Loading...
1/*
2 * linux/mm/page_alloc.c
3 *
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
6 *
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
15 */
16
17#include <linux/stddef.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/interrupt.h>
21#include <linux/pagemap.h>
22#include <linux/jiffies.h>
23#include <linux/bootmem.h>
24#include <linux/memblock.h>
25#include <linux/compiler.h>
26#include <linux/kernel.h>
27#include <linux/kasan.h>
28#include <linux/module.h>
29#include <linux/suspend.h>
30#include <linux/pagevec.h>
31#include <linux/blkdev.h>
32#include <linux/slab.h>
33#include <linux/ratelimit.h>
34#include <linux/oom.h>
35#include <linux/notifier.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/sort.h>
48#include <linux/pfn.h>
49#include <linux/backing-dev.h>
50#include <linux/fault-inject.h>
51#include <linux/page-isolation.h>
52#include <linux/page_ext.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/migrate.h>
61#include <linux/hugetlb.h>
62#include <linux/sched/rt.h>
63#include <linux/sched/mm.h>
64#include <linux/page_owner.h>
65#include <linux/kthread.h>
66#include <linux/memcontrol.h>
67#include <linux/ftrace.h>
68#include <linux/lockdep.h>
69#include <linux/nmi.h>
70
71#include <asm/sections.h>
72#include <asm/tlbflush.h>
73#include <asm/div64.h>
74#include "internal.h"
75
76/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77static DEFINE_MUTEX(pcp_batch_high_lock);
78#define MIN_PERCPU_PAGELIST_FRACTION (8)
79
80#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81DEFINE_PER_CPU(int, numa_node);
82EXPORT_PER_CPU_SYMBOL(numa_node);
83#endif
84
85DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
86
87#ifdef CONFIG_HAVE_MEMORYLESS_NODES
88/*
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
93 */
94DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96int _node_numa_mem_[MAX_NUMNODES];
97#endif
98
99/* work_structs for global per-cpu drains */
100DEFINE_MUTEX(pcpu_drain_mutex);
101DEFINE_PER_CPU(struct work_struct, pcpu_drain);
102
103#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104volatile unsigned long latent_entropy __latent_entropy;
105EXPORT_SYMBOL(latent_entropy);
106#endif
107
108/*
109 * Array of node states.
110 */
111nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
112 [N_POSSIBLE] = NODE_MASK_ALL,
113 [N_ONLINE] = { { [0] = 1UL } },
114#ifndef CONFIG_NUMA
115 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
116#ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY] = { { [0] = 1UL } },
118#endif
119 [N_MEMORY] = { { [0] = 1UL } },
120 [N_CPU] = { { [0] = 1UL } },
121#endif /* NUMA */
122};
123EXPORT_SYMBOL(node_states);
124
125/* Protect totalram_pages and zone->managed_pages */
126static DEFINE_SPINLOCK(managed_page_count_lock);
127
128unsigned long totalram_pages __read_mostly;
129unsigned long totalreserve_pages __read_mostly;
130unsigned long totalcma_pages __read_mostly;
131
132int percpu_pagelist_fraction;
133gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
134
135/*
136 * A cached value of the page's pageblock's migratetype, used when the page is
137 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
138 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
139 * Also the migratetype set in the page does not necessarily match the pcplist
140 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
141 * other index - this ensures that it will be put on the correct CMA freelist.
142 */
143static inline int get_pcppage_migratetype(struct page *page)
144{
145 return page->index;
146}
147
148static inline void set_pcppage_migratetype(struct page *page, int migratetype)
149{
150 page->index = migratetype;
151}
152
153#ifdef CONFIG_PM_SLEEP
154/*
155 * The following functions are used by the suspend/hibernate code to temporarily
156 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
157 * while devices are suspended. To avoid races with the suspend/hibernate code,
158 * they should always be called with pm_mutex held (gfp_allowed_mask also should
159 * only be modified with pm_mutex held, unless the suspend/hibernate code is
160 * guaranteed not to run in parallel with that modification).
161 */
162
163static gfp_t saved_gfp_mask;
164
165void pm_restore_gfp_mask(void)
166{
167 WARN_ON(!mutex_is_locked(&pm_mutex));
168 if (saved_gfp_mask) {
169 gfp_allowed_mask = saved_gfp_mask;
170 saved_gfp_mask = 0;
171 }
172}
173
174void pm_restrict_gfp_mask(void)
175{
176 WARN_ON(!mutex_is_locked(&pm_mutex));
177 WARN_ON(saved_gfp_mask);
178 saved_gfp_mask = gfp_allowed_mask;
179 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
180}
181
182bool pm_suspended_storage(void)
183{
184 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
185 return false;
186 return true;
187}
188#endif /* CONFIG_PM_SLEEP */
189
190#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
191unsigned int pageblock_order __read_mostly;
192#endif
193
194static void __free_pages_ok(struct page *page, unsigned int order);
195
196/*
197 * results with 256, 32 in the lowmem_reserve sysctl:
198 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
199 * 1G machine -> (16M dma, 784M normal, 224M high)
200 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
201 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
202 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 *
204 * TBD: should special case ZONE_DMA32 machines here - in those we normally
205 * don't need any ZONE_NORMAL reservation
206 */
207int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
208#ifdef CONFIG_ZONE_DMA
209 [ZONE_DMA] = 256,
210#endif
211#ifdef CONFIG_ZONE_DMA32
212 [ZONE_DMA32] = 256,
213#endif
214 [ZONE_NORMAL] = 32,
215#ifdef CONFIG_HIGHMEM
216 [ZONE_HIGHMEM] = 0,
217#endif
218 [ZONE_MOVABLE] = 0,
219};
220
221EXPORT_SYMBOL(totalram_pages);
222
223static char * const zone_names[MAX_NR_ZONES] = {
224#ifdef CONFIG_ZONE_DMA
225 "DMA",
226#endif
227#ifdef CONFIG_ZONE_DMA32
228 "DMA32",
229#endif
230 "Normal",
231#ifdef CONFIG_HIGHMEM
232 "HighMem",
233#endif
234 "Movable",
235#ifdef CONFIG_ZONE_DEVICE
236 "Device",
237#endif
238};
239
240char * const migratetype_names[MIGRATE_TYPES] = {
241 "Unmovable",
242 "Movable",
243 "Reclaimable",
244 "HighAtomic",
245#ifdef CONFIG_CMA
246 "CMA",
247#endif
248#ifdef CONFIG_MEMORY_ISOLATION
249 "Isolate",
250#endif
251};
252
253compound_page_dtor * const compound_page_dtors[] = {
254 NULL,
255 free_compound_page,
256#ifdef CONFIG_HUGETLB_PAGE
257 free_huge_page,
258#endif
259#ifdef CONFIG_TRANSPARENT_HUGEPAGE
260 free_transhuge_page,
261#endif
262};
263
264int min_free_kbytes = 1024;
265int user_min_free_kbytes = -1;
266int watermark_scale_factor = 10;
267
268static unsigned long nr_kernel_pages __meminitdata;
269static unsigned long nr_all_pages __meminitdata;
270static unsigned long dma_reserve __meminitdata;
271
272#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __meminitdata;
274static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __meminitdata;
275static unsigned long required_kernelcore __initdata;
276static unsigned long required_kernelcore_percent __initdata;
277static unsigned long required_movablecore __initdata;
278static unsigned long required_movablecore_percent __initdata;
279static unsigned long zone_movable_pfn[MAX_NUMNODES] __meminitdata;
280static bool mirrored_kernelcore __meminitdata;
281
282/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
283int movable_zone;
284EXPORT_SYMBOL(movable_zone);
285#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
286
287#if MAX_NUMNODES > 1
288int nr_node_ids __read_mostly = MAX_NUMNODES;
289int nr_online_nodes __read_mostly = 1;
290EXPORT_SYMBOL(nr_node_ids);
291EXPORT_SYMBOL(nr_online_nodes);
292#endif
293
294int page_group_by_mobility_disabled __read_mostly;
295
296#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
297/* Returns true if the struct page for the pfn is uninitialised */
298static inline bool __meminit early_page_uninitialised(unsigned long pfn)
299{
300 int nid = early_pfn_to_nid(pfn);
301
302 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
303 return true;
304
305 return false;
306}
307
308/*
309 * Returns false when the remaining initialisation should be deferred until
310 * later in the boot cycle when it can be parallelised.
311 */
312static inline bool update_defer_init(pg_data_t *pgdat,
313 unsigned long pfn, unsigned long zone_end,
314 unsigned long *nr_initialised)
315{
316 /* Always populate low zones for address-constrained allocations */
317 if (zone_end < pgdat_end_pfn(pgdat))
318 return true;
319 (*nr_initialised)++;
320 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
321 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
322 pgdat->first_deferred_pfn = pfn;
323 return false;
324 }
325
326 return true;
327}
328#else
329static inline bool early_page_uninitialised(unsigned long pfn)
330{
331 return false;
332}
333
334static inline bool update_defer_init(pg_data_t *pgdat,
335 unsigned long pfn, unsigned long zone_end,
336 unsigned long *nr_initialised)
337{
338 return true;
339}
340#endif
341
342/* Return a pointer to the bitmap storing bits affecting a block of pages */
343static inline unsigned long *get_pageblock_bitmap(struct page *page,
344 unsigned long pfn)
345{
346#ifdef CONFIG_SPARSEMEM
347 return __pfn_to_section(pfn)->pageblock_flags;
348#else
349 return page_zone(page)->pageblock_flags;
350#endif /* CONFIG_SPARSEMEM */
351}
352
353static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
354{
355#ifdef CONFIG_SPARSEMEM
356 pfn &= (PAGES_PER_SECTION-1);
357 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
358#else
359 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
360 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
361#endif /* CONFIG_SPARSEMEM */
362}
363
364/**
365 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
366 * @page: The page within the block of interest
367 * @pfn: The target page frame number
368 * @end_bitidx: The last bit of interest to retrieve
369 * @mask: mask of bits that the caller is interested in
370 *
371 * Return: pageblock_bits flags
372 */
373static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
374 unsigned long pfn,
375 unsigned long end_bitidx,
376 unsigned long mask)
377{
378 unsigned long *bitmap;
379 unsigned long bitidx, word_bitidx;
380 unsigned long word;
381
382 bitmap = get_pageblock_bitmap(page, pfn);
383 bitidx = pfn_to_bitidx(page, pfn);
384 word_bitidx = bitidx / BITS_PER_LONG;
385 bitidx &= (BITS_PER_LONG-1);
386
387 word = bitmap[word_bitidx];
388 bitidx += end_bitidx;
389 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
390}
391
392unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
393 unsigned long end_bitidx,
394 unsigned long mask)
395{
396 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
397}
398
399static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
400{
401 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
402}
403
404/**
405 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
406 * @page: The page within the block of interest
407 * @flags: The flags to set
408 * @pfn: The target page frame number
409 * @end_bitidx: The last bit of interest
410 * @mask: mask of bits that the caller is interested in
411 */
412void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
413 unsigned long pfn,
414 unsigned long end_bitidx,
415 unsigned long mask)
416{
417 unsigned long *bitmap;
418 unsigned long bitidx, word_bitidx;
419 unsigned long old_word, word;
420
421 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
422
423 bitmap = get_pageblock_bitmap(page, pfn);
424 bitidx = pfn_to_bitidx(page, pfn);
425 word_bitidx = bitidx / BITS_PER_LONG;
426 bitidx &= (BITS_PER_LONG-1);
427
428 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
429
430 bitidx += end_bitidx;
431 mask <<= (BITS_PER_LONG - bitidx - 1);
432 flags <<= (BITS_PER_LONG - bitidx - 1);
433
434 word = READ_ONCE(bitmap[word_bitidx]);
435 for (;;) {
436 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
437 if (word == old_word)
438 break;
439 word = old_word;
440 }
441}
442
443void set_pageblock_migratetype(struct page *page, int migratetype)
444{
445 if (unlikely(page_group_by_mobility_disabled &&
446 migratetype < MIGRATE_PCPTYPES))
447 migratetype = MIGRATE_UNMOVABLE;
448
449 set_pageblock_flags_group(page, (unsigned long)migratetype,
450 PB_migrate, PB_migrate_end);
451}
452
453#ifdef CONFIG_DEBUG_VM
454static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
455{
456 int ret = 0;
457 unsigned seq;
458 unsigned long pfn = page_to_pfn(page);
459 unsigned long sp, start_pfn;
460
461 do {
462 seq = zone_span_seqbegin(zone);
463 start_pfn = zone->zone_start_pfn;
464 sp = zone->spanned_pages;
465 if (!zone_spans_pfn(zone, pfn))
466 ret = 1;
467 } while (zone_span_seqretry(zone, seq));
468
469 if (ret)
470 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
471 pfn, zone_to_nid(zone), zone->name,
472 start_pfn, start_pfn + sp);
473
474 return ret;
475}
476
477static int page_is_consistent(struct zone *zone, struct page *page)
478{
479 if (!pfn_valid_within(page_to_pfn(page)))
480 return 0;
481 if (zone != page_zone(page))
482 return 0;
483
484 return 1;
485}
486/*
487 * Temporary debugging check for pages not lying within a given zone.
488 */
489static int __maybe_unused bad_range(struct zone *zone, struct page *page)
490{
491 if (page_outside_zone_boundaries(zone, page))
492 return 1;
493 if (!page_is_consistent(zone, page))
494 return 1;
495
496 return 0;
497}
498#else
499static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
500{
501 return 0;
502}
503#endif
504
505static void bad_page(struct page *page, const char *reason,
506 unsigned long bad_flags)
507{
508 static unsigned long resume;
509 static unsigned long nr_shown;
510 static unsigned long nr_unshown;
511
512 /*
513 * Allow a burst of 60 reports, then keep quiet for that minute;
514 * or allow a steady drip of one report per second.
515 */
516 if (nr_shown == 60) {
517 if (time_before(jiffies, resume)) {
518 nr_unshown++;
519 goto out;
520 }
521 if (nr_unshown) {
522 pr_alert(
523 "BUG: Bad page state: %lu messages suppressed\n",
524 nr_unshown);
525 nr_unshown = 0;
526 }
527 nr_shown = 0;
528 }
529 if (nr_shown++ == 0)
530 resume = jiffies + 60 * HZ;
531
532 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
533 current->comm, page_to_pfn(page));
534 __dump_page(page, reason);
535 bad_flags &= page->flags;
536 if (bad_flags)
537 pr_alert("bad because of flags: %#lx(%pGp)\n",
538 bad_flags, &bad_flags);
539 dump_page_owner(page);
540
541 print_modules();
542 dump_stack();
543out:
544 /* Leave bad fields for debug, except PageBuddy could make trouble */
545 page_mapcount_reset(page); /* remove PageBuddy */
546 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
547}
548
549/*
550 * Higher-order pages are called "compound pages". They are structured thusly:
551 *
552 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
553 *
554 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
555 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
556 *
557 * The first tail page's ->compound_dtor holds the offset in array of compound
558 * page destructors. See compound_page_dtors.
559 *
560 * The first tail page's ->compound_order holds the order of allocation.
561 * This usage means that zero-order pages may not be compound.
562 */
563
564void free_compound_page(struct page *page)
565{
566 __free_pages_ok(page, compound_order(page));
567}
568
569void prep_compound_page(struct page *page, unsigned int order)
570{
571 int i;
572 int nr_pages = 1 << order;
573
574 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
575 set_compound_order(page, order);
576 __SetPageHead(page);
577 for (i = 1; i < nr_pages; i++) {
578 struct page *p = page + i;
579 set_page_count(p, 0);
580 p->mapping = TAIL_MAPPING;
581 set_compound_head(p, page);
582 }
583 atomic_set(compound_mapcount_ptr(page), -1);
584}
585
586#ifdef CONFIG_DEBUG_PAGEALLOC
587unsigned int _debug_guardpage_minorder;
588bool _debug_pagealloc_enabled __read_mostly
589 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
590EXPORT_SYMBOL(_debug_pagealloc_enabled);
591bool _debug_guardpage_enabled __read_mostly;
592
593static int __init early_debug_pagealloc(char *buf)
594{
595 if (!buf)
596 return -EINVAL;
597 return kstrtobool(buf, &_debug_pagealloc_enabled);
598}
599early_param("debug_pagealloc", early_debug_pagealloc);
600
601static bool need_debug_guardpage(void)
602{
603 /* If we don't use debug_pagealloc, we don't need guard page */
604 if (!debug_pagealloc_enabled())
605 return false;
606
607 if (!debug_guardpage_minorder())
608 return false;
609
610 return true;
611}
612
613static void init_debug_guardpage(void)
614{
615 if (!debug_pagealloc_enabled())
616 return;
617
618 if (!debug_guardpage_minorder())
619 return;
620
621 _debug_guardpage_enabled = true;
622}
623
624struct page_ext_operations debug_guardpage_ops = {
625 .need = need_debug_guardpage,
626 .init = init_debug_guardpage,
627};
628
629static int __init debug_guardpage_minorder_setup(char *buf)
630{
631 unsigned long res;
632
633 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
634 pr_err("Bad debug_guardpage_minorder value\n");
635 return 0;
636 }
637 _debug_guardpage_minorder = res;
638 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
639 return 0;
640}
641early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
642
643static inline bool set_page_guard(struct zone *zone, struct page *page,
644 unsigned int order, int migratetype)
645{
646 struct page_ext *page_ext;
647
648 if (!debug_guardpage_enabled())
649 return false;
650
651 if (order >= debug_guardpage_minorder())
652 return false;
653
654 page_ext = lookup_page_ext(page);
655 if (unlikely(!page_ext))
656 return false;
657
658 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
659
660 INIT_LIST_HEAD(&page->lru);
661 set_page_private(page, order);
662 /* Guard pages are not available for any usage */
663 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
664
665 return true;
666}
667
668static inline void clear_page_guard(struct zone *zone, struct page *page,
669 unsigned int order, int migratetype)
670{
671 struct page_ext *page_ext;
672
673 if (!debug_guardpage_enabled())
674 return;
675
676 page_ext = lookup_page_ext(page);
677 if (unlikely(!page_ext))
678 return;
679
680 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
681
682 set_page_private(page, 0);
683 if (!is_migrate_isolate(migratetype))
684 __mod_zone_freepage_state(zone, (1 << order), migratetype);
685}
686#else
687struct page_ext_operations debug_guardpage_ops;
688static inline bool set_page_guard(struct zone *zone, struct page *page,
689 unsigned int order, int migratetype) { return false; }
690static inline void clear_page_guard(struct zone *zone, struct page *page,
691 unsigned int order, int migratetype) {}
692#endif
693
694static inline void set_page_order(struct page *page, unsigned int order)
695{
696 set_page_private(page, order);
697 __SetPageBuddy(page);
698}
699
700static inline void rmv_page_order(struct page *page)
701{
702 __ClearPageBuddy(page);
703 set_page_private(page, 0);
704}
705
706/*
707 * This function checks whether a page is free && is the buddy
708 * we can do coalesce a page and its buddy if
709 * (a) the buddy is not in a hole (check before calling!) &&
710 * (b) the buddy is in the buddy system &&
711 * (c) a page and its buddy have the same order &&
712 * (d) a page and its buddy are in the same zone.
713 *
714 * For recording whether a page is in the buddy system, we set ->_mapcount
715 * PAGE_BUDDY_MAPCOUNT_VALUE.
716 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
717 * serialized by zone->lock.
718 *
719 * For recording page's order, we use page_private(page).
720 */
721static inline int page_is_buddy(struct page *page, struct page *buddy,
722 unsigned int order)
723{
724 if (page_is_guard(buddy) && page_order(buddy) == order) {
725 if (page_zone_id(page) != page_zone_id(buddy))
726 return 0;
727
728 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
729
730 return 1;
731 }
732
733 if (PageBuddy(buddy) && page_order(buddy) == order) {
734 /*
735 * zone check is done late to avoid uselessly
736 * calculating zone/node ids for pages that could
737 * never merge.
738 */
739 if (page_zone_id(page) != page_zone_id(buddy))
740 return 0;
741
742 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
743
744 return 1;
745 }
746 return 0;
747}
748
749/*
750 * Freeing function for a buddy system allocator.
751 *
752 * The concept of a buddy system is to maintain direct-mapped table
753 * (containing bit values) for memory blocks of various "orders".
754 * The bottom level table contains the map for the smallest allocatable
755 * units of memory (here, pages), and each level above it describes
756 * pairs of units from the levels below, hence, "buddies".
757 * At a high level, all that happens here is marking the table entry
758 * at the bottom level available, and propagating the changes upward
759 * as necessary, plus some accounting needed to play nicely with other
760 * parts of the VM system.
761 * At each level, we keep a list of pages, which are heads of continuous
762 * free pages of length of (1 << order) and marked with _mapcount
763 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
764 * field.
765 * So when we are allocating or freeing one, we can derive the state of the
766 * other. That is, if we allocate a small block, and both were
767 * free, the remainder of the region must be split into blocks.
768 * If a block is freed, and its buddy is also free, then this
769 * triggers coalescing into a block of larger size.
770 *
771 * -- nyc
772 */
773
774static inline void __free_one_page(struct page *page,
775 unsigned long pfn,
776 struct zone *zone, unsigned int order,
777 int migratetype)
778{
779 unsigned long combined_pfn;
780 unsigned long uninitialized_var(buddy_pfn);
781 struct page *buddy;
782 unsigned int max_order;
783
784 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
785
786 VM_BUG_ON(!zone_is_initialized(zone));
787 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
788
789 VM_BUG_ON(migratetype == -1);
790 if (likely(!is_migrate_isolate(migratetype)))
791 __mod_zone_freepage_state(zone, 1 << order, migratetype);
792
793 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
794 VM_BUG_ON_PAGE(bad_range(zone, page), page);
795
796continue_merging:
797 while (order < max_order - 1) {
798 buddy_pfn = __find_buddy_pfn(pfn, order);
799 buddy = page + (buddy_pfn - pfn);
800
801 if (!pfn_valid_within(buddy_pfn))
802 goto done_merging;
803 if (!page_is_buddy(page, buddy, order))
804 goto done_merging;
805 /*
806 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
807 * merge with it and move up one order.
808 */
809 if (page_is_guard(buddy)) {
810 clear_page_guard(zone, buddy, order, migratetype);
811 } else {
812 list_del(&buddy->lru);
813 zone->free_area[order].nr_free--;
814 rmv_page_order(buddy);
815 }
816 combined_pfn = buddy_pfn & pfn;
817 page = page + (combined_pfn - pfn);
818 pfn = combined_pfn;
819 order++;
820 }
821 if (max_order < MAX_ORDER) {
822 /* If we are here, it means order is >= pageblock_order.
823 * We want to prevent merge between freepages on isolate
824 * pageblock and normal pageblock. Without this, pageblock
825 * isolation could cause incorrect freepage or CMA accounting.
826 *
827 * We don't want to hit this code for the more frequent
828 * low-order merging.
829 */
830 if (unlikely(has_isolate_pageblock(zone))) {
831 int buddy_mt;
832
833 buddy_pfn = __find_buddy_pfn(pfn, order);
834 buddy = page + (buddy_pfn - pfn);
835 buddy_mt = get_pageblock_migratetype(buddy);
836
837 if (migratetype != buddy_mt
838 && (is_migrate_isolate(migratetype) ||
839 is_migrate_isolate(buddy_mt)))
840 goto done_merging;
841 }
842 max_order++;
843 goto continue_merging;
844 }
845
846done_merging:
847 set_page_order(page, order);
848
849 /*
850 * If this is not the largest possible page, check if the buddy
851 * of the next-highest order is free. If it is, it's possible
852 * that pages are being freed that will coalesce soon. In case,
853 * that is happening, add the free page to the tail of the list
854 * so it's less likely to be used soon and more likely to be merged
855 * as a higher order page
856 */
857 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
858 struct page *higher_page, *higher_buddy;
859 combined_pfn = buddy_pfn & pfn;
860 higher_page = page + (combined_pfn - pfn);
861 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
862 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
863 if (pfn_valid_within(buddy_pfn) &&
864 page_is_buddy(higher_page, higher_buddy, order + 1)) {
865 list_add_tail(&page->lru,
866 &zone->free_area[order].free_list[migratetype]);
867 goto out;
868 }
869 }
870
871 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
872out:
873 zone->free_area[order].nr_free++;
874}
875
876/*
877 * A bad page could be due to a number of fields. Instead of multiple branches,
878 * try and check multiple fields with one check. The caller must do a detailed
879 * check if necessary.
880 */
881static inline bool page_expected_state(struct page *page,
882 unsigned long check_flags)
883{
884 if (unlikely(atomic_read(&page->_mapcount) != -1))
885 return false;
886
887 if (unlikely((unsigned long)page->mapping |
888 page_ref_count(page) |
889#ifdef CONFIG_MEMCG
890 (unsigned long)page->mem_cgroup |
891#endif
892 (page->flags & check_flags)))
893 return false;
894
895 return true;
896}
897
898static void free_pages_check_bad(struct page *page)
899{
900 const char *bad_reason;
901 unsigned long bad_flags;
902
903 bad_reason = NULL;
904 bad_flags = 0;
905
906 if (unlikely(atomic_read(&page->_mapcount) != -1))
907 bad_reason = "nonzero mapcount";
908 if (unlikely(page->mapping != NULL))
909 bad_reason = "non-NULL mapping";
910 if (unlikely(page_ref_count(page) != 0))
911 bad_reason = "nonzero _refcount";
912 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
913 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
914 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
915 }
916#ifdef CONFIG_MEMCG
917 if (unlikely(page->mem_cgroup))
918 bad_reason = "page still charged to cgroup";
919#endif
920 bad_page(page, bad_reason, bad_flags);
921}
922
923static inline int free_pages_check(struct page *page)
924{
925 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
926 return 0;
927
928 /* Something has gone sideways, find it */
929 free_pages_check_bad(page);
930 return 1;
931}
932
933static int free_tail_pages_check(struct page *head_page, struct page *page)
934{
935 int ret = 1;
936
937 /*
938 * We rely page->lru.next never has bit 0 set, unless the page
939 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
940 */
941 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
942
943 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
944 ret = 0;
945 goto out;
946 }
947 switch (page - head_page) {
948 case 1:
949 /* the first tail page: ->mapping is compound_mapcount() */
950 if (unlikely(compound_mapcount(page))) {
951 bad_page(page, "nonzero compound_mapcount", 0);
952 goto out;
953 }
954 break;
955 case 2:
956 /*
957 * the second tail page: ->mapping is
958 * page_deferred_list().next -- ignore value.
959 */
960 break;
961 default:
962 if (page->mapping != TAIL_MAPPING) {
963 bad_page(page, "corrupted mapping in tail page", 0);
964 goto out;
965 }
966 break;
967 }
968 if (unlikely(!PageTail(page))) {
969 bad_page(page, "PageTail not set", 0);
970 goto out;
971 }
972 if (unlikely(compound_head(page) != head_page)) {
973 bad_page(page, "compound_head not consistent", 0);
974 goto out;
975 }
976 ret = 0;
977out:
978 page->mapping = NULL;
979 clear_compound_head(page);
980 return ret;
981}
982
983static __always_inline bool free_pages_prepare(struct page *page,
984 unsigned int order, bool check_free)
985{
986 int bad = 0;
987
988 VM_BUG_ON_PAGE(PageTail(page), page);
989
990 trace_mm_page_free(page, order);
991
992 /*
993 * Check tail pages before head page information is cleared to
994 * avoid checking PageCompound for order-0 pages.
995 */
996 if (unlikely(order)) {
997 bool compound = PageCompound(page);
998 int i;
999
1000 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1001
1002 if (compound)
1003 ClearPageDoubleMap(page);
1004 for (i = 1; i < (1 << order); i++) {
1005 if (compound)
1006 bad += free_tail_pages_check(page, page + i);
1007 if (unlikely(free_pages_check(page + i))) {
1008 bad++;
1009 continue;
1010 }
1011 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1012 }
1013 }
1014 if (PageMappingFlags(page))
1015 page->mapping = NULL;
1016 if (memcg_kmem_enabled() && PageKmemcg(page))
1017 memcg_kmem_uncharge(page, order);
1018 if (check_free)
1019 bad += free_pages_check(page);
1020 if (bad)
1021 return false;
1022
1023 page_cpupid_reset_last(page);
1024 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1025 reset_page_owner(page, order);
1026
1027 if (!PageHighMem(page)) {
1028 debug_check_no_locks_freed(page_address(page),
1029 PAGE_SIZE << order);
1030 debug_check_no_obj_freed(page_address(page),
1031 PAGE_SIZE << order);
1032 }
1033 arch_free_page(page, order);
1034 kernel_poison_pages(page, 1 << order, 0);
1035 kernel_map_pages(page, 1 << order, 0);
1036 kasan_free_pages(page, order);
1037
1038 return true;
1039}
1040
1041#ifdef CONFIG_DEBUG_VM
1042static inline bool free_pcp_prepare(struct page *page)
1043{
1044 return free_pages_prepare(page, 0, true);
1045}
1046
1047static inline bool bulkfree_pcp_prepare(struct page *page)
1048{
1049 return false;
1050}
1051#else
1052static bool free_pcp_prepare(struct page *page)
1053{
1054 return free_pages_prepare(page, 0, false);
1055}
1056
1057static bool bulkfree_pcp_prepare(struct page *page)
1058{
1059 return free_pages_check(page);
1060}
1061#endif /* CONFIG_DEBUG_VM */
1062
1063static inline void prefetch_buddy(struct page *page)
1064{
1065 unsigned long pfn = page_to_pfn(page);
1066 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1067 struct page *buddy = page + (buddy_pfn - pfn);
1068
1069 prefetch(buddy);
1070}
1071
1072/*
1073 * Frees a number of pages from the PCP lists
1074 * Assumes all pages on list are in same zone, and of same order.
1075 * count is the number of pages to free.
1076 *
1077 * If the zone was previously in an "all pages pinned" state then look to
1078 * see if this freeing clears that state.
1079 *
1080 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1081 * pinned" detection logic.
1082 */
1083static void free_pcppages_bulk(struct zone *zone, int count,
1084 struct per_cpu_pages *pcp)
1085{
1086 int migratetype = 0;
1087 int batch_free = 0;
1088 int prefetch_nr = 0;
1089 bool isolated_pageblocks;
1090 struct page *page, *tmp;
1091 LIST_HEAD(head);
1092
1093 while (count) {
1094 struct list_head *list;
1095
1096 /*
1097 * Remove pages from lists in a round-robin fashion. A
1098 * batch_free count is maintained that is incremented when an
1099 * empty list is encountered. This is so more pages are freed
1100 * off fuller lists instead of spinning excessively around empty
1101 * lists
1102 */
1103 do {
1104 batch_free++;
1105 if (++migratetype == MIGRATE_PCPTYPES)
1106 migratetype = 0;
1107 list = &pcp->lists[migratetype];
1108 } while (list_empty(list));
1109
1110 /* This is the only non-empty list. Free them all. */
1111 if (batch_free == MIGRATE_PCPTYPES)
1112 batch_free = count;
1113
1114 do {
1115 page = list_last_entry(list, struct page, lru);
1116 /* must delete to avoid corrupting pcp list */
1117 list_del(&page->lru);
1118 pcp->count--;
1119
1120 if (bulkfree_pcp_prepare(page))
1121 continue;
1122
1123 list_add_tail(&page->lru, &head);
1124
1125 /*
1126 * We are going to put the page back to the global
1127 * pool, prefetch its buddy to speed up later access
1128 * under zone->lock. It is believed the overhead of
1129 * an additional test and calculating buddy_pfn here
1130 * can be offset by reduced memory latency later. To
1131 * avoid excessive prefetching due to large count, only
1132 * prefetch buddy for the first pcp->batch nr of pages.
1133 */
1134 if (prefetch_nr++ < pcp->batch)
1135 prefetch_buddy(page);
1136 } while (--count && --batch_free && !list_empty(list));
1137 }
1138
1139 spin_lock(&zone->lock);
1140 isolated_pageblocks = has_isolate_pageblock(zone);
1141
1142 /*
1143 * Use safe version since after __free_one_page(),
1144 * page->lru.next will not point to original list.
1145 */
1146 list_for_each_entry_safe(page, tmp, &head, lru) {
1147 int mt = get_pcppage_migratetype(page);
1148 /* MIGRATE_ISOLATE page should not go to pcplists */
1149 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1150 /* Pageblock could have been isolated meanwhile */
1151 if (unlikely(isolated_pageblocks))
1152 mt = get_pageblock_migratetype(page);
1153
1154 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1155 trace_mm_page_pcpu_drain(page, 0, mt);
1156 }
1157 spin_unlock(&zone->lock);
1158}
1159
1160static void free_one_page(struct zone *zone,
1161 struct page *page, unsigned long pfn,
1162 unsigned int order,
1163 int migratetype)
1164{
1165 spin_lock(&zone->lock);
1166 if (unlikely(has_isolate_pageblock(zone) ||
1167 is_migrate_isolate(migratetype))) {
1168 migratetype = get_pfnblock_migratetype(page, pfn);
1169 }
1170 __free_one_page(page, pfn, zone, order, migratetype);
1171 spin_unlock(&zone->lock);
1172}
1173
1174static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1175 unsigned long zone, int nid)
1176{
1177 mm_zero_struct_page(page);
1178 set_page_links(page, zone, nid, pfn);
1179 init_page_count(page);
1180 page_mapcount_reset(page);
1181 page_cpupid_reset_last(page);
1182
1183 INIT_LIST_HEAD(&page->lru);
1184#ifdef WANT_PAGE_VIRTUAL
1185 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1186 if (!is_highmem_idx(zone))
1187 set_page_address(page, __va(pfn << PAGE_SHIFT));
1188#endif
1189}
1190
1191#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1192static void __meminit init_reserved_page(unsigned long pfn)
1193{
1194 pg_data_t *pgdat;
1195 int nid, zid;
1196
1197 if (!early_page_uninitialised(pfn))
1198 return;
1199
1200 nid = early_pfn_to_nid(pfn);
1201 pgdat = NODE_DATA(nid);
1202
1203 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1204 struct zone *zone = &pgdat->node_zones[zid];
1205
1206 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1207 break;
1208 }
1209 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1210}
1211#else
1212static inline void init_reserved_page(unsigned long pfn)
1213{
1214}
1215#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1216
1217/*
1218 * Initialised pages do not have PageReserved set. This function is
1219 * called for each range allocated by the bootmem allocator and
1220 * marks the pages PageReserved. The remaining valid pages are later
1221 * sent to the buddy page allocator.
1222 */
1223void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1224{
1225 unsigned long start_pfn = PFN_DOWN(start);
1226 unsigned long end_pfn = PFN_UP(end);
1227
1228 for (; start_pfn < end_pfn; start_pfn++) {
1229 if (pfn_valid(start_pfn)) {
1230 struct page *page = pfn_to_page(start_pfn);
1231
1232 init_reserved_page(start_pfn);
1233
1234 /* Avoid false-positive PageTail() */
1235 INIT_LIST_HEAD(&page->lru);
1236
1237 SetPageReserved(page);
1238 }
1239 }
1240}
1241
1242static void __free_pages_ok(struct page *page, unsigned int order)
1243{
1244 unsigned long flags;
1245 int migratetype;
1246 unsigned long pfn = page_to_pfn(page);
1247
1248 if (!free_pages_prepare(page, order, true))
1249 return;
1250
1251 migratetype = get_pfnblock_migratetype(page, pfn);
1252 local_irq_save(flags);
1253 __count_vm_events(PGFREE, 1 << order);
1254 free_one_page(page_zone(page), page, pfn, order, migratetype);
1255 local_irq_restore(flags);
1256}
1257
1258static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1259{
1260 unsigned int nr_pages = 1 << order;
1261 struct page *p = page;
1262 unsigned int loop;
1263
1264 prefetchw(p);
1265 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1266 prefetchw(p + 1);
1267 __ClearPageReserved(p);
1268 set_page_count(p, 0);
1269 }
1270 __ClearPageReserved(p);
1271 set_page_count(p, 0);
1272
1273 page_zone(page)->managed_pages += nr_pages;
1274 set_page_refcounted(page);
1275 __free_pages(page, order);
1276}
1277
1278#if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1279 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1280
1281static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1282
1283int __meminit early_pfn_to_nid(unsigned long pfn)
1284{
1285 static DEFINE_SPINLOCK(early_pfn_lock);
1286 int nid;
1287
1288 spin_lock(&early_pfn_lock);
1289 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1290 if (nid < 0)
1291 nid = first_online_node;
1292 spin_unlock(&early_pfn_lock);
1293
1294 return nid;
1295}
1296#endif
1297
1298#ifdef CONFIG_NODES_SPAN_OTHER_NODES
1299static inline bool __meminit __maybe_unused
1300meminit_pfn_in_nid(unsigned long pfn, int node,
1301 struct mminit_pfnnid_cache *state)
1302{
1303 int nid;
1304
1305 nid = __early_pfn_to_nid(pfn, state);
1306 if (nid >= 0 && nid != node)
1307 return false;
1308 return true;
1309}
1310
1311/* Only safe to use early in boot when initialisation is single-threaded */
1312static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1313{
1314 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1315}
1316
1317#else
1318
1319static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1320{
1321 return true;
1322}
1323static inline bool __meminit __maybe_unused
1324meminit_pfn_in_nid(unsigned long pfn, int node,
1325 struct mminit_pfnnid_cache *state)
1326{
1327 return true;
1328}
1329#endif
1330
1331
1332void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1333 unsigned int order)
1334{
1335 if (early_page_uninitialised(pfn))
1336 return;
1337 return __free_pages_boot_core(page, order);
1338}
1339
1340/*
1341 * Check that the whole (or subset of) a pageblock given by the interval of
1342 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1343 * with the migration of free compaction scanner. The scanners then need to
1344 * use only pfn_valid_within() check for arches that allow holes within
1345 * pageblocks.
1346 *
1347 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1348 *
1349 * It's possible on some configurations to have a setup like node0 node1 node0
1350 * i.e. it's possible that all pages within a zones range of pages do not
1351 * belong to a single zone. We assume that a border between node0 and node1
1352 * can occur within a single pageblock, but not a node0 node1 node0
1353 * interleaving within a single pageblock. It is therefore sufficient to check
1354 * the first and last page of a pageblock and avoid checking each individual
1355 * page in a pageblock.
1356 */
1357struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1358 unsigned long end_pfn, struct zone *zone)
1359{
1360 struct page *start_page;
1361 struct page *end_page;
1362
1363 /* end_pfn is one past the range we are checking */
1364 end_pfn--;
1365
1366 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1367 return NULL;
1368
1369 start_page = pfn_to_online_page(start_pfn);
1370 if (!start_page)
1371 return NULL;
1372
1373 if (page_zone(start_page) != zone)
1374 return NULL;
1375
1376 end_page = pfn_to_page(end_pfn);
1377
1378 /* This gives a shorter code than deriving page_zone(end_page) */
1379 if (page_zone_id(start_page) != page_zone_id(end_page))
1380 return NULL;
1381
1382 return start_page;
1383}
1384
1385void set_zone_contiguous(struct zone *zone)
1386{
1387 unsigned long block_start_pfn = zone->zone_start_pfn;
1388 unsigned long block_end_pfn;
1389
1390 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1391 for (; block_start_pfn < zone_end_pfn(zone);
1392 block_start_pfn = block_end_pfn,
1393 block_end_pfn += pageblock_nr_pages) {
1394
1395 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1396
1397 if (!__pageblock_pfn_to_page(block_start_pfn,
1398 block_end_pfn, zone))
1399 return;
1400 }
1401
1402 /* We confirm that there is no hole */
1403 zone->contiguous = true;
1404}
1405
1406void clear_zone_contiguous(struct zone *zone)
1407{
1408 zone->contiguous = false;
1409}
1410
1411#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1412static void __init deferred_free_range(unsigned long pfn,
1413 unsigned long nr_pages)
1414{
1415 struct page *page;
1416 unsigned long i;
1417
1418 if (!nr_pages)
1419 return;
1420
1421 page = pfn_to_page(pfn);
1422
1423 /* Free a large naturally-aligned chunk if possible */
1424 if (nr_pages == pageblock_nr_pages &&
1425 (pfn & (pageblock_nr_pages - 1)) == 0) {
1426 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1427 __free_pages_boot_core(page, pageblock_order);
1428 return;
1429 }
1430
1431 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1432 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1433 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1434 __free_pages_boot_core(page, 0);
1435 }
1436}
1437
1438/* Completion tracking for deferred_init_memmap() threads */
1439static atomic_t pgdat_init_n_undone __initdata;
1440static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1441
1442static inline void __init pgdat_init_report_one_done(void)
1443{
1444 if (atomic_dec_and_test(&pgdat_init_n_undone))
1445 complete(&pgdat_init_all_done_comp);
1446}
1447
1448/*
1449 * Returns true if page needs to be initialized or freed to buddy allocator.
1450 *
1451 * First we check if pfn is valid on architectures where it is possible to have
1452 * holes within pageblock_nr_pages. On systems where it is not possible, this
1453 * function is optimized out.
1454 *
1455 * Then, we check if a current large page is valid by only checking the validity
1456 * of the head pfn.
1457 *
1458 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1459 * within a node: a pfn is between start and end of a node, but does not belong
1460 * to this memory node.
1461 */
1462static inline bool __init
1463deferred_pfn_valid(int nid, unsigned long pfn,
1464 struct mminit_pfnnid_cache *nid_init_state)
1465{
1466 if (!pfn_valid_within(pfn))
1467 return false;
1468 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1469 return false;
1470 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1471 return false;
1472 return true;
1473}
1474
1475/*
1476 * Free pages to buddy allocator. Try to free aligned pages in
1477 * pageblock_nr_pages sizes.
1478 */
1479static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1480 unsigned long end_pfn)
1481{
1482 struct mminit_pfnnid_cache nid_init_state = { };
1483 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1484 unsigned long nr_free = 0;
1485
1486 for (; pfn < end_pfn; pfn++) {
1487 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1488 deferred_free_range(pfn - nr_free, nr_free);
1489 nr_free = 0;
1490 } else if (!(pfn & nr_pgmask)) {
1491 deferred_free_range(pfn - nr_free, nr_free);
1492 nr_free = 1;
1493 touch_nmi_watchdog();
1494 } else {
1495 nr_free++;
1496 }
1497 }
1498 /* Free the last block of pages to allocator */
1499 deferred_free_range(pfn - nr_free, nr_free);
1500}
1501
1502/*
1503 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1504 * by performing it only once every pageblock_nr_pages.
1505 * Return number of pages initialized.
1506 */
1507static unsigned long __init deferred_init_pages(int nid, int zid,
1508 unsigned long pfn,
1509 unsigned long end_pfn)
1510{
1511 struct mminit_pfnnid_cache nid_init_state = { };
1512 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1513 unsigned long nr_pages = 0;
1514 struct page *page = NULL;
1515
1516 for (; pfn < end_pfn; pfn++) {
1517 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1518 page = NULL;
1519 continue;
1520 } else if (!page || !(pfn & nr_pgmask)) {
1521 page = pfn_to_page(pfn);
1522 touch_nmi_watchdog();
1523 } else {
1524 page++;
1525 }
1526 __init_single_page(page, pfn, zid, nid);
1527 nr_pages++;
1528 }
1529 return (nr_pages);
1530}
1531
1532/* Initialise remaining memory on a node */
1533static int __init deferred_init_memmap(void *data)
1534{
1535 pg_data_t *pgdat = data;
1536 int nid = pgdat->node_id;
1537 unsigned long start = jiffies;
1538 unsigned long nr_pages = 0;
1539 unsigned long spfn, epfn, first_init_pfn, flags;
1540 phys_addr_t spa, epa;
1541 int zid;
1542 struct zone *zone;
1543 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1544 u64 i;
1545
1546 /* Bind memory initialisation thread to a local node if possible */
1547 if (!cpumask_empty(cpumask))
1548 set_cpus_allowed_ptr(current, cpumask);
1549
1550 pgdat_resize_lock(pgdat, &flags);
1551 first_init_pfn = pgdat->first_deferred_pfn;
1552 if (first_init_pfn == ULONG_MAX) {
1553 pgdat_resize_unlock(pgdat, &flags);
1554 pgdat_init_report_one_done();
1555 return 0;
1556 }
1557
1558 /* Sanity check boundaries */
1559 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1560 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1561 pgdat->first_deferred_pfn = ULONG_MAX;
1562
1563 /* Only the highest zone is deferred so find it */
1564 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1565 zone = pgdat->node_zones + zid;
1566 if (first_init_pfn < zone_end_pfn(zone))
1567 break;
1568 }
1569 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1570
1571 /*
1572 * Initialize and free pages. We do it in two loops: first we initialize
1573 * struct page, than free to buddy allocator, because while we are
1574 * freeing pages we can access pages that are ahead (computing buddy
1575 * page in __free_one_page()).
1576 */
1577 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1578 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1579 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1580 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1581 }
1582 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1583 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1584 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1585 deferred_free_pages(nid, zid, spfn, epfn);
1586 }
1587 pgdat_resize_unlock(pgdat, &flags);
1588
1589 /* Sanity check that the next zone really is unpopulated */
1590 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1591
1592 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1593 jiffies_to_msecs(jiffies - start));
1594
1595 pgdat_init_report_one_done();
1596 return 0;
1597}
1598
1599/*
1600 * During boot we initialize deferred pages on-demand, as needed, but once
1601 * page_alloc_init_late() has finished, the deferred pages are all initialized,
1602 * and we can permanently disable that path.
1603 */
1604static DEFINE_STATIC_KEY_TRUE(deferred_pages);
1605
1606/*
1607 * If this zone has deferred pages, try to grow it by initializing enough
1608 * deferred pages to satisfy the allocation specified by order, rounded up to
1609 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1610 * of SECTION_SIZE bytes by initializing struct pages in increments of
1611 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1612 *
1613 * Return true when zone was grown, otherwise return false. We return true even
1614 * when we grow less than requested, to let the caller decide if there are
1615 * enough pages to satisfy the allocation.
1616 *
1617 * Note: We use noinline because this function is needed only during boot, and
1618 * it is called from a __ref function _deferred_grow_zone. This way we are
1619 * making sure that it is not inlined into permanent text section.
1620 */
1621static noinline bool __init
1622deferred_grow_zone(struct zone *zone, unsigned int order)
1623{
1624 int zid = zone_idx(zone);
1625 int nid = zone_to_nid(zone);
1626 pg_data_t *pgdat = NODE_DATA(nid);
1627 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1628 unsigned long nr_pages = 0;
1629 unsigned long first_init_pfn, spfn, epfn, t, flags;
1630 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1631 phys_addr_t spa, epa;
1632 u64 i;
1633
1634 /* Only the last zone may have deferred pages */
1635 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1636 return false;
1637
1638 pgdat_resize_lock(pgdat, &flags);
1639
1640 /*
1641 * If deferred pages have been initialized while we were waiting for
1642 * the lock, return true, as the zone was grown. The caller will retry
1643 * this zone. We won't return to this function since the caller also
1644 * has this static branch.
1645 */
1646 if (!static_branch_unlikely(&deferred_pages)) {
1647 pgdat_resize_unlock(pgdat, &flags);
1648 return true;
1649 }
1650
1651 /*
1652 * If someone grew this zone while we were waiting for spinlock, return
1653 * true, as there might be enough pages already.
1654 */
1655 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1656 pgdat_resize_unlock(pgdat, &flags);
1657 return true;
1658 }
1659
1660 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1661
1662 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1663 pgdat_resize_unlock(pgdat, &flags);
1664 return false;
1665 }
1666
1667 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1668 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1669 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1670
1671 while (spfn < epfn && nr_pages < nr_pages_needed) {
1672 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1673 first_deferred_pfn = min(t, epfn);
1674 nr_pages += deferred_init_pages(nid, zid, spfn,
1675 first_deferred_pfn);
1676 spfn = first_deferred_pfn;
1677 }
1678
1679 if (nr_pages >= nr_pages_needed)
1680 break;
1681 }
1682
1683 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1684 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1685 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1686 deferred_free_pages(nid, zid, spfn, epfn);
1687
1688 if (first_deferred_pfn == epfn)
1689 break;
1690 }
1691 pgdat->first_deferred_pfn = first_deferred_pfn;
1692 pgdat_resize_unlock(pgdat, &flags);
1693
1694 return nr_pages > 0;
1695}
1696
1697/*
1698 * deferred_grow_zone() is __init, but it is called from
1699 * get_page_from_freelist() during early boot until deferred_pages permanently
1700 * disables this call. This is why we have refdata wrapper to avoid warning,
1701 * and to ensure that the function body gets unloaded.
1702 */
1703static bool __ref
1704_deferred_grow_zone(struct zone *zone, unsigned int order)
1705{
1706 return deferred_grow_zone(zone, order);
1707}
1708
1709#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1710
1711void __init page_alloc_init_late(void)
1712{
1713 struct zone *zone;
1714
1715#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1716 int nid;
1717
1718 /* There will be num_node_state(N_MEMORY) threads */
1719 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1720 for_each_node_state(nid, N_MEMORY) {
1721 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1722 }
1723
1724 /* Block until all are initialised */
1725 wait_for_completion(&pgdat_init_all_done_comp);
1726
1727 /*
1728 * We initialized the rest of the deferred pages. Permanently disable
1729 * on-demand struct page initialization.
1730 */
1731 static_branch_disable(&deferred_pages);
1732
1733 /* Reinit limits that are based on free pages after the kernel is up */
1734 files_maxfiles_init();
1735#endif
1736#ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1737 /* Discard memblock private memory */
1738 memblock_discard();
1739#endif
1740
1741 for_each_populated_zone(zone)
1742 set_zone_contiguous(zone);
1743}
1744
1745#ifdef CONFIG_CMA
1746/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1747void __init init_cma_reserved_pageblock(struct page *page)
1748{
1749 unsigned i = pageblock_nr_pages;
1750 struct page *p = page;
1751
1752 do {
1753 __ClearPageReserved(p);
1754 set_page_count(p, 0);
1755 } while (++p, --i);
1756
1757 set_pageblock_migratetype(page, MIGRATE_CMA);
1758
1759 if (pageblock_order >= MAX_ORDER) {
1760 i = pageblock_nr_pages;
1761 p = page;
1762 do {
1763 set_page_refcounted(p);
1764 __free_pages(p, MAX_ORDER - 1);
1765 p += MAX_ORDER_NR_PAGES;
1766 } while (i -= MAX_ORDER_NR_PAGES);
1767 } else {
1768 set_page_refcounted(page);
1769 __free_pages(page, pageblock_order);
1770 }
1771
1772 adjust_managed_page_count(page, pageblock_nr_pages);
1773}
1774#endif
1775
1776/*
1777 * The order of subdivision here is critical for the IO subsystem.
1778 * Please do not alter this order without good reasons and regression
1779 * testing. Specifically, as large blocks of memory are subdivided,
1780 * the order in which smaller blocks are delivered depends on the order
1781 * they're subdivided in this function. This is the primary factor
1782 * influencing the order in which pages are delivered to the IO
1783 * subsystem according to empirical testing, and this is also justified
1784 * by considering the behavior of a buddy system containing a single
1785 * large block of memory acted on by a series of small allocations.
1786 * This behavior is a critical factor in sglist merging's success.
1787 *
1788 * -- nyc
1789 */
1790static inline void expand(struct zone *zone, struct page *page,
1791 int low, int high, struct free_area *area,
1792 int migratetype)
1793{
1794 unsigned long size = 1 << high;
1795
1796 while (high > low) {
1797 area--;
1798 high--;
1799 size >>= 1;
1800 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1801
1802 /*
1803 * Mark as guard pages (or page), that will allow to
1804 * merge back to allocator when buddy will be freed.
1805 * Corresponding page table entries will not be touched,
1806 * pages will stay not present in virtual address space
1807 */
1808 if (set_page_guard(zone, &page[size], high, migratetype))
1809 continue;
1810
1811 list_add(&page[size].lru, &area->free_list[migratetype]);
1812 area->nr_free++;
1813 set_page_order(&page[size], high);
1814 }
1815}
1816
1817static void check_new_page_bad(struct page *page)
1818{
1819 const char *bad_reason = NULL;
1820 unsigned long bad_flags = 0;
1821
1822 if (unlikely(atomic_read(&page->_mapcount) != -1))
1823 bad_reason = "nonzero mapcount";
1824 if (unlikely(page->mapping != NULL))
1825 bad_reason = "non-NULL mapping";
1826 if (unlikely(page_ref_count(page) != 0))
1827 bad_reason = "nonzero _count";
1828 if (unlikely(page->flags & __PG_HWPOISON)) {
1829 bad_reason = "HWPoisoned (hardware-corrupted)";
1830 bad_flags = __PG_HWPOISON;
1831 /* Don't complain about hwpoisoned pages */
1832 page_mapcount_reset(page); /* remove PageBuddy */
1833 return;
1834 }
1835 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1836 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1837 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1838 }
1839#ifdef CONFIG_MEMCG
1840 if (unlikely(page->mem_cgroup))
1841 bad_reason = "page still charged to cgroup";
1842#endif
1843 bad_page(page, bad_reason, bad_flags);
1844}
1845
1846/*
1847 * This page is about to be returned from the page allocator
1848 */
1849static inline int check_new_page(struct page *page)
1850{
1851 if (likely(page_expected_state(page,
1852 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1853 return 0;
1854
1855 check_new_page_bad(page);
1856 return 1;
1857}
1858
1859static inline bool free_pages_prezeroed(void)
1860{
1861 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1862 page_poisoning_enabled();
1863}
1864
1865#ifdef CONFIG_DEBUG_VM
1866static bool check_pcp_refill(struct page *page)
1867{
1868 return false;
1869}
1870
1871static bool check_new_pcp(struct page *page)
1872{
1873 return check_new_page(page);
1874}
1875#else
1876static bool check_pcp_refill(struct page *page)
1877{
1878 return check_new_page(page);
1879}
1880static bool check_new_pcp(struct page *page)
1881{
1882 return false;
1883}
1884#endif /* CONFIG_DEBUG_VM */
1885
1886static bool check_new_pages(struct page *page, unsigned int order)
1887{
1888 int i;
1889 for (i = 0; i < (1 << order); i++) {
1890 struct page *p = page + i;
1891
1892 if (unlikely(check_new_page(p)))
1893 return true;
1894 }
1895
1896 return false;
1897}
1898
1899inline void post_alloc_hook(struct page *page, unsigned int order,
1900 gfp_t gfp_flags)
1901{
1902 set_page_private(page, 0);
1903 set_page_refcounted(page);
1904
1905 arch_alloc_page(page, order);
1906 kernel_map_pages(page, 1 << order, 1);
1907 kernel_poison_pages(page, 1 << order, 1);
1908 kasan_alloc_pages(page, order);
1909 set_page_owner(page, order, gfp_flags);
1910}
1911
1912static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1913 unsigned int alloc_flags)
1914{
1915 int i;
1916
1917 post_alloc_hook(page, order, gfp_flags);
1918
1919 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1920 for (i = 0; i < (1 << order); i++)
1921 clear_highpage(page + i);
1922
1923 if (order && (gfp_flags & __GFP_COMP))
1924 prep_compound_page(page, order);
1925
1926 /*
1927 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1928 * allocate the page. The expectation is that the caller is taking
1929 * steps that will free more memory. The caller should avoid the page
1930 * being used for !PFMEMALLOC purposes.
