2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
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)
17 #include <linux/stddef.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/kmemcheck.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/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.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/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
77 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
78 static DEFINE_MUTEX(pcp_batch_high_lock);
79 #define MIN_PERCPU_PAGELIST_FRACTION (8)
81 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
82 DEFINE_PER_CPU(int, numa_node);
83 EXPORT_PER_CPU_SYMBOL(numa_node);
86 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
88 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
89 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
90 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
91 * defined in <linux/topology.h>.
93 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
94 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
95 int _node_numa_mem_[MAX_NUMNODES];
98 /* work_structs for global per-cpu drains */
99 DEFINE_MUTEX(pcpu_drain_mutex);
100 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
102 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
103 volatile unsigned long latent_entropy __latent_entropy;
104 EXPORT_SYMBOL(latent_entropy);
108 * Array of node states.
110 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
111 [N_POSSIBLE] = NODE_MASK_ALL,
112 [N_ONLINE] = { { [0] = 1UL } },
114 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
115 #ifdef CONFIG_HIGHMEM
116 [N_HIGH_MEMORY] = { { [0] = 1UL } },
118 [N_MEMORY] = { { [0] = 1UL } },
119 [N_CPU] = { { [0] = 1UL } },
122 EXPORT_SYMBOL(node_states);
124 /* Protect totalram_pages and zone->managed_pages */
125 static DEFINE_SPINLOCK(managed_page_count_lock);
127 unsigned long totalram_pages __read_mostly;
128 unsigned long totalreserve_pages __read_mostly;
129 unsigned long totalcma_pages __read_mostly;
131 int percpu_pagelist_fraction;
132 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
135 * A cached value of the page's pageblock's migratetype, used when the page is
136 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
137 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
138 * Also the migratetype set in the page does not necessarily match the pcplist
139 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
140 * other index - this ensures that it will be put on the correct CMA freelist.
142 static inline int get_pcppage_migratetype(struct page *page)
147 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
149 page->index = migratetype;
152 #ifdef CONFIG_PM_SLEEP
154 * The following functions are used by the suspend/hibernate code to temporarily
155 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
156 * while devices are suspended. To avoid races with the suspend/hibernate code,
157 * they should always be called with pm_mutex held (gfp_allowed_mask also should
158 * only be modified with pm_mutex held, unless the suspend/hibernate code is
159 * guaranteed not to run in parallel with that modification).
162 static gfp_t saved_gfp_mask;
164 void pm_restore_gfp_mask(void)
166 WARN_ON(!mutex_is_locked(&pm_mutex));
167 if (saved_gfp_mask) {
168 gfp_allowed_mask = saved_gfp_mask;
173 void pm_restrict_gfp_mask(void)
175 WARN_ON(!mutex_is_locked(&pm_mutex));
176 WARN_ON(saved_gfp_mask);
177 saved_gfp_mask = gfp_allowed_mask;
178 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
181 bool pm_suspended_storage(void)
183 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
187 #endif /* CONFIG_PM_SLEEP */
189 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
190 unsigned int pageblock_order __read_mostly;
193 static void __free_pages_ok(struct page *page, unsigned int order);
196 * results with 256, 32 in the lowmem_reserve sysctl:
197 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
198 * 1G machine -> (16M dma, 784M normal, 224M high)
199 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
200 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
201 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 * TBD: should special case ZONE_DMA32 machines here - in those we normally
204 * don't need any ZONE_NORMAL reservation
206 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
207 #ifdef CONFIG_ZONE_DMA
210 #ifdef CONFIG_ZONE_DMA32
213 #ifdef CONFIG_HIGHMEM
219 EXPORT_SYMBOL(totalram_pages);
221 static char * const zone_names[MAX_NR_ZONES] = {
222 #ifdef CONFIG_ZONE_DMA
225 #ifdef CONFIG_ZONE_DMA32
229 #ifdef CONFIG_HIGHMEM
233 #ifdef CONFIG_ZONE_DEVICE
238 char * const migratetype_names[MIGRATE_TYPES] = {
246 #ifdef CONFIG_MEMORY_ISOLATION
251 compound_page_dtor * const compound_page_dtors[] = {
254 #ifdef CONFIG_HUGETLB_PAGE
257 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
262 int min_free_kbytes = 1024;
263 int user_min_free_kbytes = -1;
264 int watermark_scale_factor = 10;
266 static unsigned long __meminitdata nr_kernel_pages;
267 static unsigned long __meminitdata nr_all_pages;
268 static unsigned long __meminitdata dma_reserve;
270 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
271 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
272 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
273 static unsigned long __initdata required_kernelcore;
274 static unsigned long __initdata required_movablecore;
275 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
276 static bool mirrored_kernelcore;
278 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
280 EXPORT_SYMBOL(movable_zone);
281 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
284 int nr_node_ids __read_mostly = MAX_NUMNODES;
285 int nr_online_nodes __read_mostly = 1;
286 EXPORT_SYMBOL(nr_node_ids);
287 EXPORT_SYMBOL(nr_online_nodes);
290 int page_group_by_mobility_disabled __read_mostly;
292 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
293 static inline void reset_deferred_meminit(pg_data_t *pgdat)
295 unsigned long max_initialise;
296 unsigned long reserved_lowmem;
299 * Initialise at least 2G of a node but also take into account that
300 * two large system hashes that can take up 1GB for 0.25TB/node.
302 max_initialise = max(2UL << (30 - PAGE_SHIFT),
303 (pgdat->node_spanned_pages >> 8));
306 * Compensate the all the memblock reservations (e.g. crash kernel)
307 * from the initial estimation to make sure we will initialize enough
310 reserved_lowmem = memblock_reserved_memory_within(pgdat->node_start_pfn,
311 pgdat->node_start_pfn + max_initialise);
312 max_initialise += reserved_lowmem;
314 pgdat->static_init_size = min(max_initialise, pgdat->node_spanned_pages);
315 pgdat->first_deferred_pfn = ULONG_MAX;
318 /* Returns true if the struct page for the pfn is uninitialised */
319 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
321 int nid = early_pfn_to_nid(pfn);
323 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
330 * Returns false when the remaining initialisation should be deferred until
331 * later in the boot cycle when it can be parallelised.
333 static inline bool update_defer_init(pg_data_t *pgdat,
334 unsigned long pfn, unsigned long zone_end,
335 unsigned long *nr_initialised)
337 /* Always populate low zones for address-contrained allocations */
338 if (zone_end < pgdat_end_pfn(pgdat))
341 if ((*nr_initialised > pgdat->static_init_size) &&
342 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
343 pgdat->first_deferred_pfn = pfn;
350 static inline void reset_deferred_meminit(pg_data_t *pgdat)
354 static inline bool early_page_uninitialised(unsigned long pfn)
359 static inline bool update_defer_init(pg_data_t *pgdat,
360 unsigned long pfn, unsigned long zone_end,
361 unsigned long *nr_initialised)
367 /* Return a pointer to the bitmap storing bits affecting a block of pages */
368 static inline unsigned long *get_pageblock_bitmap(struct page *page,
371 #ifdef CONFIG_SPARSEMEM
372 return __pfn_to_section(pfn)->pageblock_flags;
374 return page_zone(page)->pageblock_flags;
375 #endif /* CONFIG_SPARSEMEM */
378 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
380 #ifdef CONFIG_SPARSEMEM
381 pfn &= (PAGES_PER_SECTION-1);
382 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
384 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
385 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
386 #endif /* CONFIG_SPARSEMEM */
390 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
391 * @page: The page within the block of interest
392 * @pfn: The target page frame number
393 * @end_bitidx: The last bit of interest to retrieve
394 * @mask: mask of bits that the caller is interested in
396 * Return: pageblock_bits flags
398 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
400 unsigned long end_bitidx,
403 unsigned long *bitmap;
404 unsigned long bitidx, word_bitidx;
407 bitmap = get_pageblock_bitmap(page, pfn);
408 bitidx = pfn_to_bitidx(page, pfn);
409 word_bitidx = bitidx / BITS_PER_LONG;
410 bitidx &= (BITS_PER_LONG-1);
412 word = bitmap[word_bitidx];
413 bitidx += end_bitidx;
414 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
417 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
418 unsigned long end_bitidx,
421 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
424 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
426 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
430 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
431 * @page: The page within the block of interest
432 * @flags: The flags to set
433 * @pfn: The target page frame number
434 * @end_bitidx: The last bit of interest
435 * @mask: mask of bits that the caller is interested in
437 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
439 unsigned long end_bitidx,
442 unsigned long *bitmap;
443 unsigned long bitidx, word_bitidx;
444 unsigned long old_word, word;
446 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
448 bitmap = get_pageblock_bitmap(page, pfn);
449 bitidx = pfn_to_bitidx(page, pfn);
450 word_bitidx = bitidx / BITS_PER_LONG;
451 bitidx &= (BITS_PER_LONG-1);
453 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
455 bitidx += end_bitidx;
456 mask <<= (BITS_PER_LONG - bitidx - 1);
457 flags <<= (BITS_PER_LONG - bitidx - 1);
459 word = READ_ONCE(bitmap[word_bitidx]);
461 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
462 if (word == old_word)
468 void set_pageblock_migratetype(struct page *page, int migratetype)
470 if (unlikely(page_group_by_mobility_disabled &&
471 migratetype < MIGRATE_PCPTYPES))
472 migratetype = MIGRATE_UNMOVABLE;
474 set_pageblock_flags_group(page, (unsigned long)migratetype,
475 PB_migrate, PB_migrate_end);
478 #ifdef CONFIG_DEBUG_VM
479 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
483 unsigned long pfn = page_to_pfn(page);
484 unsigned long sp, start_pfn;
487 seq = zone_span_seqbegin(zone);
488 start_pfn = zone->zone_start_pfn;
489 sp = zone->spanned_pages;
490 if (!zone_spans_pfn(zone, pfn))
492 } while (zone_span_seqretry(zone, seq));
495 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
496 pfn, zone_to_nid(zone), zone->name,
497 start_pfn, start_pfn + sp);
502 static int page_is_consistent(struct zone *zone, struct page *page)
504 if (!pfn_valid_within(page_to_pfn(page)))
506 if (zone != page_zone(page))
512 * Temporary debugging check for pages not lying within a given zone.
