaim can help.
4566 */
4567 if (compaction_needs_reclaim(compact_result)) {
4568 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4569 goto out;
4570 }
4571
4572 /*
4573 * make sure the compaction wasn't deferred or didn't bail out early
4574 * due to locks contention before we declare that we should give up.
4575 * But the next retry should use a higher priority if allowed, so
4576 * we don't just keep bailing out endlessly.
4577 */
4578 if (compaction_withdrawn(compact_result)) {
4579 goto check_priority;
4580 }
4581
4582 /*
4583 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4584 * costly ones because they are de facto nofail and invoke OOM
4585 * killer to move on while costly can fail and users are ready
4586 * to cope with that. 1/4 retries is rather arbitrary but we
4587 * would need much more detailed feedback from compaction to
4588 * make a better decision.
4589 */
4590 if (order > PAGE_ALLOC_COSTLY_ORDER)
4591 max_retries /= 4;
4592 if (*compaction_retries <= max_retries) {
4593 ret = true;
4594 goto out;
4595 }
4596
4597 /*
4598 * Make sure there are attempts at the highest priority if we exhausted
4599 * all retries or failed at the lower priorities.
4600 */
4601check_priority:
4602 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4603 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4604
4605 if (*compact_priority > min_priority) {
4606 (*compact_priority)--;
4607 *compaction_retries = 0;
4608 ret = true;
4609 }
4610out:
4611 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4612 return ret;
4613}
4614#else
4615static inline struct page *
4616__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4617 unsigned int alloc_flags, const struct alloc_context *ac,
4618 enum compact_priority prio, enum compact_result *compact_result)
4619{
4620 *compact_result = COMPACT_SKIPPED;
4621 return NULL;
4622}
4623
4624static inline bool
4625should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4626 enum compact_result compact_result,
4627 enum compact_priority *compact_priority,
4628 int *compaction_retries)
4629{
4630 struct zone *zone;
4631 struct zoneref *z;
4632
4633 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4634 return false;
4635
4636 /*
4637 * There are setups with compaction disabled which would prefer to loop
4638 * inside the allocator rather than hit the oom killer prematurely.
4639 * Let's give them a good hope and keep retrying while the order-0
4640 * watermarks are OK.
4641 */
4642 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4643 ac->highest_zoneidx, ac->nodemask) {
4644 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4645 ac->highest_zoneidx, alloc_flags))
4646 return true;
4647 }
4648 return false;
4649}
4650#endif /* CONFIG_COMPACTION */
4651
4652#ifdef CONFIG_LOCKDEP
4653static struct lockdep_map __fs_reclaim_map =
4654 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4655
4656static bool __need_reclaim(gfp_t gfp_mask)
4657{
4658 /* no reclaim without waiting on it */
4659 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4660 return false;
4661
4662 /* this guy won't enter reclaim */
4663 if (current->flags & PF_MEMALLOC)
4664 return false;
4665
4666 if (gfp_mask & __GFP_NOLOCKDEP)
4667 return false;
4668
4669 return true;
4670}
4671
4672void __fs_reclaim_acquire(unsigned long ip)
4673{
4674 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4675}
4676
4677void __fs_reclaim_release(unsigned long ip)
4678{
4679 lock_release(&__fs_reclaim_map, ip);
4680}
4681
4682void fs_reclaim_acquire(gfp_t gfp_mask)
4683{
4684 gfp_mask = current_gfp_context(gfp_mask);
4685
4686 if (__need_reclaim(gfp_mask)) {
4687 if (gfp_mask & __GFP_FS)
4688 __fs_reclaim_acquire(_RET_IP_);
4689
4690#ifdef CONFIG_MMU_NOTIFIER
4691 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4692 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4693#endif
4694
4695 }
4696}
4697EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4698
4699void fs_reclaim_release(gfp_t gfp_mask)
4700{
4701 gfp_mask = current_gfp_context(gfp_mask);
4702
4703 if (__need_reclaim(gfp_mask)) {
4704 if (gfp_mask & __GFP_FS)
4705 __fs_reclaim_release(_RET_IP_);
4706 }
4707}
4708EXPORT_SYMBOL_GPL(fs_reclaim_release);
4709#endif
4710
4711/*
4712 * Zonelists may change due to hotplug during allocation. Detect when zonelists
4713 * have been rebuilt so allocation retries. Reader side does not lock and
4714 * retries the allocation if zonelist changes. Writer side is protected by the
4715 * embedded spin_lock.
4716 */
4717static DEFINE_SEQLOCK(zonelist_update_seq);
4718
4719static unsigned int zonelist_iter_begin(void)
4720{
4721 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4722 return read_seqbegin(&zonelist_update_seq);
4723
4724 return 0;
4725}
4726
4727static unsigned int check_retry_zonelist(unsigned int seq)
4728{
4729 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4730 return read_seqretry(&zonelist_update_seq, seq);
4731
4732 return seq;
4733}
4734
4735/* Perform direct synchronous page reclaim */
4736static unsigned long
4737__perform_reclaim(gfp_t gfp_mask, unsigned int order,
4738 const struct alloc_context *ac)
4739{
4740 unsigned int noreclaim_flag;
4741 unsigned long progress;
4742
4743 cond_resched();
4744
4745 /* We now go into synchronous reclaim */
4746 cpuset_memory_pressure_bump();
4747 fs_reclaim_acquire(gfp_mask);
4748 noreclaim_flag = memalloc_noreclaim_save();
4749
4750 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4751 ac->nodemask);
4752
4753 memalloc_noreclaim_restore(noreclaim_flag);
4754 fs_reclaim_release(gfp_mask);
4755
4756 cond_resched();
4757
4758 return progress;
4759}
4760
4761/* The really slow allocator path where we enter direct reclaim */
4762static inline struct page *
4763__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4764 unsigned int alloc_flags, const struct alloc_context *ac,
4765 unsigned long *did_some_progress)
4766{
4767 struct page *page = NULL;
4768 unsigned long pflags;
4769 bool drained = false;
4770
4771 psi_memstall_enter(&pflags);
4772 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4773 if (unlikely(!(*did_some_progress)))
4774 goto out;
4775
4776retry:
4777 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4778
4779 /*
4780 * If an allocation failed after direct reclaim, it could be because
4781 * pages are pinned on the per-cpu lists or in high alloc reserves.
4782 * Shrink them and try again
4783 */
4784 if (!page && !drained) {
4785 unreserve_highatomic_pageblock(ac, false);
4786 drain_all_pages(NULL);
4787 drained = true;
4788 goto retry;
4789 }
4790out:
4791 psi_memstall_leave(&pflags);
4792
4793 return page;
4794}
4795
4796static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4797 const struct alloc_context *ac)
4798{
4799 struct zoneref *z;
4800 struct zone *zone;
4801 pg_data_t *last_pgdat = NULL;
4802 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4803
4804 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4805 ac->nodemask) {
4806 if (!managed_zone(zone))
4807 continue;
4808 if (last_pgdat != zone->zone_pgdat) {
4809 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4810 last_pgdat = zone->zone_pgdat;
4811 }
4812 }
4813}
4814
4815static inline unsigned int
4816gfp_to_alloc_flags(gfp_t gfp_mask)
4817{
4818 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4819
4820 /*
4821 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4822 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4823 * to save two branches.
4824 */
4825 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4826 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4827
4828 /*
4829 * The caller may dip into page reserves a bit more if the caller
4830 * cannot run direct reclaim, or if the caller has realtime scheduling
4831 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4832 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4833 */
4834 alloc_flags |= (__force int)
4835 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4836
4837 if (gfp_mask & __GFP_ATOMIC) {
4838 /*
4839 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4840 * if it can't schedule.
4841 */
4842 if (!(gfp_mask & __GFP_NOMEMALLOC))
4843 alloc_flags |= ALLOC_HARDER;
4844 /*
4845 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4846 * comment for __cpuset_node_allowed().
4847 */
4848 alloc_flags &= ~ALLOC_CPUSET;
4849 } else if (unlikely(rt_task(current)) && in_task())
4850 alloc_flags |= ALLOC_HARDER;
4851
4852 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4853
4854 return alloc_flags;
4855}
4856
4857static bool oom_reserves_allowed(struct task_struct *tsk)
4858{
4859 if (!tsk_is_oom_victim(tsk))
4860 return false;
4861
4862 /*
4863 * !MMU doesn't have oom reaper so give access to memory reserves
4864 * only to the thread with TIF_MEMDIE set
4865 */
4866 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4867 return false;
4868
4869 return true;
4870}
4871
4872/*
4873 * Distinguish requests which really need access to full memory
4874 * reserves from oom victims which can live with a portion of it
4875 */
4876static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4877{
4878 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4879 return 0;
4880 if (gfp_mask & __GFP_MEMALLOC)
4881 return ALLOC_NO_WATERMARKS;
4882 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4883 return ALLOC_NO_WATERMARKS;
4884 if (!in_interrupt()) {
4885 if (current->flags & PF_MEMALLOC)
4886 return ALLOC_NO_WATERMARKS;
4887 else if (oom_reserves_allowed(current))
4888 return ALLOC_OOM;
4889 }
4890
4891 return 0;
4892}
4893
4894bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4895{
4896 return !!__gfp_pfmemalloc_flags(gfp_mask);
4897}
4898
4899/*
4900 * Checks whether it makes sense to retry the reclaim to make a forward progress
4901 * for the given allocation request.
4902 *
4903 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4904 * without success, or when we couldn't even meet the watermark if we
4905 * reclaimed all remaining pages on the LRU lists.
4906 *
4907 * Returns true if a retry is viable or false to enter the oom path.
4908 */
4909static inline bool
4910should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4911 struct alloc_context *ac, int alloc_flags,
4912 bool did_some_progress, int *no_progress_loops)
4913{
4914 struct zone *zone;
4915 struct zoneref *z;
4916 bool ret = false;
4917
4918 /*
4919 * Costly allocations might have made a progress but this doesn't mean
4920 * their order will become available due to high fragmentation so
4921 * always increment the no progress counter for them
4922 */
4923 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4924 *no_progress_loops = 0;
4925 else
4926 (*no_progress_loops)++;
4927
4928 /*
4929 * Make sure we converge to OOM if we cannot make any progress
4930 * several times in the row.
4931 */
4932 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4933 /* Before OOM, exhaust highatomic_reserve */
4934 return unreserve_highatomic_pageblock(ac, true);
4935 }
4936
4937 /*
4938 * Keep reclaiming pages while there is a chance this will lead
4939 * somewhere. If none of the target zones can satisfy our allocation
4940 * request even if all reclaimable pages are considered then we are
4941 * screwed and have to go OOM.
4942 */
4943 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4944 ac->highest_zoneidx, ac->nodemask) {
4945 unsigned long available;
4946 unsigned long reclaimable;
4947 unsigned long min_wmark = min_wmark_pages(zone);
4948 bool wmark;
4949
4950 available = reclaimable = zone_reclaimable_pages(zone);
4951 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4952
4953 /*
4954 * Would the allocation succeed if we reclaimed all
4955 * reclaimable pages?
4956 */
4957 wmark = __zone_watermark_ok(zone, order, min_wmark,
4958 ac->highest_zoneidx, alloc_flags, available);
4959 trace_reclaim_retry_zone(z, order, reclaimable,
4960 available, min_wmark, *no_progress_loops, wmark);
4961 if (wmark) {
4962 ret = true;
4963 break;
4964 }
4965 }
4966
4967 /*
4968 * Memory allocation/reclaim might be called from a WQ context and the
4969 * current implementation of the WQ concurrency control doesn't
4970 * recognize that a particular WQ is congested if the worker thread is
4971 * looping without ever sleeping. Therefore we have to do a short sleep
4972 * here rather than calling cond_resched().
4973 */
4974 if (current->flags & PF_WQ_WORKER)
4975 schedule_timeout_uninterruptible(1);
4976 else
4977 cond_resched();
4978 return ret;
4979}
4980
4981static inline bool
4982check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4983{
4984 /*
4985 * It's possible that cpuset's mems_allowed and the nodemask from
4986 * mempolicy don't intersect. This should be normally dealt with by
4987 * policy_nodemask(), but it's possible to race with cpuset update in
4988 * such a way the check therein was true, and then it became false
4989 * before we got our cpuset_mems_cookie here.
4990 * This assumes that for all allocations, ac->nodemask can come only
4991 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4992 * when it does not intersect with the cpuset restrictions) or the
4993 * caller can deal with a violated nodemask.
4994 */
4995 if (cpusets_enabled() && ac->nodemask &&
4996 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4997 ac->nodemask = NULL;
4998 return true;
4999 }
5000
5001 /*
5002 * When updating a task's mems_allowed or mempolicy nodemask, it is
5003 * possible to race with parallel threads in such a way that our
5004 * allocation can fail while the mask is being updated. If we are about
5005 * to fail, check if the cpuset changed during allocation and if so,
5006 * retry.
5007 */
5008 if (read_mems_allowed_retry(cpuset_mems_cookie))
5009 return true;
5010
5011 return false;
5012}
5013
5014static inline struct page *
5015__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
5016 struct alloc_context *ac)
5017{
5018 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
5019 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
5020 struct page *page = NULL;
5021 unsigned int alloc_flags;
5022 unsigned long did_some_progress;
5023 enum compact_priority compact_priority;
5024 enum compact_result compact_result;
5025 int compaction_retries;
5026 int no_progress_loops;
5027 unsigned int cpuset_mems_cookie;
5028 unsigned int zonelist_iter_cookie;
5029 int reserve_flags;
5030
5031 /*
5032 * We also sanity check to catch abuse of atomic reserves being used by
5033 * callers that are not in atomic context.
5034 */
5035 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
5036 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
5037 gfp_mask &= ~__GFP_ATOMIC;
5038
5039restart:
5040 compaction_retries = 0;
5041 no_progress_loops = 0;
5042 compact_priority = DEF_COMPACT_PRIORITY;
5043 cpuset_mems_cookie = read_mems_allowed_begin();
5044 zonelist_iter_cookie = zonelist_iter_begin();
5045
5046 /*
5047 * The fast path uses conservative alloc_flags to succeed only until
5048 * kswapd needs to be woken up, and to avoid the cost of setting up
5049 * alloc_flags precisely. So we do that now.
5050 */
5051 alloc_flags = gfp_to_alloc_flags(gfp_mask);
5052
5053 /*
5054 * We need to recalculate the starting point for the zonelist iterator
5055 * because we might have used different nodemask in the fast path, or
5056 * there was a cpuset modification and we are retrying - otherwise we
5057 * could end up iterating over non-eligible zones endlessly.
5058 */
5059 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5060 ac->highest_zoneidx, ac->nodemask);
5061 if (!ac->preferred_zoneref->zone)
5062 goto nopage;
5063
5064 /*
5065 * Check for insane configurations where the cpuset doesn't contain
5066 * any suitable zone to satisfy the request - e.g. non-movable
5067 * GFP_HIGHUSER allocations from MOVABLE nodes only.
5068 */
5069 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
5070 struct zoneref *z = first_zones_zonelist(ac->zonelist,
5071 ac->highest_zoneidx,
5072 &cpuset_current_mems_allowed);
5073 if (!z->zone)
5074 goto nopage;
5075 }
5076
5077 if (alloc_flags & ALLOC_KSWAPD)
5078 wake_all_kswapds(order, gfp_mask, ac);
5079
5080 /*
5081 * The adjusted alloc_flags might result in immediate success, so try
5082 * that first
5083 */
5084 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5085 if (page)
5086 goto got_pg;
5087
5088 /*
5089 * For costly allocations, try direct compaction first, as it's likely
5090 * that we have enough base pages and don't need to reclaim. For non-
5091 * movable high-order allocations, do that as well, as compaction will
5092 * try prevent permanent fragmentation by migrating from blocks of the
5093 * same migratetype.
5094 * Don't try this for allocations that are allowed to ignore
5095 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
5096 */
5097 if (can_direct_reclaim &&
5098 (costly_order ||
5099 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
5100 && !gfp_pfmemalloc_allowed(gfp_mask)) {
5101 page = __alloc_pages_direct_compact(gfp_mask, order,
5102 alloc_flags, ac,
5103 INIT_COMPACT_PRIORITY,
5104 &compact_result);
5105 if (page)
5106 goto got_pg;
5107
5108 /*
5109 * Checks for costly allocations with __GFP_NORETRY, which
5110 * includes some THP page fault allocations
5111 */
5112 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
5113 /*
5114 * If allocating entire pageblock(s) and compaction
5115 * failed because all zones are below low watermarks
5116 * or is prohibited because it recently failed at this
5117 * order, fail immediately unless the allocator has
5118 * requested compaction and reclaim retry.
5119 *
5120 * Reclaim is
5121 * - potentially very expensive because zones are far
5122 * below their low watermarks or this is part of very
5123 * bursty high order allocations,
5124 * - not guaranteed to help because isolate_freepages()
5125 * may not iterate over freed pages as part of its
5126 * linear scan, and
5127 * - unlikely to make entire pageblocks free on its
5128 * own.
5129 */
5130 if (compact_result == COMPACT_SKIPPED ||
5131 compact_result == COMPACT_DEFERRED)
5132 goto nopage;
5133
5134 /*
5135 * Looks like reclaim/compaction is worth trying, but
5136 * sync compaction could be very expensive, so keep
5137 * using async compaction.
5138 */
5139 compact_priority = INIT_COMPACT_PRIORITY;
5140 }
5141 }
5142
5143retry:
5144 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5145 if (alloc_flags & ALLOC_KSWAPD)
5146 wake_all_kswapds(order, gfp_mask, ac);
5147
5148 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5149 if (reserve_flags)
5150 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
5151 (alloc_flags & ALLOC_KSWAPD);
5152
5153 /*
5154 * Reset the nodemask and zonelist iterators if memory policies can be
5155 * ignored. These allocations are high priority and system rather than
5156 * user oriented.
5157 */
5158 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5159 ac->nodemask = NULL;
5160 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5161 ac->highest_zoneidx, ac->nodemask);
5162 }
5163
5164 /* Attempt with potentially adjusted zonelist and alloc_flags */
5165 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5166 if (page)
5167 goto got_pg;
5168
5169 /* Caller is not willing to reclaim, we can't balance anything */
5170 if (!can_direct_reclaim)
5171 goto nopage;
5172
5173 /* Avoid recursion of direct reclaim */
5174 if (current->flags & PF_MEMALLOC)
5175 goto nopage;
5176
5177 /* Try direct reclaim and then allocating */
5178 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5179 &did_some_progress);
5180 if (page)
5181 goto got_pg;
5182
5183 /* Try direct compaction and then allocating */
5184 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5185 compact_priority, &compact_result);
5186 if (page)
5187 goto got_pg;
5188
5189 /* Do not loop if specifically requested */
5190 if (gfp_mask & __GFP_NORETRY)
5191 goto nopage;
5192
5193 /*
5194 * Do not retry costly high order allocations unless they are
5195 * __GFP_RETRY_MAYFAIL
5196 */
5197 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5198 goto nopage;
5199
5200 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5201 did_some_progress > 0, &no_progress_loops))
5202 goto retry;
5203
5204 /*
5205 * It doesn't make any sense to retry for the compaction if the order-0
5206 * reclaim is not able to make any progress because the current
5207 * implementation of the compaction depends on the sufficient amount
5208 * of free memory (see __compaction_suitable)
5209 */
5210 if (did_some_progress > 0 &&
5211 should_compact_retry(ac, order, alloc_flags,
5212 compact_result, &compact_priority,
5213 &compaction_retries))
5214 goto retry;
5215
5216
5217 /*
5218 * Deal with possible cpuset update races or zonelist updates to avoid
5219 * a unnecessary OOM kill.
