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/swap.h>
22#include <linux/interrupt.h>
23#include <linux/pagemap.h>
24#include <linux/jiffies.h>
25#include <linux/memblock.h>
26#include <linux/compiler.h>
27#include <linux/kernel.h>
28#include <linux/kasan.h>
29#include <linux/module.h>
30#include <linux/suspend.h>
31#include <linux/pagevec.h>
32#include <linux/blkdev.h>
33#include <linux/slab.h>
34#include <linux/ratelimit.h>
35#include <linux/oom.h>
36#include <linux/topology.h>
37#include <linux/sysctl.h>
38#include <linux/cpu.h>
39#include <linux/cpuset.h>
40#include <linux/memory_hotplug.h>
41#include <linux/nodemask.h>
42#include <linux/vmalloc.h>
43#include <linux/vmstat.h>
44#include <linux/mempolicy.h>
45#include <linux/memremap.h>
46#include <linux/stop_machine.h>
47#include <linux/random.h>
48#include <linux/sort.h>
49#include <linux/pfn.h>
50#include <linux/backing-dev.h>
51#include <linux/fault-inject.h>
52#include <linux/page-isolation.h>
53#include <linux/debugobjects.h>
54#include <linux/kmemleak.h>
55#include <linux/compaction.h>
56#include <trace/events/kmem.h>
57#include <trace/events/oom.h>
58#include <linux/prefetch.h>
59#include <linux/mm_inline.h>
60#include <linux/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#include <linux/psi.h>
71
72#include <asm/sections.h>
73#include <asm/tlbflush.h>
74#include <asm/div64.h>
75#include "internal.h"
76#include "shuffle.h"
77
78/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
79static DEFINE_MUTEX(pcp_batch_high_lock);
80#define MIN_PERCPU_PAGELIST_FRACTION (8)
81
82#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
83DEFINE_PER_CPU(int, numa_node);
84EXPORT_PER_CPU_SYMBOL(numa_node);
85#endif
86
87DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
88
89#ifdef CONFIG_HAVE_MEMORYLESS_NODES
90/*
91 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
92 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
93 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
94 * defined in <linux/topology.h>.
95 */
96DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
97EXPORT_PER_CPU_SYMBOL(_numa_mem_);
98int _node_numa_mem_[MAX_NUMNODES];
99#endif
100
101/* work_structs for global per-cpu drains */
102struct pcpu_drain {
103 struct zone *zone;
104 struct work_struct work;
105};
106DEFINE_MUTEX(pcpu_drain_mutex);
107DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
108
109#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
110volatile unsigned long latent_entropy __latent_entropy;
111EXPORT_SYMBOL(latent_entropy);
112#endif
113
114/*
115 * Array of node states.
116 */
117nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
118 [N_POSSIBLE] = NODE_MASK_ALL,
119 [N_ONLINE] = { { [0] = 1UL } },
120#ifndef CONFIG_NUMA
121 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
122#ifdef CONFIG_HIGHMEM
123 [N_HIGH_MEMORY] = { { [0] = 1UL } },
124#endif
125 [N_MEMORY] = { { [0] = 1UL } },
126 [N_CPU] = { { [0] = 1UL } },
127#endif /* NUMA */
128};
129EXPORT_SYMBOL(node_states);
130
131atomic_long_t _totalram_pages __read_mostly;
132EXPORT_SYMBOL(_totalram_pages);
133unsigned long totalreserve_pages __read_mostly;
134unsigned long totalcma_pages __read_mostly;
135
136int percpu_pagelist_fraction;
137gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
138#ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
139DEFINE_STATIC_KEY_TRUE(init_on_alloc);
140#else
141DEFINE_STATIC_KEY_FALSE(init_on_alloc);
142#endif
143EXPORT_SYMBOL(init_on_alloc);
144
145#ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
146DEFINE_STATIC_KEY_TRUE(init_on_free);
147#else
148DEFINE_STATIC_KEY_FALSE(init_on_free);
149#endif
150EXPORT_SYMBOL(init_on_free);
151
152static int __init early_init_on_alloc(char *buf)
153{
154 int ret;
155 bool bool_result;
156
157 if (!buf)
158 return -EINVAL;
159 ret = kstrtobool(buf, &bool_result);
160 if (bool_result && page_poisoning_enabled())
161 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
162 if (bool_result)
163 static_branch_enable(&init_on_alloc);
164 else
165 static_branch_disable(&init_on_alloc);
166 return ret;
167}
168early_param("init_on_alloc", early_init_on_alloc);
169
170static int __init early_init_on_free(char *buf)
171{
172 int ret;
173 bool bool_result;
174
175 if (!buf)
176 return -EINVAL;
177 ret = kstrtobool(buf, &bool_result);
178 if (bool_result && page_poisoning_enabled())
179 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
180 if (bool_result)
181 static_branch_enable(&init_on_free);
182 else
183 static_branch_disable(&init_on_free);
184 return ret;
185}
186early_param("init_on_free", early_init_on_free);
187
188/*
189 * A cached value of the page's pageblock's migratetype, used when the page is
190 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
191 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
192 * Also the migratetype set in the page does not necessarily match the pcplist
193 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
194 * other index - this ensures that it will be put on the correct CMA freelist.
195 */
196static inline int get_pcppage_migratetype(struct page *page)
197{
198 return page->index;
199}
200
201static inline void set_pcppage_migratetype(struct page *page, int migratetype)
202{
203 page->index = migratetype;
204}
205
206#ifdef CONFIG_PM_SLEEP
207/*
208 * The following functions are used by the suspend/hibernate code to temporarily
209 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
210 * while devices are suspended. To avoid races with the suspend/hibernate code,
211 * they should always be called with system_transition_mutex held
212 * (gfp_allowed_mask also should only be modified with system_transition_mutex
213 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
214 * with that modification).
215 */
216
217static gfp_t saved_gfp_mask;
218
219void pm_restore_gfp_mask(void)
220{
221 WARN_ON(!mutex_is_locked(&system_transition_mutex));
222 if (saved_gfp_mask) {
223 gfp_allowed_mask = saved_gfp_mask;
224 saved_gfp_mask = 0;
225 }
226}
227
228void pm_restrict_gfp_mask(void)
229{
230 WARN_ON(!mutex_is_locked(&system_transition_mutex));
231 WARN_ON(saved_gfp_mask);
232 saved_gfp_mask = gfp_allowed_mask;
233 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
234}
235
236bool pm_suspended_storage(void)
237{
238 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
239 return false;
240 return true;
241}
242#endif /* CONFIG_PM_SLEEP */
243
244#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
245unsigned int pageblock_order __read_mostly;
246#endif
247
248static void __free_pages_ok(struct page *page, unsigned int order);
249
250/*
251 * results with 256, 32 in the lowmem_reserve sysctl:
252 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
253 * 1G machine -> (16M dma, 784M normal, 224M high)
254 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
255 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
256 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
257 *
258 * TBD: should special case ZONE_DMA32 machines here - in those we normally
259 * don't need any ZONE_NORMAL reservation
260 */
261int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
262#ifdef CONFIG_ZONE_DMA
263 [ZONE_DMA] = 256,
264#endif
265#ifdef CONFIG_ZONE_DMA32
266 [ZONE_DMA32] = 256,
267#endif
268 [ZONE_NORMAL] = 32,
269#ifdef CONFIG_HIGHMEM
270 [ZONE_HIGHMEM] = 0,
271#endif
272 [ZONE_MOVABLE] = 0,
273};
274
275static char * const zone_names[MAX_NR_ZONES] = {
276#ifdef CONFIG_ZONE_DMA
277 "DMA",
278#endif
279#ifdef CONFIG_ZONE_DMA32
280 "DMA32",
281#endif
282 "Normal",
283#ifdef CONFIG_HIGHMEM
284 "HighMem",
285#endif
286 "Movable",
287#ifdef CONFIG_ZONE_DEVICE
288 "Device",
289#endif
290};
291
292const char * const migratetype_names[MIGRATE_TYPES] = {
293 "Unmovable",
294 "Movable",
295 "Reclaimable",
296 "HighAtomic",
297#ifdef CONFIG_CMA
298 "CMA",
299#endif
300#ifdef CONFIG_MEMORY_ISOLATION
301 "Isolate",
302#endif
303};
304
305compound_page_dtor * const compound_page_dtors[] = {
306 NULL,
307 free_compound_page,
308#ifdef CONFIG_HUGETLB_PAGE
309 free_huge_page,
310#endif
311#ifdef CONFIG_TRANSPARENT_HUGEPAGE
312 free_transhuge_page,
313#endif
314};
315
316int min_free_kbytes = 1024;
317int user_min_free_kbytes = -1;
318#ifdef CONFIG_DISCONTIGMEM
319/*
320 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
321 * are not on separate NUMA nodes. Functionally this works but with
322 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
323 * quite small. By default, do not boost watermarks on discontigmem as in
324 * many cases very high-order allocations like THP are likely to be
325 * unsupported and the premature reclaim offsets the advantage of long-term
326 * fragmentation avoidance.
327 */
328int watermark_boost_factor __read_mostly;
329#else
330int watermark_boost_factor __read_mostly = 15000;
331#endif
332int watermark_scale_factor = 10;
333
334static unsigned long nr_kernel_pages __initdata;
335static unsigned long nr_all_pages __initdata;
336static unsigned long dma_reserve __initdata;
337
338#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
339static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
340static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
341static unsigned long required_kernelcore __initdata;
342static unsigned long required_kernelcore_percent __initdata;
343static unsigned long required_movablecore __initdata;
344static unsigned long required_movablecore_percent __initdata;
345static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
346static bool mirrored_kernelcore __meminitdata;
347
348/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
349int movable_zone;
350EXPORT_SYMBOL(movable_zone);
351#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
352
353#if MAX_NUMNODES > 1
354unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
355unsigned int nr_online_nodes __read_mostly = 1;
356EXPORT_SYMBOL(nr_node_ids);
357EXPORT_SYMBOL(nr_online_nodes);
358#endif
359
360int page_group_by_mobility_disabled __read_mostly;
361
362#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
363/*
364 * During boot we initialize deferred pages on-demand, as needed, but once
365 * page_alloc_init_late() has finished, the deferred pages are all initialized,
366 * and we can permanently disable that path.
367 */
368static DEFINE_STATIC_KEY_TRUE(deferred_pages);
369
370/*
371 * Calling kasan_free_pages() only after deferred memory initialization
372 * has completed. Poisoning pages during deferred memory init will greatly
373 * lengthen the process and cause problem in large memory systems as the
374 * deferred pages initialization is done with interrupt disabled.
375 *
376 * Assuming that there will be no reference to those newly initialized
377 * pages before they are ever allocated, this should have no effect on
378 * KASAN memory tracking as the poison will be properly inserted at page
379 * allocation time. The only corner case is when pages are allocated by
380 * on-demand allocation and then freed again before the deferred pages
381 * initialization is done, but this is not likely to happen.
382 */
383static inline void kasan_free_nondeferred_pages(struct page *page, int order)
384{
385 if (!static_branch_unlikely(&deferred_pages))
386 kasan_free_pages(page, order);
387}
388
389/* Returns true if the struct page for the pfn is uninitialised */
390static inline bool __meminit early_page_uninitialised(unsigned long pfn)
391{
392 int nid = early_pfn_to_nid(pfn);
393
394 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
395 return true;
396
397 return false;
398}
399
400/*
401 * Returns true when the remaining initialisation should be deferred until
402 * later in the boot cycle when it can be parallelised.
403 */
404static bool __meminit
405defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
406{
407 static unsigned long prev_end_pfn, nr_initialised;
408
409 /*
410 * prev_end_pfn static that contains the end of previous zone
411 * No need to protect because called very early in boot before smp_init.
412 */
413 if (prev_end_pfn != end_pfn) {
414 prev_end_pfn = end_pfn;
415 nr_initialised = 0;
416 }
417
418 /* Always populate low zones for address-constrained allocations */
419 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
420 return false;
421
422 /*
423 * We start only with one section of pages, more pages are added as
424 * needed until the rest of deferred pages are initialized.
425 */
426 nr_initialised++;
427 if ((nr_initialised > PAGES_PER_SECTION) &&
428 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
429 NODE_DATA(nid)->first_deferred_pfn = pfn;
430 return true;
431 }
432 return false;
433}
434#else
435#define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
436
437static inline bool early_page_uninitialised(unsigned long pfn)
438{
439 return false;
440}
441
442static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
443{
444 return false;
445}
446#endif
447
448/* Return a pointer to the bitmap storing bits affecting a block of pages */
449static inline unsigned long *get_pageblock_bitmap(struct page *page,
450 unsigned long pfn)
451{
452#ifdef CONFIG_SPARSEMEM
453 return section_to_usemap(__pfn_to_section(pfn));
454#else
455 return page_zone(page)->pageblock_flags;
456#endif /* CONFIG_SPARSEMEM */
457}
458
459static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
460{
461#ifdef CONFIG_SPARSEMEM
462 pfn &= (PAGES_PER_SECTION-1);
463 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
464#else
465 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
466 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
467#endif /* CONFIG_SPARSEMEM */
468}
469
470/**
471 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
472 * @page: The page within the block of interest
473 * @pfn: The target page frame number
474 * @end_bitidx: The last bit of interest to retrieve
475 * @mask: mask of bits that the caller is interested in
476 *
477 * Return: pageblock_bits flags
478 */
479static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
480 unsigned long pfn,
481 unsigned long end_bitidx,
482 unsigned long mask)
483{
484 unsigned long *bitmap;
485 unsigned long bitidx, word_bitidx;
486 unsigned long word;
487
488 bitmap = get_pageblock_bitmap(page, pfn);
489 bitidx = pfn_to_bitidx(page, pfn);
490 word_bitidx = bitidx / BITS_PER_LONG;
491 bitidx &= (BITS_PER_LONG-1);
492
493 word = bitmap[word_bitidx];
494 bitidx += end_bitidx;
495 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
496}
497
498unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
499 unsigned long end_bitidx,
500 unsigned long mask)
501{
502 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
503}
504
505static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
506{
507 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
508}
509
510/**
511 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
512 * @page: The page within the block of interest
513 * @flags: The flags to set
514 * @pfn: The target page frame number
515 * @end_bitidx: The last bit of interest
516 * @mask: mask of bits that the caller is interested in
517 */
518void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
519 unsigned long pfn,
520 unsigned long end_bitidx,
521 unsigned long mask)
522{
523 unsigned long *bitmap;
524 unsigned long bitidx, word_bitidx;
525 unsigned long old_word, word;
526
527 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
528 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
529
530 bitmap = get_pageblock_bitmap(page, pfn);
531 bitidx = pfn_to_bitidx(page, pfn);
532 word_bitidx = bitidx / BITS_PER_LONG;
533 bitidx &= (BITS_PER_LONG-1);
534
535 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
536
537 bitidx += end_bitidx;
538 mask <<= (BITS_PER_LONG - bitidx - 1);
539 flags <<= (BITS_PER_LONG - bitidx - 1);
540
541 word = READ_ONCE(bitmap[word_bitidx]);
542 for (;;) {
543 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
544 if (word == old_word)
545 break;
546 word = old_word;
547 }
548}
549
550void set_pageblock_migratetype(struct page *page, int migratetype)
551{
552 if (unlikely(page_group_by_mobility_disabled &&
553 migratetype < MIGRATE_PCPTYPES))
554 migratetype = MIGRATE_UNMOVABLE;
555
556 set_pageblock_flags_group(page, (unsigned long)migratetype,
557 PB_migrate, PB_migrate_end);
558}
559
560#ifdef CONFIG_DEBUG_VM
561static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
562{
563 int ret = 0;
564 unsigned seq;
565 unsigned long pfn = page_to_pfn(page);
566 unsigned long sp, start_pfn;
567
568 do {
569 seq = zone_span_seqbegin(zone);
570 start_pfn = zone->zone_start_pfn;
571 sp = zone->spanned_pages;
572 if (!zone_spans_pfn(zone, pfn))
573 ret = 1;
574 } while (zone_span_seqretry(zone, seq));
575
576 if (ret)
577 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
578 pfn, zone_to_nid(zone), zone->name,
579 start_pfn, start_pfn + sp);
580
581 return ret;
582}
583
584static int page_is_consistent(struct zone *zone, struct page *page)
585{
586 if (!pfn_valid_within(page_to_pfn(page)))
587 return 0;
588 if (zone != page_zone(page))
589 return 0;
590
591 return 1;
592}
593/*
594 * Temporary debugging check for pages not lying within a given zone.
595 */
596static int __maybe_unused bad_range(struct zone *zone, struct page *page)
597{
598 if (page_outside_zone_boundaries(zone, page))
599 return 1;
600 if (!page_is_consistent(zone, page))
601 return 1;
602
603 return 0;
604}
605#else
606static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
607{
608 return 0;
609}
610#endif
611
612static void bad_page(struct page *page, const char *reason,
613 unsigned long bad_flags)
614{
615 static unsigned long resume;
616 static unsigned long nr_shown;
617 static unsigned long nr_unshown;
618
619 /*
620 * Allow a burst of 60 reports, then keep quiet for that minute;
621 * or allow a steady drip of one report per second.
622 */
623 if (nr_shown == 60) {
624 if (time_before(jiffies, resume)) {
625 nr_unshown++;
626 goto out;
627 }
628 if (nr_unshown) {
629 pr_alert(
630 "BUG: Bad page state: %lu messages suppressed\n",
631 nr_unshown);
632 nr_unshown = 0;
633 }
634 nr_shown = 0;
635 }
636 if (nr_shown++ == 0)
637 resume = jiffies + 60 * HZ;
638
639 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
640 current->comm, page_to_pfn(page));
641 __dump_page(page, reason);
642 bad_flags &= page->flags;
643 if (bad_flags)
644 pr_alert("bad because of flags: %#lx(%pGp)\n",
645 bad_flags, &bad_flags);
646 dump_page_owner(page);
647
648 print_modules();
649 dump_stack();
650out:
651 /* Leave bad fields for debug, except PageBuddy could make trouble */
652 page_mapcount_reset(page); /* remove PageBuddy */
653 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
654}
655
656/*
657 * Higher-order pages are called "compound pages". They are structured thusly:
658 *
659 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
660 *
661 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
662 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
663 *
664 * The first tail page's ->compound_dtor holds the offset in array of compound
665 * page destructors. See compound_page_dtors.
666 *
667 * The first tail page's ->compound_order holds the order of allocation.
668 * This usage means that zero-order pages may not be compound.
669 */
670
671void free_compound_page(struct page *page)
672{
673 mem_cgroup_uncharge(page);
674 __free_pages_ok(page, compound_order(page));
675}
676
677void prep_compound_page(struct page *page, unsigned int order)
678{
679 int i;
680 int nr_pages = 1 << order;
681
682 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
683 set_compound_order(page, order);
684 __SetPageHead(page);
685 for (i = 1; i < nr_pages; i++) {
686 struct page *p = page + i;
687 set_page_count(p, 0);
688 p->mapping = TAIL_MAPPING;
689 set_compound_head(p, page);
690 }
691 atomic_set(compound_mapcount_ptr(page), -1);
692}
693
694#ifdef CONFIG_DEBUG_PAGEALLOC
695unsigned int _debug_guardpage_minorder;
696
697#ifdef CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
698DEFINE_STATIC_KEY_TRUE(_debug_pagealloc_enabled);
699#else
700DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
701#endif
702EXPORT_SYMBOL(_debug_pagealloc_enabled);
703
704DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
705
706static int __init early_debug_pagealloc(char *buf)
707{
708 bool enable = false;
709
710 if (kstrtobool(buf, &enable))
711 return -EINVAL;
712
713 if (enable)
714 static_branch_enable(&_debug_pagealloc_enabled);
715
716 return 0;
717}
718early_param("debug_pagealloc", early_debug_pagealloc);
719
720static void init_debug_guardpage(void)
721{
722 if (!debug_pagealloc_enabled())
723 return;
724
725 if (!debug_guardpage_minorder())
726 return;
727
728 static_branch_enable(&_debug_guardpage_enabled);
729}
730
731static int __init debug_guardpage_minorder_setup(char *buf)
732{
733 unsigned long res;
734
735 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
736 pr_err("Bad debug_guardpage_minorder value\n");
737 return 0;
738 }
739 _debug_guardpage_minorder = res;
740 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
741 return 0;
742}
743early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
744
745static inline bool set_page_guard(struct zone *zone, struct page *page,
746 unsigned int order, int migratetype)
747{
748 if (!debug_guardpage_enabled())
749 return false;
750
751 if (order >= debug_guardpage_minorder())
752 return false;
753
754 __SetPageGuard(page);
755 INIT_LIST_HEAD(&page->lru);
756 set_page_private(page, order);
757 /* Guard pages are not available for any usage */
758 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
759
760 return true;
761}
762
763static inline void clear_page_guard(struct zone *zone, struct page *page,
764 unsigned int order, int migratetype)
765{
766 if (!debug_guardpage_enabled())
767 return;
768
769 __ClearPageGuard(page);
770
771 set_page_private(page, 0);
772 if (!is_migrate_isolate(migratetype))
773 __mod_zone_freepage_state(zone, (1 << order), migratetype);
774}
775#else
776static inline bool set_page_guard(struct zone *zone, struct page *page,
777 unsigned int order, int migratetype) { return false; }
778static inline void clear_page_guard(struct zone *zone, struct page *page,
779 unsigned int order, int migratetype) {}
780#endif
781
782static inline void set_page_order(struct page *page, unsigned int order)
783{
784 set_page_private(page, order);
785 __SetPageBuddy(page);
786}
787
788/*
789 * This function checks whether a page is free && is the buddy
790 * we can coalesce a page and its buddy if
791 * (a) the buddy is not in a hole (check before calling!) &&
792 * (b) the buddy is in the buddy system &&
793 * (c) a page and its buddy have the same order &&
794 * (d) a page and its buddy are in the same zone.
795 *
796 * For recording whether a page is in the buddy system, we set PageBuddy.
797 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
798 *
799 * For recording page's order, we use page_private(page).
800 */
801static inline int page_is_buddy(struct page *page, struct page *buddy,
802 unsigned int order)
803{
804 if (page_is_guard(buddy) && page_order(buddy) == order) {
805 if (page_zone_id(page) != page_zone_id(buddy))
806 return 0;
807
808 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
809
810 return 1;
811 }
812
813 if (PageBuddy(buddy) && page_order(buddy) == order) {
814 /*
815 * zone check is done late to avoid uselessly
816 * calculating zone/node ids for pages that could
817 * never merge.
818 */
819 if (page_zone_id(page) != page_zone_id(buddy))
820 return 0;
821
822 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
823
824 return 1;
825 }
826 return 0;
827}
828
829#ifdef CONFIG_COMPACTION
830static inline struct capture_control *task_capc(struct zone *zone)
831{
832 struct capture_control *capc = current->capture_control;
833
834 return capc &&
835 !(current->flags & PF_KTHREAD) &&
836 !capc->page &&
837 capc->cc->zone == zone &&
838 capc->cc->direct_compaction ? capc : NULL;
839}
840
841static inline bool
842compaction_capture(struct capture_control *capc, struct page *page,
843 int order, int migratetype)
844{
845 if (!capc || order != capc->cc->order)
846 return false;
847
848 /* Do not accidentally pollute CMA or isolated regions*/
849 if (is_migrate_cma(migratetype) ||
850 is_migrate_isolate(migratetype))
851 return false;
852
853 /*
854 * Do not let lower order allocations polluate a movable pageblock.
855 * This might let an unmovable request use a reclaimable pageblock
856 * and vice-versa but no more than normal fallback logic which can
857 * have trouble finding a high-order free page.
858 */
859 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
860 return false;
861
862 capc->page = page;
863 return true;
864}
865
866#else
867static inline struct capture_control *task_capc(struct zone *zone)
868{
869 return NULL;
870}
871
872static inline bool
873compaction_capture(struct capture_control *capc, struct page *page,
874 int order, int migratetype)
875{
876 return false;
877}
878#endif /* CONFIG_COMPACTION */
879
880/*
881 * Freeing function for a buddy system allocator.
882 *
883 * The concept of a buddy system is to maintain direct-mapped table
884 * (containing bit values) for memory blocks of various "orders".
885 * The bottom level table contains the map for the smallest allocatable
886 * units of memory (here, pages), and each level above it describes
887 * pairs of units from the levels below, hence, "buddies".
888 * At a high level, all that happens here is marking the table entry
889 * at the bottom level available, and propagating the changes upward
890 * as necessary, plus some accounting needed to play nicely with other
891 * parts of the VM system.
892 * At each level, we keep a list of pages, which are heads of continuous
893 * free pages of length of (1 << order) and marked with PageBuddy.
894 * Page's order is recorded in page_private(page) field.
895 * So when we are allocating or freeing one, we can derive the state of the
896 * other. That is, if we allocate a small block, and both were
897 * free, the remainder of the region must be split into blocks.
898 * If a block is freed, and its buddy is also free, then this
899 * triggers coalescing into a block of larger size.
900 *
901 * -- nyc
902 */
903
904static inline void __free_one_page(struct page *page,
905 unsigned long pfn,
906 struct zone *zone, unsigned int order,
907 int migratetype)
908{
909 unsigned long combined_pfn;
910 unsigned long uninitialized_var(buddy_pfn);
911 struct page *buddy;
912 unsigned int max_order;
913 struct capture_control *capc = task_capc(zone);
914
915 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
916
917 VM_BUG_ON(!zone_is_initialized(zone));
918 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
919
920 VM_BUG_ON(migratetype == -1);
921 if (likely(!is_migrate_isolate(migratetype)))
922 __mod_zone_freepage_state(zone, 1 << order, migratetype);
923
924 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
925 VM_BUG_ON_PAGE(bad_range(zone, page), page);
926
927continue_merging:
928 while (order < max_order - 1) {
929 if (compaction_capture(capc, page, order, migratetype)) {
930 __mod_zone_freepage_state(zone, -(1 << order),
931 migratetype);
932 return;
933 }
934 buddy_pfn = __find_buddy_pfn(pfn, order);
935 buddy = page + (buddy_pfn - pfn);
936
937 if (!pfn_valid_within(buddy_pfn))
938 goto done_merging;
939 if (!page_is_buddy(page, buddy, order))
940 goto done_merging;
941 /*
942 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
943 * merge with it and move up one order.
944 */
945 if (page_is_guard(buddy))
946 clear_page_guard(zone, buddy, order, migratetype);
947 else
948 del_page_from_free_area(buddy, &zone->free_area[order]);
949 combined_pfn = buddy_pfn & pfn;
950 page = page + (combined_pfn - pfn);
951 pfn = combined_pfn;
952 order++;
953 }
954 if (max_order < MAX_ORDER) {
955 /* If we are here, it means order is >= pageblock_order.
956 * We want to prevent merge between freepages on isolate
957 * pageblock and normal pageblock. Without this, pageblock
958 * isolation could cause incorrect freepage or CMA accounting.
959 *
960 * We don't want to hit this code for the more frequent
961 * low-order merging.
962 */
963 if (unlikely(has_isolate_pageblock(zone))) {
964 int buddy_mt;
965
966 buddy_pfn = __find_buddy_pfn(pfn, order);
967 buddy = page + (buddy_pfn - pfn);
968 buddy_mt = get_pageblock_migratetype(buddy);
969
970 if (migratetype != buddy_mt
971 && (is_migrate_isolate(migratetype) ||
972 is_migrate_isolate(buddy_mt)))
973 goto done_merging;
974 }
975 max_order++;
976 goto continue_merging;
977 }
978
979done_merging:
980 set_page_order(page, order);
981
982 /*
983 * If this is not the largest possible page, check if the buddy
984 * of the next-highest order is free. If it is, it's possible
985 * that pages are being freed that will coalesce soon. In case,
986 * that is happening, add the free page to the tail of the list
987 * so it's less likely to be used soon and more likely to be merged
988 * as a higher order page
989 */
990 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)
991 && !is_shuffle_order(order)) {
992 struct page *higher_page, *higher_buddy;
993 combined_pfn = buddy_pfn & pfn;
994 higher_page = page + (combined_pfn - pfn);
995 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
996 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
997 if (pfn_valid_within(buddy_pfn) &&
998 page_is_buddy(higher_page, higher_buddy, order + 1)) {
999 add_to_free_area_tail(page, &zone->free_area[order],
1000 migratetype);
1001 return;
1002 }
1003 }
1004
1005 if (is_shuffle_order(order))
1006 add_to_free_area_random(page, &zone->free_area[order],
1007 migratetype);
1008 else
1009 add_to_free_area(page, &zone->free_area[order], migratetype);
1010
1011}
1012
1013/*
1014 * A bad page could be due to a number of fields. Instead of multiple branches,
1015 * try and check multiple fields with one check. The caller must do a detailed
1016 * check if necessary.
1017 */
1018static inline bool page_expected_state(struct page *page,
1019 unsigned long check_flags)
1020{
1021 if (unlikely(atomic_read(&page->_mapcount) != -1))
1022 return false;
1023
1024 if (unlikely((unsigned long)page->mapping |
1025 page_ref_count(page) |
1026#ifdef CONFIG_MEMCG
1027 (unsigned long)page->mem_cgroup |
1028#endif
1029 (page->flags & check_flags)))
1030 return false;
1031
1032 return true;
1033}
1034
1035static void free_pages_check_bad(struct page *page)
1036{
1037 const char *bad_reason;
1038 unsigned long bad_flags;
1039
1040 bad_reason = NULL;
1041 bad_flags = 0;
1042
1043 if (unlikely(atomic_read(&page->_mapcount) != -1))
1044 bad_reason = "nonzero mapcount";
1045 if (unlikely(page->mapping != NULL))
1046 bad_reason = "non-NULL mapping";
1047 if (unlikely(page_ref_count(page) != 0))
1048 bad_reason = "nonzero _refcount";
1049 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1050 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1051 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1052 }
1053#ifdef CONFIG_MEMCG
1054 if (unlikely(page->mem_cgroup))
1055 bad_reason = "page still charged to cgroup";
1056#endif
1057 bad_page(page, bad_reason, bad_flags);
1058}
1059
1060static inline int free_pages_check(struct page *page)
1061{
1062 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1063 return 0;
1064
1065 /* Something has gone sideways, find it */
1066 free_pages_check_bad(page);
1067 return 1;
1068}
1069
1070static int free_tail_pages_check(struct page *head_page, struct page *page)
1071{
1072 int ret = 1;
1073
1074 /*
1075 * We rely page->lru.next never has bit 0 set, unless the page
1076 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1077 */
1078 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1079
1080 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1081 ret = 0;
1082 goto out;
1083 }
1084 switch (page - head_page) {
1085 case 1:
1086 /* the first tail page: ->mapping may be compound_mapcount() */
1087 if (unlikely(compound_mapcount(page))) {
1088 bad_page(page, "nonzero compound_mapcount", 0);
1089 goto out;
1090 }
1091 break;
1092 case 2:
1093 /*
1094 * the second tail page: ->mapping is
1095 * deferred_list.next -- ignore value.
1096 */
1097 break;
1098 default:
1099 if (page->mapping != TAIL_MAPPING) {
1100 bad_page(page, "corrupted mapping in tail page", 0);
1101 goto out;
1102 }
1103 break;
1104 }
1105 if (unlikely(!PageTail(page))) {
1106 bad_page(page, "PageTail not set", 0);
1107 goto out;
1108 }
1109 if (unlikely(compound_head(page) != head_page)) {
1110 bad_page(page, "compound_head not consistent", 0);
1111 goto out;
1112 }
1113 ret = 0;
1114out:
1115 page->mapping = NULL;
1116 clear_compound_head(page);
1117 return ret;
1118}
1119
1120static void kernel_init_free_pages(struct page *page, int numpages)
1121{
1122 int i;
1123
1124 for (i = 0; i < numpages; i++)
1125 clear_highpage(page + i);
1126}
1127
1128static __always_inline bool free_pages_prepare(struct page *page,
1129 unsigned int order, bool check_free)
1130{
1131 int bad = 0;
1132
1133 VM_BUG_ON_PAGE(PageTail(page), page);
1134
1135 trace_mm_page_free(page, order);
1136
1137 /*
1138 * Check tail pages before head page information is cleared to
1139 * avoid checking PageCompound for order-0 pages.