514 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
516 if (page_outside_zone_boundaries(zone, page))
518 if (!page_is_consistent(zone, page))
524 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
530 static void bad_page(struct page *page, const char *reason,
531 unsigned long bad_flags)
533 static unsigned long resume;
534 static unsigned long nr_shown;
535 static unsigned long nr_unshown;
538 * Allow a burst of 60 reports, then keep quiet for that minute;
539 * or allow a steady drip of one report per second.
541 if (nr_shown == 60) {
542 if (time_before(jiffies, resume)) {
548 "BUG: Bad page state: %lu messages suppressed\n",
555 resume = jiffies + 60 * HZ;
557 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
558 current->comm, page_to_pfn(page));
559 __dump_page(page, reason);
560 bad_flags &= page->flags;
562 pr_alert("bad because of flags: %#lx(%pGp)\n",
563 bad_flags, &bad_flags);
564 dump_page_owner(page);
569 /* Leave bad fields for debug, except PageBuddy could make trouble */
570 page_mapcount_reset(page); /* remove PageBuddy */
571 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
575 * Higher-order pages are called "compound pages". They are structured thusly:
577 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
579 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
580 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
582 * The first tail page's ->compound_dtor holds the offset in array of compound
583 * page destructors. See compound_page_dtors.
585 * The first tail page's ->compound_order holds the order of allocation.
586 * This usage means that zero-order pages may not be compound.
589 void free_compound_page(struct page *page)
591 __free_pages_ok(page, compound_order(page));
594 void prep_compound_page(struct page *page, unsigned int order)
597 int nr_pages = 1 << order;
599 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
600 set_compound_order(page, order);
602 for (i = 1; i < nr_pages; i++) {
603 struct page *p = page + i;
604 set_page_count(p, 0);
605 p->mapping = TAIL_MAPPING;
606 set_compound_head(p, page);
608 atomic_set(compound_mapcount_ptr(page), -1);
611 #ifdef CONFIG_DEBUG_PAGEALLOC
612 unsigned int _debug_guardpage_minorder;
613 bool _debug_pagealloc_enabled __read_mostly
614 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
615 EXPORT_SYMBOL(_debug_pagealloc_enabled);
616 bool _debug_guardpage_enabled __read_mostly;
618 static int __init early_debug_pagealloc(char *buf)
622 return kstrtobool(buf, &_debug_pagealloc_enabled);
624 early_param("debug_pagealloc", early_debug_pagealloc);
626 static bool need_debug_guardpage(void)
628 /* If we don't use debug_pagealloc, we don't need guard page */
629 if (!debug_pagealloc_enabled())
632 if (!debug_guardpage_minorder())
638 static void init_debug_guardpage(void)
640 if (!debug_pagealloc_enabled())
643 if (!debug_guardpage_minorder())
646 _debug_guardpage_enabled = true;
649 struct page_ext_operations debug_guardpage_ops = {
650 .need = need_debug_guardpage,
651 .init = init_debug_guardpage,
654 static int __init debug_guardpage_minorder_setup(char *buf)
658 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
659 pr_err("Bad debug_guardpage_minorder value\n");
662 _debug_guardpage_minorder = res;
663 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
666 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
668 static inline bool set_page_guard(struct zone *zone, struct page *page,
669 unsigned int order, int migratetype)
671 struct page_ext *page_ext;
673 if (!debug_guardpage_enabled())
676 if (order >= debug_guardpage_minorder())
679 page_ext = lookup_page_ext(page);
680 if (unlikely(!page_ext))
683 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
685 INIT_LIST_HEAD(&page->lru);
686 set_page_private(page, order);
687 /* Guard pages are not available for any usage */
688 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
693 static inline void clear_page_guard(struct zone *zone, struct page *page,
694 unsigned int order, int migratetype)
696 struct page_ext *page_ext;
698 if (!debug_guardpage_enabled())
701 page_ext = lookup_page_ext(page);
702 if (unlikely(!page_ext))
705 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
707 set_page_private(page, 0);
708 if (!is_migrate_isolate(migratetype))
709 __mod_zone_freepage_state(zone, (1 << order), migratetype);
712 struct page_ext_operations debug_guardpage_ops;
713 static inline bool set_page_guard(struct zone *zone, struct page *page,
714 unsigned int order, int migratetype) { return false; }
715 static inline void clear_page_guard(struct zone *zone, struct page *page,
716 unsigned int order, int migratetype) {}
719 static inline void set_page_order(struct page *page, unsigned int order)
721 set_page_private(page, order);
722 __SetPageBuddy(page);
725 static inline void rmv_page_order(struct page *page)
727 __ClearPageBuddy(page);
728 set_page_private(page, 0);
732 * This function checks whether a page is free && is the buddy
733 * we can do coalesce a page and its buddy if
734 * (a) the buddy is not in a hole (check before calling!) &&
735 * (b) the buddy is in the buddy system &&
736 * (c) a page and its buddy have the same order &&
737 * (d) a page and its buddy are in the same zone.
739 * For recording whether a page is in the buddy system, we set ->_mapcount
740 * PAGE_BUDDY_MAPCOUNT_VALUE.
741 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
742 * serialized by zone->lock.
744 * For recording page's order, we use page_private(page).
746 static inline int page_is_buddy(struct page *page, struct page *buddy,
749 if (page_is_guard(buddy) && page_order(buddy) == order) {
750 if (page_zone_id(page) != page_zone_id(buddy))
753 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
758 if (PageBuddy(buddy) && page_order(buddy) == order) {
760 * zone check is done late to avoid uselessly
761 * calculating zone/node ids for pages that could
764 if (page_zone_id(page) != page_zone_id(buddy))
767 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
775 * Freeing function for a buddy system allocator.
777 * The concept of a buddy system is to maintain direct-mapped table
778 * (containing bit values) for memory blocks of various "orders".
779 * The bottom level table contains the map for the smallest allocatable
780 * units of memory (here, pages), and each level above it describes
781 * pairs of units from the levels below, hence, "buddies".
782 * At a high level, all that happens here is marking the table entry
783 * at the bottom level available, and propagating the changes upward
784 * as necessary, plus some accounting needed to play nicely with other
785 * parts of the VM system.
786 * At each level, we keep a list of pages, which are heads of continuous
787 * free pages of length of (1 << order) and marked with _mapcount
788 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
790 * So when we are allocating or freeing one, we can derive the state of the
791 * other. That is, if we allocate a small block, and both were
792 * free, the remainder of the region must be split into blocks.