1931 */
1932 if (alloc_flags & ALLOC_NO_WATERMARKS)
1933 set_page_pfmemalloc(page);
1934 else
1935 clear_page_pfmemalloc(page);
1936}
1937
1938/*
1939 * Go through the free lists for the given migratetype and remove
1940 * the smallest available page from the freelists
1941 */
1942static __always_inline
1943struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1944 int migratetype)
1945{
1946 unsigned int current_order;
1947 struct free_area *area;
1948 struct page *page;
1949
1950 /* Find a page of the appropriate size in the preferred list */
1951 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1952 area = &(zone->free_area[current_order]);
1953 page = list_first_entry_or_null(&area->free_list[migratetype],
1954 struct page, lru);
1955 if (!page)
1956 continue;
1957 list_del(&page->lru);
1958 rmv_page_order(page);
1959 area->nr_free--;
1960 expand(zone, page, order, current_order, area, migratetype);
1961 set_pcppage_migratetype(page, migratetype);
1962 return page;
1963 }
1964
1965 return NULL;
1966}
1967
1968
1969/*
1970 * This array describes the order lists are fallen back to when
1971 * the free lists for the desirable migrate type are depleted
1972 */
1973static int fallbacks[MIGRATE_TYPES][4] = {
1974 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1975 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1976 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1977#ifdef CONFIG_CMA
1978 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1979#endif
1980#ifdef CONFIG_MEMORY_ISOLATION
1981 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1982#endif
1983};
1984
1985#ifdef CONFIG_CMA
1986static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1987 unsigned int order)
1988{
1989 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1990}
1991#else
1992static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1993 unsigned int order) { return NULL; }
1994#endif
1995
1996/*
1997 * Move the free pages in a range to the free lists of the requested type.
1998 * Note that start_page and end_pages are not aligned on a pageblock
1999 * boundary. If alignment is required, use move_freepages_block()
2000 */
2001static int move_freepages(struct zone *zone,
2002 struct page *start_page, struct page *end_page,
2003 int migratetype, int *num_movable)
2004{
2005 struct page *page;
2006 unsigned int order;
2007 int pages_moved = 0;
2008
2009#ifndef CONFIG_HOLES_IN_ZONE
2010 /*
2011 * page_zone is not safe to call in this context when
2012 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2013 * anyway as we check zone boundaries in move_freepages_block().
2014 * Remove at a later date when no bug reports exist related to
2015 * grouping pages by mobility
2016 */
2017 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2018 pfn_valid(page_to_pfn(end_page)) &&
2019 page_zone(start_page) != page_zone(end_page));
2020#endif
2021
2022 if (num_movable)
2023 *num_movable = 0;
2024
2025 for (page = start_page; page <= end_page;) {
2026 if (!pfn_valid_within(page_to_pfn(page))) {
2027 page++;
2028 continue;
2029 }
2030
2031 /* Make sure we are not inadvertently changing nodes */
2032 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2033
2034 if (!PageBuddy(page)) {
2035 /*
2036 * We assume that pages that could be isolated for
2037 * migration are movable. But we don't actually try
2038 * isolating, as that would be expensive.
2039 */
2040 if (num_movable &&
2041 (PageLRU(page) || __PageMovable(page)))
2042 (*num_movable)++;
2043
2044 page++;
2045 continue;
2046 }
2047
2048 order = page_order(page);
2049 list_move(&page->lru,
2050 &zone->free_area[order].free_list[migratetype]);
2051 page += 1 << order;
2052 pages_moved += 1 << order;
2053 }
2054
2055 return pages_moved;
2056}
2057
2058int move_freepages_block(struct zone *zone, struct page *page,
2059 int migratetype, int *num_movable)
2060{
2061 unsigned long start_pfn, end_pfn;
2062 struct page *start_page, *end_page;
2063
2064 start_pfn = page_to_pfn(page);
2065 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2066 start_page = pfn_to_page(start_pfn);
2067 end_page = start_page + pageblock_nr_pages - 1;
2068 end_pfn = start_pfn + pageblock_nr_pages - 1;
2069
2070 /* Do not cross zone boundaries */
2071 if (!zone_spans_pfn(zone, start_pfn))
2072 start_page = page;
2073 if (!zone_spans_pfn(zone, end_pfn))
2074 return 0;
2075
2076 return move_freepages(zone, start_page, end_page, migratetype,
2077 num_movable);
2078}
2079
2080static void change_pageblock_range(struct page *pageblock_page,
2081 int start_order, int migratetype)
2082{
2083 int nr_pageblocks = 1 << (start_order - pageblock_order);
2084
2085 while (nr_pageblocks--) {
2086 set_pageblock_migratetype(pageblock_page, migratetype);
2087 pageblock_page += pageblock_nr_pages;
2088 }
2089}
2090
2091/*
2092 * When we are falling back to another migratetype during allocation, try to
2093 * steal extra free pages from the same pageblocks to satisfy further
2094 * allocations, instead of polluting multiple pageblocks.
2095 *
2096 * If we are stealing a relatively large buddy page, it is likely there will
2097 * be more free pages in the pageblock, so try to steal them all. For
2098 * reclaimable and unmovable allocations, we steal regardless of page size,
2099 * as fragmentation caused by those allocations polluting movable pageblocks
2100 * is worse than movable allocations stealing from unmovable and reclaimable
2101 * pageblocks.
2102 */
2103static bool can_steal_fallback(unsigned int order, int start_mt)
2104{
2105 /*
2106 * Leaving this order check is intended, although there is
2107 * relaxed order check in next check. The reason is that
2108 * we can actually steal whole pageblock if this condition met,
2109 * but, below check doesn't guarantee it and that is just heuristic
2110 * so could be changed anytime.
2111 */
2112 if (order >= pageblock_order)
2113 return true;
2114
2115 if (order >= pageblock_order / 2 ||
2116 start_mt == MIGRATE_RECLAIMABLE ||
2117 start_mt == MIGRATE_UNMOVABLE ||
2118 page_group_by_mobility_disabled)
2119 return true;
2120
2121 return false;
2122}
2123
2124/*
2125 * This function implements actual steal behaviour. If order is large enough,
2126 * we can steal whole pageblock. If not, we first move freepages in this
2127 * pageblock to our migratetype and determine how many already-allocated pages
2128 * are there in the pageblock with a compatible migratetype. If at least half
2129 * of pages are free or compatible, we can change migratetype of the pageblock
2130 * itself, so pages freed in the future will be put on the correct free list.
2131 */
2132static void steal_suitable_fallback(struct zone *zone, struct page *page,
2133 int start_type, bool whole_block)
2134{
2135 unsigned int current_order = page_order(page);
2136 struct free_area *area;
2137 int free_pages, movable_pages, alike_pages;
2138 int old_block_type;
2139
2140 old_block_type = get_pageblock_migratetype(page);
2141
2142 /*
2143 * This can happen due to races and we want to prevent broken
2144 * highatomic accounting.
2145 */
2146 if (is_migrate_highatomic(old_block_type))
2147 goto single_page;
2148
2149 /* Take ownership for orders >= pageblock_order */
2150 if (current_order >= pageblock_order) {
2151 change_pageblock_range(page, current_order, start_type);
2152 goto single_page;
2153 }
2154
2155 /* We are not allowed to try stealing from the whole block */
2156 if (!whole_block)
2157 goto single_page;
2158
2159 free_pages = move_freepages_block(zone, page, start_type,
2160 &movable_pages);
2161 /*
2162 * Determine how many pages are compatible with our allocation.
2163 * For movable allocation, it's the number of movable pages which
2164 * we just obtained. For other types it's a bit more tricky.
2165 */
2166 if (start_type == MIGRATE_MOVABLE) {
2167 alike_pages = movable_pages;
2168 } else {
2169 /*
2170 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2171 * to MOVABLE pageblock, consider all non-movable pages as
2172 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2173 * vice versa, be conservative since we can't distinguish the
2174 * exact migratetype of non-movable pages.
2175 */
2176 if (old_block_type == MIGRATE_MOVABLE)
2177 alike_pages = pageblock_nr_pages
2178 - (free_pages + movable_pages);
2179 else
2180 alike_pages = 0;
2181 }
2182
2183 /* moving whole block can fail due to zone boundary conditions */
2184 if (!free_pages)
2185 goto single_page;
2186
2187 /*
2188 * If a sufficient number of pages in the block are either free or of
2189 * comparable migratability as our allocation, claim the whole block.
2190 */
2191 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2192 page_group_by_mobility_disabled)
2193 set_pageblock_migratetype(page, start_type);
2194
2195 return;
2196
2197single_page:
2198 area = &zone->free_area[current_order];
2199 list_move(&page->lru, &area->free_list[start_type]);
2200}
2201
2202/*
2203 * Check whether there is a suitable fallback freepage with requested order.
2204 * If only_stealable is true, this function returns fallback_mt only if
2205 * we can steal other freepages all together. This would help to reduce
2206 * fragmentation due to mixed migratetype pages in one pageblock.
2207 */
2208int find_suitable_fallback(struct free_area *area, unsigned int order,
2209 int migratetype, bool only_stealable, bool *can_steal)
2210{
2211 int i;
2212 int fallback_mt;
2213
2214 if (area->nr_free == 0)
2215 return -1;
2216
2217 *can_steal = false;
2218 for (i = 0;; i++) {
2219 fallback_mt = fallbacks[migratetype][i];
2220 if (fallback_mt == MIGRATE_TYPES)
2221 break;
2222
2223 if (list_empty(&area->free_list[fallback_mt]))
2224 continue;
2225
2226 if (can_steal_fallback(order, migratetype))
2227 *can_steal = true;
2228
2229 if (!only_stealable)
2230 return fallback_mt;
2231
2232 if (*can_steal)
2233 return fallback_mt;
2234 }
2235
2236 return -1;
2237}
2238
2239/*
2240 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2241 * there are no empty page blocks that contain a page with a suitable order
2242 */
2243static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2244 unsigned int alloc_order)
2245{
2246 int mt;
2247 unsigned long max_managed, flags;
2248
2249 /*
2250 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2251 * Check is race-prone but harmless.
2252 */
2253 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2254 if (zone->nr_reserved_highatomic >= max_managed)
2255 return;
2256
2257 spin_lock_irqsave(&zone->lock, flags);
2258
2259 /* Recheck the nr_reserved_highatomic limit under the lock */
2260 if (zone->nr_reserved_highatomic >= max_managed)
2261 goto out_unlock;
2262
2263 /* Yoink! */
2264 mt = get_pageblock_migratetype(page);
2265 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2266 && !is_migrate_cma(mt)) {
2267 zone->nr_reserved_highatomic += pageblock_nr_pages;
2268 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2269 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2270 }
2271
2272out_unlock:
2273 spin_unlock_irqrestore(&zone->lock, flags);
2274}
2275
2276/*
2277 * Used when an allocation is about to fail under memory pressure. This
2278 * potentially hurts the reliability of high-order allocations when under
2279 * intense memory pressure but failed atomic allocations should be easier
2280 * to recover from than an OOM.
2281 *
2282 * If @force is true, try to unreserve a pageblock even though highatomic
2283 * pageblock is exhausted.
2284 */
2285static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2286 bool force)
2287{
2288 struct zonelist *zonelist = ac->zonelist;
2289 unsigned long flags;
2290 struct zoneref *z;
2291 struct zone *zone;
2292 struct page *page;
2293 int order;
2294 bool ret;
2295
2296 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2297 ac->nodemask) {
2298 /*
2299 * Preserve at least one pageblock unless memory pressure
2300 * is really high.
2301 */
2302 if (!force && zone->nr_reserved_highatomic <=
2303 pageblock_nr_pages)
2304 continue;
2305
2306 spin_lock_irqsave(&zone->lock, flags);
2307 for (order = 0; order < MAX_ORDER; order++) {
2308 struct free_area *area = &(zone->free_area[order]);
2309
2310 page = list_first_entry_or_null(
2311 &area->free_list[MIGRATE_HIGHATOMIC],
2312 struct page, lru);
2313 if (!page)
2314 continue;
2315
2316 /*
2317 * In page freeing path, migratetype change is racy so
2318 * we can counter several free pages in a pageblock
2319 * in this loop althoug we changed the pageblock type
2320 * from highatomic to ac->migratetype. So we should
2321 * adjust the count once.
2322 */
2323 if (is_migrate_highatomic_page(page)) {
2324 /*
2325 * It should never happen but changes to
2326 * locking could inadvertently allow a per-cpu
2327 * drain to add pages to MIGRATE_HIGHATOMIC
2328 * while unreserving so be safe and watch for
2329 * underflows.
2330 */
2331 zone->nr_reserved_highatomic -= min(
2332 pageblock_nr_pages,
2333 zone->nr_reserved_highatomic);
2334 }
2335
2336 /*
2337 * Convert to ac->migratetype and avoid the normal
2338 * pageblock stealing heuristics. Minimally, the caller
2339 * is doing the work and needs the pages. More
2340 * importantly, if the block was always converted to
2341 * MIGRATE_UNMOVABLE or another type then the number
2342 * of pageblocks that cannot be completely freed
2343 * may increase.
2344 */
2345 set_pageblock_migratetype(page, ac->migratetype);
2346 ret = move_freepages_block(zone, page, ac->migratetype,
2347 NULL);
2348 if (ret) {
2349 spin_unlock_irqrestore(&zone->lock, flags);
2350 return ret;
2351 }
2352 }
2353 spin_unlock_irqrestore(&zone->lock, flags);
2354 }
2355
2356 return false;
2357}
2358
2359/*
2360 * Try finding a free buddy page on the fallback list and put it on the free
2361 * list of requested migratetype, possibly along with other pages from the same
2362 * block, depending on fragmentation avoidance heuristics. Returns true if
2363 * fallback was found so that __rmqueue_smallest() can grab it.
2364 *
2365 * The use of signed ints for order and current_order is a deliberate
2366 * deviation from the rest of this file, to make the for loop
2367 * condition simpler.
2368 */
2369static __always_inline bool
2370__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2371{
2372 struct free_area *area;
2373 int current_order;
2374 struct page *page;
2375 int fallback_mt;
2376 bool can_steal;
2377
2378 /*
2379 * Find the largest available free page in the other list. This roughly
2380 * approximates finding the pageblock with the most free pages, which
2381 * would be too costly to do exactly.
2382 */
2383 for (current_order = MAX_ORDER - 1; current_order >= order;
2384 --current_order) {
2385 area = &(zone->free_area[current_order]);
2386 fallback_mt = find_suitable_fallback(area, current_order,
2387 start_migratetype, false, &can_steal);
2388 if (fallback_mt == -1)
2389 continue;
2390
2391 /*
2392 * We cannot steal all free pages from the pageblock and the
2393 * requested migratetype is movable. In that case it's better to
2394 * steal and split the smallest available page instead of the
2395 * largest available page, because even if the next movable
2396 * allocation falls back into a different pageblock than this
2397 * one, it won't cause permanent fragmentation.
2398 */
2399 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2400 && current_order > order)
2401 goto find_smallest;
2402
2403 goto do_steal;
2404 }
2405
2406 return false;
2407
2408find_smallest:
2409 for (current_order = order; current_order < MAX_ORDER;
2410 current_order++) {
2411 area = &(zone->free_area[current_order]);
2412 fallback_mt = find_suitable_fallback(area, current_order,
2413 start_migratetype, false, &can_steal);
2414 if (fallback_mt != -1)
2415 break;
2416 }
2417
2418 /*
2419 * This should not happen - we already found a suitable fallback
2420 * when looking for the largest page.
2421 */
2422 VM_BUG_ON(current_order == MAX_ORDER);
2423
2424do_steal:
2425 page = list_first_entry(&area->free_list[fallback_mt],
2426 struct page, lru);
2427
2428 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2429
2430 trace_mm_page_alloc_extfrag(page, order, current_order,
2431 start_migratetype, fallback_mt);
2432
2433 return true;
2434
2435}
2436
2437/*
2438 * Do the hard work of removing an element from the buddy allocator.
2439 * Call me with the zone->lock already held.
2440 */
2441static __always_inline struct page *
2442__rmqueue(struct zone *zone, unsigned int order, int migratetype)
2443{
2444 struct page *page;
2445
2446retry:
2447 page = __rmqueue_smallest(zone, order, migratetype);
2448 if (unlikely(!page)) {
2449 if (migratetype == MIGRATE_MOVABLE)
2450 page = __rmqueue_cma_fallback(zone, order);
2451
2452 if (!page && __rmqueue_fallback(zone, order, migratetype))
2453 goto retry;
2454 }
2455
2456 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2457 return page;
2458}
2459
2460/*
2461 * Obtain a specified number of elements from the buddy allocator, all under
2462 * a single hold of the lock, for efficiency. Add them to the supplied list.
2463 * Returns the number of new pages which were placed at *list.
2464 */
2465static int rmqueue_bulk(struct zone *zone, unsigned int order,
2466 unsigned long count, struct list_head *list,
2467 int migratetype)
2468{
2469 int i, alloced = 0;
2470
2471 spin_lock(&zone->lock);
2472 for (i = 0; i < count; ++i) {
2473 struct page *page = __rmqueue(zone, order, migratetype);
2474 if (unlikely(page == NULL))
2475 break;
2476
2477 if (unlikely(check_pcp_refill(page)))
2478 continue;
2479
2480 /*
2481 * Split buddy pages returned by expand() are received here in
2482 * physical page order. The page is added to the tail of
2483 * caller's list. From the callers perspective, the linked list
2484 * is ordered by page number under some conditions. This is
2485 * useful for IO devices that can forward direction from the
2486 * head, thus also in the physical page order. This is useful
2487 * for IO devices that can merge IO requests if the physical
2488 * pages are ordered properly.
2489 */
2490 list_add_tail(&page->lru, list);
2491 alloced++;
2492 if (is_migrate_cma(get_pcppage_migratetype(page)))
2493 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2494 -(1 << order));
2495 }
2496
2497 /*
2498 * i pages were removed from the buddy list even if some leak due
2499 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2500 * on i. Do not confuse with 'alloced' which is the number of
2501 * pages added to the pcp list.
2502 */
2503 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2504 spin_unlock(&zone->lock);
2505 return alloced;
2506}
2507
2508#ifdef CONFIG_NUMA
2509/*
2510 * Called from the vmstat counter updater to drain pagesets of this
2511 * currently executing processor on remote nodes after they have
2512 * expired.
2513 *
2514 * Note that this function must be called with the thread pinned to
2515 * a single processor.
2516 */
2517void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2518{
2519 unsigned long flags;
2520 int to_drain, batch;
2521
2522 local_irq_save(flags);
2523 batch = READ_ONCE(pcp->batch);
2524 to_drain = min(pcp->count, batch);
2525 if (to_drain > 0)
2526 free_pcppages_bulk(zone, to_drain, pcp);
2527 local_irq_restore(flags);
2528}
2529#endif
2530
2531/*
2532 * Drain pcplists of the indicated processor and zone.
2533 *
2534 * The processor must either be the current processor and the
2535 * thread pinned to the current processor or a processor that
2536 * is not online.
2537 */
2538static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2539{
2540 unsigned long flags;
2541 struct per_cpu_pageset *pset;
2542 struct per_cpu_pages *pcp;
2543
2544 local_irq_save(flags);
2545 pset = per_cpu_ptr(zone->pageset, cpu);
2546
2547 pcp = &pset->pcp;
2548 if (pcp->count)
2549 free_pcppages_bulk(zone, pcp->count, pcp);
2550 local_irq_restore(flags);
2551}
2552
2553/*
2554 * Drain pcplists of all zones on the indicated processor.
2555 *
2556 * The processor must either be the current processor and the
2557 * thread pinned to the current processor or a processor that
2558 * is not online.
2559 */
2560static void drain_pages(unsigned int cpu)
2561{
2562 struct zone *zone;
2563
2564 for_each_populated_zone(zone) {
2565 drain_pages_zone(cpu, zone);
2566 }
2567}
2568
2569/*
2570 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2571 *
2572 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2573 * the single zone's pages.
2574 */
2575void drain_local_pages(struct zone *zone)
2576{
2577 int cpu = smp_processor_id();
2578
2579 if (zone)
2580 drain_pages_zone(cpu, zone);
2581 else
2582 drain_pages(cpu);
2583}
2584
2585static void drain_local_pages_wq(struct work_struct *work)
2586{
2587 /*
2588 * drain_all_pages doesn't use proper cpu hotplug protection so
2589 * we can race with cpu offline when the WQ can move this from
2590 * a cpu pinned worker to an unbound one. We can operate on a different
2591 * cpu which is allright but we also have to make sure to not move to
2592 * a different one.
2593 */
2594 preempt_disable();
2595 drain_local_pages(NULL);
2596 preempt_enable();
2597}
2598
2599/*
2600 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2601 *
2602 * When zone parameter is non-NULL, spill just the single zone's pages.
2603 *
2604 * Note that this can be extremely slow as the draining happens in a workqueue.
2605 */
2606void drain_all_pages(struct zone *zone)
2607{
2608 int cpu;
2609
2610 /*
2611 * Allocate in the BSS so we wont require allocation in
2612 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2613 */
2614 static cpumask_t cpus_with_pcps;
2615
2616 /*
2617 * Make sure nobody triggers this path before mm_percpu_wq is fully
2618 * initialized.
2619 */
2620 if (WARN_ON_ONCE(!mm_percpu_wq))
2621 return;
2622
2623 /*
2624 * Do not drain if one is already in progress unless it's specific to
2625 * a zone. Such callers are primarily CMA and memory hotplug and need
2626 * the drain to be complete when the call returns.
2627 */
2628 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2629 if (!zone)
2630 return;
2631 mutex_lock(&pcpu_drain_mutex);
2632 }
2633
2634 /*
2635 * We don't care about racing with CPU hotplug event
2636 * as offline notification will cause the notified
2637 * cpu to drain that CPU pcps and on_each_cpu_mask
2638 * disables preemption as part of its processing
2639 */
2640 for_each_online_cpu(cpu) {
2641 struct per_cpu_pageset *pcp;
2642 struct zone *z;
2643 bool has_pcps = false;
2644
2645 if (zone) {
2646 pcp = per_cpu_ptr(zone->pageset, cpu);
2647 if (pcp->pcp.count)
2648 has_pcps = true;
2649 } else {
2650 for_each_populated_zone(z) {
2651 pcp = per_cpu_ptr(z->pageset, cpu);
2652 if (pcp->pcp.count) {
2653 has_pcps = true;
2654 break;
2655 }
2656 }
2657 }
2658
2659 if (has_pcps)
2660 cpumask_set_cpu(cpu, &cpus_with_pcps);
2661 else
2662 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2663 }
2664
2665 for_each_cpu(cpu, &cpus_with_pcps) {
2666 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2667 INIT_WORK(work, drain_local_pages_wq);
2668 queue_work_on(cpu, mm_percpu_wq, work);
2669 }
2670 for_each_cpu(cpu, &cpus_with_pcps)
2671 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2672
2673 mutex_unlock(&pcpu_drain_mutex);
2674}
2675
2676#ifdef CONFIG_HIBERNATION
2677
2678/*
2679 * Touch the watchdog for every WD_PAGE_COUNT pages.
2680 */
2681#define WD_PAGE_COUNT (128*1024)
2682
2683void mark_free_pages(struct zone *zone)
2684{
2685 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2686 unsigned long flags;
2687 unsigned int order, t;
2688 struct page *page;
2689
2690 if (zone_is_empty(zone))
2691 return;
2692
2693 spin_lock_irqsave(&zone->lock, flags);
2694
2695 max_zone_pfn = zone_end_pfn(zone);
2696 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2697 if (pfn_valid(pfn)) {
2698 page = pfn_to_page(pfn);
2699
2700 if (!--page_count) {
2701 touch_nmi_watchdog();
2702 page_count = WD_PAGE_COUNT;
2703 }
2704
2705 if (page_zone(page) != zone)
2706 continue;
2707
2708 if (!swsusp_page_is_forbidden(page))
2709 swsusp_unset_page_free(page);
2710 }
2711
2712 for_each_migratetype_order(order, t) {
2713 list_for_each_entry(page,
2714 &zone->free_area[order].free_list[t], lru) {
2715 unsigned long i;
2716
2717 pfn = page_to_pfn(page);
2718 for (i = 0; i < (1UL << order); i++) {
2719 if (!--page_count) {
2720 touch_nmi_watchdog();
2721 page_count = WD_PAGE_COUNT;
2722 }
2723 swsusp_set_page_free(pfn_to_page(pfn + i));
2724 }
2725 }
2726 }
2727 spin_unlock_irqrestore(&zone->lock, flags);
2728}
2729#endif /* CONFIG_PM */
2730
2731static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2732{
2733 int migratetype;
2734
2735 if (!free_pcp_prepare(page))
2736 return false;
2737
2738 migratetype = get_pfnblock_migratetype(page, pfn);
2739 set_pcppage_migratetype(page, migratetype);
2740 return true;
2741}
2742
2743static void free_unref_page_commit(struct page *page, unsigned long pfn)
2744{
2745 struct zone *zone = page_zone(page);
2746 struct per_cpu_pages *pcp;
2747 int migratetype;
2748
2749 migratetype = get_pcppage_migratetype(page);
2750 __count_vm_event(PGFREE);
2751
2752 /*
2753 * We only track unmovable, reclaimable and movable on pcp lists.
2754 * Free ISOLATE pages back to the allocator because they are being
2755 * offlined but treat HIGHATOMIC as movable pages so we can get those
2756 * areas back if necessary. Otherwise, we may have to free
2757 * excessively into the page allocator
2758 */
2759 if (migratetype >= MIGRATE_PCPTYPES) {
2760 if (unlikely(is_migrate_isolate(migratetype))) {
2761 free_one_page(zone, page, pfn, 0, migratetype);
2762 return;
2763 }
2764 migratetype = MIGRATE_MOVABLE;
2765 }
2766
2767 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2768 list_add(&page->lru, &pcp->lists[migratetype]);
2769 pcp->count++;
2770 if (pcp->count >= pcp->high) {
2771 unsigned long batch = READ_ONCE(pcp->batch);
2772 free_pcppages_bulk(zone, batch, pcp);
2773 }
2774}
2775
2776/*
2777 * Free a 0-order page
2778 */
2779void free_unref_page(struct page *page)
2780{
2781 unsigned long flags;
2782 unsigned long pfn = page_to_pfn(page);
2783
2784 if (!free_unref_page_prepare(page, pfn))
2785 return;
2786
2787 local_irq_save(flags);
2788 free_unref_page_commit(page, pfn);
2789 local_irq_restore(flags);
2790}
2791
2792/*
2793 * Free a list of 0-order pages
2794 */
2795void free_unref_page_list(struct list_head *list)
2796{
2797 struct page *page, *next;
2798 unsigned long flags, pfn;
2799 int batch_count = 0;
2800
2801 /* Prepare pages for freeing */
2802 list_for_each_entry_safe(page, next, list, lru) {
2803 pfn = page_to_pfn(page);
2804 if (!free_unref_page_prepare(page, pfn))
2805 list_del(&page->lru);
2806 set_page_private(page, pfn);
2807 }
2808
2809 local_irq_save(flags);
2810 list_for_each_entry_safe(page, next, list, lru) {
2811 unsigned long pfn = page_private(page);
2812
2813 set_page_private(page, 0);
2814 trace_mm_page_free_batched(page);
2815 free_unref_page_commit(page, pfn);
2816
2817 /*
2818 * Guard against excessive IRQ disabled times when we get
2819 * a large list of pages to free.
2820 */
2821 if (++batch_count == SWAP_CLUSTER_MAX) {
2822 local_irq_restore(flags);
2823 batch_count = 0;
2824 local_irq_save(flags);
2825 }
2826 }
2827 local_irq_restore(flags);
2828}
2829
2830/*
2831 * split_page takes a non-compound higher-order page, and splits it into
2832 * n (1<<order) sub-pages: page[0..n]
2833 * Each sub-page must be freed individually.
2834 *
2835 * Note: this is probably too low level an operation for use in drivers.
2836 * Please consult with lkml before using this in your driver.
2837 */
2838void split_page(struct page *page, unsigned int order)
2839{
2840 int i;
2841
2842 VM_BUG_ON_PAGE(PageCompound(page), page);
2843 VM_BUG_ON_PAGE(!page_count(page), page);
2844
2845 for (i = 1; i < (1 << order); i++)
2846 set_page_refcounted(page + i);
2847 split_page_owner(page, order);
2848}
2849EXPORT_SYMBOL_GPL(split_page);
2850
2851int __isolate_free_page(struct page *page, unsigned int order)
2852{
2853 unsigned long watermark;
2854 struct zone *zone;
2855 int mt;
2856
2857 BUG_ON(!PageBuddy(page));
2858
2859 zone = page_zone(page);
2860 mt = get_pageblock_migratetype(page);
2861
2862 if (!is_migrate_isolate(mt)) {
2863 /*
2864 * Obey watermarks as if the page was being allocated. We can
2865 * emulate a high-order watermark check with a raised order-0
2866 * watermark, because we already know our high-order page
2867 * exists.
2868 */
2869 watermark = min_wmark_pages(zone) + (1UL << order);
2870 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2871 return 0;
2872
2873 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2874 }
2875
2876 /* Remove page from free list */
2877 list_del(&page->lru);
2878 zone->free_area[order].nr_free--;
2879 rmv_page_order(page);
2880
2881 /*
2882 * Set the pageblock if the isolated page is at least half of a
2883 * pageblock
2884 */
2885 if (order >= pageblock_order - 1) {
2886 struct page *endpage = page + (1 << order) - 1;
2887 for (; page < endpage; page += pageblock_nr_pages) {
2888 int mt = get_pageblock_migratetype(page);
2889 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2890 && !is_migrate_highatomic(mt))
2891 set_pageblock_migratetype(page,
2892 MIGRATE_MOVABLE);
2893 }
2894 }
2895
2896
2897 return 1UL << order;
2898}
2899
2900/*
2901 * Update NUMA hit/miss statistics
2902 *
2903 * Must be called with interrupts disabled.
2904 */
2905static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2906{
2907#ifdef CONFIG_NUMA
2908 enum numa_stat_item local_stat = NUMA_LOCAL;
2909
2910 /* skip numa counters update if numa stats is disabled */
2911 if (!static_branch_likely(&vm_numa_stat_key))
2912 return;
2913
2914 if (z->node != numa_node_id())
2915 local_stat = NUMA_OTHER;
2916
2917 if (z->node == preferred_zone->node)
2918 __inc_numa_state(z, NUMA_HIT);
2919 else {
2920 __inc_numa_state(z, NUMA_MISS);
2921 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2922 }
2923 __inc_numa_state(z, local_stat);
2924#endif
2925}
2926
2927/* Remove page from the per-cpu list, caller must protect the list */
2928static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2929 struct per_cpu_pages *pcp,
2930 struct list_head *list)
2931{
2932 struct page *page;
2933
2934 do {
2935 if (list_empty(list)) {
2936 pcp->count += rmqueue_bulk(zone, 0,
2937 pcp->batch, list,
2938 migratetype);
2939 if (unlikely(list_empty(list)))
2940 return NULL;
2941 }
2942
2943 page = list_first_entry(list, struct page, lru);
2944 list_del(&page->lru);
2945 pcp->count--;
2946 } while (check_new_pcp(page));
2947
2948 return page;
2949}
2950
2951/* Lock and remove page from the per-cpu list */
2952static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2953 struct zone *zone, unsigned int order,
2954 gfp_t gfp_flags, int migratetype)
2955{
2956 struct per_cpu_pages *pcp;
2957 struct list_head *list;
2958 struct page *page;
2959 unsigned long flags;
2960
2961 local_irq_save(flags);
2962 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2963 list = &pcp->lists[migratetype];
2964 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2965 if (page) {
2966 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2967 zone_statistics(preferred_zone, zone);
2968 }
2969 local_irq_restore(flags);
2970 return page;
2971}
2972
2973/*
2974 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2975 */
2976static inline
2977struct page *rmqueue(struct zone *preferred_zone,
2978 struct zone *zone, unsigned int order,
2979 gfp_t gfp_flags, unsigned int alloc_flags,
2980 int migratetype)
2981{
2982 unsigned long flags;
2983 struct page *page;
2984
2985 if (likely(order == 0)) {
2986 page = rmqueue_pcplist(preferred_zone, zone, order,
2987 gfp_flags, migratetype);
2988 goto out;
2989 }
2990
2991 /*
2992 * We most definitely don't want callers attempting to
2993 * allocate greater than order-1 page units with __GFP_NOFAIL.
2994 */
2995 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2996 spin_lock_irqsave(&zone->lock, flags);
2997
2998 do {
2999 page = NULL;
3000 if (alloc_flags & ALLOC_HARDER) {
3001 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3002 if (page)
3003 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3004 }
3005 if (!page)
3006 page = __rmqueue(zone, order, migratetype);
3007 } while (page && check_new_pages(page, order));
3008 spin_unlock(&zone->lock);
3009 if (!page)
3010 goto failed;
3011 __mod_zone_freepage_state(zone, -(1 << order),
3012 get_pcppage_migratetype(page));
3013
3014 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3015 zone_statistics(preferred_zone, zone);
3016 local_irq_restore(flags);
3017
3018out:
3019 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3020 return page;
3021
3022failed:
3023 local_irq_restore(flags);
3024 return NULL;
3025}
3026
3027#ifdef CONFIG_FAIL_PAGE_ALLOC
3028
3029static struct {
3030 struct fault_attr attr;
3031
3032 bool ignore_gfp_highmem;
3033 bool ignore_gfp_reclaim;
3034 u32 min_order;
3035} fail_page_alloc = {
3036 .attr = FAULT_ATTR_INITIALIZER,
3037 .ignore_gfp_reclaim = true,
3038 .ignore_gfp_highmem = true,
3039 .min_order = 1,
3040};
3041
3042static int __init setup_fail_page_alloc(char *str)
3043{
3044 return setup_fault_attr(&fail_page_alloc.attr, str);
3045}
3046__setup("fail_page_alloc=", setup_fail_page_alloc);
3047
3048static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3049{
3050 if (order < fail_page_alloc.min_order)
3051 return false;
3052 if (gfp_mask & __GFP_NOFAIL)
3053 return false;
3054 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3055 return false;
3056 if (fail_page_alloc.ignore_gfp_reclaim &&
3057 (gfp_mask & __GFP_DIRECT_RECLAIM))
3058 return false;
3059
3060 return should_fail(&fail_page_alloc.attr, 1 << order);
3061}
3062
3063#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3064
3065static int __init fail_page_alloc_debugfs(void)
3066{
3067 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3068 struct dentry *dir;
3069
3070 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3071 &fail_page_alloc.attr);
3072 if (IS_ERR(dir))
3073 return PTR_ERR(dir);
3074
3075 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3076 &fail_page_alloc.ignore_gfp_reclaim))
3077 goto fail;
3078 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3079 &fail_page_alloc.ignore_gfp_highmem))
3080 goto fail;
3081 if (!debugfs_create_u32("min-order", mode, dir,
3082 &fail_page_alloc.min_order))
3083 goto fail;
3084
3085 return 0;
3086fail:
3087 debugfs_remove_recursive(dir);
3088
3089 return -ENOMEM;
3090}
3091
3092late_initcall(fail_page_alloc_debugfs);
3093
3094#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3095
3096#else /* CONFIG_FAIL_PAGE_ALLOC */
3097
3098static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3099{
3100 return false;
3101}
3102
3103#endif /* CONFIG_FAIL_PAGE_ALLOC */
3104
3105/*
3106 * Return true if free base pages are above 'mark'. For high-order checks it
3107 * will return true of the order-0 watermark is reached and there is at least
3108 * one free page of a suitable size. Checking now avoids taking the zone lock
3109 * to check in the allocation paths if no pages are free.
3110 */
3111bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3112 int classzone_idx, unsigned int alloc_flags,
3113 long free_pages)
3114{
3115 long min = mark;
3116 int o;
3117 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3118
3119 /* free_pages may go negative - that's OK */
3120 free_pages -= (1 << order) - 1;
3121
3122 if (alloc_flags & ALLOC_HIGH)
3123 min -= min / 2;
3124
3125 /*
3126 * If the caller does not have rights to ALLOC_HARDER then subtract
3127 * the high-atomic reserves. This will over-estimate the size of the
3128 * atomic reserve but it avoids a search.
3129 */
3130 if (likely(!alloc_harder)) {
3131 free_pages -= z->nr_reserved_highatomic;
3132 } else {
3133 /*
3134 * OOM victims can try even harder than normal ALLOC_HARDER
3135 * users on the grounds that it's definitely going to be in
3136 * the exit path shortly and free memory. Any allocation it
3137 * makes during the free path will be small and short-lived.
3138 */
3139 if (alloc_flags & ALLOC_OOM)
3140 min -= min / 2;
3141 else
3142 min -= min / 4;
3143 }
3144
3145
3146#ifdef CONFIG_CMA
3147 /* If allocation can't use CMA areas don't use free CMA pages */
3148 if (!(alloc_flags & ALLOC_CMA))
3149 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3150#endif
3151
3152 /*
3153 * Check watermarks for an order-0 allocation request. If these
3154 * are not met, then a high-order request also cannot go ahead
3155 * even if a suitable page happened to be free.
3156 */
3157 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3158 return false;
3159
3160 /* If this is an order-0 request then the watermark is fine */
3161 if (!order)
3162 return true;
3163
3164 /* For a high-order request, check at least one suitable page is free */
3165 for (o = order; o < MAX_ORDER; o++) {
3166 struct free_area *area = &z->free_area[o];
3167 int mt;
3168
3169 if (!area->nr_free)
3170 continue;
3171
3172 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3173 if (!list_empty(&area->free_list[mt]))
3174 return true;
3175 }
3176
3177#ifdef CONFIG_CMA
3178 if ((alloc_flags & ALLOC_CMA) &&
3179 !list_empty(&area->free_list[MIGRATE_CMA])) {
3180 return true;
3181 }
3182#endif
3183 if (alloc_harder &&
3184 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3185 return true;
3186 }
3187 return false;
3188}
3189
3190bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3191 int classzone_idx, unsigned int alloc_flags)
3192{
3193 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3194 zone_page_state(z, NR_FREE_PAGES));
3195}
3196
3197static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3198 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3199{
3200 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3201 long cma_pages = 0;
3202
3203#ifdef CONFIG_CMA
3204 /* If allocation can't use CMA areas don't use free CMA pages */
3205 if (!(alloc_flags & ALLOC_CMA))
3206 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3207#endif
3208
3209 /*
3210 * Fast check for order-0 only. If this fails then the reserves
3211 * need to be calculated. There is a corner case where the check
3212 * passes but only the high-order atomic reserve are free. If
3213 * the caller is !atomic then it'll uselessly search the free
3214 * list. That corner case is then slower but it is harmless.
3215 */
3216 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3217 return true;
3218
3219 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3220 free_pages);
3221}
3222
3223bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3224 unsigned long mark, int classzone_idx)
3225{
3226 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3227
3228 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3229 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3230
3231 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3232 free_pages);
3233}
3234
3235#ifdef CONFIG_NUMA
3236static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3237{
3238 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3239 RECLAIM_DISTANCE;
3240}
3241#else /* CONFIG_NUMA */
3242static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3243{
3244 return true;
3245}
3246#endif /* CONFIG_NUMA */
3247
3248/*
3249 * get_page_from_freelist goes through the zonelist trying to allocate
3250 * a page.
3251 */
3252static struct page *
3253get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3254 const struct alloc_context *ac)
3255{
3256 struct zoneref *z = ac->preferred_zoneref;
3257 struct zone *zone;
3258 struct pglist_data *last_pgdat_dirty_limit = NULL;
3259
3260 /*
3261 * Scan zonelist, looking for a zone with enough free.
3262 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3263 */
3264 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3265 ac->nodemask) {
3266 struct page *page;
3267 unsigned long mark;
3268
3269 if (cpusets_enabled() &&
3270 (alloc_flags & ALLOC_CPUSET) &&
3271 !__cpuset_zone_allowed(zone, gfp_mask))
3272 continue;
3273 /*
3274 * When allocating a page cache page for writing, we
3275 * want to get it from a node that is within its dirty
3276 * limit, such that no single node holds more than its
3277 * proportional share of globally allowed dirty pages.
3278 * The dirty limits take into account the node's
3279 * lowmem reserves and high watermark so that kswapd
3280 * should be able to balance it without having to
3281 * write pages from its LRU list.
3282 *
3283 * XXX: For now, allow allocations to potentially
3284 * exceed the per-node dirty limit in the slowpath
3285 * (spread_dirty_pages unset) before going into reclaim,
3286 * which is important when on a NUMA setup the allowed
3287 * nodes are together not big enough to reach the
3288 * global limit. The proper fix for these situations
3289 * will require awareness of nodes in the
3290 * dirty-throttling and the flusher threads.
3291 */
3292 if (ac->spread_dirty_pages) {
3293 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3294 continue;
3295
3296 if (!node_dirty_ok(zone->zone_pgdat)) {
3297 last_pgdat_dirty_limit = zone->zone_pgdat;
3298 continue;
3299 }
3300 }
3301
3302 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3303 if (!zone_watermark_fast(zone, order, mark,
3304 ac_classzone_idx(ac), alloc_flags)) {
3305 int ret;
3306
3307#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3308 /*
3309 * Watermark failed for this zone, but see if we can
3310 * grow this zone if it contains deferred pages.
3311 */
3312 if (static_branch_unlikely(&deferred_pages)) {
3313 if (_deferred_grow_zone(zone, order))
3314 goto try_this_zone;
3315 }
3316#endif
3317 /* Checked here to keep the fast path fast */
3318 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3319 if (alloc_flags & ALLOC_NO_WATERMARKS)
3320 goto try_this_zone;
3321
3322 if (node_reclaim_mode == 0 ||
3323 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3324 continue;
3325
3326 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3327 switch (ret) {
3328 case NODE_RECLAIM_NOSCAN:
3329 /* did not scan */
3330 continue;
3331 case NODE_RECLAIM_FULL:
3332 /* scanned but unreclaimable */
3333 continue;
3334 default:
3335 /* did we reclaim enough */
3336 if (zone_watermark_ok(zone, order, mark,
3337 ac_classzone_idx(ac), alloc_flags))
3338 goto try_this_zone;
3339
3340 continue;
3341 }
3342 }
3343
3344try_this_zone:
3345 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3346 gfp_mask, alloc_flags, ac->migratetype);
3347 if (page) {
3348 prep_new_page(page, order, gfp_mask, alloc_flags);
3349
3350 /*
3351 * If this is a high-order atomic allocation then check
3352 * if the pageblock should be reserved for the future
3353 */
3354 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3355 reserve_highatomic_pageblock(page, zone, order);
3356
3357 return page;
3358 } else {
3359#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3360 /* Try again if zone has deferred pages */
3361 if (static_branch_unlikely(&deferred_pages)) {
3362 if (_deferred_grow_zone(zone, order))
3363 goto try_this_zone;
3364 }
3365#endif
3366 }
3367 }
3368
3369 return NULL;
3370}
3371
3372/*
3373 * Large machines with many possible nodes should not always dump per-node
3374 * meminfo in irq context.
3375 */
3376static inline bool should_suppress_show_mem(void)
3377{
3378 bool ret = false;
3379
3380#if NODES_SHIFT > 8
3381 ret = in_interrupt();
3382#endif
3383 return ret;
3384}
3385
3386static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3387{
3388 unsigned int filter = SHOW_MEM_FILTER_NODES;
3389 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3390
3391 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3392 return;
3393
3394 /*
3395 * This documents exceptions given to allocations in certain
3396 * contexts that are allowed to allocate outside current's set
3397 * of allowed nodes.
3398 */
3399 if (!(gfp_mask & __GFP_NOMEMALLOC))
3400 if (tsk_is_oom_victim(current) ||
3401 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3402 filter &= ~SHOW_MEM_FILTER_NODES;
3403 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3404 filter &= ~SHOW_MEM_FILTER_NODES;
3405
3406 show_mem(filter, nodemask);
3407}
3408
3409void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3410{
3411 struct va_format vaf;
3412 va_list args;
3413 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3414 DEFAULT_RATELIMIT_BURST);
3415
3416 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3417 return;
3418
3419 va_start(args, fmt);
3420 vaf.fmt = fmt;
3421 vaf.va = &args;
3422 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3423 current->comm, &vaf, gfp_mask, &gfp_mask,
3424 nodemask_pr_args(nodemask));
3425 va_end(args);
3426
3427 cpuset_print_current_mems_allowed();
3428
3429 dump_stack();
3430 warn_alloc_show_mem(gfp_mask, nodemask);
3431}
3432
3433static inline struct page *
3434__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3435 unsigned int alloc_flags,
3436 const struct alloc_context *ac)
3437{
3438 struct page *page;
3439
3440 page = get_page_from_freelist(gfp_mask, order,
3441 alloc_flags|ALLOC_CPUSET, ac);
3442 /*
3443 * fallback to ignore cpuset restriction if our nodes
3444 * are depleted
3445 */
3446 if (!page)
3447 page = get_page_from_freelist(gfp_mask, order,
3448 alloc_flags, ac);
3449
3450 return page;
3451}
3452
3453static inline struct page *
3454__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3455 const struct alloc_context *ac, unsigned long *did_some_progress)
3456{
3457 struct oom_control oc = {
3458 .zonelist = ac->zonelist,
3459 .nodemask = ac->nodemask,
3460 .memcg = NULL,
3461 .gfp_mask = gfp_mask,
3462 .order = order,
3463 };
3464 struct page *page;
3465
3466 *did_some_progress = 0;
3467
3468 /*
3469 * Acquire the oom lock. If that fails, somebody else is
3470 * making progress for us.
3471 */
3472 if (!mutex_trylock(&oom_lock)) {
3473 *did_some_progress = 1;
3474 schedule_timeout_uninterruptible(1);
3475 return NULL;
3476 }
3477
3478 /*
3479 * Go through the zonelist yet one more time, keep very high watermark
3480 * here, this is only to catch a parallel oom killing, we must fail if
3481 * we're still under heavy pressure. But make sure that this reclaim
3482 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3483 * allocation which will never fail due to oom_lock already held.
3484 */
3485 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3486 ~__GFP_DIRECT_RECLAIM, order,
3487 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3488 if (page)
3489 goto out;
3490
3491 /* Coredumps can quickly deplete all memory reserves */
3492 if (current->flags & PF_DUMPCORE)
3493 goto out;
3494 /* The OOM killer will not help higher order allocs */
3495 if (order > PAGE_ALLOC_COSTLY_ORDER)
3496 goto out;
3497 /*
3498 * We have already exhausted all our reclaim opportunities without any
3499 * success so it is time to admit defeat. We will skip the OOM killer
3500 * because it is very likely that the caller has a more reasonable
3501 * fallback than shooting a random task.
3502 */
3503 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3504 goto out;
3505 /* The OOM killer does not needlessly kill tasks for lowmem */
3506 if (ac->high_zoneidx < ZONE_NORMAL)
3507 goto out;
3508 if (pm_suspended_storage())
3509 goto out;
3510 /*
3511 * XXX: GFP_NOFS allocations should rather fail than rely on
3512 * other request to make a forward progress.
3513 * We are in an unfortunate situation where out_of_memory cannot
3514 * do much for this context but let's try it to at least get
3515 * access to memory reserved if the current task is killed (see
3516 * out_of_memory). Once filesystems are ready to handle allocation
3517 * failures more gracefully we should just bail out here.
3518 */
3519
3520 /* The OOM killer may not free memory on a specific node */
3521 if (gfp_mask & __GFP_THISNODE)
3522 goto out;
3523
3524 /* Exhausted what can be done so it's blame time */
3525 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3526 *did_some_progress = 1;
3527
3528 /*
3529 * Help non-failing allocations by giving them access to memory
3530 * reserves
3531 */
3532 if (gfp_mask & __GFP_NOFAIL)
3533 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3534 ALLOC_NO_WATERMARKS, ac);
3535 }
3536out:
3537 mutex_unlock(&oom_lock);
3538 return page;
3539}
3540
3541/*
3542 * Maximum number of compaction retries wit a progress before OOM
3543 * killer is consider as the only way to move forward.
3544 */
3545#define MAX_COMPACT_RETRIES 16
3546
3547#ifdef CONFIG_COMPACTION
3548/* Try memory compaction for high-order allocations before reclaim */
3549static struct page *
3550__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3551 unsigned int alloc_flags, const struct alloc_context *ac,
3552 enum compact_priority prio, enum compact_result *compact_result)
3553{
3554 struct page *page;
3555 unsigned int noreclaim_flag;
3556
3557 if (!order)
3558 return NULL;
3559
3560 noreclaim_flag = memalloc_noreclaim_save();
3561 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3562 prio);
3563 memalloc_noreclaim_restore(noreclaim_flag);
3564
3565 if (*compact_result <= COMPACT_INACTIVE)
3566 return NULL;
3567
3568 /*
3569 * At least in one zone compaction wasn't deferred or skipped, so let's
3570 * count a compaction stall
3571 */
3572 count_vm_event(COMPACTSTALL);
3573
3574 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3575
3576 if (page) {
3577 struct zone *zone = page_zone(page);
3578
3579 zone->compact_blockskip_flush = false;
3580 compaction_defer_reset(zone, order, true);
3581 count_vm_event(COMPACTSUCCESS);
3582 return page;
3583 }
3584
3585 /*
3586 * It's bad if compaction run occurs and fails. The most likely reason
3587 * is that pages exist, but not enough to satisfy watermarks.
3588 */
3589 count_vm_event(COMPACTFAIL);
3590
3591 cond_resched();
3592
3593 return NULL;
3594}
3595
3596static inline bool
3597should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3598 enum compact_result compact_result,
3599 enum compact_priority *compact_priority,
3600 int *compaction_retries)
3601{
3602 int max_retries = MAX_COMPACT_RETRIES;
3603 int min_priority;
3604 bool ret = false;
3605 int retries = *compaction_retries;
3606 enum compact_priority priority = *compact_priority;
3607
3608 if (!order)
3609 return false;
3610
3611 if (compaction_made_progress(compact_result))
3612 (*compaction_retries)++;
3613
3614 /*
3615 * compaction considers all the zone as desperately out of memory
3616 * so it doesn't really make much sense to retry except when the
3617 * failure could be caused by insufficient priority
3618 */
3619 if (compaction_failed(compact_result))
3620 goto check_priority;
3621
3622 /*
3623 * make sure the compaction wasn't deferred or didn't bail out early
3624 * due to locks contention before we declare that we should give up.
3625 * But do not retry if the given zonelist is not suitable for
3626 * compaction.
3627 */
3628 if (compaction_withdrawn(compact_result)) {
3629 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3630 goto out;
3631 }
3632
3633 /*
3634 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3635 * costly ones because they are de facto nofail and invoke OOM
3636 * killer to move on while costly can fail and users are ready
3637 * to cope with that. 1/4 retries is rather arbitrary but we
3638 * would need much more detailed feedback from compaction to
3639 * make a better decision.
3640 */
3641 if (order > PAGE_ALLOC_COSTLY_ORDER)
3642 max_retries /= 4;
3643 if (*compaction_retries <= max_retries) {
3644 ret = true;
3645 goto out;
3646 }
3647
3648 /*
3649 * Make sure there are attempts at the highest priority if we exhausted
3650 * all retries or failed at the lower priorities.
3651 */
3652check_priority:
3653 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3654 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3655
3656 if (*compact_priority > min_priority) {
3657 (*compact_priority)--;
3658 *compaction_retries = 0;
3659 ret = true;
3660 }
3661out:
3662 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3663 return ret;
3664}
3665#else
3666static inline struct page *
3667__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3668 unsigned int alloc_flags, const struct alloc_context *ac,
3669 enum compact_priority prio, enum compact_result *compact_result)
3670{
3671 *compact_result = COMPACT_SKIPPED;
3672 return NULL;
3673}
3674
3675static inline bool
3676should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3677 enum compact_result compact_result,
3678 enum compact_priority *compact_priority,
3679 int *compaction_retries)
3680{
3681 struct zone *zone;
3682 struct zoneref *z;
3683
3684 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3685 return false;
3686
3687 /*
3688 * There are setups with compaction disabled which would prefer to loop
3689 * inside the allocator rather than hit the oom killer prematurely.
3690 * Let's give them a good hope and keep retrying while the order-0
3691 * watermarks are OK.
3692 */
3693 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3694 ac->nodemask) {
3695 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3696 ac_classzone_idx(ac), alloc_flags))
3697 return true;
3698 }
3699 return false;
3700}
3701#endif /* CONFIG_COMPACTION */
3702
3703#ifdef CONFIG_LOCKDEP
3704struct lockdep_map __fs_reclaim_map =
3705 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3706
3707static bool __need_fs_reclaim(gfp_t gfp_mask)
3708{
3709 gfp_mask = current_gfp_context(gfp_mask);
3710
3711 /* no reclaim without waiting on it */
3712 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3713 return false;
3714
3715 /* this guy won't enter reclaim */
3716 if (current->flags & PF_MEMALLOC)
3717 return false;
3718
3719 /* We're only interested __GFP_FS allocations for now */
3720 if (!(gfp_mask & __GFP_FS))
3721 return false;
3722
3723 if (gfp_mask & __GFP_NOLOCKDEP)
3724 return false;
3725
3726 return true;
3727}
3728
3729void fs_reclaim_acquire(gfp_t gfp_mask)
3730{
3731 if (__need_fs_reclaim(gfp_mask))
3732 lock_map_acquire(&__fs_reclaim_map);
3733}
3734EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3735
3736void fs_reclaim_release(gfp_t gfp_mask)
3737{
3738 if (__need_fs_reclaim(gfp_mask))
3739 lock_map_release(&__fs_reclaim_map);
3740}
3741EXPORT_SYMBOL_GPL(fs_reclaim_release);
3742#endif
3743
3744/* Perform direct synchronous page reclaim */
3745static int
3746__perform_reclaim(gfp_t gfp_mask, unsigned int order,
3747 const struct alloc_context *ac)
3748{
3749 struct reclaim_state reclaim_state;
3750 int progress;
3751 unsigned int noreclaim_flag;
3752
3753 cond_resched();
3754
3755 /* We now go into synchronous reclaim */
3756 cpuset_memory_pressure_bump();
3757 noreclaim_flag = memalloc_noreclaim_save();
3758 fs_reclaim_acquire(gfp_mask);
3759 reclaim_state.reclaimed_slab = 0;
3760 current->reclaim_state = &reclaim_state;
3761
3762 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3763 ac->nodemask);
3764
3765 current->reclaim_state = NULL;
3766 fs_reclaim_release(gfp_mask);
3767 memalloc_noreclaim_restore(noreclaim_flag);
3768
3769 cond_resched();
3770
3771 return progress;
3772}
3773
3774/* The really slow allocator path where we enter direct reclaim */
3775static inline struct page *
3776__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3777 unsigned int alloc_flags, const struct alloc_context *ac,
3778 unsigned long *did_some_progress)
3779{
3780 struct page *page = NULL;
3781 bool drained = false;
3782
3783 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3784 if (unlikely(!(*did_some_progress)))
3785 return NULL;
3786
3787retry:
3788 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3789
3790 /*
3791 * If an allocation failed after direct reclaim, it could be because
3792 * pages are pinned on the per-cpu lists or in high alloc reserves.
3793 * Shrink them them and try again
3794 */
3795 if (!page && !drained) {
3796 unreserve_highatomic_pageblock(ac, false);
3797 drain_all_pages(NULL);
3798 drained = true;
3799 goto retry;
3800 }
3801
3802 return page;
3803}
3804
3805static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3806 const struct alloc_context *ac)
3807{
3808 struct zoneref *z;
3809 struct zone *zone;
3810 pg_data_t *last_pgdat = NULL;
3811 enum zone_type high_zoneidx = ac->high_zoneidx;
3812
3813 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3814 ac->nodemask) {
3815 if (last_pgdat != zone->zone_pgdat)
3816 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3817 last_pgdat = zone->zone_pgdat;
3818 }
3819}
3820
3821static inline unsigned int
3822gfp_to_alloc_flags(gfp_t gfp_mask)
3823{
3824 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3825
3826 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3827 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3828
3829 /*
3830 * The caller may dip into page reserves a bit more if the caller
3831 * cannot run direct reclaim, or if the caller has realtime scheduling
3832 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3833 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3834 */
3835 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3836
3837 if (gfp_mask & __GFP_ATOMIC) {
3838 /*
3839 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3840 * if it can't schedule.