1140 */
1141 if (unlikely(order)) {
1142 bool compound = PageCompound(page);
1143 int i;
1144
1145 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1146
1147 if (compound)
1148 ClearPageDoubleMap(page);
1149 for (i = 1; i < (1 << order); i++) {
1150 if (compound)
1151 bad += free_tail_pages_check(page, page + i);
1152 if (unlikely(free_pages_check(page + i))) {
1153 bad++;
1154 continue;
1155 }
1156 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1157 }
1158 }
1159 if (PageMappingFlags(page))
1160 page->mapping = NULL;
1161 if (memcg_kmem_enabled() && PageKmemcg(page))
1162 __memcg_kmem_uncharge(page, order);
1163 if (check_free)
1164 bad += free_pages_check(page);
1165 if (bad)
1166 return false;
1167
1168 page_cpupid_reset_last(page);
1169 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1170 reset_page_owner(page, order);
1171
1172 if (!PageHighMem(page)) {
1173 debug_check_no_locks_freed(page_address(page),
1174 PAGE_SIZE << order);
1175 debug_check_no_obj_freed(page_address(page),
1176 PAGE_SIZE << order);
1177 }
1178 if (want_init_on_free())
1179 kernel_init_free_pages(page, 1 << order);
1180
1181 kernel_poison_pages(page, 1 << order, 0);
1182 /*
1183 * arch_free_page() can make the page's contents inaccessible. s390
1184 * does this. So nothing which can access the page's contents should
1185 * happen after this.
1186 */
1187 arch_free_page(page, order);
1188
1189 if (debug_pagealloc_enabled())
1190 kernel_map_pages(page, 1 << order, 0);
1191
1192 kasan_free_nondeferred_pages(page, order);
1193
1194 return true;
1195}
1196
1197#ifdef CONFIG_DEBUG_VM
1198/*
1199 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1200 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1201 * moved from pcp lists to free lists.
1202 */
1203static bool free_pcp_prepare(struct page *page)
1204{
1205 return free_pages_prepare(page, 0, true);
1206}
1207
1208static bool bulkfree_pcp_prepare(struct page *page)
1209{
1210 if (debug_pagealloc_enabled())
1211 return free_pages_check(page);
1212 else
1213 return false;
1214}
1215#else
1216/*
1217 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1218 * moving from pcp lists to free list in order to reduce overhead. With
1219 * debug_pagealloc enabled, they are checked also immediately when being freed
1220 * to the pcp lists.
1221 */
1222static bool free_pcp_prepare(struct page *page)
1223{
1224 if (debug_pagealloc_enabled())
1225 return free_pages_prepare(page, 0, true);
1226 else
1227 return free_pages_prepare(page, 0, false);
1228}
1229
1230static bool bulkfree_pcp_prepare(struct page *page)
1231{
1232 return free_pages_check(page);
1233}
1234#endif /* CONFIG_DEBUG_VM */
1235
1236static inline void prefetch_buddy(struct page *page)
1237{
1238 unsigned long pfn = page_to_pfn(page);
1239 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1240 struct page *buddy = page + (buddy_pfn - pfn);
1241
1242 prefetch(buddy);
1243}
1244
1245/*
1246 * Frees a number of pages from the PCP lists
1247 * Assumes all pages on list are in same zone, and of same order.
1248 * count is the number of pages to free.
1249 *
1250 * If the zone was previously in an "all pages pinned" state then look to
1251 * see if this freeing clears that state.
1252 *
1253 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1254 * pinned" detection logic.
1255 */
1256static void free_pcppages_bulk(struct zone *zone, int count,
1257 struct per_cpu_pages *pcp)
1258{
1259 int migratetype = 0;
1260 int batch_free = 0;
1261 int prefetch_nr = 0;
1262 bool isolated_pageblocks;
1263 struct page *page, *tmp;
1264 LIST_HEAD(head);
1265
1266 while (count) {
1267 struct list_head *list;
1268
1269 /*
1270 * Remove pages from lists in a round-robin fashion. A
1271 * batch_free count is maintained that is incremented when an
1272 * empty list is encountered. This is so more pages are freed
1273 * off fuller lists instead of spinning excessively around empty
1274 * lists
1275 */
1276 do {
1277 batch_free++;
1278 if (++migratetype == MIGRATE_PCPTYPES)
1279 migratetype = 0;
1280 list = &pcp->lists[migratetype];
1281 } while (list_empty(list));
1282
1283 /* This is the only non-empty list. Free them all. */
1284 if (batch_free == MIGRATE_PCPTYPES)
1285 batch_free = count;
1286
1287 do {
1288 page = list_last_entry(list, struct page, lru);
1289 /* must delete to avoid corrupting pcp list */
1290 list_del(&page->lru);
1291 pcp->count--;
1292
1293 if (bulkfree_pcp_prepare(page))
1294 continue;
1295
1296 list_add_tail(&page->lru, &head);
1297
1298 /*
1299 * We are going to put the page back to the global
1300 * pool, prefetch its buddy to speed up later access
1301 * under zone->lock. It is believed the overhead of
1302 * an additional test and calculating buddy_pfn here
1303 * can be offset by reduced memory latency later. To
1304 * avoid excessive prefetching due to large count, only
1305 * prefetch buddy for the first pcp->batch nr of pages.
1306 */
1307 if (prefetch_nr++ < pcp->batch)
1308 prefetch_buddy(page);
1309 } while (--count && --batch_free && !list_empty(list));
1310 }
1311
1312 spin_lock(&zone->lock);
1313 isolated_pageblocks = has_isolate_pageblock(zone);
1314
1315 /*
1316 * Use safe version since after __free_one_page(),
1317 * page->lru.next will not point to original list.
1318 */
1319 list_for_each_entry_safe(page, tmp, &head, lru) {
1320 int mt = get_pcppage_migratetype(page);
1321 /* MIGRATE_ISOLATE page should not go to pcplists */
1322 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1323 /* Pageblock could have been isolated meanwhile */
1324 if (unlikely(isolated_pageblocks))
1325 mt = get_pageblock_migratetype(page);
1326
1327 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1328 trace_mm_page_pcpu_drain(page, 0, mt);
1329 }
1330 spin_unlock(&zone->lock);
1331}
1332
1333static void free_one_page(struct zone *zone,
1334 struct page *page, unsigned long pfn,
1335 unsigned int order,
1336 int migratetype)
1337{
1338 spin_lock(&zone->lock);
1339 if (unlikely(has_isolate_pageblock(zone) ||
1340 is_migrate_isolate(migratetype))) {
1341 migratetype = get_pfnblock_migratetype(page, pfn);
1342 }
1343 __free_one_page(page, pfn, zone, order, migratetype);
1344 spin_unlock(&zone->lock);
1345}
1346
1347static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1348 unsigned long zone, int nid)
1349{
1350 mm_zero_struct_page(page);
1351 set_page_links(page, zone, nid, pfn);
1352 init_page_count(page);
1353 page_mapcount_reset(page);
1354 page_cpupid_reset_last(page);
1355 page_kasan_tag_reset(page);
1356
1357 INIT_LIST_HEAD(&page->lru);
1358#ifdef WANT_PAGE_VIRTUAL
1359 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1360 if (!is_highmem_idx(zone))
1361 set_page_address(page, __va(pfn << PAGE_SHIFT));
1362#endif
1363}
1364
1365#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1366static void __meminit init_reserved_page(unsigned long pfn)
1367{
1368 pg_data_t *pgdat;
1369 int nid, zid;
1370
1371 if (!early_page_uninitialised(pfn))
1372 return;
1373
1374 nid = early_pfn_to_nid(pfn);
1375 pgdat = NODE_DATA(nid);
1376
1377 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1378 struct zone *zone = &pgdat->node_zones[zid];
1379
1380 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1381 break;
1382 }
1383 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1384}
1385#else
1386static inline void init_reserved_page(unsigned long pfn)
1387{
1388}
1389#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1390
1391/*
1392 * Initialised pages do not have PageReserved set. This function is
1393 * called for each range allocated by the bootmem allocator and
1394 * marks the pages PageReserved. The remaining valid pages are later
1395 * sent to the buddy page allocator.
1396 */
1397void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1398{
1399 unsigned long start_pfn = PFN_DOWN(start);
1400 unsigned long end_pfn = PFN_UP(end);
1401
1402 for (; start_pfn < end_pfn; start_pfn++) {
1403 if (pfn_valid(start_pfn)) {
1404 struct page *page = pfn_to_page(start_pfn);
1405
1406 init_reserved_page(start_pfn);
1407
1408 /* Avoid false-positive PageTail() */
1409 INIT_LIST_HEAD(&page->lru);
1410
1411 /*
1412 * no need for atomic set_bit because the struct
1413 * page is not visible yet so nobody should
1414 * access it yet.
1415 */
1416 __SetPageReserved(page);
1417 }
1418 }
1419}
1420
1421static void __free_pages_ok(struct page *page, unsigned int order)
1422{
1423 unsigned long flags;
1424 int migratetype;
1425 unsigned long pfn = page_to_pfn(page);
1426
1427 if (!free_pages_prepare(page, order, true))
1428 return;
1429
1430 migratetype = get_pfnblock_migratetype(page, pfn);
1431 local_irq_save(flags);
1432 __count_vm_events(PGFREE, 1 << order);
1433 free_one_page(page_zone(page), page, pfn, order, migratetype);
1434 local_irq_restore(flags);
1435}
1436
1437void __free_pages_core(struct page *page, unsigned int order)
1438{
1439 unsigned int nr_pages = 1 << order;
1440 struct page *p = page;
1441 unsigned int loop;
1442
1443 prefetchw(p);
1444 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1445 prefetchw(p + 1);
1446 __ClearPageReserved(p);
1447 set_page_count(p, 0);
1448 }
1449 __ClearPageReserved(p);
1450 set_page_count(p, 0);
1451
1452 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1453 set_page_refcounted(page);
1454 __free_pages(page, order);
1455}
1456
1457#if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1458 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1459
1460static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1461
1462int __meminit early_pfn_to_nid(unsigned long pfn)
1463{
1464 static DEFINE_SPINLOCK(early_pfn_lock);
1465 int nid;
1466
1467 spin_lock(&early_pfn_lock);
1468 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1469 if (nid < 0)
1470 nid = first_online_node;
1471 spin_unlock(&early_pfn_lock);
1472
1473 return nid;
1474}
1475#endif
1476
1477#ifdef CONFIG_NODES_SPAN_OTHER_NODES
1478/* Only safe to use early in boot when initialisation is single-threaded */
1479static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1480{
1481 int nid;
1482
1483 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1484 if (nid >= 0 && nid != node)
1485 return false;
1486 return true;
1487}
1488
1489#else
1490static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1491{
1492 return true;
1493}
1494#endif
1495
1496
1497void __init memblock_free_pages(struct page *page, unsigned long pfn,
1498 unsigned int order)
1499{
1500 if (early_page_uninitialised(pfn))
1501 return;
1502 __free_pages_core(page, order);
1503}
1504
1505/*
1506 * Check that the whole (or subset of) a pageblock given by the interval of
1507 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1508 * with the migration of free compaction scanner. The scanners then need to
1509 * use only pfn_valid_within() check for arches that allow holes within
1510 * pageblocks.
1511 *
1512 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1513 *
1514 * It's possible on some configurations to have a setup like node0 node1 node0
1515 * i.e. it's possible that all pages within a zones range of pages do not
1516 * belong to a single zone. We assume that a border between node0 and node1
1517 * can occur within a single pageblock, but not a node0 node1 node0
1518 * interleaving within a single pageblock. It is therefore sufficient to check
1519 * the first and last page of a pageblock and avoid checking each individual
1520 * page in a pageblock.
1521 */
1522struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1523 unsigned long end_pfn, struct zone *zone)
1524{
1525 struct page *start_page;
1526 struct page *end_page;
1527
1528 /* end_pfn is one past the range we are checking */
1529 end_pfn--;
1530
1531 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1532 return NULL;
1533
1534 start_page = pfn_to_online_page(start_pfn);
1535 if (!start_page)
1536 return NULL;
1537
1538 if (page_zone(start_page) != zone)
1539 return NULL;
1540
1541 end_page = pfn_to_page(end_pfn);
1542
1543 /* This gives a shorter code than deriving page_zone(end_page) */
1544 if (page_zone_id(start_page) != page_zone_id(end_page))
1545 return NULL;
1546
1547 return start_page;
1548}
1549
1550void set_zone_contiguous(struct zone *zone)
1551{
1552 unsigned long block_start_pfn = zone->zone_start_pfn;
1553 unsigned long block_end_pfn;
1554
1555 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1556 for (; block_start_pfn < zone_end_pfn(zone);
1557 block_start_pfn = block_end_pfn,
1558 block_end_pfn += pageblock_nr_pages) {
1559
1560 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1561
1562 if (!__pageblock_pfn_to_page(block_start_pfn,
1563 block_end_pfn, zone))
1564 return;
1565 }
1566
1567 /* We confirm that there is no hole */
1568 zone->contiguous = true;
1569}
1570
1571void clear_zone_contiguous(struct zone *zone)
1572{
1573 zone->contiguous = false;
1574}
1575
1576#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1577static void __init deferred_free_range(unsigned long pfn,
1578 unsigned long nr_pages)
1579{
1580 struct page *page;
1581 unsigned long i;
1582
1583 if (!nr_pages)
1584 return;
1585
1586 page = pfn_to_page(pfn);
1587
1588 /* Free a large naturally-aligned chunk if possible */
1589 if (nr_pages == pageblock_nr_pages &&
1590 (pfn & (pageblock_nr_pages - 1)) == 0) {
1591 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1592 __free_pages_core(page, pageblock_order);
1593 return;
1594 }
1595
1596 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1597 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1598 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1599 __free_pages_core(page, 0);
1600 }
1601}
1602
1603/* Completion tracking for deferred_init_memmap() threads */
1604static atomic_t pgdat_init_n_undone __initdata;
1605static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1606
1607static inline void __init pgdat_init_report_one_done(void)
1608{
1609 if (atomic_dec_and_test(&pgdat_init_n_undone))
1610 complete(&pgdat_init_all_done_comp);
1611}
1612
1613/*
1614 * Returns true if page needs to be initialized or freed to buddy allocator.
1615 *
1616 * First we check if pfn is valid on architectures where it is possible to have
1617 * holes within pageblock_nr_pages. On systems where it is not possible, this
1618 * function is optimized out.
1619 *
1620 * Then, we check if a current large page is valid by only checking the validity
1621 * of the head pfn.
1622 */
1623static inline bool __init deferred_pfn_valid(unsigned long pfn)
1624{
1625 if (!pfn_valid_within(pfn))
1626 return false;
1627 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1628 return false;
1629 return true;
1630}
1631
1632/*
1633 * Free pages to buddy allocator. Try to free aligned pages in
1634 * pageblock_nr_pages sizes.
1635 */
1636static void __init deferred_free_pages(unsigned long pfn,
1637 unsigned long end_pfn)
1638{
1639 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1640 unsigned long nr_free = 0;
1641
1642 for (; pfn < end_pfn; pfn++) {
1643 if (!deferred_pfn_valid(pfn)) {
1644 deferred_free_range(pfn - nr_free, nr_free);
1645 nr_free = 0;
1646 } else if (!(pfn & nr_pgmask)) {
1647 deferred_free_range(pfn - nr_free, nr_free);
1648 nr_free = 1;
1649 touch_nmi_watchdog();
1650 } else {
1651 nr_free++;
1652 }
1653 }
1654 /* Free the last block of pages to allocator */
1655 deferred_free_range(pfn - nr_free, nr_free);
1656}
1657
1658/*
1659 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1660 * by performing it only once every pageblock_nr_pages.
1661 * Return number of pages initialized.
1662 */
1663static unsigned long __init deferred_init_pages(struct zone *zone,
1664 unsigned long pfn,
1665 unsigned long end_pfn)
1666{
1667 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1668 int nid = zone_to_nid(zone);
1669 unsigned long nr_pages = 0;
1670 int zid = zone_idx(zone);
1671 struct page *page = NULL;
1672
1673 for (; pfn < end_pfn; pfn++) {
1674 if (!deferred_pfn_valid(pfn)) {
1675 page = NULL;
1676 continue;
1677 } else if (!page || !(pfn & nr_pgmask)) {
1678 page = pfn_to_page(pfn);
1679 touch_nmi_watchdog();
1680 } else {
1681 page++;
1682 }
1683 __init_single_page(page, pfn, zid, nid);
1684 nr_pages++;
1685 }
1686 return (nr_pages);
1687}
1688
1689/*
1690 * This function is meant to pre-load the iterator for the zone init.
1691 * Specifically it walks through the ranges until we are caught up to the
1692 * first_init_pfn value and exits there. If we never encounter the value we
1693 * return false indicating there are no valid ranges left.
1694 */
1695static bool __init
1696deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1697 unsigned long *spfn, unsigned long *epfn,
1698 unsigned long first_init_pfn)
1699{
1700 u64 j;
1701
1702 /*
1703 * Start out by walking through the ranges in this zone that have
1704 * already been initialized. We don't need to do anything with them
1705 * so we just need to flush them out of the system.
1706 */
1707 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1708 if (*epfn <= first_init_pfn)
1709 continue;
1710 if (*spfn < first_init_pfn)
1711 *spfn = first_init_pfn;
1712 *i = j;
1713 return true;
1714 }
1715
1716 return false;
1717}
1718
1719/*
1720 * Initialize and free pages. We do it in two loops: first we initialize
1721 * struct page, then free to buddy allocator, because while we are
1722 * freeing pages we can access pages that are ahead (computing buddy
1723 * page in __free_one_page()).
1724 *
1725 * In order to try and keep some memory in the cache we have the loop
1726 * broken along max page order boundaries. This way we will not cause
1727 * any issues with the buddy page computation.
1728 */
1729static unsigned long __init
1730deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1731 unsigned long *end_pfn)
1732{
1733 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1734 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1735 unsigned long nr_pages = 0;
1736 u64 j = *i;
1737
1738 /* First we loop through and initialize the page values */
1739 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1740 unsigned long t;
1741
1742 if (mo_pfn <= *start_pfn)
1743 break;
1744
1745 t = min(mo_pfn, *end_pfn);
1746 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1747
1748 if (mo_pfn < *end_pfn) {
1749 *start_pfn = mo_pfn;
1750 break;
1751 }
1752 }
1753
1754 /* Reset values and now loop through freeing pages as needed */
1755 swap(j, *i);
1756
1757 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1758 unsigned long t;
1759
1760 if (mo_pfn <= spfn)
1761 break;
1762
1763 t = min(mo_pfn, epfn);
1764 deferred_free_pages(spfn, t);
1765
1766 if (mo_pfn <= epfn)
1767 break;
1768 }
1769
1770 return nr_pages;
1771}
1772
1773/* Initialise remaining memory on a node */
1774static int __init deferred_init_memmap(void *data)
1775{
1776 pg_data_t *pgdat = data;
1777 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1778 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1779 unsigned long first_init_pfn, flags;
1780 unsigned long start = jiffies;
1781 struct zone *zone;
1782 int zid;
1783 u64 i;
1784
1785 /* Bind memory initialisation thread to a local node if possible */
1786 if (!cpumask_empty(cpumask))
1787 set_cpus_allowed_ptr(current, cpumask);
1788
1789 pgdat_resize_lock(pgdat, &flags);
1790 first_init_pfn = pgdat->first_deferred_pfn;
1791 if (first_init_pfn == ULONG_MAX) {
1792 pgdat_resize_unlock(pgdat, &flags);
1793 pgdat_init_report_one_done();
1794 return 0;
1795 }
1796
1797 /* Sanity check boundaries */
1798 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1799 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1800 pgdat->first_deferred_pfn = ULONG_MAX;
1801
1802 /* Only the highest zone is deferred so find it */
1803 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1804 zone = pgdat->node_zones + zid;
1805 if (first_init_pfn < zone_end_pfn(zone))
1806 break;
1807 }
1808
1809 /* If the zone is empty somebody else may have cleared out the zone */
1810 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1811 first_init_pfn))
1812 goto zone_empty;
1813
1814 /*
1815 * Initialize and free pages in MAX_ORDER sized increments so
1816 * that we can avoid introducing any issues with the buddy
1817 * allocator.
1818 */
1819 while (spfn < epfn)
1820 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1821zone_empty:
1822 pgdat_resize_unlock(pgdat, &flags);
1823
1824 /* Sanity check that the next zone really is unpopulated */
1825 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1826
1827 pr_info("node %d initialised, %lu pages in %ums\n",
1828 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1829
1830 pgdat_init_report_one_done();
1831 return 0;
1832}
1833
1834/*
1835 * If this zone has deferred pages, try to grow it by initializing enough
1836 * deferred pages to satisfy the allocation specified by order, rounded up to
1837 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1838 * of SECTION_SIZE bytes by initializing struct pages in increments of
1839 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1840 *
1841 * Return true when zone was grown, otherwise return false. We return true even
1842 * when we grow less than requested, to let the caller decide if there are
1843 * enough pages to satisfy the allocation.
1844 *
1845 * Note: We use noinline because this function is needed only during boot, and
1846 * it is called from a __ref function _deferred_grow_zone. This way we are
1847 * making sure that it is not inlined into permanent text section.
1848 */
1849static noinline bool __init
1850deferred_grow_zone(struct zone *zone, unsigned int order)
1851{
1852 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1853 pg_data_t *pgdat = zone->zone_pgdat;
1854 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1855 unsigned long spfn, epfn, flags;
1856 unsigned long nr_pages = 0;
1857 u64 i;
1858
1859 /* Only the last zone may have deferred pages */
1860 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1861 return false;
1862
1863 pgdat_resize_lock(pgdat, &flags);
1864
1865 /*
1866 * If deferred pages have been initialized while we were waiting for
1867 * the lock, return true, as the zone was grown. The caller will retry
1868 * this zone. We won't return to this function since the caller also
1869 * has this static branch.
1870 */
1871 if (!static_branch_unlikely(&deferred_pages)) {
1872 pgdat_resize_unlock(pgdat, &flags);
1873 return true;
1874 }
1875
1876 /*
1877 * If someone grew this zone while we were waiting for spinlock, return
1878 * true, as there might be enough pages already.
1879 */
1880 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1881 pgdat_resize_unlock(pgdat, &flags);
1882 return true;
1883 }
1884
1885 /* If the zone is empty somebody else may have cleared out the zone */
1886 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1887 first_deferred_pfn)) {
1888 pgdat->first_deferred_pfn = ULONG_MAX;
1889 pgdat_resize_unlock(pgdat, &flags);
1890 /* Retry only once. */
1891 return first_deferred_pfn != ULONG_MAX;
1892 }
1893
1894 /*
1895 * Initialize and free pages in MAX_ORDER sized increments so
1896 * that we can avoid introducing any issues with the buddy
1897 * allocator.
1898 */
1899 while (spfn < epfn) {
1900 /* update our first deferred PFN for this section */
1901 first_deferred_pfn = spfn;
1902
1903 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1904
1905 /* We should only stop along section boundaries */
1906 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1907 continue;
1908
1909 /* If our quota has been met we can stop here */
1910 if (nr_pages >= nr_pages_needed)
1911 break;
1912 }
1913
1914 pgdat->first_deferred_pfn = spfn;
1915 pgdat_resize_unlock(pgdat, &flags);
1916
1917 return nr_pages > 0;
1918}
1919
1920/*
1921 * deferred_grow_zone() is __init, but it is called from
1922 * get_page_from_freelist() during early boot until deferred_pages permanently
1923 * disables this call. This is why we have refdata wrapper to avoid warning,
1924 * and to ensure that the function body gets unloaded.
1925 */
1926static bool __ref
1927_deferred_grow_zone(struct zone *zone, unsigned int order)
1928{
1929 return deferred_grow_zone(zone, order);
1930}
1931
1932#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1933
1934void __init page_alloc_init_late(void)
1935{
1936 struct zone *zone;
1937 int nid;
1938
1939#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1940
1941 /* There will be num_node_state(N_MEMORY) threads */
1942 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1943 for_each_node_state(nid, N_MEMORY) {
1944 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1945 }
1946
1947 /* Block until all are initialised */
1948 wait_for_completion(&pgdat_init_all_done_comp);
1949
1950 /*
1951 * The number of managed pages has changed due to the initialisation
1952 * so the pcpu batch and high limits needs to be updated or the limits
1953 * will be artificially small.
1954 */
1955 for_each_populated_zone(zone)
1956 zone_pcp_update(zone);
1957
1958 /*
1959 * We initialized the rest of the deferred pages. Permanently disable
1960 * on-demand struct page initialization.
1961 */
1962 static_branch_disable(&deferred_pages);
1963
1964 /* Reinit limits that are based on free pages after the kernel is up */
1965 files_maxfiles_init();
1966#endif
1967
1968 /* Discard memblock private memory */
1969 memblock_discard();
1970
1971 for_each_node_state(nid, N_MEMORY)
1972 shuffle_free_memory(NODE_DATA(nid));
1973
1974 for_each_populated_zone(zone)
1975 set_zone_contiguous(zone);
1976
1977#ifdef CONFIG_DEBUG_PAGEALLOC
1978 init_debug_guardpage();
1979#endif
1980}
1981
1982#ifdef CONFIG_CMA
1983/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1984void __init init_cma_reserved_pageblock(struct page *page)
1985{
1986 unsigned i = pageblock_nr_pages;
1987 struct page *p = page;
1988
1989 do {
1990 __ClearPageReserved(p);
1991 set_page_count(p, 0);
1992 } while (++p, --i);
1993
1994 set_pageblock_migratetype(page, MIGRATE_CMA);
1995
1996 if (pageblock_order >= MAX_ORDER) {
1997 i = pageblock_nr_pages;
1998 p = page;
1999 do {
2000 set_page_refcounted(p);
2001 __free_pages(p, MAX_ORDER - 1);
2002 p += MAX_ORDER_NR_PAGES;
2003 } while (i -= MAX_ORDER_NR_PAGES);
2004 } else {
2005 set_page_refcounted(page);
2006 __free_pages(page, pageblock_order);
2007 }
2008
2009 adjust_managed_page_count(page, pageblock_nr_pages);
2010}
2011#endif
2012
2013/*
2014 * The order of subdivision here is critical for the IO subsystem.
2015 * Please do not alter this order without good reasons and regression
2016 * testing. Specifically, as large blocks of memory are subdivided,
2017 * the order in which smaller blocks are delivered depends on the order
2018 * they're subdivided in this function. This is the primary factor
2019 * influencing the order in which pages are delivered to the IO
2020 * subsystem according to empirical testing, and this is also justified
2021 * by considering the behavior of a buddy system containing a single
2022 * large block of memory acted on by a series of small allocations.
2023 * This behavior is a critical factor in sglist merging's success.
2024 *
2025 * -- nyc
2026 */
2027static inline void expand(struct zone *zone, struct page *page,
2028 int low, int high, struct free_area *area,
2029 int migratetype)
2030{
2031 unsigned long size = 1 << high;
2032
2033 while (high > low) {
2034 area--;
2035 high--;
2036 size >>= 1;
2037 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2038
2039 /*
2040 * Mark as guard pages (or page), that will allow to
2041 * merge back to allocator when buddy will be freed.
2042 * Corresponding page table entries will not be touched,
2043 * pages will stay not present in virtual address space
2044 */
2045 if (set_page_guard(zone, &page[size], high, migratetype))
2046 continue;
2047
2048 add_to_free_area(&page[size], area, migratetype);
2049 set_page_order(&page[size], high);
2050 }
2051}
2052
2053static void check_new_page_bad(struct page *page)
2054{
2055 const char *bad_reason = NULL;
2056 unsigned long bad_flags = 0;
2057
2058 if (unlikely(atomic_read(&page->_mapcount) != -1))
2059 bad_reason = "nonzero mapcount";
2060 if (unlikely(page->mapping != NULL))
2061 bad_reason = "non-NULL mapping";
2062 if (unlikely(page_ref_count(page) != 0))
2063 bad_reason = "nonzero _refcount";
2064 if (unlikely(page->flags & __PG_HWPOISON)) {
2065 bad_reason = "HWPoisoned (hardware-corrupted)";
2066 bad_flags = __PG_HWPOISON;
2067 /* Don't complain about hwpoisoned pages */
2068 page_mapcount_reset(page); /* remove PageBuddy */
2069 return;
2070 }
2071 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2072 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2073 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2074 }
2075#ifdef CONFIG_MEMCG
2076 if (unlikely(page->mem_cgroup))
2077 bad_reason = "page still charged to cgroup";
2078#endif
2079 bad_page(page, bad_reason, bad_flags);
2080}
2081
2082/*
2083 * This page is about to be returned from the page allocator
2084 */
2085static inline int check_new_page(struct page *page)
2086{
2087 if (likely(page_expected_state(page,
2088 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2089 return 0;
2090
2091 check_new_page_bad(page);
2092 return 1;
2093}
2094
2095static inline bool free_pages_prezeroed(void)
2096{
2097 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2098 page_poisoning_enabled()) || want_init_on_free();
2099}
2100
2101#ifdef CONFIG_DEBUG_VM
2102/*
2103 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2104 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2105 * also checked when pcp lists are refilled from the free lists.
2106 */
2107static inline bool check_pcp_refill(struct page *page)
2108{
2109 if (debug_pagealloc_enabled())
2110 return check_new_page(page);
2111 else
2112 return false;
2113}
2114
2115static inline bool check_new_pcp(struct page *page)
2116{
2117 return check_new_page(page);
2118}
2119#else
2120/*
2121 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2122 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2123 * enabled, they are also checked when being allocated from the pcp lists.
2124 */
2125static inline bool check_pcp_refill(struct page *page)
2126{
2127 return check_new_page(page);
2128}
2129static inline bool check_new_pcp(struct page *page)
2130{
2131 if (debug_pagealloc_enabled())
2132 return check_new_page(page);
2133 else
2134 return false;
2135}
2136#endif /* CONFIG_DEBUG_VM */
2137
2138static bool check_new_pages(struct page *page, unsigned int order)
2139{
2140 int i;
2141 for (i = 0; i < (1 << order); i++) {
2142 struct page *p = page + i;
2143
2144 if (unlikely(check_new_page(p)))
2145 return true;
2146 }
2147
2148 return false;
2149}
2150
2151inline void post_alloc_hook(struct page *page, unsigned int order,
2152 gfp_t gfp_flags)
2153{
2154 set_page_private(page, 0);
2155 set_page_refcounted(page);
2156
2157 arch_alloc_page(page, order);
2158 if (debug_pagealloc_enabled())
2159 kernel_map_pages(page, 1 << order, 1);
2160 kasan_alloc_pages(page, order);
2161 kernel_poison_pages(page, 1 << order, 1);
2162 set_page_owner(page, order, gfp_flags);
2163}
2164
2165static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2166 unsigned int alloc_flags)
2167{
2168 post_alloc_hook(page, order, gfp_flags);
2169
2170 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2171 kernel_init_free_pages(page, 1 << order);
2172
2173 if (order && (gfp_flags & __GFP_COMP))
2174 prep_compound_page(page, order);
2175
2176 /*
2177 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2178 * allocate the page. The expectation is that the caller is taking
2179 * steps that will free more memory. The caller should avoid the page
2180 * being used for !PFMEMALLOC purposes.
2181 */
2182 if (alloc_flags & ALLOC_NO_WATERMARKS)
2183 set_page_pfmemalloc(page);
2184 else
2185 clear_page_pfmemalloc(page);
2186}
2187
2188/*
2189 * Go through the free lists for the given migratetype and remove
2190 * the smallest available page from the freelists
2191 */
2192static __always_inline
2193struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2194 int migratetype)
2195{
2196 unsigned int current_order;
2197 struct free_area *area;
2198 struct page *page;
2199
2200 /* Find a page of the appropriate size in the preferred list */
2201 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2202 area = &(zone->free_area[current_order]);
2203 page = get_page_from_free_area(area, migratetype);
2204 if (!page)
2205 continue;
2206 del_page_from_free_area(page, area);
2207 expand(zone, page, order, current_order, area, migratetype);
2208 set_pcppage_migratetype(page, migratetype);
2209 return page;
2210 }
2211
2212 return NULL;
2213}
2214
2215
2216/*
2217 * This array describes the order lists are fallen back to when
2218 * the free lists for the desirable migrate type are depleted
2219 */
2220static int fallbacks[MIGRATE_TYPES][4] = {
2221 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2222 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2223 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2224#ifdef CONFIG_CMA
2225 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2226#endif
2227#ifdef CONFIG_MEMORY_ISOLATION
2228 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2229#endif
2230};
2231
2232#ifdef CONFIG_CMA
2233static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2234 unsigned int order)
2235{
2236 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2237}
2238#else
2239static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2240 unsigned int order) { return NULL; }
2241#endif
2242
2243/*
2244 * Move the free pages in a range to the free lists of the requested type.