793 * If a block is freed, and its buddy is also free, then this
794 * triggers coalescing into a block of larger size.
799 static inline void __free_one_page(struct page *page,
801 struct zone *zone, unsigned int order,
804 unsigned long combined_pfn;
805 unsigned long uninitialized_var(buddy_pfn);
807 unsigned int max_order;
809 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
811 VM_BUG_ON(!zone_is_initialized(zone));
812 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
814 VM_BUG_ON(migratetype == -1);
815 if (likely(!is_migrate_isolate(migratetype)))
816 __mod_zone_freepage_state(zone, 1 << order, migratetype);
818 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
819 VM_BUG_ON_PAGE(bad_range(zone, page), page);
822 while (order < max_order - 1) {
823 buddy_pfn = __find_buddy_pfn(pfn, order);
824 buddy = page + (buddy_pfn - pfn);
826 if (!pfn_valid_within(buddy_pfn))
828 if (!page_is_buddy(page, buddy, order))
831 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
832 * merge with it and move up one order.
834 if (page_is_guard(buddy)) {
835 clear_page_guard(zone, buddy, order, migratetype);
837 list_del(&buddy->lru);
838 zone->free_area[order].nr_free--;
839 rmv_page_order(buddy);
841 combined_pfn = buddy_pfn & pfn;
842 page = page + (combined_pfn - pfn);
846 if (max_order < MAX_ORDER) {
847 /* If we are here, it means order is >= pageblock_order.
848 * We want to prevent merge between freepages on isolate
849 * pageblock and normal pageblock. Without this, pageblock
850 * isolation could cause incorrect freepage or CMA accounting.
852 * We don't want to hit this code for the more frequent
855 if (unlikely(has_isolate_pageblock(zone))) {
858 buddy_pfn = __find_buddy_pfn(pfn, order);
859 buddy = page + (buddy_pfn - pfn);
860 buddy_mt = get_pageblock_migratetype(buddy);
862 if (migratetype != buddy_mt
863 && (is_migrate_isolate(migratetype) ||
864 is_migrate_isolate(buddy_mt)))
868 goto continue_merging;
872 set_page_order(page, order);
875 * If this is not the largest possible page, check if the buddy
876 * of the next-highest order is free. If it is, it's possible
877 * that pages are being freed that will coalesce soon. In case,
878 * that is happening, add the free page to the tail of the list
879 * so it's less likely to be used soon and more likely to be merged
880 * as a higher order page
882 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
883 struct page *higher_page, *higher_buddy;
884 combined_pfn = buddy_pfn & pfn;
885 higher_page = page + (combined_pfn - pfn);
886 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
887 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
888 if (pfn_valid_within(buddy_pfn) &&
889 page_is_buddy(higher_page, higher_buddy, order + 1)) {
890 list_add_tail(&page->lru,
891 &zone->free_area[order].free_list[migratetype]);
896 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
898 zone->free_area[order].nr_free++;
902 * A bad page could be due to a number of fields. Instead of multiple branches,
903 * try and check multiple fields with one check. The caller must do a detailed
904 * check if necessary.
906 static inline bool page_expected_state(struct page *page,
907 unsigned long check_flags)
909 if (unlikely(atomic_read(&page->_mapcount) != -1))
912 if (unlikely((unsigned long)page->mapping |
913 page_ref_count(page) |
915 (unsigned long)page->mem_cgroup |
917 (page->flags & check_flags)))
923 static void free_pages_check_bad(struct page *page)
925 const char *bad_reason;
926 unsigned long bad_flags;
931 if (unlikely(atomic_read(&page->_mapcount) != -1))
932 bad_reason = "nonzero mapcount";
933 if (unlikely(page->mapping != NULL))
934 bad_reason = "non-NULL mapping";
935 if (unlikely(page_ref_count(page) != 0))
936 bad_reason = "nonzero _refcount";
937 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
938 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
939 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
942 if (unlikely(page->mem_cgroup))
943 bad_reason = "page still charged to cgroup";
945 bad_page(page, bad_reason, bad_flags);
948 static inline int free_pages_check(struct page *page)
950 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
953 /* Something has gone sideways, find it */
954 free_pages_check_bad(page);
958 static int free_tail_pages_check(struct page *head_page, struct page *page)
963 * We rely page->lru.next never has bit 0 set, unless the page
964 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
966 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
968 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
972 switch (page - head_page) {
974 /* the first tail page: ->mapping is compound_mapcount() */
975 if (unlikely(compound_mapcount(page))) {
976 bad_page(page, "nonzero compound_mapcount", 0);
982 * the second tail page: ->mapping is
983 * page_deferred_list().next -- ignore value.
987 if (page->mapping != TAIL_MAPPING) {
988 bad_page(page, "corrupted mapping in tail page", 0);
993 if (unlikely(!PageTail(page))) {
994 bad_page(page, "PageTail not set", 0);
997 if (unlikely(compound_head(page) != head_page)) {
998 bad_page(page, "compound_head not consistent", 0);
1003 page->mapping = NULL;
1004 clear_compound_head(page);
1008 static __always_inline bool free_pages_prepare(struct page *page,
1009 unsigned int order, bool check_free)
1013 VM_BUG_ON_PAGE(PageTail(page), page);
1015 trace_mm_page_free(page, order);
1016 kmemcheck_free_shadow(page, order);
1019 * Check tail pages before head page information is cleared to
1020 * avoid checking PageCompound for order-0 pages.
1022 if (unlikely(order)) {
1023 bool compound = PageCompound(page);
1026 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1029 ClearPageDoubleMap(page);
1030 for (i = 1; i < (1 << order); i++) {
1032 bad += free_tail_pages_check(page, page + i);
1033 if (unlikely(free_pages_check(page + i))) {
1037 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1040 if (PageMappingFlags(page))
1041 page->mapping = NULL;
1042 if (memcg_kmem_enabled() && PageKmemcg(page))
1043 memcg_kmem_uncharge(page, order);
1045 bad += free_pages_check(page);
1049 page_cpupid_reset_last(page);
1050 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1051 reset_page_owner(page, order);
1053 if (!PageHighMem(page)) {
1054 debug_check_no_locks_freed(page_address(page),
1055 PAGE_SIZE << order);
1056 debug_check_no_obj_freed(page_address(page),
1057 PAGE_SIZE << order);
1059 arch_free_page(page, order);
1060 kernel_poison_pages(page, 1 << order, 0);
1061 kernel_map_pages(page, 1 << order, 0);
1062 kasan_free_pages(page, order);
1067 #ifdef CONFIG_DEBUG_VM
1068 static inline bool free_pcp_prepare(struct page *page)
1070 return free_pages_prepare(page, 0, true);
1073 static inline bool bulkfree_pcp_prepare(struct page *page)
1078 static bool free_pcp_prepare(struct page *page)
1080 return free_pages_prepare(page, 0, false);
1083 static bool bulkfree_pcp_prepare(struct page *page)
1085 return free_pages_check(page);
1087 #endif /* CONFIG_DEBUG_VM */
1090 * Frees a number of pages from the PCP lists
1091 * Assumes all pages on list are in same zone, and of same order.
1092 * count is the number of pages to free.
1094 * If the zone was previously in an "all pages pinned" state then look to
1095 * see if this freeing clears that state.
1097 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1098 * pinned" detection logic.