3841 */
3842 if (!(gfp_mask & __GFP_NOMEMALLOC))
3843 alloc_flags |= ALLOC_HARDER;
3844 /*
3845 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3846 * comment for __cpuset_node_allowed().
3847 */
3848 alloc_flags &= ~ALLOC_CPUSET;
3849 } else if (unlikely(rt_task(current)) && !in_interrupt())
3850 alloc_flags |= ALLOC_HARDER;
3851
3852#ifdef CONFIG_CMA
3853 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3854 alloc_flags |= ALLOC_CMA;
3855#endif
3856 return alloc_flags;
3857}
3858
3859static bool oom_reserves_allowed(struct task_struct *tsk)
3860{
3861 if (!tsk_is_oom_victim(tsk))
3862 return false;
3863
3864 /*
3865 * !MMU doesn't have oom reaper so give access to memory reserves
3866 * only to the thread with TIF_MEMDIE set
3867 */
3868 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3869 return false;
3870
3871 return true;
3872}
3873
3874/*
3875 * Distinguish requests which really need access to full memory
3876 * reserves from oom victims which can live with a portion of it
3877 */
3878static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3879{
3880 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3881 return 0;
3882 if (gfp_mask & __GFP_MEMALLOC)
3883 return ALLOC_NO_WATERMARKS;
3884 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3885 return ALLOC_NO_WATERMARKS;
3886 if (!in_interrupt()) {
3887 if (current->flags & PF_MEMALLOC)
3888 return ALLOC_NO_WATERMARKS;
3889 else if (oom_reserves_allowed(current))
3890 return ALLOC_OOM;
3891 }
3892
3893 return 0;
3894}
3895
3896bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3897{
3898 return !!__gfp_pfmemalloc_flags(gfp_mask);
3899}
3900
3901/*
3902 * Checks whether it makes sense to retry the reclaim to make a forward progress
3903 * for the given allocation request.
3904 *
3905 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3906 * without success, or when we couldn't even meet the watermark if we
3907 * reclaimed all remaining pages on the LRU lists.
3908 *
3909 * Returns true if a retry is viable or false to enter the oom path.
3910 */
3911static inline bool
3912should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3913 struct alloc_context *ac, int alloc_flags,
3914 bool did_some_progress, int *no_progress_loops)
3915{
3916 struct zone *zone;
3917 struct zoneref *z;
3918
3919 /*
3920 * Costly allocations might have made a progress but this doesn't mean
3921 * their order will become available due to high fragmentation so
3922 * always increment the no progress counter for them
3923 */
3924 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3925 *no_progress_loops = 0;
3926 else
3927 (*no_progress_loops)++;
3928
3929 /*
3930 * Make sure we converge to OOM if we cannot make any progress
3931 * several times in the row.
3932 */
3933 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3934 /* Before OOM, exhaust highatomic_reserve */
3935 return unreserve_highatomic_pageblock(ac, true);
3936 }
3937
3938 /*
3939 * Keep reclaiming pages while there is a chance this will lead
3940 * somewhere. If none of the target zones can satisfy our allocation
3941 * request even if all reclaimable pages are considered then we are
3942 * screwed and have to go OOM.
3943 */
3944 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3945 ac->nodemask) {
3946 unsigned long available;
3947 unsigned long reclaimable;
3948 unsigned long min_wmark = min_wmark_pages(zone);
3949 bool wmark;
3950
3951 available = reclaimable = zone_reclaimable_pages(zone);
3952 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3953
3954 /*
3955 * Would the allocation succeed if we reclaimed all
3956 * reclaimable pages?
3957 */
3958 wmark = __zone_watermark_ok(zone, order, min_wmark,
3959 ac_classzone_idx(ac), alloc_flags, available);
3960 trace_reclaim_retry_zone(z, order, reclaimable,
3961 available, min_wmark, *no_progress_loops, wmark);
3962 if (wmark) {
3963 /*
3964 * If we didn't make any progress and have a lot of
3965 * dirty + writeback pages then we should wait for
3966 * an IO to complete to slow down the reclaim and
3967 * prevent from pre mature OOM
3968 */
3969 if (!did_some_progress) {
3970 unsigned long write_pending;
3971
3972 write_pending = zone_page_state_snapshot(zone,
3973 NR_ZONE_WRITE_PENDING);
3974
3975 if (2 * write_pending > reclaimable) {
3976 congestion_wait(BLK_RW_ASYNC, HZ/10);
3977 return true;
3978 }
3979 }
3980
3981 /*
3982 * Memory allocation/reclaim might be called from a WQ
3983 * context and the current implementation of the WQ
3984 * concurrency control doesn't recognize that
3985 * a particular WQ is congested if the worker thread is
3986 * looping without ever sleeping. Therefore we have to
3987 * do a short sleep here rather than calling
3988 * cond_resched().
3989 */
3990 if (current->flags & PF_WQ_WORKER)
3991 schedule_timeout_uninterruptible(1);
3992 else
3993 cond_resched();
3994
3995 return true;
3996 }
3997 }
3998
3999 return false;
4000}
4001
4002static inline bool
4003check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4004{
4005 /*
4006 * It's possible that cpuset's mems_allowed and the nodemask from
4007 * mempolicy don't intersect. This should be normally dealt with by
4008 * policy_nodemask(), but it's possible to race with cpuset update in
4009 * such a way the check therein was true, and then it became false
4010 * before we got our cpuset_mems_cookie here.
4011 * This assumes that for all allocations, ac->nodemask can come only
4012 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4013 * when it does not intersect with the cpuset restrictions) or the
4014 * caller can deal with a violated nodemask.
4015 */
4016 if (cpusets_enabled() && ac->nodemask &&
4017 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4018 ac->nodemask = NULL;
4019 return true;
4020 }
4021
4022 /*
4023 * When updating a task's mems_allowed or mempolicy nodemask, it is
4024 * possible to race with parallel threads in such a way that our
4025 * allocation can fail while the mask is being updated. If we are about
4026 * to fail, check if the cpuset changed during allocation and if so,
4027 * retry.
4028 */
4029 if (read_mems_allowed_retry(cpuset_mems_cookie))
4030 return true;
4031
4032 return false;
4033}
4034
4035static inline struct page *
4036__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4037 struct alloc_context *ac)
4038{
4039 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4040 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4041 struct page *page = NULL;
4042 unsigned int alloc_flags;
4043 unsigned long did_some_progress;
4044 enum compact_priority compact_priority;
4045 enum compact_result compact_result;
4046 int compaction_retries;
4047 int no_progress_loops;
4048 unsigned int cpuset_mems_cookie;
4049 int reserve_flags;
4050
4051 /*
4052 * In the slowpath, we sanity check order to avoid ever trying to
4053 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
4054 * be using allocators in order of preference for an area that is
4055 * too large.
4056 */
4057 if (order >= MAX_ORDER) {
4058 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4059 return NULL;
4060 }
4061
4062 /*
4063 * We also sanity check to catch abuse of atomic reserves being used by
4064 * callers that are not in atomic context.
4065 */
4066 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4067 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4068 gfp_mask &= ~__GFP_ATOMIC;
4069
4070retry_cpuset:
4071 compaction_retries = 0;
4072 no_progress_loops = 0;
4073 compact_priority = DEF_COMPACT_PRIORITY;
4074 cpuset_mems_cookie = read_mems_allowed_begin();
4075
4076 /*
4077 * The fast path uses conservative alloc_flags to succeed only until
4078 * kswapd needs to be woken up, and to avoid the cost of setting up
4079 * alloc_flags precisely. So we do that now.
4080 */
4081 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4082
4083 /*
4084 * We need to recalculate the starting point for the zonelist iterator
4085 * because we might have used different nodemask in the fast path, or
4086 * there was a cpuset modification and we are retrying - otherwise we
4087 * could end up iterating over non-eligible zones endlessly.
4088 */
4089 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4090 ac->high_zoneidx, ac->nodemask);
4091 if (!ac->preferred_zoneref->zone)
4092 goto nopage;
4093
4094 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4095 wake_all_kswapds(order, gfp_mask, ac);
4096
4097 /*
4098 * The adjusted alloc_flags might result in immediate success, so try
4099 * that first
4100 */
4101 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4102 if (page)
4103 goto got_pg;
4104
4105 /*
4106 * For costly allocations, try direct compaction first, as it's likely
4107 * that we have enough base pages and don't need to reclaim. For non-
4108 * movable high-order allocations, do that as well, as compaction will
4109 * try prevent permanent fragmentation by migrating from blocks of the
4110 * same migratetype.
4111 * Don't try this for allocations that are allowed to ignore
4112 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4113 */
4114 if (can_direct_reclaim &&
4115 (costly_order ||
4116 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4117 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4118 page = __alloc_pages_direct_compact(gfp_mask, order,
4119 alloc_flags, ac,
4120 INIT_COMPACT_PRIORITY,
4121 &compact_result);
4122 if (page)
4123 goto got_pg;
4124
4125 /*
4126 * Checks for costly allocations with __GFP_NORETRY, which
4127 * includes THP page fault allocations
4128 */
4129 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4130 /*
4131 * If compaction is deferred for high-order allocations,
4132 * it is because sync compaction recently failed. If
4133 * this is the case and the caller requested a THP
4134 * allocation, we do not want to heavily disrupt the
4135 * system, so we fail the allocation instead of entering
4136 * direct reclaim.
4137 */
4138 if (compact_result == COMPACT_DEFERRED)
4139 goto nopage;
4140
4141 /*
4142 * Looks like reclaim/compaction is worth trying, but
4143 * sync compaction could be very expensive, so keep
4144 * using async compaction.
4145 */
4146 compact_priority = INIT_COMPACT_PRIORITY;
4147 }
4148 }
4149
4150retry:
4151 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4152 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4153 wake_all_kswapds(order, gfp_mask, ac);
4154
4155 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4156 if (reserve_flags)
4157 alloc_flags = reserve_flags;
4158
4159 /*
4160 * Reset the zonelist iterators if memory policies can be ignored.
4161 * These allocations are high priority and system rather than user
4162 * orientated.
4163 */
4164 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4165 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
4166 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4167 ac->high_zoneidx, ac->nodemask);
4168 }
4169
4170 /* Attempt with potentially adjusted zonelist and alloc_flags */
4171 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4172 if (page)
4173 goto got_pg;
4174
4175 /* Caller is not willing to reclaim, we can't balance anything */
4176 if (!can_direct_reclaim)
4177 goto nopage;
4178
4179 /* Avoid recursion of direct reclaim */
4180 if (current->flags & PF_MEMALLOC)
4181 goto nopage;
4182
4183 /* Try direct reclaim and then allocating */
4184 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4185 &did_some_progress);
4186 if (page)
4187 goto got_pg;
4188
4189 /* Try direct compaction and then allocating */
4190 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4191 compact_priority, &compact_result);
4192 if (page)
4193 goto got_pg;
4194
4195 /* Do not loop if specifically requested */
4196 if (gfp_mask & __GFP_NORETRY)
4197 goto nopage;
4198
4199 /*
4200 * Do not retry costly high order allocations unless they are
4201 * __GFP_RETRY_MAYFAIL
4202 */
4203 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4204 goto nopage;
4205
4206 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4207 did_some_progress > 0, &no_progress_loops))
4208 goto retry;
4209
4210 /*
4211 * It doesn't make any sense to retry for the compaction if the order-0
4212 * reclaim is not able to make any progress because the current
4213 * implementation of the compaction depends on the sufficient amount
4214 * of free memory (see __compaction_suitable)
4215 */
4216 if (did_some_progress > 0 &&
4217 should_compact_retry(ac, order, alloc_flags,
4218 compact_result, &compact_priority,
4219 &compaction_retries))
4220 goto retry;
4221
4222
4223 /* Deal with possible cpuset update races before we start OOM killing */
4224 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4225 goto retry_cpuset;
4226
4227 /* Reclaim has failed us, start killing things */
4228 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4229 if (page)
4230 goto got_pg;
4231
4232 /* Avoid allocations with no watermarks from looping endlessly */
4233 if (tsk_is_oom_victim(current) &&
4234 (alloc_flags == ALLOC_OOM ||
4235 (gfp_mask & __GFP_NOMEMALLOC)))
4236 goto nopage;
4237
4238 /* Retry as long as the OOM killer is making progress */
4239 if (did_some_progress) {
4240 no_progress_loops = 0;
4241 goto retry;
4242 }
4243
4244nopage:
4245 /* Deal with possible cpuset update races before we fail */
4246 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4247 goto retry_cpuset;
4248
4249 /*
4250 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4251 * we always retry
4252 */
4253 if (gfp_mask & __GFP_NOFAIL) {
4254 /*
4255 * All existing users of the __GFP_NOFAIL are blockable, so warn
4256 * of any new users that actually require GFP_NOWAIT
4257 */
4258 if (WARN_ON_ONCE(!can_direct_reclaim))
4259 goto fail;
4260
4261 /*
4262 * PF_MEMALLOC request from this context is rather bizarre
4263 * because we cannot reclaim anything and only can loop waiting
4264 * for somebody to do a work for us
4265 */
4266 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4267
4268 /*
4269 * non failing costly orders are a hard requirement which we
4270 * are not prepared for much so let's warn about these users
4271 * so that we can identify them and convert them to something
4272 * else.
4273 */
4274 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4275
4276 /*
4277 * Help non-failing allocations by giving them access to memory
4278 * reserves but do not use ALLOC_NO_WATERMARKS because this
4279 * could deplete whole memory reserves which would just make
4280 * the situation worse
4281 */
4282 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4283 if (page)
4284 goto got_pg;
4285
4286 cond_resched();
4287 goto retry;
4288 }
4289fail:
4290 warn_alloc(gfp_mask, ac->nodemask,
4291 "page allocation failure: order:%u", order);
4292got_pg:
4293 return page;
4294}
4295
4296static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4297 int preferred_nid, nodemask_t *nodemask,
4298 struct alloc_context *ac, gfp_t *alloc_mask,
4299 unsigned int *alloc_flags)
4300{
4301 ac->high_zoneidx = gfp_zone(gfp_mask);
4302 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4303 ac->nodemask = nodemask;
4304 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4305
4306 if (cpusets_enabled()) {
4307 *alloc_mask |= __GFP_HARDWALL;
4308 if (!ac->nodemask)
4309 ac->nodemask = &cpuset_current_mems_allowed;
4310 else
4311 *alloc_flags |= ALLOC_CPUSET;
4312 }
4313
4314 fs_reclaim_acquire(gfp_mask);
4315 fs_reclaim_release(gfp_mask);
4316
4317 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4318
4319 if (should_fail_alloc_page(gfp_mask, order))
4320 return false;
4321
4322 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4323 *alloc_flags |= ALLOC_CMA;
4324
4325 return true;
4326}
4327
4328/* Determine whether to spread dirty pages and what the first usable zone */
4329static inline void finalise_ac(gfp_t gfp_mask,
4330 unsigned int order, struct alloc_context *ac)
4331{
4332 /* Dirty zone balancing only done in the fast path */
4333 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4334
4335 /*
4336 * The preferred zone is used for statistics but crucially it is
4337 * also used as the starting point for the zonelist iterator. It
4338 * may get reset for allocations that ignore memory policies.
4339 */
4340 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4341 ac->high_zoneidx, ac->nodemask);
4342}
4343
4344/*
4345 * This is the 'heart' of the zoned buddy allocator.
4346 */
4347struct page *
4348__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4349 nodemask_t *nodemask)
4350{
4351 struct page *page;
4352 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4353 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4354 struct alloc_context ac = { };
4355
4356 gfp_mask &= gfp_allowed_mask;
4357 alloc_mask = gfp_mask;
4358 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4359 return NULL;
4360
4361 finalise_ac(gfp_mask, order, &ac);
4362
4363 /* First allocation attempt */
4364 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4365 if (likely(page))
4366 goto out;
4367
4368 /*
4369 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4370 * resp. GFP_NOIO which has to be inherited for all allocation requests
4371 * from a particular context which has been marked by
4372 * memalloc_no{fs,io}_{save,restore}.
4373 */
4374 alloc_mask = current_gfp_context(gfp_mask);
4375 ac.spread_dirty_pages = false;
4376
4377 /*
4378 * Restore the original nodemask if it was potentially replaced with
4379 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4380 */
4381 if (unlikely(ac.nodemask != nodemask))
4382 ac.nodemask = nodemask;
4383
4384 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4385
4386out:
4387 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4388 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4389 __free_pages(page, order);
4390 page = NULL;
4391 }
4392
4393 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4394
4395 return page;
4396}
4397EXPORT_SYMBOL(__alloc_pages_nodemask);
4398
4399/*
4400 * Common helper functions.
4401 */
4402unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4403{
4404 struct page *page;
4405
4406 /*
4407 * __get_free_pages() returns a virtual address, which cannot represent
4408 * a highmem page
4409 */
4410 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4411
4412 page = alloc_pages(gfp_mask, order);
4413 if (!page)
4414 return 0;
4415 return (unsigned long) page_address(page);
4416}
4417EXPORT_SYMBOL(__get_free_pages);
4418
4419unsigned long get_zeroed_page(gfp_t gfp_mask)
4420{
4421 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4422}
4423EXPORT_SYMBOL(get_zeroed_page);
4424
4425void __free_pages(struct page *page, unsigned int order)
4426{
4427 if (put_page_testzero(page)) {
4428 if (order == 0)
4429 free_unref_page(page);
4430 else
4431 __free_pages_ok(page, order);
4432 }
4433}
4434
4435EXPORT_SYMBOL(__free_pages);
4436
4437void free_pages(unsigned long addr, unsigned int order)
4438{
4439 if (addr != 0) {
4440 VM_BUG_ON(!virt_addr_valid((void *)addr));
4441 __free_pages(virt_to_page((void *)addr), order);
4442 }
4443}
4444
4445EXPORT_SYMBOL(free_pages);
4446
4447/*
4448 * Page Fragment:
4449 * An arbitrary-length arbitrary-offset area of memory which resides
4450 * within a 0 or higher order page. Multiple fragments within that page
4451 * are individually refcounted, in the page's reference counter.
4452 *
4453 * The page_frag functions below provide a simple allocation framework for
4454 * page fragments. This is used by the network stack and network device
4455 * drivers to provide a backing region of memory for use as either an
4456 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4457 */
4458static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4459 gfp_t gfp_mask)
4460{
4461 struct page *page = NULL;
4462 gfp_t gfp = gfp_mask;
4463
4464#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4465 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4466 __GFP_NOMEMALLOC;
4467 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4468 PAGE_FRAG_CACHE_MAX_ORDER);
4469 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4470#endif
4471 if (unlikely(!page))
4472 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4473
4474 nc->va = page ? page_address(page) : NULL;
4475
4476 return page;
4477}
4478
4479void __page_frag_cache_drain(struct page *page, unsigned int count)
4480{
4481 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4482
4483 if (page_ref_sub_and_test(page, count)) {
4484 unsigned int order = compound_order(page);
4485
4486 if (order == 0)
4487 free_unref_page(page);
4488 else
4489 __free_pages_ok(page, order);
4490 }
4491}
4492EXPORT_SYMBOL(__page_frag_cache_drain);
4493
4494void *page_frag_alloc(struct page_frag_cache *nc,
4495 unsigned int fragsz, gfp_t gfp_mask)
4496{
4497 unsigned int size = PAGE_SIZE;
4498 struct page *page;
4499 int offset;
4500
4501 if (unlikely(!nc->va)) {
4502refill:
4503 page = __page_frag_cache_refill(nc, gfp_mask);
4504 if (!page)
4505 return NULL;
4506
4507#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4508 /* if size can vary use size else just use PAGE_SIZE */
4509 size = nc->size;
4510#endif
4511 /* Even if we own the page, we do not use atomic_set().
4512 * This would break get_page_unless_zero() users.
4513 */
4514 page_ref_add(page, size - 1);
4515
4516 /* reset page count bias and offset to start of new frag */
4517 nc->pfmemalloc = page_is_pfmemalloc(page);
4518 nc->pagecnt_bias = size;
4519 nc->offset = size;
4520 }
4521
4522 offset = nc->offset - fragsz;
4523 if (unlikely(offset < 0)) {
4524 page = virt_to_page(nc->va);
4525
4526 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4527 goto refill;
4528
4529#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4530 /* if size can vary use size else just use PAGE_SIZE */
4531 size = nc->size;
4532#endif
4533 /* OK, page count is 0, we can safely set it */
4534 set_page_count(page, size);
4535
4536 /* reset page count bias and offset to start of new frag */
4537 nc->pagecnt_bias = size;
4538 offset = size - fragsz;
4539 }
4540
4541 nc->pagecnt_bias--;
4542 nc->offset = offset;
4543
4544 return nc->va + offset;
4545}
4546EXPORT_SYMBOL(page_frag_alloc);
4547
4548/*
4549 * Frees a page fragment allocated out of either a compound or order 0 page.
4550 */
4551void page_frag_free(void *addr)
4552{
4553 struct page *page = virt_to_head_page(addr);
4554
4555 if (unlikely(put_page_testzero(page)))
4556 __free_pages_ok(page, compound_order(page));
4557}
4558EXPORT_SYMBOL(page_frag_free);
4559
4560static void *make_alloc_exact(unsigned long addr, unsigned int order,
4561 size_t size)
4562{
4563 if (addr) {
4564 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4565 unsigned long used = addr + PAGE_ALIGN(size);
4566
4567 split_page(virt_to_page((void *)addr), order);
4568 while (used < alloc_end) {
4569 free_page(used);
4570 used += PAGE_SIZE;
4571 }
4572 }
4573 return (void *)addr;
4574}
4575
4576/**
4577 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4578 * @size: the number of bytes to allocate
4579 * @gfp_mask: GFP flags for the allocation
4580 *
4581 * This function is similar to alloc_pages(), except that it allocates the
4582 * minimum number of pages to satisfy the request. alloc_pages() can only
4583 * allocate memory in power-of-two pages.
4584 *
4585 * This function is also limited by MAX_ORDER.
4586 *
4587 * Memory allocated by this function must be released by free_pages_exact().
4588 */
4589void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4590{
4591 unsigned int order = get_order(size);
4592 unsigned long addr;
4593
4594 addr = __get_free_pages(gfp_mask, order);
4595 return make_alloc_exact(addr, order, size);
4596}
4597EXPORT_SYMBOL(alloc_pages_exact);
4598
4599/**
4600 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4601 * pages on a node.
4602 * @nid: the preferred node ID where memory should be allocated
4603 * @size: the number of bytes to allocate
4604 * @gfp_mask: GFP flags for the allocation
4605 *
4606 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4607 * back.
4608 */
4609void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4610{
4611 unsigned int order = get_order(size);
4612 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4613 if (!p)
4614 return NULL;
4615 return make_alloc_exact((unsigned long)page_address(p), order, size);
4616}
4617
4618/**
4619 * free_pages_exact - release memory allocated via alloc_pages_exact()
4620 * @virt: the value returned by alloc_pages_exact.
4621 * @size: size of allocation, same value as passed to alloc_pages_exact().
4622 *
4623 * Release the memory allocated by a previous call to alloc_pages_exact.
4624 */
4625void free_pages_exact(void *virt, size_t size)
4626{
4627 unsigned long addr = (unsigned long)virt;
4628 unsigned long end = addr + PAGE_ALIGN(size);
4629
4630 while (addr < end) {
4631 free_page(addr);
4632 addr += PAGE_SIZE;
4633 }
4634}
4635EXPORT_SYMBOL(free_pages_exact);
4636
4637/**
4638 * nr_free_zone_pages - count number of pages beyond high watermark
4639 * @offset: The zone index of the highest zone
4640 *
4641 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4642 * high watermark within all zones at or below a given zone index. For each
4643 * zone, the number of pages is calculated as:
4644 *
4645 * nr_free_zone_pages = managed_pages - high_pages
4646 */
4647static unsigned long nr_free_zone_pages(int offset)
4648{
4649 struct zoneref *z;
4650 struct zone *zone;
4651
4652 /* Just pick one node, since fallback list is circular */
4653 unsigned long sum = 0;
4654
4655 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4656
4657 for_each_zone_zonelist(zone, z, zonelist, offset) {
4658 unsigned long size = zone->managed_pages;
4659 unsigned long high = high_wmark_pages(zone);
4660 if (size > high)
4661 sum += size - high;
4662 }
4663
4664 return sum;
4665}
4666
4667/**
4668 * nr_free_buffer_pages - count number of pages beyond high watermark
4669 *
4670 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4671 * watermark within ZONE_DMA and ZONE_NORMAL.
4672 */
4673unsigned long nr_free_buffer_pages(void)
4674{
4675 return nr_free_zone_pages(gfp_zone(GFP_USER));
4676}
4677EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4678
4679/**
4680 * nr_free_pagecache_pages - count number of pages beyond high watermark
4681 *
4682 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4683 * high watermark within all zones.
4684 */
4685unsigned long nr_free_pagecache_pages(void)
4686{
4687 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4688}
4689
4690static inline void show_node(struct zone *zone)
4691{
4692 if (IS_ENABLED(CONFIG_NUMA))
4693 printk("Node %d ", zone_to_nid(zone));
4694}
4695
4696long si_mem_available(void)
4697{
4698 long available;
4699 unsigned long pagecache;
4700 unsigned long wmark_low = 0;
4701 unsigned long pages[NR_LRU_LISTS];
4702 struct zone *zone;
4703 int lru;
4704
4705 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4706 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4707
4708 for_each_zone(zone)
4709 wmark_low += zone->watermark[WMARK_LOW];
4710
4711 /*
4712 * Estimate the amount of memory available for userspace allocations,
4713 * without causing swapping.
4714 */
4715 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4716
4717 /*
4718 * Not all the page cache can be freed, otherwise the system will
4719 * start swapping. Assume at least half of the page cache, or the
4720 * low watermark worth of cache, needs to stay.
4721 */
4722 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4723 pagecache -= min(pagecache / 2, wmark_low);
4724 available += pagecache;
4725
4726 /*
4727 * Part of the reclaimable slab consists of items that are in use,
4728 * and cannot be freed. Cap this estimate at the low watermark.
4729 */
4730 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4731 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4732 wmark_low);
4733
4734 /*
4735 * Part of the kernel memory, which can be released under memory
4736 * pressure.
4737 */
4738 available += global_node_page_state(NR_INDIRECTLY_RECLAIMABLE_BYTES) >>
4739 PAGE_SHIFT;
4740
4741 if (available < 0)
4742 available = 0;
4743 return available;
4744}
4745EXPORT_SYMBOL_GPL(si_mem_available);
4746
4747void si_meminfo(struct sysinfo *val)
4748{
4749 val->totalram = totalram_pages;
4750 val->sharedram = global_node_page_state(NR_SHMEM);
4751 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4752 val->bufferram = nr_blockdev_pages();
4753 val->totalhigh = totalhigh_pages;
4754 val->freehigh = nr_free_highpages();
4755 val->mem_unit = PAGE_SIZE;
4756}
4757
4758EXPORT_SYMBOL(si_meminfo);
4759
4760#ifdef CONFIG_NUMA
4761void si_meminfo_node(struct sysinfo *val, int nid)
4762{
4763 int zone_type; /* needs to be signed */
4764 unsigned long managed_pages = 0;
4765 unsigned long managed_highpages = 0;
4766 unsigned long free_highpages = 0;
4767 pg_data_t *pgdat = NODE_DATA(nid);
4768
4769 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4770 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4771 val->totalram = managed_pages;
4772 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4773 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4774#ifdef CONFIG_HIGHMEM
4775 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4776 struct zone *zone = &pgdat->node_zones[zone_type];
4777
4778 if (is_highmem(zone)) {
4779 managed_highpages += zone->managed_pages;
4780 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4781 }
4782 }
4783 val->totalhigh = managed_highpages;
4784 val->freehigh = free_highpages;
4785#else
4786 val->totalhigh = managed_highpages;
4787 val->freehigh = free_highpages;
4788#endif
4789 val->mem_unit = PAGE_SIZE;
4790}
4791#endif
4792
4793/*
4794 * Determine whether the node should be displayed or not, depending on whether
4795 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4796 */
4797static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4798{
4799 if (!(flags & SHOW_MEM_FILTER_NODES))
4800 return false;
4801
4802 /*
4803 * no node mask - aka implicit memory numa policy. Do not bother with
4804 * the synchronization - read_mems_allowed_begin - because we do not
4805 * have to be precise here.
4806 */
4807 if (!nodemask)
4808 nodemask = &cpuset_current_mems_allowed;
4809
4810 return !node_isset(nid, *nodemask);
4811}
4812
4813#define K(x) ((x) << (PAGE_SHIFT-10))
4814
4815static void show_migration_types(unsigned char type)
4816{
4817 static const char types[MIGRATE_TYPES] = {
4818 [MIGRATE_UNMOVABLE] = 'U',
4819 [MIGRATE_MOVABLE] = 'M',
4820 [MIGRATE_RECLAIMABLE] = 'E',
4821 [MIGRATE_HIGHATOMIC] = 'H',
4822#ifdef CONFIG_CMA
4823 [MIGRATE_CMA] = 'C',
4824#endif
4825#ifdef CONFIG_MEMORY_ISOLATION
4826 [MIGRATE_ISOLATE] = 'I',
4827#endif
4828 };
4829 char tmp[MIGRATE_TYPES + 1];
4830 char *p = tmp;
4831 int i;
4832
4833 for (i = 0; i < MIGRATE_TYPES; i++) {
4834 if (type & (1 << i))
4835 *p++ = types[i];
4836 }
4837
4838 *p = '\0';
4839 printk(KERN_CONT "(%s) ", tmp);
4840}
4841
4842/*
4843 * Show free area list (used inside shift_scroll-lock stuff)
4844 * We also calculate the percentage fragmentation. We do this by counting the
4845 * memory on each free list with the exception of the first item on the list.
4846 *
4847 * Bits in @filter:
4848 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4849 * cpuset.
4850 */
4851void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4852{
4853 unsigned long free_pcp = 0;
4854 int cpu;
4855 struct zone *zone;
4856 pg_data_t *pgdat;
4857
4858 for_each_populated_zone(zone) {
4859 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4860 continue;
4861
4862 for_each_online_cpu(cpu)
4863 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4864 }
4865
4866 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4867 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4868 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4869 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4870 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4871 " free:%lu free_pcp:%lu free_cma:%lu\n",
4872 global_node_page_state(NR_ACTIVE_ANON),
4873 global_node_page_state(NR_INACTIVE_ANON),
4874 global_node_page_state(NR_ISOLATED_ANON),
4875 global_node_page_state(NR_ACTIVE_FILE),
4876 global_node_page_state(NR_INACTIVE_FILE),
4877 global_node_page_state(NR_ISOLATED_FILE),
4878 global_node_page_state(NR_UNEVICTABLE),
4879 global_node_page_state(NR_FILE_DIRTY),
4880 global_node_page_state(NR_WRITEBACK),
4881 global_node_page_state(NR_UNSTABLE_NFS),
4882 global_node_page_state(NR_SLAB_RECLAIMABLE),
4883 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4884 global_node_page_state(NR_FILE_MAPPED),
4885 global_node_page_state(NR_SHMEM),
4886 global_zone_page_state(NR_PAGETABLE),
4887 global_zone_page_state(NR_BOUNCE),
4888 global_zone_page_state(NR_FREE_PAGES),
4889 free_pcp,
4890 global_zone_page_state(NR_FREE_CMA_PAGES));
4891
4892 for_each_online_pgdat(pgdat) {
4893 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4894 continue;
4895
4896 printk("Node %d"
4897 " active_anon:%lukB"
4898 " inactive_anon:%lukB"
4899 " active_file:%lukB"
4900 " inactive_file:%lukB"
4901 " unevictable:%lukB"
4902 " isolated(anon):%lukB"
4903 " isolated(file):%lukB"
4904 " mapped:%lukB"
4905 " dirty:%lukB"
4906 " writeback:%lukB"
4907 " shmem:%lukB"
4908#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4909 " shmem_thp: %lukB"
4910 " shmem_pmdmapped: %lukB"
4911 " anon_thp: %lukB"
4912#endif
4913 " writeback_tmp:%lukB"
4914 " unstable:%lukB"
4915 " all_unreclaimable? %s"
4916 "\n",
4917 pgdat->node_id,
4918 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4919 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4920 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4921 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4922 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4923 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4924 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4925 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4926 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4927 K(node_page_state(pgdat, NR_WRITEBACK)),
4928 K(node_page_state(pgdat, NR_SHMEM)),
4929#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4930 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4931 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4932 * HPAGE_PMD_NR),
4933 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4934#endif
4935 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4936 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4937 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4938 "yes" : "no");
4939 }
4940
4941 for_each_populated_zone(zone) {
4942 int i;
4943
4944 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4945 continue;
4946
4947 free_pcp = 0;
4948 for_each_online_cpu(cpu)
4949 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4950
4951 show_node(zone);
4952 printk(KERN_CONT
4953 "%s"
4954 " free:%lukB"
4955 " min:%lukB"
4956 " low:%lukB"
4957 " high:%lukB"
4958 " active_anon:%lukB"
4959 " inactive_anon:%lukB"
4960 " active_file:%lukB"
4961 " inactive_file:%lukB"
4962 " unevictable:%lukB"
4963 " writepending:%lukB"
4964 " present:%lukB"
4965 " managed:%lukB"
4966 " mlocked:%lukB"
4967 " kernel_stack:%lukB"
4968 " pagetables:%lukB"
4969 " bounce:%lukB"
4970 " free_pcp:%lukB"
4971 " local_pcp:%ukB"
4972 " free_cma:%lukB"
4973 "\n",
4974 zone->name,
4975 K(zone_page_state(zone, NR_FREE_PAGES)),
4976 K(min_wmark_pages(zone)),
4977 K(low_wmark_pages(zone)),
4978 K(high_wmark_pages(zone)),
4979 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4980 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4981 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4982 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4983 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4984 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4985 K(zone->present_pages),
4986 K(zone->managed_pages),
4987 K(zone_page_state(zone, NR_MLOCK)),
4988 zone_page_state(zone, NR_KERNEL_STACK_KB),
4989 K(zone_page_state(zone, NR_PAGETABLE)),
4990 K(zone_page_state(zone, NR_BOUNCE)),
4991 K(free_pcp),
4992 K(this_cpu_read(zone->pageset->pcp.count)),
4993 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4994 printk("lowmem_reserve[]:");
4995 for (i = 0; i < MAX_NR_ZONES; i++)
4996 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4997 printk(KERN_CONT "\n");
4998 }
4999
5000 for_each_populated_zone(zone) {
5001 unsigned int order;
5002 unsigned long nr[MAX_ORDER], flags, total = 0;
5003 unsigned char types[MAX_ORDER];
5004
5005 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5006 continue;
5007 show_node(zone);
5008 printk(KERN_CONT "%s: ", zone->name);
5009
5010 spin_lock_irqsave(&zone->lock, flags);
5011 for (order = 0; order < MAX_ORDER; order++) {
5012 struct free_area *area = &zone->free_area[order];
5013 int type;
5014
5015 nr[order] = area->nr_free;
5016 total += nr[order] << order;
5017
5018 types[order] = 0;
5019 for (type = 0; type < MIGRATE_TYPES; type++) {
5020 if (!list_empty(&area->free_list[type]))
5021 types[order] |= 1 << type;
5022 }
5023 }
5024 spin_unlock_irqrestore(&zone->lock, flags);
5025 for (order = 0; order < MAX_ORDER; order++) {
5026 printk(KERN_CONT "%lu*%lukB ",
5027 nr[order], K(1UL) << order);
5028 if (nr[order])
5029 show_migration_types(types[order]);
5030 }
5031 printk(KERN_CONT "= %lukB\n", K(total));
5032 }
5033
5034 hugetlb_show_meminfo();
5035
5036 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5037
5038 show_swap_cache_info();
5039}
5040
5041static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5042{
5043 zoneref->zone = zone;
5044 zoneref->zone_idx = zone_idx(zone);
5045}
5046
5047/*
5048 * Builds allocation fallback zone lists.
5049 *
5050 * Add all populated zones of a node to the zonelist.
5051 */
5052static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5053{
5054 struct zone *zone;
5055 enum zone_type zone_type = MAX_NR_ZONES;
5056 int nr_zones = 0;
5057
5058 do {
5059 zone_type--;
5060 zone = pgdat->node_zones + zone_type;
5061 if (managed_zone(zone)) {
5062 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5063 check_highest_zone(zone_type);
5064 }
5065 } while (zone_type);
5066
5067 return nr_zones;
5068}
5069
5070#ifdef CONFIG_NUMA
5071
5072static int __parse_numa_zonelist_order(char *s)
5073{
5074 /*
5075 * We used to support different zonlists modes but they turned
5076 * out to be just not useful. Let's keep the warning in place
5077 * if somebody still use the cmd line parameter so that we do
5078 * not fail it silently
5079 */
5080 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5081 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5082 return -EINVAL;
5083 }
5084 return 0;
5085}
5086
5087static __init int setup_numa_zonelist_order(char *s)
5088{
5089 if (!s)
5090 return 0;
5091
5092 return __parse_numa_zonelist_order(s);
5093}
5094early_param("numa_zonelist_order", setup_numa_zonelist_order);
5095
5096char numa_zonelist_order[] = "Node";
5097
5098/*
5099 * sysctl handler for numa_zonelist_order
5100 */
5101int numa_zonelist_order_handler(struct ctl_table *table, int write,
5102 void __user *buffer, size_t *length,
5103 loff_t *ppos)
5104{
5105 char *str;
5106 int ret;
5107
5108 if (!write)
5109 return proc_dostring(table, write, buffer, length, ppos);
5110 str = memdup_user_nul(buffer, 16);
5111 if (IS_ERR(str))
5112 return PTR_ERR(str);
5113
5114 ret = __parse_numa_zonelist_order(str);
5115 kfree(str);
5116 return ret;
5117}
5118
5119
5120#define MAX_NODE_LOAD (nr_online_nodes)
5121static int node_load[MAX_NUMNODES];
5122
5123/**
5124 * find_next_best_node - find the next node that should appear in a given node's fallback list
5125 * @node: node whose fallback list we're appending
5126 * @used_node_mask: nodemask_t of already used nodes
5127 *
5128 * We use a number of factors to determine which is the next node that should
5129 * appear on a given node's fallback list. The node should not have appeared
5130 * already in @node's fallback list, and it should be the next closest node
5131 * according to the distance array (which contains arbitrary distance values
5132 * from each node to each node in the system), and should also prefer nodes
5133 * with no CPUs, since presumably they'll have very little allocation pressure
5134 * on them otherwise.
5135 * It returns -1 if no node is found.
5136 */
5137static int find_next_best_node(int node, nodemask_t *used_node_mask)
5138{
5139 int n, val;
5140 int min_val = INT_MAX;
5141 int best_node = NUMA_NO_NODE;
5142 const struct cpumask *tmp = cpumask_of_node(0);
5143
5144 /* Use the local node if we haven't already */
5145 if (!node_isset(node, *used_node_mask)) {
5146 node_set(node, *used_node_mask);
5147 return node;
5148 }
5149
5150 for_each_node_state(n, N_MEMORY) {
5151
5152 /* Don't want a node to appear more than once */
5153 if (node_isset(n, *used_node_mask))
5154 continue;
5155
5156 /* Use the distance array to find the distance */
5157 val = node_distance(node, n);
5158
5159 /* Penalize nodes under us ("prefer the next node") */
5160 val += (n < node);
5161
5162 /* Give preference to headless and unused nodes */
5163 tmp = cpumask_of_node(n);
5164 if (!cpumask_empty(tmp))
5165 val += PENALTY_FOR_NODE_WITH_CPUS;
5166
5167 /* Slight preference for less loaded node */
5168 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5169 val += node_load[n];
5170
5171 if (val < min_val) {
5172 min_val = val;
5173 best_node = n;
5174 }
5175 }
5176
5177 if (best_node >= 0)
5178 node_set(best_node, *used_node_mask);
5179
5180 return best_node;
5181}
5182
5183
5184/*
5185 * Build zonelists ordered by node and zones within node.
5186 * This results in maximum locality--normal zone overflows into local
5187 * DMA zone, if any--but risks exhausting DMA zone.
5188 */
5189static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5190 unsigned nr_nodes)
5191{
5192 struct zoneref *zonerefs;
5193 int i;
5194
5195 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5196
5197 for (i = 0; i < nr_nodes; i++) {
5198 int nr_zones;
5199
5200 pg_data_t *node = NODE_DATA(node_order[i]);
5201
5202 nr_zones = build_zonerefs_node(node, zonerefs);
5203 zonerefs += nr_zones;
5204 }
5205 zonerefs->zone = NULL;
5206 zonerefs->zone_idx = 0;
5207}
5208
5209/*
5210 * Build gfp_thisnode zonelists
5211 */
5212static void build_thisnode_zonelists(pg_data_t *pgdat)
5213{
5214 struct zoneref *zonerefs;
5215 int nr_zones;
5216
5217 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5218 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5219 zonerefs += nr_zones;
5220 zonerefs->zone = NULL;
5221 zonerefs->zone_idx = 0;
5222}
5223
5224/*
5225 * Build zonelists ordered by zone and nodes within zones.
5226 * This results in conserving DMA zone[s] until all Normal memory is
5227 * exhausted, but results in overflowing to remote node while memory
5228 * may still exist in local DMA zone.
5229 */
5230
5231static void build_zonelists(pg_data_t *pgdat)
5232{
5233 static int node_order[MAX_NUMNODES];
5234 int node, load, nr_nodes = 0;
5235 nodemask_t used_mask;
5236 int local_node, prev_node;
5237
5238 /* NUMA-aware ordering of nodes */
5239 local_node = pgdat->node_id;
5240 load = nr_online_nodes;
5241 prev_node = local_node;
5242 nodes_clear(used_mask);
5243
5244 memset(node_order, 0, sizeof(node_order));
5245 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5246 /*
5247 * We don't want to pressure a particular node.
5248 * So adding penalty to the first node in same
5249 * distance group to make it round-robin.
5250 */
5251 if (node_distance(local_node, node) !=
5252 node_distance(local_node, prev_node))
5253 node_load[node] = load;
5254
5255 node_order[nr_nodes++] = node;
5256 prev_node = node;
5257 load--;
5258 }
5259
5260 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5261 build_thisnode_zonelists(pgdat);
5262}
5263
5264#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5265/*
5266 * Return node id of node used for "local" allocations.
5267 * I.e., first node id of first zone in arg node's generic zonelist.
5268 * Used for initializing percpu 'numa_mem', which is used primarily
5269 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5270 */
5271int local_memory_node(int node)
5272{
5273 struct zoneref *z;
5274
5275 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5276 gfp_zone(GFP_KERNEL),
5277 NULL);
5278 return z->zone->node;
5279}
5280#endif
5281
5282static void setup_min_unmapped_ratio(void);
5283static void setup_min_slab_ratio(void);
5284#else /* CONFIG_NUMA */
5285
5286static void build_zonelists(pg_data_t *pgdat)
5287{
5288 int node, local_node;
5289 struct zoneref *zonerefs;
5290 int nr_zones;
5291
5292 local_node = pgdat->node_id;
5293
5294 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5295 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5296 zonerefs += nr_zones;
5297
5298 /*
5299 * Now we build the zonelist so that it contains the zones
5300 * of all the other nodes.
5301 * We don't want to pressure a particular node, so when
5302 * building the zones for node N, we make sure that the
5303 * zones coming right after the local ones are those from
5304 * node N+1 (modulo N)
5305 */
5306 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5307 if (!node_online(node))
5308 continue;
5309 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5310 zonerefs += nr_zones;
5311 }
5312 for (node = 0; node < local_node; node++) {
5313 if (!node_online(node))
5314 continue;
5315 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5316 zonerefs += nr_zones;
5317 }
5318
5319 zonerefs->zone = NULL;
5320 zonerefs->zone_idx = 0;
5321}
5322
5323#endif /* CONFIG_NUMA */
5324
5325/*
5326 * Boot pageset table. One per cpu which is going to be used for all
5327 * zones and all nodes. The parameters will be set in such a way
5328 * that an item put on a list will immediately be handed over to
5329 * the buddy list. This is safe since pageset manipulation is done
5330 * with interrupts disabled.
5331 *
5332 * The boot_pagesets must be kept even after bootup is complete for
5333 * unused processors and/or zones. They do play a role for bootstrapping
5334 * hotplugged processors.
5335 *
5336 * zoneinfo_show() and maybe other functions do
5337 * not check if the processor is online before following the pageset pointer.
5338 * Other parts of the kernel may not check if the zone is available.
5339 */
5340static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5341static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5342static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5343
5344static void __build_all_zonelists(void *data)
5345{
5346 int nid;
5347 int __maybe_unused cpu;
5348 pg_data_t *self = data;
5349 static DEFINE_SPINLOCK(lock);
5350
5351 spin_lock(&lock);
5352
5353#ifdef CONFIG_NUMA
5354 memset(node_load, 0, sizeof(node_load));
5355#endif
5356
5357 /*
5358 * This node is hotadded and no memory is yet present. So just
5359 * building zonelists is fine - no need to touch other nodes.
5360 */
5361 if (self && !node_online(self->node_id)) {
5362 build_zonelists(self);
5363 } else {
5364 for_each_online_node(nid) {
5365 pg_data_t *pgdat = NODE_DATA(nid);
5366
5367 build_zonelists(pgdat);
5368 }
5369
5370#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5371 /*
5372 * We now know the "local memory node" for each node--
5373 * i.e., the node of the first zone in the generic zonelist.
5374 * Set up numa_mem percpu variable for on-line cpus. During
5375 * boot, only the boot cpu should be on-line; we'll init the
5376 * secondary cpus' numa_mem as they come on-line. During
5377 * node/memory hotplug, we'll fixup all on-line cpus.
5378 */
5379 for_each_online_cpu(cpu)
5380 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5381#endif
5382 }
5383
5384 spin_unlock(&lock);
5385}
5386
5387static noinline void __init
5388build_all_zonelists_init(void)
5389{
5390 int cpu;
5391
5392 __build_all_zonelists(NULL);
5393
5394 /*
5395 * Initialize the boot_pagesets that are going to be used
5396 * for bootstrapping processors. The real pagesets for
5397 * each zone will be allocated later when the per cpu
5398 * allocator is available.
5399 *
5400 * boot_pagesets are used also for bootstrapping offline
5401 * cpus if the system is already booted because the pagesets
5402 * are needed to initialize allocators on a specific cpu too.
5403 * F.e. the percpu allocator needs the page allocator which
5404 * needs the percpu allocator in order to allocate its pagesets
5405 * (a chicken-egg dilemma).
5406 */
5407 for_each_possible_cpu(cpu)
5408 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5409
5410 mminit_verify_zonelist();
5411 cpuset_init_current_mems_allowed();
5412}
5413
5414/*
5415 * unless system_state == SYSTEM_BOOTING.
5416 *
5417 * __ref due to call of __init annotated helper build_all_zonelists_init
5418 * [protected by SYSTEM_BOOTING].
5419 */
5420void __ref build_all_zonelists(pg_data_t *pgdat)
5421{
5422 if (system_state == SYSTEM_BOOTING) {
5423 build_all_zonelists_init();
5424 } else {
5425 __build_all_zonelists(pgdat);
5426 /* cpuset refresh routine should be here */
5427 }
5428 vm_total_pages = nr_free_pagecache_pages();
5429 /*
5430 * Disable grouping by mobility if the number of pages in the
5431 * system is too low to allow the mechanism to work. It would be
5432 * more accurate, but expensive to check per-zone. This check is
5433 * made on memory-hotadd so a system can start with mobility
5434 * disabled and enable it later
5435 */
5436 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5437 page_group_by_mobility_disabled = 1;
5438 else
5439 page_group_by_mobility_disabled = 0;
5440
5441 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5442 nr_online_nodes,
5443 page_group_by_mobility_disabled ? "off" : "on",
5444 vm_total_pages);
5445#ifdef CONFIG_NUMA
5446 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5447#endif
5448}
5449
5450/*
5451 * Initially all pages are reserved - free ones are freed
5452 * up by free_all_bootmem() once the early boot process is
5453 * done. Non-atomic initialization, single-pass.
5454 */
5455void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5456 unsigned long start_pfn, enum memmap_context context,
5457 struct vmem_altmap *altmap)
5458{
5459 unsigned long end_pfn = start_pfn + size;
5460 pg_data_t *pgdat = NODE_DATA(nid);
5461 unsigned long pfn;
5462 unsigned long nr_initialised = 0;
5463 struct page *page;
5464#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5465 struct memblock_region *r = NULL, *tmp;
5466#endif
5467
5468 if (highest_memmap_pfn < end_pfn - 1)
5469 highest_memmap_pfn = end_pfn - 1;
5470
5471 /*
5472 * Honor reservation requested by the driver for this ZONE_DEVICE
5473 * memory
5474 */
5475 if (altmap && start_pfn == altmap->base_pfn)
5476 start_pfn += altmap->reserve;
5477
5478 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5479 /*
5480 * There can be holes in boot-time mem_map[]s handed to this
5481 * function. They do not exist on hotplugged memory.
5482 */
5483 if (context != MEMMAP_EARLY)
5484 goto not_early;
5485
5486 if (!early_pfn_valid(pfn))
5487 continue;
5488 if (!early_pfn_in_nid(pfn, nid))
5489 continue;
5490 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5491 break;
5492
5493#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5494 /*
5495 * Check given memblock attribute by firmware which can affect
5496 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5497 * mirrored, it's an overlapped memmap init. skip it.
5498 */
5499 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5500 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5501 for_each_memblock(memory, tmp)
5502 if (pfn < memblock_region_memory_end_pfn(tmp))
5503 break;
5504 r = tmp;
5505 }
5506 if (pfn >= memblock_region_memory_base_pfn(r) &&
5507 memblock_is_mirror(r)) {
5508 /* already initialized as NORMAL */
5509 pfn = memblock_region_memory_end_pfn(r);
5510 continue;
5511 }
5512 }
5513#endif
5514
5515not_early:
5516 page = pfn_to_page(pfn);
5517 __init_single_page(page, pfn, zone, nid);
5518 if (context == MEMMAP_HOTPLUG)
5519 SetPageReserved(page);
5520
5521 /*
5522 * Mark the block movable so that blocks are reserved for
5523 * movable at startup. This will force kernel allocations
5524 * to reserve their blocks rather than leaking throughout
5525 * the address space during boot when many long-lived
5526 * kernel allocations are made.
5527 *
5528 * bitmap is created for zone's valid pfn range. but memmap
5529 * can be created for invalid pages (for alignment)
5530 * check here not to call set_pageblock_migratetype() against
5531 * pfn out of zone.
5532 *
5533 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5534 * because this is done early in sparse_add_one_section
5535 */
5536 if (!(pfn & (pageblock_nr_pages - 1))) {
5537 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5538 cond_resched();
5539 }
5540 }
5541}
5542
5543static void __meminit zone_init_free_lists(struct zone *zone)
5544{
5545 unsigned int order, t;
5546 for_each_migratetype_order(order, t) {
5547 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5548 zone->free_area[order].nr_free = 0;
5549 }
5550}
5551
5552#ifndef __HAVE_ARCH_MEMMAP_INIT
5553#define memmap_init(size, nid, zone, start_pfn) \
5554 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY, NULL)
5555#endif
5556
5557static int zone_batchsize(struct zone *zone)
5558{
5559#ifdef CONFIG_MMU
5560 int batch;
5561
5562 /*
5563 * The per-cpu-pages pools are set to around 1000th of the
5564 * size of the zone. But no more than 1/2 of a meg.
5565 *
5566 * OK, so we don't know how big the cache is. So guess.
5567 */
5568 batch = zone->managed_pages / 1024;
5569 if (batch * PAGE_SIZE > 512 * 1024)
5570 batch = (512 * 1024) / PAGE_SIZE;
5571 batch /= 4; /* We effectively *= 4 below */
5572 if (batch < 1)
5573 batch = 1;
5574
5575 /*
5576 * Clamp the batch to a 2^n - 1 value. Having a power
5577 * of 2 value was found to be more likely to have
5578 * suboptimal cache aliasing properties in some cases.
5579 *
5580 * For example if 2 tasks are alternately allocating
5581 * batches of pages, one task can end up with a lot
5582 * of pages of one half of the possible page colors
5583 * and the other with pages of the other colors.
5584 */
5585 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5586
5587 return batch;
5588
5589#else
5590 /* The deferral and batching of frees should be suppressed under NOMMU
5591 * conditions.
5592 *
5593 * The problem is that NOMMU needs to be able to allocate large chunks
5594 * of contiguous memory as there's no hardware page translation to
5595 * assemble apparent contiguous memory from discontiguous pages.
5596 *
5597 * Queueing large contiguous runs of pages for batching, however,
5598 * causes the pages to actually be freed in smaller chunks. As there
5599 * can be a significant delay between the individual batches being
5600 * recycled, this leads to the once large chunks of space being
5601 * fragmented and becoming unavailable for high-order allocations.
5602 */
5603 return 0;
5604#endif
5605}
5606
5607/*
5608 * pcp->high and pcp->batch values are related and dependent on one another:
5609 * ->batch must never be higher then ->high.
5610 * The following function updates them in a safe manner without read side
5611 * locking.
5612 *
5613 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5614 * those fields changing asynchronously (acording the the above rule).
5615 *
5616 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5617 * outside of boot time (or some other assurance that no concurrent updaters
5618 * exist).
5619 */
5620static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5621 unsigned long batch)
5622{
5623 /* start with a fail safe value for batch */
5624 pcp->batch = 1;
5625 smp_wmb();
5626
5627 /* Update high, then batch, in order */
5628 pcp->high = high;
5629 smp_wmb();
5630
5631 pcp->batch = batch;
5632}
5633
5634/* a companion to pageset_set_high() */
5635static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5636{
5637 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5638}
5639
5640static void pageset_init(struct per_cpu_pageset *p)
5641{
5642 struct per_cpu_pages *pcp;
5643 int migratetype;
5644
5645 memset(p, 0, sizeof(*p));
5646
5647 pcp = &p->pcp;
5648 pcp->count = 0;
5649 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5650 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5651}
5652
5653static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5654{
5655 pageset_init(p);
5656 pageset_set_batch(p, batch);
5657}
5658
5659/*
5660 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5661 * to the value high for the pageset p.
5662 */
5663static void pageset_set_high(struct per_cpu_pageset *p,
5664 unsigned long high)
5665{
5666 unsigned long batch = max(1UL, high / 4);
5667 if ((high / 4) > (PAGE_SHIFT * 8))
5668 batch = PAGE_SHIFT * 8;
5669
5670 pageset_update(&p->pcp, high, batch);
5671}
5672
5673static void pageset_set_high_and_batch(struct zone *zone,
5674 struct per_cpu_pageset *pcp)
5675{
5676 if (percpu_pagelist_fraction)
5677 pageset_set_high(pcp,
5678 (zone->managed_pages /
5679 percpu_pagelist_fraction));
5680 else
5681 pageset_set_batch(pcp, zone_batchsize(zone));
5682}
5683
5684static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5685{
5686 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5687
5688 pageset_init(pcp);
5689 pageset_set_high_and_batch(zone, pcp);
5690}
5691
5692void __meminit setup_zone_pageset(struct zone *zone)
5693{
5694 int cpu;
5695 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5696 for_each_possible_cpu(cpu)
5697 zone_pageset_init(zone, cpu);
5698}
5699
5700/*
5701 * Allocate per cpu pagesets and initialize them.