2245 * Note that start_page and end_pages are not aligned on a pageblock
2246 * boundary. If alignment is required, use move_freepages_block()
2247 */
2248static int move_freepages(struct zone *zone,
2249 struct page *start_page, struct page *end_page,
2250 int migratetype, int *num_movable)
2251{
2252 struct page *page;
2253 unsigned int order;
2254 int pages_moved = 0;
2255
2256 for (page = start_page; page <= end_page;) {
2257 if (!pfn_valid_within(page_to_pfn(page))) {
2258 page++;
2259 continue;
2260 }
2261
2262 if (!PageBuddy(page)) {
2263 /*
2264 * We assume that pages that could be isolated for
2265 * migration are movable. But we don't actually try
2266 * isolating, as that would be expensive.
2267 */
2268 if (num_movable &&
2269 (PageLRU(page) || __PageMovable(page)))
2270 (*num_movable)++;
2271
2272 page++;
2273 continue;
2274 }
2275
2276 /* Make sure we are not inadvertently changing nodes */
2277 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2278 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2279
2280 order = page_order(page);
2281 move_to_free_area(page, &zone->free_area[order], migratetype);
2282 page += 1 << order;
2283 pages_moved += 1 << order;
2284 }
2285
2286 return pages_moved;
2287}
2288
2289int move_freepages_block(struct zone *zone, struct page *page,
2290 int migratetype, int *num_movable)
2291{
2292 unsigned long start_pfn, end_pfn;
2293 struct page *start_page, *end_page;
2294
2295 if (num_movable)
2296 *num_movable = 0;
2297
2298 start_pfn = page_to_pfn(page);
2299 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2300 start_page = pfn_to_page(start_pfn);
2301 end_page = start_page + pageblock_nr_pages - 1;
2302 end_pfn = start_pfn + pageblock_nr_pages - 1;
2303
2304 /* Do not cross zone boundaries */
2305 if (!zone_spans_pfn(zone, start_pfn))
2306 start_page = page;
2307 if (!zone_spans_pfn(zone, end_pfn))
2308 return 0;
2309
2310 return move_freepages(zone, start_page, end_page, migratetype,
2311 num_movable);
2312}
2313
2314static void change_pageblock_range(struct page *pageblock_page,
2315 int start_order, int migratetype)
2316{
2317 int nr_pageblocks = 1 << (start_order - pageblock_order);
2318
2319 while (nr_pageblocks--) {
2320 set_pageblock_migratetype(pageblock_page, migratetype);
2321 pageblock_page += pageblock_nr_pages;
2322 }
2323}
2324
2325/*
2326 * When we are falling back to another migratetype during allocation, try to
2327 * steal extra free pages from the same pageblocks to satisfy further
2328 * allocations, instead of polluting multiple pageblocks.
2329 *
2330 * If we are stealing a relatively large buddy page, it is likely there will
2331 * be more free pages in the pageblock, so try to steal them all. For
2332 * reclaimable and unmovable allocations, we steal regardless of page size,
2333 * as fragmentation caused by those allocations polluting movable pageblocks
2334 * is worse than movable allocations stealing from unmovable and reclaimable
2335 * pageblocks.
2336 */
2337static bool can_steal_fallback(unsigned int order, int start_mt)
2338{
2339 /*
2340 * Leaving this order check is intended, although there is
2341 * relaxed order check in next check. The reason is that
2342 * we can actually steal whole pageblock if this condition met,
2343 * but, below check doesn't guarantee it and that is just heuristic
2344 * so could be changed anytime.
2345 */
2346 if (order >= pageblock_order)
2347 return true;
2348
2349 if (order >= pageblock_order / 2 ||
2350 start_mt == MIGRATE_RECLAIMABLE ||
2351 start_mt == MIGRATE_UNMOVABLE ||
2352 page_group_by_mobility_disabled)
2353 return true;
2354
2355 return false;
2356}
2357
2358static inline void boost_watermark(struct zone *zone)
2359{
2360 unsigned long max_boost;
2361
2362 if (!watermark_boost_factor)
2363 return;
2364
2365 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2366 watermark_boost_factor, 10000);
2367
2368 /*
2369 * high watermark may be uninitialised if fragmentation occurs
2370 * very early in boot so do not boost. We do not fall
2371 * through and boost by pageblock_nr_pages as failing
2372 * allocations that early means that reclaim is not going
2373 * to help and it may even be impossible to reclaim the
2374 * boosted watermark resulting in a hang.
2375 */
2376 if (!max_boost)
2377 return;
2378
2379 max_boost = max(pageblock_nr_pages, max_boost);
2380
2381 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2382 max_boost);
2383}
2384
2385/*
2386 * This function implements actual steal behaviour. If order is large enough,
2387 * we can steal whole pageblock. If not, we first move freepages in this
2388 * pageblock to our migratetype and determine how many already-allocated pages
2389 * are there in the pageblock with a compatible migratetype. If at least half
2390 * of pages are free or compatible, we can change migratetype of the pageblock
2391 * itself, so pages freed in the future will be put on the correct free list.
2392 */
2393static void steal_suitable_fallback(struct zone *zone, struct page *page,
2394 unsigned int alloc_flags, int start_type, bool whole_block)
2395{
2396 unsigned int current_order = page_order(page);
2397 struct free_area *area;
2398 int free_pages, movable_pages, alike_pages;
2399 int old_block_type;
2400
2401 old_block_type = get_pageblock_migratetype(page);
2402
2403 /*
2404 * This can happen due to races and we want to prevent broken
2405 * highatomic accounting.
2406 */
2407 if (is_migrate_highatomic(old_block_type))
2408 goto single_page;
2409
2410 /* Take ownership for orders >= pageblock_order */
2411 if (current_order >= pageblock_order) {
2412 change_pageblock_range(page, current_order, start_type);
2413 goto single_page;
2414 }
2415
2416 /*
2417 * Boost watermarks to increase reclaim pressure to reduce the
2418 * likelihood of future fallbacks. Wake kswapd now as the node
2419 * may be balanced overall and kswapd will not wake naturally.
2420 */
2421 boost_watermark(zone);
2422 if (alloc_flags & ALLOC_KSWAPD)
2423 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2424
2425 /* We are not allowed to try stealing from the whole block */
2426 if (!whole_block)
2427 goto single_page;
2428
2429 free_pages = move_freepages_block(zone, page, start_type,
2430 &movable_pages);
2431 /*
2432 * Determine how many pages are compatible with our allocation.
2433 * For movable allocation, it's the number of movable pages which
2434 * we just obtained. For other types it's a bit more tricky.
2435 */
2436 if (start_type == MIGRATE_MOVABLE) {
2437 alike_pages = movable_pages;
2438 } else {
2439 /*
2440 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2441 * to MOVABLE pageblock, consider all non-movable pages as
2442 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2443 * vice versa, be conservative since we can't distinguish the
2444 * exact migratetype of non-movable pages.
2445 */
2446 if (old_block_type == MIGRATE_MOVABLE)
2447 alike_pages = pageblock_nr_pages
2448 - (free_pages + movable_pages);
2449 else
2450 alike_pages = 0;
2451 }
2452
2453 /* moving whole block can fail due to zone boundary conditions */
2454 if (!free_pages)
2455 goto single_page;
2456
2457 /*
2458 * If a sufficient number of pages in the block are either free or of
2459 * comparable migratability as our allocation, claim the whole block.
2460 */
2461 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2462 page_group_by_mobility_disabled)
2463 set_pageblock_migratetype(page, start_type);
2464
2465 return;
2466
2467single_page:
2468 area = &zone->free_area[current_order];
2469 move_to_free_area(page, area, start_type);
2470}
2471
2472/*
2473 * Check whether there is a suitable fallback freepage with requested order.
2474 * If only_stealable is true, this function returns fallback_mt only if
2475 * we can steal other freepages all together. This would help to reduce
2476 * fragmentation due to mixed migratetype pages in one pageblock.
2477 */
2478int find_suitable_fallback(struct free_area *area, unsigned int order,
2479 int migratetype, bool only_stealable, bool *can_steal)
2480{
2481 int i;
2482 int fallback_mt;
2483
2484 if (area->nr_free == 0)
2485 return -1;
2486
2487 *can_steal = false;
2488 for (i = 0;; i++) {
2489 fallback_mt = fallbacks[migratetype][i];
2490 if (fallback_mt == MIGRATE_TYPES)
2491 break;
2492
2493 if (free_area_empty(area, fallback_mt))
2494 continue;
2495
2496 if (can_steal_fallback(order, migratetype))
2497 *can_steal = true;
2498
2499 if (!only_stealable)
2500 return fallback_mt;
2501
2502 if (*can_steal)
2503 return fallback_mt;
2504 }
2505
2506 return -1;
2507}
2508
2509/*
2510 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2511 * there are no empty page blocks that contain a page with a suitable order
2512 */
2513static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2514 unsigned int alloc_order)
2515{
2516 int mt;
2517 unsigned long max_managed, flags;
2518
2519 /*
2520 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2521 * Check is race-prone but harmless.
2522 */
2523 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2524 if (zone->nr_reserved_highatomic >= max_managed)
2525 return;
2526
2527 spin_lock_irqsave(&zone->lock, flags);
2528
2529 /* Recheck the nr_reserved_highatomic limit under the lock */
2530 if (zone->nr_reserved_highatomic >= max_managed)
2531 goto out_unlock;
2532
2533 /* Yoink! */
2534 mt = get_pageblock_migratetype(page);
2535 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2536 && !is_migrate_cma(mt)) {
2537 zone->nr_reserved_highatomic += pageblock_nr_pages;
2538 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2539 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2540 }
2541
2542out_unlock:
2543 spin_unlock_irqrestore(&zone->lock, flags);
2544}
2545
2546/*
2547 * Used when an allocation is about to fail under memory pressure. This
2548 * potentially hurts the reliability of high-order allocations when under
2549 * intense memory pressure but failed atomic allocations should be easier
2550 * to recover from than an OOM.
2551 *
2552 * If @force is true, try to unreserve a pageblock even though highatomic
2553 * pageblock is exhausted.
2554 */
2555static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2556 bool force)
2557{
2558 struct zonelist *zonelist = ac->zonelist;
2559 unsigned long flags;
2560 struct zoneref *z;
2561 struct zone *zone;
2562 struct page *page;
2563 int order;
2564 bool ret;
2565
2566 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2567 ac->nodemask) {
2568 /*
2569 * Preserve at least one pageblock unless memory pressure
2570 * is really high.
2571 */
2572 if (!force && zone->nr_reserved_highatomic <=
2573 pageblock_nr_pages)
2574 continue;
2575
2576 spin_lock_irqsave(&zone->lock, flags);
2577 for (order = 0; order < MAX_ORDER; order++) {
2578 struct free_area *area = &(zone->free_area[order]);
2579
2580 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2581 if (!page)
2582 continue;
2583
2584 /*
2585 * In page freeing path, migratetype change is racy so
2586 * we can counter several free pages in a pageblock
2587 * in this loop althoug we changed the pageblock type
2588 * from highatomic to ac->migratetype. So we should
2589 * adjust the count once.
2590 */
2591 if (is_migrate_highatomic_page(page)) {
2592 /*
2593 * It should never happen but changes to
2594 * locking could inadvertently allow a per-cpu
2595 * drain to add pages to MIGRATE_HIGHATOMIC
2596 * while unreserving so be safe and watch for
2597 * underflows.
2598 */
2599 zone->nr_reserved_highatomic -= min(
2600 pageblock_nr_pages,
2601 zone->nr_reserved_highatomic);
2602 }
2603
2604 /*
2605 * Convert to ac->migratetype and avoid the normal
2606 * pageblock stealing heuristics. Minimally, the caller
2607 * is doing the work and needs the pages. More
2608 * importantly, if the block was always converted to
2609 * MIGRATE_UNMOVABLE or another type then the number
2610 * of pageblocks that cannot be completely freed
2611 * may increase.
2612 */
2613 set_pageblock_migratetype(page, ac->migratetype);
2614 ret = move_freepages_block(zone, page, ac->migratetype,
2615 NULL);
2616 if (ret) {
2617 spin_unlock_irqrestore(&zone->lock, flags);
2618 return ret;
2619 }
2620 }
2621 spin_unlock_irqrestore(&zone->lock, flags);
2622 }
2623
2624 return false;
2625}
2626
2627/*
2628 * Try finding a free buddy page on the fallback list and put it on the free
2629 * list of requested migratetype, possibly along with other pages from the same
2630 * block, depending on fragmentation avoidance heuristics. Returns true if
2631 * fallback was found so that __rmqueue_smallest() can grab it.
2632 *
2633 * The use of signed ints for order and current_order is a deliberate
2634 * deviation from the rest of this file, to make the for loop
2635 * condition simpler.
2636 */
2637static __always_inline bool
2638__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2639 unsigned int alloc_flags)
2640{
2641 struct free_area *area;
2642 int current_order;
2643 int min_order = order;
2644 struct page *page;
2645 int fallback_mt;
2646 bool can_steal;
2647
2648 /*
2649 * Do not steal pages from freelists belonging to other pageblocks
2650 * i.e. orders < pageblock_order. If there are no local zones free,
2651 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2652 */
2653 if (alloc_flags & ALLOC_NOFRAGMENT)
2654 min_order = pageblock_order;
2655
2656 /*
2657 * Find the largest available free page in the other list. This roughly
2658 * approximates finding the pageblock with the most free pages, which
2659 * would be too costly to do exactly.
2660 */
2661 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2662 --current_order) {
2663 area = &(zone->free_area[current_order]);
2664 fallback_mt = find_suitable_fallback(area, current_order,
2665 start_migratetype, false, &can_steal);
2666 if (fallback_mt == -1)
2667 continue;
2668
2669 /*
2670 * We cannot steal all free pages from the pageblock and the
2671 * requested migratetype is movable. In that case it's better to
2672 * steal and split the smallest available page instead of the
2673 * largest available page, because even if the next movable
2674 * allocation falls back into a different pageblock than this
2675 * one, it won't cause permanent fragmentation.
2676 */
2677 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2678 && current_order > order)
2679 goto find_smallest;
2680
2681 goto do_steal;
2682 }
2683
2684 return false;
2685
2686find_smallest:
2687 for (current_order = order; current_order < MAX_ORDER;
2688 current_order++) {
2689 area = &(zone->free_area[current_order]);
2690 fallback_mt = find_suitable_fallback(area, current_order,
2691 start_migratetype, false, &can_steal);
2692 if (fallback_mt != -1)
2693 break;
2694 }
2695
2696 /*
2697 * This should not happen - we already found a suitable fallback
2698 * when looking for the largest page.
2699 */
2700 VM_BUG_ON(current_order == MAX_ORDER);
2701
2702do_steal:
2703 page = get_page_from_free_area(area, fallback_mt);
2704
2705 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2706 can_steal);
2707
2708 trace_mm_page_alloc_extfrag(page, order, current_order,
2709 start_migratetype, fallback_mt);
2710
2711 return true;
2712
2713}
2714
2715/*
2716 * Do the hard work of removing an element from the buddy allocator.
2717 * Call me with the zone->lock already held.
2718 */
2719static __always_inline struct page *
2720__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2721 unsigned int alloc_flags)
2722{
2723 struct page *page;
2724
2725retry:
2726 page = __rmqueue_smallest(zone, order, migratetype);
2727 if (unlikely(!page)) {
2728 if (migratetype == MIGRATE_MOVABLE)
2729 page = __rmqueue_cma_fallback(zone, order);
2730
2731 if (!page && __rmqueue_fallback(zone, order, migratetype,
2732 alloc_flags))
2733 goto retry;
2734 }
2735
2736 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2737 return page;
2738}
2739
2740/*
2741 * Obtain a specified number of elements from the buddy allocator, all under
2742 * a single hold of the lock, for efficiency. Add them to the supplied list.
2743 * Returns the number of new pages which were placed at *list.
2744 */
2745static int rmqueue_bulk(struct zone *zone, unsigned int order,
2746 unsigned long count, struct list_head *list,
2747 int migratetype, unsigned int alloc_flags)
2748{
2749 int i, alloced = 0;
2750
2751 spin_lock(&zone->lock);
2752 for (i = 0; i < count; ++i) {
2753 struct page *page = __rmqueue(zone, order, migratetype,
2754 alloc_flags);
2755 if (unlikely(page == NULL))
2756 break;
2757
2758 if (unlikely(check_pcp_refill(page)))
2759 continue;
2760
2761 /*
2762 * Split buddy pages returned by expand() are received here in
2763 * physical page order. The page is added to the tail of
2764 * caller's list. From the callers perspective, the linked list
2765 * is ordered by page number under some conditions. This is
2766 * useful for IO devices that can forward direction from the
2767 * head, thus also in the physical page order. This is useful
2768 * for IO devices that can merge IO requests if the physical
2769 * pages are ordered properly.
2770 */
2771 list_add_tail(&page->lru, list);
2772 alloced++;
2773 if (is_migrate_cma(get_pcppage_migratetype(page)))
2774 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2775 -(1 << order));
2776 }
2777
2778 /*
2779 * i pages were removed from the buddy list even if some leak due
2780 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2781 * on i. Do not confuse with 'alloced' which is the number of
2782 * pages added to the pcp list.
2783 */
2784 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2785 spin_unlock(&zone->lock);
2786 return alloced;
2787}
2788
2789#ifdef CONFIG_NUMA
2790/*
2791 * Called from the vmstat counter updater to drain pagesets of this
2792 * currently executing processor on remote nodes after they have
2793 * expired.
2794 *
2795 * Note that this function must be called with the thread pinned to
2796 * a single processor.
2797 */
2798void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2799{
2800 unsigned long flags;
2801 int to_drain, batch;
2802
2803 local_irq_save(flags);
2804 batch = READ_ONCE(pcp->batch);
2805 to_drain = min(pcp->count, batch);
2806 if (to_drain > 0)
2807 free_pcppages_bulk(zone, to_drain, pcp);
2808 local_irq_restore(flags);
2809}
2810#endif
2811
2812/*
2813 * Drain pcplists of the indicated processor and zone.
2814 *
2815 * The processor must either be the current processor and the
2816 * thread pinned to the current processor or a processor that
2817 * is not online.
2818 */
2819static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2820{
2821 unsigned long flags;
2822 struct per_cpu_pageset *pset;
2823 struct per_cpu_pages *pcp;
2824
2825 local_irq_save(flags);
2826 pset = per_cpu_ptr(zone->pageset, cpu);
2827
2828 pcp = &pset->pcp;
2829 if (pcp->count)
2830 free_pcppages_bulk(zone, pcp->count, pcp);
2831 local_irq_restore(flags);
2832}
2833
2834/*
2835 * Drain pcplists of all zones on the indicated processor.
2836 *
2837 * The processor must either be the current processor and the
2838 * thread pinned to the current processor or a processor that
2839 * is not online.
2840 */
2841static void drain_pages(unsigned int cpu)
2842{
2843 struct zone *zone;
2844
2845 for_each_populated_zone(zone) {
2846 drain_pages_zone(cpu, zone);
2847 }
2848}
2849
2850/*
2851 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2852 *
2853 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2854 * the single zone's pages.
2855 */
2856void drain_local_pages(struct zone *zone)
2857{
2858 int cpu = smp_processor_id();
2859
2860 if (zone)
2861 drain_pages_zone(cpu, zone);
2862 else
2863 drain_pages(cpu);
2864}
2865
2866static void drain_local_pages_wq(struct work_struct *work)
2867{
2868 struct pcpu_drain *drain;
2869
2870 drain = container_of(work, struct pcpu_drain, work);
2871
2872 /*
2873 * drain_all_pages doesn't use proper cpu hotplug protection so
2874 * we can race with cpu offline when the WQ can move this from
2875 * a cpu pinned worker to an unbound one. We can operate on a different
2876 * cpu which is allright but we also have to make sure to not move to
2877 * a different one.
2878 */
2879 preempt_disable();
2880 drain_local_pages(drain->zone);
2881 preempt_enable();
2882}
2883
2884/*
2885 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2886 *
2887 * When zone parameter is non-NULL, spill just the single zone's pages.
2888 *
2889 * Note that this can be extremely slow as the draining happens in a workqueue.
2890 */
2891void drain_all_pages(struct zone *zone)
2892{
2893 int cpu;
2894
2895 /*
2896 * Allocate in the BSS so we wont require allocation in
2897 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2898 */
2899 static cpumask_t cpus_with_pcps;
2900
2901 /*
2902 * Make sure nobody triggers this path before mm_percpu_wq is fully
2903 * initialized.
2904 */
2905 if (WARN_ON_ONCE(!mm_percpu_wq))
2906 return;
2907
2908 /*
2909 * Do not drain if one is already in progress unless it's specific to
2910 * a zone. Such callers are primarily CMA and memory hotplug and need
2911 * the drain to be complete when the call returns.
2912 */
2913 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2914 if (!zone)
2915 return;
2916 mutex_lock(&pcpu_drain_mutex);
2917 }
2918
2919 /*
2920 * We don't care about racing with CPU hotplug event
2921 * as offline notification will cause the notified
2922 * cpu to drain that CPU pcps and on_each_cpu_mask
2923 * disables preemption as part of its processing
2924 */
2925 for_each_online_cpu(cpu) {
2926 struct per_cpu_pageset *pcp;
2927 struct zone *z;
2928 bool has_pcps = false;
2929
2930 if (zone) {
2931 pcp = per_cpu_ptr(zone->pageset, cpu);
2932 if (pcp->pcp.count)
2933 has_pcps = true;
2934 } else {
2935 for_each_populated_zone(z) {
2936 pcp = per_cpu_ptr(z->pageset, cpu);
2937 if (pcp->pcp.count) {
2938 has_pcps = true;
2939 break;
2940 }
2941 }
2942 }
2943
2944 if (has_pcps)
2945 cpumask_set_cpu(cpu, &cpus_with_pcps);
2946 else
2947 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2948 }
2949
2950 for_each_cpu(cpu, &cpus_with_pcps) {
2951 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2952
2953 drain->zone = zone;
2954 INIT_WORK(&drain->work, drain_local_pages_wq);
2955 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2956 }
2957 for_each_cpu(cpu, &cpus_with_pcps)
2958 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2959
2960 mutex_unlock(&pcpu_drain_mutex);
2961}
2962
2963#ifdef CONFIG_HIBERNATION
2964
2965/*
2966 * Touch the watchdog for every WD_PAGE_COUNT pages.
2967 */
2968#define WD_PAGE_COUNT (128*1024)
2969
2970void mark_free_pages(struct zone *zone)
2971{
2972 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2973 unsigned long flags;
2974 unsigned int order, t;
2975 struct page *page;
2976
2977 if (zone_is_empty(zone))
2978 return;
2979
2980 spin_lock_irqsave(&zone->lock, flags);
2981
2982 max_zone_pfn = zone_end_pfn(zone);
2983 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2984 if (pfn_valid(pfn)) {
2985 page = pfn_to_page(pfn);
2986
2987 if (!--page_count) {
2988 touch_nmi_watchdog();
2989 page_count = WD_PAGE_COUNT;
2990 }
2991
2992 if (page_zone(page) != zone)
2993 continue;
2994
2995 if (!swsusp_page_is_forbidden(page))
2996 swsusp_unset_page_free(page);
2997 }
2998
2999 for_each_migratetype_order(order, t) {
3000 list_for_each_entry(page,
3001 &zone->free_area[order].free_list[t], lru) {
3002 unsigned long i;
3003
3004 pfn = page_to_pfn(page);
3005 for (i = 0; i < (1UL << order); i++) {
3006 if (!--page_count) {
3007 touch_nmi_watchdog();
3008 page_count = WD_PAGE_COUNT;
3009 }
3010 swsusp_set_page_free(pfn_to_page(pfn + i));
3011 }
3012 }
3013 }
3014 spin_unlock_irqrestore(&zone->lock, flags);
3015}
3016#endif /* CONFIG_PM */
3017
3018static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3019{
3020 int migratetype;
3021
3022 if (!free_pcp_prepare(page))
3023 return false;
3024
3025 migratetype = get_pfnblock_migratetype(page, pfn);
3026 set_pcppage_migratetype(page, migratetype);
3027 return true;
3028}
3029
3030static void free_unref_page_commit(struct page *page, unsigned long pfn)
3031{
3032 struct zone *zone = page_zone(page);
3033 struct per_cpu_pages *pcp;
3034 int migratetype;
3035
3036 migratetype = get_pcppage_migratetype(page);
3037 __count_vm_event(PGFREE);
3038
3039 /*
3040 * We only track unmovable, reclaimable and movable on pcp lists.
3041 * Free ISOLATE pages back to the allocator because they are being
3042 * offlined but treat HIGHATOMIC as movable pages so we can get those
3043 * areas back if necessary. Otherwise, we may have to free
3044 * excessively into the page allocator
3045 */
3046 if (migratetype >= MIGRATE_PCPTYPES) {
3047 if (unlikely(is_migrate_isolate(migratetype))) {
3048 free_one_page(zone, page, pfn, 0, migratetype);
3049 return;
3050 }
3051 migratetype = MIGRATE_MOVABLE;
3052 }
3053
3054 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3055 list_add(&page->lru, &pcp->lists[migratetype]);
3056 pcp->count++;
3057 if (pcp->count >= pcp->high) {
3058 unsigned long batch = READ_ONCE(pcp->batch);
3059 free_pcppages_bulk(zone, batch, pcp);
3060 }
3061}
3062
3063/*
3064 * Free a 0-order page
3065 */
3066void free_unref_page(struct page *page)
3067{
3068 unsigned long flags;
3069 unsigned long pfn = page_to_pfn(page);
3070
3071 if (!free_unref_page_prepare(page, pfn))
3072 return;
3073
3074 local_irq_save(flags);
3075 free_unref_page_commit(page, pfn);
3076 local_irq_restore(flags);
3077}
3078
3079/*
3080 * Free a list of 0-order pages
3081 */
3082void free_unref_page_list(struct list_head *list)
3083{
3084 struct page *page, *next;
3085 unsigned long flags, pfn;
3086 int batch_count = 0;
3087
3088 /* Prepare pages for freeing */
3089 list_for_each_entry_safe(page, next, list, lru) {
3090 pfn = page_to_pfn(page);
3091 if (!free_unref_page_prepare(page, pfn))
3092 list_del(&page->lru);
3093 set_page_private(page, pfn);
3094 }
3095
3096 local_irq_save(flags);
3097 list_for_each_entry_safe(page, next, list, lru) {
3098 unsigned long pfn = page_private(page);
3099
3100 set_page_private(page, 0);
3101 trace_mm_page_free_batched(page);
3102 free_unref_page_commit(page, pfn);
3103
3104 /*
3105 * Guard against excessive IRQ disabled times when we get
3106 * a large list of pages to free.
3107 */
3108 if (++batch_count == SWAP_CLUSTER_MAX) {
3109 local_irq_restore(flags);
3110 batch_count = 0;
3111 local_irq_save(flags);
3112 }
3113 }
3114 local_irq_restore(flags);
3115}
3116
3117/*
3118 * split_page takes a non-compound higher-order page, and splits it into
3119 * n (1<<order) sub-pages: page[0..n]
3120 * Each sub-page must be freed individually.
3121 *
3122 * Note: this is probably too low level an operation for use in drivers.
3123 * Please consult with lkml before using this in your driver.
3124 */
3125void split_page(struct page *page, unsigned int order)
3126{
3127 int i;
3128
3129 VM_BUG_ON_PAGE(PageCompound(page), page);
3130 VM_BUG_ON_PAGE(!page_count(page), page);
3131
3132 for (i = 1; i < (1 << order); i++)
3133 set_page_refcounted(page + i);
3134 split_page_owner(page, order);
3135}
3136EXPORT_SYMBOL_GPL(split_page);
3137
3138int __isolate_free_page(struct page *page, unsigned int order)
3139{
3140 struct free_area *area = &page_zone(page)->free_area[order];
3141 unsigned long watermark;
3142 struct zone *zone;
3143 int mt;
3144
3145 BUG_ON(!PageBuddy(page));
3146
3147 zone = page_zone(page);
3148 mt = get_pageblock_migratetype(page);
3149
3150 if (!is_migrate_isolate(mt)) {
3151 /*
3152 * Obey watermarks as if the page was being allocated. We can
3153 * emulate a high-order watermark check with a raised order-0
3154 * watermark, because we already know our high-order page
3155 * exists.
3156 */
3157 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3158 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3159 return 0;
3160
3161 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3162 }
3163
3164 /* Remove page from free list */
3165
3166 del_page_from_free_area(page, area);
3167
3168 /*
3169 * Set the pageblock if the isolated page is at least half of a
3170 * pageblock
3171 */
3172 if (order >= pageblock_order - 1) {
3173 struct page *endpage = page + (1 << order) - 1;
3174 for (; page < endpage; page += pageblock_nr_pages) {
3175 int mt = get_pageblock_migratetype(page);
3176 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3177 && !is_migrate_highatomic(mt))
3178 set_pageblock_migratetype(page,
3179 MIGRATE_MOVABLE);
3180 }
3181 }
3182
3183
3184 return 1UL << order;
3185}
3186
3187/*
3188 * Update NUMA hit/miss statistics
3189 *
3190 * Must be called with interrupts disabled.
3191 */
3192static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3193{
3194#ifdef CONFIG_NUMA
3195 enum numa_stat_item local_stat = NUMA_LOCAL;
3196
3197 /* skip numa counters update if numa stats is disabled */
3198 if (!static_branch_likely(&vm_numa_stat_key))
3199 return;
3200
3201 if (zone_to_nid(z) != numa_node_id())
3202 local_stat = NUMA_OTHER;
3203
3204 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3205 __inc_numa_state(z, NUMA_HIT);
3206 else {
3207 __inc_numa_state(z, NUMA_MISS);
3208 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3209 }
3210 __inc_numa_state(z, local_stat);
3211#endif
3212}
3213
3214/* Remove page from the per-cpu list, caller must protect the list */
3215static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3216 unsigned int alloc_flags,
3217 struct per_cpu_pages *pcp,
3218 struct list_head *list)
3219{
3220 struct page *page;
3221
3222 do {
3223 if (list_empty(list)) {
3224 pcp->count += rmqueue_bulk(zone, 0,
3225 pcp->batch, list,
3226 migratetype, alloc_flags);
3227 if (unlikely(list_empty(list)))
3228 return NULL;
3229 }
3230
3231 page = list_first_entry(list, struct page, lru);
3232 list_del(&page->lru);
3233 pcp->count--;
3234 } while (check_new_pcp(page));
3235
3236 return page;
3237}
3238
3239/* Lock and remove page from the per-cpu list */
3240static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3241 struct zone *zone, gfp_t gfp_flags,
3242 int migratetype, unsigned int alloc_flags)
3243{
3244 struct per_cpu_pages *pcp;
3245 struct list_head *list;
3246 struct page *page;
3247 unsigned long flags;
3248
3249 local_irq_save(flags);
3250 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3251 list = &pcp->lists[migratetype];
3252 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3253 if (page) {
3254 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3255 zone_statistics(preferred_zone, zone);
3256 }
3257 local_irq_restore(flags);
3258 return page;
3259}
3260
3261/*
3262 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3263 */
3264static inline
3265struct page *rmqueue(struct zone *preferred_zone,
3266 struct zone *zone, unsigned int order,
3267 gfp_t gfp_flags, unsigned int alloc_flags,
3268 int migratetype)
3269{
3270 unsigned long flags;
3271 struct page *page;
3272
3273 if (likely(order == 0)) {
3274 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3275 migratetype, alloc_flags);
3276 goto out;
3277 }
3278
3279 /*
3280 * We most definitely don't want callers attempting to
3281 * allocate greater than order-1 page units with __GFP_NOFAIL.