1100 static void free_pcppages_bulk(struct zone *zone, int count,
1101 struct per_cpu_pages *pcp)
1103 int migratetype = 0;
1105 bool isolated_pageblocks;
1107 spin_lock(&zone->lock);
1108 isolated_pageblocks = has_isolate_pageblock(zone);
1112 struct list_head *list;
1115 * Remove pages from lists in a round-robin fashion. A
1116 * batch_free count is maintained that is incremented when an
1117 * empty list is encountered. This is so more pages are freed
1118 * off fuller lists instead of spinning excessively around empty
1123 if (++migratetype == MIGRATE_PCPTYPES)
1125 list = &pcp->lists[migratetype];
1126 } while (list_empty(list));
1128 /* This is the only non-empty list. Free them all. */
1129 if (batch_free == MIGRATE_PCPTYPES)
1133 int mt; /* migratetype of the to-be-freed page */
1135 page = list_last_entry(list, struct page, lru);
1136 /* must delete as __free_one_page list manipulates */
1137 list_del(&page->lru);
1139 mt = get_pcppage_migratetype(page);
1140 /* MIGRATE_ISOLATE page should not go to pcplists */
1141 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1142 /* Pageblock could have been isolated meanwhile */
1143 if (unlikely(isolated_pageblocks))
1144 mt = get_pageblock_migratetype(page);
1146 if (bulkfree_pcp_prepare(page))
1149 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1150 trace_mm_page_pcpu_drain(page, 0, mt);
1151 } while (--count && --batch_free && !list_empty(list));
1153 spin_unlock(&zone->lock);
1156 static void free_one_page(struct zone *zone,
1157 struct page *page, unsigned long pfn,
1161 spin_lock(&zone->lock);
1162 if (unlikely(has_isolate_pageblock(zone) ||
1163 is_migrate_isolate(migratetype))) {
1164 migratetype = get_pfnblock_migratetype(page, pfn);
1166 __free_one_page(page, pfn, zone, order, migratetype);
1167 spin_unlock(&zone->lock);
1170 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1171 unsigned long zone, int nid)
1173 set_page_links(page, zone, nid, pfn);
1174 init_page_count(page);
1175 page_mapcount_reset(page);
1176 page_cpupid_reset_last(page);
1178 INIT_LIST_HEAD(&page->lru);
1179 #ifdef WANT_PAGE_VIRTUAL
1180 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1181 if (!is_highmem_idx(zone))
1182 set_page_address(page, __va(pfn << PAGE_SHIFT));
1186 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1189 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1192 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1193 static void __meminit init_reserved_page(unsigned long pfn)
1198 if (!early_page_uninitialised(pfn))
1201 nid = early_pfn_to_nid(pfn);
1202 pgdat = NODE_DATA(nid);
1204 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1205 struct zone *zone = &pgdat->node_zones[zid];
1207 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1210 __init_single_pfn(pfn, zid, nid);
1213 static inline void init_reserved_page(unsigned long pfn)
1216 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1219 * Initialised pages do not have PageReserved set. This function is
1220 * called for each range allocated by the bootmem allocator and
1221 * marks the pages PageReserved. The remaining valid pages are later
1222 * sent to the buddy page allocator.
1224 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1226 unsigned long start_pfn = PFN_DOWN(start);
1227 unsigned long end_pfn = PFN_UP(end);
1229 for (; start_pfn < end_pfn; start_pfn++) {
1230 if (pfn_valid(start_pfn)) {
1231 struct page *page = pfn_to_page(start_pfn);
1233 init_reserved_page(start_pfn);
1235 /* Avoid false-positive PageTail() */
1236 INIT_LIST_HEAD(&page->lru);
1238 SetPageReserved(page);
1243 static void __free_pages_ok(struct page *page, unsigned int order)
1245 unsigned long flags;
1247 unsigned long pfn = page_to_pfn(page);
1249 if (!free_pages_prepare(page, order, true))
1252 migratetype = get_pfnblock_migratetype(page, pfn);
1253 local_irq_save(flags);
1254 __count_vm_events(PGFREE, 1 << order);
1255 free_one_page(page_zone(page), page, pfn, order, migratetype);
1256 local_irq_restore(flags);
1259 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1261 unsigned int nr_pages = 1 << order;
1262 struct page *p = page;
1266 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1268 __ClearPageReserved(p);
1269 set_page_count(p, 0);
1271 __ClearPageReserved(p);
1272 set_page_count(p, 0);
1274 page_zone(page)->managed_pages += nr_pages;
1275 set_page_refcounted(page);
1276 __free_pages(page, order);
1279 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1280 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1282 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1284 int __meminit early_pfn_to_nid(unsigned long pfn)
1286 static DEFINE_SPINLOCK(early_pfn_lock);
1289 spin_lock(&early_pfn_lock);
1290 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1292 nid = first_online_node;
1293 spin_unlock(&early_pfn_lock);
1299 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1300 static inline bool __meminit __maybe_unused
1301 meminit_pfn_in_nid(unsigned long pfn, int node,
1302 struct mminit_pfnnid_cache *state)
1306 nid = __early_pfn_to_nid(pfn, state);
1307 if (nid >= 0 && nid != node)
1312 /* Only safe to use early in boot when initialisation is single-threaded */
1313 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1315 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1320 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1324 static inline bool __meminit __maybe_unused
1325 meminit_pfn_in_nid(unsigned long pfn, int node,
1326 struct mminit_pfnnid_cache *state)
1333 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1336 if (early_page_uninitialised(pfn))
1338 return __free_pages_boot_core(page, order);
1342 * Check that the whole (or subset of) a pageblock given by the interval of
1343 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1344 * with the migration of free compaction scanner. The scanners then need to
1345 * use only pfn_valid_within() check for arches that allow holes within
1348 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1350 * It's possible on some configurations to have a setup like node0 node1 node0
1351 * i.e. it's possible that all pages within a zones range of pages do not
1352 * belong to a single zone. We assume that a border between node0 and node1
1353 * can occur within a single pageblock, but not a node0 node1 node0
1354 * interleaving within a single pageblock. It is therefore sufficient to check
1355 * the first and last page of a pageblock and avoid checking each individual
1356 * page in a pageblock.
1358 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1359 unsigned long end_pfn, struct zone *zone)
1361 struct page *start_page;
1362 struct page *end_page;
1364 /* end_pfn is one past the range we are checking */
1367 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1370 start_page = pfn_to_online_page(start_pfn);
1374 if (page_zone(start_page) != zone)
1377 end_page = pfn_to_page(end_pfn);
1379 /* This gives a shorter code than deriving page_zone(end_page) */
1380 if (page_zone_id(start_page) != page_zone_id(end_page))
1386 void set_zone_contiguous(struct zone *zone)
1388 unsigned long block_start_pfn = zone->zone_start_pfn;
1389 unsigned long block_end_pfn;
1391 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1392 for (; block_start_pfn < zone_end_pfn(zone);
1393 block_start_pfn = block_end_pfn,
1394 block_end_pfn += pageblock_nr_pages) {
1396 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1398 if (!__pageblock_pfn_to_page(block_start_pfn,
1399 block_end_pfn, zone))
1403 /* We confirm that there is no hole */
1404 zone->contiguous = true;
1407 void clear_zone_contiguous(struct zone *zone)
1409 zone->contiguous = false;
1412 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1413 static void __init deferred_free_range(struct page *page,
1414 unsigned long pfn, int nr_pages)
1421 /* Free a large naturally-aligned chunk if possible */
1422 if (nr_pages == pageblock_nr_pages &&
1423 (pfn & (pageblock_nr_pages - 1)) == 0) {
1424 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1425 __free_pages_boot_core(page, pageblock_order);
1429 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1430 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1431 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1432 __free_pages_boot_core(page, 0);
1436 /* Completion tracking for deferred_init_memmap() threads */
1437 static atomic_t pgdat_init_n_undone __initdata;
1438 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1440 static inline void __init pgdat_init_report_one_done(void)
1442 if (atomic_dec_and_test(&pgdat_init_n_undone))
1443 complete(&pgdat_init_all_done_comp);
1446 /* Initialise remaining memory on a node */
1447 static int __init deferred_init_memmap(void *data)
1449 pg_data_t *pgdat = data;
1450 int nid = pgdat->node_id;
1451 struct mminit_pfnnid_cache nid_init_state = { };
1452 unsigned long start = jiffies;
1453 unsigned long nr_pages = 0;
1454 unsigned long walk_start, walk_end;
1457 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1458 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1460 if (first_init_pfn == ULONG_MAX) {
1461 pgdat_init_report_one_done();
1465 /* Bind memory initialisation thread to a local node if possible */
1466 if (!cpumask_empty(cpumask))
1467 set_cpus_allowed_ptr(current, cpumask);
1469 /* Sanity check boundaries */
1470 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1471 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1472 pgdat->first_deferred_pfn = ULONG_MAX;
1474 /* Only the highest zone is deferred so find it */