5702 * Before this call only boot pagesets were available.
5703 */
5704void __init setup_per_cpu_pageset(void)
5705{
5706 struct pglist_data *pgdat;
5707 struct zone *zone;
5708
5709 for_each_populated_zone(zone)
5710 setup_zone_pageset(zone);
5711
5712 for_each_online_pgdat(pgdat)
5713 pgdat->per_cpu_nodestats =
5714 alloc_percpu(struct per_cpu_nodestat);
5715}
5716
5717static __meminit void zone_pcp_init(struct zone *zone)
5718{
5719 /*
5720 * per cpu subsystem is not up at this point. The following code
5721 * relies on the ability of the linker to provide the
5722 * offset of a (static) per cpu variable into the per cpu area.
5723 */
5724 zone->pageset = &boot_pageset;
5725
5726 if (populated_zone(zone))
5727 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5728 zone->name, zone->present_pages,
5729 zone_batchsize(zone));
5730}
5731
5732void __meminit init_currently_empty_zone(struct zone *zone,
5733 unsigned long zone_start_pfn,
5734 unsigned long size)
5735{
5736 struct pglist_data *pgdat = zone->zone_pgdat;
5737
5738 pgdat->nr_zones = zone_idx(zone) + 1;
5739
5740 zone->zone_start_pfn = zone_start_pfn;
5741
5742 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5743 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5744 pgdat->node_id,
5745 (unsigned long)zone_idx(zone),
5746 zone_start_pfn, (zone_start_pfn + size));
5747
5748 zone_init_free_lists(zone);
5749 zone->initialized = 1;
5750}
5751
5752#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5753#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5754
5755/*
5756 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5757 */
5758int __meminit __early_pfn_to_nid(unsigned long pfn,
5759 struct mminit_pfnnid_cache *state)
5760{
5761 unsigned long start_pfn, end_pfn;
5762 int nid;
5763
5764 if (state->last_start <= pfn && pfn < state->last_end)
5765 return state->last_nid;
5766
5767 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5768 if (nid != -1) {
5769 state->last_start = start_pfn;
5770 state->last_end = end_pfn;
5771 state->last_nid = nid;
5772 }
5773
5774 return nid;
5775}
5776#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5777
5778/**
5779 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5780 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5781 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5782 *
5783 * If an architecture guarantees that all ranges registered contain no holes
5784 * and may be freed, this this function may be used instead of calling
5785 * memblock_free_early_nid() manually.
5786 */
5787void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5788{
5789 unsigned long start_pfn, end_pfn;
5790 int i, this_nid;
5791
5792 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5793 start_pfn = min(start_pfn, max_low_pfn);
5794 end_pfn = min(end_pfn, max_low_pfn);
5795
5796 if (start_pfn < end_pfn)
5797 memblock_free_early_nid(PFN_PHYS(start_pfn),
5798 (end_pfn - start_pfn) << PAGE_SHIFT,
5799 this_nid);
5800 }
5801}
5802
5803/**
5804 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5805 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5806 *
5807 * If an architecture guarantees that all ranges registered contain no holes and may
5808 * be freed, this function may be used instead of calling memory_present() manually.
5809 */
5810void __init sparse_memory_present_with_active_regions(int nid)
5811{
5812 unsigned long start_pfn, end_pfn;
5813 int i, this_nid;
5814
5815 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5816 memory_present(this_nid, start_pfn, end_pfn);
5817}
5818
5819/**
5820 * get_pfn_range_for_nid - Return the start and end page frames for a node
5821 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5822 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5823 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5824 *
5825 * It returns the start and end page frame of a node based on information
5826 * provided by memblock_set_node(). If called for a node
5827 * with no available memory, a warning is printed and the start and end
5828 * PFNs will be 0.
5829 */
5830void __meminit get_pfn_range_for_nid(unsigned int nid,
5831 unsigned long *start_pfn, unsigned long *end_pfn)
5832{
5833 unsigned long this_start_pfn, this_end_pfn;
5834 int i;
5835
5836 *start_pfn = -1UL;
5837 *end_pfn = 0;
5838
5839 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5840 *start_pfn = min(*start_pfn, this_start_pfn);
5841 *end_pfn = max(*end_pfn, this_end_pfn);
5842 }
5843
5844 if (*start_pfn == -1UL)
5845 *start_pfn = 0;
5846}
5847
5848/*
5849 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5850 * assumption is made that zones within a node are ordered in monotonic
5851 * increasing memory addresses so that the "highest" populated zone is used
5852 */
5853static void __init find_usable_zone_for_movable(void)
5854{
5855 int zone_index;
5856 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5857 if (zone_index == ZONE_MOVABLE)
5858 continue;
5859
5860 if (arch_zone_highest_possible_pfn[zone_index] >
5861 arch_zone_lowest_possible_pfn[zone_index])
5862 break;
5863 }
5864
5865 VM_BUG_ON(zone_index == -1);
5866 movable_zone = zone_index;
5867}
5868
5869/*
5870 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5871 * because it is sized independent of architecture. Unlike the other zones,
5872 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5873 * in each node depending on the size of each node and how evenly kernelcore
5874 * is distributed. This helper function adjusts the zone ranges
5875 * provided by the architecture for a given node by using the end of the
5876 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5877 * zones within a node are in order of monotonic increases memory addresses
5878 */
5879static void __meminit adjust_zone_range_for_zone_movable(int nid,
5880 unsigned long zone_type,
5881 unsigned long node_start_pfn,
5882 unsigned long node_end_pfn,
5883 unsigned long *zone_start_pfn,
5884 unsigned long *zone_end_pfn)
5885{
5886 /* Only adjust if ZONE_MOVABLE is on this node */
5887 if (zone_movable_pfn[nid]) {
5888 /* Size ZONE_MOVABLE */
5889 if (zone_type == ZONE_MOVABLE) {
5890 *zone_start_pfn = zone_movable_pfn[nid];
5891 *zone_end_pfn = min(node_end_pfn,
5892 arch_zone_highest_possible_pfn[movable_zone]);
5893
5894 /* Adjust for ZONE_MOVABLE starting within this range */
5895 } else if (!mirrored_kernelcore &&
5896 *zone_start_pfn < zone_movable_pfn[nid] &&
5897 *zone_end_pfn > zone_movable_pfn[nid]) {
5898 *zone_end_pfn = zone_movable_pfn[nid];
5899
5900 /* Check if this whole range is within ZONE_MOVABLE */
5901 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5902 *zone_start_pfn = *zone_end_pfn;
5903 }
5904}
5905
5906/*
5907 * Return the number of pages a zone spans in a node, including holes
5908 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5909 */
5910static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5911 unsigned long zone_type,
5912 unsigned long node_start_pfn,
5913 unsigned long node_end_pfn,
5914 unsigned long *zone_start_pfn,
5915 unsigned long *zone_end_pfn,
5916 unsigned long *ignored)
5917{
5918 /* When hotadd a new node from cpu_up(), the node should be empty */
5919 if (!node_start_pfn && !node_end_pfn)
5920 return 0;
5921
5922 /* Get the start and end of the zone */
5923 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5924 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5925 adjust_zone_range_for_zone_movable(nid, zone_type,
5926 node_start_pfn, node_end_pfn,
5927 zone_start_pfn, zone_end_pfn);
5928
5929 /* Check that this node has pages within the zone's required range */
5930 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5931 return 0;
5932
5933 /* Move the zone boundaries inside the node if necessary */
5934 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5935 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5936
5937 /* Return the spanned pages */
5938 return *zone_end_pfn - *zone_start_pfn;
5939}
5940
5941/*
5942 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5943 * then all holes in the requested range will be accounted for.
5944 */
5945unsigned long __meminit __absent_pages_in_range(int nid,
5946 unsigned long range_start_pfn,
5947 unsigned long range_end_pfn)
5948{
5949 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5950 unsigned long start_pfn, end_pfn;
5951 int i;
5952
5953 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5954 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5955 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5956 nr_absent -= end_pfn - start_pfn;
5957 }
5958 return nr_absent;
5959}
5960
5961/**
5962 * absent_pages_in_range - Return number of page frames in holes within a range
5963 * @start_pfn: The start PFN to start searching for holes
5964 * @end_pfn: The end PFN to stop searching for holes
5965 *
5966 * It returns the number of pages frames in memory holes within a range.
5967 */
5968unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5969 unsigned long end_pfn)
5970{
5971 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5972}
5973
5974/* Return the number of page frames in holes in a zone on a node */
5975static unsigned long __meminit zone_absent_pages_in_node(int nid,
5976 unsigned long zone_type,
5977 unsigned long node_start_pfn,
5978 unsigned long node_end_pfn,
5979 unsigned long *ignored)
5980{
5981 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5982 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5983 unsigned long zone_start_pfn, zone_end_pfn;
5984 unsigned long nr_absent;
5985
5986 /* When hotadd a new node from cpu_up(), the node should be empty */
5987 if (!node_start_pfn && !node_end_pfn)
5988 return 0;
5989
5990 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5991 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5992
5993 adjust_zone_range_for_zone_movable(nid, zone_type,
5994 node_start_pfn, node_end_pfn,
5995 &zone_start_pfn, &zone_end_pfn);
5996 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5997
5998 /*
5999 * ZONE_MOVABLE handling.
6000 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6001 * and vice versa.
6002 */
6003 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6004 unsigned long start_pfn, end_pfn;
6005 struct memblock_region *r;
6006
6007 for_each_memblock(memory, r) {
6008 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6009 zone_start_pfn, zone_end_pfn);
6010 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6011 zone_start_pfn, zone_end_pfn);
6012
6013 if (zone_type == ZONE_MOVABLE &&
6014 memblock_is_mirror(r))
6015 nr_absent += end_pfn - start_pfn;
6016
6017 if (zone_type == ZONE_NORMAL &&
6018 !memblock_is_mirror(r))
6019 nr_absent += end_pfn - start_pfn;
6020 }
6021 }
6022
6023 return nr_absent;
6024}
6025
6026#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6027static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
6028 unsigned long zone_type,
6029 unsigned long node_start_pfn,
6030 unsigned long node_end_pfn,
6031 unsigned long *zone_start_pfn,
6032 unsigned long *zone_end_pfn,
6033 unsigned long *zones_size)
6034{
6035 unsigned int zone;
6036
6037 *zone_start_pfn = node_start_pfn;
6038 for (zone = 0; zone < zone_type; zone++)
6039 *zone_start_pfn += zones_size[zone];
6040
6041 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6042
6043 return zones_size[zone_type];
6044}
6045
6046static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
6047 unsigned long zone_type,
6048 unsigned long node_start_pfn,
6049 unsigned long node_end_pfn,
6050 unsigned long *zholes_size)
6051{
6052 if (!zholes_size)
6053 return 0;
6054
6055 return zholes_size[zone_type];
6056}
6057
6058#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6059
6060static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
6061 unsigned long node_start_pfn,
6062 unsigned long node_end_pfn,
6063 unsigned long *zones_size,
6064 unsigned long *zholes_size)
6065{
6066 unsigned long realtotalpages = 0, totalpages = 0;
6067 enum zone_type i;
6068
6069 for (i = 0; i < MAX_NR_ZONES; i++) {
6070 struct zone *zone = pgdat->node_zones + i;
6071 unsigned long zone_start_pfn, zone_end_pfn;
6072 unsigned long size, real_size;
6073
6074 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6075 node_start_pfn,
6076 node_end_pfn,
6077 &zone_start_pfn,
6078 &zone_end_pfn,
6079 zones_size);
6080 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6081 node_start_pfn, node_end_pfn,
6082 zholes_size);
6083 if (size)
6084 zone->zone_start_pfn = zone_start_pfn;
6085 else
6086 zone->zone_start_pfn = 0;
6087 zone->spanned_pages = size;
6088 zone->present_pages = real_size;
6089
6090 totalpages += size;
6091 realtotalpages += real_size;
6092 }
6093
6094 pgdat->node_spanned_pages = totalpages;
6095 pgdat->node_present_pages = realtotalpages;
6096 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6097 realtotalpages);
6098}
6099
6100#ifndef CONFIG_SPARSEMEM
6101/*
6102 * Calculate the size of the zone->blockflags rounded to an unsigned long
6103 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6104 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6105 * round what is now in bits to nearest long in bits, then return it in
6106 * bytes.
6107 */
6108static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6109{
6110 unsigned long usemapsize;
6111
6112 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6113 usemapsize = roundup(zonesize, pageblock_nr_pages);
6114 usemapsize = usemapsize >> pageblock_order;
6115 usemapsize *= NR_PAGEBLOCK_BITS;
6116 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6117
6118 return usemapsize / 8;
6119}
6120
6121static void __init setup_usemap(struct pglist_data *pgdat,
6122 struct zone *zone,
6123 unsigned long zone_start_pfn,
6124 unsigned long zonesize)
6125{
6126 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6127 zone->pageblock_flags = NULL;
6128 if (usemapsize)
6129 zone->pageblock_flags =
6130 memblock_virt_alloc_node_nopanic(usemapsize,
6131 pgdat->node_id);
6132}
6133#else
6134static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6135 unsigned long zone_start_pfn, unsigned long zonesize) {}
6136#endif /* CONFIG_SPARSEMEM */
6137
6138#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6139
6140/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6141void __paginginit set_pageblock_order(void)
6142{
6143 unsigned int order;
6144
6145 /* Check that pageblock_nr_pages has not already been setup */
6146 if (pageblock_order)
6147 return;
6148
6149 if (HPAGE_SHIFT > PAGE_SHIFT)
6150 order = HUGETLB_PAGE_ORDER;
6151 else
6152 order = MAX_ORDER - 1;
6153
6154 /*
6155 * Assume the largest contiguous order of interest is a huge page.
6156 * This value may be variable depending on boot parameters on IA64 and
6157 * powerpc.
6158 */
6159 pageblock_order = order;
6160}
6161#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6162
6163/*
6164 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6165 * is unused as pageblock_order is set at compile-time. See
6166 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6167 * the kernel config
6168 */
6169void __paginginit set_pageblock_order(void)
6170{
6171}
6172
6173#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6174
6175static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6176 unsigned long present_pages)
6177{
6178 unsigned long pages = spanned_pages;
6179
6180 /*
6181 * Provide a more accurate estimation if there are holes within
6182 * the zone and SPARSEMEM is in use. If there are holes within the
6183 * zone, each populated memory region may cost us one or two extra
6184 * memmap pages due to alignment because memmap pages for each
6185 * populated regions may not be naturally aligned on page boundary.
6186 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6187 */
6188 if (spanned_pages > present_pages + (present_pages >> 4) &&
6189 IS_ENABLED(CONFIG_SPARSEMEM))
6190 pages = present_pages;
6191
6192 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6193}
6194
6195/*
6196 * Set up the zone data structures:
6197 * - mark all pages reserved
6198 * - mark all memory queues empty
6199 * - clear the memory bitmaps
6200 *
6201 * NOTE: pgdat should get zeroed by caller.
6202 */
6203static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6204{
6205 enum zone_type j;
6206 int nid = pgdat->node_id;
6207
6208 pgdat_resize_init(pgdat);
6209#ifdef CONFIG_NUMA_BALANCING
6210 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6211 pgdat->numabalancing_migrate_nr_pages = 0;
6212 pgdat->numabalancing_migrate_next_window = jiffies;
6213#endif
6214#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6215 spin_lock_init(&pgdat->split_queue_lock);
6216 INIT_LIST_HEAD(&pgdat->split_queue);
6217 pgdat->split_queue_len = 0;
6218#endif
6219 init_waitqueue_head(&pgdat->kswapd_wait);
6220 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6221#ifdef CONFIG_COMPACTION
6222 init_waitqueue_head(&pgdat->kcompactd_wait);
6223#endif
6224 pgdat_page_ext_init(pgdat);
6225 spin_lock_init(&pgdat->lru_lock);
6226 lruvec_init(node_lruvec(pgdat));
6227
6228 pgdat->per_cpu_nodestats = &boot_nodestats;
6229
6230 for (j = 0; j < MAX_NR_ZONES; j++) {
6231 struct zone *zone = pgdat->node_zones + j;
6232 unsigned long size, realsize, freesize, memmap_pages;
6233 unsigned long zone_start_pfn = zone->zone_start_pfn;
6234
6235 size = zone->spanned_pages;
6236 realsize = freesize = zone->present_pages;
6237
6238 /*
6239 * Adjust freesize so that it accounts for how much memory
6240 * is used by this zone for memmap. This affects the watermark
6241 * and per-cpu initialisations
6242 */
6243 memmap_pages = calc_memmap_size(size, realsize);
6244 if (!is_highmem_idx(j)) {
6245 if (freesize >= memmap_pages) {
6246 freesize -= memmap_pages;
6247 if (memmap_pages)
6248 printk(KERN_DEBUG
6249 " %s zone: %lu pages used for memmap\n",
6250 zone_names[j], memmap_pages);
6251 } else
6252 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6253 zone_names[j], memmap_pages, freesize);
6254 }
6255
6256 /* Account for reserved pages */
6257 if (j == 0 && freesize > dma_reserve) {
6258 freesize -= dma_reserve;
6259 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6260 zone_names[0], dma_reserve);
6261 }
6262
6263 if (!is_highmem_idx(j))
6264 nr_kernel_pages += freesize;
6265 /* Charge for highmem memmap if there are enough kernel pages */
6266 else if (nr_kernel_pages > memmap_pages * 2)
6267 nr_kernel_pages -= memmap_pages;
6268 nr_all_pages += freesize;
6269
6270 /*
6271 * Set an approximate value for lowmem here, it will be adjusted
6272 * when the bootmem allocator frees pages into the buddy system.
6273 * And all highmem pages will be managed by the buddy system.
6274 */
6275 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6276#ifdef CONFIG_NUMA
6277 zone->node = nid;
6278#endif
6279 zone->name = zone_names[j];
6280 zone->zone_pgdat = pgdat;
6281 spin_lock_init(&zone->lock);
6282 zone_seqlock_init(zone);
6283 zone_pcp_init(zone);
6284
6285 if (!size)
6286 continue;
6287
6288 set_pageblock_order();
6289 setup_usemap(pgdat, zone, zone_start_pfn, size);
6290 init_currently_empty_zone(zone, zone_start_pfn, size);
6291 memmap_init(size, nid, j, zone_start_pfn);
6292 }
6293}
6294
6295#ifdef CONFIG_FLAT_NODE_MEM_MAP
6296static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6297{
6298 unsigned long __maybe_unused start = 0;
6299 unsigned long __maybe_unused offset = 0;
6300
6301 /* Skip empty nodes */
6302 if (!pgdat->node_spanned_pages)
6303 return;
6304
6305 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6306 offset = pgdat->node_start_pfn - start;
6307 /* ia64 gets its own node_mem_map, before this, without bootmem */
6308 if (!pgdat->node_mem_map) {
6309 unsigned long size, end;
6310 struct page *map;
6311
6312 /*
6313 * The zone's endpoints aren't required to be MAX_ORDER
6314 * aligned but the node_mem_map endpoints must be in order
6315 * for the buddy allocator to function correctly.
6316 */
6317 end = pgdat_end_pfn(pgdat);
6318 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6319 size = (end - start) * sizeof(struct page);
6320 map = memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
6321 pgdat->node_mem_map = map + offset;
6322 }
6323 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6324 __func__, pgdat->node_id, (unsigned long)pgdat,
6325 (unsigned long)pgdat->node_mem_map);
6326#ifndef CONFIG_NEED_MULTIPLE_NODES
6327 /*
6328 * With no DISCONTIG, the global mem_map is just set as node 0's
6329 */
6330 if (pgdat == NODE_DATA(0)) {
6331 mem_map = NODE_DATA(0)->node_mem_map;
6332#if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6333 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6334 mem_map -= offset;
6335#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6336 }
6337#endif
6338}
6339#else
6340static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6341#endif /* CONFIG_FLAT_NODE_MEM_MAP */
6342
6343void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6344 unsigned long node_start_pfn, unsigned long *zholes_size)
6345{
6346 pg_data_t *pgdat = NODE_DATA(nid);
6347 unsigned long start_pfn = 0;
6348 unsigned long end_pfn = 0;
6349
6350 /* pg_data_t should be reset to zero when it's allocated */
6351 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6352
6353 pgdat->node_id = nid;
6354 pgdat->node_start_pfn = node_start_pfn;
6355 pgdat->per_cpu_nodestats = NULL;
6356#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6357 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6358 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6359 (u64)start_pfn << PAGE_SHIFT,
6360 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6361#else
6362 start_pfn = node_start_pfn;
6363#endif
6364 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6365 zones_size, zholes_size);
6366
6367 alloc_node_mem_map(pgdat);
6368
6369#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6370 /*
6371 * We start only with one section of pages, more pages are added as
6372 * needed until the rest of deferred pages are initialized.
6373 */
6374 pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
6375 pgdat->node_spanned_pages);
6376 pgdat->first_deferred_pfn = ULONG_MAX;
6377#endif
6378 free_area_init_core(pgdat);
6379}
6380
6381#ifdef CONFIG_HAVE_MEMBLOCK
6382/*
6383 * Only struct pages that are backed by physical memory are zeroed and
6384 * initialized by going through __init_single_page(). But, there are some
6385 * struct pages which are reserved in memblock allocator and their fields
6386 * may be accessed (for example page_to_pfn() on some configuration accesses
6387 * flags). We must explicitly zero those struct pages.
6388 */
6389void __paginginit zero_resv_unavail(void)
6390{
6391 phys_addr_t start, end;
6392 unsigned long pfn;
6393 u64 i, pgcnt;
6394
6395 /*
6396 * Loop through ranges that are reserved, but do not have reported
6397 * physical memory backing.
6398 */
6399 pgcnt = 0;
6400 for_each_resv_unavail_range(i, &start, &end) {
6401 for (pfn = PFN_DOWN(start); pfn < PFN_UP(end); pfn++) {
6402 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages)))
6403 continue;
6404 mm_zero_struct_page(pfn_to_page(pfn));
6405 pgcnt++;
6406 }
6407 }
6408
6409 /*
6410 * Struct pages that do not have backing memory. This could be because
6411 * firmware is using some of this memory, or for some other reasons.
6412 * Once memblock is changed so such behaviour is not allowed: i.e.
6413 * list of "reserved" memory must be a subset of list of "memory", then
6414 * this code can be removed.
6415 */
6416 if (pgcnt)
6417 pr_info("Reserved but unavailable: %lld pages", pgcnt);
6418}
6419#endif /* CONFIG_HAVE_MEMBLOCK */
6420
6421#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6422
6423#if MAX_NUMNODES > 1
6424/*
6425 * Figure out the number of possible node ids.
6426 */
6427void __init setup_nr_node_ids(void)
6428{
6429 unsigned int highest;
6430
6431 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6432 nr_node_ids = highest + 1;
6433}
6434#endif
6435
6436/**
6437 * node_map_pfn_alignment - determine the maximum internode alignment
6438 *
6439 * This function should be called after node map is populated and sorted.
6440 * It calculates the maximum power of two alignment which can distinguish
6441 * all the nodes.
6442 *
6443 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6444 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6445 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6446 * shifted, 1GiB is enough and this function will indicate so.
6447 *
6448 * This is used to test whether pfn -> nid mapping of the chosen memory
6449 * model has fine enough granularity to avoid incorrect mapping for the
6450 * populated node map.
6451 *
6452 * Returns the determined alignment in pfn's. 0 if there is no alignment
6453 * requirement (single node).
6454 */
6455unsigned long __init node_map_pfn_alignment(void)
6456{
6457 unsigned long accl_mask = 0, last_end = 0;
6458 unsigned long start, end, mask;
6459 int last_nid = -1;
6460 int i, nid;
6461
6462 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6463 if (!start || last_nid < 0 || last_nid == nid) {
6464 last_nid = nid;
6465 last_end = end;
6466 continue;
6467 }
6468
6469 /*
6470 * Start with a mask granular enough to pin-point to the
6471 * start pfn and tick off bits one-by-one until it becomes
6472 * too coarse to separate the current node from the last.
6473 */
6474 mask = ~((1 << __ffs(start)) - 1);
6475 while (mask && last_end <= (start & (mask << 1)))
6476 mask <<= 1;
6477
6478 /* accumulate all internode masks */
6479 accl_mask |= mask;
6480 }
6481
6482 /* convert mask to number of pages */
6483 return ~accl_mask + 1;
6484}
6485
6486/* Find the lowest pfn for a node */
6487static unsigned long __init find_min_pfn_for_node(int nid)
6488{
6489 unsigned long min_pfn = ULONG_MAX;
6490 unsigned long start_pfn;
6491 int i;
6492
6493 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6494 min_pfn = min(min_pfn, start_pfn);
6495
6496 if (min_pfn == ULONG_MAX) {
6497 pr_warn("Could not find start_pfn for node %d\n", nid);
6498 return 0;
6499 }
6500
6501 return min_pfn;
6502}
6503
6504/**
6505 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6506 *
6507 * It returns the minimum PFN based on information provided via
6508 * memblock_set_node().
6509 */
6510unsigned long __init find_min_pfn_with_active_regions(void)
6511{
6512 return find_min_pfn_for_node(MAX_NUMNODES);
6513}
6514
6515/*
6516 * early_calculate_totalpages()
6517 * Sum pages in active regions for movable zone.
6518 * Populate N_MEMORY for calculating usable_nodes.
6519 */
6520static unsigned long __init early_calculate_totalpages(void)
6521{
6522 unsigned long totalpages = 0;
6523 unsigned long start_pfn, end_pfn;
6524 int i, nid;
6525
6526 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6527 unsigned long pages = end_pfn - start_pfn;
6528
6529 totalpages += pages;
6530 if (pages)
6531 node_set_state(nid, N_MEMORY);
6532 }
6533 return totalpages;
6534}
6535
6536/*
6537 * Find the PFN the Movable zone begins in each node. Kernel memory
6538 * is spread evenly between nodes as long as the nodes have enough
6539 * memory. When they don't, some nodes will have more kernelcore than
6540 * others
6541 */
6542static void __init find_zone_movable_pfns_for_nodes(void)
6543{
6544 int i, nid;
6545 unsigned long usable_startpfn;
6546 unsigned long kernelcore_node, kernelcore_remaining;
6547 /* save the state before borrow the nodemask */
6548 nodemask_t saved_node_state = node_states[N_MEMORY];
6549 unsigned long totalpages = early_calculate_totalpages();
6550 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6551 struct memblock_region *r;
6552
6553 /* Need to find movable_zone earlier when movable_node is specified. */
6554 find_usable_zone_for_movable();
6555
6556 /*
6557 * If movable_node is specified, ignore kernelcore and movablecore
6558 * options.
6559 */
6560 if (movable_node_is_enabled()) {
6561 for_each_memblock(memory, r) {
6562 if (!memblock_is_hotpluggable(r))
6563 continue;
6564
6565 nid = r->nid;
6566
6567 usable_startpfn = PFN_DOWN(r->base);
6568 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6569 min(usable_startpfn, zone_movable_pfn[nid]) :
6570 usable_startpfn;
6571 }
6572
6573 goto out2;
6574 }
6575
6576 /*
6577 * If kernelcore=mirror is specified, ignore movablecore option
6578 */
6579 if (mirrored_kernelcore) {
6580 bool mem_below_4gb_not_mirrored = false;
6581
6582 for_each_memblock(memory, r) {
6583 if (memblock_is_mirror(r))
6584 continue;
6585
6586 nid = r->nid;
6587
6588 usable_startpfn = memblock_region_memory_base_pfn(r);
6589
6590 if (usable_startpfn < 0x100000) {
6591 mem_below_4gb_not_mirrored = true;
6592 continue;
6593 }
6594
6595 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6596 min(usable_startpfn, zone_movable_pfn[nid]) :
6597 usable_startpfn;
6598 }
6599
6600 if (mem_below_4gb_not_mirrored)
6601 pr_warn("This configuration results in unmirrored kernel memory.");
6602
6603 goto out2;
6604 }
6605
6606 /*
6607 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6608 * amount of necessary memory.
6609 */
6610 if (required_kernelcore_percent)
6611 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6612 10000UL;
6613 if (required_movablecore_percent)
6614 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6615 10000UL;
6616
6617 /*
6618 * If movablecore= was specified, calculate what size of
6619 * kernelcore that corresponds so that memory usable for
6620 * any allocation type is evenly spread. If both kernelcore
6621 * and movablecore are specified, then the value of kernelcore
6622 * will be used for required_kernelcore if it's greater than
6623 * what movablecore would have allowed.
6624 */
6625 if (required_movablecore) {
6626 unsigned long corepages;
6627
6628 /*
6629 * Round-up so that ZONE_MOVABLE is at least as large as what
6630 * was requested by the user
6631 */
6632 required_movablecore =
6633 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6634 required_movablecore = min(totalpages, required_movablecore);
6635 corepages = totalpages - required_movablecore;
6636
6637 required_kernelcore = max(required_kernelcore, corepages);
6638 }
6639
6640 /*
6641 * If kernelcore was not specified or kernelcore size is larger
6642 * than totalpages, there is no ZONE_MOVABLE.
6643 */
6644 if (!required_kernelcore || required_kernelcore >= totalpages)
6645 goto out;
6646
6647 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6648 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6649
6650restart:
6651 /* Spread kernelcore memory as evenly as possible throughout nodes */
6652 kernelcore_node = required_kernelcore / usable_nodes;
6653 for_each_node_state(nid, N_MEMORY) {
6654 unsigned long start_pfn, end_pfn;
6655
6656 /*
6657 * Recalculate kernelcore_node if the division per node
6658 * now exceeds what is necessary to satisfy the requested
6659 * amount of memory for the kernel
6660 */
6661 if (required_kernelcore < kernelcore_node)
6662 kernelcore_node = required_kernelcore / usable_nodes;
6663
6664 /*
6665 * As the map is walked, we track how much memory is usable
6666 * by the kernel using kernelcore_remaining. When it is
6667 * 0, the rest of the node is usable by ZONE_MOVABLE
6668 */
6669 kernelcore_remaining = kernelcore_node;
6670
6671 /* Go through each range of PFNs within this node */
6672 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6673 unsigned long size_pages;
6674
6675 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6676 if (start_pfn >= end_pfn)
6677 continue;
6678
6679 /* Account for what is only usable for kernelcore */
6680 if (start_pfn < usable_startpfn) {
6681 unsigned long kernel_pages;
6682 kernel_pages = min(end_pfn, usable_startpfn)
6683 - start_pfn;
6684
6685 kernelcore_remaining -= min(kernel_pages,
6686 kernelcore_remaining);
6687 required_kernelcore -= min(kernel_pages,
6688 required_kernelcore);
6689
6690 /* Continue if range is now fully accounted */
6691 if (end_pfn <= usable_startpfn) {
6692
6693 /*
6694 * Push zone_movable_pfn to the end so
6695 * that if we have to rebalance
6696 * kernelcore across nodes, we will
6697 * not double account here
6698 */
6699 zone_movable_pfn[nid] = end_pfn;
6700 continue;
6701 }
6702 start_pfn = usable_startpfn;
6703 }
6704
6705 /*
6706 * The usable PFN range for ZONE_MOVABLE is from
6707 * start_pfn->end_pfn. Calculate size_pages as the
6708 * number of pages used as kernelcore
6709 */
6710 size_pages = end_pfn - start_pfn;
6711 if (size_pages > kernelcore_remaining)
6712 size_pages = kernelcore_remaining;
6713 zone_movable_pfn[nid] = start_pfn + size_pages;
6714
6715 /*
6716 * Some kernelcore has been met, update counts and
6717 * break if the kernelcore for this node has been
6718 * satisfied
6719 */
6720 required_kernelcore -= min(required_kernelcore,
6721 size_pages);
6722 kernelcore_remaining -= size_pages;
6723 if (!kernelcore_remaining)
6724 break;
6725 }
6726 }
6727
6728 /*
6729 * If there is still required_kernelcore, we do another pass with one
6730 * less node in the count. This will push zone_movable_pfn[nid] further
6731 * along on the nodes that still have memory until kernelcore is
6732 * satisfied
6733 */
6734 usable_nodes--;
6735 if (usable_nodes && required_kernelcore > usable_nodes)
6736 goto restart;
6737
6738out2:
6739 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6740 for (nid = 0; nid < MAX_NUMNODES; nid++)
6741 zone_movable_pfn[nid] =
6742 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6743
6744out:
6745 /* restore the node_state */
6746 node_states[N_MEMORY] = saved_node_state;
6747}
6748
6749/* Any regular or high memory on that node ? */
6750static void check_for_memory(pg_data_t *pgdat, int nid)
6751{
6752 enum zone_type zone_type;
6753
6754 if (N_MEMORY == N_NORMAL_MEMORY)
6755 return;
6756
6757 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6758 struct zone *zone = &pgdat->node_zones[zone_type];
6759 if (populated_zone(zone)) {
6760 node_set_state(nid, N_HIGH_MEMORY);
6761 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6762 zone_type <= ZONE_NORMAL)
6763 node_set_state(nid, N_NORMAL_MEMORY);
6764 break;
6765 }
6766 }
6767}
6768
6769/**
6770 * free_area_init_nodes - Initialise all pg_data_t and zone data
6771 * @max_zone_pfn: an array of max PFNs for each zone
6772 *
6773 * This will call free_area_init_node() for each active node in the system.
6774 * Using the page ranges provided by memblock_set_node(), the size of each
6775 * zone in each node and their holes is calculated. If the maximum PFN
6776 * between two adjacent zones match, it is assumed that the zone is empty.
6777 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6778 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6779 * starts where the previous one ended. For example, ZONE_DMA32 starts
6780 * at arch_max_dma_pfn.
6781 */
6782void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6783{
6784 unsigned long start_pfn, end_pfn;
6785 int i, nid;
6786
6787 /* Record where the zone boundaries are */
6788 memset(arch_zone_lowest_possible_pfn, 0,
6789 sizeof(arch_zone_lowest_possible_pfn));
6790 memset(arch_zone_highest_possible_pfn, 0,
6791 sizeof(arch_zone_highest_possible_pfn));
6792
6793 start_pfn = find_min_pfn_with_active_regions();
6794
6795 for (i = 0; i < MAX_NR_ZONES; i++) {
6796 if (i == ZONE_MOVABLE)
6797 continue;
6798
6799 end_pfn = max(max_zone_pfn[i], start_pfn);
6800 arch_zone_lowest_possible_pfn[i] = start_pfn;
6801 arch_zone_highest_possible_pfn[i] = end_pfn;
6802
6803 start_pfn = end_pfn;
6804 }
6805
6806 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6807 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6808 find_zone_movable_pfns_for_nodes();
6809
6810 /* Print out the zone ranges */
6811 pr_info("Zone ranges:\n");
6812 for (i = 0; i < MAX_NR_ZONES; i++) {
6813 if (i == ZONE_MOVABLE)
6814 continue;
6815 pr_info(" %-8s ", zone_names[i]);
6816 if (arch_zone_lowest_possible_pfn[i] ==
6817 arch_zone_highest_possible_pfn[i])
6818 pr_cont("empty\n");
6819 else
6820 pr_cont("[mem %#018Lx-%#018Lx]\n",
6821 (u64)arch_zone_lowest_possible_pfn[i]
6822 << PAGE_SHIFT,
6823 ((u64)arch_zone_highest_possible_pfn[i]
6824 << PAGE_SHIFT) - 1);
6825 }
6826
6827 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6828 pr_info("Movable zone start for each node\n");
6829 for (i = 0; i < MAX_NUMNODES; i++) {
6830 if (zone_movable_pfn[i])
6831 pr_info(" Node %d: %#018Lx\n", i,
6832 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6833 }
6834
6835 /* Print out the early node map */
6836 pr_info("Early memory node ranges\n");
6837 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6838 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6839 (u64)start_pfn << PAGE_SHIFT,
6840 ((u64)end_pfn << PAGE_SHIFT) - 1);
6841
6842 /* Initialise every node */
6843 mminit_verify_pageflags_layout();
6844 setup_nr_node_ids();
6845 for_each_online_node(nid) {
6846 pg_data_t *pgdat = NODE_DATA(nid);
6847 free_area_init_node(nid, NULL,
6848 find_min_pfn_for_node(nid), NULL);
6849
6850 /* Any memory on that node */
6851 if (pgdat->node_present_pages)
6852 node_set_state(nid, N_MEMORY);
6853 check_for_memory(pgdat, nid);
6854 }
6855 zero_resv_unavail();
6856}
6857
6858static int __init cmdline_parse_core(char *p, unsigned long *core,
6859 unsigned long *percent)
6860{
6861 unsigned long long coremem;
6862 char *endptr;
6863
6864 if (!p)
6865 return -EINVAL;
6866
6867 /* Value may be a percentage of total memory, otherwise bytes */
6868 coremem = simple_strtoull(p, &endptr, 0);
6869 if (*endptr == '%') {
6870 /* Paranoid check for percent values greater than 100 */
6871 WARN_ON(coremem > 100);
6872
6873 *percent = coremem;
6874 } else {
6875 coremem = memparse(p, &p);
6876 /* Paranoid check that UL is enough for the coremem value */
6877 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6878
6879 *core = coremem >> PAGE_SHIFT;
6880 *percent = 0UL;
6881 }
6882 return 0;
6883}
6884
6885/*
6886 * kernelcore=size sets the amount of memory for use for allocations that
6887 * cannot be reclaimed or migrated.
6888 */
6889static int __init cmdline_parse_kernelcore(char *p)
6890{
6891 /* parse kernelcore=mirror */
6892 if (parse_option_str(p, "mirror")) {
6893 mirrored_kernelcore = true;
6894 return 0;
6895 }
6896
6897 return cmdline_parse_core(p, &required_kernelcore,
6898 &required_kernelcore_percent);
6899}
6900
6901/*
6902 * movablecore=size sets the amount of memory for use for allocations that
6903 * can be reclaimed or migrated.
6904 */
6905static int __init cmdline_parse_movablecore(char *p)
6906{
6907 return cmdline_parse_core(p, &required_movablecore,
6908 &required_movablecore_percent);
6909}
6910
6911early_param("kernelcore", cmdline_parse_kernelcore);
6912early_param("movablecore", cmdline_parse_movablecore);
6913
6914#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6915
6916void adjust_managed_page_count(struct page *page, long count)
6917{
6918 spin_lock(&managed_page_count_lock);
6919 page_zone(page)->managed_pages += count;
6920 totalram_pages += count;
6921#ifdef CONFIG_HIGHMEM
6922 if (PageHighMem(page))
6923 totalhigh_pages += count;
6924#endif
6925 spin_unlock(&managed_page_count_lock);
6926}
6927EXPORT_SYMBOL(adjust_managed_page_count);
6928
6929unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6930{
6931 void *pos;
6932 unsigned long pages = 0;
6933
6934 start = (void *)PAGE_ALIGN((unsigned long)start);
6935 end = (void *)((unsigned long)end & PAGE_MASK);
6936 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6937 if ((unsigned int)poison <= 0xFF)
6938 memset(pos, poison, PAGE_SIZE);
6939 free_reserved_page(virt_to_page(pos));
6940 }
6941
6942 if (pages && s)
6943 pr_info("Freeing %s memory: %ldK\n",
6944 s, pages << (PAGE_SHIFT - 10));
6945
6946 return pages;
6947}
6948EXPORT_SYMBOL(free_reserved_area);
6949
6950#ifdef CONFIG_HIGHMEM
6951void free_highmem_page(struct page *page)
6952{
6953 __free_reserved_page(page);
6954 totalram_pages++;
6955 page_zone(page)->managed_pages++;
6956 totalhigh_pages++;
6957}
6958#endif
6959
6960
6961void __init mem_init_print_info(const char *str)
6962{
6963 unsigned long physpages, codesize, datasize, rosize, bss_size;
6964 unsigned long init_code_size, init_data_size;
6965
6966 physpages = get_num_physpages();
6967 codesize = _etext - _stext;
6968 datasize = _edata - _sdata;
6969 rosize = __end_rodata - __start_rodata;
6970 bss_size = __bss_stop - __bss_start;
6971 init_data_size = __init_end - __init_begin;
6972 init_code_size = _einittext - _sinittext;
6973
6974 /*
6975 * Detect special cases and adjust section sizes accordingly:
6976 * 1) .init.* may be embedded into .data sections
6977 * 2) .init.text.* may be out of [__init_begin, __init_end],
6978 * please refer to arch/tile/kernel/vmlinux.lds.S.
6979 * 3) .rodata.* may be embedded into .text or .data sections.
6980 */
6981#define adj_init_size(start, end, size, pos, adj) \
6982 do { \
6983 if (start <= pos && pos < end && size > adj) \
6984 size -= adj; \
6985 } while (0)
6986
6987 adj_init_size(__init_begin, __init_end, init_data_size,
6988 _sinittext, init_code_size);
6989 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6990 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6991 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6992 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6993
6994#undef adj_init_size
6995
6996 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6997#ifdef CONFIG_HIGHMEM
6998 ", %luK highmem"
6999#endif
7000 "%s%s)\n",
7001 nr_free_pages() << (PAGE_SHIFT - 10),
7002 physpages << (PAGE_SHIFT - 10),
7003 codesize >> 10, datasize >> 10, rosize >> 10,
7004 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7005 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
7006 totalcma_pages << (PAGE_SHIFT - 10),
7007#ifdef CONFIG_HIGHMEM
7008 totalhigh_pages << (PAGE_SHIFT - 10),
7009#endif
7010 str ? ", " : "", str ? str : "");
7011}
7012
7013/**
7014 * set_dma_reserve - set the specified number of pages reserved in the first zone
7015 * @new_dma_reserve: The number of pages to mark reserved
7016 *
7017 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7018 * In the DMA zone, a significant percentage may be consumed by kernel image
7019 * and other unfreeable allocations which can skew the watermarks badly. This
7020 * function may optionally be used to account for unfreeable pages in the
7021 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7022 * smaller per-cpu batchsize.
7023 */
7024void __init set_dma_reserve(unsigned long new_dma_reserve)
7025{
7026 dma_reserve = new_dma_reserve;
7027}
7028
7029void __init free_area_init(unsigned long *zones_size)
7030{
7031 free_area_init_node(0, zones_size,
7032 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7033 zero_resv_unavail();
7034}
7035
7036static int page_alloc_cpu_dead(unsigned int cpu)
7037{
7038
7039 lru_add_drain_cpu(cpu);
7040 drain_pages(cpu);
7041
7042 /*
7043 * Spill the event counters of the dead processor
7044 * into the current processors event counters.
7045 * This artificially elevates the count of the current
7046 * processor.
7047 */
7048 vm_events_fold_cpu(cpu);
7049
7050 /*
7051 * Zero the differential counters of the dead processor
7052 * so that the vm statistics are consistent.
7053 *
7054 * This is only okay since the processor is dead and cannot
7055 * race with what we are doing.
7056 */
7057 cpu_vm_stats_fold(cpu);
7058 return 0;
7059}
7060
7061void __init page_alloc_init(void)
7062{
7063 int ret;
7064
7065 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7066 "mm/page_alloc:dead", NULL,
7067 page_alloc_cpu_dead);
7068 WARN_ON(ret < 0);
7069}
7070
7071/*
7072 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7073 * or min_free_kbytes changes.
7074 */
7075static void calculate_totalreserve_pages(void)
7076{
7077 struct pglist_data *pgdat;
7078 unsigned long reserve_pages = 0;
7079 enum zone_type i, j;
7080
7081 for_each_online_pgdat(pgdat) {
7082
7083 pgdat->totalreserve_pages = 0;
7084
7085 for (i = 0; i < MAX_NR_ZONES; i++) {
7086 struct zone *zone = pgdat->node_zones + i;
7087 long max = 0;
7088
7089 /* Find valid and maximum lowmem_reserve in the zone */
7090 for (j = i; j < MAX_NR_ZONES; j++) {
7091 if (zone->lowmem_reserve[j] > max)
7092 max = zone->lowmem_reserve[j];
7093 }
7094
7095 /* we treat the high watermark as reserved pages. */
7096 max += high_wmark_pages(zone);
7097
7098 if (max > zone->managed_pages)
7099 max = zone->managed_pages;
7100
7101 pgdat->totalreserve_pages += max;
7102
7103 reserve_pages += max;
7104 }
7105 }
7106 totalreserve_pages = reserve_pages;
7107}
7108
7109/*
7110 * setup_per_zone_lowmem_reserve - called whenever
7111 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7112 * has a correct pages reserved value, so an adequate number of
7113 * pages are left in the zone after a successful __alloc_pages().
7114 */
7115static void setup_per_zone_lowmem_reserve(void)
7116{
7117 struct pglist_data *pgdat;
7118 enum zone_type j, idx;
7119
7120 for_each_online_pgdat(pgdat) {
7121 for (j = 0; j < MAX_NR_ZONES; j++) {
7122 struct zone *zone = pgdat->node_zones + j;
7123 unsigned long managed_pages = zone->managed_pages;
7124
7125 zone->lowmem_reserve[j] = 0;
7126
7127 idx = j;
7128 while (idx) {
7129 struct zone *lower_zone;
7130
7131 idx--;
7132 lower_zone = pgdat->node_zones + idx;
7133
7134 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7135 sysctl_lowmem_reserve_ratio[idx] = 0;
7136 lower_zone->lowmem_reserve[j] = 0;
7137 } else {
7138 lower_zone->lowmem_reserve[j] =
7139 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7140 }
7141 managed_pages += lower_zone->managed_pages;
7142 }
7143 }
7144 }
7145
7146 /* update totalreserve_pages */
7147 calculate_totalreserve_pages();
7148}
7149
7150static void __setup_per_zone_wmarks(void)
7151{
7152 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7153 unsigned long lowmem_pages = 0;
7154 struct zone *zone;
7155 unsigned long flags;
7156
7157 /* Calculate total number of !ZONE_HIGHMEM pages */
7158 for_each_zone(zone) {
7159 if (!is_highmem(zone))
7160 lowmem_pages += zone->managed_pages;
7161 }
7162
7163 for_each_zone(zone) {
7164 u64 tmp;
7165
7166 spin_lock_irqsave(&zone->lock, flags);
7167 tmp = (u64)pages_min * zone->managed_pages;
7168 do_div(tmp, lowmem_pages);
7169 if (is_highmem(zone)) {
7170 /*
7171 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7172 * need highmem pages, so cap pages_min to a small
7173 * value here.
7174 *
7175 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7176 * deltas control asynch page reclaim, and so should
7177 * not be capped for highmem.
7178 */
7179 unsigned long min_pages;
7180
7181 min_pages = zone->managed_pages / 1024;
7182 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7183 zone->watermark[WMARK_MIN] = min_pages;
7184 } else {
7185 /*
7186 * If it's a lowmem zone, reserve a number of pages
7187 * proportionate to the zone's size.
7188 */
7189 zone->watermark[WMARK_MIN] = tmp;
7190 }
7191
7192 /*
7193 * Set the kswapd watermarks distance according to the
7194 * scale factor in proportion to available memory, but
7195 * ensure a minimum size on small systems.
7196 */
7197 tmp = max_t(u64, tmp >> 2,
7198 mult_frac(zone->managed_pages,
7199 watermark_scale_factor, 10000));
7200
7201 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7202 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7203
7204 spin_unlock_irqrestore(&zone->lock, flags);
7205 }
7206
7207 /* update totalreserve_pages */
7208 calculate_totalreserve_pages();
7209}
7210
7211/**
7212 * setup_per_zone_wmarks - called when min_free_kbytes changes
7213 * or when memory is hot-{added|removed}
7214 *
7215 * Ensures that the watermark[min,low,high] values for each zone are set
7216 * correctly with respect to min_free_kbytes.
7217 */
7218void setup_per_zone_wmarks(void)
7219{
7220 static DEFINE_SPINLOCK(lock);
7221
7222 spin_lock(&lock);
7223 __setup_per_zone_wmarks();
7224 spin_unlock(&lock);
7225}
7226
7227/*
7228 * Initialise min_free_kbytes.
7229 *
7230 * For small machines we want it small (128k min). For large machines
7231 * we want it large (64MB max). But it is not linear, because network
7232 * bandwidth does not increase linearly with machine size. We use
7233 *
7234 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7235 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7236 *
7237 * which yields
7238 *
7239 * 16MB: 512k
7240 * 32MB: 724k
7241 * 64MB: 1024k
7242 * 128MB: 1448k
7243 * 256MB: 2048k
7244 * 512MB: 2896k
7245 * 1024MB: 4096k
7246 * 2048MB: 5792k
7247 * 4096MB: 8192k
7248 * 8192MB: 11584k
7249 * 16384MB: 16384k
7250 */
7251int __meminit init_per_zone_wmark_min(void)
7252{
7253 unsigned long lowmem_kbytes;
7254 int new_min_free_kbytes;
7255
7256 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7257 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7258
7259 if (new_min_free_kbytes > user_min_free_kbytes) {
7260 min_free_kbytes = new_min_free_kbytes;
7261 if (min_free_kbytes < 128)
7262 min_free_kbytes = 128;
7263 if (min_free_kbytes > 65536)
7264 min_free_kbytes = 65536;
7265 } else {
7266 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7267 new_min_free_kbytes, user_min_free_kbytes);
7268 }
7269 setup_per_zone_wmarks();
7270 refresh_zone_stat_thresholds();
7271 setup_per_zone_lowmem_reserve();
7272
7273#ifdef CONFIG_NUMA
7274 setup_min_unmapped_ratio();
7275 setup_min_slab_ratio();
7276#endif
7277
7278 return 0;
7279}
7280core_initcall(init_per_zone_wmark_min)
7281
7282/*
7283 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7284 * that we can call two helper functions whenever min_free_kbytes
7285 * changes.
7286 */
7287int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7288 void __user *buffer, size_t *length, loff_t *ppos)
7289{
7290 int rc;
7291
7292 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7293 if (rc)
7294 return rc;
7295
7296 if (write) {
7297 user_min_free_kbytes = min_free_kbytes;
7298 setup_per_zone_wmarks();
7299 }
7300 return 0;
7301}
7302
7303int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7304 void __user *buffer, size_t *length, loff_t *ppos)
7305{
7306 int rc;
7307
7308 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7309 if (rc)
7310 return rc;
7311
7312 if (write)
7313 setup_per_zone_wmarks();
7314
7315 return 0;
7316}
7317
7318#ifdef CONFIG_NUMA
7319static void setup_min_unmapped_ratio(void)
7320{
7321 pg_data_t *pgdat;
7322 struct zone *zone;
7323
7324 for_each_online_pgdat(pgdat)
7325 pgdat->min_unmapped_pages = 0;
7326
7327 for_each_zone(zone)
7328 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7329 sysctl_min_unmapped_ratio) / 100;
7330}
7331
7332
7333int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7334 void __user *buffer, size_t *length, loff_t *ppos)
7335{
7336 int rc;
7337
7338 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7339 if (rc)
7340 return rc;
7341
7342 setup_min_unmapped_ratio();
7343
7344 return 0;
7345}
7346
7347static void setup_min_slab_ratio(void)
7348{
7349 pg_data_t *pgdat;
7350 struct zone *zone;
7351
7352 for_each_online_pgdat(pgdat)
7353 pgdat->min_slab_pages = 0;
7354
7355 for_each_zone(zone)
7356 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7357 sysctl_min_slab_ratio) / 100;
7358}
7359
7360int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7361 void __user *buffer, size_t *length, loff_t *ppos)
7362{
7363 int rc;
7364
7365 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7366 if (rc)
7367 return rc;
7368
7369 setup_min_slab_ratio();
7370
7371 return 0;
7372}
7373#endif
7374
7375/*
7376 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7377 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7378 * whenever sysctl_lowmem_reserve_ratio changes.
7379 *
7380 * The reserve ratio obviously has absolutely no relation with the
7381 * minimum watermarks. The lowmem reserve ratio can only make sense
7382 * if in function of the boot time zone sizes.
7383 */
7384int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7385 void __user *buffer, size_t *length, loff_t *ppos)
7386{
7387 proc_dointvec_minmax(table, write, buffer, length, ppos);
7388 setup_per_zone_lowmem_reserve();
7389 return 0;
7390}
7391
7392/*
7393 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7394 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7395 * pagelist can have before it gets flushed back to buddy allocator.
7396 */
7397int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7398 void __user *buffer, size_t *length, loff_t *ppos)
7399{
7400 struct zone *zone;
7401 int old_percpu_pagelist_fraction;
7402 int ret;
7403
7404 mutex_lock(&pcp_batch_high_lock);
7405 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7406
7407 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7408 if (!write || ret < 0)
7409 goto out;
7410
7411 /* Sanity checking to avoid pcp imbalance */
7412 if (percpu_pagelist_fraction &&
7413 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7414 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7415 ret = -EINVAL;
7416 goto out;
7417 }
7418
7419 /* No change? */
7420 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7421 goto out;
7422
7423 for_each_populated_zone(zone) {
7424 unsigned int cpu;
7425
7426 for_each_possible_cpu(cpu)
7427 pageset_set_high_and_batch(zone,
7428 per_cpu_ptr(zone->pageset, cpu));
7429 }
7430out:
7431 mutex_unlock(&pcp_batch_high_lock);
7432 return ret;
7433}
7434
7435#ifdef CONFIG_NUMA
7436int hashdist = HASHDIST_DEFAULT;
7437
7438static int __init set_hashdist(char *str)
7439{
7440 if (!str)
7441 return 0;
7442 hashdist = simple_strtoul(str, &str, 0);
7443 return 1;
7444}
7445__setup("hashdist=", set_hashdist);
7446#endif
7447
7448#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7449/*
7450 * Returns the number of pages that arch has reserved but
7451 * is not known to alloc_large_system_hash().
7452 */
7453static unsigned long __init arch_reserved_kernel_pages(void)
7454{
7455 return 0;
7456}
7457#endif
7458
7459/*
7460 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7461 * machines. As memory size is increased the scale is also increased but at
7462 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7463 * quadruples the scale is increased by one, which means the size of hash table
7464 * only doubles, instead of quadrupling as well.
7465 * Because 32-bit systems cannot have large physical memory, where this scaling
7466 * makes sense, it is disabled on such platforms.