3282 */
3283 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3284 spin_lock_irqsave(&zone->lock, flags);
3285
3286 do {
3287 page = NULL;
3288 if (alloc_flags & ALLOC_HARDER) {
3289 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3290 if (page)
3291 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3292 }
3293 if (!page)
3294 page = __rmqueue(zone, order, migratetype, alloc_flags);
3295 } while (page && check_new_pages(page, order));
3296 spin_unlock(&zone->lock);
3297 if (!page)
3298 goto failed;
3299 __mod_zone_freepage_state(zone, -(1 << order),
3300 get_pcppage_migratetype(page));
3301
3302 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3303 zone_statistics(preferred_zone, zone);
3304 local_irq_restore(flags);
3305
3306out:
3307 /* Separate test+clear to avoid unnecessary atomics */
3308 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3309 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3310 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3311 }
3312
3313 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3314 return page;
3315
3316failed:
3317 local_irq_restore(flags);
3318 return NULL;
3319}
3320
3321#ifdef CONFIG_FAIL_PAGE_ALLOC
3322
3323static struct {
3324 struct fault_attr attr;
3325
3326 bool ignore_gfp_highmem;
3327 bool ignore_gfp_reclaim;
3328 u32 min_order;
3329} fail_page_alloc = {
3330 .attr = FAULT_ATTR_INITIALIZER,
3331 .ignore_gfp_reclaim = true,
3332 .ignore_gfp_highmem = true,
3333 .min_order = 1,
3334};
3335
3336static int __init setup_fail_page_alloc(char *str)
3337{
3338 return setup_fault_attr(&fail_page_alloc.attr, str);
3339}
3340__setup("fail_page_alloc=", setup_fail_page_alloc);
3341
3342static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3343{
3344 if (order < fail_page_alloc.min_order)
3345 return false;
3346 if (gfp_mask & __GFP_NOFAIL)
3347 return false;
3348 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3349 return false;
3350 if (fail_page_alloc.ignore_gfp_reclaim &&
3351 (gfp_mask & __GFP_DIRECT_RECLAIM))
3352 return false;
3353
3354 return should_fail(&fail_page_alloc.attr, 1 << order);
3355}
3356
3357#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3358
3359static int __init fail_page_alloc_debugfs(void)
3360{
3361 umode_t mode = S_IFREG | 0600;
3362 struct dentry *dir;
3363
3364 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3365 &fail_page_alloc.attr);
3366
3367 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3368 &fail_page_alloc.ignore_gfp_reclaim);
3369 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3370 &fail_page_alloc.ignore_gfp_highmem);
3371 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3372
3373 return 0;
3374}
3375
3376late_initcall(fail_page_alloc_debugfs);
3377
3378#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3379
3380#else /* CONFIG_FAIL_PAGE_ALLOC */
3381
3382static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3383{
3384 return false;
3385}
3386
3387#endif /* CONFIG_FAIL_PAGE_ALLOC */
3388
3389static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3390{
3391 return __should_fail_alloc_page(gfp_mask, order);
3392}
3393ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3394
3395/*
3396 * Return true if free base pages are above 'mark'. For high-order checks it
3397 * will return true of the order-0 watermark is reached and there is at least
3398 * one free page of a suitable size. Checking now avoids taking the zone lock
3399 * to check in the allocation paths if no pages are free.
3400 */
3401bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3402 int classzone_idx, unsigned int alloc_flags,
3403 long free_pages)
3404{
3405 long min = mark;
3406 int o;
3407 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3408
3409 /* free_pages may go negative - that's OK */
3410 free_pages -= (1 << order) - 1;
3411
3412 if (alloc_flags & ALLOC_HIGH)
3413 min -= min / 2;
3414
3415 /*
3416 * If the caller does not have rights to ALLOC_HARDER then subtract
3417 * the high-atomic reserves. This will over-estimate the size of the
3418 * atomic reserve but it avoids a search.
3419 */
3420 if (likely(!alloc_harder)) {
3421 free_pages -= z->nr_reserved_highatomic;
3422 } else {
3423 /*
3424 * OOM victims can try even harder than normal ALLOC_HARDER
3425 * users on the grounds that it's definitely going to be in
3426 * the exit path shortly and free memory. Any allocation it
3427 * makes during the free path will be small and short-lived.
3428 */
3429 if (alloc_flags & ALLOC_OOM)
3430 min -= min / 2;
3431 else
3432 min -= min / 4;
3433 }
3434
3435
3436#ifdef CONFIG_CMA
3437 /* If allocation can't use CMA areas don't use free CMA pages */
3438 if (!(alloc_flags & ALLOC_CMA))
3439 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3440#endif
3441
3442 /*
3443 * Check watermarks for an order-0 allocation request. If these
3444 * are not met, then a high-order request also cannot go ahead
3445 * even if a suitable page happened to be free.
3446 */
3447 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3448 return false;
3449
3450 /* If this is an order-0 request then the watermark is fine */
3451 if (!order)
3452 return true;
3453
3454 /* For a high-order request, check at least one suitable page is free */
3455 for (o = order; o < MAX_ORDER; o++) {
3456 struct free_area *area = &z->free_area[o];
3457 int mt;
3458
3459 if (!area->nr_free)
3460 continue;
3461
3462 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3463 if (!free_area_empty(area, mt))
3464 return true;
3465 }
3466
3467#ifdef CONFIG_CMA
3468 if ((alloc_flags & ALLOC_CMA) &&
3469 !free_area_empty(area, MIGRATE_CMA)) {
3470 return true;
3471 }
3472#endif
3473 if (alloc_harder &&
3474 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3475 return true;
3476 }
3477 return false;
3478}
3479
3480bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3481 int classzone_idx, unsigned int alloc_flags)
3482{
3483 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3484 zone_page_state(z, NR_FREE_PAGES));
3485}
3486
3487static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3488 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3489{
3490 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3491 long cma_pages = 0;
3492
3493#ifdef CONFIG_CMA
3494 /* If allocation can't use CMA areas don't use free CMA pages */
3495 if (!(alloc_flags & ALLOC_CMA))
3496 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3497#endif
3498
3499 /*
3500 * Fast check for order-0 only. If this fails then the reserves
3501 * need to be calculated. There is a corner case where the check
3502 * passes but only the high-order atomic reserve are free. If
3503 * the caller is !atomic then it'll uselessly search the free
3504 * list. That corner case is then slower but it is harmless.
3505 */
3506 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3507 return true;
3508
3509 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3510 free_pages);
3511}
3512
3513bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3514 unsigned long mark, int classzone_idx)
3515{
3516 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3517
3518 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3519 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3520
3521 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3522 free_pages);
3523}
3524
3525#ifdef CONFIG_NUMA
3526static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3527{
3528 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3529 node_reclaim_distance;
3530}
3531#else /* CONFIG_NUMA */
3532static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3533{
3534 return true;
3535}
3536#endif /* CONFIG_NUMA */
3537
3538/*
3539 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3540 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3541 * premature use of a lower zone may cause lowmem pressure problems that
3542 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3543 * probably too small. It only makes sense to spread allocations to avoid
3544 * fragmentation between the Normal and DMA32 zones.
3545 */
3546static inline unsigned int
3547alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3548{
3549 unsigned int alloc_flags = 0;
3550
3551 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3552 alloc_flags |= ALLOC_KSWAPD;
3553
3554#ifdef CONFIG_ZONE_DMA32
3555 if (!zone)
3556 return alloc_flags;
3557
3558 if (zone_idx(zone) != ZONE_NORMAL)
3559 return alloc_flags;
3560
3561 /*
3562 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3563 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3564 * on UMA that if Normal is populated then so is DMA32.
3565 */
3566 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3567 if (nr_online_nodes > 1 && !populated_zone(--zone))
3568 return alloc_flags;
3569
3570 alloc_flags |= ALLOC_NOFRAGMENT;
3571#endif /* CONFIG_ZONE_DMA32 */
3572 return alloc_flags;
3573}
3574
3575/*
3576 * get_page_from_freelist goes through the zonelist trying to allocate
3577 * a page.
3578 */
3579static struct page *
3580get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3581 const struct alloc_context *ac)
3582{
3583 struct zoneref *z;
3584 struct zone *zone;
3585 struct pglist_data *last_pgdat_dirty_limit = NULL;
3586 bool no_fallback;
3587
3588retry:
3589 /*
3590 * Scan zonelist, looking for a zone with enough free.
3591 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3592 */
3593 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3594 z = ac->preferred_zoneref;
3595 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3596 ac->nodemask) {
3597 struct page *page;
3598 unsigned long mark;
3599
3600 if (cpusets_enabled() &&
3601 (alloc_flags & ALLOC_CPUSET) &&
3602 !__cpuset_zone_allowed(zone, gfp_mask))
3603 continue;
3604 /*
3605 * When allocating a page cache page for writing, we
3606 * want to get it from a node that is within its dirty
3607 * limit, such that no single node holds more than its
3608 * proportional share of globally allowed dirty pages.
3609 * The dirty limits take into account the node's
3610 * lowmem reserves and high watermark so that kswapd
3611 * should be able to balance it without having to
3612 * write pages from its LRU list.
3613 *
3614 * XXX: For now, allow allocations to potentially
3615 * exceed the per-node dirty limit in the slowpath
3616 * (spread_dirty_pages unset) before going into reclaim,
3617 * which is important when on a NUMA setup the allowed
3618 * nodes are together not big enough to reach the
3619 * global limit. The proper fix for these situations
3620 * will require awareness of nodes in the
3621 * dirty-throttling and the flusher threads.
3622 */
3623 if (ac->spread_dirty_pages) {
3624 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3625 continue;
3626
3627 if (!node_dirty_ok(zone->zone_pgdat)) {
3628 last_pgdat_dirty_limit = zone->zone_pgdat;
3629 continue;
3630 }
3631 }
3632
3633 if (no_fallback && nr_online_nodes > 1 &&
3634 zone != ac->preferred_zoneref->zone) {
3635 int local_nid;
3636
3637 /*
3638 * If moving to a remote node, retry but allow
3639 * fragmenting fallbacks. Locality is more important
3640 * than fragmentation avoidance.
3641 */
3642 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3643 if (zone_to_nid(zone) != local_nid) {
3644 alloc_flags &= ~ALLOC_NOFRAGMENT;
3645 goto retry;
3646 }
3647 }
3648
3649 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3650 if (!zone_watermark_fast(zone, order, mark,
3651 ac_classzone_idx(ac), alloc_flags)) {
3652 int ret;
3653
3654#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3655 /*
3656 * Watermark failed for this zone, but see if we can
3657 * grow this zone if it contains deferred pages.
3658 */
3659 if (static_branch_unlikely(&deferred_pages)) {
3660 if (_deferred_grow_zone(zone, order))
3661 goto try_this_zone;
3662 }
3663#endif
3664 /* Checked here to keep the fast path fast */
3665 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3666 if (alloc_flags & ALLOC_NO_WATERMARKS)
3667 goto try_this_zone;
3668
3669 if (node_reclaim_mode == 0 ||
3670 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3671 continue;
3672
3673 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3674 switch (ret) {
3675 case NODE_RECLAIM_NOSCAN:
3676 /* did not scan */
3677 continue;
3678 case NODE_RECLAIM_FULL:
3679 /* scanned but unreclaimable */
3680 continue;
3681 default:
3682 /* did we reclaim enough */
3683 if (zone_watermark_ok(zone, order, mark,
3684 ac_classzone_idx(ac), alloc_flags))
3685 goto try_this_zone;
3686
3687 continue;
3688 }
3689 }
3690
3691try_this_zone:
3692 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3693 gfp_mask, alloc_flags, ac->migratetype);
3694 if (page) {
3695 prep_new_page(page, order, gfp_mask, alloc_flags);
3696
3697 /*
3698 * If this is a high-order atomic allocation then check
3699 * if the pageblock should be reserved for the future
3700 */
3701 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3702 reserve_highatomic_pageblock(page, zone, order);
3703
3704 return page;
3705 } else {
3706#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3707 /* Try again if zone has deferred pages */
3708 if (static_branch_unlikely(&deferred_pages)) {
3709 if (_deferred_grow_zone(zone, order))
3710 goto try_this_zone;
3711 }
3712#endif
3713 }
3714 }
3715
3716 /*
3717 * It's possible on a UMA machine to get through all zones that are
3718 * fragmented. If avoiding fragmentation, reset and try again.
3719 */
3720 if (no_fallback) {
3721 alloc_flags &= ~ALLOC_NOFRAGMENT;
3722 goto retry;
3723 }
3724
3725 return NULL;
3726}
3727
3728static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3729{
3730 unsigned int filter = SHOW_MEM_FILTER_NODES;
3731
3732 /*
3733 * This documents exceptions given to allocations in certain
3734 * contexts that are allowed to allocate outside current's set
3735 * of allowed nodes.
3736 */
3737 if (!(gfp_mask & __GFP_NOMEMALLOC))
3738 if (tsk_is_oom_victim(current) ||
3739 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3740 filter &= ~SHOW_MEM_FILTER_NODES;
3741 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3742 filter &= ~SHOW_MEM_FILTER_NODES;
3743
3744 show_mem(filter, nodemask);
3745}
3746
3747void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3748{
3749 struct va_format vaf;
3750 va_list args;
3751 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3752
3753 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3754 return;
3755
3756 va_start(args, fmt);
3757 vaf.fmt = fmt;
3758 vaf.va = &args;
3759 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3760 current->comm, &vaf, gfp_mask, &gfp_mask,
3761 nodemask_pr_args(nodemask));
3762 va_end(args);
3763
3764 cpuset_print_current_mems_allowed();
3765 pr_cont("\n");
3766 dump_stack();
3767 warn_alloc_show_mem(gfp_mask, nodemask);
3768}
3769
3770static inline struct page *
3771__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3772 unsigned int alloc_flags,
3773 const struct alloc_context *ac)
3774{
3775 struct page *page;
3776
3777 page = get_page_from_freelist(gfp_mask, order,
3778 alloc_flags|ALLOC_CPUSET, ac);
3779 /*
3780 * fallback to ignore cpuset restriction if our nodes
3781 * are depleted
3782 */
3783 if (!page)
3784 page = get_page_from_freelist(gfp_mask, order,
3785 alloc_flags, ac);
3786
3787 return page;
3788}
3789
3790static inline struct page *
3791__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3792 const struct alloc_context *ac, unsigned long *did_some_progress)
3793{
3794 struct oom_control oc = {
3795 .zonelist = ac->zonelist,
3796 .nodemask = ac->nodemask,
3797 .memcg = NULL,
3798 .gfp_mask = gfp_mask,
3799 .order = order,
3800 };
3801 struct page *page;
3802
3803 *did_some_progress = 0;
3804
3805 /*
3806 * Acquire the oom lock. If that fails, somebody else is
3807 * making progress for us.
3808 */
3809 if (!mutex_trylock(&oom_lock)) {
3810 *did_some_progress = 1;
3811 schedule_timeout_uninterruptible(1);
3812 return NULL;
3813 }
3814
3815 /*
3816 * Go through the zonelist yet one more time, keep very high watermark
3817 * here, this is only to catch a parallel oom killing, we must fail if
3818 * we're still under heavy pressure. But make sure that this reclaim
3819 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3820 * allocation which will never fail due to oom_lock already held.
3821 */
3822 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3823 ~__GFP_DIRECT_RECLAIM, order,
3824 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3825 if (page)
3826 goto out;
3827
3828 /* Coredumps can quickly deplete all memory reserves */
3829 if (current->flags & PF_DUMPCORE)
3830 goto out;
3831 /* The OOM killer will not help higher order allocs */
3832 if (order > PAGE_ALLOC_COSTLY_ORDER)
3833 goto out;
3834 /*
3835 * We have already exhausted all our reclaim opportunities without any
3836 * success so it is time to admit defeat. We will skip the OOM killer
3837 * because it is very likely that the caller has a more reasonable
3838 * fallback than shooting a random task.
3839 */
3840 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3841 goto out;
3842 /* The OOM killer does not needlessly kill tasks for lowmem */
3843 if (ac->high_zoneidx < ZONE_NORMAL)
3844 goto out;
3845 if (pm_suspended_storage())
3846 goto out;
3847 /*
3848 * XXX: GFP_NOFS allocations should rather fail than rely on
3849 * other request to make a forward progress.
3850 * We are in an unfortunate situation where out_of_memory cannot
3851 * do much for this context but let's try it to at least get
3852 * access to memory reserved if the current task is killed (see
3853 * out_of_memory). Once filesystems are ready to handle allocation
3854 * failures more gracefully we should just bail out here.
3855 */
3856
3857 /* The OOM killer may not free memory on a specific node */
3858 if (gfp_mask & __GFP_THISNODE)
3859 goto out;
3860
3861 /* Exhausted what can be done so it's blame time */
3862 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3863 *did_some_progress = 1;
3864
3865 /*
3866 * Help non-failing allocations by giving them access to memory
3867 * reserves
3868 */
3869 if (gfp_mask & __GFP_NOFAIL)
3870 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3871 ALLOC_NO_WATERMARKS, ac);
3872 }
3873out:
3874 mutex_unlock(&oom_lock);
3875 return page;
3876}
3877
3878/*
3879 * Maximum number of compaction retries wit a progress before OOM
3880 * killer is consider as the only way to move forward.
3881 */
3882#define MAX_COMPACT_RETRIES 16
3883
3884#ifdef CONFIG_COMPACTION
3885/* Try memory compaction for high-order allocations before reclaim */
3886static struct page *
3887__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3888 unsigned int alloc_flags, const struct alloc_context *ac,
3889 enum compact_priority prio, enum compact_result *compact_result)
3890{
3891 struct page *page = NULL;
3892 unsigned long pflags;
3893 unsigned int noreclaim_flag;
3894
3895 if (!order)
3896 return NULL;
3897
3898 psi_memstall_enter(&pflags);
3899 noreclaim_flag = memalloc_noreclaim_save();
3900
3901 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3902 prio, &page);
3903
3904 memalloc_noreclaim_restore(noreclaim_flag);
3905 psi_memstall_leave(&pflags);
3906
3907 /*
3908 * At least in one zone compaction wasn't deferred or skipped, so let's
3909 * count a compaction stall
3910 */
3911 count_vm_event(COMPACTSTALL);
3912
3913 /* Prep a captured page if available */
3914 if (page)
3915 prep_new_page(page, order, gfp_mask, alloc_flags);
3916
3917 /* Try get a page from the freelist if available */
3918 if (!page)
3919 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3920
3921 if (page) {
3922 struct zone *zone = page_zone(page);
3923
3924 zone->compact_blockskip_flush = false;
3925 compaction_defer_reset(zone, order, true);
3926 count_vm_event(COMPACTSUCCESS);
3927 return page;
3928 }
3929
3930 /*
3931 * It's bad if compaction run occurs and fails. The most likely reason
3932 * is that pages exist, but not enough to satisfy watermarks.
3933 */
3934 count_vm_event(COMPACTFAIL);
3935
3936 cond_resched();
3937
3938 return NULL;
3939}
3940
3941static inline bool
3942should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3943 enum compact_result compact_result,
3944 enum compact_priority *compact_priority,
3945 int *compaction_retries)
3946{
3947 int max_retries = MAX_COMPACT_RETRIES;
3948 int min_priority;
3949 bool ret = false;
3950 int retries = *compaction_retries;
3951 enum compact_priority priority = *compact_priority;
3952
3953 if (!order)
3954 return false;
3955
3956 if (compaction_made_progress(compact_result))
3957 (*compaction_retries)++;
3958
3959 /*
3960 * compaction considers all the zone as desperately out of memory
3961 * so it doesn't really make much sense to retry except when the
3962 * failure could be caused by insufficient priority
3963 */
3964 if (compaction_failed(compact_result))
3965 goto check_priority;
3966
3967 /*
3968 * compaction was skipped because there are not enough order-0 pages
3969 * to work with, so we retry only if it looks like reclaim can help.
3970 */
3971 if (compaction_needs_reclaim(compact_result)) {
3972 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3973 goto out;
3974 }
3975
3976 /*
3977 * make sure the compaction wasn't deferred or didn't bail out early
3978 * due to locks contention before we declare that we should give up.
3979 * But the next retry should use a higher priority if allowed, so
3980 * we don't just keep bailing out endlessly.
3981 */
3982 if (compaction_withdrawn(compact_result)) {
3983 goto check_priority;
3984 }
3985
3986 /*
3987 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3988 * costly ones because they are de facto nofail and invoke OOM
3989 * killer to move on while costly can fail and users are ready
3990 * to cope with that. 1/4 retries is rather arbitrary but we
3991 * would need much more detailed feedback from compaction to
3992 * make a better decision.
3993 */
3994 if (order > PAGE_ALLOC_COSTLY_ORDER)
3995 max_retries /= 4;
3996 if (*compaction_retries <= max_retries) {
3997 ret = true;
3998 goto out;
3999 }
4000
4001 /*
4002 * Make sure there are attempts at the highest priority if we exhausted
4003 * all retries or failed at the lower priorities.
4004 */
4005check_priority:
4006 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4007 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4008
4009 if (*compact_priority > min_priority) {
4010 (*compact_priority)--;
4011 *compaction_retries = 0;
4012 ret = true;
4013 }
4014out:
4015 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4016 return ret;
4017}
4018#else
4019static inline struct page *
4020__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4021 unsigned int alloc_flags, const struct alloc_context *ac,
4022 enum compact_priority prio, enum compact_result *compact_result)
4023{
4024 *compact_result = COMPACT_SKIPPED;
4025 return NULL;
4026}
4027
4028static inline bool
4029should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4030 enum compact_result compact_result,
4031 enum compact_priority *compact_priority,
4032 int *compaction_retries)
4033{
4034 struct zone *zone;
4035 struct zoneref *z;
4036
4037 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4038 return false;
4039
4040 /*
4041 * There are setups with compaction disabled which would prefer to loop
4042 * inside the allocator rather than hit the oom killer prematurely.
4043 * Let's give them a good hope and keep retrying while the order-0
4044 * watermarks are OK.
4045 */
4046 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4047 ac->nodemask) {
4048 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4049 ac_classzone_idx(ac), alloc_flags))
4050 return true;
4051 }
4052 return false;
4053}
4054#endif /* CONFIG_COMPACTION */
4055
4056#ifdef CONFIG_LOCKDEP
4057static struct lockdep_map __fs_reclaim_map =
4058 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4059
4060static bool __need_fs_reclaim(gfp_t gfp_mask)
4061{
4062 gfp_mask = current_gfp_context(gfp_mask);
4063
4064 /* no reclaim without waiting on it */
4065 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4066 return false;
4067
4068 /* this guy won't enter reclaim */
4069 if (current->flags & PF_MEMALLOC)
4070 return false;
4071
4072 /* We're only interested __GFP_FS allocations for now */
4073 if (!(gfp_mask & __GFP_FS))
4074 return false;
4075
4076 if (gfp_mask & __GFP_NOLOCKDEP)
4077 return false;
4078
4079 return true;
4080}
4081
4082void __fs_reclaim_acquire(void)
4083{
4084 lock_map_acquire(&__fs_reclaim_map);
4085}
4086
4087void __fs_reclaim_release(void)
4088{
4089 lock_map_release(&__fs_reclaim_map);
4090}
4091
4092void fs_reclaim_acquire(gfp_t gfp_mask)
4093{
4094 if (__need_fs_reclaim(gfp_mask))
4095 __fs_reclaim_acquire();
4096}
4097EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4098
4099void fs_reclaim_release(gfp_t gfp_mask)
4100{
4101 if (__need_fs_reclaim(gfp_mask))
4102 __fs_reclaim_release();
4103}
4104EXPORT_SYMBOL_GPL(fs_reclaim_release);
4105#endif
4106
4107/* Perform direct synchronous page reclaim */
4108static int
4109__perform_reclaim(gfp_t gfp_mask, unsigned int order,
4110 const struct alloc_context *ac)
4111{
4112 int progress;
4113 unsigned int noreclaim_flag;
4114 unsigned long pflags;
4115
4116 cond_resched();
4117
4118 /* We now go into synchronous reclaim */
4119 cpuset_memory_pressure_bump();
4120 psi_memstall_enter(&pflags);
4121 fs_reclaim_acquire(gfp_mask);
4122 noreclaim_flag = memalloc_noreclaim_save();
4123
4124 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4125 ac->nodemask);
4126
4127 memalloc_noreclaim_restore(noreclaim_flag);
4128 fs_reclaim_release(gfp_mask);
4129 psi_memstall_leave(&pflags);
4130
4131 cond_resched();
4132
4133 return progress;
4134}
4135
4136/* The really slow allocator path where we enter direct reclaim */
4137static inline struct page *
4138__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4139 unsigned int alloc_flags, const struct alloc_context *ac,
4140 unsigned long *did_some_progress)
4141{
4142 struct page *page = NULL;
4143 bool drained = false;
4144
4145 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4146 if (unlikely(!(*did_some_progress)))
4147 return NULL;
4148
4149retry:
4150 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4151
4152 /*
4153 * If an allocation failed after direct reclaim, it could be because
4154 * pages are pinned on the per-cpu lists or in high alloc reserves.
4155 * Shrink them them and try again
4156 */
4157 if (!page && !drained) {
4158 unreserve_highatomic_pageblock(ac, false);
4159 drain_all_pages(NULL);
4160 drained = true;
4161 goto retry;
4162 }
4163
4164 return page;
4165}
4166
4167static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4168 const struct alloc_context *ac)
4169{
4170 struct zoneref *z;
4171 struct zone *zone;
4172 pg_data_t *last_pgdat = NULL;
4173 enum zone_type high_zoneidx = ac->high_zoneidx;
4174
4175 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4176 ac->nodemask) {
4177 if (last_pgdat != zone->zone_pgdat)
4178 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4179 last_pgdat = zone->zone_pgdat;
4180 }
4181}
4182
4183static inline unsigned int
4184gfp_to_alloc_flags(gfp_t gfp_mask)
4185{
4186 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4187
4188 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4189 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4190
4191 /*
4192 * The caller may dip into page reserves a bit more if the caller
4193 * cannot run direct reclaim, or if the caller has realtime scheduling
4194 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4195 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4196 */
4197 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4198
4199 if (gfp_mask & __GFP_ATOMIC) {
4200 /*
4201 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4202 * if it can't schedule.
4203 */
4204 if (!(gfp_mask & __GFP_NOMEMALLOC))
4205 alloc_flags |= ALLOC_HARDER;
4206 /*
4207 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4208 * comment for __cpuset_node_allowed().
4209 */
4210 alloc_flags &= ~ALLOC_CPUSET;
4211 } else if (unlikely(rt_task(current)) && !in_interrupt())
4212 alloc_flags |= ALLOC_HARDER;
4213
4214 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4215 alloc_flags |= ALLOC_KSWAPD;
4216
4217#ifdef CONFIG_CMA
4218 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4219 alloc_flags |= ALLOC_CMA;
4220#endif
4221 return alloc_flags;
4222}
4223
4224static bool oom_reserves_allowed(struct task_struct *tsk)
4225{
4226 if (!tsk_is_oom_victim(tsk))
4227 return false;
4228
4229 /*
4230 * !MMU doesn't have oom reaper so give access to memory reserves
4231 * only to the thread with TIF_MEMDIE set
4232 */
4233 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4234 return false;
4235
4236 return true;
4237}
4238
4239/*
4240 * Distinguish requests which really need access to full memory
4241 * reserves from oom victims which can live with a portion of it
4242 */
4243static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4244{
4245 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4246 return 0;
4247 if (gfp_mask & __GFP_MEMALLOC)
4248 return ALLOC_NO_WATERMARKS;
4249 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4250 return ALLOC_NO_WATERMARKS;
4251 if (!in_interrupt()) {
4252 if (current->flags & PF_MEMALLOC)
4253 return ALLOC_NO_WATERMARKS;
4254 else if (oom_reserves_allowed(current))
4255 return ALLOC_OOM;
4256 }
4257
4258 return 0;
4259}
4260
4261bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4262{
4263 return !!__gfp_pfmemalloc_flags(gfp_mask);
4264}
4265
4266/*
4267 * Checks whether it makes sense to retry the reclaim to make a forward progress
4268 * for the given allocation request.
4269 *
4270 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4271 * without success, or when we couldn't even meet the watermark if we
4272 * reclaimed all remaining pages on the LRU lists.
4273 *
4274 * Returns true if a retry is viable or false to enter the oom path.
4275 */
4276static inline bool
4277should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4278 struct alloc_context *ac, int alloc_flags,
4279 bool did_some_progress, int *no_progress_loops)
4280{
4281 struct zone *zone;
4282 struct zoneref *z;
4283 bool ret = false;
4284
4285 /*
4286 * Costly allocations might have made a progress but this doesn't mean
4287 * their order will become available due to high fragmentation so
4288 * always increment the no progress counter for them
4289 */
4290 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4291 *no_progress_loops = 0;
4292 else
4293 (*no_progress_loops)++;
4294
4295 /*
4296 * Make sure we converge to OOM if we cannot make any progress
4297 * several times in the row.
4298 */
4299 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4300 /* Before OOM, exhaust highatomic_reserve */
4301 return unreserve_highatomic_pageblock(ac, true);
4302 }
4303
4304 /*
4305 * Keep reclaiming pages while there is a chance this will lead
4306 * somewhere. If none of the target zones can satisfy our allocation
4307 * request even if all reclaimable pages are considered then we are
4308 * screwed and have to go OOM.
4309 */
4310 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4311 ac->nodemask) {
4312 unsigned long available;
4313 unsigned long reclaimable;
4314 unsigned long min_wmark = min_wmark_pages(zone);
4315 bool wmark;
4316
4317 available = reclaimable = zone_reclaimable_pages(zone);
4318 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4319
4320 /*
4321 * Would the allocation succeed if we reclaimed all
4322 * reclaimable pages?
4323 */
4324 wmark = __zone_watermark_ok(zone, order, min_wmark,
4325 ac_classzone_idx(ac), alloc_flags, available);
4326 trace_reclaim_retry_zone(z, order, reclaimable,
4327 available, min_wmark, *no_progress_loops, wmark);
4328 if (wmark) {
4329 /*
4330 * If we didn't make any progress and have a lot of
4331 * dirty + writeback pages then we should wait for
4332 * an IO to complete to slow down the reclaim and
4333 * prevent from pre mature OOM
4334 */
4335 if (!did_some_progress) {
4336 unsigned long write_pending;
4337
4338 write_pending = zone_page_state_snapshot(zone,
4339 NR_ZONE_WRITE_PENDING);
4340
4341 if (2 * write_pending > reclaimable) {
4342 congestion_wait(BLK_RW_ASYNC, HZ/10);
4343 return true;
4344 }
4345 }
4346
4347 ret = true;
4348 goto out;
4349 }
4350 }
4351
4352out:
4353 /*
4354 * Memory allocation/reclaim might be called from a WQ context and the
4355 * current implementation of the WQ concurrency control doesn't
4356 * recognize that a particular WQ is congested if the worker thread is
4357 * looping without ever sleeping. Therefore we have to do a short sleep
4358 * here rather than calling cond_resched().
4359 */
4360 if (current->flags & PF_WQ_WORKER)
4361 schedule_timeout_uninterruptible(1);
4362 else
4363 cond_resched();
4364 return ret;
4365}
4366
4367static inline bool
4368check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4369{
4370 /*
4371 * It's possible that cpuset's mems_allowed and the nodemask from
4372 * mempolicy don't intersect. This should be normally dealt with by
4373 * policy_nodemask(), but it's possible to race with cpuset update in
4374 * such a way the check therein was true, and then it became false
4375 * before we got our cpuset_mems_cookie here.
4376 * This assumes that for all allocations, ac->nodemask can come only
4377 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4378 * when it does not intersect with the cpuset restrictions) or the
4379 * caller can deal with a violated nodemask.
4380 */
4381 if (cpusets_enabled() && ac->nodemask &&
4382 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4383 ac->nodemask = NULL;
4384 return true;
4385 }
4386
4387 /*
4388 * When updating a task's mems_allowed or mempolicy nodemask, it is
4389 * possible to race with parallel threads in such a way that our
4390 * allocation can fail while the mask is being updated. If we are about
4391 * to fail, check if the cpuset changed during allocation and if so,
4392 * retry.