1475 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1476 zone = pgdat->node_zones + zid;
1477 if (first_init_pfn < zone_end_pfn(zone))
1481 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1482 unsigned long pfn, end_pfn;
1483 struct page *page = NULL;
1484 struct page *free_base_page = NULL;
1485 unsigned long free_base_pfn = 0;
1488 end_pfn = min(walk_end, zone_end_pfn(zone));
1489 pfn = first_init_pfn;
1490 if (pfn < walk_start)
1492 if (pfn < zone->zone_start_pfn)
1493 pfn = zone->zone_start_pfn;
1495 for (; pfn < end_pfn; pfn++) {
1496 if (!pfn_valid_within(pfn))
1500 * Ensure pfn_valid is checked every
1501 * pageblock_nr_pages for memory holes
1503 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1504 if (!pfn_valid(pfn)) {
1510 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1515 /* Minimise pfn page lookups and scheduler checks */
1516 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1519 nr_pages += nr_to_free;
1520 deferred_free_range(free_base_page,
1521 free_base_pfn, nr_to_free);
1522 free_base_page = NULL;
1523 free_base_pfn = nr_to_free = 0;
1525 page = pfn_to_page(pfn);
1530 VM_BUG_ON(page_zone(page) != zone);
1534 __init_single_page(page, pfn, zid, nid);
1535 if (!free_base_page) {
1536 free_base_page = page;
1537 free_base_pfn = pfn;
1542 /* Where possible, batch up pages for a single free */
1545 /* Free the current block of pages to allocator */
1546 nr_pages += nr_to_free;
1547 deferred_free_range(free_base_page, free_base_pfn,
1549 free_base_page = NULL;
1550 free_base_pfn = nr_to_free = 0;
1552 /* Free the last block of pages to allocator */
1553 nr_pages += nr_to_free;
1554 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1556 first_init_pfn = max(end_pfn, first_init_pfn);
1559 /* Sanity check that the next zone really is unpopulated */
1560 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1562 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1563 jiffies_to_msecs(jiffies - start));
1565 pgdat_init_report_one_done();
1568 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1570 void __init page_alloc_init_late(void)
1574 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1577 /* There will be num_node_state(N_MEMORY) threads */
1578 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1579 for_each_node_state(nid, N_MEMORY) {
1580 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1583 /* Block until all are initialised */
1584 wait_for_completion(&pgdat_init_all_done_comp);
1586 /* Reinit limits that are based on free pages after the kernel is up */
1587 files_maxfiles_init();
1589 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1590 /* Discard memblock private memory */
1594 for_each_populated_zone(zone)
1595 set_zone_contiguous(zone);
1599 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1600 void __init init_cma_reserved_pageblock(struct page *page)
1602 unsigned i = pageblock_nr_pages;
1603 struct page *p = page;
1606 __ClearPageReserved(p);
1607 set_page_count(p, 0);
1610 set_pageblock_migratetype(page, MIGRATE_CMA);
1612 if (pageblock_order >= MAX_ORDER) {
1613 i = pageblock_nr_pages;
1616 set_page_refcounted(p);
1617 __free_pages(p, MAX_ORDER - 1);
1618 p += MAX_ORDER_NR_PAGES;
1619 } while (i -= MAX_ORDER_NR_PAGES);
1621 set_page_refcounted(page);
1622 __free_pages(page, pageblock_order);
1625 adjust_managed_page_count(page, pageblock_nr_pages);
1630 * The order of subdivision here is critical for the IO subsystem.
1631 * Please do not alter this order without good reasons and regression
1632 * testing. Specifically, as large blocks of memory are subdivided,
1633 * the order in which smaller blocks are delivered depends on the order
1634 * they're subdivided in this function. This is the primary factor
1635 * influencing the order in which pages are delivered to the IO
1636 * subsystem according to empirical testing, and this is also justified
1637 * by considering the behavior of a buddy system containing a single
1638 * large block of memory acted on by a series of small allocations.
1639 * This behavior is a critical factor in sglist merging's success.
1643 static inline void expand(struct zone *zone, struct page *page,
1644 int low, int high, struct free_area *area,
1647 unsigned long size = 1 << high;
1649 while (high > low) {
1653 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1656 * Mark as guard pages (or page), that will allow to
1657 * merge back to allocator when buddy will be freed.
1658 * Corresponding page table entries will not be touched,
1659 * pages will stay not present in virtual address space
1661 if (set_page_guard(zone, &page[size], high, migratetype))
1664 list_add(&page[size].lru, &area->free_list[migratetype]);
1666 set_page_order(&page[size], high);
1670 static void check_new_page_bad(struct page *page)
1672 const char *bad_reason = NULL;
1673 unsigned long bad_flags = 0;
1675 if (unlikely(atomic_read(&page->_mapcount) != -1))
1676 bad_reason = "nonzero mapcount";
1677 if (unlikely(page->mapping != NULL))
1678 bad_reason = "non-NULL mapping";
1679 if (unlikely(page_ref_count(page) != 0))
1680 bad_reason = "nonzero _count";
1681 if (unlikely(page->flags & __PG_HWPOISON)) {
1682 bad_reason = "HWPoisoned (hardware-corrupted)";
1683 bad_flags = __PG_HWPOISON;
1684 /* Don't complain about hwpoisoned pages */
1685 page_mapcount_reset(page); /* remove PageBuddy */
1688 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1689 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1690 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1693 if (unlikely(page->mem_cgroup))
1694 bad_reason = "page still charged to cgroup";
1696 bad_page(page, bad_reason, bad_flags);
1700 * This page is about to be returned from the page allocator
1702 static inline int check_new_page(struct page *page)
1704 if (likely(page_expected_state(page,
1705 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1708 check_new_page_bad(page);
1712 static inline bool free_pages_prezeroed(void)
1714 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1715 page_poisoning_enabled();
1718 #ifdef CONFIG_DEBUG_VM
1719 static bool check_pcp_refill(struct page *page)
1724 static bool check_new_pcp(struct page *page)
1726 return check_new_page(page);
1729 static bool check_pcp_refill(struct page *page)
1731 return check_new_page(page);
1733 static bool check_new_pcp(struct page *page)
1737 #endif /* CONFIG_DEBUG_VM */
1739 static bool check_new_pages(struct page *page, unsigned int order)
1742 for (i = 0; i < (1 << order); i++) {
1743 struct page *p = page + i;
1745 if (unlikely(check_new_page(p)))
1752 inline void post_alloc_hook(struct page *page, unsigned int order,
1755 set_page_private(page, 0);
1756 set_page_refcounted(page);
1758 arch_alloc_page(page, order);
1759 kernel_map_pages(page, 1 << order, 1);
1760 kernel_poison_pages(page, 1 << order, 1);
1761 kasan_alloc_pages(page, order);
1762 set_page_owner(page, order, gfp_flags);
1765 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1766 unsigned int alloc_flags)
1770 post_alloc_hook(page, order, gfp_flags);
1772 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1773 for (i = 0; i < (1 << order); i++)
1774 clear_highpage(page + i);
1776 if (order && (gfp_flags & __GFP_COMP))
1777 prep_compound_page(page, order);
1780 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1781 * allocate the page. The expectation is that the caller is taking
1782 * steps that will free more memory. The caller should avoid the page
1783 * being used for !PFMEMALLOC purposes.
1785 if (alloc_flags & ALLOC_NO_WATERMARKS)
1786 set_page_pfmemalloc(page);
1788 clear_page_pfmemalloc(page);
1792 * Go through the free lists for the given migratetype and remove
1793 * the smallest available page from the freelists
1796 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1799 unsigned int current_order;
1800 struct free_area *area;
1803 /* Find a page of the appropriate size in the preferred list */
1804 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1805 area = &(zone->free_area[current_order]);
1806 page = list_first_entry_or_null(&area->free_list[migratetype],
1810 list_del(&page->lru);
1811 rmv_page_order(page);
1813 expand(zone, page, order, current_order, area, migratetype);
1814 set_pcppage_migratetype(page, migratetype);
1823 * This array describes the order lists are fallen back to when
1824 * the free lists for the desirable migrate type are depleted
1826 static int fallbacks[MIGRATE_TYPES][4] = {
1827 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1828 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1829 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1831 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1833 #ifdef CONFIG_MEMORY_ISOLATION
1834 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1839 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1842 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1845 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1846 unsigned int order) { return NULL; }
1850 * Move the free pages in a range to the free lists of the requested type.
1851 * Note that start_page and end_pages are not aligned on a pageblock
1852 * boundary. If alignment is required, use move_freepages_block()
1854 static int move_freepages(struct zone *zone,
1855 struct page *start_page, struct page *end_page,
1856 int migratetype, int *num_movable)
1860 int pages_moved = 0;
1862 #ifndef CONFIG_HOLES_IN_ZONE
1864 * page_zone is not safe to call in this context when
1865 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1866 * anyway as we check zone boundaries in move_freepages_block().