5220 */
5221 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5222 check_retry_zonelist(zonelist_iter_cookie))
5223 goto restart;
5224
5225 /* Reclaim has failed us, start killing things */
5226 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5227 if (page)
5228 goto got_pg;
5229
5230 /* Avoid allocations with no watermarks from looping endlessly */
5231 if (tsk_is_oom_victim(current) &&
5232 (alloc_flags & ALLOC_OOM ||
5233 (gfp_mask & __GFP_NOMEMALLOC)))
5234 goto nopage;
5235
5236 /* Retry as long as the OOM killer is making progress */
5237 if (did_some_progress) {
5238 no_progress_loops = 0;
5239 goto retry;
5240 }
5241
5242nopage:
5243 /*
5244 * Deal with possible cpuset update races or zonelist updates to avoid
5245 * a unnecessary OOM kill.
5246 */
5247 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5248 check_retry_zonelist(zonelist_iter_cookie))
5249 goto restart;
5250
5251 /*
5252 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5253 * we always retry
5254 */
5255 if (gfp_mask & __GFP_NOFAIL) {
5256 /*
5257 * All existing users of the __GFP_NOFAIL are blockable, so warn
5258 * of any new users that actually require GFP_NOWAIT
5259 */
5260 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
5261 goto fail;
5262
5263 /*
5264 * PF_MEMALLOC request from this context is rather bizarre
5265 * because we cannot reclaim anything and only can loop waiting
5266 * for somebody to do a work for us
5267 */
5268 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
5269
5270 /*
5271 * non failing costly orders are a hard requirement which we
5272 * are not prepared for much so let's warn about these users
5273 * so that we can identify them and convert them to something
5274 * else.
5275 */
5276 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
5277
5278 /*
5279 * Help non-failing allocations by giving them access to memory
5280 * reserves but do not use ALLOC_NO_WATERMARKS because this
5281 * could deplete whole memory reserves which would just make
5282 * the situation worse
5283 */
5284 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5285 if (page)
5286 goto got_pg;
5287
5288 cond_resched();
5289 goto retry;
5290 }
5291fail:
5292 warn_alloc(gfp_mask, ac->nodemask,
5293 "page allocation failure: order:%u", order);
5294got_pg:
5295 return page;
5296}
5297
5298static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5299 int preferred_nid, nodemask_t *nodemask,
5300 struct alloc_context *ac, gfp_t *alloc_gfp,
5301 unsigned int *alloc_flags)
5302{
5303 ac->highest_zoneidx = gfp_zone(gfp_mask);
5304 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5305 ac->nodemask = nodemask;
5306 ac->migratetype = gfp_migratetype(gfp_mask);
5307
5308 if (cpusets_enabled()) {
5309 *alloc_gfp |= __GFP_HARDWALL;
5310 /*
5311 * When we are in the interrupt context, it is irrelevant
5312 * to the current task context. It means that any node ok.
5313 */
5314 if (in_task() && !ac->nodemask)
5315 ac->nodemask = &cpuset_current_mems_allowed;
5316 else
5317 *alloc_flags |= ALLOC_CPUSET;
5318 }
5319
5320 might_alloc(gfp_mask);
5321
5322 if (should_fail_alloc_page(gfp_mask, order))
5323 return false;
5324
5325 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5326
5327 /* Dirty zone balancing only done in the fast path */
5328 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5329
5330 /*
5331 * The preferred zone is used for statistics but crucially it is
5332 * also used as the starting point for the zonelist iterator. It
5333 * may get reset for allocations that ignore memory policies.
5334 */
5335 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5336 ac->highest_zoneidx, ac->nodemask);
5337
5338 return true;
5339}
5340
5341/*
5342 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5343 * @gfp: GFP flags for the allocation
5344 * @preferred_nid: The preferred NUMA node ID to allocate from
5345 * @nodemask: Set of nodes to allocate from, may be NULL
5346 * @nr_pages: The number of pages desired on the list or array
5347 * @page_list: Optional list to store the allocated pages
5348 * @page_array: Optional array to store the pages
5349 *
5350 * This is a batched version of the page allocator that attempts to
5351 * allocate nr_pages quickly. Pages are added to page_list if page_list
5352 * is not NULL, otherwise it is assumed that the page_array is valid.
5353 *
5354 * For lists, nr_pages is the number of pages that should be allocated.
5355 *
5356 * For arrays, only NULL elements are populated with pages and nr_pages
5357 * is the maximum number of pages that will be stored in the array.
5358 *
5359 * Returns the number of pages on the list or array.
5360 */
5361unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5362 nodemask_t *nodemask, int nr_pages,
5363 struct list_head *page_list,
5364 struct page **page_array)
5365{
5366 struct page *page;
5367 unsigned long __maybe_unused UP_flags;
5368 struct zone *zone;
5369 struct zoneref *z;
5370 struct per_cpu_pages *pcp;
5371 struct list_head *pcp_list;
5372 struct alloc_context ac;
5373 gfp_t alloc_gfp;
5374 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5375 int nr_populated = 0, nr_account = 0;
5376
5377 /*
5378 * Skip populated array elements to determine if any pages need
5379 * to be allocated before disabling IRQs.
5380 */
5381 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5382 nr_populated++;
5383
5384 /* No pages requested? */
5385 if (unlikely(nr_pages <= 0))
5386 goto out;
5387
5388 /* Already populated array? */
5389 if (unlikely(page_array && nr_pages - nr_populated == 0))
5390 goto out;
5391
5392 /* Bulk allocator does not support memcg accounting. */
5393 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5394 goto failed;
5395
5396 /* Use the single page allocator for one page. */
5397 if (nr_pages - nr_populated == 1)
5398 goto failed;
5399
5400#ifdef CONFIG_PAGE_OWNER
5401 /*
5402 * PAGE_OWNER may recurse into the allocator to allocate space to
5403 * save the stack with pagesets.lock held. Releasing/reacquiring
5404 * removes much of the performance benefit of bulk allocation so
5405 * force the caller to allocate one page at a time as it'll have
5406 * similar performance to added complexity to the bulk allocator.
5407 */
5408 if (static_branch_unlikely(&page_owner_inited))
5409 goto failed;
5410#endif
5411
5412 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5413 gfp &= gfp_allowed_mask;
5414 alloc_gfp = gfp;
5415 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5416 goto out;
5417 gfp = alloc_gfp;
5418
5419 /* Find an allowed local zone that meets the low watermark. */
5420 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5421 unsigned long mark;
5422
5423 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5424 !__cpuset_zone_allowed(zone, gfp)) {
5425 continue;
5426 }
5427
5428 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5429 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5430 goto failed;
5431 }
5432
5433 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5434 if (zone_watermark_fast(zone, 0, mark,
5435 zonelist_zone_idx(ac.preferred_zoneref),
5436 alloc_flags, gfp)) {
5437 break;
5438 }
5439 }
5440
5441 /*
5442 * If there are no allowed local zones that meets the watermarks then
5443 * try to allocate a single page and reclaim if necessary.
5444 */
5445 if (unlikely(!zone))
5446 goto failed;
5447
5448 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
5449 pcp_trylock_prepare(UP_flags);
5450 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
5451 if (!pcp)
5452 goto failed_irq;
5453
5454 /* Attempt the batch allocation */
5455 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5456 while (nr_populated < nr_pages) {
5457
5458 /* Skip existing pages */
5459 if (page_array && page_array[nr_populated]) {
5460 nr_populated++;
5461 continue;
5462 }
5463
5464 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5465 pcp, pcp_list);
5466 if (unlikely(!page)) {
5467 /* Try and allocate at least one page */
5468 if (!nr_account) {
5469 pcp_spin_unlock(pcp);
5470 goto failed_irq;
5471 }
5472 break;
5473 }
5474 nr_account++;
5475
5476 prep_new_page(page, 0, gfp, 0);
5477 if (page_list)
5478 list_add(&page->lru, page_list);
5479 else
5480 page_array[nr_populated] = page;
5481 nr_populated++;
5482 }
5483
5484 pcp_spin_unlock(pcp);
5485 pcp_trylock_finish(UP_flags);
5486
5487 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5488 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5489
5490out:
5491 return nr_populated;
5492
5493failed_irq:
5494 pcp_trylock_finish(UP_flags);
5495
5496failed:
5497 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5498 if (page) {
5499 if (page_list)
5500 list_add(&page->lru, page_list);
5501 else
5502 page_array[nr_populated] = page;
5503 nr_populated++;
5504 }
5505
5506 goto out;
5507}
5508EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5509
5510/*
5511 * This is the 'heart' of the zoned buddy allocator.
5512 */
5513struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5514 nodemask_t *nodemask)
5515{
5516 struct page *page;
5517 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5518 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5519 struct alloc_context ac = { };
5520
5521 /*
5522 * There are several places where we assume that the order value is sane
5523 * so bail out early if the request is out of bound.
5524 */
5525 if (WARN_ON_ONCE_GFP(order >= MAX_ORDER, gfp))
5526 return NULL;
5527
5528 gfp &= gfp_allowed_mask;
5529 /*
5530 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5531 * resp. GFP_NOIO which has to be inherited for all allocation requests
5532 * from a particular context which has been marked by
5533 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5534 * movable zones are not used during allocation.
5535 */
5536 gfp = current_gfp_context(gfp);
5537 alloc_gfp = gfp;
5538 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5539 &alloc_gfp, &alloc_flags))
5540 return NULL;
5541
5542 /*
5543 * Forbid the first pass from falling back to types that fragment
5544 * memory until all local zones are considered.
5545 */
5546 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5547
5548 /* First allocation attempt */
5549 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5550 if (likely(page))
5551 goto out;
5552
5553 alloc_gfp = gfp;
5554 ac.spread_dirty_pages = false;
5555
5556 /*
5557 * Restore the original nodemask if it was potentially replaced with
5558 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5559 */
5560 ac.nodemask = nodemask;
5561
5562 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5563
5564out:
5565 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5566 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5567 __free_pages(page, order);
5568 page = NULL;
5569 }
5570
5571 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5572 kmsan_alloc_page(page, order, alloc_gfp);
5573
5574 return page;
5575}
5576EXPORT_SYMBOL(__alloc_pages);
5577
5578struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5579 nodemask_t *nodemask)
5580{
5581 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5582 preferred_nid, nodemask);
5583
5584 if (page && order > 1)
5585 prep_transhuge_page(page);
5586 return (struct folio *)page;
5587}
5588EXPORT_SYMBOL(__folio_alloc);
5589
5590/*
5591 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5592 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5593 * you need to access high mem.
5594 */
5595unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5596{
5597 struct page *page;
5598
5599 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5600 if (!page)
5601 return 0;
5602 return (unsigned long) page_address(page);
5603}
5604EXPORT_SYMBOL(__get_free_pages);
5605
5606unsigned long get_zeroed_page(gfp_t gfp_mask)
5607{
5608 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5609}
5610EXPORT_SYMBOL(get_zeroed_page);
5611
5612/**
5613 * __free_pages - Free pages allocated with alloc_pages().
5614 * @page: The page pointer returned from alloc_pages().
5615 * @order: The order of the allocation.
5616 *
5617 * This function can free multi-page allocations that are not compound
5618 * pages. It does not check that the @order passed in matches that of
5619 * the allocation, so it is easy to leak memory. Freeing more memory
5620 * than was allocated will probably emit a warning.
5621 *
5622 * If the last reference to this page is speculative, it will be released
5623 * by put_page() which only frees the first page of a non-compound
5624 * allocation. To prevent the remaining pages from being leaked, we free
5625 * the subsequent pages here. If you want to use the page's reference
5626 * count to decide when to free the allocation, you should allocate a
5627 * compound page, and use put_page() instead of __free_pages().
5628 *
5629 * Context: May be called in interrupt context or while holding a normal
5630 * spinlock, but not in NMI context or while holding a raw spinlock.
5631 */
5632void __free_pages(struct page *page, unsigned int order)
5633{
5634 /* get PageHead before we drop reference */
5635 int head = PageHead(page);
5636
5637 if (put_page_testzero(page))
5638 free_the_page(page, order);
5639 else if (!head)
5640 while (order-- > 0)
5641 free_the_page(page + (1 << order), order);
5642}
5643EXPORT_SYMBOL(__free_pages);
5644
5645void free_pages(unsigned long addr, unsigned int order)
5646{
5647 if (addr != 0) {
5648 VM_BUG_ON(!virt_addr_valid((void *)addr));
5649 __free_pages(virt_to_page((void *)addr), order);
5650 }
5651}
5652
5653EXPORT_SYMBOL(free_pages);
5654
5655/*
5656 * Page Fragment:
5657 * An arbitrary-length arbitrary-offset area of memory which resides
5658 * within a 0 or higher order page. Multiple fragments within that page
5659 * are individually refcounted, in the page's reference counter.
5660 *
5661 * The page_frag functions below provide a simple allocation framework for
5662 * page fragments. This is used by the network stack and network device
5663 * drivers to provide a backing region of memory for use as either an
5664 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5665 */
5666static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5667 gfp_t gfp_mask)
5668{
5669 struct page *page = NULL;
5670 gfp_t gfp = gfp_mask;
5671
5672#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5673 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5674 __GFP_NOMEMALLOC;
5675 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5676 PAGE_FRAG_CACHE_MAX_ORDER);
5677 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5678#endif
5679 if (unlikely(!page))
5680 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5681
5682 nc->va = page ? page_address(page) : NULL;
5683
5684 return page;
5685}
5686
5687void __page_frag_cache_drain(struct page *page, unsigned int count)
5688{
5689 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5690
5691 if (page_ref_sub_and_test(page, count))
5692 free_the_page(page, compound_order(page));
5693}
5694EXPORT_SYMBOL(__page_frag_cache_drain);
5695
5696void *page_frag_alloc_align(struct page_frag_cache *nc,
5697 unsigned int fragsz, gfp_t gfp_mask,
5698 unsigned int align_mask)
5699{
5700 unsigned int size = PAGE_SIZE;
5701 struct page *page;
5702 int offset;
5703
5704 if (unlikely(!nc->va)) {
5705refill:
5706 page = __page_frag_cache_refill(nc, gfp_mask);
5707 if (!page)
5708 return NULL;
5709
5710#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5711 /* if size can vary use size else just use PAGE_SIZE */
5712 size = nc->size;
5713#endif
5714 /* Even if we own the page, we do not use atomic_set().
5715 * This would break get_page_unless_zero() users.
5716 */
5717 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5718
5719 /* reset page count bias and offset to start of new frag */
5720 nc->pfmemalloc = page_is_pfmemalloc(page);
5721 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5722 nc->offset = size;
5723 }
5724
5725 offset = nc->offset - fragsz;
5726 if (unlikely(offset < 0)) {
5727 page = virt_to_page(nc->va);
5728
5729 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5730 goto refill;
5731
5732 if (unlikely(nc->pfmemalloc)) {
5733 free_the_page(page, compound_order(page));
5734 goto refill;
5735 }
5736
5737#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5738 /* if size can vary use size else just use PAGE_SIZE */
5739 size = nc->size;
5740#endif
5741 /* OK, page count is 0, we can safely set it */
5742 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5743
5744 /* reset page count bias and offset to start of new frag */
5745 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5746 offset = size - fragsz;
5747 if (unlikely(offset < 0)) {
5748 /*
5749 * The caller is trying to allocate a fragment
5750 * with fragsz > PAGE_SIZE but the cache isn't big
5751 * enough to satisfy the request, this may
5752 * happen in low memory conditions.
5753 * We don't release the cache page because
5754 * it could make memory pressure worse
5755 * so we simply return NULL here.
5756 */
5757 return NULL;
5758 }
5759 }
5760
5761 nc->pagecnt_bias--;
5762 offset &= align_mask;
5763 nc->offset = offset;
5764
5765 return nc->va + offset;
5766}
5767EXPORT_SYMBOL(page_frag_alloc_align);
5768
5769/*
5770 * Frees a page fragment allocated out of either a compound or order 0 page.
5771 */
5772void page_frag_free(void *addr)
5773{
5774 struct page *page = virt_to_head_page(addr);
5775
5776 if (unlikely(put_page_testzero(page)))
5777 free_the_page(page, compound_order(page));
5778}
5779EXPORT_SYMBOL(page_frag_free);
5780
5781static void *make_alloc_exact(unsigned long addr, unsigned int order,
5782 size_t size)
5783{
5784 if (addr) {
5785 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
5786 struct page *page = virt_to_page((void *)addr);
5787 struct page *last = page + nr;
5788
5789 split_page_owner(page, 1 << order);
5790 split_page_memcg(page, 1 << order);
5791 while (page < --last)
5792 set_page_refcounted(last);
5793
5794 last = page + (1UL << order);
5795 for (page += nr; page < last; page++)
5796 __free_pages_ok(page, 0, FPI_TO_TAIL);
5797 }
5798 return (void *)addr;
5799}
5800
5801/**
5802 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5803 * @size: the number of bytes to allocate
5804 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5805 *
5806 * This function is similar to alloc_pages(), except that it allocates the
5807 * minimum number of pages to satisfy the request. alloc_pages() can only
5808 * allocate memory in power-of-two pages.
5809 *
5810 * This function is also limited by MAX_ORDER.
5811 *
5812 * Memory allocated by this function must be released by free_pages_exact().
5813 *
5814 * Return: pointer to the allocated area or %NULL in case of error.
5815 */
5816void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5817{
5818 unsigned int order = get_order(size);
5819 unsigned long addr;
5820
5821 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5822 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5823
5824 addr = __get_free_pages(gfp_mask, order);
5825 return make_alloc_exact(addr, order, size);
5826}
5827EXPORT_SYMBOL(alloc_pages_exact);
5828
5829/**
5830 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5831 * pages on a node.
5832 * @nid: the preferred node ID where memory should be allocated
5833 * @size: the number of bytes to allocate
5834 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5835 *
5836 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5837 * back.
5838 *
5839 * Return: pointer to the allocated area or %NULL in case of error.
5840 */
5841void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5842{
5843 unsigned int order = get_order(size);
5844 struct page *p;
5845
5846 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5847 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5848
5849 p = alloc_pages_node(nid, gfp_mask, order);
5850 if (!p)
5851 return NULL;
5852 return make_alloc_exact((unsigned long)page_address(p), order, size);
5853}
5854
5855/**
5856 * free_pages_exact - release memory allocated via alloc_pages_exact()
5857 * @virt: the value returned by alloc_pages_exact.
5858 * @size: size of allocation, same value as passed to alloc_pages_exact().
5859 *
5860 * Release the memory allocated by a previous call to alloc_pages_exact.
5861 */
5862void free_pages_exact(void *virt, size_t size)
5863{
5864 unsigned long addr = (unsigned long)virt;
5865 unsigned long end = addr + PAGE_ALIGN(size);
5866
5867 while (addr < end) {
5868 free_page(addr);
5869 addr += PAGE_SIZE;
5870 }
5871}
5872EXPORT_SYMBOL(free_pages_exact);
5873
5874/**
5875 * nr_free_zone_pages - count number of pages beyond high watermark
5876 * @offset: The zone index of the highest zone
5877 *
5878 * nr_free_zone_pages() counts the number of pages which are beyond the
5879 * high watermark within all zones at or below a given zone index. For each
5880 * zone, the number of pages is calculated as:
5881 *
5882 * nr_free_zone_pages = managed_pages - high_pages
5883 *
5884 * Return: number of pages beyond high watermark.
5885 */
5886static unsigned long nr_free_zone_pages(int offset)
5887{
5888 struct zoneref *z;
5889 struct zone *zone;
5890
5891 /* Just pick one node, since fallback list is circular */
5892 unsigned long sum = 0;
5893
5894 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5895
5896 for_each_zone_zonelist(zone, z, zonelist, offset) {
5897 unsigned long size = zone_managed_pages(zone);
5898 unsigned long high = high_wmark_pages(zone);
5899 if (size > high)
5900 sum += size - high;
5901 }
5902
5903 return sum;
5904}
5905
5906/**
5907 * nr_free_buffer_pages - count number of pages beyond high watermark
5908 *
5909 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5910 * watermark within ZONE_DMA and ZONE_NORMAL.
5911 *
5912 * Return: number of pages beyond high watermark within ZONE_DMA and
5913 * ZONE_NORMAL.