4393 */
4394 if (read_mems_allowed_retry(cpuset_mems_cookie))
4395 return true;
4396
4397 return false;
4398}
4399
4400static inline struct page *
4401__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4402 struct alloc_context *ac)
4403{
4404 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4405 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4406 struct page *page = NULL;
4407 unsigned int alloc_flags;
4408 unsigned long did_some_progress;
4409 enum compact_priority compact_priority;
4410 enum compact_result compact_result;
4411 int compaction_retries;
4412 int no_progress_loops;
4413 unsigned int cpuset_mems_cookie;
4414 int reserve_flags;
4415
4416 /*
4417 * We also sanity check to catch abuse of atomic reserves being used by
4418 * callers that are not in atomic context.
4419 */
4420 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4421 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4422 gfp_mask &= ~__GFP_ATOMIC;
4423
4424retry_cpuset:
4425 compaction_retries = 0;
4426 no_progress_loops = 0;
4427 compact_priority = DEF_COMPACT_PRIORITY;
4428 cpuset_mems_cookie = read_mems_allowed_begin();
4429
4430 /*
4431 * The fast path uses conservative alloc_flags to succeed only until
4432 * kswapd needs to be woken up, and to avoid the cost of setting up
4433 * alloc_flags precisely. So we do that now.
4434 */
4435 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4436
4437 /*
4438 * We need to recalculate the starting point for the zonelist iterator
4439 * because we might have used different nodemask in the fast path, or
4440 * there was a cpuset modification and we are retrying - otherwise we
4441 * could end up iterating over non-eligible zones endlessly.
4442 */
4443 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4444 ac->high_zoneidx, ac->nodemask);
4445 if (!ac->preferred_zoneref->zone)
4446 goto nopage;
4447
4448 if (alloc_flags & ALLOC_KSWAPD)
4449 wake_all_kswapds(order, gfp_mask, ac);
4450
4451 /*
4452 * The adjusted alloc_flags might result in immediate success, so try
4453 * that first
4454 */
4455 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4456 if (page)
4457 goto got_pg;
4458
4459 /*
4460 * For costly allocations, try direct compaction first, as it's likely
4461 * that we have enough base pages and don't need to reclaim. For non-
4462 * movable high-order allocations, do that as well, as compaction will
4463 * try prevent permanent fragmentation by migrating from blocks of the
4464 * same migratetype.
4465 * Don't try this for allocations that are allowed to ignore
4466 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4467 */
4468 if (can_direct_reclaim &&
4469 (costly_order ||
4470 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4471 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4472 page = __alloc_pages_direct_compact(gfp_mask, order,
4473 alloc_flags, ac,
4474 INIT_COMPACT_PRIORITY,
4475 &compact_result);
4476 if (page)
4477 goto got_pg;
4478
4479 if (order >= pageblock_order && (gfp_mask & __GFP_IO) &&
4480 !(gfp_mask & __GFP_RETRY_MAYFAIL)) {
4481 /*
4482 * If allocating entire pageblock(s) and compaction
4483 * failed because all zones are below low watermarks
4484 * or is prohibited because it recently failed at this
4485 * order, fail immediately unless the allocator has
4486 * requested compaction and reclaim retry.
4487 *
4488 * Reclaim is
4489 * - potentially very expensive because zones are far
4490 * below their low watermarks or this is part of very
4491 * bursty high order allocations,
4492 * - not guaranteed to help because isolate_freepages()
4493 * may not iterate over freed pages as part of its
4494 * linear scan, and
4495 * - unlikely to make entire pageblocks free on its
4496 * own.
4497 */
4498 if (compact_result == COMPACT_SKIPPED ||
4499 compact_result == COMPACT_DEFERRED)
4500 goto nopage;
4501 }
4502
4503 /*
4504 * Checks for costly allocations with __GFP_NORETRY, which
4505 * includes THP page fault allocations
4506 */
4507 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4508 /*
4509 * If compaction is deferred for high-order allocations,
4510 * it is because sync compaction recently failed. If
4511 * this is the case and the caller requested a THP
4512 * allocation, we do not want to heavily disrupt the
4513 * system, so we fail the allocation instead of entering
4514 * direct reclaim.
4515 */
4516 if (compact_result == COMPACT_DEFERRED)
4517 goto nopage;
4518
4519 /*
4520 * Looks like reclaim/compaction is worth trying, but
4521 * sync compaction could be very expensive, so keep
4522 * using async compaction.
4523 */
4524 compact_priority = INIT_COMPACT_PRIORITY;
4525 }
4526 }
4527
4528retry:
4529 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4530 if (alloc_flags & ALLOC_KSWAPD)
4531 wake_all_kswapds(order, gfp_mask, ac);
4532
4533 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4534 if (reserve_flags)
4535 alloc_flags = reserve_flags;
4536
4537 /*
4538 * Reset the nodemask and zonelist iterators if memory policies can be
4539 * ignored. These allocations are high priority and system rather than
4540 * user oriented.
4541 */
4542 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4543 ac->nodemask = NULL;
4544 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4545 ac->high_zoneidx, ac->nodemask);
4546 }
4547
4548 /* Attempt with potentially adjusted zonelist and alloc_flags */
4549 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4550 if (page)
4551 goto got_pg;
4552
4553 /* Caller is not willing to reclaim, we can't balance anything */
4554 if (!can_direct_reclaim)
4555 goto nopage;
4556
4557 /* Avoid recursion of direct reclaim */
4558 if (current->flags & PF_MEMALLOC)
4559 goto nopage;
4560
4561 /* Try direct reclaim and then allocating */
4562 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4563 &did_some_progress);
4564 if (page)
4565 goto got_pg;
4566
4567 /* Try direct compaction and then allocating */
4568 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4569 compact_priority, &compact_result);
4570 if (page)
4571 goto got_pg;
4572
4573 /* Do not loop if specifically requested */
4574 if (gfp_mask & __GFP_NORETRY)
4575 goto nopage;
4576
4577 /*
4578 * Do not retry costly high order allocations unless they are
4579 * __GFP_RETRY_MAYFAIL
4580 */
4581 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4582 goto nopage;
4583
4584 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4585 did_some_progress > 0, &no_progress_loops))
4586 goto retry;
4587
4588 /*
4589 * It doesn't make any sense to retry for the compaction if the order-0
4590 * reclaim is not able to make any progress because the current
4591 * implementation of the compaction depends on the sufficient amount
4592 * of free memory (see __compaction_suitable)
4593 */
4594 if (did_some_progress > 0 &&
4595 should_compact_retry(ac, order, alloc_flags,
4596 compact_result, &compact_priority,
4597 &compaction_retries))
4598 goto retry;
4599
4600
4601 /* Deal with possible cpuset update races before we start OOM killing */
4602 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4603 goto retry_cpuset;
4604
4605 /* Reclaim has failed us, start killing things */
4606 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4607 if (page)
4608 goto got_pg;
4609
4610 /* Avoid allocations with no watermarks from looping endlessly */
4611 if (tsk_is_oom_victim(current) &&
4612 (alloc_flags == ALLOC_OOM ||
4613 (gfp_mask & __GFP_NOMEMALLOC)))
4614 goto nopage;
4615
4616 /* Retry as long as the OOM killer is making progress */
4617 if (did_some_progress) {
4618 no_progress_loops = 0;
4619 goto retry;
4620 }
4621
4622nopage:
4623 /* Deal with possible cpuset update races before we fail */
4624 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4625 goto retry_cpuset;
4626
4627 /*
4628 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4629 * we always retry
4630 */
4631 if (gfp_mask & __GFP_NOFAIL) {
4632 /*
4633 * All existing users of the __GFP_NOFAIL are blockable, so warn
4634 * of any new users that actually require GFP_NOWAIT
4635 */
4636 if (WARN_ON_ONCE(!can_direct_reclaim))
4637 goto fail;
4638
4639 /*
4640 * PF_MEMALLOC request from this context is rather bizarre
4641 * because we cannot reclaim anything and only can loop waiting
4642 * for somebody to do a work for us
4643 */
4644 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4645
4646 /*
4647 * non failing costly orders are a hard requirement which we
4648 * are not prepared for much so let's warn about these users
4649 * so that we can identify them and convert them to something
4650 * else.
4651 */
4652 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4653
4654 /*
4655 * Help non-failing allocations by giving them access to memory
4656 * reserves but do not use ALLOC_NO_WATERMARKS because this
4657 * could deplete whole memory reserves which would just make
4658 * the situation worse
4659 */
4660 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4661 if (page)
4662 goto got_pg;
4663
4664 cond_resched();
4665 goto retry;
4666 }
4667fail:
4668 warn_alloc(gfp_mask, ac->nodemask,
4669 "page allocation failure: order:%u", order);
4670got_pg:
4671 return page;
4672}
4673
4674static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4675 int preferred_nid, nodemask_t *nodemask,
4676 struct alloc_context *ac, gfp_t *alloc_mask,
4677 unsigned int *alloc_flags)
4678{
4679 ac->high_zoneidx = gfp_zone(gfp_mask);
4680 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4681 ac->nodemask = nodemask;
4682 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4683
4684 if (cpusets_enabled()) {
4685 *alloc_mask |= __GFP_HARDWALL;
4686 if (!ac->nodemask)
4687 ac->nodemask = &cpuset_current_mems_allowed;
4688 else
4689 *alloc_flags |= ALLOC_CPUSET;
4690 }
4691
4692 fs_reclaim_acquire(gfp_mask);
4693 fs_reclaim_release(gfp_mask);
4694
4695 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4696
4697 if (should_fail_alloc_page(gfp_mask, order))
4698 return false;
4699
4700 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4701 *alloc_flags |= ALLOC_CMA;
4702
4703 return true;
4704}
4705
4706/* Determine whether to spread dirty pages and what the first usable zone */
4707static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4708{
4709 /* Dirty zone balancing only done in the fast path */
4710 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4711
4712 /*
4713 * The preferred zone is used for statistics but crucially it is
4714 * also used as the starting point for the zonelist iterator. It
4715 * may get reset for allocations that ignore memory policies.
4716 */
4717 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4718 ac->high_zoneidx, ac->nodemask);
4719}
4720
4721/*
4722 * This is the 'heart' of the zoned buddy allocator.
4723 */
4724struct page *
4725__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4726 nodemask_t *nodemask)
4727{
4728 struct page *page;
4729 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4730 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4731 struct alloc_context ac = { };
4732
4733 /*
4734 * There are several places where we assume that the order value is sane
4735 * so bail out early if the request is out of bound.
4736 */
4737 if (unlikely(order >= MAX_ORDER)) {
4738 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4739 return NULL;
4740 }
4741
4742 gfp_mask &= gfp_allowed_mask;
4743 alloc_mask = gfp_mask;
4744 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4745 return NULL;
4746
4747 finalise_ac(gfp_mask, &ac);
4748
4749 /*
4750 * Forbid the first pass from falling back to types that fragment
4751 * memory until all local zones are considered.
4752 */
4753 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4754
4755 /* First allocation attempt */
4756 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4757 if (likely(page))
4758 goto out;
4759
4760 /*
4761 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4762 * resp. GFP_NOIO which has to be inherited for all allocation requests
4763 * from a particular context which has been marked by
4764 * memalloc_no{fs,io}_{save,restore}.
4765 */
4766 alloc_mask = current_gfp_context(gfp_mask);
4767 ac.spread_dirty_pages = false;
4768
4769 /*
4770 * Restore the original nodemask if it was potentially replaced with
4771 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4772 */
4773 if (unlikely(ac.nodemask != nodemask))
4774 ac.nodemask = nodemask;
4775
4776 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4777
4778out:
4779 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4780 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4781 __free_pages(page, order);
4782 page = NULL;
4783 }
4784
4785 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4786
4787 return page;
4788}
4789EXPORT_SYMBOL(__alloc_pages_nodemask);
4790
4791/*
4792 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4793 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4794 * you need to access high mem.
4795 */
4796unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4797{
4798 struct page *page;
4799
4800 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4801 if (!page)
4802 return 0;
4803 return (unsigned long) page_address(page);
4804}
4805EXPORT_SYMBOL(__get_free_pages);
4806
4807unsigned long get_zeroed_page(gfp_t gfp_mask)
4808{
4809 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4810}
4811EXPORT_SYMBOL(get_zeroed_page);
4812
4813static inline void free_the_page(struct page *page, unsigned int order)
4814{
4815 if (order == 0) /* Via pcp? */
4816 free_unref_page(page);
4817 else
4818 __free_pages_ok(page, order);
4819}
4820
4821void __free_pages(struct page *page, unsigned int order)
4822{
4823 if (put_page_testzero(page))
4824 free_the_page(page, order);
4825}
4826EXPORT_SYMBOL(__free_pages);
4827
4828void free_pages(unsigned long addr, unsigned int order)
4829{
4830 if (addr != 0) {
4831 VM_BUG_ON(!virt_addr_valid((void *)addr));
4832 __free_pages(virt_to_page((void *)addr), order);
4833 }
4834}
4835
4836EXPORT_SYMBOL(free_pages);
4837
4838/*
4839 * Page Fragment:
4840 * An arbitrary-length arbitrary-offset area of memory which resides
4841 * within a 0 or higher order page. Multiple fragments within that page
4842 * are individually refcounted, in the page's reference counter.
4843 *
4844 * The page_frag functions below provide a simple allocation framework for
4845 * page fragments. This is used by the network stack and network device
4846 * drivers to provide a backing region of memory for use as either an
4847 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4848 */
4849static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4850 gfp_t gfp_mask)
4851{
4852 struct page *page = NULL;
4853 gfp_t gfp = gfp_mask;
4854
4855#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4856 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4857 __GFP_NOMEMALLOC;
4858 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4859 PAGE_FRAG_CACHE_MAX_ORDER);
4860 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4861#endif
4862 if (unlikely(!page))
4863 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4864
4865 nc->va = page ? page_address(page) : NULL;
4866
4867 return page;
4868}
4869
4870void __page_frag_cache_drain(struct page *page, unsigned int count)
4871{
4872 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4873
4874 if (page_ref_sub_and_test(page, count))
4875 free_the_page(page, compound_order(page));
4876}
4877EXPORT_SYMBOL(__page_frag_cache_drain);
4878
4879void *page_frag_alloc(struct page_frag_cache *nc,
4880 unsigned int fragsz, gfp_t gfp_mask)
4881{
4882 unsigned int size = PAGE_SIZE;
4883 struct page *page;
4884 int offset;
4885
4886 if (unlikely(!nc->va)) {
4887refill:
4888 page = __page_frag_cache_refill(nc, gfp_mask);
4889 if (!page)
4890 return NULL;
4891
4892#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4893 /* if size can vary use size else just use PAGE_SIZE */
4894 size = nc->size;
4895#endif
4896 /* Even if we own the page, we do not use atomic_set().
4897 * This would break get_page_unless_zero() users.
4898 */
4899 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4900
4901 /* reset page count bias and offset to start of new frag */
4902 nc->pfmemalloc = page_is_pfmemalloc(page);
4903 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4904 nc->offset = size;
4905 }
4906
4907 offset = nc->offset - fragsz;
4908 if (unlikely(offset < 0)) {
4909 page = virt_to_page(nc->va);
4910
4911 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4912 goto refill;
4913
4914#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4915 /* if size can vary use size else just use PAGE_SIZE */
4916 size = nc->size;
4917#endif
4918 /* OK, page count is 0, we can safely set it */
4919 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4920
4921 /* reset page count bias and offset to start of new frag */
4922 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4923 offset = size - fragsz;
4924 }
4925
4926 nc->pagecnt_bias--;
4927 nc->offset = offset;
4928
4929 return nc->va + offset;
4930}
4931EXPORT_SYMBOL(page_frag_alloc);
4932
4933/*
4934 * Frees a page fragment allocated out of either a compound or order 0 page.
4935 */
4936void page_frag_free(void *addr)
4937{
4938 struct page *page = virt_to_head_page(addr);
4939
4940 if (unlikely(put_page_testzero(page)))
4941 free_the_page(page, compound_order(page));
4942}
4943EXPORT_SYMBOL(page_frag_free);
4944
4945static void *make_alloc_exact(unsigned long addr, unsigned int order,
4946 size_t size)
4947{
4948 if (addr) {
4949 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4950 unsigned long used = addr + PAGE_ALIGN(size);
4951
4952 split_page(virt_to_page((void *)addr), order);
4953 while (used < alloc_end) {
4954 free_page(used);
4955 used += PAGE_SIZE;
4956 }
4957 }
4958 return (void *)addr;
4959}
4960
4961/**
4962 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4963 * @size: the number of bytes to allocate
4964 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4965 *
4966 * This function is similar to alloc_pages(), except that it allocates the
4967 * minimum number of pages to satisfy the request. alloc_pages() can only
4968 * allocate memory in power-of-two pages.
4969 *
4970 * This function is also limited by MAX_ORDER.
4971 *
4972 * Memory allocated by this function must be released by free_pages_exact().
4973 *
4974 * Return: pointer to the allocated area or %NULL in case of error.
4975 */
4976void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4977{
4978 unsigned int order = get_order(size);
4979 unsigned long addr;
4980
4981 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4982 gfp_mask &= ~__GFP_COMP;
4983
4984 addr = __get_free_pages(gfp_mask, order);
4985 return make_alloc_exact(addr, order, size);
4986}
4987EXPORT_SYMBOL(alloc_pages_exact);
4988
4989/**
4990 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4991 * pages on a node.
4992 * @nid: the preferred node ID where memory should be allocated
4993 * @size: the number of bytes to allocate
4994 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4995 *
4996 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4997 * back.
4998 *
4999 * Return: pointer to the allocated area or %NULL in case of error.
5000 */
5001void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5002{
5003 unsigned int order = get_order(size);
5004 struct page *p;
5005
5006 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5007 gfp_mask &= ~__GFP_COMP;
5008
5009 p = alloc_pages_node(nid, gfp_mask, order);
5010 if (!p)
5011 return NULL;
5012 return make_alloc_exact((unsigned long)page_address(p), order, size);
5013}
5014
5015/**
5016 * free_pages_exact - release memory allocated via alloc_pages_exact()
5017 * @virt: the value returned by alloc_pages_exact.
5018 * @size: size of allocation, same value as passed to alloc_pages_exact().
5019 *
5020 * Release the memory allocated by a previous call to alloc_pages_exact.
5021 */
5022void free_pages_exact(void *virt, size_t size)
5023{
5024 unsigned long addr = (unsigned long)virt;
5025 unsigned long end = addr + PAGE_ALIGN(size);
5026
5027 while (addr < end) {
5028 free_page(addr);
5029 addr += PAGE_SIZE;
5030 }
5031}
5032EXPORT_SYMBOL(free_pages_exact);
5033
5034/**
5035 * nr_free_zone_pages - count number of pages beyond high watermark
5036 * @offset: The zone index of the highest zone
5037 *
5038 * nr_free_zone_pages() counts the number of pages which are beyond the
5039 * high watermark within all zones at or below a given zone index. For each
5040 * zone, the number of pages is calculated as:
5041 *
5042 * nr_free_zone_pages = managed_pages - high_pages
5043 *
5044 * Return: number of pages beyond high watermark.
5045 */
5046static unsigned long nr_free_zone_pages(int offset)
5047{
5048 struct zoneref *z;
5049 struct zone *zone;
5050
5051 /* Just pick one node, since fallback list is circular */
5052 unsigned long sum = 0;
5053
5054 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5055
5056 for_each_zone_zonelist(zone, z, zonelist, offset) {
5057 unsigned long size = zone_managed_pages(zone);
5058 unsigned long high = high_wmark_pages(zone);
5059 if (size > high)
5060 sum += size - high;
5061 }
5062
5063 return sum;
5064}
5065
5066/**
5067 * nr_free_buffer_pages - count number of pages beyond high watermark
5068 *
5069 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5070 * watermark within ZONE_DMA and ZONE_NORMAL.
5071 *
5072 * Return: number of pages beyond high watermark within ZONE_DMA and
5073 * ZONE_NORMAL.
5074 */
5075unsigned long nr_free_buffer_pages(void)
5076{
5077 return nr_free_zone_pages(gfp_zone(GFP_USER));
5078}
5079EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5080
5081/**
5082 * nr_free_pagecache_pages - count number of pages beyond high watermark
5083 *
5084 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5085 * high watermark within all zones.
5086 *
5087 * Return: number of pages beyond high watermark within all zones.
5088 */
5089unsigned long nr_free_pagecache_pages(void)
5090{
5091 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5092}
5093
5094static inline void show_node(struct zone *zone)
5095{
5096 if (IS_ENABLED(CONFIG_NUMA))
5097 printk("Node %d ", zone_to_nid(zone));
5098}
5099
5100long si_mem_available(void)
5101{
5102 long available;
5103 unsigned long pagecache;
5104 unsigned long wmark_low = 0;
5105 unsigned long pages[NR_LRU_LISTS];
5106 unsigned long reclaimable;
5107 struct zone *zone;
5108 int lru;
5109
5110 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5111 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5112
5113 for_each_zone(zone)
5114 wmark_low += low_wmark_pages(zone);
5115
5116 /*
5117 * Estimate the amount of memory available for userspace allocations,
5118 * without causing swapping.
5119 */
5120 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5121
5122 /*
5123 * Not all the page cache can be freed, otherwise the system will
5124 * start swapping. Assume at least half of the page cache, or the
5125 * low watermark worth of cache, needs to stay.
5126 */
5127 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5128 pagecache -= min(pagecache / 2, wmark_low);
5129 available += pagecache;
5130
5131 /*
5132 * Part of the reclaimable slab and other kernel memory consists of
5133 * items that are in use, and cannot be freed. Cap this estimate at the
5134 * low watermark.
5135 */
5136 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5137 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5138 available += reclaimable - min(reclaimable / 2, wmark_low);
5139
5140 if (available < 0)
5141 available = 0;
5142 return available;
5143}
5144EXPORT_SYMBOL_GPL(si_mem_available);
5145
5146void si_meminfo(struct sysinfo *val)
5147{
5148 val->totalram = totalram_pages();
5149 val->sharedram = global_node_page_state(NR_SHMEM);
5150 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5151 val->bufferram = nr_blockdev_pages();
5152 val->totalhigh = totalhigh_pages();
5153 val->freehigh = nr_free_highpages();
5154 val->mem_unit = PAGE_SIZE;
5155}
5156
5157EXPORT_SYMBOL(si_meminfo);
5158
5159#ifdef CONFIG_NUMA
5160void si_meminfo_node(struct sysinfo *val, int nid)
5161{
5162 int zone_type; /* needs to be signed */
5163 unsigned long managed_pages = 0;
5164 unsigned long managed_highpages = 0;
5165 unsigned long free_highpages = 0;
5166 pg_data_t *pgdat = NODE_DATA(nid);
5167
5168 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5169 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5170 val->totalram = managed_pages;
5171 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5172 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5173#ifdef CONFIG_HIGHMEM
5174 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5175 struct zone *zone = &pgdat->node_zones[zone_type];
5176
5177 if (is_highmem(zone)) {
5178 managed_highpages += zone_managed_pages(zone);
5179 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5180 }
5181 }
5182 val->totalhigh = managed_highpages;
5183 val->freehigh = free_highpages;
5184#else
5185 val->totalhigh = managed_highpages;
5186 val->freehigh = free_highpages;
5187#endif
5188 val->mem_unit = PAGE_SIZE;
5189}
5190#endif
5191
5192/*
5193 * Determine whether the node should be displayed or not, depending on whether
5194 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5195 */
5196static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5197{
5198 if (!(flags & SHOW_MEM_FILTER_NODES))
5199 return false;
5200
5201 /*
5202 * no node mask - aka implicit memory numa policy. Do not bother with
5203 * the synchronization - read_mems_allowed_begin - because we do not
5204 * have to be precise here.
5205 */
5206 if (!nodemask)
5207 nodemask = &cpuset_current_mems_allowed;
5208
5209 return !node_isset(nid, *nodemask);
5210}
5211
5212#define K(x) ((x) << (PAGE_SHIFT-10))
5213
5214static void show_migration_types(unsigned char type)
5215{
5216 static const char types[MIGRATE_TYPES] = {
5217 [MIGRATE_UNMOVABLE] = 'U',
5218 [MIGRATE_MOVABLE] = 'M',
5219 [MIGRATE_RECLAIMABLE] = 'E',
5220 [MIGRATE_HIGHATOMIC] = 'H',
5221#ifdef CONFIG_CMA
5222 [MIGRATE_CMA] = 'C',
5223#endif
5224#ifdef CONFIG_MEMORY_ISOLATION
5225 [MIGRATE_ISOLATE] = 'I',
5226#endif
5227 };
5228 char tmp[MIGRATE_TYPES + 1];
5229 char *p = tmp;
5230 int i;
5231
5232 for (i = 0; i < MIGRATE_TYPES; i++) {
5233 if (type & (1 << i))
5234 *p++ = types[i];
5235 }
5236
5237 *p = '\0';
5238 printk(KERN_CONT "(%s) ", tmp);
5239}
5240
5241/*
5242 * Show free area list (used inside shift_scroll-lock stuff)
5243 * We also calculate the percentage fragmentation. We do this by counting the
5244 * memory on each free list with the exception of the first item on the list.
5245 *
5246 * Bits in @filter:
5247 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5248 * cpuset.
5249 */
5250void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5251{
5252 unsigned long free_pcp = 0;
5253 int cpu;
5254 struct zone *zone;
5255 pg_data_t *pgdat;
5256
5257 for_each_populated_zone(zone) {
5258 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5259 continue;
5260
5261 for_each_online_cpu(cpu)
5262 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5263 }
5264
5265 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5266 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5267 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5268 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5269 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5270 " free:%lu free_pcp:%lu free_cma:%lu\n",
5271 global_node_page_state(NR_ACTIVE_ANON),
5272 global_node_page_state(NR_INACTIVE_ANON),
5273 global_node_page_state(NR_ISOLATED_ANON),
5274 global_node_page_state(NR_ACTIVE_FILE),
5275 global_node_page_state(NR_INACTIVE_FILE),
5276 global_node_page_state(NR_ISOLATED_FILE),
5277 global_node_page_state(NR_UNEVICTABLE),
5278 global_node_page_state(NR_FILE_DIRTY),
5279 global_node_page_state(NR_WRITEBACK),
5280 global_node_page_state(NR_UNSTABLE_NFS),
5281 global_node_page_state(NR_SLAB_RECLAIMABLE),
5282 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5283 global_node_page_state(NR_FILE_MAPPED),
5284 global_node_page_state(NR_SHMEM),
5285 global_zone_page_state(NR_PAGETABLE),
5286 global_zone_page_state(NR_BOUNCE),
5287 global_zone_page_state(NR_FREE_PAGES),
5288 free_pcp,
5289 global_zone_page_state(NR_FREE_CMA_PAGES));
5290
5291 for_each_online_pgdat(pgdat) {
5292 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5293 continue;
5294
5295 printk("Node %d"
5296 " active_anon:%lukB"
5297 " inactive_anon:%lukB"
5298 " active_file:%lukB"
5299 " inactive_file:%lukB"
5300 " unevictable:%lukB"
5301 " isolated(anon):%lukB"
5302 " isolated(file):%lukB"
5303 " mapped:%lukB"
5304 " dirty:%lukB"
5305 " writeback:%lukB"
5306 " shmem:%lukB"
5307#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5308 " shmem_thp: %lukB"
5309 " shmem_pmdmapped: %lukB"
5310 " anon_thp: %lukB"
5311#endif
5312 " writeback_tmp:%lukB"
5313 " unstable:%lukB"
5314 " all_unreclaimable? %s"
5315 "\n",
5316 pgdat->node_id,
5317 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5318 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5319 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5320 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5321 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5322 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5323 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5324 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5325 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5326 K(node_page_state(pgdat, NR_WRITEBACK)),
5327 K(node_page_state(pgdat, NR_SHMEM)),
5328#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5329 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5330 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5331 * HPAGE_PMD_NR),
5332 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5333#endif
5334 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5335 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5336 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5337 "yes" : "no");
5338 }
5339
5340 for_each_populated_zone(zone) {
5341 int i;
5342
5343 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5344 continue;
5345
5346 free_pcp = 0;
5347 for_each_online_cpu(cpu)
5348 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5349
5350 show_node(zone);
5351 printk(KERN_CONT
5352 "%s"
5353 " free:%lukB"
5354 " min:%lukB"
5355 " low:%lukB"
5356 " high:%lukB"
5357 " active_anon:%lukB"
5358 " inactive_anon:%lukB"
5359 " active_file:%lukB"
5360 " inactive_file:%lukB"
5361 " unevictable:%lukB"
5362 " writepending:%lukB"
5363 " present:%lukB"
5364 " managed:%lukB"
5365 " mlocked:%lukB"
5366 " kernel_stack:%lukB"
5367 " pagetables:%lukB"
5368 " bounce:%lukB"
5369 " free_pcp:%lukB"
5370 " local_pcp:%ukB"
5371 " free_cma:%lukB"
5372 "\n",
5373 zone->name,
5374 K(zone_page_state(zone, NR_FREE_PAGES)),
5375 K(min_wmark_pages(zone)),
5376 K(low_wmark_pages(zone)),
5377 K(high_wmark_pages(zone)),
5378 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5379 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5380 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5381 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5382 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5383 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5384 K(zone->present_pages),
5385 K(zone_managed_pages(zone)),
5386 K(zone_page_state(zone, NR_MLOCK)),
5387 zone_page_state(zone, NR_KERNEL_STACK_KB),
5388 K(zone_page_state(zone, NR_PAGETABLE)),
5389 K(zone_page_state(zone, NR_BOUNCE)),
5390 K(free_pcp),
5391 K(this_cpu_read(zone->pageset->pcp.count)),
5392 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5393 printk("lowmem_reserve[]:");
5394 for (i = 0; i < MAX_NR_ZONES; i++)
5395 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5396 printk(KERN_CONT "\n");
5397 }
5398
5399 for_each_populated_zone(zone) {
5400 unsigned int order;
5401 unsigned long nr[MAX_ORDER], flags, total = 0;
5402 unsigned char types[MAX_ORDER];
5403
5404 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5405 continue;
5406 show_node(zone);
5407 printk(KERN_CONT "%s: ", zone->name);
5408
5409 spin_lock_irqsave(&zone->lock, flags);
5410 for (order = 0; order < MAX_ORDER; order++) {
5411 struct free_area *area = &zone->free_area[order];
5412 int type;
5413
5414 nr[order] = area->nr_free;
5415 total += nr[order] << order;
5416
5417 types[order] = 0;
5418 for (type = 0; type < MIGRATE_TYPES; type++) {
5419 if (!free_area_empty(area, type))
5420 types[order] |= 1 << type;
5421 }
5422 }
5423 spin_unlock_irqrestore(&zone->lock, flags);
5424 for (order = 0; order < MAX_ORDER; order++) {
5425 printk(KERN_CONT "%lu*%lukB ",
5426 nr[order], K(1UL) << order);
5427 if (nr[order])
5428 show_migration_types(types[order]);
5429 }
5430 printk(KERN_CONT "= %lukB\n", K(total));
5431 }
5432
5433 hugetlb_show_meminfo();
5434
5435 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5436
5437 show_swap_cache_info();
5438}
5439
5440static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5441{
5442 zoneref->zone = zone;
5443 zoneref->zone_idx = zone_idx(zone);
5444}
5445
5446/*
5447 * Builds allocation fallback zone lists.
5448 *
5449 * Add all populated zones of a node to the zonelist.