7467 */
7468#if __BITS_PER_LONG > 32
7469#define ADAPT_SCALE_BASE (64ul << 30)
7470#define ADAPT_SCALE_SHIFT 2
7471#define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7472#endif
7473
7474/*
7475 * allocate a large system hash table from bootmem
7476 * - it is assumed that the hash table must contain an exact power-of-2
7477 * quantity of entries
7478 * - limit is the number of hash buckets, not the total allocation size
7479 */
7480void *__init alloc_large_system_hash(const char *tablename,
7481 unsigned long bucketsize,
7482 unsigned long numentries,
7483 int scale,
7484 int flags,
7485 unsigned int *_hash_shift,
7486 unsigned int *_hash_mask,
7487 unsigned long low_limit,
7488 unsigned long high_limit)
7489{
7490 unsigned long long max = high_limit;
7491 unsigned long log2qty, size;
7492 void *table = NULL;
7493 gfp_t gfp_flags;
7494
7495 /* allow the kernel cmdline to have a say */
7496 if (!numentries) {
7497 /* round applicable memory size up to nearest megabyte */
7498 numentries = nr_kernel_pages;
7499 numentries -= arch_reserved_kernel_pages();
7500
7501 /* It isn't necessary when PAGE_SIZE >= 1MB */
7502 if (PAGE_SHIFT < 20)
7503 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7504
7505#if __BITS_PER_LONG > 32
7506 if (!high_limit) {
7507 unsigned long adapt;
7508
7509 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7510 adapt <<= ADAPT_SCALE_SHIFT)
7511 scale++;
7512 }
7513#endif
7514
7515 /* limit to 1 bucket per 2^scale bytes of low memory */
7516 if (scale > PAGE_SHIFT)
7517 numentries >>= (scale - PAGE_SHIFT);
7518 else
7519 numentries <<= (PAGE_SHIFT - scale);
7520
7521 /* Make sure we've got at least a 0-order allocation.. */
7522 if (unlikely(flags & HASH_SMALL)) {
7523 /* Makes no sense without HASH_EARLY */
7524 WARN_ON(!(flags & HASH_EARLY));
7525 if (!(numentries >> *_hash_shift)) {
7526 numentries = 1UL << *_hash_shift;
7527 BUG_ON(!numentries);
7528 }
7529 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7530 numentries = PAGE_SIZE / bucketsize;
7531 }
7532 numentries = roundup_pow_of_two(numentries);
7533
7534 /* limit allocation size to 1/16 total memory by default */
7535 if (max == 0) {
7536 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7537 do_div(max, bucketsize);
7538 }
7539 max = min(max, 0x80000000ULL);
7540
7541 if (numentries < low_limit)
7542 numentries = low_limit;
7543 if (numentries > max)
7544 numentries = max;
7545
7546 log2qty = ilog2(numentries);
7547
7548 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7549 do {
7550 size = bucketsize << log2qty;
7551 if (flags & HASH_EARLY) {
7552 if (flags & HASH_ZERO)
7553 table = memblock_virt_alloc_nopanic(size, 0);
7554 else
7555 table = memblock_virt_alloc_raw(size, 0);
7556 } else if (hashdist) {
7557 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7558 } else {
7559 /*
7560 * If bucketsize is not a power-of-two, we may free
7561 * some pages at the end of hash table which
7562 * alloc_pages_exact() automatically does
7563 */
7564 if (get_order(size) < MAX_ORDER) {
7565 table = alloc_pages_exact(size, gfp_flags);
7566 kmemleak_alloc(table, size, 1, gfp_flags);
7567 }
7568 }
7569 } while (!table && size > PAGE_SIZE && --log2qty);
7570
7571 if (!table)
7572 panic("Failed to allocate %s hash table\n", tablename);
7573
7574 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7575 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7576
7577 if (_hash_shift)
7578 *_hash_shift = log2qty;
7579 if (_hash_mask)
7580 *_hash_mask = (1 << log2qty) - 1;
7581
7582 return table;
7583}
7584
7585/*
7586 * This function checks whether pageblock includes unmovable pages or not.
7587 * If @count is not zero, it is okay to include less @count unmovable pages
7588 *
7589 * PageLRU check without isolation or lru_lock could race so that
7590 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7591 * check without lock_page also may miss some movable non-lru pages at
7592 * race condition. So you can't expect this function should be exact.
7593 */
7594bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7595 int migratetype,
7596 bool skip_hwpoisoned_pages)
7597{
7598 unsigned long pfn, iter, found;
7599
7600 /*
7601 * TODO we could make this much more efficient by not checking every
7602 * page in the range if we know all of them are in MOVABLE_ZONE and
7603 * that the movable zone guarantees that pages are migratable but
7604 * the later is not the case right now unfortunatelly. E.g. movablecore
7605 * can still lead to having bootmem allocations in zone_movable.
7606 */
7607
7608 /*
7609 * CMA allocations (alloc_contig_range) really need to mark isolate
7610 * CMA pageblocks even when they are not movable in fact so consider
7611 * them movable here.
7612 */
7613 if (is_migrate_cma(migratetype) &&
7614 is_migrate_cma(get_pageblock_migratetype(page)))
7615 return false;
7616
7617 pfn = page_to_pfn(page);
7618 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7619 unsigned long check = pfn + iter;
7620
7621 if (!pfn_valid_within(check))
7622 continue;
7623
7624 page = pfn_to_page(check);
7625
7626 if (PageReserved(page))
7627 goto unmovable;
7628
7629 /*
7630 * Hugepages are not in LRU lists, but they're movable.
7631 * We need not scan over tail pages bacause we don't
7632 * handle each tail page individually in migration.
7633 */
7634 if (PageHuge(page)) {
7635 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7636 continue;
7637 }
7638
7639 /*
7640 * We can't use page_count without pin a page
7641 * because another CPU can free compound page.
7642 * This check already skips compound tails of THP
7643 * because their page->_refcount is zero at all time.
7644 */
7645 if (!page_ref_count(page)) {
7646 if (PageBuddy(page))
7647 iter += (1 << page_order(page)) - 1;
7648 continue;
7649 }
7650
7651 /*
7652 * The HWPoisoned page may be not in buddy system, and
7653 * page_count() is not 0.
7654 */
7655 if (skip_hwpoisoned_pages && PageHWPoison(page))
7656 continue;
7657
7658 if (__PageMovable(page))
7659 continue;
7660
7661 if (!PageLRU(page))
7662 found++;
7663 /*
7664 * If there are RECLAIMABLE pages, we need to check
7665 * it. But now, memory offline itself doesn't call
7666 * shrink_node_slabs() and it still to be fixed.
7667 */
7668 /*
7669 * If the page is not RAM, page_count()should be 0.
7670 * we don't need more check. This is an _used_ not-movable page.
7671 *
7672 * The problematic thing here is PG_reserved pages. PG_reserved
7673 * is set to both of a memory hole page and a _used_ kernel
7674 * page at boot.
7675 */
7676 if (found > count)
7677 goto unmovable;
7678 }
7679 return false;
7680unmovable:
7681 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
7682 return true;
7683}
7684
7685bool is_pageblock_removable_nolock(struct page *page)
7686{
7687 struct zone *zone;
7688 unsigned long pfn;
7689
7690 /*
7691 * We have to be careful here because we are iterating over memory
7692 * sections which are not zone aware so we might end up outside of
7693 * the zone but still within the section.
7694 * We have to take care about the node as well. If the node is offline
7695 * its NODE_DATA will be NULL - see page_zone.
7696 */
7697 if (!node_online(page_to_nid(page)))
7698 return false;
7699
7700 zone = page_zone(page);
7701 pfn = page_to_pfn(page);
7702 if (!zone_spans_pfn(zone, pfn))
7703 return false;
7704
7705 return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, true);
7706}
7707
7708#if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7709
7710static unsigned long pfn_max_align_down(unsigned long pfn)
7711{
7712 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7713 pageblock_nr_pages) - 1);
7714}
7715
7716static unsigned long pfn_max_align_up(unsigned long pfn)
7717{
7718 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7719 pageblock_nr_pages));
7720}
7721
7722/* [start, end) must belong to a single zone. */
7723static int __alloc_contig_migrate_range(struct compact_control *cc,
7724 unsigned long start, unsigned long end)
7725{
7726 /* This function is based on compact_zone() from compaction.c. */
7727 unsigned long nr_reclaimed;
7728 unsigned long pfn = start;
7729 unsigned int tries = 0;
7730 int ret = 0;
7731
7732 migrate_prep();
7733
7734 while (pfn < end || !list_empty(&cc->migratepages)) {
7735 if (fatal_signal_pending(current)) {
7736 ret = -EINTR;
7737 break;
7738 }
7739
7740 if (list_empty(&cc->migratepages)) {
7741 cc->nr_migratepages = 0;
7742 pfn = isolate_migratepages_range(cc, pfn, end);
7743 if (!pfn) {
7744 ret = -EINTR;
7745 break;
7746 }
7747 tries = 0;
7748 } else if (++tries == 5) {
7749 ret = ret < 0 ? ret : -EBUSY;
7750 break;
7751 }
7752
7753 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7754 &cc->migratepages);
7755 cc->nr_migratepages -= nr_reclaimed;
7756
7757 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7758 NULL, 0, cc->mode, MR_CONTIG_RANGE);
7759 }
7760 if (ret < 0) {
7761 putback_movable_pages(&cc->migratepages);
7762 return ret;
7763 }
7764 return 0;
7765}
7766
7767/**
7768 * alloc_contig_range() -- tries to allocate given range of pages
7769 * @start: start PFN to allocate
7770 * @end: one-past-the-last PFN to allocate
7771 * @migratetype: migratetype of the underlaying pageblocks (either
7772 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7773 * in range must have the same migratetype and it must
7774 * be either of the two.
7775 * @gfp_mask: GFP mask to use during compaction
7776 *
7777 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7778 * aligned. The PFN range must belong to a single zone.
7779 *
7780 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
7781 * pageblocks in the range. Once isolated, the pageblocks should not
7782 * be modified by others.
7783 *
7784 * Returns zero on success or negative error code. On success all
7785 * pages which PFN is in [start, end) are allocated for the caller and
7786 * need to be freed with free_contig_range().
7787 */
7788int alloc_contig_range(unsigned long start, unsigned long end,
7789 unsigned migratetype, gfp_t gfp_mask)
7790{
7791 unsigned long outer_start, outer_end;
7792 unsigned int order;
7793 int ret = 0;
7794
7795 struct compact_control cc = {
7796 .nr_migratepages = 0,
7797 .order = -1,
7798 .zone = page_zone(pfn_to_page(start)),
7799 .mode = MIGRATE_SYNC,
7800 .ignore_skip_hint = true,
7801 .no_set_skip_hint = true,
7802 .gfp_mask = current_gfp_context(gfp_mask),
7803 };
7804 INIT_LIST_HEAD(&cc.migratepages);
7805
7806 /*
7807 * What we do here is we mark all pageblocks in range as
7808 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7809 * have different sizes, and due to the way page allocator
7810 * work, we align the range to biggest of the two pages so
7811 * that page allocator won't try to merge buddies from
7812 * different pageblocks and change MIGRATE_ISOLATE to some
7813 * other migration type.
7814 *
7815 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7816 * migrate the pages from an unaligned range (ie. pages that
7817 * we are interested in). This will put all the pages in
7818 * range back to page allocator as MIGRATE_ISOLATE.
7819 *
7820 * When this is done, we take the pages in range from page
7821 * allocator removing them from the buddy system. This way
7822 * page allocator will never consider using them.
7823 *
7824 * This lets us mark the pageblocks back as
7825 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7826 * aligned range but not in the unaligned, original range are
7827 * put back to page allocator so that buddy can use them.
7828 */
7829
7830 ret = start_isolate_page_range(pfn_max_align_down(start),
7831 pfn_max_align_up(end), migratetype,
7832 false);
7833 if (ret)
7834 return ret;
7835
7836 /*
7837 * In case of -EBUSY, we'd like to know which page causes problem.
7838 * So, just fall through. test_pages_isolated() has a tracepoint
7839 * which will report the busy page.
7840 *
7841 * It is possible that busy pages could become available before
7842 * the call to test_pages_isolated, and the range will actually be
7843 * allocated. So, if we fall through be sure to clear ret so that
7844 * -EBUSY is not accidentally used or returned to caller.
7845 */
7846 ret = __alloc_contig_migrate_range(&cc, start, end);
7847 if (ret && ret != -EBUSY)
7848 goto done;
7849 ret =0;
7850
7851 /*
7852 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7853 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7854 * more, all pages in [start, end) are free in page allocator.
7855 * What we are going to do is to allocate all pages from
7856 * [start, end) (that is remove them from page allocator).
7857 *
7858 * The only problem is that pages at the beginning and at the
7859 * end of interesting range may be not aligned with pages that
7860 * page allocator holds, ie. they can be part of higher order
7861 * pages. Because of this, we reserve the bigger range and
7862 * once this is done free the pages we are not interested in.
7863 *
7864 * We don't have to hold zone->lock here because the pages are
7865 * isolated thus they won't get removed from buddy.
7866 */
7867
7868 lru_add_drain_all();
7869 drain_all_pages(cc.zone);
7870
7871 order = 0;
7872 outer_start = start;
7873 while (!PageBuddy(pfn_to_page(outer_start))) {
7874 if (++order >= MAX_ORDER) {
7875 outer_start = start;
7876 break;
7877 }
7878 outer_start &= ~0UL << order;
7879 }
7880
7881 if (outer_start != start) {
7882 order = page_order(pfn_to_page(outer_start));
7883
7884 /*
7885 * outer_start page could be small order buddy page and
7886 * it doesn't include start page. Adjust outer_start
7887 * in this case to report failed page properly
7888 * on tracepoint in test_pages_isolated()
7889 */
7890 if (outer_start + (1UL << order) <= start)
7891 outer_start = start;
7892 }
7893
7894 /* Make sure the range is really isolated. */
7895 if (test_pages_isolated(outer_start, end, false)) {
7896 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7897 __func__, outer_start, end);
7898 ret = -EBUSY;
7899 goto done;
7900 }
7901
7902 /* Grab isolated pages from freelists. */
7903 outer_end = isolate_freepages_range(&cc, outer_start, end);
7904 if (!outer_end) {
7905 ret = -EBUSY;
7906 goto done;
7907 }
7908
7909 /* Free head and tail (if any) */
7910 if (start != outer_start)
7911 free_contig_range(outer_start, start - outer_start);
7912 if (end != outer_end)
7913 free_contig_range(end, outer_end - end);
7914
7915done:
7916 undo_isolate_page_range(pfn_max_align_down(start),
7917 pfn_max_align_up(end), migratetype);
7918 return ret;
7919}
7920
7921void free_contig_range(unsigned long pfn, unsigned nr_pages)
7922{
7923 unsigned int count = 0;
7924
7925 for (; nr_pages--; pfn++) {
7926 struct page *page = pfn_to_page(pfn);
7927
7928 count += page_count(page) != 1;
7929 __free_page(page);
7930 }
7931 WARN(count != 0, "%d pages are still in use!\n", count);
7932}
7933#endif
7934
7935#ifdef CONFIG_MEMORY_HOTPLUG
7936/*
7937 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7938 * page high values need to be recalulated.
7939 */
7940void __meminit zone_pcp_update(struct zone *zone)
7941{
7942 unsigned cpu;
7943 mutex_lock(&pcp_batch_high_lock);
7944 for_each_possible_cpu(cpu)
7945 pageset_set_high_and_batch(zone,
7946 per_cpu_ptr(zone->pageset, cpu));
7947 mutex_unlock(&pcp_batch_high_lock);
7948}
7949#endif
7950
7951void zone_pcp_reset(struct zone *zone)
7952{
7953 unsigned long flags;
7954 int cpu;
7955 struct per_cpu_pageset *pset;
7956
7957 /* avoid races with drain_pages() */
7958 local_irq_save(flags);
7959 if (zone->pageset != &boot_pageset) {
7960 for_each_online_cpu(cpu) {
7961 pset = per_cpu_ptr(zone->pageset, cpu);
7962 drain_zonestat(zone, pset);
7963 }
7964 free_percpu(zone->pageset);
7965 zone->pageset = &boot_pageset;
7966 }
7967 local_irq_restore(flags);
7968}
7969
7970#ifdef CONFIG_MEMORY_HOTREMOVE
7971/*
7972 * All pages in the range must be in a single zone and isolated
7973 * before calling this.
7974 */
7975void
7976__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7977{
7978 struct page *page;
7979 struct zone *zone;
7980 unsigned int order, i;
7981 unsigned long pfn;
7982 unsigned long flags;
7983 /* find the first valid pfn */
7984 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7985 if (pfn_valid(pfn))
7986 break;
7987 if (pfn == end_pfn)
7988 return;
7989 offline_mem_sections(pfn, end_pfn);
7990 zone = page_zone(pfn_to_page(pfn));
7991 spin_lock_irqsave(&zone->lock, flags);
7992 pfn = start_pfn;
7993 while (pfn < end_pfn) {
7994 if (!pfn_valid(pfn)) {
7995 pfn++;
7996 continue;
7997 }
7998 page = pfn_to_page(pfn);
7999 /*
8000 * The HWPoisoned page may be not in buddy system, and
8001 * page_count() is not 0.
8002 */
8003 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8004 pfn++;
8005 SetPageReserved(page);
8006 continue;
8007 }
8008
8009 BUG_ON(page_count(page));
8010 BUG_ON(!PageBuddy(page));
8011 order = page_order(page);
8012#ifdef CONFIG_DEBUG_VM
8013 pr_info("remove from free list %lx %d %lx\n",
8014 pfn, 1 << order, end_pfn);
8015#endif
8016 list_del(&page->lru);
8017 rmv_page_order(page);
8018 zone->free_area[order].nr_free--;
8019 for (i = 0; i < (1 << order); i++)
8020 SetPageReserved((page+i));
8021 pfn += (1 << order);
8022 }
8023 spin_unlock_irqrestore(&zone->lock, flags);
8024}
8025#endif
8026
8027bool is_free_buddy_page(struct page *page)
8028{
8029 struct zone *zone = page_zone(page);
8030 unsigned long pfn = page_to_pfn(page);
8031 unsigned long flags;
8032 unsigned int order;
8033
8034 spin_lock_irqsave(&zone->lock, flags);
8035 for (order = 0; order < MAX_ORDER; order++) {
8036 struct page *page_head = page - (pfn & ((1 << order) - 1));
8037
8038 if (PageBuddy(page_head) && page_order(page_head) >= order)
8039 break;
8040 }
8041 spin_unlock_irqrestore(&zone->lock, flags);
8042
8043 return order < MAX_ORDER;
8044}
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/interrupt.h>
22#include <linux/jiffies.h>
23#include <linux/compiler.h>
24#include <linux/kernel.h>
25#include <linux/kasan.h>
26#include <linux/kmsan.h>
27#include <linux/module.h>
28#include <linux/suspend.h>
29#include <linux/ratelimit.h>
30#include <linux/oom.h>
31#include <linux/topology.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/memory_hotplug.h>
36#include <linux/nodemask.h>
37#include <linux/vmstat.h>
38#include <linux/fault-inject.h>
39#include <linux/compaction.h>
40#include <trace/events/kmem.h>
41#include <trace/events/oom.h>
42#include <linux/prefetch.h>
43#include <linux/mm_inline.h>
44#include <linux/mmu_notifier.h>
45#include <linux/migrate.h>
46#include <linux/sched/mm.h>
47#include <linux/page_owner.h>
48#include <linux/page_table_check.h>
49#include <linux/memcontrol.h>
50#include <linux/ftrace.h>
51#include <linux/lockdep.h>
52#include <linux/psi.h>
53#include <linux/khugepaged.h>
54#include <linux/delayacct.h>
55#include <linux/cacheinfo.h>
56#include <asm/div64.h>
57#include "internal.h"
58#include "shuffle.h"
59#include "page_reporting.h"
60
61/* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
62typedef int __bitwise fpi_t;
63
64/* No special request */
65#define FPI_NONE ((__force fpi_t)0)
66
67/*
68 * Skip free page reporting notification for the (possibly merged) page.
69 * This does not hinder free page reporting from grabbing the page,
70 * reporting it and marking it "reported" - it only skips notifying
71 * the free page reporting infrastructure about a newly freed page. For
72 * example, used when temporarily pulling a page from a freelist and
73 * putting it back unmodified.
74 */
75#define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
76
77/*
78 * Place the (possibly merged) page to the tail of the freelist. Will ignore
79 * page shuffling (relevant code - e.g., memory onlining - is expected to
80 * shuffle the whole zone).
81 *
82 * Note: No code should rely on this flag for correctness - it's purely
83 * to allow for optimizations when handing back either fresh pages
84 * (memory onlining) or untouched pages (page isolation, free page
85 * reporting).
86 */
87#define FPI_TO_TAIL ((__force fpi_t)BIT(1))
88
89/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
90static DEFINE_MUTEX(pcp_batch_high_lock);
91#define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
92
93#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
94/*
95 * On SMP, spin_trylock is sufficient protection.
96 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
97 */
98#define pcp_trylock_prepare(flags) do { } while (0)
99#define pcp_trylock_finish(flag) do { } while (0)
100#else
101
102/* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
103#define pcp_trylock_prepare(flags) local_irq_save(flags)
104#define pcp_trylock_finish(flags) local_irq_restore(flags)
105#endif
106
107/*
108 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
109 * a migration causing the wrong PCP to be locked and remote memory being
110 * potentially allocated, pin the task to the CPU for the lookup+lock.
111 * preempt_disable is used on !RT because it is faster than migrate_disable.
112 * migrate_disable is used on RT because otherwise RT spinlock usage is
113 * interfered with and a high priority task cannot preempt the allocator.
114 */
115#ifndef CONFIG_PREEMPT_RT
116#define pcpu_task_pin() preempt_disable()
117#define pcpu_task_unpin() preempt_enable()
118#else
119#define pcpu_task_pin() migrate_disable()
120#define pcpu_task_unpin() migrate_enable()
121#endif
122
123/*
124 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
125 * Return value should be used with equivalent unlock helper.
126 */
127#define pcpu_spin_lock(type, member, ptr) \
128({ \
129 type *_ret; \
130 pcpu_task_pin(); \
131 _ret = this_cpu_ptr(ptr); \
132 spin_lock(&_ret->member); \
133 _ret; \
134})
135
136#define pcpu_spin_trylock(type, member, ptr) \
137({ \
138 type *_ret; \
139 pcpu_task_pin(); \
140 _ret = this_cpu_ptr(ptr); \
141 if (!spin_trylock(&_ret->member)) { \
142 pcpu_task_unpin(); \
143 _ret = NULL; \
144 } \
145 _ret; \
146})
147
148#define pcpu_spin_unlock(member, ptr) \
149({ \
150 spin_unlock(&ptr->member); \
151 pcpu_task_unpin(); \
152})
153
154/* struct per_cpu_pages specific helpers. */
155#define pcp_spin_lock(ptr) \
156 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
157
158#define pcp_spin_trylock(ptr) \
159 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
160
161#define pcp_spin_unlock(ptr) \
162 pcpu_spin_unlock(lock, ptr)
163
164#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
165DEFINE_PER_CPU(int, numa_node);
166EXPORT_PER_CPU_SYMBOL(numa_node);
167#endif
168
169DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
170
171#ifdef CONFIG_HAVE_MEMORYLESS_NODES
172/*
173 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
174 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
175 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
176 * defined in <linux/topology.h>.
177 */
178DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
179EXPORT_PER_CPU_SYMBOL(_numa_mem_);
180#endif
181
182static DEFINE_MUTEX(pcpu_drain_mutex);
183
184#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
185volatile unsigned long latent_entropy __latent_entropy;
186EXPORT_SYMBOL(latent_entropy);
187#endif
188
189/*
190 * Array of node states.
191 */
192nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
193 [N_POSSIBLE] = NODE_MASK_ALL,
194 [N_ONLINE] = { { [0] = 1UL } },
195#ifndef CONFIG_NUMA
196 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
197#ifdef CONFIG_HIGHMEM
198 [N_HIGH_MEMORY] = { { [0] = 1UL } },
199#endif
200 [N_MEMORY] = { { [0] = 1UL } },
201 [N_CPU] = { { [0] = 1UL } },
202#endif /* NUMA */
203};
204EXPORT_SYMBOL(node_states);
205
206gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
207
208/*
209 * A cached value of the page's pageblock's migratetype, used when the page is
210 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
211 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
212 * Also the migratetype set in the page does not necessarily match the pcplist
213 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
214 * other index - this ensures that it will be put on the correct CMA freelist.
215 */
216static inline int get_pcppage_migratetype(struct page *page)
217{
218 return page->index;
219}
220
221static inline void set_pcppage_migratetype(struct page *page, int migratetype)
222{
223 page->index = migratetype;
224}
225
226#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
227unsigned int pageblock_order __read_mostly;
228#endif
229
230static void __free_pages_ok(struct page *page, unsigned int order,
231 fpi_t fpi_flags);
232
233/*
234 * results with 256, 32 in the lowmem_reserve sysctl:
235 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
236 * 1G machine -> (16M dma, 784M normal, 224M high)
237 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
238 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
239 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
240 *
241 * TBD: should special case ZONE_DMA32 machines here - in those we normally
242 * don't need any ZONE_NORMAL reservation
243 */
244static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
245#ifdef CONFIG_ZONE_DMA
246 [ZONE_DMA] = 256,
247#endif
248#ifdef CONFIG_ZONE_DMA32
249 [ZONE_DMA32] = 256,
250#endif
251 [ZONE_NORMAL] = 32,
252#ifdef CONFIG_HIGHMEM
253 [ZONE_HIGHMEM] = 0,
254#endif
255 [ZONE_MOVABLE] = 0,
256};
257
258char * const zone_names[MAX_NR_ZONES] = {
259#ifdef CONFIG_ZONE_DMA
260 "DMA",
261#endif
262#ifdef CONFIG_ZONE_DMA32
263 "DMA32",
264#endif
265 "Normal",
266#ifdef CONFIG_HIGHMEM
267 "HighMem",
268#endif
269 "Movable",
270#ifdef CONFIG_ZONE_DEVICE
271 "Device",
272#endif
273};
274
275const char * const migratetype_names[MIGRATE_TYPES] = {
276 "Unmovable",
277 "Movable",
278 "Reclaimable",
279 "HighAtomic",
280#ifdef CONFIG_CMA
281 "CMA",
282#endif
283#ifdef CONFIG_MEMORY_ISOLATION
284 "Isolate",
285#endif
286};
287
288int min_free_kbytes = 1024;
289int user_min_free_kbytes = -1;
290static int watermark_boost_factor __read_mostly = 15000;
291static int watermark_scale_factor = 10;
292
293/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
294int movable_zone;
295EXPORT_SYMBOL(movable_zone);
296
297#if MAX_NUMNODES > 1
298unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
299unsigned int nr_online_nodes __read_mostly = 1;
300EXPORT_SYMBOL(nr_node_ids);
301EXPORT_SYMBOL(nr_online_nodes);
302#endif
303
304static bool page_contains_unaccepted(struct page *page, unsigned int order);
305static void accept_page(struct page *page, unsigned int order);
306static bool try_to_accept_memory(struct zone *zone, unsigned int order);
307static inline bool has_unaccepted_memory(void);
308static bool __free_unaccepted(struct page *page);
309
310int page_group_by_mobility_disabled __read_mostly;
311
312#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
313/*
314 * During boot we initialize deferred pages on-demand, as needed, but once
315 * page_alloc_init_late() has finished, the deferred pages are all initialized,
316 * and we can permanently disable that path.
317 */
318DEFINE_STATIC_KEY_TRUE(deferred_pages);
319
320static inline bool deferred_pages_enabled(void)
321{
322 return static_branch_unlikely(&deferred_pages);
323}
324
325/*
326 * deferred_grow_zone() is __init, but it is called from
327 * get_page_from_freelist() during early boot until deferred_pages permanently
328 * disables this call. This is why we have refdata wrapper to avoid warning,
329 * and to ensure that the function body gets unloaded.
330 */
331static bool __ref
332_deferred_grow_zone(struct zone *zone, unsigned int order)
333{
334 return deferred_grow_zone(zone, order);
335}
336#else
337static inline bool deferred_pages_enabled(void)
338{
339 return false;
340}
341#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
342
343/* Return a pointer to the bitmap storing bits affecting a block of pages */
344static inline unsigned long *get_pageblock_bitmap(const struct page *page,
345 unsigned long pfn)
346{
347#ifdef CONFIG_SPARSEMEM
348 return section_to_usemap(__pfn_to_section(pfn));
349#else
350 return page_zone(page)->pageblock_flags;
351#endif /* CONFIG_SPARSEMEM */
352}
353
354static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
355{
356#ifdef CONFIG_SPARSEMEM
357 pfn &= (PAGES_PER_SECTION-1);
358#else
359 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
360#endif /* CONFIG_SPARSEMEM */
361 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
362}
363
364/**
365 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
366 * @page: The page within the block of interest
367 * @pfn: The target page frame number
368 * @mask: mask of bits that the caller is interested in
369 *
370 * Return: pageblock_bits flags
371 */
372unsigned long get_pfnblock_flags_mask(const struct page *page,
373 unsigned long pfn, unsigned long mask)
374{
375 unsigned long *bitmap;
376 unsigned long bitidx, word_bitidx;
377 unsigned long word;
378
379 bitmap = get_pageblock_bitmap(page, pfn);
380 bitidx = pfn_to_bitidx(page, pfn);
381 word_bitidx = bitidx / BITS_PER_LONG;
382 bitidx &= (BITS_PER_LONG-1);
383 /*
384 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
385 * a consistent read of the memory array, so that results, even though
386 * racy, are not corrupted.
387 */
388 word = READ_ONCE(bitmap[word_bitidx]);
389 return (word >> bitidx) & mask;
390}
391
392static __always_inline int get_pfnblock_migratetype(const struct page *page,
393 unsigned long pfn)
394{
395 return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
396}
397
398/**
399 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
400 * @page: The page within the block of interest
401 * @flags: The flags to set
402 * @pfn: The target page frame number
403 * @mask: mask of bits that the caller is interested in
404 */
405void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
406 unsigned long pfn,
407 unsigned long mask)
408{
409 unsigned long *bitmap;
410 unsigned long bitidx, word_bitidx;
411 unsigned long word;
412
413 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
414 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
415
416 bitmap = get_pageblock_bitmap(page, pfn);
417 bitidx = pfn_to_bitidx(page, pfn);
418 word_bitidx = bitidx / BITS_PER_LONG;
419 bitidx &= (BITS_PER_LONG-1);
420
421 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
422
423 mask <<= bitidx;
424 flags <<= bitidx;
425
426 word = READ_ONCE(bitmap[word_bitidx]);
427 do {
428 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
429}
430
431void set_pageblock_migratetype(struct page *page, int migratetype)
432{
433 if (unlikely(page_group_by_mobility_disabled &&
434 migratetype < MIGRATE_PCPTYPES))
435 migratetype = MIGRATE_UNMOVABLE;
436
437 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
438 page_to_pfn(page), MIGRATETYPE_MASK);
439}
440
441#ifdef CONFIG_DEBUG_VM
442static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
443{
444 int ret;
445 unsigned seq;
446 unsigned long pfn = page_to_pfn(page);
447 unsigned long sp, start_pfn;
448
449 do {
450 seq = zone_span_seqbegin(zone);
451 start_pfn = zone->zone_start_pfn;
452 sp = zone->spanned_pages;
453 ret = !zone_spans_pfn(zone, pfn);
454 } while (zone_span_seqretry(zone, seq));
455
456 if (ret)
457 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
458 pfn, zone_to_nid(zone), zone->name,
459 start_pfn, start_pfn + sp);
460
461 return ret;
462}
463
464/*
465 * Temporary debugging check for pages not lying within a given zone.
466 */
467static int __maybe_unused bad_range(struct zone *zone, struct page *page)
468{
469 if (page_outside_zone_boundaries(zone, page))
470 return 1;
471 if (zone != page_zone(page))
472 return 1;
473
474 return 0;
475}
476#else
477static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
478{
479 return 0;
480}
481#endif
482
483static void bad_page(struct page *page, const char *reason)
484{
485 static unsigned long resume;
486 static unsigned long nr_shown;
487 static unsigned long nr_unshown;
488
489 /*
490 * Allow a burst of 60 reports, then keep quiet for that minute;
491 * or allow a steady drip of one report per second.
492 */
493 if (nr_shown == 60) {
494 if (time_before(jiffies, resume)) {
495 nr_unshown++;
496 goto out;
497 }
498 if (nr_unshown) {
499 pr_alert(
500 "BUG: Bad page state: %lu messages suppressed\n",
501 nr_unshown);
502 nr_unshown = 0;
503 }
504 nr_shown = 0;
505 }
506 if (nr_shown++ == 0)
507 resume = jiffies + 60 * HZ;
508
509 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
510 current->comm, page_to_pfn(page));
511 dump_page(page, reason);
512
513 print_modules();
514 dump_stack();
515out:
516 /* Leave bad fields for debug, except PageBuddy could make trouble */
517 page_mapcount_reset(page); /* remove PageBuddy */
518 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
519}
520
521static inline unsigned int order_to_pindex(int migratetype, int order)
522{
523#ifdef CONFIG_TRANSPARENT_HUGEPAGE
524 if (order > PAGE_ALLOC_COSTLY_ORDER) {
525 VM_BUG_ON(order != pageblock_order);
526 return NR_LOWORDER_PCP_LISTS;
527 }
528#else
529 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
530#endif
531
532 return (MIGRATE_PCPTYPES * order) + migratetype;
533}
534
535static inline int pindex_to_order(unsigned int pindex)
536{
537 int order = pindex / MIGRATE_PCPTYPES;
538
539#ifdef CONFIG_TRANSPARENT_HUGEPAGE
540 if (pindex == NR_LOWORDER_PCP_LISTS)
541 order = pageblock_order;
542#else
543 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
544#endif
545
546 return order;
547}
548
549static inline bool pcp_allowed_order(unsigned int order)
550{
551 if (order <= PAGE_ALLOC_COSTLY_ORDER)
552 return true;
553#ifdef CONFIG_TRANSPARENT_HUGEPAGE
554 if (order == pageblock_order)
555 return true;
556#endif
557 return false;
558}
559
560static inline void free_the_page(struct page *page, unsigned int order)
561{
562 if (pcp_allowed_order(order)) /* Via pcp? */
563 free_unref_page(page, order);
564 else
565 __free_pages_ok(page, order, FPI_NONE);
566}
567
568/*
569 * Higher-order pages are called "compound pages". They are structured thusly:
570 *
571 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
572 *
573 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
574 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
575 *
576 * The first tail page's ->compound_order holds the order of allocation.
577 * This usage means that zero-order pages may not be compound.
578 */
579
580void prep_compound_page(struct page *page, unsigned int order)
581{
582 int i;
583 int nr_pages = 1 << order;
584
585 __SetPageHead(page);
586 for (i = 1; i < nr_pages; i++)
587 prep_compound_tail(page, i);
588
589 prep_compound_head(page, order);
590}
591
592void destroy_large_folio(struct folio *folio)
593{
594 if (folio_test_hugetlb(folio)) {
595 free_huge_folio(folio);
596 return;
597 }
598
599 if (folio_test_large_rmappable(folio))
600 folio_undo_large_rmappable(folio);
601
602 mem_cgroup_uncharge(folio);
603 free_the_page(&folio->page, folio_order(folio));
604}
605
606static inline void set_buddy_order(struct page *page, unsigned int order)
607{
608 set_page_private(page, order);
609 __SetPageBuddy(page);
610}
611
612#ifdef CONFIG_COMPACTION
613static inline struct capture_control *task_capc(struct zone *zone)
614{
615 struct capture_control *capc = current->capture_control;
616
617 return unlikely(capc) &&
618 !(current->flags & PF_KTHREAD) &&
619 !capc->page &&
620 capc->cc->zone == zone ? capc : NULL;
621}
622
623static inline bool
624compaction_capture(struct capture_control *capc, struct page *page,
625 int order, int migratetype)
626{
627 if (!capc || order != capc->cc->order)
628 return false;
629
630 /* Do not accidentally pollute CMA or isolated regions*/
631 if (is_migrate_cma(migratetype) ||
632 is_migrate_isolate(migratetype))
633 return false;
634
635 /*
636 * Do not let lower order allocations pollute a movable pageblock.
637 * This might let an unmovable request use a reclaimable pageblock
638 * and vice-versa but no more than normal fallback logic which can
639 * have trouble finding a high-order free page.
640 */
641 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
642 return false;
643
644 capc->page = page;
645 return true;
646}
647
648#else
649static inline struct capture_control *task_capc(struct zone *zone)
650{
651 return NULL;
652}
653
654static inline bool
655compaction_capture(struct capture_control *capc, struct page *page,
656 int order, int migratetype)
657{
658 return false;
659}
660#endif /* CONFIG_COMPACTION */
661
662/* Used for pages not on another list */
663static inline void add_to_free_list(struct page *page, struct zone *zone,
664 unsigned int order, int migratetype)
665{
666 struct free_area *area = &zone->free_area[order];
667
668 list_add(&page->buddy_list, &area->free_list[migratetype]);
669 area->nr_free++;
670}
671
672/* Used for pages not on another list */
673static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
674 unsigned int order, int migratetype)
675{
676 struct free_area *area = &zone->free_area[order];
677
678 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
679 area->nr_free++;
680}
681
682/*
683 * Used for pages which are on another list. Move the pages to the tail
684 * of the list - so the moved pages won't immediately be considered for
685 * allocation again (e.g., optimization for memory onlining).
686 */
687static inline void move_to_free_list(struct page *page, struct zone *zone,
688 unsigned int order, int migratetype)
689{
690 struct free_area *area = &zone->free_area[order];
691
692 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
693}
694
695static inline void del_page_from_free_list(struct page *page, struct zone *zone,
696 unsigned int order)
697{
698 /* clear reported state and update reported page count */
699 if (page_reported(page))
700 __ClearPageReported(page);
701
702 list_del(&page->buddy_list);
703 __ClearPageBuddy(page);
704 set_page_private(page, 0);
705 zone->free_area[order].nr_free--;
706}
707
708static inline struct page *get_page_from_free_area(struct free_area *area,
709 int migratetype)
710{
711 return list_first_entry_or_null(&area->free_list[migratetype],
712 struct page, buddy_list);
713}
714
715/*
716 * If this is not the largest possible page, check if the buddy
717 * of the next-highest order is free. If it is, it's possible
718 * that pages are being freed that will coalesce soon. In case,
719 * that is happening, add the free page to the tail of the list
720 * so it's less likely to be used soon and more likely to be merged
721 * as a higher order page
722 */
723static inline bool
724buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
725 struct page *page, unsigned int order)
726{
727 unsigned long higher_page_pfn;
728 struct page *higher_page;
729
730 if (order >= MAX_PAGE_ORDER - 1)
731 return false;
732
733 higher_page_pfn = buddy_pfn & pfn;
734 higher_page = page + (higher_page_pfn - pfn);
735
736 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
737 NULL) != NULL;
738}
739
740/*
741 * Freeing function for a buddy system allocator.
742 *
743 * The concept of a buddy system is to maintain direct-mapped table
744 * (containing bit values) for memory blocks of various "orders".
745 * The bottom level table contains the map for the smallest allocatable
746 * units of memory (here, pages), and each level above it describes
747 * pairs of units from the levels below, hence, "buddies".
748 * At a high level, all that happens here is marking the table entry
749 * at the bottom level available, and propagating the changes upward
750 * as necessary, plus some accounting needed to play nicely with other
751 * parts of the VM system.
752 * At each level, we keep a list of pages, which are heads of continuous
753 * free pages of length of (1 << order) and marked with PageBuddy.
754 * Page's order is recorded in page_private(page) field.
755 * So when we are allocating or freeing one, we can derive the state of the
756 * other. That is, if we allocate a small block, and both were
757 * free, the remainder of the region must be split into blocks.
758 * If a block is freed, and its buddy is also free, then this
759 * triggers coalescing into a block of larger size.
760 *
761 * -- nyc
762 */
763
764static inline void __free_one_page(struct page *page,
765 unsigned long pfn,
766 struct zone *zone, unsigned int order,
767 int migratetype, fpi_t fpi_flags)
768{
769 struct capture_control *capc = task_capc(zone);
770 unsigned long buddy_pfn = 0;
771 unsigned long combined_pfn;
772 struct page *buddy;
773 bool to_tail;
774
775 VM_BUG_ON(!zone_is_initialized(zone));
776 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
777
778 VM_BUG_ON(migratetype == -1);
779 if (likely(!is_migrate_isolate(migratetype)))
780 __mod_zone_freepage_state(zone, 1 << order, migratetype);
781
782 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
783 VM_BUG_ON_PAGE(bad_range(zone, page), page);
784
785 while (order < MAX_PAGE_ORDER) {
786 if (compaction_capture(capc, page, order, migratetype)) {
787 __mod_zone_freepage_state(zone, -(1 << order),
788 migratetype);
789 return;
790 }
791
792 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
793 if (!buddy)
794 goto done_merging;
795
796 if (unlikely(order >= pageblock_order)) {
797 /*
798 * We want to prevent merge between freepages on pageblock
799 * without fallbacks and normal pageblock. Without this,
800 * pageblock isolation could cause incorrect freepage or CMA
801 * accounting or HIGHATOMIC accounting.
802 */
803 int buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
804
805 if (migratetype != buddy_mt
806 && (!migratetype_is_mergeable(migratetype) ||
807 !migratetype_is_mergeable(buddy_mt)))
808 goto done_merging;
809 }
810
811 /*
812 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
813 * merge with it and move up one order.
814 */
815 if (page_is_guard(buddy))
816 clear_page_guard(zone, buddy, order, migratetype);
817 else
818 del_page_from_free_list(buddy, zone, order);
819 combined_pfn = buddy_pfn & pfn;
820 page = page + (combined_pfn - pfn);
821 pfn = combined_pfn;
822 order++;
823 }
824
825done_merging:
826 set_buddy_order(page, order);
827
828 if (fpi_flags & FPI_TO_TAIL)
829 to_tail = true;
830 else if (is_shuffle_order(order))
831 to_tail = shuffle_pick_tail();
832 else
833 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
834
835 if (to_tail)
836 add_to_free_list_tail(page, zone, order, migratetype);
837 else
838 add_to_free_list(page, zone, order, migratetype);
839
840 /* Notify page reporting subsystem of freed page */
841 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
842 page_reporting_notify_free(order);
843}
844
845/**
846 * split_free_page() -- split a free page at split_pfn_offset
847 * @free_page: the original free page
848 * @order: the order of the page
849 * @split_pfn_offset: split offset within the page
850 *
851 * Return -ENOENT if the free page is changed, otherwise 0
852 *
853 * It is used when the free page crosses two pageblocks with different migratetypes
854 * at split_pfn_offset within the page. The split free page will be put into
855 * separate migratetype lists afterwards. Otherwise, the function achieves
856 * nothing.
857 */
858int split_free_page(struct page *free_page,
859 unsigned int order, unsigned long split_pfn_offset)
860{
861 struct zone *zone = page_zone(free_page);
862 unsigned long free_page_pfn = page_to_pfn(free_page);
863 unsigned long pfn;
864 unsigned long flags;
865 int free_page_order;
866 int mt;
867 int ret = 0;
868
869 if (split_pfn_offset == 0)
870 return ret;
871
872 spin_lock_irqsave(&zone->lock, flags);
873
874 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
875 ret = -ENOENT;
876 goto out;
877 }
878
879 mt = get_pfnblock_migratetype(free_page, free_page_pfn);
880 if (likely(!is_migrate_isolate(mt)))
881 __mod_zone_freepage_state(zone, -(1UL << order), mt);
882
883 del_page_from_free_list(free_page, zone, order);
884 for (pfn = free_page_pfn;
885 pfn < free_page_pfn + (1UL << order);) {
886 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
887
888 free_page_order = min_t(unsigned int,
889 pfn ? __ffs(pfn) : order,
890 __fls(split_pfn_offset));
891 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
892 mt, FPI_NONE);
893 pfn += 1UL << free_page_order;
894 split_pfn_offset -= (1UL << free_page_order);
895 /* we have done the first part, now switch to second part */
896 if (split_pfn_offset == 0)
897 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
898 }
899out:
900 spin_unlock_irqrestore(&zone->lock, flags);
901 return ret;
902}
903/*
904 * A bad page could be due to a number of fields. Instead of multiple branches,
905 * try and check multiple fields with one check. The caller must do a detailed
906 * check if necessary.
907 */
908static inline bool page_expected_state(struct page *page,
909 unsigned long check_flags)
910{
911 if (unlikely(atomic_read(&page->_mapcount) != -1))
912 return false;
913
914 if (unlikely((unsigned long)page->mapping |
915 page_ref_count(page) |
916#ifdef CONFIG_MEMCG
917 page->memcg_data |
918#endif
919#ifdef CONFIG_PAGE_POOL
920 ((page->pp_magic & ~0x3UL) == PP_SIGNATURE) |
921#endif
922 (page->flags & check_flags)))
923 return false;
924
925 return true;
926}
927
928static const char *page_bad_reason(struct page *page, unsigned long flags)
929{
930 const char *bad_reason = NULL;
931
932 if (unlikely(atomic_read(&page->_mapcount) != -1))
933 bad_reason = "nonzero mapcount";
934 if (unlikely(page->mapping != NULL))
935 bad_reason = "non-NULL mapping";
936 if (unlikely(page_ref_count(page) != 0))
937 bad_reason = "nonzero _refcount";
938 if (unlikely(page->flags & flags)) {
939 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
940 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
941 else
942 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
943 }
944#ifdef CONFIG_MEMCG
945 if (unlikely(page->memcg_data))
946 bad_reason = "page still charged to cgroup";
947#endif
948#ifdef CONFIG_PAGE_POOL
949 if (unlikely((page->pp_magic & ~0x3UL) == PP_SIGNATURE))
950 bad_reason = "page_pool leak";
951#endif
952 return bad_reason;
953}
954
955static void free_page_is_bad_report(struct page *page)
956{
957 bad_page(page,
958 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
959}
960
961static inline bool free_page_is_bad(struct page *page)
962{
963 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
964 return false;
965
966 /* Something has gone sideways, find it */
967 free_page_is_bad_report(page);
968 return true;
969}
970
971static inline bool is_check_pages_enabled(void)
972{
973 return static_branch_unlikely(&check_pages_enabled);
974}
975
976static int free_tail_page_prepare(struct page *head_page, struct page *page)
977{
978 struct folio *folio = (struct folio *)head_page;
979 int ret = 1;
980
981 /*
982 * We rely page->lru.next never has bit 0 set, unless the page
983 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
984 */
985 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
986
987 if (!is_check_pages_enabled()) {
988 ret = 0;
989 goto out;
990 }
991 switch (page - head_page) {
992 case 1:
993 /* the first tail page: these may be in place of ->mapping */
994 if (unlikely(folio_entire_mapcount(folio))) {
995 bad_page(page, "nonzero entire_mapcount");
996 goto out;
997 }
998 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
999 bad_page(page, "nonzero nr_pages_mapped");
1000 goto out;
1001 }
1002 if (unlikely(atomic_read(&folio->_pincount))) {
1003 bad_page(page, "nonzero pincount");
1004 goto out;
1005 }
1006 break;
1007 case 2:
1008 /*
1009 * the second tail page: ->mapping is
1010 * deferred_list.next -- ignore value.
1011 */
1012 break;
1013 default:
1014 if (page->mapping != TAIL_MAPPING) {
1015 bad_page(page, "corrupted mapping in tail page");
1016 goto out;
1017 }
1018 break;
1019 }
1020 if (unlikely(!PageTail(page))) {
1021 bad_page(page, "PageTail not set");
1022 goto out;
1023 }
1024 if (unlikely(compound_head(page) != head_page)) {
1025 bad_page(page, "compound_head not consistent");
1026 goto out;
1027 }
1028 ret = 0;
1029out:
1030 page->mapping = NULL;
1031 clear_compound_head(page);
1032 return ret;
1033}
1034
1035/*
1036 * Skip KASAN memory poisoning when either:
1037 *
1038 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1039 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1040 * using page tags instead (see below).
1041 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1042 * that error detection is disabled for accesses via the page address.
1043 *
1044 * Pages will have match-all tags in the following circumstances:
1045 *
1046 * 1. Pages are being initialized for the first time, including during deferred
1047 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1048 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1049 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1050 * 3. The allocation was excluded from being checked due to sampling,
1051 * see the call to kasan_unpoison_pages.
1052 *
1053 * Poisoning pages during deferred memory init will greatly lengthen the
1054 * process and cause problem in large memory systems as the deferred pages
1055 * initialization is done with interrupt disabled.
1056 *
1057 * Assuming that there will be no reference to those newly initialized
1058 * pages before they are ever allocated, this should have no effect on
1059 * KASAN memory tracking as the poison will be properly inserted at page
1060 * allocation time. The only corner case is when pages are allocated by
1061 * on-demand allocation and then freed again before the deferred pages
1062 * initialization is done, but this is not likely to happen.
1063 */
1064static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1065{
1066 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1067 return deferred_pages_enabled();
1068
1069 return page_kasan_tag(page) == KASAN_TAG_KERNEL;
1070}
1071
1072static void kernel_init_pages(struct page *page, int numpages)
1073{
1074 int i;
1075
1076 /* s390's use of memset() could override KASAN redzones. */
1077 kasan_disable_current();
1078 for (i = 0; i < numpages; i++)
1079 clear_highpage_kasan_tagged(page + i);
1080 kasan_enable_current();
1081}
1082
1083static __always_inline bool free_pages_prepare(struct page *page,
1084 unsigned int order, fpi_t fpi_flags)
1085{
1086 int bad = 0;
1087 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1088 bool init = want_init_on_free();
1089 bool compound = PageCompound(page);
1090
1091 VM_BUG_ON_PAGE(PageTail(page), page);
1092
1093 trace_mm_page_free(page, order);
1094 kmsan_free_page(page, order);
1095
1096 if (memcg_kmem_online() && PageMemcgKmem(page))
1097 __memcg_kmem_uncharge_page(page, order);
1098
1099 if (unlikely(PageHWPoison(page)) && !order) {
1100 /* Do not let hwpoison pages hit pcplists/buddy */
1101 reset_page_owner(page, order);
1102 page_table_check_free(page, order);
1103 return false;
1104 }
1105
1106 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1107
1108 /*
1109 * Check tail pages before head page information is cleared to
1110 * avoid checking PageCompound for order-0 pages.
1111 */
1112 if (unlikely(order)) {
1113 int i;
1114
1115 if (compound)
1116 page[1].flags &= ~PAGE_FLAGS_SECOND;
1117 for (i = 1; i < (1 << order); i++) {
1118 if (compound)
1119 bad += free_tail_page_prepare(page, page + i);
1120 if (is_check_pages_enabled()) {
1121 if (free_page_is_bad(page + i)) {
1122 bad++;
1123 continue;
1124 }
1125 }
1126 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1127 }
1128 }
1129 if (PageMappingFlags(page))
1130 page->mapping = NULL;
1131 if (is_check_pages_enabled()) {
1132 if (free_page_is_bad(page))
1133 bad++;
1134 if (bad)
1135 return false;
1136 }
1137
1138 page_cpupid_reset_last(page);
1139 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1140 reset_page_owner(page, order);
1141 page_table_check_free(page, order);
1142
1143 if (!PageHighMem(page)) {
1144 debug_check_no_locks_freed(page_address(page),
1145 PAGE_SIZE << order);
1146 debug_check_no_obj_freed(page_address(page),
1147 PAGE_SIZE << order);
1148 }
1149
1150 kernel_poison_pages(page, 1 << order);
1151
1152 /*
1153 * As memory initialization might be integrated into KASAN,
1154 * KASAN poisoning and memory initialization code must be
1155 * kept together to avoid discrepancies in behavior.
1156 *
1157 * With hardware tag-based KASAN, memory tags must be set before the
1158 * page becomes unavailable via debug_pagealloc or arch_free_page.
1159 */
1160 if (!skip_kasan_poison) {
1161 kasan_poison_pages(page, order, init);
1162
1163 /* Memory is already initialized if KASAN did it internally. */
1164 if (kasan_has_integrated_init())
1165 init = false;
1166 }
1167 if (init)
1168 kernel_init_pages(page, 1 << order);
1169
1170 /*
1171 * arch_free_page() can make the page's contents inaccessible. s390
1172 * does this. So nothing which can access the page's contents should
1173 * happen after this.
1174 */
1175 arch_free_page(page, order);
1176
1177 debug_pagealloc_unmap_pages(page, 1 << order);
1178
1179 return true;
1180}
1181
1182/*
1183 * Frees a number of pages from the PCP lists
1184 * Assumes all pages on list are in same zone.
1185 * count is the number of pages to free.
1186 */
1187static void free_pcppages_bulk(struct zone *zone, int count,
1188 struct per_cpu_pages *pcp,
1189 int pindex)
1190{
1191 unsigned long flags;
1192 unsigned int order;
1193 bool isolated_pageblocks;
1194 struct page *page;
1195
1196 /*
1197 * Ensure proper count is passed which otherwise would stuck in the
1198 * below while (list_empty(list)) loop.
1199 */
1200 count = min(pcp->count, count);
1201
1202 /* Ensure requested pindex is drained first. */
1203 pindex = pindex - 1;
1204
1205 spin_lock_irqsave(&zone->lock, flags);
1206 isolated_pageblocks = has_isolate_pageblock(zone);
1207
1208 while (count > 0) {
1209 struct list_head *list;
1210 int nr_pages;
1211
1212 /* Remove pages from lists in a round-robin fashion. */
1213 do {
1214 if (++pindex > NR_PCP_LISTS - 1)
1215 pindex = 0;
1216 list = &pcp->lists[pindex];
1217 } while (list_empty(list));
1218
1219 order = pindex_to_order(pindex);
1220 nr_pages = 1 << order;
1221 do {
1222 int mt;
1223
1224 page = list_last_entry(list, struct page, pcp_list);
1225 mt = get_pcppage_migratetype(page);
1226
1227 /* must delete to avoid corrupting pcp list */
1228 list_del(&page->pcp_list);
1229 count -= nr_pages;
1230 pcp->count -= nr_pages;
1231
1232 /* MIGRATE_ISOLATE page should not go to pcplists */
1233 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1234 /* Pageblock could have been isolated meanwhile */
1235 if (unlikely(isolated_pageblocks))
1236 mt = get_pageblock_migratetype(page);
1237
1238 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1239 trace_mm_page_pcpu_drain(page, order, mt);
1240 } while (count > 0 && !list_empty(list));
1241 }
1242
1243 spin_unlock_irqrestore(&zone->lock, flags);
1244}
1245
1246static void free_one_page(struct zone *zone,
1247 struct page *page, unsigned long pfn,
1248 unsigned int order,
1249 int migratetype, fpi_t fpi_flags)
1250{
1251 unsigned long flags;
1252
1253 spin_lock_irqsave(&zone->lock, flags);
1254 if (unlikely(has_isolate_pageblock(zone) ||
1255 is_migrate_isolate(migratetype))) {
1256 migratetype = get_pfnblock_migratetype(page, pfn);
1257 }
1258 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1259 spin_unlock_irqrestore(&zone->lock, flags);
1260}
1261
1262static void __free_pages_ok(struct page *page, unsigned int order,
1263 fpi_t fpi_flags)
1264{
1265 int migratetype;
1266 unsigned long pfn = page_to_pfn(page);
1267 struct zone *zone = page_zone(page);
1268
1269 if (!free_pages_prepare(page, order, fpi_flags))
1270 return;
1271
1272 /*
1273 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1274 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1275 * This will reduce the lock holding time.
1276 */
1277 migratetype = get_pfnblock_migratetype(page, pfn);
1278
1279 free_one_page(zone, page, pfn, order, migratetype, fpi_flags);
1280
1281 __count_vm_events(PGFREE, 1 << order);
1282}
1283
1284void __free_pages_core(struct page *page, unsigned int order)
1285{
1286 unsigned int nr_pages = 1 << order;
1287 struct page *p = page;
1288 unsigned int loop;
1289
1290 /*
1291 * When initializing the memmap, __init_single_page() sets the refcount
1292 * of all pages to 1 ("allocated"/"not free"). We have to set the
1293 * refcount of all involved pages to 0.
1294 */
1295 prefetchw(p);
1296 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1297 prefetchw(p + 1);
1298 __ClearPageReserved(p);
1299 set_page_count(p, 0);
1300 }
1301 __ClearPageReserved(p);
1302 set_page_count(p, 0);
1303
1304 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1305
1306 if (page_contains_unaccepted(page, order)) {
1307 if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1308 return;
1309
1310 accept_page(page, order);
1311 }
1312
1313 /*
1314 * Bypass PCP and place fresh pages right to the tail, primarily
1315 * relevant for memory onlining.
1316 */
1317 __free_pages_ok(page, order, FPI_TO_TAIL);
1318}
1319
1320/*
1321 * Check that the whole (or subset of) a pageblock given by the interval of
1322 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1323 * with the migration of free compaction scanner.