1867 * Remove at a later date when no bug reports exist related to
1868 * grouping pages by mobility
1870 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1876 for (page = start_page; page <= end_page;) {
1877 if (!pfn_valid_within(page_to_pfn(page))) {
1882 /* Make sure we are not inadvertently changing nodes */
1883 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1885 if (!PageBuddy(page)) {
1887 * We assume that pages that could be isolated for
1888 * migration are movable. But we don't actually try
1889 * isolating, as that would be expensive.
1892 (PageLRU(page) || __PageMovable(page)))
1899 order = page_order(page);
1900 list_move(&page->lru,
1901 &zone->free_area[order].free_list[migratetype]);
1903 pages_moved += 1 << order;
1909 int move_freepages_block(struct zone *zone, struct page *page,
1910 int migratetype, int *num_movable)
1912 unsigned long start_pfn, end_pfn;
1913 struct page *start_page, *end_page;
1915 start_pfn = page_to_pfn(page);
1916 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1917 start_page = pfn_to_page(start_pfn);
1918 end_page = start_page + pageblock_nr_pages - 1;
1919 end_pfn = start_pfn + pageblock_nr_pages - 1;
1921 /* Do not cross zone boundaries */
1922 if (!zone_spans_pfn(zone, start_pfn))
1924 if (!zone_spans_pfn(zone, end_pfn))
1927 return move_freepages(zone, start_page, end_page, migratetype,
1931 static void change_pageblock_range(struct page *pageblock_page,
1932 int start_order, int migratetype)
1934 int nr_pageblocks = 1 << (start_order - pageblock_order);
1936 while (nr_pageblocks--) {
1937 set_pageblock_migratetype(pageblock_page, migratetype);
1938 pageblock_page += pageblock_nr_pages;
1943 * When we are falling back to another migratetype during allocation, try to
1944 * steal extra free pages from the same pageblocks to satisfy further
1945 * allocations, instead of polluting multiple pageblocks.
1947 * If we are stealing a relatively large buddy page, it is likely there will
1948 * be more free pages in the pageblock, so try to steal them all. For
1949 * reclaimable and unmovable allocations, we steal regardless of page size,
1950 * as fragmentation caused by those allocations polluting movable pageblocks
1951 * is worse than movable allocations stealing from unmovable and reclaimable
1954 static bool can_steal_fallback(unsigned int order, int start_mt)
1957 * Leaving this order check is intended, although there is
1958 * relaxed order check in next check. The reason is that
1959 * we can actually steal whole pageblock if this condition met,
1960 * but, below check doesn't guarantee it and that is just heuristic
1961 * so could be changed anytime.
1963 if (order >= pageblock_order)
1966 if (order >= pageblock_order / 2 ||
1967 start_mt == MIGRATE_RECLAIMABLE ||
1968 start_mt == MIGRATE_UNMOVABLE ||
1969 page_group_by_mobility_disabled)
1976 * This function implements actual steal behaviour. If order is large enough,
1977 * we can steal whole pageblock. If not, we first move freepages in this
1978 * pageblock to our migratetype and determine how many already-allocated pages
1979 * are there in the pageblock with a compatible migratetype. If at least half
1980 * of pages are free or compatible, we can change migratetype of the pageblock
1981 * itself, so pages freed in the future will be put on the correct free list.
1983 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1984 int start_type, bool whole_block)
1986 unsigned int current_order = page_order(page);
1987 struct free_area *area;
1988 int free_pages, movable_pages, alike_pages;
1991 old_block_type = get_pageblock_migratetype(page);
1994 * This can happen due to races and we want to prevent broken
1995 * highatomic accounting.
1997 if (is_migrate_highatomic(old_block_type))
2000 /* Take ownership for orders >= pageblock_order */
2001 if (current_order >= pageblock_order) {
2002 change_pageblock_range(page, current_order, start_type);
2006 /* We are not allowed to try stealing from the whole block */
2010 free_pages = move_freepages_block(zone, page, start_type,
2013 * Determine how many pages are compatible with our allocation.
2014 * For movable allocation, it's the number of movable pages which
2015 * we just obtained. For other types it's a bit more tricky.
2017 if (start_type == MIGRATE_MOVABLE) {
2018 alike_pages = movable_pages;
2021 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2022 * to MOVABLE pageblock, consider all non-movable pages as
2023 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2024 * vice versa, be conservative since we can't distinguish the
2025 * exact migratetype of non-movable pages.
2027 if (old_block_type == MIGRATE_MOVABLE)
2028 alike_pages = pageblock_nr_pages
2029 - (free_pages + movable_pages);
2034 /* moving whole block can fail due to zone boundary conditions */
2039 * If a sufficient number of pages in the block are either free or of
2040 * comparable migratability as our allocation, claim the whole block.
2042 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2043 page_group_by_mobility_disabled)
2044 set_pageblock_migratetype(page, start_type);
2049 area = &zone->free_area[current_order];
2050 list_move(&page->lru, &area->free_list[start_type]);
2054 * Check whether there is a suitable fallback freepage with requested order.
2055 * If only_stealable is true, this function returns fallback_mt only if
2056 * we can steal other freepages all together. This would help to reduce
2057 * fragmentation due to mixed migratetype pages in one pageblock.
2059 int find_suitable_fallback(struct free_area *area, unsigned int order,
2060 int migratetype, bool only_stealable, bool *can_steal)
2065 if (area->nr_free == 0)
2070 fallback_mt = fallbacks[migratetype][i];
2071 if (fallback_mt == MIGRATE_TYPES)
2074 if (list_empty(&area->free_list[fallback_mt]))
2077 if (can_steal_fallback(order, migratetype))
2080 if (!only_stealable)
2091 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2092 * there are no empty page blocks that contain a page with a suitable order
2094 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2095 unsigned int alloc_order)
2098 unsigned long max_managed, flags;
2101 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2102 * Check is race-prone but harmless.
2104 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2105 if (zone->nr_reserved_highatomic >= max_managed)
2108 spin_lock_irqsave(&zone->lock, flags);
2110 /* Recheck the nr_reserved_highatomic limit under the lock */
2111 if (zone->nr_reserved_highatomic >= max_managed)
2115 mt = get_pageblock_migratetype(page);
2116 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2117 && !is_migrate_cma(mt)) {
2118 zone->nr_reserved_highatomic += pageblock_nr_pages;
2119 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2120 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2124 spin_unlock_irqrestore(&zone->lock, flags);
2128 * Used when an allocation is about to fail under memory pressure. This
2129 * potentially hurts the reliability of high-order allocations when under
2130 * intense memory pressure but failed atomic allocations should be easier
2131 * to recover from than an OOM.
2133 * If @force is true, try to unreserve a pageblock even though highatomic
2134 * pageblock is exhausted.
2136 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2139 struct zonelist *zonelist = ac->zonelist;
2140 unsigned long flags;
2147 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2150 * Preserve at least one pageblock unless memory pressure
2153 if (!force && zone->nr_reserved_highatomic <=
2157 spin_lock_irqsave(&zone->lock, flags);
2158 for (order = 0; order < MAX_ORDER; order++) {
2159 struct free_area *area = &(zone->free_area[order]);
2161 page = list_first_entry_or_null(
2162 &area->free_list[MIGRATE_HIGHATOMIC],
2168 * In page freeing path, migratetype change is racy so
2169 * we can counter several free pages in a pageblock
2170 * in this loop althoug we changed the pageblock type
2171 * from highatomic to ac->migratetype. So we should
2172 * adjust the count once.
2174 if (is_migrate_highatomic_page(page)) {
2176 * It should never happen but changes to
2177 * locking could inadvertently allow a per-cpu
2178 * drain to add pages to MIGRATE_HIGHATOMIC
2179 * while unreserving so be safe and watch for
2182 zone->nr_reserved_highatomic -= min(
2184 zone->nr_reserved_highatomic);
2188 * Convert to ac->migratetype and avoid the normal
2189 * pageblock stealing heuristics. Minimally, the caller
2190 * is doing the work and needs the pages. More
2191 * importantly, if the block was always converted to
2192 * MIGRATE_UNMOVABLE or another type then the number
2193 * of pageblocks that cannot be completely freed
2196 set_pageblock_migratetype(page, ac->migratetype);
2197 ret = move_freepages_block(zone, page, ac->migratetype,
2200 spin_unlock_irqrestore(&zone->lock, flags);
2204 spin_unlock_irqrestore(&zone->lock, flags);
2211 * Try finding a free buddy page on the fallback list and put it on the free
2212 * list of requested migratetype, possibly along with other pages from the same
2213 * block, depending on fragmentation avoidance heuristics. Returns true if
2214 * fallback was found so that __rmqueue_smallest() can grab it.