5914 */
5915unsigned long nr_free_buffer_pages(void)
5916{
5917 return nr_free_zone_pages(gfp_zone(GFP_USER));
5918}
5919EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5920
5921static inline void show_node(struct zone *zone)
5922{
5923 if (IS_ENABLED(CONFIG_NUMA))
5924 printk("Node %d ", zone_to_nid(zone));
5925}
5926
5927long si_mem_available(void)
5928{
5929 long available;
5930 unsigned long pagecache;
5931 unsigned long wmark_low = 0;
5932 unsigned long pages[NR_LRU_LISTS];
5933 unsigned long reclaimable;
5934 struct zone *zone;
5935 int lru;
5936
5937 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5938 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5939
5940 for_each_zone(zone)
5941 wmark_low += low_wmark_pages(zone);
5942
5943 /*
5944 * Estimate the amount of memory available for userspace allocations,
5945 * without causing swapping or OOM.
5946 */
5947 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5948
5949 /*
5950 * Not all the page cache can be freed, otherwise the system will
5951 * start swapping or thrashing. Assume at least half of the page
5952 * cache, or the low watermark worth of cache, needs to stay.
5953 */
5954 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5955 pagecache -= min(pagecache / 2, wmark_low);
5956 available += pagecache;
5957
5958 /*
5959 * Part of the reclaimable slab and other kernel memory consists of
5960 * items that are in use, and cannot be freed. Cap this estimate at the
5961 * low watermark.
5962 */
5963 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5964 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5965 available += reclaimable - min(reclaimable / 2, wmark_low);
5966
5967 if (available < 0)
5968 available = 0;
5969 return available;
5970}
5971EXPORT_SYMBOL_GPL(si_mem_available);
5972
5973void si_meminfo(struct sysinfo *val)
5974{
5975 val->totalram = totalram_pages();
5976 val->sharedram = global_node_page_state(NR_SHMEM);
5977 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5978 val->bufferram = nr_blockdev_pages();
5979 val->totalhigh = totalhigh_pages();
5980 val->freehigh = nr_free_highpages();
5981 val->mem_unit = PAGE_SIZE;
5982}
5983
5984EXPORT_SYMBOL(si_meminfo);
5985
5986#ifdef CONFIG_NUMA
5987void si_meminfo_node(struct sysinfo *val, int nid)
5988{
5989 int zone_type; /* needs to be signed */
5990 unsigned long managed_pages = 0;
5991 unsigned long managed_highpages = 0;
5992 unsigned long free_highpages = 0;
5993 pg_data_t *pgdat = NODE_DATA(nid);
5994
5995 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5996 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5997 val->totalram = managed_pages;
5998 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5999 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
6000#ifdef CONFIG_HIGHMEM
6001 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
6002 struct zone *zone = &pgdat->node_zones[zone_type];
6003
6004 if (is_highmem(zone)) {
6005 managed_highpages += zone_managed_pages(zone);
6006 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
6007 }
6008 }
6009 val->totalhigh = managed_highpages;
6010 val->freehigh = free_highpages;
6011#else
6012 val->totalhigh = managed_highpages;
6013 val->freehigh = free_highpages;
6014#endif
6015 val->mem_unit = PAGE_SIZE;
6016}
6017#endif
6018
6019/*
6020 * Determine whether the node should be displayed or not, depending on whether
6021 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
6022 */
6023static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
6024{
6025 if (!(flags & SHOW_MEM_FILTER_NODES))
6026 return false;
6027
6028 /*
6029 * no node mask - aka implicit memory numa policy. Do not bother with
6030 * the synchronization - read_mems_allowed_begin - because we do not
6031 * have to be precise here.
6032 */
6033 if (!nodemask)
6034 nodemask = &cpuset_current_mems_allowed;
6035
6036 return !node_isset(nid, *nodemask);
6037}
6038
6039#define K(x) ((x) << (PAGE_SHIFT-10))
6040
6041static void show_migration_types(unsigned char type)
6042{
6043 static const char types[MIGRATE_TYPES] = {
6044 [MIGRATE_UNMOVABLE] = 'U',
6045 [MIGRATE_MOVABLE] = 'M',
6046 [MIGRATE_RECLAIMABLE] = 'E',
6047 [MIGRATE_HIGHATOMIC] = 'H',
6048#ifdef CONFIG_CMA
6049 [MIGRATE_CMA] = 'C',
6050#endif
6051#ifdef CONFIG_MEMORY_ISOLATION
6052 [MIGRATE_ISOLATE] = 'I',
6053#endif
6054 };
6055 char tmp[MIGRATE_TYPES + 1];
6056 char *p = tmp;
6057 int i;
6058
6059 for (i = 0; i < MIGRATE_TYPES; i++) {
6060 if (type & (1 << i))
6061 *p++ = types[i];
6062 }
6063
6064 *p = '\0';
6065 printk(KERN_CONT "(%s) ", tmp);
6066}
6067
6068static bool node_has_managed_zones(pg_data_t *pgdat, int max_zone_idx)
6069{
6070 int zone_idx;
6071 for (zone_idx = 0; zone_idx <= max_zone_idx; zone_idx++)
6072 if (zone_managed_pages(pgdat->node_zones + zone_idx))
6073 return true;
6074 return false;
6075}
6076
6077/*
6078 * Show free area list (used inside shift_scroll-lock stuff)
6079 * We also calculate the percentage fragmentation. We do this by counting the
6080 * memory on each free list with the exception of the first item on the list.
6081 *
6082 * Bits in @filter:
6083 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
6084 * cpuset.
6085 */
6086void __show_free_areas(unsigned int filter, nodemask_t *nodemask, int max_zone_idx)
6087{
6088 unsigned long free_pcp = 0;
6089 int cpu, nid;
6090 struct zone *zone;
6091 pg_data_t *pgdat;
6092
6093 for_each_populated_zone(zone) {
6094 if (zone_idx(zone) > max_zone_idx)
6095 continue;
6096 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6097 continue;
6098
6099 for_each_online_cpu(cpu)
6100 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6101 }
6102
6103 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
6104 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
6105 " unevictable:%lu dirty:%lu writeback:%lu\n"
6106 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
6107 " mapped:%lu shmem:%lu pagetables:%lu\n"
6108 " sec_pagetables:%lu bounce:%lu\n"
6109 " kernel_misc_reclaimable:%lu\n"
6110 " free:%lu free_pcp:%lu free_cma:%lu\n",
6111 global_node_page_state(NR_ACTIVE_ANON),
6112 global_node_page_state(NR_INACTIVE_ANON),
6113 global_node_page_state(NR_ISOLATED_ANON),
6114 global_node_page_state(NR_ACTIVE_FILE),
6115 global_node_page_state(NR_INACTIVE_FILE),
6116 global_node_page_state(NR_ISOLATED_FILE),
6117 global_node_page_state(NR_UNEVICTABLE),
6118 global_node_page_state(NR_FILE_DIRTY),
6119 global_node_page_state(NR_WRITEBACK),
6120 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
6121 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
6122 global_node_page_state(NR_FILE_MAPPED),
6123 global_node_page_state(NR_SHMEM),
6124 global_node_page_state(NR_PAGETABLE),
6125 global_node_page_state(NR_SECONDARY_PAGETABLE),
6126 global_zone_page_state(NR_BOUNCE),
6127 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
6128 global_zone_page_state(NR_FREE_PAGES),
6129 free_pcp,
6130 global_zone_page_state(NR_FREE_CMA_PAGES));
6131
6132 for_each_online_pgdat(pgdat) {
6133 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
6134 continue;
6135 if (!node_has_managed_zones(pgdat, max_zone_idx))
6136 continue;
6137
6138 printk("Node %d"
6139 " active_anon:%lukB"
6140 " inactive_anon:%lukB"
6141 " active_file:%lukB"
6142 " inactive_file:%lukB"
6143 " unevictable:%lukB"
6144 " isolated(anon):%lukB"
6145 " isolated(file):%lukB"
6146 " mapped:%lukB"
6147 " dirty:%lukB"
6148 " writeback:%lukB"
6149 " shmem:%lukB"
6150#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6151 " shmem_thp: %lukB"
6152 " shmem_pmdmapped: %lukB"
6153 " anon_thp: %lukB"
6154#endif
6155 " writeback_tmp:%lukB"
6156 " kernel_stack:%lukB"
6157#ifdef CONFIG_SHADOW_CALL_STACK
6158 " shadow_call_stack:%lukB"
6159#endif
6160 " pagetables:%lukB"
6161 " sec_pagetables:%lukB"
6162 " all_unreclaimable? %s"
6163 "\n",
6164 pgdat->node_id,
6165 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
6166 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
6167 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
6168 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
6169 K(node_page_state(pgdat, NR_UNEVICTABLE)),
6170 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
6171 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
6172 K(node_page_state(pgdat, NR_FILE_MAPPED)),
6173 K(node_page_state(pgdat, NR_FILE_DIRTY)),
6174 K(node_page_state(pgdat, NR_WRITEBACK)),
6175 K(node_page_state(pgdat, NR_SHMEM)),
6176#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6177 K(node_page_state(pgdat, NR_SHMEM_THPS)),
6178 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6179 K(node_page_state(pgdat, NR_ANON_THPS)),
6180#endif
6181 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6182 node_page_state(pgdat, NR_KERNEL_STACK_KB),
6183#ifdef CONFIG_SHADOW_CALL_STACK
6184 node_page_state(pgdat, NR_KERNEL_SCS_KB),
6185#endif
6186 K(node_page_state(pgdat, NR_PAGETABLE)),
6187 K(node_page_state(pgdat, NR_SECONDARY_PAGETABLE)),
6188 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6189 "yes" : "no");
6190 }
6191
6192 for_each_populated_zone(zone) {
6193 int i;
6194
6195 if (zone_idx(zone) > max_zone_idx)
6196 continue;
6197 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6198 continue;
6199
6200 free_pcp = 0;
6201 for_each_online_cpu(cpu)
6202 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6203
6204 show_node(zone);
6205 printk(KERN_CONT
6206 "%s"
6207 " free:%lukB"
6208 " boost:%lukB"
6209 " min:%lukB"
6210 " low:%lukB"
6211 " high:%lukB"
6212 " reserved_highatomic:%luKB"
6213 " active_anon:%lukB"
6214 " inactive_anon:%lukB"
6215 " active_file:%lukB"
6216 " inactive_file:%lukB"
6217 " unevictable:%lukB"
6218 " writepending:%lukB"
6219 " present:%lukB"
6220 " managed:%lukB"
6221 " mlocked:%lukB"
6222 " bounce:%lukB"
6223 " free_pcp:%lukB"
6224 " local_pcp:%ukB"
6225 " free_cma:%lukB"
6226 "\n",
6227 zone->name,
6228 K(zone_page_state(zone, NR_FREE_PAGES)),
6229 K(zone->watermark_boost),
6230 K(min_wmark_pages(zone)),
6231 K(low_wmark_pages(zone)),
6232 K(high_wmark_pages(zone)),
6233 K(zone->nr_reserved_highatomic),
6234 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6235 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6236 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6237 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6238 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6239 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6240 K(zone->present_pages),
6241 K(zone_managed_pages(zone)),
6242 K(zone_page_state(zone, NR_MLOCK)),
6243 K(zone_page_state(zone, NR_BOUNCE)),
6244 K(free_pcp),
6245 K(this_cpu_read(zone->per_cpu_pageset->count)),
6246 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6247 printk("lowmem_reserve[]:");
6248 for (i = 0; i < MAX_NR_ZONES; i++)
6249 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6250 printk(KERN_CONT "\n");
6251 }
6252
6253 for_each_populated_zone(zone) {
6254 unsigned int order;
6255 unsigned long nr[MAX_ORDER], flags, total = 0;
6256 unsigned char types[MAX_ORDER];
6257
6258 if (zone_idx(zone) > max_zone_idx)
6259 continue;
6260 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6261 continue;
6262 show_node(zone);
6263 printk(KERN_CONT "%s: ", zone->name);
6264
6265 spin_lock_irqsave(&zone->lock, flags);
6266 for (order = 0; order < MAX_ORDER; order++) {
6267 struct free_area *area = &zone->free_area[order];
6268 int type;
6269
6270 nr[order] = area->nr_free;
6271 total += nr[order] << order;
6272
6273 types[order] = 0;
6274 for (type = 0; type < MIGRATE_TYPES; type++) {
6275 if (!free_area_empty(area, type))
6276 types[order] |= 1 << type;
6277 }
6278 }
6279 spin_unlock_irqrestore(&zone->lock, flags);
6280 for (order = 0; order < MAX_ORDER; order++) {
6281 printk(KERN_CONT "%lu*%lukB ",
6282 nr[order], K(1UL) << order);
6283 if (nr[order])
6284 show_migration_types(types[order]);
6285 }
6286 printk(KERN_CONT "= %lukB\n", K(total));
6287 }
6288
6289 for_each_online_node(nid) {
6290 if (show_mem_node_skip(filter, nid, nodemask))
6291 continue;
6292 hugetlb_show_meminfo_node(nid);
6293 }
6294
6295 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6296
6297 show_swap_cache_info();
6298}
6299
6300static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6301{
6302 zoneref->zone = zone;
6303 zoneref->zone_idx = zone_idx(zone);
6304}
6305
6306/*
6307 * Builds allocation fallback zone lists.
6308 *
6309 * Add all populated zones of a node to the zonelist.
6310 */
6311static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6312{
6313 struct zone *zone;
6314 enum zone_type zone_type = MAX_NR_ZONES;
6315 int nr_zones = 0;
6316
6317 do {
6318 zone_type--;
6319 zone = pgdat->node_zones + zone_type;
6320 if (populated_zone(zone)) {
6321 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6322 check_highest_zone(zone_type);
6323 }
6324 } while (zone_type);
6325
6326 return nr_zones;
6327}
6328
6329#ifdef CONFIG_NUMA
6330
6331static int __parse_numa_zonelist_order(char *s)
6332{
6333 /*
6334 * We used to support different zonelists modes but they turned
6335 * out to be just not useful. Let's keep the warning in place
6336 * if somebody still use the cmd line parameter so that we do
6337 * not fail it silently
6338 */
6339 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6340 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6341 return -EINVAL;
6342 }
6343 return 0;
6344}
6345
6346char numa_zonelist_order[] = "Node";
6347
6348/*
6349 * sysctl handler for numa_zonelist_order
6350 */
6351int numa_zonelist_order_handler(struct ctl_table *table, int write,
6352 void *buffer, size_t *length, loff_t *ppos)
6353{
6354 if (write)
6355 return __parse_numa_zonelist_order(buffer);
6356 return proc_dostring(table, write, buffer, length, ppos);
6357}
6358
6359
6360static int node_load[MAX_NUMNODES];
6361
6362/**
6363 * find_next_best_node - find the next node that should appear in a given node's fallback list
6364 * @node: node whose fallback list we're appending
6365 * @used_node_mask: nodemask_t of already used nodes
6366 *
6367 * We use a number of factors to determine which is the next node that should
6368 * appear on a given node's fallback list. The node should not have appeared
6369 * already in @node's fallback list, and it should be the next closest node
6370 * according to the distance array (which contains arbitrary distance values
6371 * from each node to each node in the system), and should also prefer nodes
6372 * with no CPUs, since presumably they'll have very little allocation pressure
6373 * on them otherwise.
6374 *
6375 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6376 */
6377int find_next_best_node(int node, nodemask_t *used_node_mask)
6378{
6379 int n, val;
6380 int min_val = INT_MAX;
6381 int best_node = NUMA_NO_NODE;
6382
6383 /* Use the local node if we haven't already */
6384 if (!node_isset(node, *used_node_mask)) {
6385 node_set(node, *used_node_mask);
6386 return node;
6387 }
6388
6389 for_each_node_state(n, N_MEMORY) {
6390
6391 /* Don't want a node to appear more than once */
6392 if (node_isset(n, *used_node_mask))
6393 continue;
6394
6395 /* Use the distance array to find the distance */
6396 val = node_distance(node, n);
6397
6398 /* Penalize nodes under us ("prefer the next node") */
6399 val += (n < node);
6400
6401 /* Give preference to headless and unused nodes */
6402 if (!cpumask_empty(cpumask_of_node(n)))
6403 val += PENALTY_FOR_NODE_WITH_CPUS;
6404
6405 /* Slight preference for less loaded node */
6406 val *= MAX_NUMNODES;
6407 val += node_load[n];
6408
6409 if (val < min_val) {
6410 min_val = val;
6411 best_node = n;
6412 }
6413 }
6414
6415 if (best_node >= 0)
6416 node_set(best_node, *used_node_mask);
6417
6418 return best_node;
6419}
6420
6421
6422/*
6423 * Build zonelists ordered by node and zones within node.
6424 * This results in maximum locality--normal zone overflows into local
6425 * DMA zone, if any--but risks exhausting DMA zone.
6426 */
6427static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6428 unsigned nr_nodes)
6429{
6430 struct zoneref *zonerefs;
6431 int i;
6432
6433 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6434
6435 for (i = 0; i < nr_nodes; i++) {
6436 int nr_zones;
6437
6438 pg_data_t *node = NODE_DATA(node_order[i]);
6439
6440 nr_zones = build_zonerefs_node(node, zonerefs);
6441 zonerefs += nr_zones;
6442 }
6443 zonerefs->zone = NULL;
6444 zonerefs->zone_idx = 0;
6445}
6446
6447/*
6448 * Build gfp_thisnode zonelists
6449 */
6450static void build_thisnode_zonelists(pg_data_t *pgdat)
6451{
6452 struct zoneref *zonerefs;
6453 int nr_zones;
6454
6455 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6456 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6457 zonerefs += nr_zones;
6458 zonerefs->zone = NULL;
6459 zonerefs->zone_idx = 0;
6460}
6461
6462/*
6463 * Build zonelists ordered by zone and nodes within zones.
6464 * This results in conserving DMA zone[s] until all Normal memory is
6465 * exhausted, but results in overflowing to remote node while memory
6466 * may still exist in local DMA zone.
6467 */
6468
6469static void build_zonelists(pg_data_t *pgdat)
6470{
6471 static int node_order[MAX_NUMNODES];
6472 int node, nr_nodes = 0;
6473 nodemask_t used_mask = NODE_MASK_NONE;
6474 int local_node, prev_node;
6475
6476 /* NUMA-aware ordering of nodes */
6477 local_node = pgdat->node_id;
6478 prev_node = local_node;
6479
6480 memset(node_order, 0, sizeof(node_order));
6481 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6482 /*
6483 * We don't want to pressure a particular node.
6484 * So adding penalty to the first node in same
6485 * distance group to make it round-robin.
6486 */
6487 if (node_distance(local_node, node) !=
6488 node_distance(local_node, prev_node))
6489 node_load[node] += 1;
6490
6491 node_order[nr_nodes++] = node;
6492 prev_node = node;
6493 }
6494
6495 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6496 build_thisnode_zonelists(pgdat);
6497 pr_info("Fallback order for Node %d: ", local_node);
6498 for (node = 0; node < nr_nodes; node++)
6499 pr_cont("%d ", node_order[node]);
6500 pr_cont("\n");
6501}
6502
6503#ifdef CONFIG_HAVE_MEMORYLESS_NODES
6504/*
6505 * Return node id of node used for "local" allocations.
6506 * I.e., first node id of first zone in arg node's generic zonelist.
6507 * Used for initializing percpu 'numa_mem', which is used primarily
6508 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6509 */
6510int local_memory_node(int node)
6511{
6512 struct zoneref *z;
6513
6514 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6515 gfp_zone(GFP_KERNEL),
6516 NULL);
6517 return zone_to_nid(z->zone);
6518}
6519#endif
6520
6521static void setup_min_unmapped_ratio(void);
6522static void setup_min_slab_ratio(void);
6523#else /* CONFIG_NUMA */
6524
6525static void build_zonelists(pg_data_t *pgdat)
6526{
6527 int node, local_node;
6528 struct zoneref *zonerefs;
6529 int nr_zones;
6530
6531 local_node = pgdat->node_id;
6532
6533 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6534 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6535 zonerefs += nr_zones;
6536
6537 /*
6538 * Now we build the zonelist so that it contains the zones
6539 * of all the other nodes.