1324 *
1325 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1326 *
1327 * It's possible on some configurations to have a setup like node0 node1 node0
1328 * i.e. it's possible that all pages within a zones range of pages do not
1329 * belong to a single zone. We assume that a border between node0 and node1
1330 * can occur within a single pageblock, but not a node0 node1 node0
1331 * interleaving within a single pageblock. It is therefore sufficient to check
1332 * the first and last page of a pageblock and avoid checking each individual
1333 * page in a pageblock.
1334 *
1335 * Note: the function may return non-NULL struct page even for a page block
1336 * which contains a memory hole (i.e. there is no physical memory for a subset
1337 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1338 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1339 * even though the start pfn is online and valid. This should be safe most of
1340 * the time because struct pages are still initialized via init_unavailable_range()
1341 * and pfn walkers shouldn't touch any physical memory range for which they do
1342 * not recognize any specific metadata in struct pages.
1343 */
1344struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1345 unsigned long end_pfn, struct zone *zone)
1346{
1347 struct page *start_page;
1348 struct page *end_page;
1349
1350 /* end_pfn is one past the range we are checking */
1351 end_pfn--;
1352
1353 if (!pfn_valid(end_pfn))
1354 return NULL;
1355
1356 start_page = pfn_to_online_page(start_pfn);
1357 if (!start_page)
1358 return NULL;
1359
1360 if (page_zone(start_page) != zone)
1361 return NULL;
1362
1363 end_page = pfn_to_page(end_pfn);
1364
1365 /* This gives a shorter code than deriving page_zone(end_page) */
1366 if (page_zone_id(start_page) != page_zone_id(end_page))
1367 return NULL;
1368
1369 return start_page;
1370}
1371
1372/*
1373 * The order of subdivision here is critical for the IO subsystem.
1374 * Please do not alter this order without good reasons and regression
1375 * testing. Specifically, as large blocks of memory are subdivided,
1376 * the order in which smaller blocks are delivered depends on the order
1377 * they're subdivided in this function. This is the primary factor
1378 * influencing the order in which pages are delivered to the IO
1379 * subsystem according to empirical testing, and this is also justified
1380 * by considering the behavior of a buddy system containing a single
1381 * large block of memory acted on by a series of small allocations.
1382 * This behavior is a critical factor in sglist merging's success.
1383 *
1384 * -- nyc
1385 */
1386static inline void expand(struct zone *zone, struct page *page,
1387 int low, int high, int migratetype)
1388{
1389 unsigned long size = 1 << high;
1390
1391 while (high > low) {
1392 high--;
1393 size >>= 1;
1394 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1395
1396 /*
1397 * Mark as guard pages (or page), that will allow to
1398 * merge back to allocator when buddy will be freed.
1399 * Corresponding page table entries will not be touched,
1400 * pages will stay not present in virtual address space
1401 */
1402 if (set_page_guard(zone, &page[size], high, migratetype))
1403 continue;
1404
1405 add_to_free_list(&page[size], zone, high, migratetype);
1406 set_buddy_order(&page[size], high);
1407 }
1408}
1409
1410static void check_new_page_bad(struct page *page)
1411{
1412 if (unlikely(page->flags & __PG_HWPOISON)) {
1413 /* Don't complain about hwpoisoned pages */
1414 page_mapcount_reset(page); /* remove PageBuddy */
1415 return;
1416 }
1417
1418 bad_page(page,
1419 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1420}
1421
1422/*
1423 * This page is about to be returned from the page allocator
1424 */
1425static int check_new_page(struct page *page)
1426{
1427 if (likely(page_expected_state(page,
1428 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1429 return 0;
1430
1431 check_new_page_bad(page);
1432 return 1;
1433}
1434
1435static inline bool check_new_pages(struct page *page, unsigned int order)
1436{
1437 if (is_check_pages_enabled()) {
1438 for (int i = 0; i < (1 << order); i++) {
1439 struct page *p = page + i;
1440
1441 if (check_new_page(p))
1442 return true;
1443 }
1444 }
1445
1446 return false;
1447}
1448
1449static inline bool should_skip_kasan_unpoison(gfp_t flags)
1450{
1451 /* Don't skip if a software KASAN mode is enabled. */
1452 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1453 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1454 return false;
1455
1456 /* Skip, if hardware tag-based KASAN is not enabled. */
1457 if (!kasan_hw_tags_enabled())
1458 return true;
1459
1460 /*
1461 * With hardware tag-based KASAN enabled, skip if this has been
1462 * requested via __GFP_SKIP_KASAN.
1463 */
1464 return flags & __GFP_SKIP_KASAN;
1465}
1466
1467static inline bool should_skip_init(gfp_t flags)
1468{
1469 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1470 if (!kasan_hw_tags_enabled())
1471 return false;
1472
1473 /* For hardware tag-based KASAN, skip if requested. */
1474 return (flags & __GFP_SKIP_ZERO);
1475}
1476
1477inline void post_alloc_hook(struct page *page, unsigned int order,
1478 gfp_t gfp_flags)
1479{
1480 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1481 !should_skip_init(gfp_flags);
1482 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1483 int i;
1484
1485 set_page_private(page, 0);
1486 set_page_refcounted(page);
1487
1488 arch_alloc_page(page, order);
1489 debug_pagealloc_map_pages(page, 1 << order);
1490
1491 /*
1492 * Page unpoisoning must happen before memory initialization.
1493 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1494 * allocations and the page unpoisoning code will complain.
1495 */
1496 kernel_unpoison_pages(page, 1 << order);
1497
1498 /*
1499 * As memory initialization might be integrated into KASAN,
1500 * KASAN unpoisoning and memory initializion code must be
1501 * kept together to avoid discrepancies in behavior.
1502 */
1503
1504 /*
1505 * If memory tags should be zeroed
1506 * (which happens only when memory should be initialized as well).
1507 */
1508 if (zero_tags) {
1509 /* Initialize both memory and memory tags. */
1510 for (i = 0; i != 1 << order; ++i)
1511 tag_clear_highpage(page + i);
1512
1513 /* Take note that memory was initialized by the loop above. */
1514 init = false;
1515 }
1516 if (!should_skip_kasan_unpoison(gfp_flags) &&
1517 kasan_unpoison_pages(page, order, init)) {
1518 /* Take note that memory was initialized by KASAN. */
1519 if (kasan_has_integrated_init())
1520 init = false;
1521 } else {
1522 /*
1523 * If memory tags have not been set by KASAN, reset the page
1524 * tags to ensure page_address() dereferencing does not fault.
1525 */
1526 for (i = 0; i != 1 << order; ++i)
1527 page_kasan_tag_reset(page + i);
1528 }
1529 /* If memory is still not initialized, initialize it now. */
1530 if (init)
1531 kernel_init_pages(page, 1 << order);
1532
1533 set_page_owner(page, order, gfp_flags);
1534 page_table_check_alloc(page, order);
1535}
1536
1537static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1538 unsigned int alloc_flags)
1539{
1540 post_alloc_hook(page, order, gfp_flags);
1541
1542 if (order && (gfp_flags & __GFP_COMP))
1543 prep_compound_page(page, order);
1544
1545 /*
1546 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1547 * allocate the page. The expectation is that the caller is taking
1548 * steps that will free more memory. The caller should avoid the page
1549 * being used for !PFMEMALLOC purposes.
1550 */
1551 if (alloc_flags & ALLOC_NO_WATERMARKS)
1552 set_page_pfmemalloc(page);
1553 else
1554 clear_page_pfmemalloc(page);
1555}
1556
1557/*
1558 * Go through the free lists for the given migratetype and remove
1559 * the smallest available page from the freelists
1560 */
1561static __always_inline
1562struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1563 int migratetype)
1564{
1565 unsigned int current_order;
1566 struct free_area *area;
1567 struct page *page;
1568
1569 /* Find a page of the appropriate size in the preferred list */
1570 for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
1571 area = &(zone->free_area[current_order]);
1572 page = get_page_from_free_area(area, migratetype);
1573 if (!page)
1574 continue;
1575 del_page_from_free_list(page, zone, current_order);
1576 expand(zone, page, order, current_order, migratetype);
1577 set_pcppage_migratetype(page, migratetype);
1578 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1579 pcp_allowed_order(order) &&
1580 migratetype < MIGRATE_PCPTYPES);
1581 return page;
1582 }
1583
1584 return NULL;
1585}
1586
1587
1588/*
1589 * This array describes the order lists are fallen back to when
1590 * the free lists for the desirable migrate type are depleted
1591 *
1592 * The other migratetypes do not have fallbacks.
1593 */
1594static int fallbacks[MIGRATE_TYPES][MIGRATE_PCPTYPES - 1] = {
1595 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1596 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1597 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1598};
1599
1600#ifdef CONFIG_CMA
1601static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1602 unsigned int order)
1603{
1604 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1605}
1606#else
1607static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1608 unsigned int order) { return NULL; }
1609#endif
1610
1611/*
1612 * Move the free pages in a range to the freelist tail of the requested type.
1613 * Note that start_page and end_pages are not aligned on a pageblock
1614 * boundary. If alignment is required, use move_freepages_block()
1615 */
1616static int move_freepages(struct zone *zone,
1617 unsigned long start_pfn, unsigned long end_pfn,
1618 int migratetype, int *num_movable)
1619{
1620 struct page *page;
1621 unsigned long pfn;
1622 unsigned int order;
1623 int pages_moved = 0;
1624
1625 for (pfn = start_pfn; pfn <= end_pfn;) {
1626 page = pfn_to_page(pfn);
1627 if (!PageBuddy(page)) {
1628 /*
1629 * We assume that pages that could be isolated for
1630 * migration are movable. But we don't actually try
1631 * isolating, as that would be expensive.
1632 */
1633 if (num_movable &&
1634 (PageLRU(page) || __PageMovable(page)))
1635 (*num_movable)++;
1636 pfn++;
1637 continue;
1638 }
1639
1640 /* Make sure we are not inadvertently changing nodes */
1641 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1642 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1643
1644 order = buddy_order(page);
1645 move_to_free_list(page, zone, order, migratetype);
1646 pfn += 1 << order;
1647 pages_moved += 1 << order;
1648 }
1649
1650 return pages_moved;
1651}
1652
1653int move_freepages_block(struct zone *zone, struct page *page,
1654 int migratetype, int *num_movable)
1655{
1656 unsigned long start_pfn, end_pfn, pfn;
1657
1658 if (num_movable)
1659 *num_movable = 0;
1660
1661 pfn = page_to_pfn(page);
1662 start_pfn = pageblock_start_pfn(pfn);
1663 end_pfn = pageblock_end_pfn(pfn) - 1;
1664
1665 /* Do not cross zone boundaries */
1666 if (!zone_spans_pfn(zone, start_pfn))
1667 start_pfn = pfn;
1668 if (!zone_spans_pfn(zone, end_pfn))
1669 return 0;
1670
1671 return move_freepages(zone, start_pfn, end_pfn, migratetype,
1672 num_movable);
1673}
1674
1675static void change_pageblock_range(struct page *pageblock_page,
1676 int start_order, int migratetype)
1677{
1678 int nr_pageblocks = 1 << (start_order - pageblock_order);
1679
1680 while (nr_pageblocks--) {
1681 set_pageblock_migratetype(pageblock_page, migratetype);
1682 pageblock_page += pageblock_nr_pages;
1683 }
1684}
1685
1686/*
1687 * When we are falling back to another migratetype during allocation, try to
1688 * steal extra free pages from the same pageblocks to satisfy further
1689 * allocations, instead of polluting multiple pageblocks.
1690 *
1691 * If we are stealing a relatively large buddy page, it is likely there will
1692 * be more free pages in the pageblock, so try to steal them all. For
1693 * reclaimable and unmovable allocations, we steal regardless of page size,
1694 * as fragmentation caused by those allocations polluting movable pageblocks
1695 * is worse than movable allocations stealing from unmovable and reclaimable
1696 * pageblocks.
1697 */
1698static bool can_steal_fallback(unsigned int order, int start_mt)
1699{
1700 /*
1701 * Leaving this order check is intended, although there is
1702 * relaxed order check in next check. The reason is that
1703 * we can actually steal whole pageblock if this condition met,
1704 * but, below check doesn't guarantee it and that is just heuristic
1705 * so could be changed anytime.
1706 */
1707 if (order >= pageblock_order)
1708 return true;
1709
1710 if (order >= pageblock_order / 2 ||
1711 start_mt == MIGRATE_RECLAIMABLE ||
1712 start_mt == MIGRATE_UNMOVABLE ||
1713 page_group_by_mobility_disabled)
1714 return true;
1715
1716 return false;
1717}
1718
1719static inline bool boost_watermark(struct zone *zone)
1720{
1721 unsigned long max_boost;
1722
1723 if (!watermark_boost_factor)
1724 return false;
1725 /*
1726 * Don't bother in zones that are unlikely to produce results.
1727 * On small machines, including kdump capture kernels running
1728 * in a small area, boosting the watermark can cause an out of
1729 * memory situation immediately.
1730 */
1731 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1732 return false;
1733
1734 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1735 watermark_boost_factor, 10000);
1736
1737 /*
1738 * high watermark may be uninitialised if fragmentation occurs
1739 * very early in boot so do not boost. We do not fall
1740 * through and boost by pageblock_nr_pages as failing
1741 * allocations that early means that reclaim is not going
1742 * to help and it may even be impossible to reclaim the
1743 * boosted watermark resulting in a hang.
1744 */
1745 if (!max_boost)
1746 return false;
1747
1748 max_boost = max(pageblock_nr_pages, max_boost);
1749
1750 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1751 max_boost);
1752
1753 return true;
1754}
1755
1756/*
1757 * This function implements actual steal behaviour. If order is large enough,
1758 * we can steal whole pageblock. If not, we first move freepages in this
1759 * pageblock to our migratetype and determine how many already-allocated pages
1760 * are there in the pageblock with a compatible migratetype. If at least half
1761 * of pages are free or compatible, we can change migratetype of the pageblock
1762 * itself, so pages freed in the future will be put on the correct free list.
1763 */
1764static void steal_suitable_fallback(struct zone *zone, struct page *page,
1765 unsigned int alloc_flags, int start_type, bool whole_block)
1766{
1767 unsigned int current_order = buddy_order(page);
1768 int free_pages, movable_pages, alike_pages;
1769 int old_block_type;
1770
1771 old_block_type = get_pageblock_migratetype(page);
1772
1773 /*
1774 * This can happen due to races and we want to prevent broken
1775 * highatomic accounting.
1776 */
1777 if (is_migrate_highatomic(old_block_type))
1778 goto single_page;
1779
1780 /* Take ownership for orders >= pageblock_order */
1781 if (current_order >= pageblock_order) {
1782 change_pageblock_range(page, current_order, start_type);
1783 goto single_page;
1784 }
1785
1786 /*
1787 * Boost watermarks to increase reclaim pressure to reduce the
1788 * likelihood of future fallbacks. Wake kswapd now as the node
1789 * may be balanced overall and kswapd will not wake naturally.
1790 */
1791 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1792 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1793
1794 /* We are not allowed to try stealing from the whole block */
1795 if (!whole_block)
1796 goto single_page;
1797
1798 free_pages = move_freepages_block(zone, page, start_type,
1799 &movable_pages);
1800 /* moving whole block can fail due to zone boundary conditions */
1801 if (!free_pages)
1802 goto single_page;
1803
1804 /*
1805 * Determine how many pages are compatible with our allocation.
1806 * For movable allocation, it's the number of movable pages which
1807 * we just obtained. For other types it's a bit more tricky.
1808 */
1809 if (start_type == MIGRATE_MOVABLE) {
1810 alike_pages = movable_pages;
1811 } else {
1812 /*
1813 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1814 * to MOVABLE pageblock, consider all non-movable pages as
1815 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1816 * vice versa, be conservative since we can't distinguish the
1817 * exact migratetype of non-movable pages.
1818 */
1819 if (old_block_type == MIGRATE_MOVABLE)
1820 alike_pages = pageblock_nr_pages
1821 - (free_pages + movable_pages);
1822 else
1823 alike_pages = 0;
1824 }
1825 /*
1826 * If a sufficient number of pages in the block are either free or of
1827 * compatible migratability as our allocation, claim the whole block.
1828 */
1829 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1830 page_group_by_mobility_disabled)
1831 set_pageblock_migratetype(page, start_type);
1832
1833 return;
1834
1835single_page:
1836 move_to_free_list(page, zone, current_order, start_type);
1837}
1838
1839/*
1840 * Check whether there is a suitable fallback freepage with requested order.
1841 * If only_stealable is true, this function returns fallback_mt only if
1842 * we can steal other freepages all together. This would help to reduce
1843 * fragmentation due to mixed migratetype pages in one pageblock.
1844 */
1845int find_suitable_fallback(struct free_area *area, unsigned int order,
1846 int migratetype, bool only_stealable, bool *can_steal)
1847{
1848 int i;
1849 int fallback_mt;
1850
1851 if (area->nr_free == 0)
1852 return -1;
1853
1854 *can_steal = false;
1855 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
1856 fallback_mt = fallbacks[migratetype][i];
1857 if (free_area_empty(area, fallback_mt))
1858 continue;
1859
1860 if (can_steal_fallback(order, migratetype))
1861 *can_steal = true;
1862
1863 if (!only_stealable)
1864 return fallback_mt;
1865
1866 if (*can_steal)
1867 return fallback_mt;
1868 }
1869
1870 return -1;
1871}
1872
1873/*
1874 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1875 * there are no empty page blocks that contain a page with a suitable order
1876 */
1877static void reserve_highatomic_pageblock(struct page *page, struct zone *zone)
1878{
1879 int mt;
1880 unsigned long max_managed, flags;
1881
1882 /*
1883 * The number reserved as: minimum is 1 pageblock, maximum is
1884 * roughly 1% of a zone. But if 1% of a zone falls below a
1885 * pageblock size, then don't reserve any pageblocks.
1886 * Check is race-prone but harmless.
1887 */
1888 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
1889 return;
1890 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
1891 if (zone->nr_reserved_highatomic >= max_managed)
1892 return;
1893
1894 spin_lock_irqsave(&zone->lock, flags);
1895
1896 /* Recheck the nr_reserved_highatomic limit under the lock */
1897 if (zone->nr_reserved_highatomic >= max_managed)
1898 goto out_unlock;
1899
1900 /* Yoink! */
1901 mt = get_pageblock_migratetype(page);
1902 /* Only reserve normal pageblocks (i.e., they can merge with others) */
1903 if (migratetype_is_mergeable(mt)) {
1904 zone->nr_reserved_highatomic += pageblock_nr_pages;
1905 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1906 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
1907 }
1908
1909out_unlock:
1910 spin_unlock_irqrestore(&zone->lock, flags);
1911}
1912
1913/*
1914 * Used when an allocation is about to fail under memory pressure. This
1915 * potentially hurts the reliability of high-order allocations when under
1916 * intense memory pressure but failed atomic allocations should be easier
1917 * to recover from than an OOM.
1918 *
1919 * If @force is true, try to unreserve a pageblock even though highatomic
1920 * pageblock is exhausted.
1921 */
1922static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
1923 bool force)
1924{
1925 struct zonelist *zonelist = ac->zonelist;
1926 unsigned long flags;
1927 struct zoneref *z;
1928 struct zone *zone;
1929 struct page *page;
1930 int order;
1931 bool ret;
1932
1933 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
1934 ac->nodemask) {
1935 /*
1936 * Preserve at least one pageblock unless memory pressure
1937 * is really high.
1938 */
1939 if (!force && zone->nr_reserved_highatomic <=
1940 pageblock_nr_pages)
1941 continue;
1942
1943 spin_lock_irqsave(&zone->lock, flags);
1944 for (order = 0; order < NR_PAGE_ORDERS; order++) {
1945 struct free_area *area = &(zone->free_area[order]);
1946
1947 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
1948 if (!page)
1949 continue;
1950
1951 /*
1952 * In page freeing path, migratetype change is racy so
1953 * we can counter several free pages in a pageblock
1954 * in this loop although we changed the pageblock type
1955 * from highatomic to ac->migratetype. So we should
1956 * adjust the count once.
1957 */
1958 if (is_migrate_highatomic_page(page)) {
1959 /*
1960 * It should never happen but changes to
1961 * locking could inadvertently allow a per-cpu
1962 * drain to add pages to MIGRATE_HIGHATOMIC
1963 * while unreserving so be safe and watch for
1964 * underflows.
1965 */
1966 zone->nr_reserved_highatomic -= min(
1967 pageblock_nr_pages,
1968 zone->nr_reserved_highatomic);
1969 }
1970
1971 /*
1972 * Convert to ac->migratetype and avoid the normal
1973 * pageblock stealing heuristics. Minimally, the caller
1974 * is doing the work and needs the pages. More
1975 * importantly, if the block was always converted to
1976 * MIGRATE_UNMOVABLE or another type then the number
1977 * of pageblocks that cannot be completely freed
1978 * may increase.
1979 */
1980 set_pageblock_migratetype(page, ac->migratetype);
1981 ret = move_freepages_block(zone, page, ac->migratetype,
1982 NULL);
1983 if (ret) {
1984 spin_unlock_irqrestore(&zone->lock, flags);
1985 return ret;
1986 }
1987 }
1988 spin_unlock_irqrestore(&zone->lock, flags);
1989 }
1990
1991 return false;
1992}
1993
1994/*
1995 * Try finding a free buddy page on the fallback list and put it on the free
1996 * list of requested migratetype, possibly along with other pages from the same
1997 * block, depending on fragmentation avoidance heuristics. Returns true if
1998 * fallback was found so that __rmqueue_smallest() can grab it.
1999 *
2000 * The use of signed ints for order and current_order is a deliberate
2001 * deviation from the rest of this file, to make the for loop
2002 * condition simpler.
2003 */
2004static __always_inline bool
2005__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2006 unsigned int alloc_flags)
2007{
2008 struct free_area *area;
2009 int current_order;
2010 int min_order = order;
2011 struct page *page;
2012 int fallback_mt;
2013 bool can_steal;
2014
2015 /*
2016 * Do not steal pages from freelists belonging to other pageblocks
2017 * i.e. orders < pageblock_order. If there are no local zones free,
2018 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2019 */
2020 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2021 min_order = pageblock_order;
2022
2023 /*
2024 * Find the largest available free page in the other list. This roughly
2025 * approximates finding the pageblock with the most free pages, which
2026 * would be too costly to do exactly.
2027 */
2028 for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2029 --current_order) {
2030 area = &(zone->free_area[current_order]);
2031 fallback_mt = find_suitable_fallback(area, current_order,
2032 start_migratetype, false, &can_steal);
2033 if (fallback_mt == -1)
2034 continue;
2035
2036 /*
2037 * We cannot steal all free pages from the pageblock and the
2038 * requested migratetype is movable. In that case it's better to
2039 * steal and split the smallest available page instead of the
2040 * largest available page, because even if the next movable
2041 * allocation falls back into a different pageblock than this
2042 * one, it won't cause permanent fragmentation.
2043 */
2044 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2045 && current_order > order)
2046 goto find_smallest;
2047
2048 goto do_steal;
2049 }
2050
2051 return false;
2052
2053find_smallest:
2054 for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
2055 area = &(zone->free_area[current_order]);
2056 fallback_mt = find_suitable_fallback(area, current_order,
2057 start_migratetype, false, &can_steal);
2058 if (fallback_mt != -1)
2059 break;
2060 }
2061
2062 /*
2063 * This should not happen - we already found a suitable fallback
2064 * when looking for the largest page.
2065 */
2066 VM_BUG_ON(current_order > MAX_PAGE_ORDER);
2067
2068do_steal:
2069 page = get_page_from_free_area(area, fallback_mt);
2070
2071 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2072 can_steal);
2073
2074 trace_mm_page_alloc_extfrag(page, order, current_order,
2075 start_migratetype, fallback_mt);
2076
2077 return true;
2078
2079}
2080
2081/*
2082 * Do the hard work of removing an element from the buddy allocator.
2083 * Call me with the zone->lock already held.
2084 */
2085static __always_inline struct page *
2086__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2087 unsigned int alloc_flags)
2088{
2089 struct page *page;
2090
2091 if (IS_ENABLED(CONFIG_CMA)) {
2092 /*
2093 * Balance movable allocations between regular and CMA areas by
2094 * allocating from CMA when over half of the zone's free memory
2095 * is in the CMA area.
2096 */
2097 if (alloc_flags & ALLOC_CMA &&
2098 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2099 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2100 page = __rmqueue_cma_fallback(zone, order);
2101 if (page)
2102 return page;
2103 }
2104 }
2105retry:
2106 page = __rmqueue_smallest(zone, order, migratetype);
2107 if (unlikely(!page)) {
2108 if (alloc_flags & ALLOC_CMA)
2109 page = __rmqueue_cma_fallback(zone, order);
2110
2111 if (!page && __rmqueue_fallback(zone, order, migratetype,
2112 alloc_flags))
2113 goto retry;
2114 }
2115 return page;
2116}
2117
2118/*
2119 * Obtain a specified number of elements from the buddy allocator, all under
2120 * a single hold of the lock, for efficiency. Add them to the supplied list.
2121 * Returns the number of new pages which were placed at *list.
2122 */
2123static int rmqueue_bulk(struct zone *zone, unsigned int order,
2124 unsigned long count, struct list_head *list,
2125 int migratetype, unsigned int alloc_flags)
2126{
2127 unsigned long flags;
2128 int i;
2129
2130 spin_lock_irqsave(&zone->lock, flags);
2131 for (i = 0; i < count; ++i) {
2132 struct page *page = __rmqueue(zone, order, migratetype,
2133 alloc_flags);
2134 if (unlikely(page == NULL))
2135 break;
2136
2137 /*
2138 * Split buddy pages returned by expand() are received here in
2139 * physical page order. The page is added to the tail of
2140 * caller's list. From the callers perspective, the linked list
2141 * is ordered by page number under some conditions. This is
2142 * useful for IO devices that can forward direction from the
2143 * head, thus also in the physical page order. This is useful
2144 * for IO devices that can merge IO requests if the physical
2145 * pages are ordered properly.
2146 */
2147 list_add_tail(&page->pcp_list, list);
2148 if (is_migrate_cma(get_pcppage_migratetype(page)))
2149 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2150 -(1 << order));
2151 }
2152
2153 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2154 spin_unlock_irqrestore(&zone->lock, flags);
2155
2156 return i;
2157}
2158
2159/*
2160 * Called from the vmstat counter updater to decay the PCP high.
2161 * Return whether there are addition works to do.
2162 */
2163int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2164{
2165 int high_min, to_drain, batch;
2166 int todo = 0;
2167
2168 high_min = READ_ONCE(pcp->high_min);
2169 batch = READ_ONCE(pcp->batch);
2170 /*
2171 * Decrease pcp->high periodically to try to free possible
2172 * idle PCP pages. And, avoid to free too many pages to
2173 * control latency. This caps pcp->high decrement too.
2174 */
2175 if (pcp->high > high_min) {
2176 pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2177 pcp->high - (pcp->high >> 3), high_min);
2178 if (pcp->high > high_min)
2179 todo++;
2180 }
2181
2182 to_drain = pcp->count - pcp->high;
2183 if (to_drain > 0) {
2184 spin_lock(&pcp->lock);
2185 free_pcppages_bulk(zone, to_drain, pcp, 0);
2186 spin_unlock(&pcp->lock);
2187 todo++;
2188 }
2189
2190 return todo;
2191}
2192
2193#ifdef CONFIG_NUMA
2194/*
2195 * Called from the vmstat counter updater to drain pagesets of this
2196 * currently executing processor on remote nodes after they have
2197 * expired.
2198 */
2199void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2200{
2201 int to_drain, batch;
2202
2203 batch = READ_ONCE(pcp->batch);
2204 to_drain = min(pcp->count, batch);
2205 if (to_drain > 0) {
2206 spin_lock(&pcp->lock);
2207 free_pcppages_bulk(zone, to_drain, pcp, 0);
2208 spin_unlock(&pcp->lock);
2209 }
2210}
2211#endif
2212
2213/*
2214 * Drain pcplists of the indicated processor and zone.
2215 */
2216static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2217{
2218 struct per_cpu_pages *pcp;
2219
2220 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2221 if (pcp->count) {
2222 spin_lock(&pcp->lock);
2223 free_pcppages_bulk(zone, pcp->count, pcp, 0);
2224 spin_unlock(&pcp->lock);
2225 }
2226}
2227
2228/*
2229 * Drain pcplists of all zones on the indicated processor.
2230 */
2231static void drain_pages(unsigned int cpu)
2232{
2233 struct zone *zone;
2234
2235 for_each_populated_zone(zone) {
2236 drain_pages_zone(cpu, zone);
2237 }
2238}
2239
2240/*
2241 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2242 */
2243void drain_local_pages(struct zone *zone)
2244{
2245 int cpu = smp_processor_id();
2246
2247 if (zone)
2248 drain_pages_zone(cpu, zone);
2249 else
2250 drain_pages(cpu);
2251}
2252
2253/*
2254 * The implementation of drain_all_pages(), exposing an extra parameter to
2255 * drain on all cpus.
2256 *
2257 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2258 * not empty. The check for non-emptiness can however race with a free to
2259 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2260 * that need the guarantee that every CPU has drained can disable the
2261 * optimizing racy check.
2262 */
2263static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2264{
2265 int cpu;
2266
2267 /*
2268 * Allocate in the BSS so we won't require allocation in
2269 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2270 */
2271 static cpumask_t cpus_with_pcps;
2272
2273 /*
2274 * Do not drain if one is already in progress unless it's specific to
2275 * a zone. Such callers are primarily CMA and memory hotplug and need
2276 * the drain to be complete when the call returns.
2277 */
2278 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2279 if (!zone)
2280 return;
2281 mutex_lock(&pcpu_drain_mutex);
2282 }
2283
2284 /*
2285 * We don't care about racing with CPU hotplug event
2286 * as offline notification will cause the notified
2287 * cpu to drain that CPU pcps and on_each_cpu_mask
2288 * disables preemption as part of its processing
2289 */
2290 for_each_online_cpu(cpu) {
2291 struct per_cpu_pages *pcp;
2292 struct zone *z;
2293 bool has_pcps = false;
2294
2295 if (force_all_cpus) {
2296 /*
2297 * The pcp.count check is racy, some callers need a
2298 * guarantee that no cpu is missed.
2299 */
2300 has_pcps = true;
2301 } else if (zone) {
2302 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2303 if (pcp->count)
2304 has_pcps = true;
2305 } else {
2306 for_each_populated_zone(z) {
2307 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2308 if (pcp->count) {
2309 has_pcps = true;
2310 break;
2311 }
2312 }
2313 }
2314
2315 if (has_pcps)
2316 cpumask_set_cpu(cpu, &cpus_with_pcps);
2317 else
2318 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2319 }
2320
2321 for_each_cpu(cpu, &cpus_with_pcps) {
2322 if (zone)
2323 drain_pages_zone(cpu, zone);
2324 else
2325 drain_pages(cpu);
2326 }
2327
2328 mutex_unlock(&pcpu_drain_mutex);
2329}
2330
2331/*
2332 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2333 *
2334 * When zone parameter is non-NULL, spill just the single zone's pages.
2335 */
2336void drain_all_pages(struct zone *zone)
2337{
2338 __drain_all_pages(zone, false);
2339}
2340
2341static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
2342 unsigned int order)
2343{
2344 int migratetype;
2345
2346 if (!free_pages_prepare(page, order, FPI_NONE))
2347 return false;
2348
2349 migratetype = get_pfnblock_migratetype(page, pfn);
2350 set_pcppage_migratetype(page, migratetype);
2351 return true;
2352}
2353
2354static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2355{
2356 int min_nr_free, max_nr_free;
2357
2358 /* Free as much as possible if batch freeing high-order pages. */
2359 if (unlikely(free_high))
2360 return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2361
2362 /* Check for PCP disabled or boot pageset */
2363 if (unlikely(high < batch))
2364 return 1;
2365
2366 /* Leave at least pcp->batch pages on the list */
2367 min_nr_free = batch;
2368 max_nr_free = high - batch;
2369
2370 /*
2371 * Increase the batch number to the number of the consecutive
2372 * freed pages to reduce zone lock contention.
2373 */
2374 batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2375
2376 return batch;
2377}
2378
2379static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2380 int batch, bool free_high)
2381{
2382 int high, high_min, high_max;
2383
2384 high_min = READ_ONCE(pcp->high_min);
2385 high_max = READ_ONCE(pcp->high_max);
2386 high = pcp->high = clamp(pcp->high, high_min, high_max);
2387
2388 if (unlikely(!high))
2389 return 0;
2390
2391 if (unlikely(free_high)) {
2392 pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2393 high_min);
2394 return 0;
2395 }
2396
2397 /*
2398 * If reclaim is active, limit the number of pages that can be
2399 * stored on pcp lists
2400 */
2401 if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2402 int free_count = max_t(int, pcp->free_count, batch);
2403
2404 pcp->high = max(high - free_count, high_min);
2405 return min(batch << 2, pcp->high);
2406 }
2407
2408 if (high_min == high_max)
2409 return high;
2410
2411 if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2412 int free_count = max_t(int, pcp->free_count, batch);
2413
2414 pcp->high = max(high - free_count, high_min);
2415 high = max(pcp->count, high_min);
2416 } else if (pcp->count >= high) {
2417 int need_high = pcp->free_count + batch;
2418
2419 /* pcp->high should be large enough to hold batch freed pages */
2420 if (pcp->high < need_high)
2421 pcp->high = clamp(need_high, high_min, high_max);
2422 }
2423
2424 return high;
2425}
2426
2427static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2428 struct page *page, int migratetype,
2429 unsigned int order)
2430{
2431 int high, batch;
2432 int pindex;
2433 bool free_high = false;
2434
2435 /*
2436 * On freeing, reduce the number of pages that are batch allocated.
2437 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2438 * allocations.
2439 */
2440 pcp->alloc_factor >>= 1;
2441 __count_vm_events(PGFREE, 1 << order);
2442 pindex = order_to_pindex(migratetype, order);
2443 list_add(&page->pcp_list, &pcp->lists[pindex]);
2444 pcp->count += 1 << order;
2445
2446 batch = READ_ONCE(pcp->batch);
2447 /*
2448 * As high-order pages other than THP's stored on PCP can contribute
2449 * to fragmentation, limit the number stored when PCP is heavily
2450 * freeing without allocation. The remainder after bulk freeing
2451 * stops will be drained from vmstat refresh context.
2452 */
2453 if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2454 free_high = (pcp->free_count >= batch &&
2455 (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2456 (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2457 pcp->count >= READ_ONCE(batch)));
2458 pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2459 } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2460 pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2461 }
2462 if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2463 pcp->free_count += (1 << order);
2464 high = nr_pcp_high(pcp, zone, batch, free_high);
2465 if (pcp->count >= high) {
2466 free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high),
2467 pcp, pindex);
2468 if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2469 zone_watermark_ok(zone, 0, high_wmark_pages(zone),
2470 ZONE_MOVABLE, 0))
2471 clear_bit(ZONE_BELOW_HIGH, &zone->flags);
2472 }
2473}
2474
2475/*
2476 * Free a pcp page
2477 */
2478void free_unref_page(struct page *page, unsigned int order)
2479{
2480 unsigned long __maybe_unused UP_flags;
2481 struct per_cpu_pages *pcp;
2482 struct zone *zone;
2483 unsigned long pfn = page_to_pfn(page);
2484 int migratetype, pcpmigratetype;
2485
2486 if (!free_unref_page_prepare(page, pfn, order))
2487 return;
2488
2489 /*
2490 * We only track unmovable, reclaimable and movable on pcp lists.
2491 * Place ISOLATE pages on the isolated list because they are being
2492 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2493 * get those areas back if necessary. Otherwise, we may have to free
2494 * excessively into the page allocator
2495 */
2496 migratetype = pcpmigratetype = get_pcppage_migratetype(page);
2497 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2498 if (unlikely(is_migrate_isolate(migratetype))) {
2499 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
2500 return;
2501 }
2502 pcpmigratetype = MIGRATE_MOVABLE;
2503 }
2504
2505 zone = page_zone(page);
2506 pcp_trylock_prepare(UP_flags);
2507 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2508 if (pcp) {
2509 free_unref_page_commit(zone, pcp, page, pcpmigratetype, order);
2510 pcp_spin_unlock(pcp);
2511 } else {
2512 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
2513 }
2514 pcp_trylock_finish(UP_flags);
2515}
2516
2517/*
2518 * Free a list of 0-order pages
2519 */
2520void free_unref_page_list(struct list_head *list)
2521{
2522 unsigned long __maybe_unused UP_flags;
2523 struct page *page, *next;
2524 struct per_cpu_pages *pcp = NULL;
2525 struct zone *locked_zone = NULL;
2526 int batch_count = 0;
2527 int migratetype;
2528
2529 /* Prepare pages for freeing */
2530 list_for_each_entry_safe(page, next, list, lru) {
2531 unsigned long pfn = page_to_pfn(page);
2532 if (!free_unref_page_prepare(page, pfn, 0)) {
2533 list_del(&page->lru);
2534 continue;
2535 }
2536
2537 /*
2538 * Free isolated pages directly to the allocator, see
2539 * comment in free_unref_page.
2540 */
2541 migratetype = get_pcppage_migratetype(page);
2542 if (unlikely(is_migrate_isolate(migratetype))) {
2543 list_del(&page->lru);
2544 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
2545 continue;
2546 }
2547 }
2548
2549 list_for_each_entry_safe(page, next, list, lru) {
2550 struct zone *zone = page_zone(page);
2551
2552 list_del(&page->lru);
2553 migratetype = get_pcppage_migratetype(page);
2554
2555 /*
2556 * Either different zone requiring a different pcp lock or
2557 * excessive lock hold times when freeing a large list of
2558 * pages.
2559 */
2560 if (zone != locked_zone || batch_count == SWAP_CLUSTER_MAX) {
2561 if (pcp) {
2562 pcp_spin_unlock(pcp);
2563 pcp_trylock_finish(UP_flags);
2564 }
2565
2566 batch_count = 0;
2567
2568 /*
2569 * trylock is necessary as pages may be getting freed
2570 * from IRQ or SoftIRQ context after an IO completion.
2571 */
2572 pcp_trylock_prepare(UP_flags);
2573 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2574 if (unlikely(!pcp)) {
2575 pcp_trylock_finish(UP_flags);
2576 free_one_page(zone, page, page_to_pfn(page),
2577 0, migratetype, FPI_NONE);
2578 locked_zone = NULL;
2579 continue;
2580 }
2581 locked_zone = zone;
2582 }
2583
2584 /*
2585 * Non-isolated types over MIGRATE_PCPTYPES get added
2586 * to the MIGRATE_MOVABLE pcp list.
2587 */
2588 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2589 migratetype = MIGRATE_MOVABLE;
2590
2591 trace_mm_page_free_batched(page);
2592 free_unref_page_commit(zone, pcp, page, migratetype, 0);
2593 batch_count++;
2594 }
2595
2596 if (pcp) {
2597 pcp_spin_unlock(pcp);
2598 pcp_trylock_finish(UP_flags);
2599 }
2600}
2601
2602/*
2603 * split_page takes a non-compound higher-order page, and splits it into
2604 * n (1<<order) sub-pages: page[0..n]
2605 * Each sub-page must be freed individually.
2606 *
2607 * Note: this is probably too low level an operation for use in drivers.
2608 * Please consult with lkml before using this in your driver.
2609 */
2610void split_page(struct page *page, unsigned int order)
2611{
2612 int i;
2613
2614 VM_BUG_ON_PAGE(PageCompound(page), page);
2615 VM_BUG_ON_PAGE(!page_count(page), page);
2616
2617 for (i = 1; i < (1 << order); i++)
2618 set_page_refcounted(page + i);
2619 split_page_owner(page, 1 << order);
2620 split_page_memcg(page, 1 << order);
2621}
2622EXPORT_SYMBOL_GPL(split_page);
2623
2624int __isolate_free_page(struct page *page, unsigned int order)
2625{
2626 struct zone *zone = page_zone(page);
2627 int mt = get_pageblock_migratetype(page);
2628
2629 if (!is_migrate_isolate(mt)) {
2630 unsigned long watermark;
2631 /*
2632 * Obey watermarks as if the page was being allocated. We can
2633 * emulate a high-order watermark check with a raised order-0
2634 * watermark, because we already know our high-order page
2635 * exists.
2636 */
2637 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2638 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2639 return 0;
2640
2641 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2642 }
2643
2644 del_page_from_free_list(page, zone, order);
2645
2646 /*
2647 * Set the pageblock if the isolated page is at least half of a
2648 * pageblock
2649 */
2650 if (order >= pageblock_order - 1) {
2651 struct page *endpage = page + (1 << order) - 1;
2652 for (; page < endpage; page += pageblock_nr_pages) {
2653 int mt = get_pageblock_migratetype(page);
2654 /*
2655 * Only change normal pageblocks (i.e., they can merge
2656 * with others)
2657 */
2658 if (migratetype_is_mergeable(mt))
2659 set_pageblock_migratetype(page,
2660 MIGRATE_MOVABLE);
2661 }
2662 }
2663
2664 return 1UL << order;
2665}
2666
2667/**
2668 * __putback_isolated_page - Return a now-isolated page back where we got it
2669 * @page: Page that was isolated
2670 * @order: Order of the isolated page
2671 * @mt: The page's pageblock's migratetype
2672 *
2673 * This function is meant to return a page pulled from the free lists via
2674 * __isolate_free_page back to the free lists they were pulled from.
2675 */
2676void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2677{
2678 struct zone *zone = page_zone(page);
2679
2680 /* zone lock should be held when this function is called */
2681 lockdep_assert_held(&zone->lock);
2682
2683 /* Return isolated page to tail of freelist. */
2684 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2685 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2686}
2687
2688/*
2689 * Update NUMA hit/miss statistics
2690 */
2691static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2692 long nr_account)
2693{
2694#ifdef CONFIG_NUMA
2695 enum numa_stat_item local_stat = NUMA_LOCAL;
2696
2697 /* skip numa counters update if numa stats is disabled */
2698 if (!static_branch_likely(&vm_numa_stat_key))
2699 return;
2700
2701 if (zone_to_nid(z) != numa_node_id())
2702 local_stat = NUMA_OTHER;
2703
2704 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2705 __count_numa_events(z, NUMA_HIT, nr_account);
2706 else {
2707 __count_numa_events(z, NUMA_MISS, nr_account);
2708 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2709 }
2710 __count_numa_events(z, local_stat, nr_account);
2711#endif
2712}
2713
2714static __always_inline
2715struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2716 unsigned int order, unsigned int alloc_flags,
2717 int migratetype)
2718{
2719 struct page *page;
2720 unsigned long flags;
2721
2722 do {
2723 page = NULL;
2724 spin_lock_irqsave(&zone->lock, flags);
2725 if (alloc_flags & ALLOC_HIGHATOMIC)
2726 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2727 if (!page) {
2728 page = __rmqueue(zone, order, migratetype, alloc_flags);
2729
2730 /*
2731 * If the allocation fails, allow OOM handling access
2732 * to HIGHATOMIC reserves as failing now is worse than
2733 * failing a high-order atomic allocation in the
2734 * future.
2735 */
2736 if (!page && (alloc_flags & ALLOC_OOM))
2737 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2738
2739 if (!page) {
2740 spin_unlock_irqrestore(&zone->lock, flags);
2741 return NULL;
2742 }
2743 }
2744 __mod_zone_freepage_state(zone, -(1 << order),
2745 get_pcppage_migratetype(page));
2746 spin_unlock_irqrestore(&zone->lock, flags);
2747 } while (check_new_pages(page, order));
2748
2749 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2750 zone_statistics(preferred_zone, zone, 1);
2751
2752 return page;
2753}
2754
2755static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
2756{
2757 int high, base_batch, batch, max_nr_alloc;
2758 int high_max, high_min;
2759
2760 base_batch = READ_ONCE(pcp->batch);
2761 high_min = READ_ONCE(pcp->high_min);
2762 high_max = READ_ONCE(pcp->high_max);
2763 high = pcp->high = clamp(pcp->high, high_min, high_max);
2764
2765 /* Check for PCP disabled or boot pageset */
2766 if (unlikely(high < base_batch))
2767 return 1;
2768
2769 if (order)
2770 batch = base_batch;
2771 else
2772 batch = (base_batch << pcp->alloc_factor);
2773
2774 /*
2775 * If we had larger pcp->high, we could avoid to allocate from
2776 * zone.
2777 */
2778 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
2779 high = pcp->high = min(high + batch, high_max);
2780
2781 if (!order) {
2782 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
2783 /*
2784 * Double the number of pages allocated each time there is
2785 * subsequent allocation of order-0 pages without any freeing.
2786 */
2787 if (batch <= max_nr_alloc &&
2788 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
2789 pcp->alloc_factor++;
2790 batch = min(batch, max_nr_alloc);
2791 }
2792
2793 /*
2794 * Scale batch relative to order if batch implies free pages
2795 * can be stored on the PCP. Batch can be 1 for small zones or
2796 * for boot pagesets which should never store free pages as
2797 * the pages may belong to arbitrary zones.
2798 */
2799 if (batch > 1)
2800 batch = max(batch >> order, 2);
2801
2802 return batch;
2803}
2804
2805/* Remove page from the per-cpu list, caller must protect the list */
2806static inline
2807struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2808 int migratetype,
2809 unsigned int alloc_flags,
2810 struct per_cpu_pages *pcp,
2811 struct list_head *list)
2812{
2813 struct page *page;
2814
2815 do {
2816 if (list_empty(list)) {
2817 int batch = nr_pcp_alloc(pcp, zone, order);
2818 int alloced;
2819
2820 alloced = rmqueue_bulk(zone, order,
2821 batch, list,
2822 migratetype, alloc_flags);
2823
2824 pcp->count += alloced << order;
2825 if (unlikely(list_empty(list)))
2826 return NULL;
2827 }
2828
2829 page = list_first_entry(list, struct page, pcp_list);
2830 list_del(&page->pcp_list);
2831 pcp->count -= 1 << order;
2832 } while (check_new_pages(page, order));
2833
2834 return page;
2835}
2836
2837/* Lock and remove page from the per-cpu list */
2838static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2839 struct zone *zone, unsigned int order,
2840 int migratetype, unsigned int alloc_flags)
2841{
2842 struct per_cpu_pages *pcp;
2843 struct list_head *list;
2844 struct page *page;
2845 unsigned long __maybe_unused UP_flags;
2846
2847 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2848 pcp_trylock_prepare(UP_flags);
2849 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2850 if (!pcp) {
2851 pcp_trylock_finish(UP_flags);
2852 return NULL;
2853 }
2854
2855 /*
2856 * On allocation, reduce the number of pages that are batch freed.
2857 * See nr_pcp_free() where free_factor is increased for subsequent
2858 * frees.
2859 */
2860 pcp->free_count >>= 1;
2861 list = &pcp->lists[order_to_pindex(migratetype, order)];
2862 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
2863 pcp_spin_unlock(pcp);
2864 pcp_trylock_finish(UP_flags);
2865 if (page) {
2866 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2867 zone_statistics(preferred_zone, zone, 1);
2868 }
2869 return page;
2870}
2871
2872/*
2873 * Allocate a page from the given zone.
2874 * Use pcplists for THP or "cheap" high-order allocations.
2875 */
2876
2877/*
2878 * Do not instrument rmqueue() with KMSAN. This function may call
2879 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
2880 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
2881 * may call rmqueue() again, which will result in a deadlock.
2882 */
2883__no_sanitize_memory
2884static inline
2885struct page *rmqueue(struct zone *preferred_zone,
2886 struct zone *zone, unsigned int order,
2887 gfp_t gfp_flags, unsigned int alloc_flags,
2888 int migratetype)
2889{
2890 struct page *page;
2891
2892 /*
2893 * We most definitely don't want callers attempting to
2894 * allocate greater than order-1 page units with __GFP_NOFAIL.
2895 */
2896 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2897
2898 if (likely(pcp_allowed_order(order))) {
2899 page = rmqueue_pcplist(preferred_zone, zone, order,
2900 migratetype, alloc_flags);
2901 if (likely(page))
2902 goto out;
2903 }
2904
2905 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
2906 migratetype);
2907
2908out:
2909 /* Separate test+clear to avoid unnecessary atomics */
2910 if ((alloc_flags & ALLOC_KSWAPD) &&
2911 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
2912 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2913 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
2914 }
2915
2916 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2917 return page;
2918}
2919
2920noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2921{
2922 return __should_fail_alloc_page(gfp_mask, order);
2923}
2924ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
2925
2926static inline long __zone_watermark_unusable_free(struct zone *z,
2927 unsigned int order, unsigned int alloc_flags)
2928{
2929 long unusable_free = (1 << order) - 1;
2930
2931 /*
2932 * If the caller does not have rights to reserves below the min
2933 * watermark then subtract the high-atomic reserves. This will
2934 * over-estimate the size of the atomic reserve but it avoids a search.
2935 */
2936 if (likely(!(alloc_flags & ALLOC_RESERVES)))
2937 unusable_free += z->nr_reserved_highatomic;
2938
2939#ifdef CONFIG_CMA
2940 /* If allocation can't use CMA areas don't use free CMA pages */
2941 if (!(alloc_flags & ALLOC_CMA))
2942 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
2943#endif
2944#ifdef CONFIG_UNACCEPTED_MEMORY
2945 unusable_free += zone_page_state(z, NR_UNACCEPTED);
2946#endif
2947
2948 return unusable_free;
2949}
2950
2951/*
2952 * Return true if free base pages are above 'mark'. For high-order checks it
2953 * will return true of the order-0 watermark is reached and there is at least
2954 * one free page of a suitable size. Checking now avoids taking the zone lock
2955 * to check in the allocation paths if no pages are free.
2956 */
2957bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2958 int highest_zoneidx, unsigned int alloc_flags,
2959 long free_pages)
2960{
2961 long min = mark;
2962 int o;
2963
2964 /* free_pages may go negative - that's OK */
2965 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
2966
2967 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
2968 /*
2969 * __GFP_HIGH allows access to 50% of the min reserve as well
2970 * as OOM.
2971 */
2972 if (alloc_flags & ALLOC_MIN_RESERVE) {
2973 min -= min / 2;
2974
2975 /*
2976 * Non-blocking allocations (e.g. GFP_ATOMIC) can
2977 * access more reserves than just __GFP_HIGH. Other
2978 * non-blocking allocations requests such as GFP_NOWAIT
2979 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
2980 * access to the min reserve.
2981 */
2982 if (alloc_flags & ALLOC_NON_BLOCK)
2983 min -= min / 4;
2984 }
2985
2986 /*
2987 * OOM victims can try even harder than the normal reserve
2988 * users on the grounds that it's definitely going to be in
2989 * the exit path shortly and free memory. Any allocation it
2990 * makes during the free path will be small and short-lived.
2991 */
2992 if (alloc_flags & ALLOC_OOM)
2993 min -= min / 2;
2994 }
2995
2996 /*
2997 * Check watermarks for an order-0 allocation request. If these
2998 * are not met, then a high-order request also cannot go ahead
2999 * even if a suitable page happened to be free.
3000 */
3001 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3002 return false;
3003
3004 /* If this is an order-0 request then the watermark is fine */
3005 if (!order)
3006 return true;
3007
3008 /* For a high-order request, check at least one suitable page is free */
3009 for (o = order; o < NR_PAGE_ORDERS; o++) {
3010 struct free_area *area = &z->free_area[o];
3011 int mt;
3012
3013 if (!area->nr_free)
3014 continue;
3015
3016 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3017 if (!free_area_empty(area, mt))
3018 return true;
3019 }
3020
3021#ifdef CONFIG_CMA
3022 if ((alloc_flags & ALLOC_CMA) &&
3023 !free_area_empty(area, MIGRATE_CMA)) {
3024 return true;
3025 }
3026#endif
3027 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3028 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3029 return true;
3030 }
3031 }
3032 return false;
3033}
3034
3035bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3036 int highest_zoneidx, unsigned int alloc_flags)
3037{
3038 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3039 zone_page_state(z, NR_FREE_PAGES));
3040}
3041
3042static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3043 unsigned long mark, int highest_zoneidx,
3044 unsigned int alloc_flags, gfp_t gfp_mask)
3045{
3046 long free_pages;
3047
3048 free_pages = zone_page_state(z, NR_FREE_PAGES);
3049
3050 /*
3051 * Fast check for order-0 only. If this fails then the reserves
3052 * need to be calculated.
3053 */
3054 if (!order) {
3055 long usable_free;
3056 long reserved;
3057
3058 usable_free = free_pages;
3059 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3060
3061 /* reserved may over estimate high-atomic reserves. */
3062 usable_free -= min(usable_free, reserved);
3063 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3064 return true;
3065 }
3066
3067 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3068 free_pages))
3069 return true;
3070
3071 /*
3072 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3073 * when checking the min watermark. The min watermark is the
3074 * point where boosting is ignored so that kswapd is woken up
3075 * when below the low watermark.
3076 */
3077 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3078 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3079 mark = z->_watermark[WMARK_MIN];
3080 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3081 alloc_flags, free_pages);
3082 }
3083
3084 return false;
3085}
3086
3087bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3088 unsigned long mark, int highest_zoneidx)
3089{
3090 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3091
3092 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3093 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3094
3095 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3096 free_pages);
3097}
3098
3099#ifdef CONFIG_NUMA
3100int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3101
3102static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3103{
3104 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3105 node_reclaim_distance;
3106}
3107#else /* CONFIG_NUMA */
3108static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3109{
3110 return true;
3111}
3112#endif /* CONFIG_NUMA */
3113
3114/*
3115 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3116 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3117 * premature use of a lower zone may cause lowmem pressure problems that
3118 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3119 * probably too small. It only makes sense to spread allocations to avoid
3120 * fragmentation between the Normal and DMA32 zones.
3121 */
3122static inline unsigned int
3123alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3124{
3125 unsigned int alloc_flags;
3126
3127 /*
3128 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3129 * to save a branch.
3130 */
3131 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3132
3133#ifdef CONFIG_ZONE_DMA32
3134 if (!zone)
3135 return alloc_flags;
3136
3137 if (zone_idx(zone) != ZONE_NORMAL)
3138 return alloc_flags;
3139
3140 /*
3141 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3142 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3143 * on UMA that if Normal is populated then so is DMA32.
3144 */
3145 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3146 if (nr_online_nodes > 1 && !populated_zone(--zone))
3147 return alloc_flags;
3148
3149 alloc_flags |= ALLOC_NOFRAGMENT;
3150#endif /* CONFIG_ZONE_DMA32 */
3151 return alloc_flags;
3152}
3153
3154/* Must be called after current_gfp_context() which can change gfp_mask */
3155static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3156 unsigned int alloc_flags)
3157{
3158#ifdef CONFIG_CMA
3159 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3160 alloc_flags |= ALLOC_CMA;
3161#endif
3162 return alloc_flags;
3163}
3164
3165/*
3166 * get_page_from_freelist goes through the zonelist trying to allocate
3167 * a page.
3168 */
3169static struct page *
3170get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3171 const struct alloc_context *ac)
3172{
3173 struct zoneref *z;
3174 struct zone *zone;
3175 struct pglist_data *last_pgdat = NULL;
3176 bool last_pgdat_dirty_ok = false;
3177 bool no_fallback;
3178
3179retry:
3180 /*
3181 * Scan zonelist, looking for a zone with enough free.