2216 * The use of signed ints for order and current_order is a deliberate
2217 * deviation from the rest of this file, to make the for loop
2218 * condition simpler.
2221 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2223 struct free_area *area;
2230 * Find the largest available free page in the other list. This roughly
2231 * approximates finding the pageblock with the most free pages, which
2232 * would be too costly to do exactly.
2234 for (current_order = MAX_ORDER - 1; current_order >= order;
2236 area = &(zone->free_area[current_order]);
2237 fallback_mt = find_suitable_fallback(area, current_order,
2238 start_migratetype, false, &can_steal);
2239 if (fallback_mt == -1)
2243 * We cannot steal all free pages from the pageblock and the
2244 * requested migratetype is movable. In that case it's better to
2245 * steal and split the smallest available page instead of the
2246 * largest available page, because even if the next movable
2247 * allocation falls back into a different pageblock than this
2248 * one, it won't cause permanent fragmentation.
2250 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2251 && current_order > order)
2260 for (current_order = order; current_order < MAX_ORDER;
2262 area = &(zone->free_area[current_order]);
2263 fallback_mt = find_suitable_fallback(area, current_order,
2264 start_migratetype, false, &can_steal);
2265 if (fallback_mt != -1)
2270 * This should not happen - we already found a suitable fallback
2271 * when looking for the largest page.
2273 VM_BUG_ON(current_order == MAX_ORDER);
2276 page = list_first_entry(&area->free_list[fallback_mt],
2279 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2281 trace_mm_page_alloc_extfrag(page, order, current_order,
2282 start_migratetype, fallback_mt);
2289 * Do the hard work of removing an element from the buddy allocator.
2290 * Call me with the zone->lock already held.
2292 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2298 page = __rmqueue_smallest(zone, order, migratetype);
2299 if (unlikely(!page)) {
2300 if (migratetype == MIGRATE_MOVABLE)
2301 page = __rmqueue_cma_fallback(zone, order);
2303 if (!page && __rmqueue_fallback(zone, order, migratetype))
2307 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2312 * Obtain a specified number of elements from the buddy allocator, all under
2313 * a single hold of the lock, for efficiency. Add them to the supplied list.
2314 * Returns the number of new pages which were placed at *list.
2316 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2317 unsigned long count, struct list_head *list,
2318 int migratetype, bool cold)
2322 spin_lock(&zone->lock);
2323 for (i = 0; i < count; ++i) {
2324 struct page *page = __rmqueue(zone, order, migratetype);
2325 if (unlikely(page == NULL))
2328 if (unlikely(check_pcp_refill(page)))
2332 * Split buddy pages returned by expand() are received here
2333 * in physical page order. The page is added to the callers and
2334 * list and the list head then moves forward. From the callers
2335 * perspective, the linked list is ordered by page number in
2336 * some conditions. This is useful for IO devices that can
2337 * merge IO requests if the physical pages are ordered
2341 list_add(&page->lru, list);
2343 list_add_tail(&page->lru, list);
2346 if (is_migrate_cma(get_pcppage_migratetype(page)))
2347 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2352 * i pages were removed from the buddy list even if some leak due
2353 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2354 * on i. Do not confuse with 'alloced' which is the number of
2355 * pages added to the pcp list.
2357 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2358 spin_unlock(&zone->lock);
2364 * Called from the vmstat counter updater to drain pagesets of this
2365 * currently executing processor on remote nodes after they have
2368 * Note that this function must be called with the thread pinned to
2369 * a single processor.
2371 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2373 unsigned long flags;
2374 int to_drain, batch;
2376 local_irq_save(flags);
2377 batch = READ_ONCE(pcp->batch);
2378 to_drain = min(pcp->count, batch);
2380 free_pcppages_bulk(zone, to_drain, pcp);
2381 pcp->count -= to_drain;
2383 local_irq_restore(flags);
2388 * Drain pcplists of the indicated processor and zone.
2390 * The processor must either be the current processor and the
2391 * thread pinned to the current processor or a processor that
2394 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2396 unsigned long flags;
2397 struct per_cpu_pageset *pset;
2398 struct per_cpu_pages *pcp;
2400 local_irq_save(flags);
2401 pset = per_cpu_ptr(zone->pageset, cpu);
2405 free_pcppages_bulk(zone, pcp->count, pcp);
2408 local_irq_restore(flags);
2412 * Drain pcplists of all zones on the indicated processor.
2414 * The processor must either be the current processor and the
2415 * thread pinned to the current processor or a processor that
2418 static void drain_pages(unsigned int cpu)
2422 for_each_populated_zone(zone) {
2423 drain_pages_zone(cpu, zone);
2428 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2430 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2431 * the single zone's pages.
2433 void drain_local_pages(struct zone *zone)
2435 int cpu = smp_processor_id();
2438 drain_pages_zone(cpu, zone);
2443 static void drain_local_pages_wq(struct work_struct *work)
2446 * drain_all_pages doesn't use proper cpu hotplug protection so
2447 * we can race with cpu offline when the WQ can move this from
2448 * a cpu pinned worker to an unbound one. We can operate on a different
2449 * cpu which is allright but we also have to make sure to not move to
2453 drain_local_pages(NULL);
2458 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2460 * When zone parameter is non-NULL, spill just the single zone's pages.
2462 * Note that this can be extremely slow as the draining happens in a workqueue.
2464 void drain_all_pages(struct zone *zone)
2469 * Allocate in the BSS so we wont require allocation in
2470 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2472 static cpumask_t cpus_with_pcps;
2475 * Make sure nobody triggers this path before mm_percpu_wq is fully
2478 if (WARN_ON_ONCE(!mm_percpu_wq))
2481 /* Workqueues cannot recurse */
2482 if (current->flags & PF_WQ_WORKER)
2486 * Do not drain if one is already in progress unless it's specific to
2487 * a zone. Such callers are primarily CMA and memory hotplug and need
2488 * the drain to be complete when the call returns.
2490 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2493 mutex_lock(&pcpu_drain_mutex);
2497 * We don't care about racing with CPU hotplug event
2498 * as offline notification will cause the notified
2499 * cpu to drain that CPU pcps and on_each_cpu_mask
2500 * disables preemption as part of its processing
2502 for_each_online_cpu(cpu) {
2503 struct per_cpu_pageset *pcp;
2505 bool has_pcps = false;
2508 pcp = per_cpu_ptr(zone->pageset, cpu);
2512 for_each_populated_zone(z) {
2513 pcp = per_cpu_ptr(z->pageset, cpu);
2514 if (pcp->pcp.count) {
2522 cpumask_set_cpu(cpu, &cpus_with_pcps);
2524 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2527 for_each_cpu(cpu, &cpus_with_pcps) {
2528 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2529 INIT_WORK(work, drain_local_pages_wq);
2530 queue_work_on(cpu, mm_percpu_wq, work);
2532 for_each_cpu(cpu, &cpus_with_pcps)
2533 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2535 mutex_unlock(&pcpu_drain_mutex);
2538 #ifdef CONFIG_HIBERNATION
2541 * Touch the watchdog for every WD_PAGE_COUNT pages.