5450 */
5451static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5452{
5453 struct zone *zone;
5454 enum zone_type zone_type = MAX_NR_ZONES;
5455 int nr_zones = 0;
5456
5457 do {
5458 zone_type--;
5459 zone = pgdat->node_zones + zone_type;
5460 if (managed_zone(zone)) {
5461 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5462 check_highest_zone(zone_type);
5463 }
5464 } while (zone_type);
5465
5466 return nr_zones;
5467}
5468
5469#ifdef CONFIG_NUMA
5470
5471static int __parse_numa_zonelist_order(char *s)
5472{
5473 /*
5474 * We used to support different zonlists modes but they turned
5475 * out to be just not useful. Let's keep the warning in place
5476 * if somebody still use the cmd line parameter so that we do
5477 * not fail it silently
5478 */
5479 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5480 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5481 return -EINVAL;
5482 }
5483 return 0;
5484}
5485
5486static __init int setup_numa_zonelist_order(char *s)
5487{
5488 if (!s)
5489 return 0;
5490
5491 return __parse_numa_zonelist_order(s);
5492}
5493early_param("numa_zonelist_order", setup_numa_zonelist_order);
5494
5495char numa_zonelist_order[] = "Node";
5496
5497/*
5498 * sysctl handler for numa_zonelist_order
5499 */
5500int numa_zonelist_order_handler(struct ctl_table *table, int write,
5501 void __user *buffer, size_t *length,
5502 loff_t *ppos)
5503{
5504 char *str;
5505 int ret;
5506
5507 if (!write)
5508 return proc_dostring(table, write, buffer, length, ppos);
5509 str = memdup_user_nul(buffer, 16);
5510 if (IS_ERR(str))
5511 return PTR_ERR(str);
5512
5513 ret = __parse_numa_zonelist_order(str);
5514 kfree(str);
5515 return ret;
5516}
5517
5518
5519#define MAX_NODE_LOAD (nr_online_nodes)
5520static int node_load[MAX_NUMNODES];
5521
5522/**
5523 * find_next_best_node - find the next node that should appear in a given node's fallback list
5524 * @node: node whose fallback list we're appending
5525 * @used_node_mask: nodemask_t of already used nodes
5526 *
5527 * We use a number of factors to determine which is the next node that should
5528 * appear on a given node's fallback list. The node should not have appeared
5529 * already in @node's fallback list, and it should be the next closest node
5530 * according to the distance array (which contains arbitrary distance values
5531 * from each node to each node in the system), and should also prefer nodes
5532 * with no CPUs, since presumably they'll have very little allocation pressure
5533 * on them otherwise.
5534 *
5535 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5536 */
5537static int find_next_best_node(int node, nodemask_t *used_node_mask)
5538{
5539 int n, val;
5540 int min_val = INT_MAX;
5541 int best_node = NUMA_NO_NODE;
5542 const struct cpumask *tmp = cpumask_of_node(0);
5543
5544 /* Use the local node if we haven't already */
5545 if (!node_isset(node, *used_node_mask)) {
5546 node_set(node, *used_node_mask);
5547 return node;
5548 }
5549
5550 for_each_node_state(n, N_MEMORY) {
5551
5552 /* Don't want a node to appear more than once */
5553 if (node_isset(n, *used_node_mask))
5554 continue;
5555
5556 /* Use the distance array to find the distance */
5557 val = node_distance(node, n);
5558
5559 /* Penalize nodes under us ("prefer the next node") */
5560 val += (n < node);
5561
5562 /* Give preference to headless and unused nodes */
5563 tmp = cpumask_of_node(n);
5564 if (!cpumask_empty(tmp))
5565 val += PENALTY_FOR_NODE_WITH_CPUS;
5566
5567 /* Slight preference for less loaded node */
5568 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5569 val += node_load[n];
5570
5571 if (val < min_val) {
5572 min_val = val;
5573 best_node = n;
5574 }
5575 }
5576
5577 if (best_node >= 0)
5578 node_set(best_node, *used_node_mask);
5579
5580 return best_node;
5581}
5582
5583
5584/*
5585 * Build zonelists ordered by node and zones within node.
5586 * This results in maximum locality--normal zone overflows into local
5587 * DMA zone, if any--but risks exhausting DMA zone.
5588 */
5589static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5590 unsigned nr_nodes)
5591{
5592 struct zoneref *zonerefs;
5593 int i;
5594
5595 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5596
5597 for (i = 0; i < nr_nodes; i++) {
5598 int nr_zones;
5599
5600 pg_data_t *node = NODE_DATA(node_order[i]);
5601
5602 nr_zones = build_zonerefs_node(node, zonerefs);
5603 zonerefs += nr_zones;
5604 }
5605 zonerefs->zone = NULL;
5606 zonerefs->zone_idx = 0;
5607}
5608
5609/*
5610 * Build gfp_thisnode zonelists
5611 */
5612static void build_thisnode_zonelists(pg_data_t *pgdat)
5613{
5614 struct zoneref *zonerefs;
5615 int nr_zones;
5616
5617 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5618 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5619 zonerefs += nr_zones;
5620 zonerefs->zone = NULL;
5621 zonerefs->zone_idx = 0;
5622}
5623
5624/*
5625 * Build zonelists ordered by zone and nodes within zones.
5626 * This results in conserving DMA zone[s] until all Normal memory is
5627 * exhausted, but results in overflowing to remote node while memory
5628 * may still exist in local DMA zone.
5629 */
5630
5631static void build_zonelists(pg_data_t *pgdat)
5632{
5633 static int node_order[MAX_NUMNODES];
5634 int node, load, nr_nodes = 0;
5635 nodemask_t used_mask;
5636 int local_node, prev_node;
5637
5638 /* NUMA-aware ordering of nodes */
5639 local_node = pgdat->node_id;
5640 load = nr_online_nodes;
5641 prev_node = local_node;
5642 nodes_clear(used_mask);
5643
5644 memset(node_order, 0, sizeof(node_order));
5645 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5646 /*
5647 * We don't want to pressure a particular node.
5648 * So adding penalty to the first node in same
5649 * distance group to make it round-robin.
5650 */
5651 if (node_distance(local_node, node) !=
5652 node_distance(local_node, prev_node))
5653 node_load[node] = load;
5654
5655 node_order[nr_nodes++] = node;
5656 prev_node = node;
5657 load--;
5658 }
5659
5660 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5661 build_thisnode_zonelists(pgdat);
5662}
5663
5664#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5665/*
5666 * Return node id of node used for "local" allocations.
5667 * I.e., first node id of first zone in arg node's generic zonelist.
5668 * Used for initializing percpu 'numa_mem', which is used primarily
5669 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5670 */
5671int local_memory_node(int node)
5672{
5673 struct zoneref *z;
5674
5675 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5676 gfp_zone(GFP_KERNEL),
5677 NULL);
5678 return zone_to_nid(z->zone);
5679}
5680#endif
5681
5682static void setup_min_unmapped_ratio(void);
5683static void setup_min_slab_ratio(void);
5684#else /* CONFIG_NUMA */
5685
5686static void build_zonelists(pg_data_t *pgdat)
5687{
5688 int node, local_node;
5689 struct zoneref *zonerefs;
5690 int nr_zones;
5691
5692 local_node = pgdat->node_id;
5693
5694 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5695 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5696 zonerefs += nr_zones;
5697
5698 /*
5699 * Now we build the zonelist so that it contains the zones
5700 * of all the other nodes.
5701 * We don't want to pressure a particular node, so when
5702 * building the zones for node N, we make sure that the
5703 * zones coming right after the local ones are those from
5704 * node N+1 (modulo N)
5705 */
5706 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5707 if (!node_online(node))
5708 continue;
5709 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5710 zonerefs += nr_zones;
5711 }
5712 for (node = 0; node < local_node; node++) {
5713 if (!node_online(node))
5714 continue;
5715 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5716 zonerefs += nr_zones;
5717 }
5718
5719 zonerefs->zone = NULL;
5720 zonerefs->zone_idx = 0;
5721}
5722
5723#endif /* CONFIG_NUMA */
5724
5725/*
5726 * Boot pageset table. One per cpu which is going to be used for all
5727 * zones and all nodes. The parameters will be set in such a way
5728 * that an item put on a list will immediately be handed over to
5729 * the buddy list. This is safe since pageset manipulation is done
5730 * with interrupts disabled.
5731 *
5732 * The boot_pagesets must be kept even after bootup is complete for
5733 * unused processors and/or zones. They do play a role for bootstrapping
5734 * hotplugged processors.
5735 *
5736 * zoneinfo_show() and maybe other functions do
5737 * not check if the processor is online before following the pageset pointer.
5738 * Other parts of the kernel may not check if the zone is available.
5739 */
5740static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5741static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5742static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5743
5744static void __build_all_zonelists(void *data)
5745{
5746 int nid;
5747 int __maybe_unused cpu;
5748 pg_data_t *self = data;
5749 static DEFINE_SPINLOCK(lock);
5750
5751 spin_lock(&lock);
5752
5753#ifdef CONFIG_NUMA
5754 memset(node_load, 0, sizeof(node_load));
5755#endif
5756
5757 /*
5758 * This node is hotadded and no memory is yet present. So just
5759 * building zonelists is fine - no need to touch other nodes.
5760 */
5761 if (self && !node_online(self->node_id)) {
5762 build_zonelists(self);
5763 } else {
5764 for_each_online_node(nid) {
5765 pg_data_t *pgdat = NODE_DATA(nid);
5766
5767 build_zonelists(pgdat);
5768 }
5769
5770#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5771 /*
5772 * We now know the "local memory node" for each node--
5773 * i.e., the node of the first zone in the generic zonelist.
5774 * Set up numa_mem percpu variable for on-line cpus. During
5775 * boot, only the boot cpu should be on-line; we'll init the
5776 * secondary cpus' numa_mem as they come on-line. During
5777 * node/memory hotplug, we'll fixup all on-line cpus.
5778 */
5779 for_each_online_cpu(cpu)
5780 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5781#endif
5782 }
5783
5784 spin_unlock(&lock);
5785}
5786
5787static noinline void __init
5788build_all_zonelists_init(void)
5789{
5790 int cpu;
5791
5792 __build_all_zonelists(NULL);
5793
5794 /*
5795 * Initialize the boot_pagesets that are going to be used
5796 * for bootstrapping processors. The real pagesets for
5797 * each zone will be allocated later when the per cpu
5798 * allocator is available.
5799 *
5800 * boot_pagesets are used also for bootstrapping offline
5801 * cpus if the system is already booted because the pagesets
5802 * are needed to initialize allocators on a specific cpu too.
5803 * F.e. the percpu allocator needs the page allocator which
5804 * needs the percpu allocator in order to allocate its pagesets
5805 * (a chicken-egg dilemma).
5806 */
5807 for_each_possible_cpu(cpu)
5808 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5809
5810 mminit_verify_zonelist();
5811 cpuset_init_current_mems_allowed();
5812}
5813
5814/*
5815 * unless system_state == SYSTEM_BOOTING.
5816 *
5817 * __ref due to call of __init annotated helper build_all_zonelists_init
5818 * [protected by SYSTEM_BOOTING].
5819 */
5820void __ref build_all_zonelists(pg_data_t *pgdat)
5821{
5822 if (system_state == SYSTEM_BOOTING) {
5823 build_all_zonelists_init();
5824 } else {
5825 __build_all_zonelists(pgdat);
5826 /* cpuset refresh routine should be here */
5827 }
5828 vm_total_pages = nr_free_pagecache_pages();
5829 /*
5830 * Disable grouping by mobility if the number of pages in the
5831 * system is too low to allow the mechanism to work. It would be
5832 * more accurate, but expensive to check per-zone. This check is
5833 * made on memory-hotadd so a system can start with mobility
5834 * disabled and enable it later
5835 */
5836 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5837 page_group_by_mobility_disabled = 1;
5838 else
5839 page_group_by_mobility_disabled = 0;
5840
5841 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5842 nr_online_nodes,
5843 page_group_by_mobility_disabled ? "off" : "on",
5844 vm_total_pages);
5845#ifdef CONFIG_NUMA
5846 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5847#endif
5848}
5849
5850/* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5851static bool __meminit
5852overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5853{
5854#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5855 static struct memblock_region *r;
5856
5857 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5858 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5859 for_each_memblock(memory, r) {
5860 if (*pfn < memblock_region_memory_end_pfn(r))
5861 break;
5862 }
5863 }
5864 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5865 memblock_is_mirror(r)) {
5866 *pfn = memblock_region_memory_end_pfn(r);
5867 return true;
5868 }
5869 }
5870#endif
5871 return false;
5872}
5873
5874/*
5875 * Initially all pages are reserved - free ones are freed
5876 * up by memblock_free_all() once the early boot process is
5877 * done. Non-atomic initialization, single-pass.
5878 */
5879void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5880 unsigned long start_pfn, enum memmap_context context,
5881 struct vmem_altmap *altmap)
5882{
5883 unsigned long pfn, end_pfn = start_pfn + size;
5884 struct page *page;
5885
5886 if (highest_memmap_pfn < end_pfn - 1)
5887 highest_memmap_pfn = end_pfn - 1;
5888
5889#ifdef CONFIG_ZONE_DEVICE
5890 /*
5891 * Honor reservation requested by the driver for this ZONE_DEVICE
5892 * memory. We limit the total number of pages to initialize to just
5893 * those that might contain the memory mapping. We will defer the
5894 * ZONE_DEVICE page initialization until after we have released
5895 * the hotplug lock.
5896 */
5897 if (zone == ZONE_DEVICE) {
5898 if (!altmap)
5899 return;
5900
5901 if (start_pfn == altmap->base_pfn)
5902 start_pfn += altmap->reserve;
5903 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5904 }
5905#endif
5906
5907 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5908 /*
5909 * There can be holes in boot-time mem_map[]s handed to this
5910 * function. They do not exist on hotplugged memory.
5911 */
5912 if (context == MEMMAP_EARLY) {
5913 if (!early_pfn_valid(pfn))
5914 continue;
5915 if (!early_pfn_in_nid(pfn, nid))
5916 continue;
5917 if (overlap_memmap_init(zone, &pfn))
5918 continue;
5919 if (defer_init(nid, pfn, end_pfn))
5920 break;
5921 }
5922
5923 page = pfn_to_page(pfn);
5924 __init_single_page(page, pfn, zone, nid);
5925 if (context == MEMMAP_HOTPLUG)
5926 __SetPageReserved(page);
5927
5928 /*
5929 * Mark the block movable so that blocks are reserved for
5930 * movable at startup. This will force kernel allocations
5931 * to reserve their blocks rather than leaking throughout
5932 * the address space during boot when many long-lived
5933 * kernel allocations are made.
5934 *
5935 * bitmap is created for zone's valid pfn range. but memmap
5936 * can be created for invalid pages (for alignment)
5937 * check here not to call set_pageblock_migratetype() against
5938 * pfn out of zone.
5939 */
5940 if (!(pfn & (pageblock_nr_pages - 1))) {
5941 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5942 cond_resched();
5943 }
5944 }
5945}
5946
5947#ifdef CONFIG_ZONE_DEVICE
5948void __ref memmap_init_zone_device(struct zone *zone,
5949 unsigned long start_pfn,
5950 unsigned long size,
5951 struct dev_pagemap *pgmap)
5952{
5953 unsigned long pfn, end_pfn = start_pfn + size;
5954 struct pglist_data *pgdat = zone->zone_pgdat;
5955 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
5956 unsigned long zone_idx = zone_idx(zone);
5957 unsigned long start = jiffies;
5958 int nid = pgdat->node_id;
5959
5960 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
5961 return;
5962
5963 /*
5964 * The call to memmap_init_zone should have already taken care
5965 * of the pages reserved for the memmap, so we can just jump to
5966 * the end of that region and start processing the device pages.
5967 */
5968 if (altmap) {
5969 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5970 size = end_pfn - start_pfn;
5971 }
5972
5973 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5974 struct page *page = pfn_to_page(pfn);
5975
5976 __init_single_page(page, pfn, zone_idx, nid);
5977
5978 /*
5979 * Mark page reserved as it will need to wait for onlining
5980 * phase for it to be fully associated with a zone.
5981 *
5982 * We can use the non-atomic __set_bit operation for setting
5983 * the flag as we are still initializing the pages.
5984 */
5985 __SetPageReserved(page);
5986
5987 /*
5988 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5989 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5990 * ever freed or placed on a driver-private list.
5991 */
5992 page->pgmap = pgmap;
5993 page->zone_device_data = NULL;
5994
5995 /*
5996 * Mark the block movable so that blocks are reserved for
5997 * movable at startup. This will force kernel allocations
5998 * to reserve their blocks rather than leaking throughout
5999 * the address space during boot when many long-lived
6000 * kernel allocations are made.
6001 *
6002 * bitmap is created for zone's valid pfn range. but memmap
6003 * can be created for invalid pages (for alignment)
6004 * check here not to call set_pageblock_migratetype() against
6005 * pfn out of zone.
6006 *
6007 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6008 * because this is done early in section_activate()
6009 */
6010 if (!(pfn & (pageblock_nr_pages - 1))) {
6011 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6012 cond_resched();
6013 }
6014 }
6015
6016 pr_info("%s initialised %lu pages in %ums\n", __func__,
6017 size, jiffies_to_msecs(jiffies - start));
6018}
6019
6020#endif
6021static void __meminit zone_init_free_lists(struct zone *zone)
6022{
6023 unsigned int order, t;
6024 for_each_migratetype_order(order, t) {
6025 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6026 zone->free_area[order].nr_free = 0;
6027 }
6028}
6029
6030void __meminit __weak memmap_init(unsigned long size, int nid,
6031 unsigned long zone, unsigned long start_pfn)
6032{
6033 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6034}
6035
6036static int zone_batchsize(struct zone *zone)
6037{
6038#ifdef CONFIG_MMU
6039 int batch;
6040
6041 /*
6042 * The per-cpu-pages pools are set to around 1000th of the
6043 * size of the zone.
6044 */
6045 batch = zone_managed_pages(zone) / 1024;
6046 /* But no more than a meg. */
6047 if (batch * PAGE_SIZE > 1024 * 1024)
6048 batch = (1024 * 1024) / PAGE_SIZE;
6049 batch /= 4; /* We effectively *= 4 below */
6050 if (batch < 1)
6051 batch = 1;
6052
6053 /*
6054 * Clamp the batch to a 2^n - 1 value. Having a power
6055 * of 2 value was found to be more likely to have
6056 * suboptimal cache aliasing properties in some cases.
6057 *
6058 * For example if 2 tasks are alternately allocating
6059 * batches of pages, one task can end up with a lot
6060 * of pages of one half of the possible page colors
6061 * and the other with pages of the other colors.
6062 */
6063 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6064
6065 return batch;
6066
6067#else
6068 /* The deferral and batching of frees should be suppressed under NOMMU
6069 * conditions.
6070 *
6071 * The problem is that NOMMU needs to be able to allocate large chunks
6072 * of contiguous memory as there's no hardware page translation to
6073 * assemble apparent contiguous memory from discontiguous pages.
6074 *
6075 * Queueing large contiguous runs of pages for batching, however,
6076 * causes the pages to actually be freed in smaller chunks. As there
6077 * can be a significant delay between the individual batches being
6078 * recycled, this leads to the once large chunks of space being
6079 * fragmented and becoming unavailable for high-order allocations.
6080 */
6081 return 0;
6082#endif
6083}
6084
6085/*
6086 * pcp->high and pcp->batch values are related and dependent on one another:
6087 * ->batch must never be higher then ->high.
6088 * The following function updates them in a safe manner without read side
6089 * locking.
6090 *
6091 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6092 * those fields changing asynchronously (acording the the above rule).
6093 *
6094 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6095 * outside of boot time (or some other assurance that no concurrent updaters
6096 * exist).
6097 */
6098static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6099 unsigned long batch)
6100{
6101 /* start with a fail safe value for batch */
6102 pcp->batch = 1;
6103 smp_wmb();
6104
6105 /* Update high, then batch, in order */
6106 pcp->high = high;
6107 smp_wmb();
6108
6109 pcp->batch = batch;
6110}
6111
6112/* a companion to pageset_set_high() */
6113static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6114{
6115 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6116}
6117
6118static void pageset_init(struct per_cpu_pageset *p)
6119{
6120 struct per_cpu_pages *pcp;
6121 int migratetype;
6122
6123 memset(p, 0, sizeof(*p));
6124
6125 pcp = &p->pcp;
6126 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6127 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6128}
6129
6130static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6131{
6132 pageset_init(p);
6133 pageset_set_batch(p, batch);
6134}
6135
6136/*
6137 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6138 * to the value high for the pageset p.
6139 */
6140static void pageset_set_high(struct per_cpu_pageset *p,
6141 unsigned long high)
6142{
6143 unsigned long batch = max(1UL, high / 4);
6144 if ((high / 4) > (PAGE_SHIFT * 8))
6145 batch = PAGE_SHIFT * 8;
6146
6147 pageset_update(&p->pcp, high, batch);
6148}
6149
6150static void pageset_set_high_and_batch(struct zone *zone,
6151 struct per_cpu_pageset *pcp)
6152{
6153 if (percpu_pagelist_fraction)
6154 pageset_set_high(pcp,
6155 (zone_managed_pages(zone) /
6156 percpu_pagelist_fraction));
6157 else
6158 pageset_set_batch(pcp, zone_batchsize(zone));
6159}
6160
6161static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6162{
6163 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6164
6165 pageset_init(pcp);
6166 pageset_set_high_and_batch(zone, pcp);
6167}
6168
6169void __meminit setup_zone_pageset(struct zone *zone)
6170{
6171 int cpu;
6172 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6173 for_each_possible_cpu(cpu)
6174 zone_pageset_init(zone, cpu);
6175}
6176
6177/*
6178 * Allocate per cpu pagesets and initialize them.
6179 * Before this call only boot pagesets were available.
6180 */
6181void __init setup_per_cpu_pageset(void)
6182{
6183 struct pglist_data *pgdat;
6184 struct zone *zone;
6185
6186 for_each_populated_zone(zone)
6187 setup_zone_pageset(zone);
6188
6189 for_each_online_pgdat(pgdat)
6190 pgdat->per_cpu_nodestats =
6191 alloc_percpu(struct per_cpu_nodestat);
6192}
6193
6194static __meminit void zone_pcp_init(struct zone *zone)
6195{
6196 /*
6197 * per cpu subsystem is not up at this point. The following code
6198 * relies on the ability of the linker to provide the
6199 * offset of a (static) per cpu variable into the per cpu area.
6200 */
6201 zone->pageset = &boot_pageset;
6202
6203 if (populated_zone(zone))
6204 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6205 zone->name, zone->present_pages,
6206 zone_batchsize(zone));
6207}
6208
6209void __meminit init_currently_empty_zone(struct zone *zone,
6210 unsigned long zone_start_pfn,
6211 unsigned long size)
6212{
6213 struct pglist_data *pgdat = zone->zone_pgdat;
6214 int zone_idx = zone_idx(zone) + 1;
6215
6216 if (zone_idx > pgdat->nr_zones)
6217 pgdat->nr_zones = zone_idx;
6218
6219 zone->zone_start_pfn = zone_start_pfn;
6220
6221 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6222 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6223 pgdat->node_id,
6224 (unsigned long)zone_idx(zone),
6225 zone_start_pfn, (zone_start_pfn + size));
6226
6227 zone_init_free_lists(zone);
6228 zone->initialized = 1;
6229}
6230
6231#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6232#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6233
6234/*
6235 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6236 */
6237int __meminit __early_pfn_to_nid(unsigned long pfn,
6238 struct mminit_pfnnid_cache *state)
6239{
6240 unsigned long start_pfn, end_pfn;
6241 int nid;
6242
6243 if (state->last_start <= pfn && pfn < state->last_end)
6244 return state->last_nid;
6245
6246 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6247 if (nid != NUMA_NO_NODE) {
6248 state->last_start = start_pfn;
6249 state->last_end = end_pfn;
6250 state->last_nid = nid;
6251 }
6252
6253 return nid;
6254}
6255#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6256
6257/**
6258 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6259 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6260 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6261 *
6262 * If an architecture guarantees that all ranges registered contain no holes
6263 * and may be freed, this this function may be used instead of calling
6264 * memblock_free_early_nid() manually.
6265 */
6266void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6267{
6268 unsigned long start_pfn, end_pfn;
6269 int i, this_nid;
6270
6271 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6272 start_pfn = min(start_pfn, max_low_pfn);
6273 end_pfn = min(end_pfn, max_low_pfn);
6274
6275 if (start_pfn < end_pfn)
6276 memblock_free_early_nid(PFN_PHYS(start_pfn),
6277 (end_pfn - start_pfn) << PAGE_SHIFT,
6278 this_nid);
6279 }
6280}
6281
6282/**
6283 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6284 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6285 *
6286 * If an architecture guarantees that all ranges registered contain no holes and may
6287 * be freed, this function may be used instead of calling memory_present() manually.
6288 */
6289void __init sparse_memory_present_with_active_regions(int nid)
6290{
6291 unsigned long start_pfn, end_pfn;
6292 int i, this_nid;
6293
6294 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6295 memory_present(this_nid, start_pfn, end_pfn);
6296}
6297
6298/**
6299 * get_pfn_range_for_nid - Return the start and end page frames for a node
6300 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6301 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6302 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6303 *
6304 * It returns the start and end page frame of a node based on information
6305 * provided by memblock_set_node(). If called for a node
6306 * with no available memory, a warning is printed and the start and end
6307 * PFNs will be 0.
6308 */
6309void __init get_pfn_range_for_nid(unsigned int nid,
6310 unsigned long *start_pfn, unsigned long *end_pfn)
6311{
6312 unsigned long this_start_pfn, this_end_pfn;
6313 int i;
6314
6315 *start_pfn = -1UL;
6316 *end_pfn = 0;
6317
6318 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6319 *start_pfn = min(*start_pfn, this_start_pfn);
6320 *end_pfn = max(*end_pfn, this_end_pfn);
6321 }
6322
6323 if (*start_pfn == -1UL)
6324 *start_pfn = 0;
6325}
6326
6327/*
6328 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6329 * assumption is made that zones within a node are ordered in monotonic
6330 * increasing memory addresses so that the "highest" populated zone is used
6331 */
6332static void __init find_usable_zone_for_movable(void)
6333{
6334 int zone_index;
6335 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6336 if (zone_index == ZONE_MOVABLE)
6337 continue;
6338
6339 if (arch_zone_highest_possible_pfn[zone_index] >
6340 arch_zone_lowest_possible_pfn[zone_index])
6341 break;
6342 }
6343
6344 VM_BUG_ON(zone_index == -1);
6345 movable_zone = zone_index;
6346}
6347
6348/*
6349 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6350 * because it is sized independent of architecture. Unlike the other zones,
6351 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6352 * in each node depending on the size of each node and how evenly kernelcore
6353 * is distributed. This helper function adjusts the zone ranges
6354 * provided by the architecture for a given node by using the end of the
6355 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6356 * zones within a node are in order of monotonic increases memory addresses
6357 */
6358static void __init adjust_zone_range_for_zone_movable(int nid,
6359 unsigned long zone_type,
6360 unsigned long node_start_pfn,
6361 unsigned long node_end_pfn,
6362 unsigned long *zone_start_pfn,
6363 unsigned long *zone_end_pfn)
6364{
6365 /* Only adjust if ZONE_MOVABLE is on this node */
6366 if (zone_movable_pfn[nid]) {
6367 /* Size ZONE_MOVABLE */
6368 if (zone_type == ZONE_MOVABLE) {
6369 *zone_start_pfn = zone_movable_pfn[nid];
6370 *zone_end_pfn = min(node_end_pfn,
6371 arch_zone_highest_possible_pfn[movable_zone]);
6372
6373 /* Adjust for ZONE_MOVABLE starting within this range */
6374 } else if (!mirrored_kernelcore &&
6375 *zone_start_pfn < zone_movable_pfn[nid] &&
6376 *zone_end_pfn > zone_movable_pfn[nid]) {
6377 *zone_end_pfn = zone_movable_pfn[nid];
6378
6379 /* Check if this whole range is within ZONE_MOVABLE */
6380 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6381 *zone_start_pfn = *zone_end_pfn;
6382 }
6383}
6384
6385/*
6386 * Return the number of pages a zone spans in a node, including holes
6387 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6388 */
6389static unsigned long __init zone_spanned_pages_in_node(int nid,
6390 unsigned long zone_type,
6391 unsigned long node_start_pfn,
6392 unsigned long node_end_pfn,
6393 unsigned long *zone_start_pfn,
6394 unsigned long *zone_end_pfn,
6395 unsigned long *ignored)
6396{
6397 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6398 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6399 /* When hotadd a new node from cpu_up(), the node should be empty */
6400 if (!node_start_pfn && !node_end_pfn)
6401 return 0;
6402
6403 /* Get the start and end of the zone */
6404 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6405 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6406 adjust_zone_range_for_zone_movable(nid, zone_type,
6407 node_start_pfn, node_end_pfn,
6408 zone_start_pfn, zone_end_pfn);
6409
6410 /* Check that this node has pages within the zone's required range */
6411 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6412 return 0;
6413
6414 /* Move the zone boundaries inside the node if necessary */
6415 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6416 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6417
6418 /* Return the spanned pages */
6419 return *zone_end_pfn - *zone_start_pfn;
6420}
6421
6422/*
6423 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6424 * then all holes in the requested range will be accounted for.
6425 */
6426unsigned long __init __absent_pages_in_range(int nid,
6427 unsigned long range_start_pfn,
6428 unsigned long range_end_pfn)
6429{
6430 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6431 unsigned long start_pfn, end_pfn;
6432 int i;
6433
6434 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6435 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6436 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6437 nr_absent -= end_pfn - start_pfn;
6438 }
6439 return nr_absent;
6440}
6441
6442/**
6443 * absent_pages_in_range - Return number of page frames in holes within a range
6444 * @start_pfn: The start PFN to start searching for holes
6445 * @end_pfn: The end PFN to stop searching for holes
6446 *
6447 * Return: the number of pages frames in memory holes within a range.
6448 */
6449unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6450 unsigned long end_pfn)
6451{
6452 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6453}
6454
6455/* Return the number of page frames in holes in a zone on a node */
6456static unsigned long __init zone_absent_pages_in_node(int nid,
6457 unsigned long zone_type,
6458 unsigned long node_start_pfn,
6459 unsigned long node_end_pfn,
6460 unsigned long *ignored)
6461{
6462 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6463 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6464 unsigned long zone_start_pfn, zone_end_pfn;
6465 unsigned long nr_absent;
6466
6467 /* When hotadd a new node from cpu_up(), the node should be empty */
6468 if (!node_start_pfn && !node_end_pfn)
6469 return 0;
6470
6471 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6472 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6473
6474 adjust_zone_range_for_zone_movable(nid, zone_type,
6475 node_start_pfn, node_end_pfn,
6476 &zone_start_pfn, &zone_end_pfn);
6477 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6478
6479 /*
6480 * ZONE_MOVABLE handling.
6481 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6482 * and vice versa.
6483 */
6484 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6485 unsigned long start_pfn, end_pfn;
6486 struct memblock_region *r;
6487
6488 for_each_memblock(memory, r) {
6489 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6490 zone_start_pfn, zone_end_pfn);
6491 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6492 zone_start_pfn, zone_end_pfn);
6493
6494 if (zone_type == ZONE_MOVABLE &&
6495 memblock_is_mirror(r))
6496 nr_absent += end_pfn - start_pfn;
6497
6498 if (zone_type == ZONE_NORMAL &&
6499 !memblock_is_mirror(r))
6500 nr_absent += end_pfn - start_pfn;
6501 }
6502 }
6503
6504 return nr_absent;
6505}
6506
6507#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6508static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6509 unsigned long zone_type,
6510 unsigned long node_start_pfn,
6511 unsigned long node_end_pfn,
6512 unsigned long *zone_start_pfn,
6513 unsigned long *zone_end_pfn,
6514 unsigned long *zones_size)
6515{
6516 unsigned int zone;
6517
6518 *zone_start_pfn = node_start_pfn;
6519 for (zone = 0; zone < zone_type; zone++)
6520 *zone_start_pfn += zones_size[zone];
6521
6522 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6523
6524 return zones_size[zone_type];
6525}
6526
6527static inline unsigned long __init zone_absent_pages_in_node(int nid,
6528 unsigned long zone_type,
6529 unsigned long node_start_pfn,
6530 unsigned long node_end_pfn,
6531 unsigned long *zholes_size)
6532{
6533 if (!zholes_size)
6534 return 0;
6535
6536 return zholes_size[zone_type];
6537}
6538
6539#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6540
6541static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6542 unsigned long node_start_pfn,
6543 unsigned long node_end_pfn,
6544 unsigned long *zones_size,
6545 unsigned long *zholes_size)
6546{
6547 unsigned long realtotalpages = 0, totalpages = 0;
6548 enum zone_type i;
6549
6550 for (i = 0; i < MAX_NR_ZONES; i++) {
6551 struct zone *zone = pgdat->node_zones + i;
6552 unsigned long zone_start_pfn, zone_end_pfn;
6553 unsigned long size, real_size;
6554
6555 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6556 node_start_pfn,
6557 node_end_pfn,
6558 &zone_start_pfn,
6559 &zone_end_pfn,
6560 zones_size);
6561 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6562 node_start_pfn, node_end_pfn,
6563 zholes_size);
6564 if (size)
6565 zone->zone_start_pfn = zone_start_pfn;
6566 else
6567 zone->zone_start_pfn = 0;
6568 zone->spanned_pages = size;
6569 zone->present_pages = real_size;
6570
6571 totalpages += size;
6572 realtotalpages += real_size;
6573 }
6574
6575 pgdat->node_spanned_pages = totalpages;
6576 pgdat->node_present_pages = realtotalpages;
6577 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6578 realtotalpages);
6579}
6580
6581#ifndef CONFIG_SPARSEMEM
6582/*
6583 * Calculate the size of the zone->blockflags rounded to an unsigned long
6584 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6585 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6586 * round what is now in bits to nearest long in bits, then return it in
6587 * bytes.