3182 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3183 */
3184 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3185 z = ac->preferred_zoneref;
3186 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3187 ac->nodemask) {
3188 struct page *page;
3189 unsigned long mark;
3190
3191 if (cpusets_enabled() &&
3192 (alloc_flags & ALLOC_CPUSET) &&
3193 !__cpuset_zone_allowed(zone, gfp_mask))
3194 continue;
3195 /*
3196 * When allocating a page cache page for writing, we
3197 * want to get it from a node that is within its dirty
3198 * limit, such that no single node holds more than its
3199 * proportional share of globally allowed dirty pages.
3200 * The dirty limits take into account the node's
3201 * lowmem reserves and high watermark so that kswapd
3202 * should be able to balance it without having to
3203 * write pages from its LRU list.
3204 *
3205 * XXX: For now, allow allocations to potentially
3206 * exceed the per-node dirty limit in the slowpath
3207 * (spread_dirty_pages unset) before going into reclaim,
3208 * which is important when on a NUMA setup the allowed
3209 * nodes are together not big enough to reach the
3210 * global limit. The proper fix for these situations
3211 * will require awareness of nodes in the
3212 * dirty-throttling and the flusher threads.
3213 */
3214 if (ac->spread_dirty_pages) {
3215 if (last_pgdat != zone->zone_pgdat) {
3216 last_pgdat = zone->zone_pgdat;
3217 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3218 }
3219
3220 if (!last_pgdat_dirty_ok)
3221 continue;
3222 }
3223
3224 if (no_fallback && nr_online_nodes > 1 &&
3225 zone != ac->preferred_zoneref->zone) {
3226 int local_nid;
3227
3228 /*
3229 * If moving to a remote node, retry but allow
3230 * fragmenting fallbacks. Locality is more important
3231 * than fragmentation avoidance.
3232 */
3233 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3234 if (zone_to_nid(zone) != local_nid) {
3235 alloc_flags &= ~ALLOC_NOFRAGMENT;
3236 goto retry;
3237 }
3238 }
3239
3240 /*
3241 * Detect whether the number of free pages is below high
3242 * watermark. If so, we will decrease pcp->high and free
3243 * PCP pages in free path to reduce the possibility of
3244 * premature page reclaiming. Detection is done here to
3245 * avoid to do that in hotter free path.
3246 */
3247 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3248 goto check_alloc_wmark;
3249
3250 mark = high_wmark_pages(zone);
3251 if (zone_watermark_fast(zone, order, mark,
3252 ac->highest_zoneidx, alloc_flags,
3253 gfp_mask))
3254 goto try_this_zone;
3255 else
3256 set_bit(ZONE_BELOW_HIGH, &zone->flags);
3257
3258check_alloc_wmark:
3259 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3260 if (!zone_watermark_fast(zone, order, mark,
3261 ac->highest_zoneidx, alloc_flags,
3262 gfp_mask)) {
3263 int ret;
3264
3265 if (has_unaccepted_memory()) {
3266 if (try_to_accept_memory(zone, order))
3267 goto try_this_zone;
3268 }
3269
3270#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3271 /*
3272 * Watermark failed for this zone, but see if we can
3273 * grow this zone if it contains deferred pages.
3274 */
3275 if (deferred_pages_enabled()) {
3276 if (_deferred_grow_zone(zone, order))
3277 goto try_this_zone;
3278 }
3279#endif
3280 /* Checked here to keep the fast path fast */
3281 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3282 if (alloc_flags & ALLOC_NO_WATERMARKS)
3283 goto try_this_zone;
3284
3285 if (!node_reclaim_enabled() ||
3286 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3287 continue;
3288
3289 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3290 switch (ret) {
3291 case NODE_RECLAIM_NOSCAN:
3292 /* did not scan */
3293 continue;
3294 case NODE_RECLAIM_FULL:
3295 /* scanned but unreclaimable */
3296 continue;
3297 default:
3298 /* did we reclaim enough */
3299 if (zone_watermark_ok(zone, order, mark,
3300 ac->highest_zoneidx, alloc_flags))
3301 goto try_this_zone;
3302
3303 continue;
3304 }
3305 }
3306
3307try_this_zone:
3308 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3309 gfp_mask, alloc_flags, ac->migratetype);
3310 if (page) {
3311 prep_new_page(page, order, gfp_mask, alloc_flags);
3312
3313 /*
3314 * If this is a high-order atomic allocation then check
3315 * if the pageblock should be reserved for the future
3316 */
3317 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3318 reserve_highatomic_pageblock(page, zone);
3319
3320 return page;
3321 } else {
3322 if (has_unaccepted_memory()) {
3323 if (try_to_accept_memory(zone, order))
3324 goto try_this_zone;
3325 }
3326
3327#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3328 /* Try again if zone has deferred pages */
3329 if (deferred_pages_enabled()) {
3330 if (_deferred_grow_zone(zone, order))
3331 goto try_this_zone;
3332 }
3333#endif
3334 }
3335 }
3336
3337 /*
3338 * It's possible on a UMA machine to get through all zones that are
3339 * fragmented. If avoiding fragmentation, reset and try again.
3340 */
3341 if (no_fallback) {
3342 alloc_flags &= ~ALLOC_NOFRAGMENT;
3343 goto retry;
3344 }
3345
3346 return NULL;
3347}
3348
3349static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3350{
3351 unsigned int filter = SHOW_MEM_FILTER_NODES;
3352
3353 /*
3354 * This documents exceptions given to allocations in certain
3355 * contexts that are allowed to allocate outside current's set
3356 * of allowed nodes.
3357 */
3358 if (!(gfp_mask & __GFP_NOMEMALLOC))
3359 if (tsk_is_oom_victim(current) ||
3360 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3361 filter &= ~SHOW_MEM_FILTER_NODES;
3362 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3363 filter &= ~SHOW_MEM_FILTER_NODES;
3364
3365 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3366}
3367
3368void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3369{
3370 struct va_format vaf;
3371 va_list args;
3372 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3373
3374 if ((gfp_mask & __GFP_NOWARN) ||
3375 !__ratelimit(&nopage_rs) ||
3376 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3377 return;
3378
3379 va_start(args, fmt);
3380 vaf.fmt = fmt;
3381 vaf.va = &args;
3382 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3383 current->comm, &vaf, gfp_mask, &gfp_mask,
3384 nodemask_pr_args(nodemask));
3385 va_end(args);
3386
3387 cpuset_print_current_mems_allowed();
3388 pr_cont("\n");
3389 dump_stack();
3390 warn_alloc_show_mem(gfp_mask, nodemask);
3391}
3392
3393static inline struct page *
3394__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3395 unsigned int alloc_flags,
3396 const struct alloc_context *ac)
3397{
3398 struct page *page;
3399
3400 page = get_page_from_freelist(gfp_mask, order,
3401 alloc_flags|ALLOC_CPUSET, ac);
3402 /*
3403 * fallback to ignore cpuset restriction if our nodes
3404 * are depleted
3405 */
3406 if (!page)
3407 page = get_page_from_freelist(gfp_mask, order,
3408 alloc_flags, ac);
3409
3410 return page;
3411}
3412
3413static inline struct page *
3414__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3415 const struct alloc_context *ac, unsigned long *did_some_progress)
3416{
3417 struct oom_control oc = {
3418 .zonelist = ac->zonelist,
3419 .nodemask = ac->nodemask,
3420 .memcg = NULL,
3421 .gfp_mask = gfp_mask,
3422 .order = order,
3423 };
3424 struct page *page;
3425
3426 *did_some_progress = 0;
3427
3428 /*
3429 * Acquire the oom lock. If that fails, somebody else is
3430 * making progress for us.
3431 */
3432 if (!mutex_trylock(&oom_lock)) {
3433 *did_some_progress = 1;
3434 schedule_timeout_uninterruptible(1);
3435 return NULL;
3436 }
3437
3438 /*
3439 * Go through the zonelist yet one more time, keep very high watermark
3440 * here, this is only to catch a parallel oom killing, we must fail if
3441 * we're still under heavy pressure. But make sure that this reclaim
3442 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3443 * allocation which will never fail due to oom_lock already held.
3444 */
3445 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3446 ~__GFP_DIRECT_RECLAIM, order,
3447 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3448 if (page)
3449 goto out;
3450
3451 /* Coredumps can quickly deplete all memory reserves */
3452 if (current->flags & PF_DUMPCORE)
3453 goto out;
3454 /* The OOM killer will not help higher order allocs */
3455 if (order > PAGE_ALLOC_COSTLY_ORDER)
3456 goto out;
3457 /*
3458 * We have already exhausted all our reclaim opportunities without any
3459 * success so it is time to admit defeat. We will skip the OOM killer
3460 * because it is very likely that the caller has a more reasonable
3461 * fallback than shooting a random task.
3462 *
3463 * The OOM killer may not free memory on a specific node.
3464 */
3465 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3466 goto out;
3467 /* The OOM killer does not needlessly kill tasks for lowmem */
3468 if (ac->highest_zoneidx < ZONE_NORMAL)
3469 goto out;
3470 if (pm_suspended_storage())
3471 goto out;
3472 /*
3473 * XXX: GFP_NOFS allocations should rather fail than rely on
3474 * other request to make a forward progress.
3475 * We are in an unfortunate situation where out_of_memory cannot
3476 * do much for this context but let's try it to at least get
3477 * access to memory reserved if the current task is killed (see
3478 * out_of_memory). Once filesystems are ready to handle allocation
3479 * failures more gracefully we should just bail out here.
3480 */
3481
3482 /* Exhausted what can be done so it's blame time */
3483 if (out_of_memory(&oc) ||
3484 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3485 *did_some_progress = 1;
3486
3487 /*
3488 * Help non-failing allocations by giving them access to memory
3489 * reserves
3490 */
3491 if (gfp_mask & __GFP_NOFAIL)
3492 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3493 ALLOC_NO_WATERMARKS, ac);
3494 }
3495out:
3496 mutex_unlock(&oom_lock);
3497 return page;
3498}
3499
3500/*
3501 * Maximum number of compaction retries with a progress before OOM
3502 * killer is consider as the only way to move forward.
3503 */
3504#define MAX_COMPACT_RETRIES 16
3505
3506#ifdef CONFIG_COMPACTION
3507/* Try memory compaction for high-order allocations before reclaim */
3508static struct page *
3509__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3510 unsigned int alloc_flags, const struct alloc_context *ac,
3511 enum compact_priority prio, enum compact_result *compact_result)
3512{
3513 struct page *page = NULL;
3514 unsigned long pflags;
3515 unsigned int noreclaim_flag;
3516
3517 if (!order)
3518 return NULL;
3519
3520 psi_memstall_enter(&pflags);
3521 delayacct_compact_start();
3522 noreclaim_flag = memalloc_noreclaim_save();
3523
3524 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3525 prio, &page);
3526
3527 memalloc_noreclaim_restore(noreclaim_flag);
3528 psi_memstall_leave(&pflags);
3529 delayacct_compact_end();
3530
3531 if (*compact_result == COMPACT_SKIPPED)
3532 return NULL;
3533 /*
3534 * At least in one zone compaction wasn't deferred or skipped, so let's
3535 * count a compaction stall
3536 */
3537 count_vm_event(COMPACTSTALL);
3538
3539 /* Prep a captured page if available */
3540 if (page)
3541 prep_new_page(page, order, gfp_mask, alloc_flags);
3542
3543 /* Try get a page from the freelist if available */
3544 if (!page)
3545 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3546
3547 if (page) {
3548 struct zone *zone = page_zone(page);
3549
3550 zone->compact_blockskip_flush = false;
3551 compaction_defer_reset(zone, order, true);
3552 count_vm_event(COMPACTSUCCESS);
3553 return page;
3554 }
3555
3556 /*
3557 * It's bad if compaction run occurs and fails. The most likely reason
3558 * is that pages exist, but not enough to satisfy watermarks.
3559 */
3560 count_vm_event(COMPACTFAIL);
3561
3562 cond_resched();
3563
3564 return NULL;
3565}
3566
3567static inline bool
3568should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3569 enum compact_result compact_result,
3570 enum compact_priority *compact_priority,
3571 int *compaction_retries)
3572{
3573 int max_retries = MAX_COMPACT_RETRIES;
3574 int min_priority;
3575 bool ret = false;
3576 int retries = *compaction_retries;
3577 enum compact_priority priority = *compact_priority;
3578
3579 if (!order)
3580 return false;
3581
3582 if (fatal_signal_pending(current))
3583 return false;
3584
3585 /*
3586 * Compaction was skipped due to a lack of free order-0
3587 * migration targets. Continue if reclaim can help.
3588 */
3589 if (compact_result == COMPACT_SKIPPED) {
3590 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3591 goto out;
3592 }
3593
3594 /*
3595 * Compaction managed to coalesce some page blocks, but the
3596 * allocation failed presumably due to a race. Retry some.
3597 */
3598 if (compact_result == COMPACT_SUCCESS) {
3599 /*
3600 * !costly requests are much more important than
3601 * __GFP_RETRY_MAYFAIL costly ones because they are de
3602 * facto nofail and invoke OOM killer to move on while
3603 * costly can fail and users are ready to cope with
3604 * that. 1/4 retries is rather arbitrary but we would
3605 * need much more detailed feedback from compaction to
3606 * make a better decision.
3607 */
3608 if (order > PAGE_ALLOC_COSTLY_ORDER)
3609 max_retries /= 4;
3610
3611 if (++(*compaction_retries) <= max_retries) {
3612 ret = true;
3613 goto out;
3614 }
3615 }
3616
3617 /*
3618 * Compaction failed. Retry with increasing priority.
3619 */
3620 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3621 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3622
3623 if (*compact_priority > min_priority) {
3624 (*compact_priority)--;
3625 *compaction_retries = 0;
3626 ret = true;
3627 }
3628out:
3629 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3630 return ret;
3631}
3632#else
3633static inline struct page *
3634__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3635 unsigned int alloc_flags, const struct alloc_context *ac,
3636 enum compact_priority prio, enum compact_result *compact_result)
3637{
3638 *compact_result = COMPACT_SKIPPED;
3639 return NULL;
3640}
3641
3642static inline bool
3643should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3644 enum compact_result compact_result,
3645 enum compact_priority *compact_priority,
3646 int *compaction_retries)
3647{
3648 struct zone *zone;
3649 struct zoneref *z;
3650
3651 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3652 return false;
3653
3654 /*
3655 * There are setups with compaction disabled which would prefer to loop
3656 * inside the allocator rather than hit the oom killer prematurely.
3657 * Let's give them a good hope and keep retrying while the order-0
3658 * watermarks are OK.
3659 */
3660 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3661 ac->highest_zoneidx, ac->nodemask) {
3662 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3663 ac->highest_zoneidx, alloc_flags))
3664 return true;
3665 }
3666 return false;
3667}
3668#endif /* CONFIG_COMPACTION */
3669
3670#ifdef CONFIG_LOCKDEP
3671static struct lockdep_map __fs_reclaim_map =
3672 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3673
3674static bool __need_reclaim(gfp_t gfp_mask)
3675{
3676 /* no reclaim without waiting on it */
3677 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3678 return false;
3679
3680 /* this guy won't enter reclaim */
3681 if (current->flags & PF_MEMALLOC)
3682 return false;
3683
3684 if (gfp_mask & __GFP_NOLOCKDEP)
3685 return false;
3686
3687 return true;
3688}
3689
3690void __fs_reclaim_acquire(unsigned long ip)
3691{
3692 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3693}
3694
3695void __fs_reclaim_release(unsigned long ip)
3696{
3697 lock_release(&__fs_reclaim_map, ip);
3698}
3699
3700void fs_reclaim_acquire(gfp_t gfp_mask)
3701{
3702 gfp_mask = current_gfp_context(gfp_mask);
3703
3704 if (__need_reclaim(gfp_mask)) {
3705 if (gfp_mask & __GFP_FS)
3706 __fs_reclaim_acquire(_RET_IP_);
3707
3708#ifdef CONFIG_MMU_NOTIFIER
3709 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3710 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3711#endif
3712
3713 }
3714}
3715EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3716
3717void fs_reclaim_release(gfp_t gfp_mask)
3718{
3719 gfp_mask = current_gfp_context(gfp_mask);
3720
3721 if (__need_reclaim(gfp_mask)) {
3722 if (gfp_mask & __GFP_FS)
3723 __fs_reclaim_release(_RET_IP_);
3724 }
3725}
3726EXPORT_SYMBOL_GPL(fs_reclaim_release);
3727#endif
3728
3729/*
3730 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3731 * have been rebuilt so allocation retries. Reader side does not lock and
3732 * retries the allocation if zonelist changes. Writer side is protected by the
3733 * embedded spin_lock.
3734 */
3735static DEFINE_SEQLOCK(zonelist_update_seq);
3736
3737static unsigned int zonelist_iter_begin(void)
3738{
3739 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3740 return read_seqbegin(&zonelist_update_seq);
3741
3742 return 0;
3743}
3744
3745static unsigned int check_retry_zonelist(unsigned int seq)
3746{
3747 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3748 return read_seqretry(&zonelist_update_seq, seq);
3749
3750 return seq;
3751}
3752
3753/* Perform direct synchronous page reclaim */
3754static unsigned long
3755__perform_reclaim(gfp_t gfp_mask, unsigned int order,
3756 const struct alloc_context *ac)
3757{
3758 unsigned int noreclaim_flag;
3759 unsigned long progress;
3760
3761 cond_resched();
3762
3763 /* We now go into synchronous reclaim */
3764 cpuset_memory_pressure_bump();
3765 fs_reclaim_acquire(gfp_mask);
3766 noreclaim_flag = memalloc_noreclaim_save();
3767
3768 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3769 ac->nodemask);
3770
3771 memalloc_noreclaim_restore(noreclaim_flag);
3772 fs_reclaim_release(gfp_mask);
3773
3774 cond_resched();
3775
3776 return progress;
3777}
3778
3779/* The really slow allocator path where we enter direct reclaim */
3780static inline struct page *
3781__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3782 unsigned int alloc_flags, const struct alloc_context *ac,
3783 unsigned long *did_some_progress)
3784{
3785 struct page *page = NULL;
3786 unsigned long pflags;
3787 bool drained = false;
3788
3789 psi_memstall_enter(&pflags);
3790 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3791 if (unlikely(!(*did_some_progress)))
3792 goto out;
3793
3794retry:
3795 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3796
3797 /*
3798 * If an allocation failed after direct reclaim, it could be because
3799 * pages are pinned on the per-cpu lists or in high alloc reserves.
3800 * Shrink them and try again
3801 */
3802 if (!page && !drained) {
3803 unreserve_highatomic_pageblock(ac, false);
3804 drain_all_pages(NULL);
3805 drained = true;
3806 goto retry;
3807 }
3808out:
3809 psi_memstall_leave(&pflags);
3810
3811 return page;
3812}
3813
3814static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3815 const struct alloc_context *ac)
3816{
3817 struct zoneref *z;
3818 struct zone *zone;
3819 pg_data_t *last_pgdat = NULL;
3820 enum zone_type highest_zoneidx = ac->highest_zoneidx;
3821
3822 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3823 ac->nodemask) {
3824 if (!managed_zone(zone))
3825 continue;
3826 if (last_pgdat != zone->zone_pgdat) {
3827 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3828 last_pgdat = zone->zone_pgdat;
3829 }
3830 }
3831}
3832
3833static inline unsigned int
3834gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3835{
3836 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3837
3838 /*
3839 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3840 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3841 * to save two branches.
3842 */
3843 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
3844 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
3845
3846 /*
3847 * The caller may dip into page reserves a bit more if the caller
3848 * cannot run direct reclaim, or if the caller has realtime scheduling
3849 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3850 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3851 */
3852 alloc_flags |= (__force int)
3853 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
3854
3855 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
3856 /*
3857 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3858 * if it can't schedule.
3859 */
3860 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
3861 alloc_flags |= ALLOC_NON_BLOCK;
3862
3863 if (order > 0)
3864 alloc_flags |= ALLOC_HIGHATOMIC;
3865 }
3866
3867 /*
3868 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
3869 * GFP_ATOMIC) rather than fail, see the comment for
3870 * cpuset_node_allowed().
3871 */
3872 if (alloc_flags & ALLOC_MIN_RESERVE)
3873 alloc_flags &= ~ALLOC_CPUSET;
3874 } else if (unlikely(rt_task(current)) && in_task())
3875 alloc_flags |= ALLOC_MIN_RESERVE;
3876
3877 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
3878
3879 return alloc_flags;
3880}
3881
3882static bool oom_reserves_allowed(struct task_struct *tsk)
3883{
3884 if (!tsk_is_oom_victim(tsk))
3885 return false;
3886
3887 /*
3888 * !MMU doesn't have oom reaper so give access to memory reserves
3889 * only to the thread with TIF_MEMDIE set
3890 */
3891 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3892 return false;
3893
3894 return true;
3895}
3896
3897/*
3898 * Distinguish requests which really need access to full memory
3899 * reserves from oom victims which can live with a portion of it
3900 */
3901static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3902{
3903 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3904 return 0;
3905 if (gfp_mask & __GFP_MEMALLOC)
3906 return ALLOC_NO_WATERMARKS;
3907 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3908 return ALLOC_NO_WATERMARKS;
3909 if (!in_interrupt()) {
3910 if (current->flags & PF_MEMALLOC)
3911 return ALLOC_NO_WATERMARKS;
3912 else if (oom_reserves_allowed(current))
3913 return ALLOC_OOM;
3914 }
3915
3916 return 0;
3917}
3918
3919bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3920{
3921 return !!__gfp_pfmemalloc_flags(gfp_mask);
3922}
3923
3924/*
3925 * Checks whether it makes sense to retry the reclaim to make a forward progress
3926 * for the given allocation request.
3927 *
3928 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3929 * without success, or when we couldn't even meet the watermark if we
3930 * reclaimed all remaining pages on the LRU lists.
3931 *
3932 * Returns true if a retry is viable or false to enter the oom path.
3933 */
3934static inline bool
3935should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3936 struct alloc_context *ac, int alloc_flags,
3937 bool did_some_progress, int *no_progress_loops)
3938{
3939 struct zone *zone;
3940 struct zoneref *z;
3941 bool ret = false;
3942
3943 /*
3944 * Costly allocations might have made a progress but this doesn't mean
3945 * their order will become available due to high fragmentation so
3946 * always increment the no progress counter for them
3947 */
3948 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3949 *no_progress_loops = 0;
3950 else
3951 (*no_progress_loops)++;
3952
3953 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
3954 goto out;
3955
3956
3957 /*
3958 * Keep reclaiming pages while there is a chance this will lead
3959 * somewhere. If none of the target zones can satisfy our allocation
3960 * request even if all reclaimable pages are considered then we are
3961 * screwed and have to go OOM.
3962 */
3963 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3964 ac->highest_zoneidx, ac->nodemask) {
3965 unsigned long available;
3966 unsigned long reclaimable;
3967 unsigned long min_wmark = min_wmark_pages(zone);
3968 bool wmark;
3969
3970 available = reclaimable = zone_reclaimable_pages(zone);
3971 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3972
3973 /*
3974 * Would the allocation succeed if we reclaimed all
3975 * reclaimable pages?
3976 */
3977 wmark = __zone_watermark_ok(zone, order, min_wmark,
3978 ac->highest_zoneidx, alloc_flags, available);
3979 trace_reclaim_retry_zone(z, order, reclaimable,
3980 available, min_wmark, *no_progress_loops, wmark);
3981 if (wmark) {
3982 ret = true;
3983 break;
3984 }
3985 }
3986
3987 /*
3988 * Memory allocation/reclaim might be called from a WQ context and the
3989 * current implementation of the WQ concurrency control doesn't
3990 * recognize that a particular WQ is congested if the worker thread is
3991 * looping without ever sleeping. Therefore we have to do a short sleep
3992 * here rather than calling cond_resched().
3993 */
3994 if (current->flags & PF_WQ_WORKER)
3995 schedule_timeout_uninterruptible(1);
3996 else
3997 cond_resched();
3998out:
3999 /* Before OOM, exhaust highatomic_reserve */
4000 if (!ret)
4001 return unreserve_highatomic_pageblock(ac, true);
4002
4003 return ret;
4004}
4005
4006static inline bool
4007check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4008{
4009 /*
4010 * It's possible that cpuset's mems_allowed and the nodemask from
4011 * mempolicy don't intersect. This should be normally dealt with by
4012 * policy_nodemask(), but it's possible to race with cpuset update in
4013 * such a way the check therein was true, and then it became false
4014 * before we got our cpuset_mems_cookie here.
4015 * This assumes that for all allocations, ac->nodemask can come only
4016 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4017 * when it does not intersect with the cpuset restrictions) or the
4018 * caller can deal with a violated nodemask.
4019 */
4020 if (cpusets_enabled() && ac->nodemask &&
4021 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4022 ac->nodemask = NULL;
4023 return true;
4024 }
4025
4026 /*
4027 * When updating a task's mems_allowed or mempolicy nodemask, it is
4028 * possible to race with parallel threads in such a way that our
4029 * allocation can fail while the mask is being updated. If we are about
4030 * to fail, check if the cpuset changed during allocation and if so,
4031 * retry.
4032 */
4033 if (read_mems_allowed_retry(cpuset_mems_cookie))
4034 return true;
4035
4036 return false;
4037}
4038
4039static inline struct page *
4040__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4041 struct alloc_context *ac)
4042{
4043 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4044 bool can_compact = gfp_compaction_allowed(gfp_mask);
4045 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4046 struct page *page = NULL;
4047 unsigned int alloc_flags;
4048 unsigned long did_some_progress;
4049 enum compact_priority compact_priority;
4050 enum compact_result compact_result;
4051 int compaction_retries;
4052 int no_progress_loops;
4053 unsigned int cpuset_mems_cookie;
4054 unsigned int zonelist_iter_cookie;
4055 int reserve_flags;
4056
4057restart:
4058 compaction_retries = 0;
4059 no_progress_loops = 0;
4060 compact_priority = DEF_COMPACT_PRIORITY;
4061 cpuset_mems_cookie = read_mems_allowed_begin();
4062 zonelist_iter_cookie = zonelist_iter_begin();
4063
4064 /*
4065 * The fast path uses conservative alloc_flags to succeed only until
4066 * kswapd needs to be woken up, and to avoid the cost of setting up
4067 * alloc_flags precisely. So we do that now.
4068 */
4069 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4070
4071 /*
4072 * We need to recalculate the starting point for the zonelist iterator
4073 * because we might have used different nodemask in the fast path, or
4074 * there was a cpuset modification and we are retrying - otherwise we
4075 * could end up iterating over non-eligible zones endlessly.
4076 */
4077 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4078 ac->highest_zoneidx, ac->nodemask);
4079 if (!ac->preferred_zoneref->zone)
4080 goto nopage;
4081
4082 /*
4083 * Check for insane configurations where the cpuset doesn't contain
4084 * any suitable zone to satisfy the request - e.g. non-movable
4085 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4086 */
4087 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4088 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4089 ac->highest_zoneidx,
4090 &cpuset_current_mems_allowed);
4091 if (!z->zone)
4092 goto nopage;
4093 }
4094
4095 if (alloc_flags & ALLOC_KSWAPD)
4096 wake_all_kswapds(order, gfp_mask, ac);
4097
4098 /*
4099 * The adjusted alloc_flags might result in immediate success, so try
4100 * that first
4101 */
4102 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4103 if (page)
4104 goto got_pg;
4105
4106 /*
4107 * For costly allocations, try direct compaction first, as it's likely
4108 * that we have enough base pages and don't need to reclaim. For non-
4109 * movable high-order allocations, do that as well, as compaction will
4110 * try prevent permanent fragmentation by migrating from blocks of the
4111 * same migratetype.
4112 * Don't try this for allocations that are allowed to ignore
4113 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4114 */
4115 if (can_direct_reclaim && can_compact &&
4116 (costly_order ||
4117 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4118 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4119 page = __alloc_pages_direct_compact(gfp_mask, order,
4120 alloc_flags, ac,
4121 INIT_COMPACT_PRIORITY,
4122 &compact_result);
4123 if (page)
4124 goto got_pg;
4125
4126 /*
4127 * Checks for costly allocations with __GFP_NORETRY, which
4128 * includes some THP page fault allocations
4129 */
4130 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4131 /*
4132 * If allocating entire pageblock(s) and compaction
4133 * failed because all zones are below low watermarks
4134 * or is prohibited because it recently failed at this
4135 * order, fail immediately unless the allocator has
4136 * requested compaction and reclaim retry.
4137 *
4138 * Reclaim is
4139 * - potentially very expensive because zones are far
4140 * below their low watermarks or this is part of very
4141 * bursty high order allocations,
4142 * - not guaranteed to help because isolate_freepages()
4143 * may not iterate over freed pages as part of its
4144 * linear scan, and
4145 * - unlikely to make entire pageblocks free on its
4146 * own.
4147 */
4148 if (compact_result == COMPACT_SKIPPED ||
4149 compact_result == COMPACT_DEFERRED)
4150 goto nopage;
4151
4152 /*
4153 * Looks like reclaim/compaction is worth trying, but
4154 * sync compaction could be very expensive, so keep
4155 * using async compaction.
4156 */
4157 compact_priority = INIT_COMPACT_PRIORITY;
4158 }
4159 }
4160
4161retry:
4162 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4163 if (alloc_flags & ALLOC_KSWAPD)
4164 wake_all_kswapds(order, gfp_mask, ac);
4165
4166 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4167 if (reserve_flags)
4168 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4169 (alloc_flags & ALLOC_KSWAPD);
4170
4171 /*
4172 * Reset the nodemask and zonelist iterators if memory policies can be
4173 * ignored. These allocations are high priority and system rather than
4174 * user oriented.
4175 */
4176 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4177 ac->nodemask = NULL;
4178 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4179 ac->highest_zoneidx, ac->nodemask);
4180 }
4181
4182 /* Attempt with potentially adjusted zonelist and alloc_flags */
4183 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4184 if (page)
4185 goto got_pg;
4186
4187 /* Caller is not willing to reclaim, we can't balance anything */
4188 if (!can_direct_reclaim)
4189 goto nopage;
4190
4191 /* Avoid recursion of direct reclaim */
4192 if (current->flags & PF_MEMALLOC)
4193 goto nopage;
4194
4195 /* Try direct reclaim and then allocating */
4196 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4197 &did_some_progress);
4198 if (page)
4199 goto got_pg;
4200
4201 /* Try direct compaction and then allocating */
4202 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4203 compact_priority, &compact_result);
4204 if (page)
4205 goto got_pg;
4206
4207 /* Do not loop if specifically requested */
4208 if (gfp_mask & __GFP_NORETRY)
4209 goto nopage;
4210
4211 /*
4212 * Do not retry costly high order allocations unless they are
4213 * __GFP_RETRY_MAYFAIL and we can compact
4214 */
4215 if (costly_order && (!can_compact ||
4216 !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4217 goto nopage;
4218
4219 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4220 did_some_progress > 0, &no_progress_loops))
4221 goto retry;
4222
4223 /*
4224 * It doesn't make any sense to retry for the compaction if the order-0
4225 * reclaim is not able to make any progress because the current
4226 * implementation of the compaction depends on the sufficient amount
4227 * of free memory (see __compaction_suitable)
4228 */
4229 if (did_some_progress > 0 && can_compact &&
4230 should_compact_retry(ac, order, alloc_flags,
4231 compact_result, &compact_priority,
4232 &compaction_retries))
4233 goto retry;
4234
4235
4236 /*
4237 * Deal with possible cpuset update races or zonelist updates to avoid
4238 * a unnecessary OOM kill.
4239 */
4240 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4241 check_retry_zonelist(zonelist_iter_cookie))
4242 goto restart;
4243
4244 /* Reclaim has failed us, start killing things */
4245 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4246 if (page)
4247 goto got_pg;
4248
4249 /* Avoid allocations with no watermarks from looping endlessly */
4250 if (tsk_is_oom_victim(current) &&
4251 (alloc_flags & ALLOC_OOM ||
4252 (gfp_mask & __GFP_NOMEMALLOC)))
4253 goto nopage;
4254
4255 /* Retry as long as the OOM killer is making progress */
4256 if (did_some_progress) {
4257 no_progress_loops = 0;
4258 goto retry;
4259 }
4260
4261nopage:
4262 /*
4263 * Deal with possible cpuset update races or zonelist updates to avoid
4264 * a unnecessary OOM kill.
4265 */
4266 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4267 check_retry_zonelist(zonelist_iter_cookie))
4268 goto restart;
4269
4270 /*
4271 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4272 * we always retry
4273 */
4274 if (gfp_mask & __GFP_NOFAIL) {
4275 /*
4276 * All existing users of the __GFP_NOFAIL are blockable, so warn
4277 * of any new users that actually require GFP_NOWAIT
4278 */
4279 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4280 goto fail;
4281
4282 /*
4283 * PF_MEMALLOC request from this context is rather bizarre
4284 * because we cannot reclaim anything and only can loop waiting
4285 * for somebody to do a work for us
4286 */
4287 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4288
4289 /*
4290 * non failing costly orders are a hard requirement which we
4291 * are not prepared for much so let's warn about these users
4292 * so that we can identify them and convert them to something
4293 * else.
4294 */
4295 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4296
4297 /*
4298 * Help non-failing allocations by giving some access to memory
4299 * reserves normally used for high priority non-blocking
4300 * allocations but do not use ALLOC_NO_WATERMARKS because this
4301 * could deplete whole memory reserves which would just make
4302 * the situation worse.
4303 */
4304 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4305 if (page)
4306 goto got_pg;
4307
4308 cond_resched();
4309 goto retry;
4310 }
4311fail:
4312 warn_alloc(gfp_mask, ac->nodemask,
4313 "page allocation failure: order:%u", order);
4314got_pg:
4315 return page;
4316}
4317
4318static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4319 int preferred_nid, nodemask_t *nodemask,
4320 struct alloc_context *ac, gfp_t *alloc_gfp,
4321 unsigned int *alloc_flags)
4322{
4323 ac->highest_zoneidx = gfp_zone(gfp_mask);
4324 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4325 ac->nodemask = nodemask;
4326 ac->migratetype = gfp_migratetype(gfp_mask);
4327
4328 if (cpusets_enabled()) {
4329 *alloc_gfp |= __GFP_HARDWALL;
4330 /*
4331 * When we are in the interrupt context, it is irrelevant
4332 * to the current task context. It means that any node ok.
4333 */
4334 if (in_task() && !ac->nodemask)
4335 ac->nodemask = &cpuset_current_mems_allowed;
4336 else
4337 *alloc_flags |= ALLOC_CPUSET;
4338 }
4339
4340 might_alloc(gfp_mask);
4341
4342 if (should_fail_alloc_page(gfp_mask, order))
4343 return false;
4344
4345 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4346
4347 /* Dirty zone balancing only done in the fast path */
4348 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4349
4350 /*
4351 * The preferred zone is used for statistics but crucially it is
4352 * also used as the starting point for the zonelist iterator. It
4353 * may get reset for allocations that ignore memory policies.
4354 */
4355 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4356 ac->highest_zoneidx, ac->nodemask);
4357
4358 return true;
4359}
4360
4361/*
4362 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4363 * @gfp: GFP flags for the allocation
4364 * @preferred_nid: The preferred NUMA node ID to allocate from
4365 * @nodemask: Set of nodes to allocate from, may be NULL
4366 * @nr_pages: The number of pages desired on the list or array
4367 * @page_list: Optional list to store the allocated pages
4368 * @page_array: Optional array to store the pages
4369 *
4370 * This is a batched version of the page allocator that attempts to
4371 * allocate nr_pages quickly. Pages are added to page_list if page_list
4372 * is not NULL, otherwise it is assumed that the page_array is valid.
4373 *
4374 * For lists, nr_pages is the number of pages that should be allocated.
4375 *
4376 * For arrays, only NULL elements are populated with pages and nr_pages
4377 * is the maximum number of pages that will be stored in the array.
4378 *
4379 * Returns the number of pages on the list or array.
4380 */
4381unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
4382 nodemask_t *nodemask, int nr_pages,
4383 struct list_head *page_list,
4384 struct page **page_array)
4385{
4386 struct page *page;
4387 unsigned long __maybe_unused UP_flags;
4388 struct zone *zone;
4389 struct zoneref *z;
4390 struct per_cpu_pages *pcp;
4391 struct list_head *pcp_list;
4392 struct alloc_context ac;
4393 gfp_t alloc_gfp;
4394 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4395 int nr_populated = 0, nr_account = 0;
4396
4397 /*
4398 * Skip populated array elements to determine if any pages need
4399 * to be allocated before disabling IRQs.
4400 */
4401 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4402 nr_populated++;
4403
4404 /* No pages requested? */
4405 if (unlikely(nr_pages <= 0))
4406 goto out;
4407
4408 /* Already populated array? */
4409 if (unlikely(page_array && nr_pages - nr_populated == 0))
4410 goto out;
4411
4412 /* Bulk allocator does not support memcg accounting. */
4413 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4414 goto failed;
4415
4416 /* Use the single page allocator for one page. */
4417 if (nr_pages - nr_populated == 1)
4418 goto failed;
4419
4420#ifdef CONFIG_PAGE_OWNER
4421 /*
4422 * PAGE_OWNER may recurse into the allocator to allocate space to
4423 * save the stack with pagesets.lock held. Releasing/reacquiring
4424 * removes much of the performance benefit of bulk allocation so
4425 * force the caller to allocate one page at a time as it'll have
4426 * similar performance to added complexity to the bulk allocator.
4427 */
4428 if (static_branch_unlikely(&page_owner_inited))
4429 goto failed;
4430#endif
4431
4432 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4433 gfp &= gfp_allowed_mask;
4434 alloc_gfp = gfp;
4435 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4436 goto out;
4437 gfp = alloc_gfp;
4438
4439 /* Find an allowed local zone that meets the low watermark. */
4440 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4441 unsigned long mark;
4442
4443 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4444 !__cpuset_zone_allowed(zone, gfp)) {
4445 continue;
4446 }
4447
4448 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4449 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4450 goto failed;
4451 }
4452
4453 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4454 if (zone_watermark_fast(zone, 0, mark,
4455 zonelist_zone_idx(ac.preferred_zoneref),
4456 alloc_flags, gfp)) {
4457 break;
4458 }
4459 }
4460
4461 /*
4462 * If there are no allowed local zones that meets the watermarks then
4463 * try to allocate a single page and reclaim if necessary.
4464 */
4465 if (unlikely(!zone))
4466 goto failed;
4467
4468 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4469 pcp_trylock_prepare(UP_flags);
4470 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4471 if (!pcp)
4472 goto failed_irq;
4473
4474 /* Attempt the batch allocation */
4475 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4476 while (nr_populated < nr_pages) {
4477
4478 /* Skip existing pages */
4479 if (page_array && page_array[nr_populated]) {
4480 nr_populated++;
4481 continue;
4482 }
4483
4484 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4485 pcp, pcp_list);
4486 if (unlikely(!page)) {
4487 /* Try and allocate at least one page */
4488 if (!nr_account) {
4489 pcp_spin_unlock(pcp);
4490 goto failed_irq;
4491 }
4492 break;
4493 }
4494 nr_account++;
4495
4496 prep_new_page(page, 0, gfp, 0);
4497 if (page_list)
4498 list_add(&page->lru, page_list);
4499 else
4500 page_array[nr_populated] = page;
4501 nr_populated++;
4502 }
4503
4504 pcp_spin_unlock(pcp);
4505 pcp_trylock_finish(UP_flags);
4506
4507 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4508 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4509
4510out:
4511 return nr_populated;
4512
4513failed_irq:
4514 pcp_trylock_finish(UP_flags);
4515
4516failed:
4517 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
4518 if (page) {
4519 if (page_list)
4520 list_add(&page->lru, page_list);
4521 else
4522 page_array[nr_populated] = page;
4523 nr_populated++;
4524 }
4525
4526 goto out;
4527}
4528EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
4529
4530/*
4531 * This is the 'heart' of the zoned buddy allocator.
4532 */
4533struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
4534 nodemask_t *nodemask)
4535{
4536 struct page *page;
4537 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4538 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4539 struct alloc_context ac = { };
4540
4541 /*
4542 * There are several places where we assume that the order value is sane
4543 * so bail out early if the request is out of bound.
4544 */
4545 if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
4546 return NULL;
4547
4548 gfp &= gfp_allowed_mask;
4549 /*
4550 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4551 * resp. GFP_NOIO which has to be inherited for all allocation requests
4552 * from a particular context which has been marked by
4553 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4554 * movable zones are not used during allocation.
4555 */
4556 gfp = current_gfp_context(gfp);
4557 alloc_gfp = gfp;
4558 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4559 &alloc_gfp, &alloc_flags))
4560 return NULL;
4561
4562 /*
4563 * Forbid the first pass from falling back to types that fragment
4564 * memory until all local zones are considered.
4565 */
4566 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4567
4568 /* First allocation attempt */
4569 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4570 if (likely(page))
4571 goto out;
4572
4573 alloc_gfp = gfp;
4574 ac.spread_dirty_pages = false;
4575
4576 /*
4577 * Restore the original nodemask if it was potentially replaced with
4578 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4579 */
4580 ac.nodemask = nodemask;
4581
4582 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4583
4584out:
4585 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4586 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4587 __free_pages(page, order);
4588 page = NULL;
4589 }
4590
4591 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4592 kmsan_alloc_page(page, order, alloc_gfp);
4593
4594 return page;
4595}
4596EXPORT_SYMBOL(__alloc_pages);
4597
4598struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
4599 nodemask_t *nodemask)
4600{
4601 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
4602 preferred_nid, nodemask);
4603 return page_rmappable_folio(page);
4604}
4605EXPORT_SYMBOL(__folio_alloc);
4606
4607/*
4608 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4609 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4610 * you need to access high mem.
4611 */
4612unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4613{
4614 struct page *page;
4615
4616 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4617 if (!page)
4618 return 0;
4619 return (unsigned long) page_address(page);
4620}
4621EXPORT_SYMBOL(__get_free_pages);
4622
4623unsigned long get_zeroed_page(gfp_t gfp_mask)
4624{
4625 return __get_free_page(gfp_mask | __GFP_ZERO);
4626}
4627EXPORT_SYMBOL(get_zeroed_page);
4628
4629/**
4630 * __free_pages - Free pages allocated with alloc_pages().
4631 * @page: The page pointer returned from alloc_pages().
4632 * @order: The order of the allocation.
4633 *
4634 * This function can free multi-page allocations that are not compound
4635 * pages. It does not check that the @order passed in matches that of
4636 * the allocation, so it is easy to leak memory. Freeing more memory
4637 * than was allocated will probably emit a warning.
4638 *
4639 * If the last reference to this page is speculative, it will be released
4640 * by put_page() which only frees the first page of a non-compound
4641 * allocation. To prevent the remaining pages from being leaked, we free
4642 * the subsequent pages here. If you want to use the page's reference
4643 * count to decide when to free the allocation, you should allocate a
4644 * compound page, and use put_page() instead of __free_pages().
4645 *
4646 * Context: May be called in interrupt context or while holding a normal
4647 * spinlock, but not in NMI context or while holding a raw spinlock.
4648 */
4649void __free_pages(struct page *page, unsigned int order)
4650{
4651 /* get PageHead before we drop reference */
4652 int head = PageHead(page);
4653
4654 if (put_page_testzero(page))
4655 free_the_page(page, order);
4656 else if (!head)
4657 while (order-- > 0)
4658 free_the_page(page + (1 << order), order);
4659}
4660EXPORT_SYMBOL(__free_pages);
4661
4662void free_pages(unsigned long addr, unsigned int order)
4663{
4664 if (addr != 0) {
4665 VM_BUG_ON(!virt_addr_valid((void *)addr));
4666 __free_pages(virt_to_page((void *)addr), order);
4667 }
4668}
4669
4670EXPORT_SYMBOL(free_pages);
4671
4672/*
4673 * Page Fragment:
4674 * An arbitrary-length arbitrary-offset area of memory which resides
4675 * within a 0 or higher order page. Multiple fragments within that page
4676 * are individually refcounted, in the page's reference counter.
4677 *
4678 * The page_frag functions below provide a simple allocation framework for
4679 * page fragments. This is used by the network stack and network device
4680 * drivers to provide a backing region of memory for use as either an
4681 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4682 */
4683static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4684 gfp_t gfp_mask)
4685{
4686 struct page *page = NULL;
4687 gfp_t gfp = gfp_mask;
4688
4689#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4690 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4691 __GFP_NOMEMALLOC;
4692 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4693 PAGE_FRAG_CACHE_MAX_ORDER);
4694 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4695#endif
4696 if (unlikely(!page))
4697 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4698
4699 nc->va = page ? page_address(page) : NULL;
4700
4701 return page;
4702}
4703
4704void __page_frag_cache_drain(struct page *page, unsigned int count)
4705{
4706 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4707
4708 if (page_ref_sub_and_test(page, count))
4709 free_the_page(page, compound_order(page));
4710}
4711EXPORT_SYMBOL(__page_frag_cache_drain);
4712
4713void *page_frag_alloc_align(struct page_frag_cache *nc,
4714 unsigned int fragsz, gfp_t gfp_mask,
4715 unsigned int align_mask)
4716{
4717 unsigned int size = PAGE_SIZE;
4718 struct page *page;
4719 int offset;
4720
4721 if (unlikely(!nc->va)) {
4722refill:
4723 page = __page_frag_cache_refill(nc, gfp_mask);
4724 if (!page)
4725 return NULL;
4726
4727#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4728 /* if size can vary use size else just use PAGE_SIZE */
4729 size = nc->size;
4730#endif
4731 /* Even if we own the page, we do not use atomic_set().
4732 * This would break get_page_unless_zero() users.
4733 */
4734 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4735
4736 /* reset page count bias and offset to start of new frag */
4737 nc->pfmemalloc = page_is_pfmemalloc(page);
4738 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4739 nc->offset = size;
4740 }
4741
4742 offset = nc->offset - fragsz;
4743 if (unlikely(offset < 0)) {
4744 page = virt_to_page(nc->va);
4745
4746 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4747 goto refill;
4748
4749 if (unlikely(nc->pfmemalloc)) {
4750 free_the_page(page, compound_order(page));
4751 goto refill;
4752 }
4753
4754#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4755 /* if size can vary use size else just use PAGE_SIZE */
4756 size = nc->size;
4757#endif
4758 /* OK, page count is 0, we can safely set it */
4759 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4760
4761 /* reset page count bias and offset to start of new frag */
4762 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4763 offset = size - fragsz;
4764 if (unlikely(offset < 0)) {
4765 /*
4766 * The caller is trying to allocate a fragment
4767 * with fragsz > PAGE_SIZE but the cache isn't big
4768 * enough to satisfy the request, this may
4769 * happen in low memory conditions.
4770 * We don't release the cache page because
4771 * it could make memory pressure worse
4772 * so we simply return NULL here.
4773 */
4774 return NULL;
4775 }
4776 }
4777
4778 nc->pagecnt_bias--;
4779 offset &= align_mask;
4780 nc->offset = offset;
4781
4782 return nc->va + offset;
4783}
4784EXPORT_SYMBOL(page_frag_alloc_align);
4785
4786/*
4787 * Frees a page fragment allocated out of either a compound or order 0 page.
4788 */
4789void page_frag_free(void *addr)
4790{
4791 struct page *page = virt_to_head_page(addr);
4792
4793 if (unlikely(put_page_testzero(page)))
4794 free_the_page(page, compound_order(page));
4795}
4796EXPORT_SYMBOL(page_frag_free);
4797
4798static void *make_alloc_exact(unsigned long addr, unsigned int order,
4799 size_t size)
4800{
4801 if (addr) {
4802 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4803 struct page *page = virt_to_page((void *)addr);
4804 struct page *last = page + nr;
4805
4806 split_page_owner(page, 1 << order);
4807 split_page_memcg(page, 1 << order);
4808 while (page < --last)
4809 set_page_refcounted(last);
4810
4811 last = page + (1UL << order);
4812 for (page += nr; page < last; page++)
4813 __free_pages_ok(page, 0, FPI_TO_TAIL);
4814 }
4815 return (void *)addr;
4816}
4817
4818/**
4819 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4820 * @size: the number of bytes to allocate
4821 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4822 *
4823 * This function is similar to alloc_pages(), except that it allocates the
4824 * minimum number of pages to satisfy the request. alloc_pages() can only
4825 * allocate memory in power-of-two pages.
4826 *
4827 * This function is also limited by MAX_PAGE_ORDER.
4828 *
4829 * Memory allocated by this function must be released by free_pages_exact().
4830 *
4831 * Return: pointer to the allocated area or %NULL in case of error.
4832 */
4833void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4834{
4835 unsigned int order = get_order(size);
4836 unsigned long addr;
4837
4838 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4839 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4840
4841 addr = __get_free_pages(gfp_mask, order);
4842 return make_alloc_exact(addr, order, size);
4843}
4844EXPORT_SYMBOL(alloc_pages_exact);
4845
4846/**
4847 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4848 * pages on a node.
4849 * @nid: the preferred node ID where memory should be allocated
4850 * @size: the number of bytes to allocate
4851 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4852 *
4853 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4854 * back.
4855 *
4856 * Return: pointer to the allocated area or %NULL in case of error.
4857 */
4858void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4859{
4860 unsigned int order = get_order(size);
4861 struct page *p;
4862
4863 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4864 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4865
4866 p = alloc_pages_node(nid, gfp_mask, order);
4867 if (!p)
4868 return NULL;
4869 return make_alloc_exact((unsigned long)page_address(p), order, size);
4870}
4871
4872/**
4873 * free_pages_exact - release memory allocated via alloc_pages_exact()
4874 * @virt: the value returned by alloc_pages_exact.
4875 * @size: size of allocation, same value as passed to alloc_pages_exact().
4876 *
4877 * Release the memory allocated by a previous call to alloc_pages_exact.
4878 */
4879void free_pages_exact(void *virt, size_t size)
4880{
4881 unsigned long addr = (unsigned long)virt;
4882 unsigned long end = addr + PAGE_ALIGN(size);
4883
4884 while (addr < end) {
4885 free_page(addr);
4886 addr += PAGE_SIZE;
4887 }
4888}
4889EXPORT_SYMBOL(free_pages_exact);
4890
4891/**
4892 * nr_free_zone_pages - count number of pages beyond high watermark
4893 * @offset: The zone index of the highest zone
4894 *
4895 * nr_free_zone_pages() counts the number of pages which are beyond the
4896 * high watermark within all zones at or below a given zone index. For each
4897 * zone, the number of pages is calculated as:
4898 *
4899 * nr_free_zone_pages = managed_pages - high_pages
4900 *
4901 * Return: number of pages beyond high watermark.
4902 */
4903static unsigned long nr_free_zone_pages(int offset)
4904{
4905 struct zoneref *z;
4906 struct zone *zone;
4907
4908 /* Just pick one node, since fallback list is circular */
4909 unsigned long sum = 0;
4910
4911 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4912
4913 for_each_zone_zonelist(zone, z, zonelist, offset) {
4914 unsigned long size = zone_managed_pages(zone);
4915 unsigned long high = high_wmark_pages(zone);
4916 if (size > high)
4917 sum += size - high;
4918 }
4919
4920 return sum;
4921}
4922
4923/**
4924 * nr_free_buffer_pages - count number of pages beyond high watermark
4925 *
4926 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4927 * watermark within ZONE_DMA and ZONE_NORMAL.
4928 *
4929 * Return: number of pages beyond high watermark within ZONE_DMA and
4930 * ZONE_NORMAL.
4931 */
4932unsigned long nr_free_buffer_pages(void)
4933{
4934 return nr_free_zone_pages(gfp_zone(GFP_USER));
4935}
4936EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4937
4938static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4939{
4940 zoneref->zone = zone;
4941 zoneref->zone_idx = zone_idx(zone);
4942}
4943
4944/*
4945 * Builds allocation fallback zone lists.
4946 *
4947 * Add all populated zones of a node to the zonelist.
4948 */
4949static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4950{
4951 struct zone *zone;
4952 enum zone_type zone_type = MAX_NR_ZONES;
4953 int nr_zones = 0;
4954
4955 do {
4956 zone_type--;
4957 zone = pgdat->node_zones + zone_type;
4958 if (populated_zone(zone)) {
4959 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4960 check_highest_zone(zone_type);
4961 }
4962 } while (zone_type);
4963
4964 return nr_zones;
4965}
4966
4967#ifdef CONFIG_NUMA
4968
4969static int __parse_numa_zonelist_order(char *s)
4970{
4971 /*
4972 * We used to support different zonelists modes but they turned
4973 * out to be just not useful. Let's keep the warning in place
4974 * if somebody still use the cmd line parameter so that we do
4975 * not fail it silently
4976 */
4977 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4978 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4979 return -EINVAL;
4980 }
4981 return 0;
4982}
4983
4984static char numa_zonelist_order[] = "Node";
4985#define NUMA_ZONELIST_ORDER_LEN 16
4986/*
4987 * sysctl handler for numa_zonelist_order
4988 */
4989static int numa_zonelist_order_handler(struct ctl_table *table, int write,
4990 void *buffer, size_t *length, loff_t *ppos)
4991{
4992 if (write)
4993 return __parse_numa_zonelist_order(buffer);
4994 return proc_dostring(table, write, buffer, length, ppos);
4995}
4996
4997static int node_load[MAX_NUMNODES];
4998
4999/**
5000 * find_next_best_node - find the next node that should appear in a given node's fallback list
5001 * @node: node whose fallback list we're appending
5002 * @used_node_mask: nodemask_t of already used nodes
5003 *
5004 * We use a number of factors to determine which is the next node that should
5005 * appear on a given node's fallback list. The node should not have appeared
5006 * already in @node's fallback list, and it should be the next closest node
5007 * according to the distance array (which contains arbitrary distance values
5008 * from each node to each node in the system), and should also prefer nodes
5009 * with no CPUs, since presumably they'll have very little allocation pressure
5010 * on them otherwise.
5011 *
5012 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5013 */
5014int find_next_best_node(int node, nodemask_t *used_node_mask)
5015{
5016 int n, val;
5017 int min_val = INT_MAX;
5018 int best_node = NUMA_NO_NODE;
5019
5020 /*
5021 * Use the local node if we haven't already, but for memoryless local
5022 * node, we should skip it and fall back to other nodes.
5023 */
5024 if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5025 node_set(node, *used_node_mask);
5026 return node;
5027 }
5028
5029 for_each_node_state(n, N_MEMORY) {
5030
5031 /* Don't want a node to appear more than once */
5032 if (node_isset(n, *used_node_mask))
5033 continue;
5034
5035 /* Use the distance array to find the distance */
5036 val = node_distance(node, n);
5037
5038 /* Penalize nodes under us ("prefer the next node") */
5039 val += (n < node);
5040
5041 /* Give preference to headless and unused nodes */
5042 if (!cpumask_empty(cpumask_of_node(n)))
5043 val += PENALTY_FOR_NODE_WITH_CPUS;
5044
5045 /* Slight preference for less loaded node */
5046 val *= MAX_NUMNODES;
5047 val += node_load[n];
5048
5049 if (val < min_val) {
5050 min_val = val;
5051 best_node = n;
5052 }
5053 }
5054
5055 if (best_node >= 0)
5056 node_set(best_node, *used_node_mask);
5057
5058 return best_node;
5059}
5060
5061
5062/*
5063 * Build zonelists ordered by node and zones within node.
5064 * This results in maximum locality--normal zone overflows into local
5065 * DMA zone, if any--but risks exhausting DMA zone.