6540 * We don't want to pressure a particular node, so when
6541 * building the zones for node N, we make sure that the
6542 * zones coming right after the local ones are those from
6543 * node N+1 (modulo N)
6544 */
6545 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6546 if (!node_online(node))
6547 continue;
6548 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6549 zonerefs += nr_zones;
6550 }
6551 for (node = 0; node < local_node; node++) {
6552 if (!node_online(node))
6553 continue;
6554 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6555 zonerefs += nr_zones;
6556 }
6557
6558 zonerefs->zone = NULL;
6559 zonerefs->zone_idx = 0;
6560}
6561
6562#endif /* CONFIG_NUMA */
6563
6564/*
6565 * Boot pageset table. One per cpu which is going to be used for all
6566 * zones and all nodes. The parameters will be set in such a way
6567 * that an item put on a list will immediately be handed over to
6568 * the buddy list. This is safe since pageset manipulation is done
6569 * with interrupts disabled.
6570 *
6571 * The boot_pagesets must be kept even after bootup is complete for
6572 * unused processors and/or zones. They do play a role for bootstrapping
6573 * hotplugged processors.
6574 *
6575 * zoneinfo_show() and maybe other functions do
6576 * not check if the processor is online before following the pageset pointer.
6577 * Other parts of the kernel may not check if the zone is available.
6578 */
6579static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6580/* These effectively disable the pcplists in the boot pageset completely */
6581#define BOOT_PAGESET_HIGH 0
6582#define BOOT_PAGESET_BATCH 1
6583static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6584static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6585static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6586
6587static void __build_all_zonelists(void *data)
6588{
6589 int nid;
6590 int __maybe_unused cpu;
6591 pg_data_t *self = data;
6592
6593 write_seqlock(&zonelist_update_seq);
6594
6595#ifdef CONFIG_NUMA
6596 memset(node_load, 0, sizeof(node_load));
6597#endif
6598
6599 /*
6600 * This node is hotadded and no memory is yet present. So just
6601 * building zonelists is fine - no need to touch other nodes.
6602 */
6603 if (self && !node_online(self->node_id)) {
6604 build_zonelists(self);
6605 } else {
6606 /*
6607 * All possible nodes have pgdat preallocated
6608 * in free_area_init
6609 */
6610 for_each_node(nid) {
6611 pg_data_t *pgdat = NODE_DATA(nid);
6612
6613 build_zonelists(pgdat);
6614 }
6615
6616#ifdef CONFIG_HAVE_MEMORYLESS_NODES
6617 /*
6618 * We now know the "local memory node" for each node--
6619 * i.e., the node of the first zone in the generic zonelist.
6620 * Set up numa_mem percpu variable for on-line cpus. During
6621 * boot, only the boot cpu should be on-line; we'll init the
6622 * secondary cpus' numa_mem as they come on-line. During
6623 * node/memory hotplug, we'll fixup all on-line cpus.
6624 */
6625 for_each_online_cpu(cpu)
6626 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6627#endif
6628 }
6629
6630 write_sequnlock(&zonelist_update_seq);
6631}
6632
6633static noinline void __init
6634build_all_zonelists_init(void)
6635{
6636 int cpu;
6637
6638 __build_all_zonelists(NULL);
6639
6640 /*
6641 * Initialize the boot_pagesets that are going to be used
6642 * for bootstrapping processors. The real pagesets for
6643 * each zone will be allocated later when the per cpu
6644 * allocator is available.
6645 *
6646 * boot_pagesets are used also for bootstrapping offline
6647 * cpus if the system is already booted because the pagesets
6648 * are needed to initialize allocators on a specific cpu too.
6649 * F.e. the percpu allocator needs the page allocator which
6650 * needs the percpu allocator in order to allocate its pagesets
6651 * (a chicken-egg dilemma).
6652 */
6653 for_each_possible_cpu(cpu)
6654 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6655
6656 mminit_verify_zonelist();
6657 cpuset_init_current_mems_allowed();
6658}
6659
6660/*
6661 * unless system_state == SYSTEM_BOOTING.
6662 *
6663 * __ref due to call of __init annotated helper build_all_zonelists_init
6664 * [protected by SYSTEM_BOOTING].
6665 */
6666void __ref build_all_zonelists(pg_data_t *pgdat)
6667{
6668 unsigned long vm_total_pages;
6669
6670 if (system_state == SYSTEM_BOOTING) {
6671 build_all_zonelists_init();
6672 } else {
6673 __build_all_zonelists(pgdat);
6674 /* cpuset refresh routine should be here */
6675 }
6676 /* Get the number of free pages beyond high watermark in all zones. */
6677 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6678 /*
6679 * Disable grouping by mobility if the number of pages in the
6680 * system is too low to allow the mechanism to work. It would be
6681 * more accurate, but expensive to check per-zone. This check is
6682 * made on memory-hotadd so a system can start with mobility
6683 * disabled and enable it later
6684 */
6685 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6686 page_group_by_mobility_disabled = 1;
6687 else
6688 page_group_by_mobility_disabled = 0;
6689
6690 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6691 nr_online_nodes,
6692 page_group_by_mobility_disabled ? "off" : "on",
6693 vm_total_pages);
6694#ifdef CONFIG_NUMA
6695 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6696#endif
6697}
6698
6699/* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6700static bool __meminit
6701overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6702{
6703 static struct memblock_region *r;
6704
6705 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6706 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6707 for_each_mem_region(r) {
6708 if (*pfn < memblock_region_memory_end_pfn(r))
6709 break;
6710 }
6711 }
6712 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6713 memblock_is_mirror(r)) {
6714 *pfn = memblock_region_memory_end_pfn(r);
6715 return true;
6716 }
6717 }
6718 return false;
6719}
6720
6721/*
6722 * Initially all pages are reserved - free ones are freed
6723 * up by memblock_free_all() once the early boot process is
6724 * done. Non-atomic initialization, single-pass.
6725 *
6726 * All aligned pageblocks are initialized to the specified migratetype
6727 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6728 * zone stats (e.g., nr_isolate_pageblock) are touched.
6729 */
6730void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6731 unsigned long start_pfn, unsigned long zone_end_pfn,
6732 enum meminit_context context,
6733 struct vmem_altmap *altmap, int migratetype)
6734{
6735 unsigned long pfn, end_pfn = start_pfn + size;
6736 struct page *page;
6737
6738 if (highest_memmap_pfn < end_pfn - 1)
6739 highest_memmap_pfn = end_pfn - 1;
6740
6741#ifdef CONFIG_ZONE_DEVICE
6742 /*
6743 * Honor reservation requested by the driver for this ZONE_DEVICE
6744 * memory. We limit the total number of pages to initialize to just
6745 * those that might contain the memory mapping. We will defer the
6746 * ZONE_DEVICE page initialization until after we have released
6747 * the hotplug lock.
6748 */
6749 if (zone == ZONE_DEVICE) {
6750 if (!altmap)
6751 return;
6752
6753 if (start_pfn == altmap->base_pfn)
6754 start_pfn += altmap->reserve;
6755 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6756 }
6757#endif
6758
6759 for (pfn = start_pfn; pfn < end_pfn; ) {
6760 /*
6761 * There can be holes in boot-time mem_map[]s handed to this
6762 * function. They do not exist on hotplugged memory.
6763 */
6764 if (context == MEMINIT_EARLY) {
6765 if (overlap_memmap_init(zone, &pfn))
6766 continue;
6767 if (defer_init(nid, pfn, zone_end_pfn))
6768 break;
6769 }
6770
6771 page = pfn_to_page(pfn);
6772 __init_single_page(page, pfn, zone, nid);
6773 if (context == MEMINIT_HOTPLUG)
6774 __SetPageReserved(page);
6775
6776 /*
6777 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6778 * such that unmovable allocations won't be scattered all
6779 * over the place during system boot.
6780 */
6781 if (pageblock_aligned(pfn)) {
6782 set_pageblock_migratetype(page, migratetype);
6783 cond_resched();
6784 }
6785 pfn++;
6786 }
6787}
6788
6789#ifdef CONFIG_ZONE_DEVICE
6790static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6791 unsigned long zone_idx, int nid,
6792 struct dev_pagemap *pgmap)
6793{
6794
6795 __init_single_page(page, pfn, zone_idx, nid);
6796
6797 /*
6798 * Mark page reserved as it will need to wait for onlining
6799 * phase for it to be fully associated with a zone.
6800 *
6801 * We can use the non-atomic __set_bit operation for setting
6802 * the flag as we are still initializing the pages.
6803 */
6804 __SetPageReserved(page);
6805
6806 /*
6807 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6808 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6809 * ever freed or placed on a driver-private list.
6810 */
6811 page->pgmap = pgmap;
6812 page->zone_device_data = NULL;
6813
6814 /*
6815 * Mark the block movable so that blocks are reserved for
6816 * movable at startup. This will force kernel allocations
6817 * to reserve their blocks rather than leaking throughout
6818 * the address space during boot when many long-lived
6819 * kernel allocations are made.
6820 *
6821 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6822 * because this is done early in section_activate()
6823 */
6824 if (pageblock_aligned(pfn)) {
6825 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6826 cond_resched();
6827 }
6828
6829 /*
6830 * ZONE_DEVICE pages are released directly to the driver page allocator
6831 * which will set the page count to 1 when allocating the page.
6832 */
6833 if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
6834 pgmap->type == MEMORY_DEVICE_COHERENT)
6835 set_page_count(page, 0);
6836}
6837
6838/*
6839 * With compound page geometry and when struct pages are stored in ram most
6840 * tail pages are reused. Consequently, the amount of unique struct pages to
6841 * initialize is a lot smaller that the total amount of struct pages being
6842 * mapped. This is a paired / mild layering violation with explicit knowledge
6843 * of how the sparse_vmemmap internals handle compound pages in the lack
6844 * of an altmap. See vmemmap_populate_compound_pages().
6845 */
6846static inline unsigned long compound_nr_pages(struct vmem_altmap *altmap,
6847 unsigned long nr_pages)
6848{
6849 return is_power_of_2(sizeof(struct page)) &&
6850 !altmap ? 2 * (PAGE_SIZE / sizeof(struct page)) : nr_pages;
6851}
6852
6853static void __ref memmap_init_compound(struct page *head,
6854 unsigned long head_pfn,
6855 unsigned long zone_idx, int nid,
6856 struct dev_pagemap *pgmap,
6857 unsigned long nr_pages)
6858{
6859 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6860 unsigned int order = pgmap->vmemmap_shift;
6861
6862 __SetPageHead(head);
6863 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6864 struct page *page = pfn_to_page(pfn);
6865
6866 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6867 prep_compound_tail(head, pfn - head_pfn);
6868 set_page_count(page, 0);
6869
6870 /*
6871 * The first tail page stores important compound page info.
6872 * Call prep_compound_head() after the first tail page has
6873 * been initialized, to not have the data overwritten.
6874 */
6875 if (pfn == head_pfn + 1)
6876 prep_compound_head(head, order);
6877 }
6878}
6879
6880void __ref memmap_init_zone_device(struct zone *zone,
6881 unsigned long start_pfn,
6882 unsigned long nr_pages,
6883 struct dev_pagemap *pgmap)
6884{
6885 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6886 struct pglist_data *pgdat = zone->zone_pgdat;
6887 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6888 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6889 unsigned long zone_idx = zone_idx(zone);
6890 unsigned long start = jiffies;
6891 int nid = pgdat->node_id;
6892
6893 if (WARN_ON_ONCE(!pgmap || zone_idx != ZONE_DEVICE))
6894 return;
6895
6896 /*
6897 * The call to memmap_init should have already taken care
6898 * of the pages reserved for the memmap, so we can just jump to
6899 * the end of that region and start processing the device pages.
6900 */
6901 if (altmap) {
6902 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6903 nr_pages = end_pfn - start_pfn;
6904 }
6905
6906 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6907 struct page *page = pfn_to_page(pfn);
6908
6909 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6910
6911 if (pfns_per_compound == 1)
6912 continue;
6913
6914 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6915 compound_nr_pages(altmap, pfns_per_compound));
6916 }
6917
6918 pr_info("%s initialised %lu pages in %ums\n", __func__,
6919 nr_pages, jiffies_to_msecs(jiffies - start));
6920}
6921
6922#endif
6923static void __meminit zone_init_free_lists(struct zone *zone)
6924{
6925 unsigned int order, t;
6926 for_each_migratetype_order(order, t) {
6927 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6928 zone->free_area[order].nr_free = 0;
6929 }
6930}
6931
6932/*
6933 * Only struct pages that correspond to ranges defined by memblock.memory
6934 * are zeroed and initialized by going through __init_single_page() during
6935 * memmap_init_zone_range().
6936 *
6937 * But, there could be struct pages that correspond to holes in
6938 * memblock.memory. This can happen because of the following reasons:
6939 * - physical memory bank size is not necessarily the exact multiple of the
6940 * arbitrary section size
6941 * - early reserved memory may not be listed in memblock.memory
6942 * - memory layouts defined with memmap= kernel parameter may not align
6943 * nicely with memmap sections
6944 *
6945 * Explicitly initialize those struct pages so that:
6946 * - PG_Reserved is set
6947 * - zone and node links point to zone and node that span the page if the
6948 * hole is in the middle of a zone
6949 * - zone and node links point to adjacent zone/node if the hole falls on
6950 * the zone boundary; the pages in such holes will be prepended to the
6951 * zone/node above the hole except for the trailing pages in the last
6952 * section that will be appended to the zone/node below.
6953 */
6954static void __init init_unavailable_range(unsigned long spfn,
6955 unsigned long epfn,
6956 int zone, int node)
6957{
6958 unsigned long pfn;
6959 u64 pgcnt = 0;
6960
6961 for (pfn = spfn; pfn < epfn; pfn++) {
6962 if (!pfn_valid(pageblock_start_pfn(pfn))) {
6963 pfn = pageblock_end_pfn(pfn) - 1;
6964 continue;
6965 }
6966 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6967 __SetPageReserved(pfn_to_page(pfn));
6968 pgcnt++;
6969 }
6970
6971 if (pgcnt)
6972 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6973 node, zone_names[zone], pgcnt);
6974}
6975
6976static void __init memmap_init_zone_range(struct zone *zone,
6977 unsigned long start_pfn,
6978 unsigned long end_pfn,
6979 unsigned long *hole_pfn)
6980{
6981 unsigned long zone_start_pfn = zone->zone_start_pfn;
6982 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6983 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6984
6985 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6986 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6987
6988 if (start_pfn >= end_pfn)
6989 return;
6990
6991 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6992 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6993
6994 if (*hole_pfn < start_pfn)
6995 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6996
6997 *hole_pfn = end_pfn;
6998}
6999
7000static void __init memmap_init(void)
7001{
7002 unsigned long start_pfn, end_pfn;
7003 unsigned long hole_pfn = 0;
7004 int i, j, zone_id = 0, nid;
7005
7006 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7007 struct pglist_data *node = NODE_DATA(nid);
7008
7009 for (j = 0; j < MAX_NR_ZONES; j++) {
7010 struct zone *zone = node->node_zones + j;
7011
7012 if (!populated_zone(zone))
7013 continue;
7014
7015 memmap_init_zone_range(zone, start_pfn, end_pfn,
7016 &hole_pfn);
7017 zone_id = j;
7018 }
7019 }
7020
7021#ifdef CONFIG_SPARSEMEM
7022 /*
7023 * Initialize the memory map for hole in the range [memory_end,
7024 * section_end].
7025 * Append the pages in this hole to the highest zone in the last
7026 * node.
7027 * The call to init_unavailable_range() is outside the ifdef to
7028 * silence the compiler warining about zone_id set but not used;
7029 * for FLATMEM it is a nop anyway
7030 */
7031 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
7032 if (hole_pfn < end_pfn)
7033#endif
7034 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
7035}
7036
7037void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
7038 phys_addr_t min_addr, int nid, bool exact_nid)
7039{
7040 void *ptr;
7041
7042 if (exact_nid)
7043 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
7044 MEMBLOCK_ALLOC_ACCESSIBLE,
7045 nid);
7046 else
7047 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
7048 MEMBLOCK_ALLOC_ACCESSIBLE,
7049 nid);
7050
7051 if (ptr && size > 0)
7052 page_init_poison(ptr, size);
7053
7054 return ptr;
7055}
7056
7057static int zone_batchsize(struct zone *zone)
7058{
7059#ifdef CONFIG_MMU
7060 int batch;
7061
7062 /*
7063 * The number of pages to batch allocate is either ~0.1%
7064 * of the zone or 1MB, whichever is smaller. The batch
7065 * size is striking a balance between allocation latency
7066 * and zone lock contention.
7067 */
7068 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
7069 batch /= 4; /* We effectively *= 4 below */
7070 if (batch < 1)
7071 batch = 1;
7072
7073 /*
7074 * Clamp the batch to a 2^n - 1 value. Having a power
7075 * of 2 value was found to be more likely to have
7076 * suboptimal cache aliasing properties in some cases.
7077 *
7078 * For example if 2 tasks are alternately allocating
7079 * batches of pages, one task can end up with a lot
7080 * of pages of one half of the possible page colors
7081 * and the other with pages of the other colors.
7082 */
7083 batch = rounddown_pow_of_two(batch + batch/2) - 1;
7084
7085 return batch;
7086
7087#else
7088 /* The deferral and batching of frees should be suppressed under NOMMU
7089 * conditions.
7090 *
7091 * The problem is that NOMMU needs to be able to allocate large chunks
7092 * of contiguous memory as there's no hardware page translation to
7093 * assemble apparent contiguous memory from discontiguous pages.
7094 *
7095 * Queueing large contiguous runs of pages for batching, however,
7096 * causes the pages to actually be freed in smaller chunks. As there
7097 * can be a significant delay between the individual batches being
7098 * recycled, this leads to the once large chunks of space being
7099 * fragmented and becoming unavailable for high-order allocations.
7100 */
7101 return 0;
7102#endif
7103}
7104
7105static int zone_highsize(struct zone *zone, int batch, int cpu_online)
7106{
7107#ifdef CONFIG_MMU
7108 int high;
7109 int nr_split_cpus;
7110 unsigned long total_pages;
7111
7112 if (!percpu_pagelist_high_fraction) {
7113 /*
7114 * By default, the high value of the pcp is based on the zone
7115 * low watermark so that if they are full then background
7116 * reclaim will not be started prematurely.
7117 */
7118 total_pages = low_wmark_pages(zone);
7119 } else {
7120 /*
7121 * If percpu_pagelist_high_fraction is configured, the high
7122 * value is based on a fraction of the managed pages in the
7123 * zone.
7124 */
7125 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
7126 }
7127
7128 /*
7129 * Split the high value across all online CPUs local to the zone. Note
7130 * that early in boot that CPUs may not be online yet and that during
7131 * CPU hotplug that the cpumask is not yet updated when a CPU is being
7132 * onlined. For memory nodes that have no CPUs, split pcp->high across
7133 * all online CPUs to mitigate the risk that reclaim is triggered
7134 * prematurely due to pages stored on pcp lists.
7135 */
7136 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
7137 if (!nr_split_cpus)
7138 nr_split_cpus = num_online_cpus();
7139 high = total_pages / nr_split_cpus;
7140
7141 /*
7142 * Ensure high is at least batch*4. The multiple is based on the
7143 * historical relationship between high and batch.
7144 */
7145 high = max(high, batch << 2);
7146
7147 return high;
7148#else
7149 return 0;
7150#endif
7151}
7152
7153/*
7154 * pcp->high and pcp->batch values are related and generally batch is lower
7155 * than high. They are also related to pcp->count such that count is lower
7156 * than high, and as soon as it reaches high, the pcplist is flushed.