6588 */
6589static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6590{
6591 unsigned long usemapsize;
6592
6593 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6594 usemapsize = roundup(zonesize, pageblock_nr_pages);
6595 usemapsize = usemapsize >> pageblock_order;
6596 usemapsize *= NR_PAGEBLOCK_BITS;
6597 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6598
6599 return usemapsize / 8;
6600}
6601
6602static void __ref setup_usemap(struct pglist_data *pgdat,
6603 struct zone *zone,
6604 unsigned long zone_start_pfn,
6605 unsigned long zonesize)
6606{
6607 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6608 zone->pageblock_flags = NULL;
6609 if (usemapsize) {
6610 zone->pageblock_flags =
6611 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6612 pgdat->node_id);
6613 if (!zone->pageblock_flags)
6614 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6615 usemapsize, zone->name, pgdat->node_id);
6616 }
6617}
6618#else
6619static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6620 unsigned long zone_start_pfn, unsigned long zonesize) {}
6621#endif /* CONFIG_SPARSEMEM */
6622
6623#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6624
6625/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6626void __init set_pageblock_order(void)
6627{
6628 unsigned int order;
6629
6630 /* Check that pageblock_nr_pages has not already been setup */
6631 if (pageblock_order)
6632 return;
6633
6634 if (HPAGE_SHIFT > PAGE_SHIFT)
6635 order = HUGETLB_PAGE_ORDER;
6636 else
6637 order = MAX_ORDER - 1;
6638
6639 /*
6640 * Assume the largest contiguous order of interest is a huge page.
6641 * This value may be variable depending on boot parameters on IA64 and
6642 * powerpc.
6643 */
6644 pageblock_order = order;
6645}
6646#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6647
6648/*
6649 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6650 * is unused as pageblock_order is set at compile-time. See
6651 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6652 * the kernel config
6653 */
6654void __init set_pageblock_order(void)
6655{
6656}
6657
6658#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6659
6660static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6661 unsigned long present_pages)
6662{
6663 unsigned long pages = spanned_pages;
6664
6665 /*
6666 * Provide a more accurate estimation if there are holes within
6667 * the zone and SPARSEMEM is in use. If there are holes within the
6668 * zone, each populated memory region may cost us one or two extra
6669 * memmap pages due to alignment because memmap pages for each
6670 * populated regions may not be naturally aligned on page boundary.
6671 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6672 */
6673 if (spanned_pages > present_pages + (present_pages >> 4) &&
6674 IS_ENABLED(CONFIG_SPARSEMEM))
6675 pages = present_pages;
6676
6677 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6678}
6679
6680#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6681static void pgdat_init_split_queue(struct pglist_data *pgdat)
6682{
6683 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6684
6685 spin_lock_init(&ds_queue->split_queue_lock);
6686 INIT_LIST_HEAD(&ds_queue->split_queue);
6687 ds_queue->split_queue_len = 0;
6688}
6689#else
6690static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6691#endif
6692
6693#ifdef CONFIG_COMPACTION
6694static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6695{
6696 init_waitqueue_head(&pgdat->kcompactd_wait);
6697}
6698#else
6699static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6700#endif
6701
6702static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6703{
6704 pgdat_resize_init(pgdat);
6705
6706 pgdat_init_split_queue(pgdat);
6707 pgdat_init_kcompactd(pgdat);
6708
6709 init_waitqueue_head(&pgdat->kswapd_wait);
6710 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6711
6712 pgdat_page_ext_init(pgdat);
6713 spin_lock_init(&pgdat->lru_lock);
6714 lruvec_init(node_lruvec(pgdat));
6715}
6716
6717static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6718 unsigned long remaining_pages)
6719{
6720 atomic_long_set(&zone->managed_pages, remaining_pages);
6721 zone_set_nid(zone, nid);
6722 zone->name = zone_names[idx];
6723 zone->zone_pgdat = NODE_DATA(nid);
6724 spin_lock_init(&zone->lock);
6725 zone_seqlock_init(zone);
6726 zone_pcp_init(zone);
6727}
6728
6729/*
6730 * Set up the zone data structures
6731 * - init pgdat internals
6732 * - init all zones belonging to this node
6733 *
6734 * NOTE: this function is only called during memory hotplug
6735 */
6736#ifdef CONFIG_MEMORY_HOTPLUG
6737void __ref free_area_init_core_hotplug(int nid)
6738{
6739 enum zone_type z;
6740 pg_data_t *pgdat = NODE_DATA(nid);
6741
6742 pgdat_init_internals(pgdat);
6743 for (z = 0; z < MAX_NR_ZONES; z++)
6744 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6745}
6746#endif
6747
6748/*
6749 * Set up the zone data structures:
6750 * - mark all pages reserved
6751 * - mark all memory queues empty
6752 * - clear the memory bitmaps
6753 *
6754 * NOTE: pgdat should get zeroed by caller.
6755 * NOTE: this function is only called during early init.
6756 */
6757static void __init free_area_init_core(struct pglist_data *pgdat)
6758{
6759 enum zone_type j;
6760 int nid = pgdat->node_id;
6761
6762 pgdat_init_internals(pgdat);
6763 pgdat->per_cpu_nodestats = &boot_nodestats;
6764
6765 for (j = 0; j < MAX_NR_ZONES; j++) {
6766 struct zone *zone = pgdat->node_zones + j;
6767 unsigned long size, freesize, memmap_pages;
6768 unsigned long zone_start_pfn = zone->zone_start_pfn;
6769
6770 size = zone->spanned_pages;
6771 freesize = zone->present_pages;
6772
6773 /*
6774 * Adjust freesize so that it accounts for how much memory
6775 * is used by this zone for memmap. This affects the watermark
6776 * and per-cpu initialisations
6777 */
6778 memmap_pages = calc_memmap_size(size, freesize);
6779 if (!is_highmem_idx(j)) {
6780 if (freesize >= memmap_pages) {
6781 freesize -= memmap_pages;
6782 if (memmap_pages)
6783 printk(KERN_DEBUG
6784 " %s zone: %lu pages used for memmap\n",
6785 zone_names[j], memmap_pages);
6786 } else
6787 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6788 zone_names[j], memmap_pages, freesize);
6789 }
6790
6791 /* Account for reserved pages */
6792 if (j == 0 && freesize > dma_reserve) {
6793 freesize -= dma_reserve;
6794 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6795 zone_names[0], dma_reserve);
6796 }
6797
6798 if (!is_highmem_idx(j))
6799 nr_kernel_pages += freesize;
6800 /* Charge for highmem memmap if there are enough kernel pages */
6801 else if (nr_kernel_pages > memmap_pages * 2)
6802 nr_kernel_pages -= memmap_pages;
6803 nr_all_pages += freesize;
6804
6805 /*
6806 * Set an approximate value for lowmem here, it will be adjusted
6807 * when the bootmem allocator frees pages into the buddy system.
6808 * And all highmem pages will be managed by the buddy system.
6809 */
6810 zone_init_internals(zone, j, nid, freesize);
6811
6812 if (!size)
6813 continue;
6814
6815 set_pageblock_order();
6816 setup_usemap(pgdat, zone, zone_start_pfn, size);
6817 init_currently_empty_zone(zone, zone_start_pfn, size);
6818 memmap_init(size, nid, j, zone_start_pfn);
6819 }
6820}
6821
6822#ifdef CONFIG_FLAT_NODE_MEM_MAP
6823static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6824{
6825 unsigned long __maybe_unused start = 0;
6826 unsigned long __maybe_unused offset = 0;
6827
6828 /* Skip empty nodes */
6829 if (!pgdat->node_spanned_pages)
6830 return;
6831
6832 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6833 offset = pgdat->node_start_pfn - start;
6834 /* ia64 gets its own node_mem_map, before this, without bootmem */
6835 if (!pgdat->node_mem_map) {
6836 unsigned long size, end;
6837 struct page *map;
6838
6839 /*
6840 * The zone's endpoints aren't required to be MAX_ORDER
6841 * aligned but the node_mem_map endpoints must be in order
6842 * for the buddy allocator to function correctly.
6843 */
6844 end = pgdat_end_pfn(pgdat);
6845 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6846 size = (end - start) * sizeof(struct page);
6847 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6848 pgdat->node_id);
6849 if (!map)
6850 panic("Failed to allocate %ld bytes for node %d memory map\n",
6851 size, pgdat->node_id);
6852 pgdat->node_mem_map = map + offset;
6853 }
6854 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6855 __func__, pgdat->node_id, (unsigned long)pgdat,
6856 (unsigned long)pgdat->node_mem_map);
6857#ifndef CONFIG_NEED_MULTIPLE_NODES
6858 /*
6859 * With no DISCONTIG, the global mem_map is just set as node 0's
6860 */
6861 if (pgdat == NODE_DATA(0)) {
6862 mem_map = NODE_DATA(0)->node_mem_map;
6863#if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6864 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6865 mem_map -= offset;
6866#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6867 }
6868#endif
6869}
6870#else
6871static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6872#endif /* CONFIG_FLAT_NODE_MEM_MAP */
6873
6874#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6875static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6876{
6877 pgdat->first_deferred_pfn = ULONG_MAX;
6878}
6879#else
6880static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6881#endif
6882
6883void __init free_area_init_node(int nid, unsigned long *zones_size,
6884 unsigned long node_start_pfn,
6885 unsigned long *zholes_size)
6886{
6887 pg_data_t *pgdat = NODE_DATA(nid);
6888 unsigned long start_pfn = 0;
6889 unsigned long end_pfn = 0;
6890
6891 /* pg_data_t should be reset to zero when it's allocated */
6892 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6893
6894 pgdat->node_id = nid;
6895 pgdat->node_start_pfn = node_start_pfn;
6896 pgdat->per_cpu_nodestats = NULL;
6897#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6898 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6899 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6900 (u64)start_pfn << PAGE_SHIFT,
6901 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6902#else
6903 start_pfn = node_start_pfn;
6904#endif
6905 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6906 zones_size, zholes_size);
6907
6908 alloc_node_mem_map(pgdat);
6909 pgdat_set_deferred_range(pgdat);
6910
6911 free_area_init_core(pgdat);
6912}
6913
6914#if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6915/*
6916 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6917 * pages zeroed
6918 */
6919static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6920{
6921 unsigned long pfn;
6922 u64 pgcnt = 0;
6923
6924 for (pfn = spfn; pfn < epfn; pfn++) {
6925 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6926 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6927 + pageblock_nr_pages - 1;
6928 continue;
6929 }
6930 mm_zero_struct_page(pfn_to_page(pfn));
6931 pgcnt++;
6932 }
6933
6934 return pgcnt;
6935}
6936
6937/*
6938 * Only struct pages that are backed by physical memory are zeroed and
6939 * initialized by going through __init_single_page(). But, there are some
6940 * struct pages which are reserved in memblock allocator and their fields
6941 * may be accessed (for example page_to_pfn() on some configuration accesses
6942 * flags). We must explicitly zero those struct pages.
6943 *
6944 * This function also addresses a similar issue where struct pages are left
6945 * uninitialized because the physical address range is not covered by
6946 * memblock.memory or memblock.reserved. That could happen when memblock
6947 * layout is manually configured via memmap=.
6948 */
6949void __init zero_resv_unavail(void)
6950{
6951 phys_addr_t start, end;
6952 u64 i, pgcnt;
6953 phys_addr_t next = 0;
6954
6955 /*
6956 * Loop through unavailable ranges not covered by memblock.memory.
6957 */
6958 pgcnt = 0;
6959 for_each_mem_range(i, &memblock.memory, NULL,
6960 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6961 if (next < start)
6962 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6963 next = end;
6964 }
6965 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6966
6967 /*
6968 * Struct pages that do not have backing memory. This could be because
6969 * firmware is using some of this memory, or for some other reasons.
6970 */
6971 if (pgcnt)
6972 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6973}
6974#endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6975
6976#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6977
6978#if MAX_NUMNODES > 1
6979/*
6980 * Figure out the number of possible node ids.
6981 */
6982void __init setup_nr_node_ids(void)
6983{
6984 unsigned int highest;
6985
6986 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6987 nr_node_ids = highest + 1;
6988}
6989#endif
6990
6991/**
6992 * node_map_pfn_alignment - determine the maximum internode alignment
6993 *
6994 * This function should be called after node map is populated and sorted.
6995 * It calculates the maximum power of two alignment which can distinguish
6996 * all the nodes.
6997 *
6998 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6999 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7000 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7001 * shifted, 1GiB is enough and this function will indicate so.
7002 *
7003 * This is used to test whether pfn -> nid mapping of the chosen memory
7004 * model has fine enough granularity to avoid incorrect mapping for the
7005 * populated node map.
7006 *
7007 * Return: the determined alignment in pfn's. 0 if there is no alignment
7008 * requirement (single node).
7009 */
7010unsigned long __init node_map_pfn_alignment(void)
7011{
7012 unsigned long accl_mask = 0, last_end = 0;
7013 unsigned long start, end, mask;
7014 int last_nid = NUMA_NO_NODE;
7015 int i, nid;
7016
7017 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7018 if (!start || last_nid < 0 || last_nid == nid) {
7019 last_nid = nid;
7020 last_end = end;
7021 continue;
7022 }
7023
7024 /*
7025 * Start with a mask granular enough to pin-point to the
7026 * start pfn and tick off bits one-by-one until it becomes
7027 * too coarse to separate the current node from the last.
7028 */
7029 mask = ~((1 << __ffs(start)) - 1);
7030 while (mask && last_end <= (start & (mask << 1)))
7031 mask <<= 1;
7032
7033 /* accumulate all internode masks */
7034 accl_mask |= mask;
7035 }
7036
7037 /* convert mask to number of pages */
7038 return ~accl_mask + 1;
7039}
7040
7041/* Find the lowest pfn for a node */
7042static unsigned long __init find_min_pfn_for_node(int nid)
7043{
7044 unsigned long min_pfn = ULONG_MAX;
7045 unsigned long start_pfn;
7046 int i;
7047
7048 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7049 min_pfn = min(min_pfn, start_pfn);
7050
7051 if (min_pfn == ULONG_MAX) {
7052 pr_warn("Could not find start_pfn for node %d\n", nid);
7053 return 0;
7054 }
7055
7056 return min_pfn;
7057}
7058
7059/**
7060 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7061 *
7062 * Return: the minimum PFN based on information provided via
7063 * memblock_set_node().
7064 */
7065unsigned long __init find_min_pfn_with_active_regions(void)
7066{
7067 return find_min_pfn_for_node(MAX_NUMNODES);
7068}
7069
7070/*
7071 * early_calculate_totalpages()
7072 * Sum pages in active regions for movable zone.
7073 * Populate N_MEMORY for calculating usable_nodes.
7074 */
7075static unsigned long __init early_calculate_totalpages(void)
7076{
7077 unsigned long totalpages = 0;
7078 unsigned long start_pfn, end_pfn;
7079 int i, nid;
7080
7081 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7082 unsigned long pages = end_pfn - start_pfn;
7083
7084 totalpages += pages;
7085 if (pages)
7086 node_set_state(nid, N_MEMORY);
7087 }
7088 return totalpages;
7089}
7090
7091/*
7092 * Find the PFN the Movable zone begins in each node. Kernel memory
7093 * is spread evenly between nodes as long as the nodes have enough
7094 * memory. When they don't, some nodes will have more kernelcore than
7095 * others
7096 */
7097static void __init find_zone_movable_pfns_for_nodes(void)
7098{
7099 int i, nid;
7100 unsigned long usable_startpfn;
7101 unsigned long kernelcore_node, kernelcore_remaining;
7102 /* save the state before borrow the nodemask */
7103 nodemask_t saved_node_state = node_states[N_MEMORY];
7104 unsigned long totalpages = early_calculate_totalpages();
7105 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7106 struct memblock_region *r;
7107
7108 /* Need to find movable_zone earlier when movable_node is specified. */
7109 find_usable_zone_for_movable();
7110
7111 /*
7112 * If movable_node is specified, ignore kernelcore and movablecore
7113 * options.
7114 */
7115 if (movable_node_is_enabled()) {
7116 for_each_memblock(memory, r) {
7117 if (!memblock_is_hotpluggable(r))
7118 continue;
7119
7120 nid = r->nid;
7121
7122 usable_startpfn = PFN_DOWN(r->base);
7123 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7124 min(usable_startpfn, zone_movable_pfn[nid]) :
7125 usable_startpfn;
7126 }
7127
7128 goto out2;
7129 }
7130
7131 /*
7132 * If kernelcore=mirror is specified, ignore movablecore option
7133 */
7134 if (mirrored_kernelcore) {
7135 bool mem_below_4gb_not_mirrored = false;
7136
7137 for_each_memblock(memory, r) {
7138 if (memblock_is_mirror(r))
7139 continue;
7140
7141 nid = r->nid;
7142
7143 usable_startpfn = memblock_region_memory_base_pfn(r);
7144
7145 if (usable_startpfn < 0x100000) {
7146 mem_below_4gb_not_mirrored = true;
7147 continue;
7148 }
7149
7150 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7151 min(usable_startpfn, zone_movable_pfn[nid]) :
7152 usable_startpfn;
7153 }
7154
7155 if (mem_below_4gb_not_mirrored)
7156 pr_warn("This configuration results in unmirrored kernel memory.");
7157
7158 goto out2;
7159 }
7160
7161 /*
7162 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7163 * amount of necessary memory.
7164 */
7165 if (required_kernelcore_percent)
7166 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7167 10000UL;
7168 if (required_movablecore_percent)
7169 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7170 10000UL;
7171
7172 /*
7173 * If movablecore= was specified, calculate what size of
7174 * kernelcore that corresponds so that memory usable for
7175 * any allocation type is evenly spread. If both kernelcore
7176 * and movablecore are specified, then the value of kernelcore
7177 * will be used for required_kernelcore if it's greater than
7178 * what movablecore would have allowed.
7179 */
7180 if (required_movablecore) {
7181 unsigned long corepages;
7182
7183 /*
7184 * Round-up so that ZONE_MOVABLE is at least as large as what
7185 * was requested by the user
7186 */
7187 required_movablecore =
7188 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7189 required_movablecore = min(totalpages, required_movablecore);
7190 corepages = totalpages - required_movablecore;
7191
7192 required_kernelcore = max(required_kernelcore, corepages);
7193 }
7194
7195 /*
7196 * If kernelcore was not specified or kernelcore size is larger
7197 * than totalpages, there is no ZONE_MOVABLE.
7198 */
7199 if (!required_kernelcore || required_kernelcore >= totalpages)
7200 goto out;
7201
7202 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7203 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7204
7205restart:
7206 /* Spread kernelcore memory as evenly as possible throughout nodes */
7207 kernelcore_node = required_kernelcore / usable_nodes;
7208 for_each_node_state(nid, N_MEMORY) {
7209 unsigned long start_pfn, end_pfn;
7210
7211 /*
7212 * Recalculate kernelcore_node if the division per node
7213 * now exceeds what is necessary to satisfy the requested
7214 * amount of memory for the kernel
7215 */
7216 if (required_kernelcore < kernelcore_node)
7217 kernelcore_node = required_kernelcore / usable_nodes;
7218
7219 /*
7220 * As the map is walked, we track how much memory is usable
7221 * by the kernel using kernelcore_remaining. When it is
7222 * 0, the rest of the node is usable by ZONE_MOVABLE
7223 */
7224 kernelcore_remaining = kernelcore_node;
7225
7226 /* Go through each range of PFNs within this node */
7227 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7228 unsigned long size_pages;
7229
7230 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7231 if (start_pfn >= end_pfn)
7232 continue;
7233
7234 /* Account for what is only usable for kernelcore */
7235 if (start_pfn < usable_startpfn) {
7236 unsigned long kernel_pages;
7237 kernel_pages = min(end_pfn, usable_startpfn)
7238 - start_pfn;
7239
7240 kernelcore_remaining -= min(kernel_pages,
7241 kernelcore_remaining);
7242 required_kernelcore -= min(kernel_pages,
7243 required_kernelcore);
7244
7245 /* Continue if range is now fully accounted */
7246 if (end_pfn <= usable_startpfn) {
7247
7248 /*
7249 * Push zone_movable_pfn to the end so
7250 * that if we have to rebalance
7251 * kernelcore across nodes, we will
7252 * not double account here
7253 */
7254 zone_movable_pfn[nid] = end_pfn;
7255 continue;
7256 }
7257 start_pfn = usable_startpfn;
7258 }
7259
7260 /*
7261 * The usable PFN range for ZONE_MOVABLE is from
7262 * start_pfn->end_pfn. Calculate size_pages as the
7263 * number of pages used as kernelcore
7264 */
7265 size_pages = end_pfn - start_pfn;
7266 if (size_pages > kernelcore_remaining)
7267 size_pages = kernelcore_remaining;
7268 zone_movable_pfn[nid] = start_pfn + size_pages;
7269
7270 /*
7271 * Some kernelcore has been met, update counts and
7272 * break if the kernelcore for this node has been
7273 * satisfied
7274 */
7275 required_kernelcore -= min(required_kernelcore,
7276 size_pages);
7277 kernelcore_remaining -= size_pages;
7278 if (!kernelcore_remaining)
7279 break;
7280 }
7281 }
7282
7283 /*
7284 * If there is still required_kernelcore, we do another pass with one
7285 * less node in the count. This will push zone_movable_pfn[nid] further
7286 * along on the nodes that still have memory until kernelcore is
7287 * satisfied
7288 */
7289 usable_nodes--;
7290 if (usable_nodes && required_kernelcore > usable_nodes)
7291 goto restart;
7292
7293out2:
7294 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7295 for (nid = 0; nid < MAX_NUMNODES; nid++)
7296 zone_movable_pfn[nid] =
7297 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7298
7299out:
7300 /* restore the node_state */
7301 node_states[N_MEMORY] = saved_node_state;
7302}
7303
7304/* Any regular or high memory on that node ? */
7305static void check_for_memory(pg_data_t *pgdat, int nid)
7306{
7307 enum zone_type zone_type;
7308
7309 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7310 struct zone *zone = &pgdat->node_zones[zone_type];
7311 if (populated_zone(zone)) {
7312 if (IS_ENABLED(CONFIG_HIGHMEM))
7313 node_set_state(nid, N_HIGH_MEMORY);
7314 if (zone_type <= ZONE_NORMAL)
7315 node_set_state(nid, N_NORMAL_MEMORY);
7316 break;
7317 }
7318 }
7319}
7320
7321/**
7322 * free_area_init_nodes - Initialise all pg_data_t and zone data
7323 * @max_zone_pfn: an array of max PFNs for each zone
7324 *
7325 * This will call free_area_init_node() for each active node in the system.
7326 * Using the page ranges provided by memblock_set_node(), the size of each
7327 * zone in each node and their holes is calculated. If the maximum PFN
7328 * between two adjacent zones match, it is assumed that the zone is empty.
7329 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7330 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7331 * starts where the previous one ended. For example, ZONE_DMA32 starts
7332 * at arch_max_dma_pfn.
7333 */
7334void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7335{
7336 unsigned long start_pfn, end_pfn;
7337 int i, nid;
7338
7339 /* Record where the zone boundaries are */
7340 memset(arch_zone_lowest_possible_pfn, 0,
7341 sizeof(arch_zone_lowest_possible_pfn));
7342 memset(arch_zone_highest_possible_pfn, 0,
7343 sizeof(arch_zone_highest_possible_pfn));
7344
7345 start_pfn = find_min_pfn_with_active_regions();
7346
7347 for (i = 0; i < MAX_NR_ZONES; i++) {
7348 if (i == ZONE_MOVABLE)
7349 continue;
7350
7351 end_pfn = max(max_zone_pfn[i], start_pfn);
7352 arch_zone_lowest_possible_pfn[i] = start_pfn;
7353 arch_zone_highest_possible_pfn[i] = end_pfn;
7354
7355 start_pfn = end_pfn;
7356 }
7357
7358 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7359 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7360 find_zone_movable_pfns_for_nodes();
7361
7362 /* Print out the zone ranges */
7363 pr_info("Zone ranges:\n");
7364 for (i = 0; i < MAX_NR_ZONES; i++) {
7365 if (i == ZONE_MOVABLE)
7366 continue;
7367 pr_info(" %-8s ", zone_names[i]);
7368 if (arch_zone_lowest_possible_pfn[i] ==
7369 arch_zone_highest_possible_pfn[i])
7370 pr_cont("empty\n");
7371 else
7372 pr_cont("[mem %#018Lx-%#018Lx]\n",
7373 (u64)arch_zone_lowest_possible_pfn[i]
7374 << PAGE_SHIFT,
7375 ((u64)arch_zone_highest_possible_pfn[i]
7376 << PAGE_SHIFT) - 1);
7377 }
7378
7379 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7380 pr_info("Movable zone start for each node\n");
7381 for (i = 0; i < MAX_NUMNODES; i++) {
7382 if (zone_movable_pfn[i])
7383 pr_info(" Node %d: %#018Lx\n", i,
7384 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7385 }
7386
7387 /*
7388 * Print out the early node map, and initialize the
7389 * subsection-map relative to active online memory ranges to
7390 * enable future "sub-section" extensions of the memory map.
7391 */
7392 pr_info("Early memory node ranges\n");
7393 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7394 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7395 (u64)start_pfn << PAGE_SHIFT,
7396 ((u64)end_pfn << PAGE_SHIFT) - 1);
7397 subsection_map_init(start_pfn, end_pfn - start_pfn);
7398 }
7399
7400 /* Initialise every node */
7401 mminit_verify_pageflags_layout();
7402 setup_nr_node_ids();
7403 zero_resv_unavail();
7404 for_each_online_node(nid) {
7405 pg_data_t *pgdat = NODE_DATA(nid);
7406 free_area_init_node(nid, NULL,
7407 find_min_pfn_for_node(nid), NULL);
7408
7409 /* Any memory on that node */
7410 if (pgdat->node_present_pages)
7411 node_set_state(nid, N_MEMORY);
7412 check_for_memory(pgdat, nid);
7413 }
7414}
7415
7416static int __init cmdline_parse_core(char *p, unsigned long *core,
7417 unsigned long *percent)
7418{
7419 unsigned long long coremem;
7420 char *endptr;
7421
7422 if (!p)
7423 return -EINVAL;
7424
7425 /* Value may be a percentage of total memory, otherwise bytes */
7426 coremem = simple_strtoull(p, &endptr, 0);
7427 if (*endptr == '%') {
7428 /* Paranoid check for percent values greater than 100 */
7429 WARN_ON(coremem > 100);
7430
7431 *percent = coremem;
7432 } else {
7433 coremem = memparse(p, &p);
7434 /* Paranoid check that UL is enough for the coremem value */
7435 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7436
7437 *core = coremem >> PAGE_SHIFT;
7438 *percent = 0UL;
7439 }
7440 return 0;
7441}
7442
7443/*
7444 * kernelcore=size sets the amount of memory for use for allocations that
7445 * cannot be reclaimed or migrated.
7446 */
7447static int __init cmdline_parse_kernelcore(char *p)
7448{
7449 /* parse kernelcore=mirror */
7450 if (parse_option_str(p, "mirror")) {
7451 mirrored_kernelcore = true;
7452 return 0;
7453 }
7454
7455 return cmdline_parse_core(p, &required_kernelcore,
7456 &required_kernelcore_percent);
7457}
7458
7459/*
7460 * movablecore=size sets the amount of memory for use for allocations that
7461 * can be reclaimed or migrated.
7462 */
7463static int __init cmdline_parse_movablecore(char *p)
7464{
7465 return cmdline_parse_core(p, &required_movablecore,
7466 &required_movablecore_percent);
7467}
7468
7469early_param("kernelcore", cmdline_parse_kernelcore);
7470early_param("movablecore", cmdline_parse_movablecore);
7471
7472#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7473
7474void adjust_managed_page_count(struct page *page, long count)
7475{
7476 atomic_long_add(count, &page_zone(page)->managed_pages);
7477 totalram_pages_add(count);
7478#ifdef CONFIG_HIGHMEM
7479 if (PageHighMem(page))
7480 totalhigh_pages_add(count);
7481#endif
7482}
7483EXPORT_SYMBOL(adjust_managed_page_count);
7484
7485unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7486{
7487 void *pos;
7488 unsigned long pages = 0;
7489
7490 start = (void *)PAGE_ALIGN((unsigned long)start);
7491 end = (void *)((unsigned long)end & PAGE_MASK);
7492 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7493 struct page *page = virt_to_page(pos);
7494 void *direct_map_addr;
7495
7496 /*
7497 * 'direct_map_addr' might be different from 'pos'
7498 * because some architectures' virt_to_page()
7499 * work with aliases. Getting the direct map
7500 * address ensures that we get a _writeable_
7501 * alias for the memset().
7502 */
7503 direct_map_addr = page_address(page);
7504 if ((unsigned int)poison <= 0xFF)
7505 memset(direct_map_addr, poison, PAGE_SIZE);
7506
7507 free_reserved_page(page);
7508 }
7509
7510 if (pages && s)
7511 pr_info("Freeing %s memory: %ldK\n",
7512 s, pages << (PAGE_SHIFT - 10));
7513
7514 return pages;
7515}
7516
7517#ifdef CONFIG_HIGHMEM
7518void free_highmem_page(struct page *page)
7519{
7520 __free_reserved_page(page);
7521 totalram_pages_inc();
7522 atomic_long_inc(&page_zone(page)->managed_pages);
7523 totalhigh_pages_inc();
7524}
7525#endif
7526
7527
7528void __init mem_init_print_info(const char *str)
7529{
7530 unsigned long physpages, codesize, datasize, rosize, bss_size;
7531 unsigned long init_code_size, init_data_size;
7532
7533 physpages = get_num_physpages();
7534 codesize = _etext - _stext;
7535 datasize = _edata - _sdata;
7536 rosize = __end_rodata - __start_rodata;
7537 bss_size = __bss_stop - __bss_start;
7538 init_data_size = __init_end - __init_begin;
7539 init_code_size = _einittext - _sinittext;
7540
7541 /*
7542 * Detect special cases and adjust section sizes accordingly:
7543 * 1) .init.* may be embedded into .data sections
7544 * 2) .init.text.* may be out of [__init_begin, __init_end],
7545 * please refer to arch/tile/kernel/vmlinux.lds.S.
7546 * 3) .rodata.* may be embedded into .text or .data sections.
7547 */
7548#define adj_init_size(start, end, size, pos, adj) \
7549 do { \
7550 if (start <= pos && pos < end && size > adj) \
7551 size -= adj; \
7552 } while (0)
7553
7554 adj_init_size(__init_begin, __init_end, init_data_size,
7555 _sinittext, init_code_size);
7556 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7557 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7558 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7559 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7560
7561#undef adj_init_size
7562
7563 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7564#ifdef CONFIG_HIGHMEM
7565 ", %luK highmem"
7566#endif
7567 "%s%s)\n",
7568 nr_free_pages() << (PAGE_SHIFT - 10),
7569 physpages << (PAGE_SHIFT - 10),
7570 codesize >> 10, datasize >> 10, rosize >> 10,
7571 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7572 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7573 totalcma_pages << (PAGE_SHIFT - 10),
7574#ifdef CONFIG_HIGHMEM
7575 totalhigh_pages() << (PAGE_SHIFT - 10),
7576#endif
7577 str ? ", " : "", str ? str : "");
7578}
7579
7580/**
7581 * set_dma_reserve - set the specified number of pages reserved in the first zone
7582 * @new_dma_reserve: The number of pages to mark reserved
7583 *
7584 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7585 * In the DMA zone, a significant percentage may be consumed by kernel image
7586 * and other unfreeable allocations which can skew the watermarks badly. This
7587 * function may optionally be used to account for unfreeable pages in the
7588 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7589 * smaller per-cpu batchsize.