5066 */
5067static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5068 unsigned nr_nodes)
5069{
5070 struct zoneref *zonerefs;
5071 int i;
5072
5073 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5074
5075 for (i = 0; i < nr_nodes; i++) {
5076 int nr_zones;
5077
5078 pg_data_t *node = NODE_DATA(node_order[i]);
5079
5080 nr_zones = build_zonerefs_node(node, zonerefs);
5081 zonerefs += nr_zones;
5082 }
5083 zonerefs->zone = NULL;
5084 zonerefs->zone_idx = 0;
5085}
5086
5087/*
5088 * Build gfp_thisnode zonelists
5089 */
5090static void build_thisnode_zonelists(pg_data_t *pgdat)
5091{
5092 struct zoneref *zonerefs;
5093 int nr_zones;
5094
5095 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5096 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5097 zonerefs += nr_zones;
5098 zonerefs->zone = NULL;
5099 zonerefs->zone_idx = 0;
5100}
5101
5102/*
5103 * Build zonelists ordered by zone and nodes within zones.
5104 * This results in conserving DMA zone[s] until all Normal memory is
5105 * exhausted, but results in overflowing to remote node while memory
5106 * may still exist in local DMA zone.
5107 */
5108
5109static void build_zonelists(pg_data_t *pgdat)
5110{
5111 static int node_order[MAX_NUMNODES];
5112 int node, nr_nodes = 0;
5113 nodemask_t used_mask = NODE_MASK_NONE;
5114 int local_node, prev_node;
5115
5116 /* NUMA-aware ordering of nodes */
5117 local_node = pgdat->node_id;
5118 prev_node = local_node;
5119
5120 memset(node_order, 0, sizeof(node_order));
5121 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5122 /*
5123 * We don't want to pressure a particular node.
5124 * So adding penalty to the first node in same
5125 * distance group to make it round-robin.
5126 */
5127 if (node_distance(local_node, node) !=
5128 node_distance(local_node, prev_node))
5129 node_load[node] += 1;
5130
5131 node_order[nr_nodes++] = node;
5132 prev_node = node;
5133 }
5134
5135 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5136 build_thisnode_zonelists(pgdat);
5137 pr_info("Fallback order for Node %d: ", local_node);
5138 for (node = 0; node < nr_nodes; node++)
5139 pr_cont("%d ", node_order[node]);
5140 pr_cont("\n");
5141}
5142
5143#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5144/*
5145 * Return node id of node used for "local" allocations.
5146 * I.e., first node id of first zone in arg node's generic zonelist.
5147 * Used for initializing percpu 'numa_mem', which is used primarily
5148 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5149 */
5150int local_memory_node(int node)
5151{
5152 struct zoneref *z;
5153
5154 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5155 gfp_zone(GFP_KERNEL),
5156 NULL);
5157 return zone_to_nid(z->zone);
5158}
5159#endif
5160
5161static void setup_min_unmapped_ratio(void);
5162static void setup_min_slab_ratio(void);
5163#else /* CONFIG_NUMA */
5164
5165static void build_zonelists(pg_data_t *pgdat)
5166{
5167 int node, local_node;
5168 struct zoneref *zonerefs;
5169 int nr_zones;
5170
5171 local_node = pgdat->node_id;
5172
5173 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5174 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5175 zonerefs += nr_zones;
5176
5177 /*
5178 * Now we build the zonelist so that it contains the zones
5179 * of all the other nodes.
5180 * We don't want to pressure a particular node, so when
5181 * building the zones for node N, we make sure that the
5182 * zones coming right after the local ones are those from
5183 * node N+1 (modulo N)
5184 */
5185 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5186 if (!node_online(node))
5187 continue;
5188 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5189 zonerefs += nr_zones;
5190 }
5191 for (node = 0; node < local_node; node++) {
5192 if (!node_online(node))
5193 continue;
5194 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5195 zonerefs += nr_zones;
5196 }
5197
5198 zonerefs->zone = NULL;
5199 zonerefs->zone_idx = 0;
5200}
5201
5202#endif /* CONFIG_NUMA */
5203
5204/*
5205 * Boot pageset table. One per cpu which is going to be used for all
5206 * zones and all nodes. The parameters will be set in such a way
5207 * that an item put on a list will immediately be handed over to
5208 * the buddy list. This is safe since pageset manipulation is done
5209 * with interrupts disabled.
5210 *
5211 * The boot_pagesets must be kept even after bootup is complete for
5212 * unused processors and/or zones. They do play a role for bootstrapping
5213 * hotplugged processors.
5214 *
5215 * zoneinfo_show() and maybe other functions do
5216 * not check if the processor is online before following the pageset pointer.
5217 * Other parts of the kernel may not check if the zone is available.
5218 */
5219static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5220/* These effectively disable the pcplists in the boot pageset completely */
5221#define BOOT_PAGESET_HIGH 0
5222#define BOOT_PAGESET_BATCH 1
5223static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5224static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5225
5226static void __build_all_zonelists(void *data)
5227{
5228 int nid;
5229 int __maybe_unused cpu;
5230 pg_data_t *self = data;
5231 unsigned long flags;
5232
5233 /*
5234 * The zonelist_update_seq must be acquired with irqsave because the
5235 * reader can be invoked from IRQ with GFP_ATOMIC.
5236 */
5237 write_seqlock_irqsave(&zonelist_update_seq, flags);
5238 /*
5239 * Also disable synchronous printk() to prevent any printk() from
5240 * trying to hold port->lock, for
5241 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5242 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5243 */
5244 printk_deferred_enter();
5245
5246#ifdef CONFIG_NUMA
5247 memset(node_load, 0, sizeof(node_load));
5248#endif
5249
5250 /*
5251 * This node is hotadded and no memory is yet present. So just
5252 * building zonelists is fine - no need to touch other nodes.
5253 */
5254 if (self && !node_online(self->node_id)) {
5255 build_zonelists(self);
5256 } else {
5257 /*
5258 * All possible nodes have pgdat preallocated
5259 * in free_area_init
5260 */
5261 for_each_node(nid) {
5262 pg_data_t *pgdat = NODE_DATA(nid);
5263
5264 build_zonelists(pgdat);
5265 }
5266
5267#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5268 /*
5269 * We now know the "local memory node" for each node--
5270 * i.e., the node of the first zone in the generic zonelist.
5271 * Set up numa_mem percpu variable for on-line cpus. During
5272 * boot, only the boot cpu should be on-line; we'll init the
5273 * secondary cpus' numa_mem as they come on-line. During
5274 * node/memory hotplug, we'll fixup all on-line cpus.
5275 */
5276 for_each_online_cpu(cpu)
5277 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5278#endif
5279 }
5280
5281 printk_deferred_exit();
5282 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5283}
5284
5285static noinline void __init
5286build_all_zonelists_init(void)
5287{
5288 int cpu;
5289
5290 __build_all_zonelists(NULL);
5291
5292 /*
5293 * Initialize the boot_pagesets that are going to be used
5294 * for bootstrapping processors. The real pagesets for
5295 * each zone will be allocated later when the per cpu
5296 * allocator is available.
5297 *
5298 * boot_pagesets are used also for bootstrapping offline
5299 * cpus if the system is already booted because the pagesets
5300 * are needed to initialize allocators on a specific cpu too.
5301 * F.e. the percpu allocator needs the page allocator which
5302 * needs the percpu allocator in order to allocate its pagesets
5303 * (a chicken-egg dilemma).
5304 */
5305 for_each_possible_cpu(cpu)
5306 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5307
5308 mminit_verify_zonelist();
5309 cpuset_init_current_mems_allowed();
5310}
5311
5312/*
5313 * unless system_state == SYSTEM_BOOTING.
5314 *
5315 * __ref due to call of __init annotated helper build_all_zonelists_init
5316 * [protected by SYSTEM_BOOTING].
5317 */
5318void __ref build_all_zonelists(pg_data_t *pgdat)
5319{
5320 unsigned long vm_total_pages;
5321
5322 if (system_state == SYSTEM_BOOTING) {
5323 build_all_zonelists_init();
5324 } else {
5325 __build_all_zonelists(pgdat);
5326 /* cpuset refresh routine should be here */
5327 }
5328 /* Get the number of free pages beyond high watermark in all zones. */
5329 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5330 /*
5331 * Disable grouping by mobility if the number of pages in the
5332 * system is too low to allow the mechanism to work. It would be
5333 * more accurate, but expensive to check per-zone. This check is
5334 * made on memory-hotadd so a system can start with mobility
5335 * disabled and enable it later
5336 */
5337 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5338 page_group_by_mobility_disabled = 1;
5339 else
5340 page_group_by_mobility_disabled = 0;
5341
5342 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5343 nr_online_nodes,
5344 page_group_by_mobility_disabled ? "off" : "on",
5345 vm_total_pages);
5346#ifdef CONFIG_NUMA
5347 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5348#endif
5349}
5350
5351static int zone_batchsize(struct zone *zone)
5352{
5353#ifdef CONFIG_MMU
5354 int batch;
5355
5356 /*
5357 * The number of pages to batch allocate is either ~0.1%
5358 * of the zone or 1MB, whichever is smaller. The batch
5359 * size is striking a balance between allocation latency
5360 * and zone lock contention.
5361 */
5362 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5363 batch /= 4; /* We effectively *= 4 below */
5364 if (batch < 1)
5365 batch = 1;
5366
5367 /*
5368 * Clamp the batch to a 2^n - 1 value. Having a power
5369 * of 2 value was found to be more likely to have
5370 * suboptimal cache aliasing properties in some cases.
5371 *
5372 * For example if 2 tasks are alternately allocating
5373 * batches of pages, one task can end up with a lot
5374 * of pages of one half of the possible page colors
5375 * and the other with pages of the other colors.
5376 */
5377 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5378
5379 return batch;
5380
5381#else
5382 /* The deferral and batching of frees should be suppressed under NOMMU
5383 * conditions.
5384 *
5385 * The problem is that NOMMU needs to be able to allocate large chunks
5386 * of contiguous memory as there's no hardware page translation to
5387 * assemble apparent contiguous memory from discontiguous pages.
5388 *
5389 * Queueing large contiguous runs of pages for batching, however,
5390 * causes the pages to actually be freed in smaller chunks. As there
5391 * can be a significant delay between the individual batches being
5392 * recycled, this leads to the once large chunks of space being
5393 * fragmented and becoming unavailable for high-order allocations.
5394 */
5395 return 0;
5396#endif
5397}
5398
5399static int percpu_pagelist_high_fraction;
5400static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5401 int high_fraction)
5402{
5403#ifdef CONFIG_MMU
5404 int high;
5405 int nr_split_cpus;
5406 unsigned long total_pages;
5407
5408 if (!high_fraction) {
5409 /*
5410 * By default, the high value of the pcp is based on the zone
5411 * low watermark so that if they are full then background
5412 * reclaim will not be started prematurely.
5413 */
5414 total_pages = low_wmark_pages(zone);
5415 } else {
5416 /*
5417 * If percpu_pagelist_high_fraction is configured, the high
5418 * value is based on a fraction of the managed pages in the
5419 * zone.
5420 */
5421 total_pages = zone_managed_pages(zone) / high_fraction;
5422 }
5423
5424 /*
5425 * Split the high value across all online CPUs local to the zone. Note
5426 * that early in boot that CPUs may not be online yet and that during
5427 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5428 * onlined. For memory nodes that have no CPUs, split the high value
5429 * across all online CPUs to mitigate the risk that reclaim is triggered
5430 * prematurely due to pages stored on pcp lists.
5431 */
5432 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5433 if (!nr_split_cpus)
5434 nr_split_cpus = num_online_cpus();
5435 high = total_pages / nr_split_cpus;
5436
5437 /*
5438 * Ensure high is at least batch*4. The multiple is based on the
5439 * historical relationship between high and batch.
5440 */
5441 high = max(high, batch << 2);
5442
5443 return high;
5444#else
5445 return 0;
5446#endif
5447}
5448
5449/*
5450 * pcp->high and pcp->batch values are related and generally batch is lower
5451 * than high. They are also related to pcp->count such that count is lower
5452 * than high, and as soon as it reaches high, the pcplist is flushed.
5453 *
5454 * However, guaranteeing these relations at all times would require e.g. write
5455 * barriers here but also careful usage of read barriers at the read side, and
5456 * thus be prone to error and bad for performance. Thus the update only prevents
5457 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5458 * should ensure they can cope with those fields changing asynchronously, and
5459 * fully trust only the pcp->count field on the local CPU with interrupts
5460 * disabled.
5461 *
5462 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5463 * outside of boot time (or some other assurance that no concurrent updaters
5464 * exist).
5465 */
5466static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5467 unsigned long high_max, unsigned long batch)
5468{
5469 WRITE_ONCE(pcp->batch, batch);
5470 WRITE_ONCE(pcp->high_min, high_min);
5471 WRITE_ONCE(pcp->high_max, high_max);
5472}
5473
5474static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5475{
5476 int pindex;
5477
5478 memset(pcp, 0, sizeof(*pcp));
5479 memset(pzstats, 0, sizeof(*pzstats));
5480
5481 spin_lock_init(&pcp->lock);
5482 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5483 INIT_LIST_HEAD(&pcp->lists[pindex]);
5484
5485 /*
5486 * Set batch and high values safe for a boot pageset. A true percpu
5487 * pageset's initialization will update them subsequently. Here we don't
5488 * need to be as careful as pageset_update() as nobody can access the
5489 * pageset yet.
5490 */
5491 pcp->high_min = BOOT_PAGESET_HIGH;
5492 pcp->high_max = BOOT_PAGESET_HIGH;
5493 pcp->batch = BOOT_PAGESET_BATCH;
5494 pcp->free_count = 0;
5495}
5496
5497static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
5498 unsigned long high_max, unsigned long batch)
5499{
5500 struct per_cpu_pages *pcp;
5501 int cpu;
5502
5503 for_each_possible_cpu(cpu) {
5504 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5505 pageset_update(pcp, high_min, high_max, batch);
5506 }
5507}
5508
5509/*
5510 * Calculate and set new high and batch values for all per-cpu pagesets of a
5511 * zone based on the zone's size.
5512 */
5513static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5514{
5515 int new_high_min, new_high_max, new_batch;
5516
5517 new_batch = max(1, zone_batchsize(zone));
5518 if (percpu_pagelist_high_fraction) {
5519 new_high_min = zone_highsize(zone, new_batch, cpu_online,
5520 percpu_pagelist_high_fraction);
5521 /*
5522 * PCP high is tuned manually, disable auto-tuning via
5523 * setting high_min and high_max to the manual value.
5524 */
5525 new_high_max = new_high_min;
5526 } else {
5527 new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
5528 new_high_max = zone_highsize(zone, new_batch, cpu_online,
5529 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
5530 }
5531
5532 if (zone->pageset_high_min == new_high_min &&
5533 zone->pageset_high_max == new_high_max &&
5534 zone->pageset_batch == new_batch)
5535 return;
5536
5537 zone->pageset_high_min = new_high_min;
5538 zone->pageset_high_max = new_high_max;
5539 zone->pageset_batch = new_batch;
5540
5541 __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
5542 new_batch);
5543}
5544
5545void __meminit setup_zone_pageset(struct zone *zone)
5546{
5547 int cpu;
5548
5549 /* Size may be 0 on !SMP && !NUMA */
5550 if (sizeof(struct per_cpu_zonestat) > 0)
5551 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5552
5553 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5554 for_each_possible_cpu(cpu) {
5555 struct per_cpu_pages *pcp;
5556 struct per_cpu_zonestat *pzstats;
5557
5558 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5559 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5560 per_cpu_pages_init(pcp, pzstats);
5561 }
5562
5563 zone_set_pageset_high_and_batch(zone, 0);
5564}
5565
5566/*
5567 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5568 * page high values need to be recalculated.
5569 */
5570static void zone_pcp_update(struct zone *zone, int cpu_online)
5571{
5572 mutex_lock(&pcp_batch_high_lock);
5573 zone_set_pageset_high_and_batch(zone, cpu_online);
5574 mutex_unlock(&pcp_batch_high_lock);
5575}
5576
5577static void zone_pcp_update_cacheinfo(struct zone *zone)
5578{
5579 int cpu;
5580 struct per_cpu_pages *pcp;
5581 struct cpu_cacheinfo *cci;
5582
5583 for_each_online_cpu(cpu) {
5584 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5585 cci = get_cpu_cacheinfo(cpu);
5586 /*
5587 * If data cache slice of CPU is large enough, "pcp->batch"
5588 * pages can be preserved in PCP before draining PCP for
5589 * consecutive high-order pages freeing without allocation.
5590 * This can reduce zone lock contention without hurting
5591 * cache-hot pages sharing.
5592 */
5593 spin_lock(&pcp->lock);
5594 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
5595 pcp->flags |= PCPF_FREE_HIGH_BATCH;
5596 else
5597 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
5598 spin_unlock(&pcp->lock);
5599 }
5600}
5601
5602void setup_pcp_cacheinfo(void)
5603{
5604 struct zone *zone;
5605
5606 for_each_populated_zone(zone)
5607 zone_pcp_update_cacheinfo(zone);
5608}
5609
5610/*
5611 * Allocate per cpu pagesets and initialize them.
5612 * Before this call only boot pagesets were available.
5613 */
5614void __init setup_per_cpu_pageset(void)
5615{
5616 struct pglist_data *pgdat;
5617 struct zone *zone;
5618 int __maybe_unused cpu;
5619
5620 for_each_populated_zone(zone)
5621 setup_zone_pageset(zone);
5622
5623#ifdef CONFIG_NUMA
5624 /*
5625 * Unpopulated zones continue using the boot pagesets.
5626 * The numa stats for these pagesets need to be reset.
5627 * Otherwise, they will end up skewing the stats of
5628 * the nodes these zones are associated with.
5629 */
5630 for_each_possible_cpu(cpu) {
5631 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5632 memset(pzstats->vm_numa_event, 0,
5633 sizeof(pzstats->vm_numa_event));
5634 }
5635#endif
5636
5637 for_each_online_pgdat(pgdat)
5638 pgdat->per_cpu_nodestats =
5639 alloc_percpu(struct per_cpu_nodestat);
5640}
5641
5642__meminit void zone_pcp_init(struct zone *zone)
5643{
5644 /*
5645 * per cpu subsystem is not up at this point. The following code
5646 * relies on the ability of the linker to provide the
5647 * offset of a (static) per cpu variable into the per cpu area.
5648 */
5649 zone->per_cpu_pageset = &boot_pageset;
5650 zone->per_cpu_zonestats = &boot_zonestats;
5651 zone->pageset_high_min = BOOT_PAGESET_HIGH;
5652 zone->pageset_high_max = BOOT_PAGESET_HIGH;
5653 zone->pageset_batch = BOOT_PAGESET_BATCH;
5654
5655 if (populated_zone(zone))
5656 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5657 zone->present_pages, zone_batchsize(zone));
5658}
5659
5660void adjust_managed_page_count(struct page *page, long count)
5661{
5662 atomic_long_add(count, &page_zone(page)->managed_pages);
5663 totalram_pages_add(count);
5664#ifdef CONFIG_HIGHMEM
5665 if (PageHighMem(page))
5666 totalhigh_pages_add(count);
5667#endif
5668}
5669EXPORT_SYMBOL(adjust_managed_page_count);
5670
5671unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5672{
5673 void *pos;
5674 unsigned long pages = 0;
5675
5676 start = (void *)PAGE_ALIGN((unsigned long)start);
5677 end = (void *)((unsigned long)end & PAGE_MASK);
5678 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5679 struct page *page = virt_to_page(pos);
5680 void *direct_map_addr;
5681
5682 /*
5683 * 'direct_map_addr' might be different from 'pos'
5684 * because some architectures' virt_to_page()
5685 * work with aliases. Getting the direct map
5686 * address ensures that we get a _writeable_
5687 * alias for the memset().
5688 */
5689 direct_map_addr = page_address(page);
5690 /*
5691 * Perform a kasan-unchecked memset() since this memory
5692 * has not been initialized.
5693 */
5694 direct_map_addr = kasan_reset_tag(direct_map_addr);
5695 if ((unsigned int)poison <= 0xFF)
5696 memset(direct_map_addr, poison, PAGE_SIZE);
5697
5698 free_reserved_page(page);
5699 }
5700
5701 if (pages && s)
5702 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5703
5704 return pages;
5705}
5706
5707static int page_alloc_cpu_dead(unsigned int cpu)
5708{
5709 struct zone *zone;
5710
5711 lru_add_drain_cpu(cpu);
5712 mlock_drain_remote(cpu);
5713 drain_pages(cpu);
5714
5715 /*
5716 * Spill the event counters of the dead processor
5717 * into the current processors event counters.
5718 * This artificially elevates the count of the current
5719 * processor.
5720 */
5721 vm_events_fold_cpu(cpu);
5722
5723 /*
5724 * Zero the differential counters of the dead processor
5725 * so that the vm statistics are consistent.
5726 *
5727 * This is only okay since the processor is dead and cannot
5728 * race with what we are doing.
5729 */
5730 cpu_vm_stats_fold(cpu);
5731
5732 for_each_populated_zone(zone)
5733 zone_pcp_update(zone, 0);
5734
5735 return 0;
5736}
5737
5738static int page_alloc_cpu_online(unsigned int cpu)
5739{
5740 struct zone *zone;
5741
5742 for_each_populated_zone(zone)
5743 zone_pcp_update(zone, 1);
5744 return 0;
5745}
5746
5747void __init page_alloc_init_cpuhp(void)
5748{
5749 int ret;
5750
5751 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5752 "mm/page_alloc:pcp",
5753 page_alloc_cpu_online,
5754 page_alloc_cpu_dead);
5755 WARN_ON(ret < 0);
5756}
5757
5758/*
5759 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5760 * or min_free_kbytes changes.
5761 */
5762static void calculate_totalreserve_pages(void)
5763{
5764 struct pglist_data *pgdat;
5765 unsigned long reserve_pages = 0;
5766 enum zone_type i, j;
5767
5768 for_each_online_pgdat(pgdat) {
5769
5770 pgdat->totalreserve_pages = 0;
5771
5772 for (i = 0; i < MAX_NR_ZONES; i++) {
5773 struct zone *zone = pgdat->node_zones + i;
5774 long max = 0;
5775 unsigned long managed_pages = zone_managed_pages(zone);
5776
5777 /* Find valid and maximum lowmem_reserve in the zone */
5778 for (j = i; j < MAX_NR_ZONES; j++) {
5779 if (zone->lowmem_reserve[j] > max)
5780 max = zone->lowmem_reserve[j];
5781 }
5782
5783 /* we treat the high watermark as reserved pages. */
5784 max += high_wmark_pages(zone);
5785
5786 if (max > managed_pages)
5787 max = managed_pages;
5788
5789 pgdat->totalreserve_pages += max;
5790
5791 reserve_pages += max;
5792 }
5793 }
5794 totalreserve_pages = reserve_pages;
5795}
5796
5797/*
5798 * setup_per_zone_lowmem_reserve - called whenever
5799 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5800 * has a correct pages reserved value, so an adequate number of
5801 * pages are left in the zone after a successful __alloc_pages().
5802 */
5803static void setup_per_zone_lowmem_reserve(void)
5804{
5805 struct pglist_data *pgdat;
5806 enum zone_type i, j;
5807
5808 for_each_online_pgdat(pgdat) {
5809 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5810 struct zone *zone = &pgdat->node_zones[i];
5811 int ratio = sysctl_lowmem_reserve_ratio[i];
5812 bool clear = !ratio || !zone_managed_pages(zone);
5813 unsigned long managed_pages = 0;
5814
5815 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5816 struct zone *upper_zone = &pgdat->node_zones[j];
5817
5818 managed_pages += zone_managed_pages(upper_zone);
5819
5820 if (clear)
5821 zone->lowmem_reserve[j] = 0;
5822 else
5823 zone->lowmem_reserve[j] = managed_pages / ratio;
5824 }
5825 }
5826 }
5827
5828 /* update totalreserve_pages */
5829 calculate_totalreserve_pages();
5830}
5831
5832static void __setup_per_zone_wmarks(void)
5833{
5834 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5835 unsigned long lowmem_pages = 0;
5836 struct zone *zone;
5837 unsigned long flags;
5838
5839 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5840 for_each_zone(zone) {
5841 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5842 lowmem_pages += zone_managed_pages(zone);
5843 }
5844
5845 for_each_zone(zone) {
5846 u64 tmp;
5847
5848 spin_lock_irqsave(&zone->lock, flags);
5849 tmp = (u64)pages_min * zone_managed_pages(zone);
5850 do_div(tmp, lowmem_pages);
5851 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5852 /*
5853 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5854 * need highmem and movable zones pages, so cap pages_min
5855 * to a small value here.
5856 *
5857 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5858 * deltas control async page reclaim, and so should
5859 * not be capped for highmem and movable zones.
5860 */
5861 unsigned long min_pages;
5862
5863 min_pages = zone_managed_pages(zone) / 1024;
5864 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5865 zone->_watermark[WMARK_MIN] = min_pages;
5866 } else {
5867 /*
5868 * If it's a lowmem zone, reserve a number of pages
5869 * proportionate to the zone's size.
5870 */
5871 zone->_watermark[WMARK_MIN] = tmp;
5872 }
5873
5874 /*
5875 * Set the kswapd watermarks distance according to the
5876 * scale factor in proportion to available memory, but
5877 * ensure a minimum size on small systems.
5878 */
5879 tmp = max_t(u64, tmp >> 2,
5880 mult_frac(zone_managed_pages(zone),
5881 watermark_scale_factor, 10000));
5882
5883 zone->watermark_boost = 0;
5884 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
5885 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5886 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5887
5888 spin_unlock_irqrestore(&zone->lock, flags);
5889 }
5890
5891 /* update totalreserve_pages */
5892 calculate_totalreserve_pages();
5893}
5894
5895/**
5896 * setup_per_zone_wmarks - called when min_free_kbytes changes
5897 * or when memory is hot-{added|removed}
5898 *
5899 * Ensures that the watermark[min,low,high] values for each zone are set
5900 * correctly with respect to min_free_kbytes.
5901 */
5902void setup_per_zone_wmarks(void)
5903{
5904 struct zone *zone;
5905 static DEFINE_SPINLOCK(lock);
5906
5907 spin_lock(&lock);
5908 __setup_per_zone_wmarks();
5909 spin_unlock(&lock);
5910
5911 /*
5912 * The watermark size have changed so update the pcpu batch
5913 * and high limits or the limits may be inappropriate.
5914 */
5915 for_each_zone(zone)
5916 zone_pcp_update(zone, 0);
5917}
5918
5919/*
5920 * Initialise min_free_kbytes.
5921 *
5922 * For small machines we want it small (128k min). For large machines
5923 * we want it large (256MB max). But it is not linear, because network
5924 * bandwidth does not increase linearly with machine size. We use
5925 *
5926 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5927 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5928 *
5929 * which yields
5930 *
5931 * 16MB: 512k
5932 * 32MB: 724k
5933 * 64MB: 1024k
5934 * 128MB: 1448k
5935 * 256MB: 2048k
5936 * 512MB: 2896k
5937 * 1024MB: 4096k
5938 * 2048MB: 5792k
5939 * 4096MB: 8192k
5940 * 8192MB: 11584k
5941 * 16384MB: 16384k
5942 */
5943void calculate_min_free_kbytes(void)
5944{
5945 unsigned long lowmem_kbytes;
5946 int new_min_free_kbytes;
5947
5948 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5949 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5950
5951 if (new_min_free_kbytes > user_min_free_kbytes)
5952 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
5953 else
5954 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5955 new_min_free_kbytes, user_min_free_kbytes);
5956
5957}
5958
5959int __meminit init_per_zone_wmark_min(void)
5960{
5961 calculate_min_free_kbytes();
5962 setup_per_zone_wmarks();
5963 refresh_zone_stat_thresholds();
5964 setup_per_zone_lowmem_reserve();
5965
5966#ifdef CONFIG_NUMA
5967 setup_min_unmapped_ratio();
5968 setup_min_slab_ratio();
5969#endif
5970
5971 khugepaged_min_free_kbytes_update();
5972
5973 return 0;
5974}
5975postcore_initcall(init_per_zone_wmark_min)
5976
5977/*
5978 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5979 * that we can call two helper functions whenever min_free_kbytes
5980 * changes.
5981 */
5982static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5983 void *buffer, size_t *length, loff_t *ppos)
5984{
5985 int rc;
5986
5987 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5988 if (rc)
5989 return rc;
5990
5991 if (write) {
5992 user_min_free_kbytes = min_free_kbytes;
5993 setup_per_zone_wmarks();
5994 }
5995 return 0;
5996}
5997
5998static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
5999 void *buffer, size_t *length, loff_t *ppos)
6000{
6001 int rc;
6002
6003 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6004 if (rc)
6005 return rc;
6006
6007 if (write)
6008 setup_per_zone_wmarks();
6009
6010 return 0;
6011}
6012
6013#ifdef CONFIG_NUMA
6014static void setup_min_unmapped_ratio(void)
6015{
6016 pg_data_t *pgdat;
6017 struct zone *zone;
6018
6019 for_each_online_pgdat(pgdat)
6020 pgdat->min_unmapped_pages = 0;
6021
6022 for_each_zone(zone)
6023 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6024 sysctl_min_unmapped_ratio) / 100;
6025}
6026
6027
6028static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6029 void *buffer, size_t *length, loff_t *ppos)
6030{
6031 int rc;
6032
6033 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6034 if (rc)
6035 return rc;
6036
6037 setup_min_unmapped_ratio();
6038
6039 return 0;
6040}
6041
6042static void setup_min_slab_ratio(void)
6043{
6044 pg_data_t *pgdat;
6045 struct zone *zone;
6046
6047 for_each_online_pgdat(pgdat)
6048 pgdat->min_slab_pages = 0;
6049
6050 for_each_zone(zone)
6051 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6052 sysctl_min_slab_ratio) / 100;
6053}
6054
6055static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6056 void *buffer, size_t *length, loff_t *ppos)
6057{
6058 int rc;
6059
6060 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6061 if (rc)
6062 return rc;
6063
6064 setup_min_slab_ratio();
6065
6066 return 0;
6067}
6068#endif
6069
6070/*
6071 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6072 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6073 * whenever sysctl_lowmem_reserve_ratio changes.
6074 *
6075 * The reserve ratio obviously has absolutely no relation with the
6076 * minimum watermarks. The lowmem reserve ratio can only make sense
6077 * if in function of the boot time zone sizes.
6078 */
6079static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table,
6080 int write, void *buffer, size_t *length, loff_t *ppos)
6081{
6082 int i;
6083
6084 proc_dointvec_minmax(table, write, buffer, length, ppos);
6085
6086 for (i = 0; i < MAX_NR_ZONES; i++) {
6087 if (sysctl_lowmem_reserve_ratio[i] < 1)
6088 sysctl_lowmem_reserve_ratio[i] = 0;
6089 }
6090
6091 setup_per_zone_lowmem_reserve();
6092 return 0;
6093}
6094
6095/*
6096 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6097 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6098 * pagelist can have before it gets flushed back to buddy allocator.
6099 */
6100static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
6101 int write, void *buffer, size_t *length, loff_t *ppos)
6102{
6103 struct zone *zone;
6104 int old_percpu_pagelist_high_fraction;
6105 int ret;
6106
6107 mutex_lock(&pcp_batch_high_lock);
6108 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6109
6110 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6111 if (!write || ret < 0)
6112 goto out;
6113
6114 /* Sanity checking to avoid pcp imbalance */
6115 if (percpu_pagelist_high_fraction &&
6116 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6117 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6118 ret = -EINVAL;
6119 goto out;
6120 }
6121
6122 /* No change? */
6123 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6124 goto out;
6125
6126 for_each_populated_zone(zone)
6127 zone_set_pageset_high_and_batch(zone, 0);
6128out:
6129 mutex_unlock(&pcp_batch_high_lock);
6130 return ret;
6131}
6132
6133static struct ctl_table page_alloc_sysctl_table[] = {
6134 {
6135 .procname = "min_free_kbytes",
6136 .data = &min_free_kbytes,
6137 .maxlen = sizeof(min_free_kbytes),
6138 .mode = 0644,
6139 .proc_handler = min_free_kbytes_sysctl_handler,
6140 .extra1 = SYSCTL_ZERO,
6141 },
6142 {
6143 .procname = "watermark_boost_factor",
6144 .data = &watermark_boost_factor,
6145 .maxlen = sizeof(watermark_boost_factor),
6146 .mode = 0644,
6147 .proc_handler = proc_dointvec_minmax,
6148 .extra1 = SYSCTL_ZERO,
6149 },
6150 {
6151 .procname = "watermark_scale_factor",
6152 .data = &watermark_scale_factor,
6153 .maxlen = sizeof(watermark_scale_factor),
6154 .mode = 0644,
6155 .proc_handler = watermark_scale_factor_sysctl_handler,
6156 .extra1 = SYSCTL_ONE,
6157 .extra2 = SYSCTL_THREE_THOUSAND,
6158 },
6159 {
6160 .procname = "percpu_pagelist_high_fraction",
6161 .data = &percpu_pagelist_high_fraction,
6162 .maxlen = sizeof(percpu_pagelist_high_fraction),
6163 .mode = 0644,
6164 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6165 .extra1 = SYSCTL_ZERO,
6166 },
6167 {
6168 .procname = "lowmem_reserve_ratio",
6169 .data = &sysctl_lowmem_reserve_ratio,
6170 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6171 .mode = 0644,
6172 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6173 },
6174#ifdef CONFIG_NUMA
6175 {
6176 .procname = "numa_zonelist_order",
6177 .data = &numa_zonelist_order,
6178 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6179 .mode = 0644,
6180 .proc_handler = numa_zonelist_order_handler,
6181 },
6182 {
6183 .procname = "min_unmapped_ratio",
6184 .data = &sysctl_min_unmapped_ratio,
6185 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6186 .mode = 0644,
6187 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6188 .extra1 = SYSCTL_ZERO,
6189 .extra2 = SYSCTL_ONE_HUNDRED,
6190 },
6191 {
6192 .procname = "min_slab_ratio",
6193 .data = &sysctl_min_slab_ratio,
6194 .maxlen = sizeof(sysctl_min_slab_ratio),
6195 .mode = 0644,
6196 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6197 .extra1 = SYSCTL_ZERO,
6198 .extra2 = SYSCTL_ONE_HUNDRED,
6199 },
6200#endif
6201 {}
6202};
6203
6204void __init page_alloc_sysctl_init(void)
6205{
6206 register_sysctl_init("vm", page_alloc_sysctl_table);
6207}
6208
6209#ifdef CONFIG_CONTIG_ALLOC
6210/* Usage: See admin-guide/dynamic-debug-howto.rst */
6211static void alloc_contig_dump_pages(struct list_head *page_list)
6212{
6213 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6214
6215 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6216 struct page *page;
6217
6218 dump_stack();
6219 list_for_each_entry(page, page_list, lru)
6220 dump_page(page, "migration failure");
6221 }
6222}
6223
6224/* [start, end) must belong to a single zone. */
6225int __alloc_contig_migrate_range(struct compact_control *cc,
6226 unsigned long start, unsigned long end)
6227{
6228 /* This function is based on compact_zone() from compaction.c. */
6229 unsigned int nr_reclaimed;
6230 unsigned long pfn = start;
6231 unsigned int tries = 0;
6232 int ret = 0;
6233 struct migration_target_control mtc = {
6234 .nid = zone_to_nid(cc->zone),
6235 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6236 };
6237
6238 lru_cache_disable();
6239
6240 while (pfn < end || !list_empty(&cc->migratepages)) {
6241 if (fatal_signal_pending(current)) {
6242 ret = -EINTR;
6243 break;
6244 }
6245
6246 if (list_empty(&cc->migratepages)) {
6247 cc->nr_migratepages = 0;
6248 ret = isolate_migratepages_range(cc, pfn, end);
6249 if (ret && ret != -EAGAIN)
6250 break;
6251 pfn = cc->migrate_pfn;
6252 tries = 0;
6253 } else if (++tries == 5) {
6254 ret = -EBUSY;
6255 break;
6256 }
6257
6258 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6259 &cc->migratepages);
6260 cc->nr_migratepages -= nr_reclaimed;
6261
6262 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6263 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6264
6265 /*
6266 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6267 * to retry again over this error, so do the same here.
6268 */
6269 if (ret == -ENOMEM)
6270 break;
6271 }
6272
6273 lru_cache_enable();
6274 if (ret < 0) {
6275 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6276 alloc_contig_dump_pages(&cc->migratepages);
6277 putback_movable_pages(&cc->migratepages);
6278 return ret;
6279 }
6280 return 0;
6281}
6282
6283/**
6284 * alloc_contig_range() -- tries to allocate given range of pages
6285 * @start: start PFN to allocate
6286 * @end: one-past-the-last PFN to allocate
6287 * @migratetype: migratetype of the underlying pageblocks (either
6288 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6289 * in range must have the same migratetype and it must
6290 * be either of the two.
6291 * @gfp_mask: GFP mask to use during compaction
6292 *
6293 * The PFN range does not have to be pageblock aligned. The PFN range must
6294 * belong to a single zone.
6295 *
6296 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6297 * pageblocks in the range. Once isolated, the pageblocks should not
6298 * be modified by others.
6299 *
6300 * Return: zero on success or negative error code. On success all
6301 * pages which PFN is in [start, end) are allocated for the caller and
6302 * need to be freed with free_contig_range().
6303 */
6304int alloc_contig_range(unsigned long start, unsigned long end,
6305 unsigned migratetype, gfp_t gfp_mask)
6306{
6307 unsigned long outer_start, outer_end;
6308 int order;
6309 int ret = 0;
6310
6311 struct compact_control cc = {
6312 .nr_migratepages = 0,
6313 .order = -1,
6314 .zone = page_zone(pfn_to_page(start)),
6315 .mode = MIGRATE_SYNC,
6316 .ignore_skip_hint = true,
6317 .no_set_skip_hint = true,
6318 .gfp_mask = current_gfp_context(gfp_mask),
6319 .alloc_contig = true,
6320 };
6321 INIT_LIST_HEAD(&cc.migratepages);
6322
6323 /*
6324 * What we do here is we mark all pageblocks in range as
6325 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6326 * have different sizes, and due to the way page allocator
6327 * work, start_isolate_page_range() has special handlings for this.
6328 *
6329 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6330 * migrate the pages from an unaligned range (ie. pages that
6331 * we are interested in). This will put all the pages in
6332 * range back to page allocator as MIGRATE_ISOLATE.
6333 *
6334 * When this is done, we take the pages in range from page
6335 * allocator removing them from the buddy system. This way
6336 * page allocator will never consider using them.
6337 *
6338 * This lets us mark the pageblocks back as
6339 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6340 * aligned range but not in the unaligned, original range are
6341 * put back to page allocator so that buddy can use them.
6342 */
6343
6344 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6345 if (ret)
6346 goto done;
6347
6348 drain_all_pages(cc.zone);
6349
6350 /*
6351 * In case of -EBUSY, we'd like to know which page causes problem.
6352 * So, just fall through. test_pages_isolated() has a tracepoint
6353 * which will report the busy page.
6354 *
6355 * It is possible that busy pages could become available before
6356 * the call to test_pages_isolated, and the range will actually be
6357 * allocated. So, if we fall through be sure to clear ret so that
6358 * -EBUSY is not accidentally used or returned to caller.
6359 */
6360 ret = __alloc_contig_migrate_range(&cc, start, end);
6361 if (ret && ret != -EBUSY)
6362 goto done;
6363 ret = 0;
6364
6365 /*
6366 * Pages from [start, end) are within a pageblock_nr_pages
6367 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6368 * more, all pages in [start, end) are free in page allocator.
6369 * What we are going to do is to allocate all pages from
6370 * [start, end) (that is remove them from page allocator).
6371 *
6372 * The only problem is that pages at the beginning and at the
6373 * end of interesting range may be not aligned with pages that
6374 * page allocator holds, ie. they can be part of higher order
6375 * pages. Because of this, we reserve the bigger range and
6376 * once this is done free the pages we are not interested in.
6377 *
6378 * We don't have to hold zone->lock here because the pages are
6379 * isolated thus they won't get removed from buddy.
6380 */
6381
6382 order = 0;
6383 outer_start = start;
6384 while (!PageBuddy(pfn_to_page(outer_start))) {
6385 if (++order > MAX_PAGE_ORDER) {
6386 outer_start = start;
6387 break;
6388 }
6389 outer_start &= ~0UL << order;
6390 }
6391
6392 if (outer_start != start) {
6393 order = buddy_order(pfn_to_page(outer_start));
6394
6395 /*
6396 * outer_start page could be small order buddy page and
6397 * it doesn't include start page. Adjust outer_start
6398 * in this case to report failed page properly
6399 * on tracepoint in test_pages_isolated()
6400 */
6401 if (outer_start + (1UL << order) <= start)
6402 outer_start = start;
6403 }
6404
6405 /* Make sure the range is really isolated. */
6406 if (test_pages_isolated(outer_start, end, 0)) {
6407 ret = -EBUSY;
6408 goto done;
6409 }
6410
6411 /* Grab isolated pages from freelists. */
6412 outer_end = isolate_freepages_range(&cc, outer_start, end);
6413 if (!outer_end) {
6414 ret = -EBUSY;
6415 goto done;
6416 }
6417
6418 /* Free head and tail (if any) */
6419 if (start != outer_start)
6420 free_contig_range(outer_start, start - outer_start);
6421 if (end != outer_end)
6422 free_contig_range(end, outer_end - end);
6423
6424done:
6425 undo_isolate_page_range(start, end, migratetype);
6426 return ret;
6427}
6428EXPORT_SYMBOL(alloc_contig_range);
6429
6430static int __alloc_contig_pages(unsigned long start_pfn,
6431 unsigned long nr_pages, gfp_t gfp_mask)
6432{
6433 unsigned long end_pfn = start_pfn + nr_pages;
6434
6435 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
6436 gfp_mask);
6437}
6438
6439static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6440 unsigned long nr_pages)
6441{
6442 unsigned long i, end_pfn = start_pfn + nr_pages;
6443 struct page *page;
6444
6445 for (i = start_pfn; i < end_pfn; i++) {
6446 page = pfn_to_online_page(i);
6447 if (!page)
6448 return false;
6449
6450 if (page_zone(page) != z)
6451 return false;
6452
6453 if (PageReserved(page))
6454 return false;
6455
6456 if (PageHuge(page))
6457 return false;
6458 }
6459 return true;
6460}
6461
6462static bool zone_spans_last_pfn(const struct zone *zone,
6463 unsigned long start_pfn, unsigned long nr_pages)
6464{
6465 unsigned long last_pfn = start_pfn + nr_pages - 1;
6466
6467 return zone_spans_pfn(zone, last_pfn);
6468}
6469
6470/**
6471 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6472 * @nr_pages: Number of contiguous pages to allocate
6473 * @gfp_mask: GFP mask to limit search and used during compaction
6474 * @nid: Target node
6475 * @nodemask: Mask for other possible nodes
6476 *
6477 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6478 * on an applicable zonelist to find a contiguous pfn range which can then be
6479 * tried for allocation with alloc_contig_range(). This routine is intended
6480 * for allocation requests which can not be fulfilled with the buddy allocator.
6481 *
6482 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6483 * power of two, then allocated range is also guaranteed to be aligned to same
6484 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6485 *
6486 * Allocated pages can be freed with free_contig_range() or by manually calling
6487 * __free_page() on each allocated page.
6488 *
6489 * Return: pointer to contiguous pages on success, or NULL if not successful.
6490 */
6491struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
6492 int nid, nodemask_t *nodemask)
6493{
6494 unsigned long ret, pfn, flags;
6495 struct zonelist *zonelist;
6496 struct zone *zone;
6497 struct zoneref *z;
6498
6499 zonelist = node_zonelist(nid, gfp_mask);
6500 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6501 gfp_zone(gfp_mask), nodemask) {
6502 spin_lock_irqsave(&zone->lock, flags);
6503
6504 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6505 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6506 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6507 /*
6508 * We release the zone lock here because
6509 * alloc_contig_range() will also lock the zone
6510 * at some point. If there's an allocation
6511 * spinning on this lock, it may win the race
6512 * and cause alloc_contig_range() to fail...
6513 */
6514 spin_unlock_irqrestore(&zone->lock, flags);
6515 ret = __alloc_contig_pages(pfn, nr_pages,
6516 gfp_mask);
6517 if (!ret)
6518 return pfn_to_page(pfn);
6519 spin_lock_irqsave(&zone->lock, flags);
6520 }
6521 pfn += nr_pages;
6522 }
6523 spin_unlock_irqrestore(&zone->lock, flags);
6524 }
6525 return NULL;
6526}
6527#endif /* CONFIG_CONTIG_ALLOC */
6528
6529void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6530{
6531 unsigned long count = 0;
6532
6533 for (; nr_pages--; pfn++) {
6534 struct page *page = pfn_to_page(pfn);
6535
6536 count += page_count(page) != 1;
6537 __free_page(page);
6538 }
6539 WARN(count != 0, "%lu pages are still in use!\n", count);
6540}
6541EXPORT_SYMBOL(free_contig_range);
6542
6543/*
6544 * Effectively disable pcplists for the zone by setting the high limit to 0
6545 * and draining all cpus. A concurrent page freeing on another CPU that's about
6546 * to put the page on pcplist will either finish before the drain and the page
6547 * will be drained, or observe the new high limit and skip the pcplist.
6548 *
6549 * Must be paired with a call to zone_pcp_enable().
6550 */
6551void zone_pcp_disable(struct zone *zone)
6552{
6553 mutex_lock(&pcp_batch_high_lock);
6554 __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
6555 __drain_all_pages(zone, true);
6556}
6557
6558void zone_pcp_enable(struct zone *zone)
6559{
6560 __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
6561 zone->pageset_high_max, zone->pageset_batch);
6562 mutex_unlock(&pcp_batch_high_lock);
6563}
6564
6565void zone_pcp_reset(struct zone *zone)
6566{
6567 int cpu;
6568 struct per_cpu_zonestat *pzstats;
6569
6570 if (zone->per_cpu_pageset != &boot_pageset) {
6571 for_each_online_cpu(cpu) {
6572 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6573 drain_zonestat(zone, pzstats);
6574 }
6575 free_percpu(zone->per_cpu_pageset);
6576 zone->per_cpu_pageset = &boot_pageset;
6577 if (zone->per_cpu_zonestats != &boot_zonestats) {
6578 free_percpu(zone->per_cpu_zonestats);
6579 zone->per_cpu_zonestats = &boot_zonestats;
6580 }
6581 }
6582}
6583
6584#ifdef CONFIG_MEMORY_HOTREMOVE
6585/*
6586 * All pages in the range must be in a single zone, must not contain holes,
6587 * must span full sections, and must be isolated before calling this function.
6588 */
6589void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6590{
6591 unsigned long pfn = start_pfn;
6592 struct page *page;
6593 struct zone *zone;
6594 unsigned int order;
6595 unsigned long flags;
6596
6597 offline_mem_sections(pfn, end_pfn);
6598 zone = page_zone(pfn_to_page(pfn));
6599 spin_lock_irqsave(&zone->lock, flags);
6600 while (pfn < end_pfn) {
6601 page = pfn_to_page(pfn);
6602 /*
6603 * The HWPoisoned page may be not in buddy system, and
6604 * page_count() is not 0.
6605 */
6606 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6607 pfn++;
6608 continue;
6609 }
6610 /*
6611 * At this point all remaining PageOffline() pages have a
6612 * reference count of 0 and can simply be skipped.
6613 */
6614 if (PageOffline(page)) {
6615 BUG_ON(page_count(page));
6616 BUG_ON(PageBuddy(page));
6617 pfn++;
6618 continue;
6619 }
6620
6621 BUG_ON(page_count(page));
6622 BUG_ON(!PageBuddy(page));
6623 order = buddy_order(page);
6624 del_page_from_free_list(page, zone, order);
6625 pfn += (1 << order);
6626 }
6627 spin_unlock_irqrestore(&zone->lock, flags);
6628}
6629#endif
6630
6631/*
6632 * This function returns a stable result only if called under zone lock.
6633 */
6634bool is_free_buddy_page(struct page *page)
6635{
6636 unsigned long pfn = page_to_pfn(page);
6637 unsigned int order;
6638
6639 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6640 struct page *page_head = page - (pfn & ((1 << order) - 1));
6641
6642 if (PageBuddy(page_head) &&
6643 buddy_order_unsafe(page_head) >= order)
6644 break;
6645 }
6646
6647 return order <= MAX_PAGE_ORDER;
6648}
6649EXPORT_SYMBOL(is_free_buddy_page);
6650
6651#ifdef CONFIG_MEMORY_FAILURE
6652/*
6653 * Break down a higher-order page in sub-pages, and keep our target out of
6654 * buddy allocator.
6655 */
6656static void break_down_buddy_pages(struct zone *zone, struct page *page,
6657 struct page *target, int low, int high,
6658 int migratetype)
6659{
6660 unsigned long size = 1 << high;
6661 struct page *current_buddy;
6662
6663 while (high > low) {
6664 high--;
6665 size >>= 1;
6666
6667 if (target >= &page[size]) {
6668 current_buddy = page;
6669 page = page + size;
6670 } else {
6671 current_buddy = page + size;
6672 }
6673
6674 if (set_page_guard(zone, current_buddy, high, migratetype))
6675 continue;
6676
6677 add_to_free_list(current_buddy, zone, high, migratetype);
6678 set_buddy_order(current_buddy, high);
6679 }
6680}
6681
6682/*
6683 * Take a page that will be marked as poisoned off the buddy allocator.
6684 */
6685bool take_page_off_buddy(struct page *page)
6686{
6687 struct zone *zone = page_zone(page);
6688 unsigned long pfn = page_to_pfn(page);
6689 unsigned long flags;
6690 unsigned int order;
6691 bool ret = false;
6692
6693 spin_lock_irqsave(&zone->lock, flags);
6694 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6695 struct page *page_head = page - (pfn & ((1 << order) - 1));
6696 int page_order = buddy_order(page_head);
6697
6698 if (PageBuddy(page_head) && page_order >= order) {
6699 unsigned long pfn_head = page_to_pfn(page_head);
6700 int migratetype = get_pfnblock_migratetype(page_head,
6701 pfn_head);
6702
6703 del_page_from_free_list(page_head, zone, page_order);
6704 break_down_buddy_pages(zone, page_head, page, 0,
6705 page_order, migratetype);
6706 SetPageHWPoisonTakenOff(page);
6707 if (!is_migrate_isolate(migratetype))
6708 __mod_zone_freepage_state(zone, -1, migratetype);
6709 ret = true;
6710 break;
6711 }
6712 if (page_count(page_head) > 0)
6713 break;
6714 }
6715 spin_unlock_irqrestore(&zone->lock, flags);
6716 return ret;
6717}
6718
6719/*
6720 * Cancel takeoff done by take_page_off_buddy().
6721 */
6722bool put_page_back_buddy(struct page *page)
6723{
6724 struct zone *zone = page_zone(page);
6725 unsigned long pfn = page_to_pfn(page);
6726 unsigned long flags;
6727 int migratetype = get_pfnblock_migratetype(page, pfn);
6728 bool ret = false;
6729
6730 spin_lock_irqsave(&zone->lock, flags);
6731 if (put_page_testzero(page)) {
6732 ClearPageHWPoisonTakenOff(page);
6733 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6734 if (TestClearPageHWPoison(page)) {
6735 ret = true;
6736 }
6737 }
6738 spin_unlock_irqrestore(&zone->lock, flags);
6739
6740 return ret;
6741}
6742#endif
6743
6744#ifdef CONFIG_ZONE_DMA
6745bool has_managed_dma(void)
6746{
6747 struct pglist_data *pgdat;
6748
6749 for_each_online_pgdat(pgdat) {
6750 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6751
6752 if (managed_zone(zone))
6753 return true;
6754 }
6755 return false;
6756}
6757#endif /* CONFIG_ZONE_DMA */
6758
6759#ifdef CONFIG_UNACCEPTED_MEMORY
6760
6761/* Counts number of zones with unaccepted pages. */
6762static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6763
6764static bool lazy_accept = true;
6765
6766static int __init accept_memory_parse(char *p)
6767{
6768 if (!strcmp(p, "lazy")) {
6769 lazy_accept = true;
6770 return 0;
6771 } else if (!strcmp(p, "eager")) {
6772 lazy_accept = false;
6773 return 0;
6774 } else {
6775 return -EINVAL;
6776 }
6777}
6778early_param("accept_memory", accept_memory_parse);
6779
6780static bool page_contains_unaccepted(struct page *page, unsigned int order)
6781{
6782 phys_addr_t start = page_to_phys(page);
6783 phys_addr_t end = start + (PAGE_SIZE << order);
6784
6785 return range_contains_unaccepted_memory(start, end);
6786}
6787
6788static void accept_page(struct page *page, unsigned int order)
6789{
6790 phys_addr_t start = page_to_phys(page);
6791
6792 accept_memory(start, start + (PAGE_SIZE << order));
6793}
6794
6795static bool try_to_accept_memory_one(struct zone *zone)
6796{
6797 unsigned long flags;
6798 struct page *page;
6799 bool last;
6800
6801 if (list_empty(&zone->unaccepted_pages))
6802 return false;
6803
6804 spin_lock_irqsave(&zone->lock, flags);
6805 page = list_first_entry_or_null(&zone->unaccepted_pages,
6806 struct page, lru);
6807 if (!page) {
6808 spin_unlock_irqrestore(&zone->lock, flags);
6809 return false;
6810 }
6811
6812 list_del(&page->lru);
6813 last = list_empty(&zone->unaccepted_pages);
6814
6815 __mod_zone_freepage_state(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6816 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6817 spin_unlock_irqrestore(&zone->lock, flags);
6818
6819 accept_page(page, MAX_PAGE_ORDER);
6820
6821 __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
6822
6823 if (last)
6824 static_branch_dec(&zones_with_unaccepted_pages);
6825
6826 return true;
6827}
6828
6829static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6830{
6831 long to_accept;
6832 int ret = false;
6833
6834 /* How much to accept to get to high watermark? */
6835 to_accept = high_wmark_pages(zone) -
6836 (zone_page_state(zone, NR_FREE_PAGES) -
6837 __zone_watermark_unusable_free(zone, order, 0));
6838
6839 /* Accept at least one page */
6840 do {
6841 if (!try_to_accept_memory_one(zone))
6842 break;
6843 ret = true;
6844 to_accept -= MAX_ORDER_NR_PAGES;
6845 } while (to_accept > 0);
6846
6847 return ret;
6848}
6849
6850static inline bool has_unaccepted_memory(void)
6851{
6852 return static_branch_unlikely(&zones_with_unaccepted_pages);
6853}
6854
6855static bool __free_unaccepted(struct page *page)
6856{
6857 struct zone *zone = page_zone(page);
6858 unsigned long flags;
6859 bool first = false;
6860
6861 if (!lazy_accept)
6862 return false;
6863
6864 spin_lock_irqsave(&zone->lock, flags);
6865 first = list_empty(&zone->unaccepted_pages);
6866 list_add_tail(&page->lru, &zone->unaccepted_pages);
6867 __mod_zone_freepage_state(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6868 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
6869 spin_unlock_irqrestore(&zone->lock, flags);
6870
6871 if (first)
6872 static_branch_inc(&zones_with_unaccepted_pages);
6873
6874 return true;
6875}
6876
6877#else
6878
6879static bool page_contains_unaccepted(struct page *page, unsigned int order)
6880{
6881 return false;
6882}
6883
6884static void accept_page(struct page *page, unsigned int order)
6885{
6886}
6887
6888static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6889{
6890 return false;
6891}
6892
6893static inline bool has_unaccepted_memory(void)
6894{
6895 return false;
6896}
6897
6898static bool __free_unaccepted(struct page *page)
6899{
6900 BUILD_BUG();
6901 return false;
6902}
6903
6904#endif /* CONFIG_UNACCEPTED_MEMORY */