7157 *
7158 * However, guaranteeing these relations at all times would require e.g. write
7159 * barriers here but also careful usage of read barriers at the read side, and
7160 * thus be prone to error and bad for performance. Thus the update only prevents
7161 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
7162 * can cope with those fields changing asynchronously, and fully trust only the
7163 * pcp->count field on the local CPU with interrupts disabled.
7164 *
7165 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
7166 * outside of boot time (or some other assurance that no concurrent updaters
7167 * exist).
7168 */
7169static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
7170 unsigned long batch)
7171{
7172 WRITE_ONCE(pcp->batch, batch);
7173 WRITE_ONCE(pcp->high, high);
7174}
7175
7176static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
7177{
7178 int pindex;
7179
7180 memset(pcp, 0, sizeof(*pcp));
7181 memset(pzstats, 0, sizeof(*pzstats));
7182
7183 spin_lock_init(&pcp->lock);
7184 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
7185 INIT_LIST_HEAD(&pcp->lists[pindex]);
7186
7187 /*
7188 * Set batch and high values safe for a boot pageset. A true percpu
7189 * pageset's initialization will update them subsequently. Here we don't
7190 * need to be as careful as pageset_update() as nobody can access the
7191 * pageset yet.
7192 */
7193 pcp->high = BOOT_PAGESET_HIGH;
7194 pcp->batch = BOOT_PAGESET_BATCH;
7195 pcp->free_factor = 0;
7196}
7197
7198static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
7199 unsigned long batch)
7200{
7201 struct per_cpu_pages *pcp;
7202 int cpu;
7203
7204 for_each_possible_cpu(cpu) {
7205 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7206 pageset_update(pcp, high, batch);
7207 }
7208}
7209
7210/*
7211 * Calculate and set new high and batch values for all per-cpu pagesets of a
7212 * zone based on the zone's size.
7213 */
7214static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
7215{
7216 int new_high, new_batch;
7217
7218 new_batch = max(1, zone_batchsize(zone));
7219 new_high = zone_highsize(zone, new_batch, cpu_online);
7220
7221 if (zone->pageset_high == new_high &&
7222 zone->pageset_batch == new_batch)
7223 return;
7224
7225 zone->pageset_high = new_high;
7226 zone->pageset_batch = new_batch;
7227
7228 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7229}
7230
7231void __meminit setup_zone_pageset(struct zone *zone)
7232{
7233 int cpu;
7234
7235 /* Size may be 0 on !SMP && !NUMA */
7236 if (sizeof(struct per_cpu_zonestat) > 0)
7237 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7238
7239 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7240 for_each_possible_cpu(cpu) {
7241 struct per_cpu_pages *pcp;
7242 struct per_cpu_zonestat *pzstats;
7243
7244 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7245 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7246 per_cpu_pages_init(pcp, pzstats);
7247 }
7248
7249 zone_set_pageset_high_and_batch(zone, 0);
7250}
7251
7252/*
7253 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7254 * page high values need to be recalculated.
7255 */
7256static void zone_pcp_update(struct zone *zone, int cpu_online)
7257{
7258 mutex_lock(&pcp_batch_high_lock);
7259 zone_set_pageset_high_and_batch(zone, cpu_online);
7260 mutex_unlock(&pcp_batch_high_lock);
7261}
7262
7263/*
7264 * Allocate per cpu pagesets and initialize them.
7265 * Before this call only boot pagesets were available.
7266 */
7267void __init setup_per_cpu_pageset(void)
7268{
7269 struct pglist_data *pgdat;
7270 struct zone *zone;
7271 int __maybe_unused cpu;
7272
7273 for_each_populated_zone(zone)
7274 setup_zone_pageset(zone);
7275
7276#ifdef CONFIG_NUMA
7277 /*
7278 * Unpopulated zones continue using the boot pagesets.
7279 * The numa stats for these pagesets need to be reset.
7280 * Otherwise, they will end up skewing the stats of
7281 * the nodes these zones are associated with.
7282 */
7283 for_each_possible_cpu(cpu) {
7284 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7285 memset(pzstats->vm_numa_event, 0,
7286 sizeof(pzstats->vm_numa_event));
7287 }
7288#endif
7289
7290 for_each_online_pgdat(pgdat)
7291 pgdat->per_cpu_nodestats =
7292 alloc_percpu(struct per_cpu_nodestat);
7293}
7294
7295static __meminit void zone_pcp_init(struct zone *zone)
7296{
7297 /*
7298 * per cpu subsystem is not up at this point. The following code
7299 * relies on the ability of the linker to provide the
7300 * offset of a (static) per cpu variable into the per cpu area.
7301 */
7302 zone->per_cpu_pageset = &boot_pageset;
7303 zone->per_cpu_zonestats = &boot_zonestats;
7304 zone->pageset_high = BOOT_PAGESET_HIGH;
7305 zone->pageset_batch = BOOT_PAGESET_BATCH;
7306
7307 if (populated_zone(zone))
7308 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7309 zone->present_pages, zone_batchsize(zone));
7310}
7311
7312void __meminit init_currently_empty_zone(struct zone *zone,
7313 unsigned long zone_start_pfn,
7314 unsigned long size)
7315{
7316 struct pglist_data *pgdat = zone->zone_pgdat;
7317 int zone_idx = zone_idx(zone) + 1;
7318
7319 if (zone_idx > pgdat->nr_zones)
7320 pgdat->nr_zones = zone_idx;
7321
7322 zone->zone_start_pfn = zone_start_pfn;
7323
7324 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7325 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7326 pgdat->node_id,
7327 (unsigned long)zone_idx(zone),
7328 zone_start_pfn, (zone_start_pfn + size));
7329
7330 zone_init_free_lists(zone);
7331 zone->initialized = 1;
7332}
7333
7334/**
7335 * get_pfn_range_for_nid - Return the start and end page frames for a node
7336 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7337 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7338 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7339 *
7340 * It returns the start and end page frame of a node based on information
7341 * provided by memblock_set_node(). If called for a node
7342 * with no available memory, a warning is printed and the start and end
7343 * PFNs will be 0.
7344 */
7345void __init get_pfn_range_for_nid(unsigned int nid,
7346 unsigned long *start_pfn, unsigned long *end_pfn)
7347{
7348 unsigned long this_start_pfn, this_end_pfn;
7349 int i;
7350
7351 *start_pfn = -1UL;
7352 *end_pfn = 0;
7353
7354 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7355 *start_pfn = min(*start_pfn, this_start_pfn);
7356 *end_pfn = max(*end_pfn, this_end_pfn);
7357 }
7358
7359 if (*start_pfn == -1UL)
7360 *start_pfn = 0;
7361}
7362
7363/*
7364 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7365 * assumption is made that zones within a node are ordered in monotonic
7366 * increasing memory addresses so that the "highest" populated zone is used
7367 */
7368static void __init find_usable_zone_for_movable(void)
7369{
7370 int zone_index;
7371 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7372 if (zone_index == ZONE_MOVABLE)
7373 continue;
7374
7375 if (arch_zone_highest_possible_pfn[zone_index] >
7376 arch_zone_lowest_possible_pfn[zone_index])
7377 break;
7378 }
7379
7380 VM_BUG_ON(zone_index == -1);
7381 movable_zone = zone_index;
7382}
7383
7384/*
7385 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7386 * because it is sized independent of architecture. Unlike the other zones,
7387 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7388 * in each node depending on the size of each node and how evenly kernelcore
7389 * is distributed. This helper function adjusts the zone ranges
7390 * provided by the architecture for a given node by using the end of the
7391 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7392 * zones within a node are in order of monotonic increases memory addresses
7393 */
7394static void __init adjust_zone_range_for_zone_movable(int nid,
7395 unsigned long zone_type,
7396 unsigned long node_start_pfn,
7397 unsigned long node_end_pfn,
7398 unsigned long *zone_start_pfn,
7399 unsigned long *zone_end_pfn)
7400{
7401 /* Only adjust if ZONE_MOVABLE is on this node */
7402 if (zone_movable_pfn[nid]) {
7403 /* Size ZONE_MOVABLE */
7404 if (zone_type == ZONE_MOVABLE) {
7405 *zone_start_pfn = zone_movable_pfn[nid];
7406 *zone_end_pfn = min(node_end_pfn,
7407 arch_zone_highest_possible_pfn[movable_zone]);
7408
7409 /* Adjust for ZONE_MOVABLE starting within this range */
7410 } else if (!mirrored_kernelcore &&
7411 *zone_start_pfn < zone_movable_pfn[nid] &&
7412 *zone_end_pfn > zone_movable_pfn[nid]) {
7413 *zone_end_pfn = zone_movable_pfn[nid];
7414
7415 /* Check if this whole range is within ZONE_MOVABLE */
7416 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7417 *zone_start_pfn = *zone_end_pfn;
7418 }
7419}
7420
7421/*
7422 * Return the number of pages a zone spans in a node, including holes
7423 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7424 */
7425static unsigned long __init zone_spanned_pages_in_node(int nid,
7426 unsigned long zone_type,
7427 unsigned long node_start_pfn,
7428 unsigned long node_end_pfn,
7429 unsigned long *zone_start_pfn,
7430 unsigned long *zone_end_pfn)
7431{
7432 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7433 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7434 /* When hotadd a new node from cpu_up(), the node should be empty */
7435 if (!node_start_pfn && !node_end_pfn)
7436 return 0;
7437
7438 /* Get the start and end of the zone */
7439 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7440 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7441 adjust_zone_range_for_zone_movable(nid, zone_type,
7442 node_start_pfn, node_end_pfn,
7443 zone_start_pfn, zone_end_pfn);
7444
7445 /* Check that this node has pages within the zone's required range */
7446 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7447 return 0;
7448
7449 /* Move the zone boundaries inside the node if necessary */
7450 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7451 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7452
7453 /* Return the spanned pages */
7454 return *zone_end_pfn - *zone_start_pfn;
7455}
7456
7457/*
7458 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7459 * then all holes in the requested range will be accounted for.
7460 */
7461unsigned long __init __absent_pages_in_range(int nid,
7462 unsigned long range_start_pfn,
7463 unsigned long range_end_pfn)
7464{
7465 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7466 unsigned long start_pfn, end_pfn;
7467 int i;
7468
7469 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7470 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7471 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7472 nr_absent -= end_pfn - start_pfn;
7473 }
7474 return nr_absent;
7475}
7476
7477/**
7478 * absent_pages_in_range - Return number of page frames in holes within a range
7479 * @start_pfn: The start PFN to start searching for holes
7480 * @end_pfn: The end PFN to stop searching for holes
7481 *
7482 * Return: the number of pages frames in memory holes within a range.
7483 */
7484unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7485 unsigned long end_pfn)
7486{
7487 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7488}
7489
7490/* Return the number of page frames in holes in a zone on a node */
7491static unsigned long __init zone_absent_pages_in_node(int nid,
7492 unsigned long zone_type,
7493 unsigned long node_start_pfn,
7494 unsigned long node_end_pfn)
7495{
7496 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7497 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7498 unsigned long zone_start_pfn, zone_end_pfn;
7499 unsigned long nr_absent;
7500
7501 /* When hotadd a new node from cpu_up(), the node should be empty */
7502 if (!node_start_pfn && !node_end_pfn)
7503 return 0;
7504
7505 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7506 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7507
7508 adjust_zone_range_for_zone_movable(nid, zone_type,
7509 node_start_pfn, node_end_pfn,
7510 &zone_start_pfn, &zone_end_pfn);
7511 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7512
7513 /*
7514 * ZONE_MOVABLE handling.
7515 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7516 * and vice versa.
7517 */
7518 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7519 unsigned long start_pfn, end_pfn;
7520 struct memblock_region *r;
7521
7522 for_each_mem_region(r) {
7523 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7524 zone_start_pfn, zone_end_pfn);
7525 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7526 zone_start_pfn, zone_end_pfn);
7527
7528 if (zone_type == ZONE_MOVABLE &&
7529 memblock_is_mirror(r))
7530 nr_absent += end_pfn - start_pfn;
7531
7532 if (zone_type == ZONE_NORMAL &&
7533 !memblock_is_mirror(r))
7534 nr_absent += end_pfn - start_pfn;
7535 }
7536 }
7537
7538 return nr_absent;
7539}
7540
7541static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7542 unsigned long node_start_pfn,
7543 unsigned long node_end_pfn)
7544{
7545 unsigned long realtotalpages = 0, totalpages = 0;
7546 enum zone_type i;
7547
7548 for (i = 0; i < MAX_NR_ZONES; i++) {
7549 struct zone *zone = pgdat->node_zones + i;
7550 unsigned long zone_start_pfn, zone_end_pfn;
7551 unsigned long spanned, absent;
7552 unsigned long size, real_size;
7553
7554 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7555 node_start_pfn,
7556 node_end_pfn,
7557 &zone_start_pfn,
7558 &zone_end_pfn);
7559 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7560 node_start_pfn,
7561 node_end_pfn);
7562
7563 size = spanned;
7564 real_size = size - absent;
7565
7566 if (size)
7567 zone->zone_start_pfn = zone_start_pfn;
7568 else
7569 zone->zone_start_pfn = 0;
7570 zone->spanned_pages = size;
7571 zone->present_pages = real_size;
7572#if defined(CONFIG_MEMORY_HOTPLUG)
7573 zone->present_early_pages = real_size;
7574#endif
7575
7576 totalpages += size;
7577 realtotalpages += real_size;
7578 }
7579
7580 pgdat->node_spanned_pages = totalpages;
7581 pgdat->node_present_pages = realtotalpages;
7582 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7583}
7584
7585#ifndef CONFIG_SPARSEMEM
7586/*
7587 * Calculate the size of the zone->blockflags rounded to an unsigned long
7588 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7589 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7590 * round what is now in bits to nearest long in bits, then return it in
7591 * bytes.
7592 */
7593static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7594{
7595 unsigned long usemapsize;
7596
7597 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7598 usemapsize = roundup(zonesize, pageblock_nr_pages);
7599 usemapsize = usemapsize >> pageblock_order;
7600 usemapsize *= NR_PAGEBLOCK_BITS;
7601 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7602
7603 return usemapsize / 8;
7604}
7605
7606static void __ref setup_usemap(struct zone *zone)
7607{
7608 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7609 zone->spanned_pages);
7610 zone->pageblock_flags = NULL;
7611 if (usemapsize) {
7612 zone->pageblock_flags =
7613 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7614 zone_to_nid(zone));
7615 if (!zone->pageblock_flags)
7616 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7617 usemapsize, zone->name, zone_to_nid(zone));
7618 }
7619}
7620#else
7621static inline void setup_usemap(struct zone *zone) {}
7622#endif /* CONFIG_SPARSEMEM */
7623
7624#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7625
7626/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7627void __init set_pageblock_order(void)
7628{
7629 unsigned int order = MAX_ORDER - 1;
7630
7631 /* Check that pageblock_nr_pages has not already been setup */
7632 if (pageblock_order)
7633 return;
7634
7635 /* Don't let pageblocks exceed the maximum allocation granularity. */
7636 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7637 order = HUGETLB_PAGE_ORDER;
7638
7639 /*
7640 * Assume the largest contiguous order of interest is a huge page.
7641 * This value may be variable depending on boot parameters on IA64 and
7642 * powerpc.
7643 */
7644 pageblock_order = order;
7645}
7646#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7647
7648/*
7649 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7650 * is unused as pageblock_order is set at compile-time. See
7651 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7652 * the kernel config
7653 */
7654void __init set_pageblock_order(void)
7655{
7656}
7657
7658#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7659
7660static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7661 unsigned long present_pages)
7662{
7663 unsigned long pages = spanned_pages;
7664
7665 /*
7666 * Provide a more accurate estimation if there are holes within
7667 * the zone and SPARSEMEM is in use. If there are holes within the
7668 * zone, each populated memory region may cost us one or two extra
7669 * memmap pages due to alignment because memmap pages for each
7670 * populated regions may not be naturally aligned on page boundary.
7671 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7672 */
7673 if (spanned_pages > present_pages + (present_pages >> 4) &&
7674 IS_ENABLED(CONFIG_SPARSEMEM))
7675 pages = present_pages;
7676
7677 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7678}
7679
7680#ifdef CONFIG_TRANSPARENT_HUGEPAGE
7681static void pgdat_init_split_queue(struct pglist_data *pgdat)
7682{
7683 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7684
7685 spin_lock_init(&ds_queue->split_queue_lock);
7686 INIT_LIST_HEAD(&ds_queue->split_queue);
7687 ds_queue->split_queue_len = 0;
7688}
7689#else
7690static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7691#endif
7692
7693#ifdef CONFIG_COMPACTION
7694static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7695{
7696 init_waitqueue_head(&pgdat->kcompactd_wait);
7697}
7698#else
7699static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7700#endif
7701
7702static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7703{
7704 int i;
7705
7706 pgdat_resize_init(pgdat);
7707 pgdat_kswapd_lock_init(pgdat);
7708
7709 pgdat_init_split_queue(pgdat);
7710 pgdat_init_kcompactd(pgdat);
7711
7712 init_waitqueue_head(&pgdat->kswapd_wait);
7713 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7714
7715 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7716 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7717
7718 pgdat_page_ext_init(pgdat);
7719 lruvec_init(&pgdat->__lruvec);
7720}
7721
7722static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7723 unsigned long remaining_pages)
7724{
7725 atomic_long_set(&zone->managed_pages, remaining_pages);
7726 zone_set_nid(zone, nid);
7727 zone->name = zone_names[idx];
7728 zone->zone_pgdat = NODE_DATA(nid);
7729 spin_lock_init(&zone->lock);
7730 zone_seqlock_init(zone);
7731 zone_pcp_init(zone);
7732}
7733
7734/*
7735 * Set up the zone data structures
7736 * - init pgdat internals
7737 * - init all zones belonging to this node
7738 *
7739 * NOTE: this function is only called during memory hotplug
7740 */
7741#ifdef CONFIG_MEMORY_HOTPLUG
7742void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7743{
7744 int nid = pgdat->node_id;
7745 enum zone_type z;
7746 int cpu;
7747
7748 pgdat_init_internals(pgdat);
7749
7750 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7751 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7752
7753 /*
7754 * Reset the nr_zones, order and highest_zoneidx before reuse.
7755 * Note that kswapd will init kswapd_highest_zoneidx properly
7756 * when it starts in the near future.
7757 */
7758 pgdat->nr_zones = 0;
7759 pgdat->kswapd_order = 0;
7760 pgdat->kswapd_highest_zoneidx = 0;
7761 pgdat->node_start_pfn = 0;
7762 for_each_online_cpu(cpu) {
7763 struct per_cpu_nodestat *p;
7764
7765 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7766 memset(p, 0, sizeof(*p));
7767 }
7768
7769 for (z = 0; z < MAX_NR_ZONES; z++)
7770 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7771}
7772#endif
7773
7774/*
7775 * Set up the zone data structures:
7776 * - mark all pages reserved
7777 * - mark all memory queues empty
7778 * - clear the memory bitmaps
7779 *
7780 * NOTE: pgdat should get zeroed by caller.
7781 * NOTE: this function is only called during early init.
7782 */
7783static void __init free_area_init_core(struct pglist_data *pgdat)
7784{
7785 enum zone_type j;
7786 int nid = pgdat->node_id;
7787
7788 pgdat_init_internals(pgdat);
7789 pgdat->per_cpu_nodestats = &boot_nodestats;
7790
7791 for (j = 0; j < MAX_NR_ZONES; j++) {
7792 struct zone *zone = pgdat->node_zones + j;
7793 unsigned long size, freesize, memmap_pages;
7794
7795 size = zone->spanned_pages;
7796 freesize = zone->present_pages;
7797
7798 /*
7799 * Adjust freesize so that it accounts for how much memory
7800 * is used by this zone for memmap. This affects the watermark
7801 * and per-cpu initialisations
7802 */
7803 memmap_pages = calc_memmap_size(size, freesize);
7804 if (!is_highmem_idx(j)) {
7805 if (freesize >= memmap_pages) {
7806 freesize -= memmap_pages;
7807 if (memmap_pages)
7808 pr_debug(" %s zone: %lu pages used for memmap\n",
7809 zone_names[j], memmap_pages);
7810 } else
7811 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7812 zone_names[j], memmap_pages, freesize);
7813 }
7814
7815 /* Account for reserved pages */
7816 if (j == 0 && freesize > dma_reserve) {
7817 freesize -= dma_reserve;
7818 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7819 }
7820
7821 if (!is_highmem_idx(j))
7822 nr_kernel_pages += freesize;
7823 /* Charge for highmem memmap if there are enough kernel pages */
7824 else if (nr_kernel_pages > memmap_pages * 2)
7825 nr_kernel_pages -= memmap_pages;
7826 nr_all_pages += freesize;
7827
7828 /*
7829 * Set an approximate value for lowmem here, it will be adjusted
7830 * when the bootmem allocator frees pages into the buddy system.