7590 */
7591void __init set_dma_reserve(unsigned long new_dma_reserve)
7592{
7593 dma_reserve = new_dma_reserve;
7594}
7595
7596void __init free_area_init(unsigned long *zones_size)
7597{
7598 zero_resv_unavail();
7599 free_area_init_node(0, zones_size,
7600 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7601}
7602
7603static int page_alloc_cpu_dead(unsigned int cpu)
7604{
7605
7606 lru_add_drain_cpu(cpu);
7607 drain_pages(cpu);
7608
7609 /*
7610 * Spill the event counters of the dead processor
7611 * into the current processors event counters.
7612 * This artificially elevates the count of the current
7613 * processor.
7614 */
7615 vm_events_fold_cpu(cpu);
7616
7617 /*
7618 * Zero the differential counters of the dead processor
7619 * so that the vm statistics are consistent.
7620 *
7621 * This is only okay since the processor is dead and cannot
7622 * race with what we are doing.
7623 */
7624 cpu_vm_stats_fold(cpu);
7625 return 0;
7626}
7627
7628#ifdef CONFIG_NUMA
7629int hashdist = HASHDIST_DEFAULT;
7630
7631static int __init set_hashdist(char *str)
7632{
7633 if (!str)
7634 return 0;
7635 hashdist = simple_strtoul(str, &str, 0);
7636 return 1;
7637}
7638__setup("hashdist=", set_hashdist);
7639#endif
7640
7641void __init page_alloc_init(void)
7642{
7643 int ret;
7644
7645#ifdef CONFIG_NUMA
7646 if (num_node_state(N_MEMORY) == 1)
7647 hashdist = 0;
7648#endif
7649
7650 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7651 "mm/page_alloc:dead", NULL,
7652 page_alloc_cpu_dead);
7653 WARN_ON(ret < 0);
7654}
7655
7656/*
7657 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7658 * or min_free_kbytes changes.
7659 */
7660static void calculate_totalreserve_pages(void)
7661{
7662 struct pglist_data *pgdat;
7663 unsigned long reserve_pages = 0;
7664 enum zone_type i, j;
7665
7666 for_each_online_pgdat(pgdat) {
7667
7668 pgdat->totalreserve_pages = 0;
7669
7670 for (i = 0; i < MAX_NR_ZONES; i++) {
7671 struct zone *zone = pgdat->node_zones + i;
7672 long max = 0;
7673 unsigned long managed_pages = zone_managed_pages(zone);
7674
7675 /* Find valid and maximum lowmem_reserve in the zone */
7676 for (j = i; j < MAX_NR_ZONES; j++) {
7677 if (zone->lowmem_reserve[j] > max)
7678 max = zone->lowmem_reserve[j];
7679 }
7680
7681 /* we treat the high watermark as reserved pages. */
7682 max += high_wmark_pages(zone);
7683
7684 if (max > managed_pages)
7685 max = managed_pages;
7686
7687 pgdat->totalreserve_pages += max;
7688
7689 reserve_pages += max;
7690 }
7691 }
7692 totalreserve_pages = reserve_pages;
7693}
7694
7695/*
7696 * setup_per_zone_lowmem_reserve - called whenever
7697 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7698 * has a correct pages reserved value, so an adequate number of
7699 * pages are left in the zone after a successful __alloc_pages().
7700 */
7701static void setup_per_zone_lowmem_reserve(void)
7702{
7703 struct pglist_data *pgdat;
7704 enum zone_type j, idx;
7705
7706 for_each_online_pgdat(pgdat) {
7707 for (j = 0; j < MAX_NR_ZONES; j++) {
7708 struct zone *zone = pgdat->node_zones + j;
7709 unsigned long managed_pages = zone_managed_pages(zone);
7710
7711 zone->lowmem_reserve[j] = 0;
7712
7713 idx = j;
7714 while (idx) {
7715 struct zone *lower_zone;
7716
7717 idx--;
7718 lower_zone = pgdat->node_zones + idx;
7719
7720 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7721 sysctl_lowmem_reserve_ratio[idx] = 0;
7722 lower_zone->lowmem_reserve[j] = 0;
7723 } else {
7724 lower_zone->lowmem_reserve[j] =
7725 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7726 }
7727 managed_pages += zone_managed_pages(lower_zone);
7728 }
7729 }
7730 }
7731
7732 /* update totalreserve_pages */
7733 calculate_totalreserve_pages();
7734}
7735
7736static void __setup_per_zone_wmarks(void)
7737{
7738 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7739 unsigned long lowmem_pages = 0;
7740 struct zone *zone;
7741 unsigned long flags;
7742
7743 /* Calculate total number of !ZONE_HIGHMEM pages */
7744 for_each_zone(zone) {
7745 if (!is_highmem(zone))
7746 lowmem_pages += zone_managed_pages(zone);
7747 }
7748
7749 for_each_zone(zone) {
7750 u64 tmp;
7751
7752 spin_lock_irqsave(&zone->lock, flags);
7753 tmp = (u64)pages_min * zone_managed_pages(zone);
7754 do_div(tmp, lowmem_pages);
7755 if (is_highmem(zone)) {
7756 /*
7757 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7758 * need highmem pages, so cap pages_min to a small
7759 * value here.
7760 *
7761 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7762 * deltas control async page reclaim, and so should
7763 * not be capped for highmem.
7764 */
7765 unsigned long min_pages;
7766
7767 min_pages = zone_managed_pages(zone) / 1024;
7768 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7769 zone->_watermark[WMARK_MIN] = min_pages;
7770 } else {
7771 /*
7772 * If it's a lowmem zone, reserve a number of pages
7773 * proportionate to the zone's size.
7774 */
7775 zone->_watermark[WMARK_MIN] = tmp;
7776 }
7777
7778 /*
7779 * Set the kswapd watermarks distance according to the
7780 * scale factor in proportion to available memory, but
7781 * ensure a minimum size on small systems.
7782 */
7783 tmp = max_t(u64, tmp >> 2,
7784 mult_frac(zone_managed_pages(zone),
7785 watermark_scale_factor, 10000));
7786
7787 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7788 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7789 zone->watermark_boost = 0;
7790
7791 spin_unlock_irqrestore(&zone->lock, flags);
7792 }
7793
7794 /* update totalreserve_pages */
7795 calculate_totalreserve_pages();
7796}
7797
7798/**
7799 * setup_per_zone_wmarks - called when min_free_kbytes changes
7800 * or when memory is hot-{added|removed}
7801 *
7802 * Ensures that the watermark[min,low,high] values for each zone are set
7803 * correctly with respect to min_free_kbytes.
7804 */
7805void setup_per_zone_wmarks(void)
7806{
7807 static DEFINE_SPINLOCK(lock);
7808
7809 spin_lock(&lock);
7810 __setup_per_zone_wmarks();
7811 spin_unlock(&lock);
7812}
7813
7814/*
7815 * Initialise min_free_kbytes.
7816 *
7817 * For small machines we want it small (128k min). For large machines
7818 * we want it large (64MB max). But it is not linear, because network
7819 * bandwidth does not increase linearly with machine size. We use
7820 *
7821 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7822 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7823 *
7824 * which yields
7825 *
7826 * 16MB: 512k
7827 * 32MB: 724k
7828 * 64MB: 1024k
7829 * 128MB: 1448k
7830 * 256MB: 2048k
7831 * 512MB: 2896k
7832 * 1024MB: 4096k
7833 * 2048MB: 5792k
7834 * 4096MB: 8192k
7835 * 8192MB: 11584k
7836 * 16384MB: 16384k
7837 */
7838int __meminit init_per_zone_wmark_min(void)
7839{
7840 unsigned long lowmem_kbytes;
7841 int new_min_free_kbytes;
7842
7843 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7844 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7845
7846 if (new_min_free_kbytes > user_min_free_kbytes) {
7847 min_free_kbytes = new_min_free_kbytes;
7848 if (min_free_kbytes < 128)
7849 min_free_kbytes = 128;
7850 if (min_free_kbytes > 65536)
7851 min_free_kbytes = 65536;
7852 } else {
7853 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7854 new_min_free_kbytes, user_min_free_kbytes);
7855 }
7856 setup_per_zone_wmarks();
7857 refresh_zone_stat_thresholds();
7858 setup_per_zone_lowmem_reserve();
7859
7860#ifdef CONFIG_NUMA
7861 setup_min_unmapped_ratio();
7862 setup_min_slab_ratio();
7863#endif
7864
7865 return 0;
7866}
7867core_initcall(init_per_zone_wmark_min)
7868
7869/*
7870 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7871 * that we can call two helper functions whenever min_free_kbytes
7872 * changes.
7873 */
7874int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7875 void __user *buffer, size_t *length, loff_t *ppos)
7876{
7877 int rc;
7878
7879 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7880 if (rc)
7881 return rc;
7882
7883 if (write) {
7884 user_min_free_kbytes = min_free_kbytes;
7885 setup_per_zone_wmarks();
7886 }
7887 return 0;
7888}
7889
7890int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7891 void __user *buffer, size_t *length, loff_t *ppos)
7892{
7893 int rc;
7894
7895 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7896 if (rc)
7897 return rc;
7898
7899 return 0;
7900}
7901
7902int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7903 void __user *buffer, size_t *length, loff_t *ppos)
7904{
7905 int rc;
7906
7907 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7908 if (rc)
7909 return rc;
7910
7911 if (write)
7912 setup_per_zone_wmarks();
7913
7914 return 0;
7915}
7916
7917#ifdef CONFIG_NUMA
7918static void setup_min_unmapped_ratio(void)
7919{
7920 pg_data_t *pgdat;
7921 struct zone *zone;
7922
7923 for_each_online_pgdat(pgdat)
7924 pgdat->min_unmapped_pages = 0;
7925
7926 for_each_zone(zone)
7927 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7928 sysctl_min_unmapped_ratio) / 100;
7929}
7930
7931
7932int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7933 void __user *buffer, size_t *length, loff_t *ppos)
7934{
7935 int rc;
7936
7937 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7938 if (rc)
7939 return rc;
7940
7941 setup_min_unmapped_ratio();
7942
7943 return 0;
7944}
7945
7946static void setup_min_slab_ratio(void)
7947{
7948 pg_data_t *pgdat;
7949 struct zone *zone;
7950
7951 for_each_online_pgdat(pgdat)
7952 pgdat->min_slab_pages = 0;
7953
7954 for_each_zone(zone)
7955 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7956 sysctl_min_slab_ratio) / 100;
7957}
7958
7959int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7960 void __user *buffer, size_t *length, loff_t *ppos)
7961{
7962 int rc;
7963
7964 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7965 if (rc)
7966 return rc;
7967
7968 setup_min_slab_ratio();
7969
7970 return 0;
7971}
7972#endif
7973
7974/*
7975 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7976 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7977 * whenever sysctl_lowmem_reserve_ratio changes.
7978 *
7979 * The reserve ratio obviously has absolutely no relation with the
7980 * minimum watermarks. The lowmem reserve ratio can only make sense
7981 * if in function of the boot time zone sizes.
7982 */
7983int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7984 void __user *buffer, size_t *length, loff_t *ppos)
7985{
7986 proc_dointvec_minmax(table, write, buffer, length, ppos);
7987 setup_per_zone_lowmem_reserve();
7988 return 0;
7989}
7990
7991/*
7992 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7993 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7994 * pagelist can have before it gets flushed back to buddy allocator.
7995 */
7996int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7997 void __user *buffer, size_t *length, loff_t *ppos)
7998{
7999 struct zone *zone;
8000 int old_percpu_pagelist_fraction;
8001 int ret;
8002
8003 mutex_lock(&pcp_batch_high_lock);
8004 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8005
8006 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8007 if (!write || ret < 0)
8008 goto out;
8009
8010 /* Sanity checking to avoid pcp imbalance */
8011 if (percpu_pagelist_fraction &&
8012 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8013 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8014 ret = -EINVAL;
8015 goto out;
8016 }
8017
8018 /* No change? */
8019 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8020 goto out;
8021
8022 for_each_populated_zone(zone) {
8023 unsigned int cpu;
8024
8025 for_each_possible_cpu(cpu)
8026 pageset_set_high_and_batch(zone,
8027 per_cpu_ptr(zone->pageset, cpu));
8028 }
8029out:
8030 mutex_unlock(&pcp_batch_high_lock);
8031 return ret;
8032}
8033
8034#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8035/*
8036 * Returns the number of pages that arch has reserved but
8037 * is not known to alloc_large_system_hash().
8038 */
8039static unsigned long __init arch_reserved_kernel_pages(void)
8040{
8041 return 0;
8042}
8043#endif
8044
8045/*
8046 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8047 * machines. As memory size is increased the scale is also increased but at
8048 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8049 * quadruples the scale is increased by one, which means the size of hash table
8050 * only doubles, instead of quadrupling as well.
8051 * Because 32-bit systems cannot have large physical memory, where this scaling
8052 * makes sense, it is disabled on such platforms.
8053 */
8054#if __BITS_PER_LONG > 32
8055#define ADAPT_SCALE_BASE (64ul << 30)
8056#define ADAPT_SCALE_SHIFT 2
8057#define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8058#endif
8059
8060/*
8061 * allocate a large system hash table from bootmem
8062 * - it is assumed that the hash table must contain an exact power-of-2
8063 * quantity of entries
8064 * - limit is the number of hash buckets, not the total allocation size
8065 */
8066void *__init alloc_large_system_hash(const char *tablename,
8067 unsigned long bucketsize,
8068 unsigned long numentries,
8069 int scale,
8070 int flags,
8071 unsigned int *_hash_shift,
8072 unsigned int *_hash_mask,
8073 unsigned long low_limit,
8074 unsigned long high_limit)
8075{
8076 unsigned long long max = high_limit;
8077 unsigned long log2qty, size;
8078 void *table = NULL;
8079 gfp_t gfp_flags;
8080 bool virt;
8081
8082 /* allow the kernel cmdline to have a say */
8083 if (!numentries) {
8084 /* round applicable memory size up to nearest megabyte */
8085 numentries = nr_kernel_pages;
8086 numentries -= arch_reserved_kernel_pages();
8087
8088 /* It isn't necessary when PAGE_SIZE >= 1MB */
8089 if (PAGE_SHIFT < 20)
8090 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8091
8092#if __BITS_PER_LONG > 32
8093 if (!high_limit) {
8094 unsigned long adapt;
8095
8096 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8097 adapt <<= ADAPT_SCALE_SHIFT)
8098 scale++;
8099 }
8100#endif
8101
8102 /* limit to 1 bucket per 2^scale bytes of low memory */
8103 if (scale > PAGE_SHIFT)
8104 numentries >>= (scale - PAGE_SHIFT);
8105 else
8106 numentries <<= (PAGE_SHIFT - scale);
8107
8108 /* Make sure we've got at least a 0-order allocation.. */
8109 if (unlikely(flags & HASH_SMALL)) {
8110 /* Makes no sense without HASH_EARLY */
8111 WARN_ON(!(flags & HASH_EARLY));
8112 if (!(numentries >> *_hash_shift)) {
8113 numentries = 1UL << *_hash_shift;
8114 BUG_ON(!numentries);
8115 }
8116 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8117 numentries = PAGE_SIZE / bucketsize;
8118 }
8119 numentries = roundup_pow_of_two(numentries);
8120
8121 /* limit allocation size to 1/16 total memory by default */
8122 if (max == 0) {
8123 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8124 do_div(max, bucketsize);
8125 }
8126 max = min(max, 0x80000000ULL);
8127
8128 if (numentries < low_limit)
8129 numentries = low_limit;
8130 if (numentries > max)
8131 numentries = max;
8132
8133 log2qty = ilog2(numentries);
8134
8135 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8136 do {
8137 virt = false;
8138 size = bucketsize << log2qty;
8139 if (flags & HASH_EARLY) {
8140 if (flags & HASH_ZERO)
8141 table = memblock_alloc(size, SMP_CACHE_BYTES);
8142 else
8143 table = memblock_alloc_raw(size,
8144 SMP_CACHE_BYTES);
8145 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8146 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8147 virt = true;
8148 } else {
8149 /*
8150 * If bucketsize is not a power-of-two, we may free
8151 * some pages at the end of hash table which
8152 * alloc_pages_exact() automatically does
8153 */
8154 table = alloc_pages_exact(size, gfp_flags);
8155 kmemleak_alloc(table, size, 1, gfp_flags);
8156 }
8157 } while (!table && size > PAGE_SIZE && --log2qty);
8158
8159 if (!table)
8160 panic("Failed to allocate %s hash table\n", tablename);
8161
8162 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8163 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8164 virt ? "vmalloc" : "linear");
8165
8166 if (_hash_shift)
8167 *_hash_shift = log2qty;
8168 if (_hash_mask)
8169 *_hash_mask = (1 << log2qty) - 1;
8170
8171 return table;
8172}
8173
8174/*
8175 * This function checks whether pageblock includes unmovable pages or not.
8176 * If @count is not zero, it is okay to include less @count unmovable pages
8177 *
8178 * PageLRU check without isolation or lru_lock could race so that
8179 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8180 * check without lock_page also may miss some movable non-lru pages at
8181 * race condition. So you can't expect this function should be exact.
8182 */
8183bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
8184 int migratetype, int flags)
8185{
8186 unsigned long found;
8187 unsigned long iter = 0;
8188 unsigned long pfn = page_to_pfn(page);
8189 const char *reason = "unmovable page";
8190
8191 /*
8192 * TODO we could make this much more efficient by not checking every
8193 * page in the range if we know all of them are in MOVABLE_ZONE and
8194 * that the movable zone guarantees that pages are migratable but
8195 * the later is not the case right now unfortunatelly. E.g. movablecore
8196 * can still lead to having bootmem allocations in zone_movable.
8197 */
8198
8199 if (is_migrate_cma_page(page)) {
8200 /*
8201 * CMA allocations (alloc_contig_range) really need to mark
8202 * isolate CMA pageblocks even when they are not movable in fact
8203 * so consider them movable here.
8204 */
8205 if (is_migrate_cma(migratetype))
8206 return false;
8207
8208 reason = "CMA page";
8209 goto unmovable;
8210 }
8211
8212 for (found = 0; iter < pageblock_nr_pages; iter++) {
8213 unsigned long check = pfn + iter;
8214
8215 if (!pfn_valid_within(check))
8216 continue;
8217
8218 page = pfn_to_page(check);
8219
8220 if (PageReserved(page))
8221 goto unmovable;
8222
8223 /*
8224 * If the zone is movable and we have ruled out all reserved
8225 * pages then it should be reasonably safe to assume the rest
8226 * is movable.
8227 */
8228 if (zone_idx(zone) == ZONE_MOVABLE)
8229 continue;
8230
8231 /*
8232 * Hugepages are not in LRU lists, but they're movable.
8233 * We need not scan over tail pages because we don't
8234 * handle each tail page individually in migration.
8235 */
8236 if (PageHuge(page)) {
8237 struct page *head = compound_head(page);
8238 unsigned int skip_pages;
8239
8240 if (!hugepage_migration_supported(page_hstate(head)))
8241 goto unmovable;
8242
8243 skip_pages = compound_nr(head) - (page - head);
8244 iter += skip_pages - 1;
8245 continue;
8246 }
8247
8248 /*
8249 * We can't use page_count without pin a page
8250 * because another CPU can free compound page.
8251 * This check already skips compound tails of THP
8252 * because their page->_refcount is zero at all time.
8253 */
8254 if (!page_ref_count(page)) {
8255 if (PageBuddy(page))
8256 iter += (1 << page_order(page)) - 1;
8257 continue;
8258 }
8259
8260 /*
8261 * The HWPoisoned page may be not in buddy system, and
8262 * page_count() is not 0.
8263 */
8264 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
8265 continue;
8266
8267 if (__PageMovable(page))
8268 continue;
8269
8270 if (!PageLRU(page))
8271 found++;
8272 /*
8273 * If there are RECLAIMABLE pages, we need to check
8274 * it. But now, memory offline itself doesn't call
8275 * shrink_node_slabs() and it still to be fixed.
8276 */
8277 /*
8278 * If the page is not RAM, page_count()should be 0.
8279 * we don't need more check. This is an _used_ not-movable page.
8280 *
8281 * The problematic thing here is PG_reserved pages. PG_reserved
8282 * is set to both of a memory hole page and a _used_ kernel
8283 * page at boot.
8284 */
8285 if (found > count)
8286 goto unmovable;
8287 }
8288 return false;
8289unmovable:
8290 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8291 if (flags & REPORT_FAILURE)
8292 dump_page(pfn_to_page(pfn + iter), reason);
8293 return true;
8294}
8295
8296#ifdef CONFIG_CONTIG_ALLOC
8297static unsigned long pfn_max_align_down(unsigned long pfn)
8298{
8299 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8300 pageblock_nr_pages) - 1);
8301}
8302
8303static unsigned long pfn_max_align_up(unsigned long pfn)
8304{
8305 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8306 pageblock_nr_pages));
8307}
8308
8309/* [start, end) must belong to a single zone. */
8310static int __alloc_contig_migrate_range(struct compact_control *cc,
8311 unsigned long start, unsigned long end)
8312{
8313 /* This function is based on compact_zone() from compaction.c. */
8314 unsigned long nr_reclaimed;
8315 unsigned long pfn = start;
8316 unsigned int tries = 0;
8317 int ret = 0;
8318
8319 migrate_prep();
8320
8321 while (pfn < end || !list_empty(&cc->migratepages)) {
8322 if (fatal_signal_pending(current)) {
8323 ret = -EINTR;
8324 break;
8325 }
8326
8327 if (list_empty(&cc->migratepages)) {
8328 cc->nr_migratepages = 0;
8329 pfn = isolate_migratepages_range(cc, pfn, end);
8330 if (!pfn) {
8331 ret = -EINTR;
8332 break;
8333 }
8334 tries = 0;
8335 } else if (++tries == 5) {
8336 ret = ret < 0 ? ret : -EBUSY;
8337 break;
8338 }
8339
8340 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8341 &cc->migratepages);
8342 cc->nr_migratepages -= nr_reclaimed;
8343
8344 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8345 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8346 }
8347 if (ret < 0) {
8348 putback_movable_pages(&cc->migratepages);
8349 return ret;
8350 }
8351 return 0;
8352}
8353
8354/**
8355 * alloc_contig_range() -- tries to allocate given range of pages
8356 * @start: start PFN to allocate
8357 * @end: one-past-the-last PFN to allocate
8358 * @migratetype: migratetype of the underlaying pageblocks (either
8359 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8360 * in range must have the same migratetype and it must
8361 * be either of the two.
8362 * @gfp_mask: GFP mask to use during compaction
8363 *
8364 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8365 * aligned. The PFN range must belong to a single zone.
8366 *
8367 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8368 * pageblocks in the range. Once isolated, the pageblocks should not
8369 * be modified by others.
8370 *
8371 * Return: zero on success or negative error code. On success all
8372 * pages which PFN is in [start, end) are allocated for the caller and
8373 * need to be freed with free_contig_range().
8374 */
8375int alloc_contig_range(unsigned long start, unsigned long end,
8376 unsigned migratetype, gfp_t gfp_mask)
8377{
8378 unsigned long outer_start, outer_end;
8379 unsigned int order;
8380 int ret = 0;
8381
8382 struct compact_control cc = {
8383 .nr_migratepages = 0,
8384 .order = -1,
8385 .zone = page_zone(pfn_to_page(start)),
8386 .mode = MIGRATE_SYNC,
8387 .ignore_skip_hint = true,
8388 .no_set_skip_hint = true,
8389 .gfp_mask = current_gfp_context(gfp_mask),
8390 };
8391 INIT_LIST_HEAD(&cc.migratepages);
8392
8393 /*
8394 * What we do here is we mark all pageblocks in range as
8395 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8396 * have different sizes, and due to the way page allocator
8397 * work, we align the range to biggest of the two pages so
8398 * that page allocator won't try to merge buddies from
8399 * different pageblocks and change MIGRATE_ISOLATE to some
8400 * other migration type.
8401 *
8402 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8403 * migrate the pages from an unaligned range (ie. pages that
8404 * we are interested in). This will put all the pages in
8405 * range back to page allocator as MIGRATE_ISOLATE.
8406 *
8407 * When this is done, we take the pages in range from page
8408 * allocator removing them from the buddy system. This way
8409 * page allocator will never consider using them.
8410 *
8411 * This lets us mark the pageblocks back as
8412 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8413 * aligned range but not in the unaligned, original range are
8414 * put back to page allocator so that buddy can use them.
8415 */
8416
8417 ret = start_isolate_page_range(pfn_max_align_down(start),
8418 pfn_max_align_up(end), migratetype, 0);
8419 if (ret < 0)
8420 return ret;
8421
8422 /*
8423 * In case of -EBUSY, we'd like to know which page causes problem.
8424 * So, just fall through. test_pages_isolated() has a tracepoint
8425 * which will report the busy page.
8426 *
8427 * It is possible that busy pages could become available before
8428 * the call to test_pages_isolated, and the range will actually be
8429 * allocated. So, if we fall through be sure to clear ret so that
8430 * -EBUSY is not accidentally used or returned to caller.
8431 */
8432 ret = __alloc_contig_migrate_range(&cc, start, end);
8433 if (ret && ret != -EBUSY)
8434 goto done;
8435 ret =0;
8436
8437 /*
8438 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8439 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8440 * more, all pages in [start, end) are free in page allocator.
8441 * What we are going to do is to allocate all pages from
8442 * [start, end) (that is remove them from page allocator).
8443 *
8444 * The only problem is that pages at the beginning and at the
8445 * end of interesting range may be not aligned with pages that
8446 * page allocator holds, ie. they can be part of higher order
8447 * pages. Because of this, we reserve the bigger range and
8448 * once this is done free the pages we are not interested in.
8449 *
8450 * We don't have to hold zone->lock here because the pages are
8451 * isolated thus they won't get removed from buddy.
8452 */
8453
8454 lru_add_drain_all();
8455
8456 order = 0;
8457 outer_start = start;
8458 while (!PageBuddy(pfn_to_page(outer_start))) {
8459 if (++order >= MAX_ORDER) {
8460 outer_start = start;
8461 break;
8462 }
8463 outer_start &= ~0UL << order;
8464 }
8465
8466 if (outer_start != start) {
8467 order = page_order(pfn_to_page(outer_start));
8468
8469 /*
8470 * outer_start page could be small order buddy page and
8471 * it doesn't include start page. Adjust outer_start
8472 * in this case to report failed page properly
8473 * on tracepoint in test_pages_isolated()
8474 */
8475 if (outer_start + (1UL << order) <= start)
8476 outer_start = start;
8477 }
8478
8479 /* Make sure the range is really isolated. */
8480 if (test_pages_isolated(outer_start, end, false)) {
8481 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8482 __func__, outer_start, end);
8483 ret = -EBUSY;
8484 goto done;
8485 }
8486
8487 /* Grab isolated pages from freelists. */
8488 outer_end = isolate_freepages_range(&cc, outer_start, end);
8489 if (!outer_end) {
8490 ret = -EBUSY;
8491 goto done;
8492 }
8493
8494 /* Free head and tail (if any) */
8495 if (start != outer_start)
8496 free_contig_range(outer_start, start - outer_start);
8497 if (end != outer_end)
8498 free_contig_range(end, outer_end - end);
8499
8500done:
8501 undo_isolate_page_range(pfn_max_align_down(start),
8502 pfn_max_align_up(end), migratetype);
8503 return ret;
8504}
8505#endif /* CONFIG_CONTIG_ALLOC */
8506
8507void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8508{
8509 unsigned int count = 0;
8510
8511 for (; nr_pages--; pfn++) {
8512 struct page *page = pfn_to_page(pfn);
8513
8514 count += page_count(page) != 1;
8515 __free_page(page);
8516 }
8517 WARN(count != 0, "%d pages are still in use!\n", count);
8518}
8519
8520/*
8521 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8522 * page high values need to be recalulated.
8523 */
8524void __meminit zone_pcp_update(struct zone *zone)
8525{
8526 unsigned cpu;
8527 mutex_lock(&pcp_batch_high_lock);
8528 for_each_possible_cpu(cpu)
8529 pageset_set_high_and_batch(zone,
8530 per_cpu_ptr(zone->pageset, cpu));
8531 mutex_unlock(&pcp_batch_high_lock);
8532}
8533
8534void zone_pcp_reset(struct zone *zone)
8535{
8536 unsigned long flags;
8537 int cpu;
8538 struct per_cpu_pageset *pset;
8539
8540 /* avoid races with drain_pages() */
8541 local_irq_save(flags);
8542 if (zone->pageset != &boot_pageset) {
8543 for_each_online_cpu(cpu) {
8544 pset = per_cpu_ptr(zone->pageset, cpu);
8545 drain_zonestat(zone, pset);
8546 }
8547 free_percpu(zone->pageset);
8548 zone->pageset = &boot_pageset;
8549 }
8550 local_irq_restore(flags);
8551}
8552
8553#ifdef CONFIG_MEMORY_HOTREMOVE
8554/*
8555 * All pages in the range must be in a single zone and isolated
8556 * before calling this.
8557 */
8558unsigned long
8559__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8560{
8561 struct page *page;
8562 struct zone *zone;
8563 unsigned int order, i;
8564 unsigned long pfn;
8565 unsigned long flags;
8566 unsigned long offlined_pages = 0;
8567
8568 /* find the first valid pfn */
8569 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8570 if (pfn_valid(pfn))
8571 break;
8572 if (pfn == end_pfn)
8573 return offlined_pages;
8574
8575 offline_mem_sections(pfn, end_pfn);
8576 zone = page_zone(pfn_to_page(pfn));
8577 spin_lock_irqsave(&zone->lock, flags);
8578 pfn = start_pfn;
8579 while (pfn < end_pfn) {
8580 if (!pfn_valid(pfn)) {
8581 pfn++;
8582 continue;
8583 }
8584 page = pfn_to_page(pfn);
8585 /*
8586 * The HWPoisoned page may be not in buddy system, and
8587 * page_count() is not 0.
8588 */
8589 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8590 pfn++;
8591 SetPageReserved(page);
8592 offlined_pages++;
8593 continue;
8594 }
8595
8596 BUG_ON(page_count(page));
8597 BUG_ON(!PageBuddy(page));
8598 order = page_order(page);
8599 offlined_pages += 1 << order;
8600#ifdef CONFIG_DEBUG_VM
8601 pr_info("remove from free list %lx %d %lx\n",
8602 pfn, 1 << order, end_pfn);
8603#endif
8604 del_page_from_free_area(page, &zone->free_area[order]);
8605 for (i = 0; i < (1 << order); i++)
8606 SetPageReserved((page+i));
8607 pfn += (1 << order);
8608 }
8609 spin_unlock_irqrestore(&zone->lock, flags);
8610
8611 return offlined_pages;
8612}
8613#endif
8614
8615bool is_free_buddy_page(struct page *page)
8616{
8617 struct zone *zone = page_zone(page);
8618 unsigned long pfn = page_to_pfn(page);
8619 unsigned long flags;
8620 unsigned int order;
8621
8622 spin_lock_irqsave(&zone->lock, flags);
8623 for (order = 0; order < MAX_ORDER; order++) {
8624 struct page *page_head = page - (pfn & ((1 << order) - 1));
8625
8626 if (PageBuddy(page_head) && page_order(page_head) >= order)
8627 break;
8628 }
8629 spin_unlock_irqrestore(&zone->lock, flags);
8630
8631 return order < MAX_ORDER;
8632}
8633
8634#ifdef CONFIG_MEMORY_FAILURE
8635/*
8636 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8637 * test is performed under the zone lock to prevent a race against page
8638 * allocation.
8639 */
8640bool set_hwpoison_free_buddy_page(struct page *page)
8641{
8642 struct zone *zone = page_zone(page);
8643 unsigned long pfn = page_to_pfn(page);
8644 unsigned long flags;
8645 unsigned int order;
8646 bool hwpoisoned = false;
8647
8648 spin_lock_irqsave(&zone->lock, flags);
8649 for (order = 0; order < MAX_ORDER; order++) {
8650 struct page *page_head = page - (pfn & ((1 << order) - 1));
8651
8652 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8653 if (!TestSetPageHWPoison(page))
8654 hwpoisoned = true;
8655 break;
8656 }
8657 }
8658 spin_unlock_irqrestore(&zone->lock, flags);
8659
8660 return hwpoisoned;
8661}
8662#endif