7831 * And all highmem pages will be managed by the buddy system.
7832 */
7833 zone_init_internals(zone, j, nid, freesize);
7834
7835 if (!size)
7836 continue;
7837
7838 set_pageblock_order();
7839 setup_usemap(zone);
7840 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7841 }
7842}
7843
7844#ifdef CONFIG_FLATMEM
7845static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7846{
7847 unsigned long __maybe_unused start = 0;
7848 unsigned long __maybe_unused offset = 0;
7849
7850 /* Skip empty nodes */
7851 if (!pgdat->node_spanned_pages)
7852 return;
7853
7854 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7855 offset = pgdat->node_start_pfn - start;
7856 /* ia64 gets its own node_mem_map, before this, without bootmem */
7857 if (!pgdat->node_mem_map) {
7858 unsigned long size, end;
7859 struct page *map;
7860
7861 /*
7862 * The zone's endpoints aren't required to be MAX_ORDER
7863 * aligned but the node_mem_map endpoints must be in order
7864 * for the buddy allocator to function correctly.
7865 */
7866 end = pgdat_end_pfn(pgdat);
7867 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7868 size = (end - start) * sizeof(struct page);
7869 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7870 pgdat->node_id, false);
7871 if (!map)
7872 panic("Failed to allocate %ld bytes for node %d memory map\n",
7873 size, pgdat->node_id);
7874 pgdat->node_mem_map = map + offset;
7875 }
7876 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7877 __func__, pgdat->node_id, (unsigned long)pgdat,
7878 (unsigned long)pgdat->node_mem_map);
7879#ifndef CONFIG_NUMA
7880 /*
7881 * With no DISCONTIG, the global mem_map is just set as node 0's
7882 */
7883 if (pgdat == NODE_DATA(0)) {
7884 mem_map = NODE_DATA(0)->node_mem_map;
7885 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7886 mem_map -= offset;
7887 }
7888#endif
7889}
7890#else
7891static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7892#endif /* CONFIG_FLATMEM */
7893
7894#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7895static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7896{
7897 pgdat->first_deferred_pfn = ULONG_MAX;
7898}
7899#else
7900static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7901#endif
7902
7903static void __init free_area_init_node(int nid)
7904{
7905 pg_data_t *pgdat = NODE_DATA(nid);
7906 unsigned long start_pfn = 0;
7907 unsigned long end_pfn = 0;
7908
7909 /* pg_data_t should be reset to zero when it's allocated */
7910 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7911
7912 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7913
7914 pgdat->node_id = nid;
7915 pgdat->node_start_pfn = start_pfn;
7916 pgdat->per_cpu_nodestats = NULL;
7917
7918 if (start_pfn != end_pfn) {
7919 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7920 (u64)start_pfn << PAGE_SHIFT,
7921 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7922 } else {
7923 pr_info("Initmem setup node %d as memoryless\n", nid);
7924 }
7925
7926 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7927
7928 alloc_node_mem_map(pgdat);
7929 pgdat_set_deferred_range(pgdat);
7930
7931 free_area_init_core(pgdat);
7932}
7933
7934static void __init free_area_init_memoryless_node(int nid)
7935{
7936 free_area_init_node(nid);
7937}
7938
7939#if MAX_NUMNODES > 1
7940/*
7941 * Figure out the number of possible node ids.
7942 */
7943void __init setup_nr_node_ids(void)
7944{
7945 unsigned int highest;
7946
7947 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7948 nr_node_ids = highest + 1;
7949}
7950#endif
7951
7952/**
7953 * node_map_pfn_alignment - determine the maximum internode alignment
7954 *
7955 * This function should be called after node map is populated and sorted.
7956 * It calculates the maximum power of two alignment which can distinguish
7957 * all the nodes.
7958 *
7959 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7960 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7961 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7962 * shifted, 1GiB is enough and this function will indicate so.
7963 *
7964 * This is used to test whether pfn -> nid mapping of the chosen memory
7965 * model has fine enough granularity to avoid incorrect mapping for the
7966 * populated node map.
7967 *
7968 * Return: the determined alignment in pfn's. 0 if there is no alignment
7969 * requirement (single node).
7970 */
7971unsigned long __init node_map_pfn_alignment(void)
7972{
7973 unsigned long accl_mask = 0, last_end = 0;
7974 unsigned long start, end, mask;
7975 int last_nid = NUMA_NO_NODE;
7976 int i, nid;
7977
7978 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7979 if (!start || last_nid < 0 || last_nid == nid) {
7980 last_nid = nid;
7981 last_end = end;
7982 continue;
7983 }
7984
7985 /*
7986 * Start with a mask granular enough to pin-point to the
7987 * start pfn and tick off bits one-by-one until it becomes
7988 * too coarse to separate the current node from the last.
7989 */
7990 mask = ~((1 << __ffs(start)) - 1);
7991 while (mask && last_end <= (start & (mask << 1)))
7992 mask <<= 1;
7993
7994 /* accumulate all internode masks */
7995 accl_mask |= mask;
7996 }
7997
7998 /* convert mask to number of pages */
7999 return ~accl_mask + 1;
8000}
8001
8002/*
8003 * early_calculate_totalpages()
8004 * Sum pages in active regions for movable zone.
8005 * Populate N_MEMORY for calculating usable_nodes.
8006 */
8007static unsigned long __init early_calculate_totalpages(void)
8008{
8009 unsigned long totalpages = 0;
8010 unsigned long start_pfn, end_pfn;
8011 int i, nid;
8012
8013 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8014 unsigned long pages = end_pfn - start_pfn;
8015
8016 totalpages += pages;
8017 if (pages)
8018 node_set_state(nid, N_MEMORY);
8019 }
8020 return totalpages;
8021}
8022
8023/*
8024 * Find the PFN the Movable zone begins in each node. Kernel memory
8025 * is spread evenly between nodes as long as the nodes have enough
8026 * memory. When they don't, some nodes will have more kernelcore than
8027 * others
8028 */
8029static void __init find_zone_movable_pfns_for_nodes(void)
8030{
8031 int i, nid;
8032 unsigned long usable_startpfn;
8033 unsigned long kernelcore_node, kernelcore_remaining;
8034 /* save the state before borrow the nodemask */
8035 nodemask_t saved_node_state = node_states[N_MEMORY];
8036 unsigned long totalpages = early_calculate_totalpages();
8037 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
8038 struct memblock_region *r;
8039
8040 /* Need to find movable_zone earlier when movable_node is specified. */
8041 find_usable_zone_for_movable();
8042
8043 /*
8044 * If movable_node is specified, ignore kernelcore and movablecore
8045 * options.
8046 */
8047 if (movable_node_is_enabled()) {
8048 for_each_mem_region(r) {
8049 if (!memblock_is_hotpluggable(r))
8050 continue;
8051
8052 nid = memblock_get_region_node(r);
8053
8054 usable_startpfn = PFN_DOWN(r->base);
8055 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
8056 min(usable_startpfn, zone_movable_pfn[nid]) :
8057 usable_startpfn;
8058 }
8059
8060 goto out2;
8061 }
8062
8063 /*
8064 * If kernelcore=mirror is specified, ignore movablecore option
8065 */
8066 if (mirrored_kernelcore) {
8067 bool mem_below_4gb_not_mirrored = false;
8068
8069 for_each_mem_region(r) {
8070 if (memblock_is_mirror(r))
8071 continue;
8072
8073 nid = memblock_get_region_node(r);
8074
8075 usable_startpfn = memblock_region_memory_base_pfn(r);
8076
8077 if (usable_startpfn < PHYS_PFN(SZ_4G)) {
8078 mem_below_4gb_not_mirrored = true;
8079 continue;
8080 }
8081
8082 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
8083 min(usable_startpfn, zone_movable_pfn[nid]) :
8084 usable_startpfn;
8085 }
8086
8087 if (mem_below_4gb_not_mirrored)
8088 pr_warn("This configuration results in unmirrored kernel memory.\n");
8089
8090 goto out2;
8091 }
8092
8093 /*
8094 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
8095 * amount of necessary memory.
8096 */
8097 if (required_kernelcore_percent)
8098 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
8099 10000UL;
8100 if (required_movablecore_percent)
8101 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
8102 10000UL;
8103
8104 /*
8105 * If movablecore= was specified, calculate what size of
8106 * kernelcore that corresponds so that memory usable for
8107 * any allocation type is evenly spread. If both kernelcore
8108 * and movablecore are specified, then the value of kernelcore
8109 * will be used for required_kernelcore if it's greater than
8110 * what movablecore would have allowed.
8111 */
8112 if (required_movablecore) {
8113 unsigned long corepages;
8114
8115 /*
8116 * Round-up so that ZONE_MOVABLE is at least as large as what
8117 * was requested by the user
8118 */
8119 required_movablecore =
8120 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
8121 required_movablecore = min(totalpages, required_movablecore);
8122 corepages = totalpages - required_movablecore;
8123
8124 required_kernelcore = max(required_kernelcore, corepages);
8125 }
8126
8127 /*
8128 * If kernelcore was not specified or kernelcore size is larger
8129 * than totalpages, there is no ZONE_MOVABLE.
8130 */
8131 if (!required_kernelcore || required_kernelcore >= totalpages)
8132 goto out;
8133
8134 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
8135 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
8136
8137restart:
8138 /* Spread kernelcore memory as evenly as possible throughout nodes */
8139 kernelcore_node = required_kernelcore / usable_nodes;
8140 for_each_node_state(nid, N_MEMORY) {
8141 unsigned long start_pfn, end_pfn;
8142
8143 /*
8144 * Recalculate kernelcore_node if the division per node
8145 * now exceeds what is necessary to satisfy the requested
8146 * amount of memory for the kernel
8147 */
8148 if (required_kernelcore < kernelcore_node)
8149 kernelcore_node = required_kernelcore / usable_nodes;
8150
8151 /*
8152 * As the map is walked, we track how much memory is usable
8153 * by the kernel using kernelcore_remaining. When it is
8154 * 0, the rest of the node is usable by ZONE_MOVABLE
8155 */
8156 kernelcore_remaining = kernelcore_node;
8157
8158 /* Go through each range of PFNs within this node */
8159 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
8160 unsigned long size_pages;
8161
8162 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
8163 if (start_pfn >= end_pfn)
8164 continue;
8165
8166 /* Account for what is only usable for kernelcore */
8167 if (start_pfn < usable_startpfn) {
8168 unsigned long kernel_pages;
8169 kernel_pages = min(end_pfn, usable_startpfn)
8170 - start_pfn;
8171
8172 kernelcore_remaining -= min(kernel_pages,
8173 kernelcore_remaining);
8174 required_kernelcore -= min(kernel_pages,
8175 required_kernelcore);
8176
8177 /* Continue if range is now fully accounted */
8178 if (end_pfn <= usable_startpfn) {
8179
8180 /*
8181 * Push zone_movable_pfn to the end so
8182 * that if we have to rebalance
8183 * kernelcore across nodes, we will
8184 * not double account here
8185 */
8186 zone_movable_pfn[nid] = end_pfn;
8187 continue;
8188 }
8189 start_pfn = usable_startpfn;
8190 }
8191
8192 /*
8193 * The usable PFN range for ZONE_MOVABLE is from
8194 * start_pfn->end_pfn. Calculate size_pages as the
8195 * number of pages used as kernelcore
8196 */
8197 size_pages = end_pfn - start_pfn;
8198 if (size_pages > kernelcore_remaining)
8199 size_pages = kernelcore_remaining;
8200 zone_movable_pfn[nid] = start_pfn + size_pages;
8201
8202 /*
8203 * Some kernelcore has been met, update counts and
8204 * break if the kernelcore for this node has been
8205 * satisfied
8206 */
8207 required_kernelcore -= min(required_kernelcore,
8208 size_pages);
8209 kernelcore_remaining -= size_pages;
8210 if (!kernelcore_remaining)
8211 break;
8212 }
8213 }
8214
8215 /*
8216 * If there is still required_kernelcore, we do another pass with one
8217 * less node in the count. This will push zone_movable_pfn[nid] further
8218 * along on the nodes that still have memory until kernelcore is
8219 * satisfied
8220 */
8221 usable_nodes--;
8222 if (usable_nodes && required_kernelcore > usable_nodes)
8223 goto restart;
8224
8225out2:
8226 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
8227 for (nid = 0; nid < MAX_NUMNODES; nid++) {
8228 unsigned long start_pfn, end_pfn;
8229
8230 zone_movable_pfn[nid] =
8231 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
8232
8233 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
8234 if (zone_movable_pfn[nid] >= end_pfn)
8235 zone_movable_pfn[nid] = 0;
8236 }
8237
8238out:
8239 /* restore the node_state */
8240 node_states[N_MEMORY] = saved_node_state;
8241}
8242
8243/* Any regular or high memory on that node ? */
8244static void check_for_memory(pg_data_t *pgdat, int nid)
8245{
8246 enum zone_type zone_type;
8247
8248 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8249 struct zone *zone = &pgdat->node_zones[zone_type];
8250 if (populated_zone(zone)) {
8251 if (IS_ENABLED(CONFIG_HIGHMEM))
8252 node_set_state(nid, N_HIGH_MEMORY);
8253 if (zone_type <= ZONE_NORMAL)
8254 node_set_state(nid, N_NORMAL_MEMORY);
8255 break;
8256 }
8257 }
8258}
8259
8260/*
8261 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8262 * such cases we allow max_zone_pfn sorted in the descending order
8263 */
8264bool __weak arch_has_descending_max_zone_pfns(void)
8265{
8266 return false;
8267}
8268
8269/**
8270 * free_area_init - Initialise all pg_data_t and zone data
8271 * @max_zone_pfn: an array of max PFNs for each zone
8272 *
8273 * This will call free_area_init_node() for each active node in the system.
8274 * Using the page ranges provided by memblock_set_node(), the size of each
8275 * zone in each node and their holes is calculated. If the maximum PFN
8276 * between two adjacent zones match, it is assumed that the zone is empty.
8277 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8278 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8279 * starts where the previous one ended. For example, ZONE_DMA32 starts
8280 * at arch_max_dma_pfn.
8281 */
8282void __init free_area_init(unsigned long *max_zone_pfn)
8283{
8284 unsigned long start_pfn, end_pfn;
8285 int i, nid, zone;
8286 bool descending;
8287
8288 /* Record where the zone boundaries are */
8289 memset(arch_zone_lowest_possible_pfn, 0,
8290 sizeof(arch_zone_lowest_possible_pfn));
8291 memset(arch_zone_highest_possible_pfn, 0,
8292 sizeof(arch_zone_highest_possible_pfn));
8293
8294 start_pfn = PHYS_PFN(memblock_start_of_DRAM());
8295 descending = arch_has_descending_max_zone_pfns();
8296
8297 for (i = 0; i < MAX_NR_ZONES; i++) {
8298 if (descending)
8299 zone = MAX_NR_ZONES - i - 1;
8300 else
8301 zone = i;
8302
8303 if (zone == ZONE_MOVABLE)
8304 continue;
8305
8306 end_pfn = max(max_zone_pfn[zone], start_pfn);
8307 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8308 arch_zone_highest_possible_pfn[zone] = end_pfn;
8309
8310 start_pfn = end_pfn;
8311 }
8312
8313 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8314 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8315 find_zone_movable_pfns_for_nodes();
8316
8317 /* Print out the zone ranges */
8318 pr_info("Zone ranges:\n");
8319 for (i = 0; i < MAX_NR_ZONES; i++) {
8320 if (i == ZONE_MOVABLE)
8321 continue;
8322 pr_info(" %-8s ", zone_names[i]);
8323 if (arch_zone_lowest_possible_pfn[i] ==
8324 arch_zone_highest_possible_pfn[i])
8325 pr_cont("empty\n");
8326 else
8327 pr_cont("[mem %#018Lx-%#018Lx]\n",
8328 (u64)arch_zone_lowest_possible_pfn[i]
8329 << PAGE_SHIFT,
8330 ((u64)arch_zone_highest_possible_pfn[i]
8331 << PAGE_SHIFT) - 1);
8332 }
8333
8334 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8335 pr_info("Movable zone start for each node\n");
8336 for (i = 0; i < MAX_NUMNODES; i++) {
8337 if (zone_movable_pfn[i])
8338 pr_info(" Node %d: %#018Lx\n", i,
8339 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8340 }
8341
8342 /*
8343 * Print out the early node map, and initialize the
8344 * subsection-map relative to active online memory ranges to
8345 * enable future "sub-section" extensions of the memory map.
8346 */
8347 pr_info("Early memory node ranges\n");
8348 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8349 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8350 (u64)start_pfn << PAGE_SHIFT,
8351 ((u64)end_pfn << PAGE_SHIFT) - 1);
8352 subsection_map_init(start_pfn, end_pfn - start_pfn);
8353 }
8354
8355 /* Initialise every node */
8356 mminit_verify_pageflags_layout();
8357 setup_nr_node_ids();
8358 for_each_node(nid) {
8359 pg_data_t *pgdat;
8360
8361 if (!node_online(nid)) {
8362 pr_info("Initializing node %d as memoryless\n", nid);
8363
8364 /* Allocator not initialized yet */
8365 pgdat = arch_alloc_nodedata(nid);
8366 if (!pgdat) {
8367 pr_err("Cannot allocate %zuB for node %d.\n",
8368 sizeof(*pgdat), nid);
8369 continue;
8370 }
8371 arch_refresh_nodedata(nid, pgdat);
8372 free_area_init_memoryless_node(nid);
8373
8374 /*
8375 * We do not want to confuse userspace by sysfs
8376 * files/directories for node without any memory
8377 * attached to it, so this node is not marked as
8378 * N_MEMORY and not marked online so that no sysfs
8379 * hierarchy will be created via register_one_node for
8380 * it. The pgdat will get fully initialized by
8381 * hotadd_init_pgdat() when memory is hotplugged into
8382 * this node.
8383 */
8384 continue;
8385 }
8386
8387 pgdat = NODE_DATA(nid);
8388 free_area_init_node(nid);
8389
8390 /* Any memory on that node */
8391 if (pgdat->node_present_pages)
8392 node_set_state(nid, N_MEMORY);
8393 check_for_memory(pgdat, nid);
8394 }
8395
8396 memmap_init();
8397}
8398
8399static int __init cmdline_parse_core(char *p, unsigned long *core,
8400 unsigned long *percent)
8401{
8402 unsigned long long coremem;
8403 char *endptr;
8404
8405 if (!p)
8406 return -EINVAL;
8407
8408 /* Value may be a percentage of total memory, otherwise bytes */
8409 coremem = simple_strtoull(p, &endptr, 0);
8410 if (*endptr == '%') {
8411 /* Paranoid check for percent values greater than 100 */
8412 WARN_ON(coremem > 100);
8413
8414 *percent = coremem;
8415 } else {
8416 coremem = memparse(p, &p);
8417 /* Paranoid check that UL is enough for the coremem value */
8418 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8419
8420 *core = coremem >> PAGE_SHIFT;
8421 *percent = 0UL;
8422 }
8423 return 0;
8424}
8425
8426/*
8427 * kernelcore=size sets the amount of memory for use for allocations that
8428 * cannot be reclaimed or migrated.
8429 */
8430static int __init cmdline_parse_kernelcore(char *p)
8431{
8432 /* parse kernelcore=mirror */
8433 if (parse_option_str(p, "mirror")) {
8434 mirrored_kernelcore = true;
8435 return 0;
8436 }
8437
8438 return cmdline_parse_core(p, &required_kernelcore,
8439 &required_kernelcore_percent);
8440}
8441
8442/*
8443 * movablecore=size sets the amount of memory for use for allocations that
8444 * can be reclaimed or migrated.
8445 */
8446static int __init cmdline_parse_movablecore(char *p)
8447{
8448 return cmdline_parse_core(p, &required_movablecore,
8449 &required_movablecore_percent);
8450}
8451
8452early_param("kernelcore", cmdline_parse_kernelcore);
8453early_param("movablecore", cmdline_parse_movablecore);
8454
8455void adjust_managed_page_count(struct page *page, long count)
8456{
8457 atomic_long_add(count, &page_zone(page)->managed_pages);
8458 totalram_pages_add(count);
8459#ifdef CONFIG_HIGHMEM
8460 if (PageHighMem(page))
8461 totalhigh_pages_add(count);
8462#endif
8463}
8464EXPORT_SYMBOL(adjust_managed_page_count);
8465
8466unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8467{
8468 void *pos;
8469 unsigned long pages = 0;
8470
8471 start = (void *)PAGE_ALIGN((unsigned long)start);
8472 end = (void *)((unsigned long)end & PAGE_MASK);
8473 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8474 struct page *page = virt_to_page(pos);
8475 void *direct_map_addr;
8476
8477 /*
8478 * 'direct_map_addr' might be different from 'pos'
8479 * because some architectures' virt_to_page()
8480 * work with aliases. Getting the direct map
8481 * address ensures that we get a _writeable_
8482 * alias for the memset().
8483 */
8484 direct_map_addr = page_address(page);
8485 /*
8486 * Perform a kasan-unchecked memset() since this memory
8487 * has not been initialized.
8488 */
8489 direct_map_addr = kasan_reset_tag(direct_map_addr);
8490 if ((unsigned int)poison <= 0xFF)
8491 memset(direct_map_addr, poison, PAGE_SIZE);
8492
8493 free_reserved_page(page);
8494 }
8495
8496 if (pages && s)
8497 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8498
8499 return pages;
8500}
8501
8502void __init mem_init_print_info(void)
8503{
8504 unsigned long physpages, codesize, datasize, rosize, bss_size;
8505 unsigned long init_code_size, init_data_size;
8506
8507 physpages = get_num_physpages();
8508 codesize = _etext - _stext;
8509 datasize = _edata - _sdata;
8510 rosize = __end_rodata - __start_rodata;
8511 bss_size = __bss_stop - __bss_start;
8512 init_data_size = __init_end - __init_begin;
8513 init_code_size = _einittext - _sinittext;
8514
8515 /*
8516 * Detect special cases and adjust section sizes accordingly:
8517 * 1) .init.* may be embedded into .data sections
8518 * 2) .init.text.* may be out of [__init_begin, __init_end],
8519 * please refer to arch/tile/kernel/vmlinux.lds.S.
8520 * 3) .rodata.* may be embedded into .text or .data sections.
8521 */
8522#define adj_init_size(start, end, size, pos, adj) \
8523 do { \
8524 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8525 size -= adj; \
8526 } while (0)
8527
8528 adj_init_size(__init_begin, __init_end, init_data_size,
8529 _sinittext, init_code_size);
8530 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8531 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8532 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8533 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8534
8535#undef adj_init_size
8536
8537 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8538#ifdef CONFIG_HIGHMEM
8539 ", %luK highmem"
8540#endif
8541 ")\n",
8542 K(nr_free_pages()), K(physpages),
8543 codesize / SZ_1K, datasize / SZ_1K, rosize / SZ_1K,
8544 (init_data_size + init_code_size) / SZ_1K, bss_size / SZ_1K,
8545 K(physpages - totalram_pages() - totalcma_pages),
8546 K(totalcma_pages)
8547#ifdef CONFIG_HIGHMEM
8548 , K(totalhigh_pages())
8549#endif
8550 );
8551}
8552
8553/**
8554 * set_dma_reserve - set the specified number of pages reserved in the first zone
8555 * @new_dma_reserve: The number of pages to mark reserved
8556 *
8557 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8558 * In the DMA zone, a significant percentage may be consumed by kernel image
8559 * and other unfreeable allocations which can skew the watermarks badly. This
8560 * function may optionally be used to account for unfreeable pages in the
8561 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8562 * smaller per-cpu batchsize.
8563 */
8564void __init set_dma_reserve(unsigned long new_dma_reserve)
8565{
8566 dma_reserve = new_dma_reserve;
8567}
8568
8569static int page_alloc_cpu_dead(unsigned int cpu)
8570{
8571 struct zone *zone;
8572
8573 lru_add_drain_cpu(cpu);
8574 mlock_page_drain_remote(cpu);
8575 drain_pages(cpu);
8576
8577 /*
8578 * Spill the event counters of the dead processor
8579 * into the current processors event counters.
8580 * This artificially elevates the count of the current
8581 * processor.
8582 */
8583 vm_events_fold_cpu(cpu);
8584
8585 /*
8586 * Zero the differential counters of the dead processor
8587 * so that the vm statistics are consistent.
8588 *
8589 * This is only okay since the processor is dead and cannot
8590 * race with what we are doing.
8591 */
8592 cpu_vm_stats_fold(cpu);
8593
8594 for_each_populated_zone(zone)
8595 zone_pcp_update(zone, 0);
8596
8597 return 0;
8598}
8599
8600static int page_alloc_cpu_online(unsigned int cpu)
8601{
8602 struct zone *zone;
8603
8604 for_each_populated_zone(zone)
8605 zone_pcp_update(zone, 1);
8606 return 0;
8607}
8608
8609#ifdef CONFIG_NUMA
8610int hashdist = HASHDIST_DEFAULT;
8611
8612static int __init set_hashdist(char *str)
8613{
8614 if (!str)
8615 return 0;
8616 hashdist = simple_strtoul(str, &str, 0);
8617 return 1;
8618}
8619__setup("hashdist=", set_hashdist);
8620#endif
8621
8622void __init page_alloc_init(void)
8623{
8624 int ret;
8625
8626#ifdef CONFIG_NUMA
8627 if (num_node_state(N_MEMORY) == 1)
8628 hashdist = 0;
8629#endif
8630
8631 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8632 "mm/page_alloc:pcp",
8633 page_alloc_cpu_online,
8634 page_alloc_cpu_dead);
8635 WARN_ON(ret < 0);
8636}
8637
8638/*
8639 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8640 * or min_free_kbytes changes.
8641 */
8642static void calculate_totalreserve_pages(void)
8643{
8644 struct pglist_data *pgdat;
8645 unsigned long reserve_pages = 0;
8646 enum zone_type i, j;
8647
8648 for_each_online_pgdat(pgdat) {
8649
8650 pgdat->totalreserve_pages = 0;
8651
8652 for (i = 0; i < MAX_NR_ZONES; i++) {
8653 struct zone *zone = pgdat->node_zones + i;
8654 long max = 0;
8655 unsigned long managed_pages = zone_managed_pages(zone);
8656
8657 /* Find valid and maximum lowmem_reserve in the zone */
8658 for (j = i; j < MAX_NR_ZONES; j++) {
8659 if (zone->lowmem_reserve[j] > max)
8660 max = zone->lowmem_reserve[j];
8661 }
8662
8663 /* we treat the high watermark as reserved pages. */
8664 max += high_wmark_pages(zone);
8665
8666 if (max > managed_pages)
8667 max = managed_pages;
8668
8669 pgdat->totalreserve_pages += max;
8670
8671 reserve_pages += max;
8672 }
8673 }
8674 totalreserve_pages = reserve_pages;
8675}
8676
8677/*
8678 * setup_per_zone_lowmem_reserve - called whenever
8679 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8680 * has a correct pages reserved value, so an adequate number of
8681 * pages are left in the zone after a successful __alloc_pages().
8682 */
8683static void setup_per_zone_lowmem_reserve(void)
8684{
8685 struct pglist_data *pgdat;
8686 enum zone_type i, j;
8687
8688 for_each_online_pgdat(pgdat) {
8689 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8690 struct zone *zone = &pgdat->node_zones[i];
8691 int ratio = sysctl_lowmem_reserve_ratio[i];
8692 bool clear = !ratio || !zone_managed_pages(zone);
8693 unsigned long managed_pages = 0;
8694
8695 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8696 struct zone *upper_zone = &pgdat->node_zones[j];
8697
8698 managed_pages += zone_managed_pages(upper_zone);
8699
8700 if (clear)
8701 zone->lowmem_reserve[j] = 0;
8702 else
8703 zone->lowmem_reserve[j] = managed_pages / ratio;
8704 }
8705 }
8706 }
8707
8708 /* update totalreserve_pages */
8709 calculate_totalreserve_pages();
8710}
8711
8712static void __setup_per_zone_wmarks(void)
8713{
8714 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8715 unsigned long lowmem_pages = 0;
8716 struct zone *zone;
8717 unsigned long flags;
8718
8719 /* Calculate total number of !ZONE_HIGHMEM pages */
8720 for_each_zone(zone) {
8721 if (!is_highmem(zone))
8722 lowmem_pages += zone_managed_pages(zone);
8723 }
8724
8725 for_each_zone(zone) {
8726 u64 tmp;
8727
8728 spin_lock_irqsave(&zone->lock, flags);
8729 tmp = (u64)pages_min * zone_managed_pages(zone);
8730 do_div(tmp, lowmem_pages);
8731 if (is_highmem(zone)) {
8732 /*
8733 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8734 * need highmem pages, so cap pages_min to a small
8735 * value here.
8736 *
8737 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8738 * deltas control async page reclaim, and so should
8739 * not be capped for highmem.
8740 */
8741 unsigned long min_pages;
8742
8743 min_pages = zone_managed_pages(zone) / 1024;
8744 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8745 zone->_watermark[WMARK_MIN] = min_pages;
8746 } else {
8747 /*
8748 * If it's a lowmem zone, reserve a number of pages
8749 * proportionate to the zone's size.
8750 */
8751 zone->_watermark[WMARK_MIN] = tmp;
8752 }
8753
8754 /*
8755 * Set the kswapd watermarks distance according to the
8756 * scale factor in proportion to available memory, but
8757 * ensure a minimum size on small systems.
8758 */
8759 tmp = max_t(u64, tmp >> 2,
8760 mult_frac(zone_managed_pages(zone),
8761 watermark_scale_factor, 10000));
8762
8763 zone->watermark_boost = 0;
8764 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8765 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8766 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8767
8768 spin_unlock_irqrestore(&zone->lock, flags);
8769 }
8770
8771 /* update totalreserve_pages */
8772 calculate_totalreserve_pages();
8773}
8774
8775/**
8776 * setup_per_zone_wmarks - called when min_free_kbytes changes
8777 * or when memory is hot-{added|removed}
8778 *
8779 * Ensures that the watermark[min,low,high] values for each zone are set
8780 * correctly with respect to min_free_kbytes.
8781 */
8782void setup_per_zone_wmarks(void)
8783{
8784 struct zone *zone;
8785 static DEFINE_SPINLOCK(lock);
8786
8787 spin_lock(&lock);
8788 __setup_per_zone_wmarks();
8789 spin_unlock(&lock);
8790
8791 /*
8792 * The watermark size have changed so update the pcpu batch
8793 * and high limits or the limits may be inappropriate.
8794 */
8795 for_each_zone(zone)
8796 zone_pcp_update(zone, 0);
8797}
8798
8799/*
8800 * Initialise min_free_kbytes.
8801 *
8802 * For small machines we want it small (128k min). For large machines
8803 * we want it large (256MB max). But it is not linear, because network
8804 * bandwidth does not increase linearly with machine size. We use
8805 *
8806 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8807 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8808 *
8809 * which yields
8810 *
8811 * 16MB: 512k
8812 * 32MB: 724k
8813 * 64MB: 1024k
8814 * 128MB: 1448k
8815 * 256MB: 2048k
8816 * 512MB: 2896k
8817 * 1024MB: 4096k
8818 * 2048MB: 5792k
8819 * 4096MB: 8192k
8820 * 8192MB: 11584k
8821 * 16384MB: 16384k
8822 */
8823void calculate_min_free_kbytes(void)
8824{
8825 unsigned long lowmem_kbytes;
8826 int new_min_free_kbytes;
8827
8828 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8829 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8830
8831 if (new_min_free_kbytes > user_min_free_kbytes)
8832 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8833 else
8834 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8835 new_min_free_kbytes, user_min_free_kbytes);
8836
8837}
8838
8839int __meminit init_per_zone_wmark_min(void)
8840{
8841 calculate_min_free_kbytes();
8842 setup_per_zone_wmarks();
8843 refresh_zone_stat_thresholds();
8844 setup_per_zone_lowmem_reserve();
8845
8846#ifdef CONFIG_NUMA
8847 setup_min_unmapped_ratio();
8848 setup_min_slab_ratio();
8849#endif
8850
8851 khugepaged_min_free_kbytes_update();
8852
8853 return 0;
8854}
8855postcore_initcall(init_per_zone_wmark_min)
8856
8857/*
8858 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8859 * that we can call two helper functions whenever min_free_kbytes
8860 * changes.
8861 */
8862int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8863 void *buffer, size_t *length, loff_t *ppos)
8864{
8865 int rc;
8866
8867 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8868 if (rc)
8869 return rc;
8870
8871 if (write) {
8872 user_min_free_kbytes = min_free_kbytes;
8873 setup_per_zone_wmarks();
8874 }
8875 return 0;
8876}
8877
8878int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8879 void *buffer, size_t *length, loff_t *ppos)
8880{
8881 int rc;
8882
8883 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8884 if (rc)
8885 return rc;
8886
8887 if (write)
8888 setup_per_zone_wmarks();
8889
8890 return 0;
8891}
8892
8893#ifdef CONFIG_NUMA
8894static void setup_min_unmapped_ratio(void)
8895{
8896 pg_data_t *pgdat;
8897 struct zone *zone;
8898
8899 for_each_online_pgdat(pgdat)
8900 pgdat->min_unmapped_pages = 0;
8901
8902 for_each_zone(zone)
8903 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8904 sysctl_min_unmapped_ratio) / 100;
8905}
8906
8907
8908int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8909 void *buffer, size_t *length, loff_t *ppos)
8910{
8911 int rc;
8912
8913 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8914 if (rc)
8915 return rc;
8916
8917 setup_min_unmapped_ratio();
8918
8919 return 0;
8920}
8921
8922static void setup_min_slab_ratio(void)
8923{
8924 pg_data_t *pgdat;
8925 struct zone *zone;
8926
8927 for_each_online_pgdat(pgdat)
8928 pgdat->min_slab_pages = 0;
8929
8930 for_each_zone(zone)
8931 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8932 sysctl_min_slab_ratio) / 100;
8933}
8934
8935int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8936 void *buffer, size_t *length, loff_t *ppos)
8937{
8938 int rc;
8939
8940 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8941 if (rc)
8942 return rc;
8943
8944 setup_min_slab_ratio();
8945
8946 return 0;
8947}
8948#endif
8949
8950/*
8951 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8952 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8953 * whenever sysctl_lowmem_reserve_ratio changes.
8954 *
8955 * The reserve ratio obviously has absolutely no relation with the
8956 * minimum watermarks. The lowmem reserve ratio can only make sense
8957 * if in function of the boot time zone sizes.
8958 */
8959int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8960 void *buffer, size_t *length, loff_t *ppos)
8961{
8962 int i;
8963
8964 proc_dointvec_minmax(table, write, buffer, length, ppos);
8965
8966 for (i = 0; i < MAX_NR_ZONES; i++) {
8967 if (sysctl_lowmem_reserve_ratio[i] < 1)
8968 sysctl_lowmem_reserve_ratio[i] = 0;
8969 }
8970
8971 setup_per_zone_lowmem_reserve();
8972 return 0;
8973}
8974
8975/*
8976 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8977 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8978 * pagelist can have before it gets flushed back to buddy allocator.
8979 */
8980int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8981 int write, void *buffer, size_t *length, loff_t *ppos)
8982{
8983 struct zone *zone;
8984 int old_percpu_pagelist_high_fraction;
8985 int ret;
8986
8987 mutex_lock(&pcp_batch_high_lock);
8988 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8989
8990 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8991 if (!write || ret < 0)
8992 goto out;
8993
8994 /* Sanity checking to avoid pcp imbalance */
8995 if (percpu_pagelist_high_fraction &&
8996 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8997 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8998 ret = -EINVAL;
8999 goto out;
9000 }
9001
9002 /* No change? */
9003 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
9004 goto out;
9005
9006 for_each_populated_zone(zone)
9007 zone_set_pageset_high_and_batch(zone, 0);
9008out:
9009 mutex_unlock(&pcp_batch_high_lock);
9010 return ret;
9011}
9012
9013#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
9014/*
9015 * Returns the number of pages that arch has reserved but
9016 * is not known to alloc_large_system_hash().
9017 */
9018static unsigned long __init arch_reserved_kernel_pages(void)
9019{
9020 return 0;
9021}
9022#endif
9023
9024/*
9025 * Adaptive scale is meant to reduce sizes of hash tables on large memory
9026 * machines. As memory size is increased the scale is also increased but at
9027 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
9028 * quadruples the scale is increased by one, which means the size of hash table
9029 * only doubles, instead of quadrupling as well.
9030 * Because 32-bit systems cannot have large physical memory, where this scaling
9031 * makes sense, it is disabled on such platforms.
9032 */
9033#if __BITS_PER_LONG > 32
9034#define ADAPT_SCALE_BASE (64ul << 30)
9035#define ADAPT_SCALE_SHIFT 2
9036#define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
9037#endif
9038
9039/*
9040 * allocate a large system hash table from bootmem
9041 * - it is assumed that the hash table must contain an exact power-of-2
9042 * quantity of entries
9043 * - limit is the number of hash buckets, not the total allocation size
9044 */
9045void *__init alloc_large_system_hash(const char *tablename,
9046 unsigned long bucketsize,
9047 unsigned long numentries,
9048 int scale,
9049 int flags,
9050 unsigned int *_hash_shift,
9051 unsigned int *_hash_mask,
9052 unsigned long low_limit,
9053 unsigned long high_limit)
9054{
9055 unsigned long long max = high_limit;
9056 unsigned long log2qty, size;
9057 void *table;
9058 gfp_t gfp_flags;
9059 bool virt;
9060 bool huge;
9061
9062 /* allow the kernel cmdline to have a say */
9063 if (!numentries) {
9064 /* round applicable memory size up to nearest megabyte */
9065 numentries = nr_kernel_pages;
9066 numentries -= arch_reserved_kernel_pages();
9067
9068 /* It isn't necessary when PAGE_SIZE >= 1MB */
9069 if (PAGE_SIZE < SZ_1M)
9070 numentries = round_up(numentries, SZ_1M / PAGE_SIZE);
9071
9072#if __BITS_PER_LONG > 32
9073 if (!high_limit) {
9074 unsigned long adapt;
9075
9076 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
9077 adapt <<= ADAPT_SCALE_SHIFT)
9078 scale++;
9079 }
9080#endif
9081
9082 /* limit to 1 bucket per 2^scale bytes of low memory */
9083 if (scale > PAGE_SHIFT)
9084 numentries >>= (scale - PAGE_SHIFT);
9085 else
9086 numentries <<= (PAGE_SHIFT - scale);
9087
9088 /* Make sure we've got at least a 0-order allocation.. */
9089 if (unlikely(flags & HASH_SMALL)) {
9090 /* Makes no sense without HASH_EARLY */
9091 WARN_ON(!(flags & HASH_EARLY));
9092 if (!(numentries >> *_hash_shift)) {
9093 numentries = 1UL << *_hash_shift;
9094 BUG_ON(!numentries);
9095 }
9096 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
9097 numentries = PAGE_SIZE / bucketsize;
9098 }
9099 numentries = roundup_pow_of_two(numentries);
9100
9101 /* limit allocation size to 1/16 total memory by default */
9102 if (max == 0) {
9103 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
9104 do_div(max, bucketsize);
9105 }
9106 max = min(max, 0x80000000ULL);
9107
9108 if (numentries < low_limit)
9109 numentries = low_limit;
9110 if (numentries > max)
9111 numentries = max;
9112
9113 log2qty = ilog2(numentries);
9114
9115 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
9116 do {
9117 virt = false;
9118 size = bucketsize << log2qty;
9119 if (flags & HASH_EARLY) {
9120 if (flags & HASH_ZERO)
9121 table = memblock_alloc(size, SMP_CACHE_BYTES);
9122 else
9123 table = memblock_alloc_raw(size,
9124 SMP_CACHE_BYTES);
9125 } else if (get_order(size) >= MAX_ORDER || hashdist) {
9126 table = vmalloc_huge(size, gfp_flags);
9127 virt = true;
9128 if (table)
9129 huge = is_vm_area_hugepages(table);
9130 } else {
9131 /*
9132 * If bucketsize is not a power-of-two, we may free
9133 * some pages at the end of hash table which
9134 * alloc_pages_exact() automatically does
9135 */
9136 table = alloc_pages_exact(size, gfp_flags);
9137 kmemleak_alloc(table, size, 1, gfp_flags);
9138 }
9139 } while (!table && size > PAGE_SIZE && --log2qty);
9140
9141 if (!table)
9142 panic("Failed to allocate %s hash table\n", tablename);
9143
9144 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
9145 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
9146 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
9147
9148 if (_hash_shift)
9149 *_hash_shift = log2qty;
9150 if (_hash_mask)
9151 *_hash_mask = (1 << log2qty) - 1;
9152
9153 return table;
9154}
9155
9156#ifdef CONFIG_CONTIG_ALLOC
9157#if defined(CONFIG_DYNAMIC_DEBUG) || \
9158 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
9159/* Usage: See admin-guide/dynamic-debug-howto.rst */
9160static void alloc_contig_dump_pages(struct list_head *page_list)
9161{
9162 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9163
9164 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9165 struct page *page;
9166
9167 dump_stack();
9168 list_for_each_entry(page, page_list, lru)
9169 dump_page(page, "migration failure");
9170 }
9171}
9172#else
9173static inline void alloc_contig_dump_pages(struct list_head *page_list)
9174{
9175}
9176#endif
9177
9178/* [start, end) must belong to a single zone. */
9179int __alloc_contig_migrate_range(struct compact_control *cc,
9180 unsigned long start, unsigned long end)
9181{
9182 /* This function is based on compact_zone() from compaction.c. */
9183 unsigned int nr_reclaimed;
9184 unsigned long pfn = start;
9185 unsigned int tries = 0;
9186 int ret = 0;
9187 struct migration_target_control mtc = {
9188 .nid = zone_to_nid(cc->zone),
9189 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9190 };
9191
9192 lru_cache_disable();
9193
9194 while (pfn < end || !list_empty(&cc->migratepages)) {
9195 if (fatal_signal_pending(current)) {
9196 ret = -EINTR;
9197 break;
9198 }
9199
9200 if (list_empty(&cc->migratepages)) {
9201 cc->nr_migratepages = 0;
9202 ret = isolate_migratepages_range(cc, pfn, end);
9203 if (ret && ret != -EAGAIN)
9204 break;
9205 pfn = cc->migrate_pfn;
9206 tries = 0;
9207 } else if (++tries == 5) {
9208 ret = -EBUSY;
9209 break;
9210 }
9211
9212 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9213 &cc->migratepages);
9214 cc->nr_migratepages -= nr_reclaimed;
9215
9216 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9217 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9218
9219 /*
9220 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9221 * to retry again over this error, so do the same here.
9222 */
9223 if (ret == -ENOMEM)
9224 break;
9225 }
9226
9227 lru_cache_enable();
9228 if (ret < 0) {
9229 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
9230 alloc_contig_dump_pages(&cc->migratepages);
9231 putback_movable_pages(&cc->migratepages);
9232 return ret;
9233 }
9234 return 0;
9235}
9236
9237/**
9238 * alloc_contig_range() -- tries to allocate given range of pages
9239 * @start: start PFN to allocate
9240 * @end: one-past-the-last PFN to allocate
9241 * @migratetype: migratetype of the underlying pageblocks (either
9242 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9243 * in range must have the same migratetype and it must
9244 * be either of the two.
9245 * @gfp_mask: GFP mask to use during compaction
9246 *
9247 * The PFN range does not have to be pageblock aligned. The PFN range must
9248 * belong to a single zone.
9249 *
9250 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9251 * pageblocks in the range. Once isolated, the pageblocks should not
9252 * be modified by others.
9253 *
9254 * Return: zero on success or negative error code. On success all
9255 * pages which PFN is in [start, end) are allocated for the caller and
9256 * need to be freed with free_contig_range().
9257 */
9258int alloc_contig_range(unsigned long start, unsigned long end,
9259 unsigned migratetype, gfp_t gfp_mask)
9260{
9261 unsigned long outer_start, outer_end;
9262 int order;
9263 int ret = 0;
9264
9265 struct compact_control cc = {
9266 .nr_migratepages = 0,
9267 .order = -1,
9268 .zone = page_zone(pfn_to_page(start)),
9269 .mode = MIGRATE_SYNC,
9270 .ignore_skip_hint = true,
9271 .no_set_skip_hint = true,
9272 .gfp_mask = current_gfp_context(gfp_mask),
9273 .alloc_contig = true,
9274 };
9275 INIT_LIST_HEAD(&cc.migratepages);
9276
9277 /*
9278 * What we do here is we mark all pageblocks in range as
9279 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9280 * have different sizes, and due to the way page allocator
9281 * work, start_isolate_page_range() has special handlings for this.
9282 *
9283 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9284 * migrate the pages from an unaligned range (ie. pages that
9285 * we are interested in). This will put all the pages in
9286 * range back to page allocator as MIGRATE_ISOLATE.
9287 *
9288 * When this is done, we take the pages in range from page
9289 * allocator removing them from the buddy system. This way
9290 * page allocator will never consider using them.
9291 *
9292 * This lets us mark the pageblocks back as
9293 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9294 * aligned range but not in the unaligned, original range are
9295 * put back to page allocator so that buddy can use them.
9296 */
9297
9298 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
9299 if (ret)
9300 goto done;
9301
9302 drain_all_pages(cc.zone);
9303
9304 /*
9305 * In case of -EBUSY, we'd like to know which page causes problem.
9306 * So, just fall through. test_pages_isolated() has a tracepoint
9307 * which will report the busy page.
9308 *
9309 * It is possible that busy pages could become available before
9310 * the call to test_pages_isolated, and the range will actually be
9311 * allocated. So, if we fall through be sure to clear ret so that
9312 * -EBUSY is not accidentally used or returned to caller.
9313 */
9314 ret = __alloc_contig_migrate_range(&cc, start, end);
9315 if (ret && ret != -EBUSY)
9316 goto done;
9317 ret = 0;
9318
9319 /*
9320 * Pages from [start, end) are within a pageblock_nr_pages
9321 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9322 * more, all pages in [start, end) are free in page allocator.
9323 * What we are going to do is to allocate all pages from
9324 * [start, end) (that is remove them from page allocator).
9325 *
9326 * The only problem is that pages at the beginning and at the
9327 * end of interesting range may be not aligned with pages that
9328 * page allocator holds, ie. they can be part of higher order
9329 * pages. Because of this, we reserve the bigger range and
9330 * once this is done free the pages we are not interested in.
9331 *
9332 * We don't have to hold zone->lock here because the pages are
9333 * isolated thus they won't get removed from buddy.
9334 */
9335
9336 order = 0;
9337 outer_start = start;
9338 while (!PageBuddy(pfn_to_page(outer_start))) {
9339 if (++order >= MAX_ORDER) {
9340 outer_start = start;
9341 break;
9342 }
9343 outer_start &= ~0UL << order;
9344 }
9345
9346 if (outer_start != start) {
9347 order = buddy_order(pfn_to_page(outer_start));
9348
9349 /*
9350 * outer_start page could be small order buddy page and
9351 * it doesn't include start page. Adjust outer_start
9352 * in this case to report failed page properly
9353 * on tracepoint in test_pages_isolated()
9354 */
9355 if (outer_start + (1UL << order) <= start)
9356 outer_start = start;
9357 }
9358
9359 /* Make sure the range is really isolated. */
9360 if (test_pages_isolated(outer_start, end, 0)) {
9361 ret = -EBUSY;
9362 goto done;
9363 }
9364
9365 /* Grab isolated pages from freelists. */
9366 outer_end = isolate_freepages_range(&cc, outer_start, end);
9367 if (!outer_end) {
9368 ret = -EBUSY;
9369 goto done;
9370 }
9371
9372 /* Free head and tail (if any) */
9373 if (start != outer_start)
9374 free_contig_range(outer_start, start - outer_start);
9375 if (end != outer_end)
9376 free_contig_range(end, outer_end - end);
9377
9378done:
9379 undo_isolate_page_range(start, end, migratetype);
9380 return ret;
9381}
9382EXPORT_SYMBOL(alloc_contig_range);
9383
9384static int __alloc_contig_pages(unsigned long start_pfn,
9385 unsigned long nr_pages, gfp_t gfp_mask)
9386{
9387 unsigned long end_pfn = start_pfn + nr_pages;
9388
9389 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9390 gfp_mask);
9391}
9392
9393static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9394 unsigned long nr_pages)
9395{
9396 unsigned long i, end_pfn = start_pfn + nr_pages;
9397 struct page *page;
9398
9399 for (i = start_pfn; i < end_pfn; i++) {
9400 page = pfn_to_online_page(i);
9401 if (!page)
9402 return false;
9403
9404 if (page_zone(page) != z)
9405 return false;
9406
9407 if (PageReserved(page))
9408 return false;
9409 }
9410 return true;
9411}
9412
9413static bool zone_spans_last_pfn(const struct zone *zone,
9414 unsigned long start_pfn, unsigned long nr_pages)
9415{
9416 unsigned long last_pfn = start_pfn + nr_pages - 1;
9417
9418 return zone_spans_pfn(zone, last_pfn);
9419}
9420
9421/**
9422 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9423 * @nr_pages: Number of contiguous pages to allocate
9424 * @gfp_mask: GFP mask to limit search and used during compaction
9425 * @nid: Target node
9426 * @nodemask: Mask for other possible nodes
9427 *
9428 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9429 * on an applicable zonelist to find a contiguous pfn range which can then be
9430 * tried for allocation with alloc_contig_range(). This routine is intended
9431 * for allocation requests which can not be fulfilled with the buddy allocator.
9432 *
9433 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9434 * power of two, then allocated range is also guaranteed to be aligned to same
9435 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9436 *
9437 * Allocated pages can be freed with free_contig_range() or by manually calling
9438 * __free_page() on each allocated page.
9439 *
9440 * Return: pointer to contiguous pages on success, or NULL if not successful.
9441 */
9442struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9443 int nid, nodemask_t *nodemask)
9444{
9445 unsigned long ret, pfn, flags;
9446 struct zonelist *zonelist;
9447 struct zone *zone;
9448 struct zoneref *z;
9449
9450 zonelist = node_zonelist(nid, gfp_mask);
9451 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9452 gfp_zone(gfp_mask), nodemask) {
9453 spin_lock_irqsave(&zone->lock, flags);
9454
9455 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9456 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9457 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9458 /*
9459 * We release the zone lock here because
9460 * alloc_contig_range() will also lock the zone
9461 * at some point. If there's an allocation
9462 * spinning on this lock, it may win the race
9463 * and cause alloc_contig_range() to fail...
9464 */
9465 spin_unlock_irqrestore(&zone->lock, flags);
9466 ret = __alloc_contig_pages(pfn, nr_pages,
9467 gfp_mask);
9468 if (!ret)
9469 return pfn_to_page(pfn);
9470 spin_lock_irqsave(&zone->lock, flags);
9471 }
9472 pfn += nr_pages;
9473 }
9474 spin_unlock_irqrestore(&zone->lock, flags);
9475 }
9476 return NULL;
9477}
9478#endif /* CONFIG_CONTIG_ALLOC */
9479
9480void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9481{
9482 unsigned long count = 0;
9483
9484 for (; nr_pages--; pfn++) {
9485 struct page *page = pfn_to_page(pfn);
9486
9487 count += page_count(page) != 1;
9488 __free_page(page);
9489 }
9490 WARN(count != 0, "%lu pages are still in use!\n", count);
9491}
9492EXPORT_SYMBOL(free_contig_range);
9493
9494/*
9495 * Effectively disable pcplists for the zone by setting the high limit to 0
9496 * and draining all cpus. A concurrent page freeing on another CPU that's about
9497 * to put the page on pcplist will either finish before the drain and the page
9498 * will be drained, or observe the new high limit and skip the pcplist.
9499 *
9500 * Must be paired with a call to zone_pcp_enable().
9501 */
9502void zone_pcp_disable(struct zone *zone)
9503{
9504 mutex_lock(&pcp_batch_high_lock);
9505 __zone_set_pageset_high_and_batch(zone, 0, 1);
9506 __drain_all_pages(zone, true);
9507}
9508
9509void zone_pcp_enable(struct zone *zone)
9510{
9511 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9512 mutex_unlock(&pcp_batch_high_lock);
9513}
9514
9515void zone_pcp_reset(struct zone *zone)
9516{
9517 int cpu;
9518 struct per_cpu_zonestat *pzstats;
9519
9520 if (zone->per_cpu_pageset != &boot_pageset) {
9521 for_each_online_cpu(cpu) {
9522 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9523 drain_zonestat(zone, pzstats);
9524 }
9525 free_percpu(zone->per_cpu_pageset);
9526 zone->per_cpu_pageset = &boot_pageset;
9527 if (zone->per_cpu_zonestats != &boot_zonestats) {
9528 free_percpu(zone->per_cpu_zonestats);
9529 zone->per_cpu_zonestats = &boot_zonestats;
9530 }
9531 }
9532}
9533
9534#ifdef CONFIG_MEMORY_HOTREMOVE
9535/*
9536 * All pages in the range must be in a single zone, must not contain holes,
9537 * must span full sections, and must be isolated before calling this function.
9538 */
9539void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9540{
9541 unsigned long pfn = start_pfn;
9542 struct page *page;
9543 struct zone *zone;
9544 unsigned int order;
9545 unsigned long flags;
9546
9547 offline_mem_sections(pfn, end_pfn);
9548 zone = page_zone(pfn_to_page(pfn));
9549 spin_lock_irqsave(&zone->lock, flags);
9550 while (pfn < end_pfn) {
9551 page = pfn_to_page(pfn);
9552 /*
9553 * The HWPoisoned page may be not in buddy system, and
9554 * page_count() is not 0.
9555 */
9556 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9557 pfn++;
9558 continue;
9559 }
9560 /*
9561 * At this point all remaining PageOffline() pages have a
9562 * reference count of 0 and can simply be skipped.
9563 */
9564 if (PageOffline(page)) {
9565 BUG_ON(page_count(page));
9566 BUG_ON(PageBuddy(page));
9567 pfn++;
9568 continue;
9569 }
9570
9571 BUG_ON(page_count(page));
9572 BUG_ON(!PageBuddy(page));
9573 order = buddy_order(page);
9574 del_page_from_free_list(page, zone, order);
9575 pfn += (1 << order);
9576 }
9577 spin_unlock_irqrestore(&zone->lock, flags);
9578}
9579#endif
9580
9581/*
9582 * This function returns a stable result only if called under zone lock.
9583 */
9584bool is_free_buddy_page(struct page *page)
9585{
9586 unsigned long pfn = page_to_pfn(page);
9587 unsigned int order;
9588
9589 for (order = 0; order < MAX_ORDER; order++) {
9590 struct page *page_head = page - (pfn & ((1 << order) - 1));
9591
9592 if (PageBuddy(page_head) &&
9593 buddy_order_unsafe(page_head) >= order)
9594 break;
9595 }
9596
9597 return order < MAX_ORDER;
9598}
9599EXPORT_SYMBOL(is_free_buddy_page);
9600
9601#ifdef CONFIG_MEMORY_FAILURE
9602/*
9603 * Break down a higher-order page in sub-pages, and keep our target out of
9604 * buddy allocator.
9605 */
9606static void break_down_buddy_pages(struct zone *zone, struct page *page,
9607 struct page *target, int low, int high,
9608 int migratetype)
9609{
9610 unsigned long size = 1 << high;
9611 struct page *current_buddy, *next_page;
9612
9613 while (high > low) {
9614 high--;
9615 size >>= 1;
9616
9617 if (target >= &page[size]) {
9618 next_page = page + size;
9619 current_buddy = page;
9620 } else {
9621 next_page = page;
9622 current_buddy = page + size;
9623 }
9624
9625 if (set_page_guard(zone, current_buddy, high, migratetype))
9626 continue;
9627
9628 if (current_buddy != target) {
9629 add_to_free_list(current_buddy, zone, high, migratetype);
9630 set_buddy_order(current_buddy, high);
9631 page = next_page;
9632 }
9633 }
9634}
9635
9636/*
9637 * Take a page that will be marked as poisoned off the buddy allocator.
9638 */
9639bool take_page_off_buddy(struct page *page)
9640{
9641 struct zone *zone = page_zone(page);
9642 unsigned long pfn = page_to_pfn(page);
9643 unsigned long flags;
9644 unsigned int order;
9645 bool ret = false;
9646
9647 spin_lock_irqsave(&zone->lock, flags);
9648 for (order = 0; order < MAX_ORDER; order++) {
9649 struct page *page_head = page - (pfn & ((1 << order) - 1));
9650 int page_order = buddy_order(page_head);
9651
9652 if (PageBuddy(page_head) && page_order >= order) {
9653 unsigned long pfn_head = page_to_pfn(page_head);
9654 int migratetype = get_pfnblock_migratetype(page_head,
9655 pfn_head);
9656
9657 del_page_from_free_list(page_head, zone, page_order);
9658 break_down_buddy_pages(zone, page_head, page, 0,
9659 page_order, migratetype);
9660 SetPageHWPoisonTakenOff(page);
9661 if (!is_migrate_isolate(migratetype))
9662 __mod_zone_freepage_state(zone, -1, migratetype);
9663 ret = true;
9664 break;
9665 }
9666 if (page_count(page_head) > 0)
9667 break;
9668 }
9669 spin_unlock_irqrestore(&zone->lock, flags);
9670 return ret;
9671}
9672
9673/*
9674 * Cancel takeoff done by take_page_off_buddy().
9675 */
9676bool put_page_back_buddy(struct page *page)
9677{
9678 struct zone *zone = page_zone(page);
9679 unsigned long pfn = page_to_pfn(page);
9680 unsigned long flags;
9681 int migratetype = get_pfnblock_migratetype(page, pfn);
9682 bool ret = false;
9683
9684 spin_lock_irqsave(&zone->lock, flags);
9685 if (put_page_testzero(page)) {
9686 ClearPageHWPoisonTakenOff(page);
9687 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9688 if (TestClearPageHWPoison(page)) {
9689 ret = true;
9690 }
9691 }
9692 spin_unlock_irqrestore(&zone->lock, flags);
9693
9694 return ret;
9695}
9696#endif
9697
9698#ifdef CONFIG_ZONE_DMA
9699bool has_managed_dma(void)
9700{
9701 struct pglist_data *pgdat;
9702
9703 for_each_online_pgdat(pgdat) {
9704 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9705
9706 if (managed_zone(zone))
9707 return true;
9708 }
9709 return false;
9710}
9711#endif /* CONFIG_ZONE_DMA */