2543 #define WD_PAGE_COUNT (128*1024)
2545 void mark_free_pages(struct zone *zone)
2547 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2548 unsigned long flags;
2549 unsigned int order, t;
2552 if (zone_is_empty(zone))
2555 spin_lock_irqsave(&zone->lock, flags);
2557 max_zone_pfn = zone_end_pfn(zone);
2558 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2559 if (pfn_valid(pfn)) {
2560 page = pfn_to_page(pfn);
2562 if (!--page_count) {
2563 touch_nmi_watchdog();
2564 page_count = WD_PAGE_COUNT;
2567 if (page_zone(page) != zone)
2570 if (!swsusp_page_is_forbidden(page))
2571 swsusp_unset_page_free(page);
2574 for_each_migratetype_order(order, t) {
2575 list_for_each_entry(page,
2576 &zone->free_area[order].free_list[t], lru) {
2579 pfn = page_to_pfn(page);
2580 for (i = 0; i < (1UL << order); i++) {
2581 if (!--page_count) {
2582 touch_nmi_watchdog();
2583 page_count = WD_PAGE_COUNT;
2585 swsusp_set_page_free(pfn_to_page(pfn + i));
2589 spin_unlock_irqrestore(&zone->lock, flags);
2591 #endif /* CONFIG_PM */
2594 * Free a 0-order page
2595 * cold == true ? free a cold page : free a hot page
2597 void free_hot_cold_page(struct page *page, bool cold)
2599 struct zone *zone = page_zone(page);
2600 struct per_cpu_pages *pcp;
2601 unsigned long flags;
2602 unsigned long pfn = page_to_pfn(page);
2605 if (!free_pcp_prepare(page))
2608 migratetype = get_pfnblock_migratetype(page, pfn);
2609 set_pcppage_migratetype(page, migratetype);
2610 local_irq_save(flags);
2611 __count_vm_event(PGFREE);
2614 * We only track unmovable, reclaimable and movable on pcp lists.
2615 * Free ISOLATE pages back to the allocator because they are being
2616 * offlined but treat HIGHATOMIC as movable pages so we can get those
2617 * areas back if necessary. Otherwise, we may have to free
2618 * excessively into the page allocator
2620 if (migratetype >= MIGRATE_PCPTYPES) {
2621 if (unlikely(is_migrate_isolate(migratetype))) {
2622 free_one_page(zone, page, pfn, 0, migratetype);
2625 migratetype = MIGRATE_MOVABLE;
2628 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2630 list_add(&page->lru, &pcp->lists[migratetype]);
2632 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2634 if (pcp->count >= pcp->high) {
2635 unsigned long batch = READ_ONCE(pcp->batch);
2636 free_pcppages_bulk(zone, batch, pcp);
2637 pcp->count -= batch;
2641 local_irq_restore(flags);
2645 * Free a list of 0-order pages
2647 void free_hot_cold_page_list(struct list_head *list, bool cold)
2649 struct page *page, *next;
2651 list_for_each_entry_safe(page, next, list, lru) {
2652 trace_mm_page_free_batched(page, cold);
2653 free_hot_cold_page(page, cold);
2658 * split_page takes a non-compound higher-order page, and splits it into
2659 * n (1<<order) sub-pages: page[0..n]
2660 * Each sub-page must be freed individually.
2662 * Note: this is probably too low level an operation for use in drivers.
2663 * Please consult with lkml before using this in your driver.
2665 void split_page(struct page *page, unsigned int order)
2669 VM_BUG_ON_PAGE(PageCompound(page), page);
2670 VM_BUG_ON_PAGE(!page_count(page), page);
2672 #ifdef CONFIG_KMEMCHECK
2674 * Split shadow pages too, because free(page[0]) would
2675 * otherwise free the whole shadow.
2677 if (kmemcheck_page_is_tracked(page))
2678 split_page(virt_to_page(page[0].shadow), order);
2681 for (i = 1; i < (1 << order); i++)
2682 set_page_refcounted(page + i);
2683 split_page_owner(page, order);
2685 EXPORT_SYMBOL_GPL(split_page);
2687 int __isolate_free_page(struct page *page, unsigned int order)
2689 unsigned long watermark;
2693 BUG_ON(!PageBuddy(page));
2695 zone = page_zone(page);
2696 mt = get_pageblock_migratetype(page);
2698 if (!is_migrate_isolate(mt)) {
2700 * Obey watermarks as if the page was being allocated. We can
2701 * emulate a high-order watermark check with a raised order-0
2702 * watermark, because we already know our high-order page
2705 watermark = min_wmark_pages(zone) + (1UL << order);
2706 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2709 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2712 /* Remove page from free list */
2713 list_del(&page->lru);
2714 zone->free_area[order].nr_free--;
2715 rmv_page_order(page);
2718 * Set the pageblock if the isolated page is at least half of a
2721 if (order >= pageblock_order - 1) {
2722 struct page *endpage = page + (1 << order) - 1;
2723 for (; page < endpage; page += pageblock_nr_pages) {
2724 int mt = get_pageblock_migratetype(page);
2725 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2726 && !is_migrate_highatomic(mt))
2727 set_pageblock_migratetype(page,
2733 return 1UL << order;
2737 * Update NUMA hit/miss statistics
2739 * Must be called with interrupts disabled.
2741 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2744 enum numa_stat_item local_stat = NUMA_LOCAL;
2746 if (z->node != numa_node_id())
2747 local_stat = NUMA_OTHER;
2749 if (z->node == preferred_zone->node)
2750 __inc_numa_state(z, NUMA_HIT);
2752 __inc_numa_state(z, NUMA_MISS);
2753 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2755 __inc_numa_state(z, local_stat);
2759 /* Remove page from the per-cpu list, caller must protect the list */
2760 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2761 bool cold, struct per_cpu_pages *pcp,
2762 struct list_head *list)
2767 if (list_empty(list)) {
2768 pcp->count += rmqueue_bulk(zone, 0,
2771 if (unlikely(list_empty(list)))
2776 page = list_last_entry(list, struct page, lru);
2778 page = list_first_entry(list, struct page, lru);
2780 list_del(&page->lru);
2782 } while (check_new_pcp(page));
2787 /* Lock and remove page from the per-cpu list */
2788 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2789 struct zone *zone, unsigned int order,
2790 gfp_t gfp_flags, int migratetype)
2792 struct per_cpu_pages *pcp;
2793 struct list_head *list;
2794 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2796 unsigned long flags;
2798 local_irq_save(flags);
2799 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2800 list = &pcp->lists[migratetype];
2801 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2803 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2804 zone_statistics(preferred_zone, zone);
2806 local_irq_restore(flags);
2811 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2814 struct page *rmqueue(struct zone *preferred_zone,
2815 struct zone *zone, unsigned int order,
2816 gfp_t gfp_flags, unsigned int alloc_flags,
2819 unsigned long flags;
2822 if (likely(order == 0)) {
2823 page = rmqueue_pcplist(preferred_zone, zone, order,
2824 gfp_flags, migratetype);
2829 * We most definitely don't want callers attempting to
2830 * allocate greater than order-1 page units with __GFP_NOFAIL.
2832 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2833 spin_lock_irqsave(&zone->lock, flags);
2837 if (alloc_flags & ALLOC_HARDER) {
2838 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2840 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2843 page = __rmqueue(zone, order, migratetype);
2844 } while (page && check_new_pages(page, order));
2845 spin_unlock(&zone->lock);
2848 __mod_zone_freepage_state(zone, -(1 << order),
2849 get_pcppage_migratetype(page));
2851 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2852 zone_statistics(preferred_zone, zone);
2853 local_irq_restore(flags);
2856 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2860 local_irq_restore(flags);
2864 #ifdef CONFIG_FAIL_PAGE_ALLOC
2867 struct fault_attr attr;
2869 bool ignore_gfp_highmem;
2870 bool ignore_gfp_reclaim;
2872 } fail_page_alloc = {
2873 .attr = FAULT_ATTR_INITIALIZER,
2874 .ignore_gfp_reclaim = true,
2875 .ignore_gfp_highmem = true,
2879 static int __init setup_fail_page_alloc(char *str)
2881 return setup_fault_attr(&fail_page_alloc.attr, str);
2883 __setup("fail_page_alloc=", setup_fail_page_alloc);
2885 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2887 if (order < fail_page_alloc.min_order)
2889 if (gfp_mask & __GFP_NOFAIL)
2891 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2893 if (fail_page_alloc.ignore_gfp_reclaim &&
2894 (gfp_mask & __GFP_DIRECT_RECLAIM))
2897 return should_fail(&fail_page_alloc.attr, 1 << order);
2900 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2902 static int __init fail_page_alloc_debugfs(void)
2904 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2907 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2908 &fail_page_alloc.attr);
2910 return PTR_ERR(dir);
2912 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2913 &fail_page_alloc.ignore_gfp_reclaim))
2915 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2916 &fail_page_alloc.ignore_gfp_highmem))
2918 if (!debugfs_create_u32("min-order", mode, dir,
2919 &fail_page_alloc.min_order))
2924 debugfs_remove_recursive(dir);
2929 late_initcall(fail_page_alloc_debugfs);
2931 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2933 #else /* CONFIG_FAIL_PAGE_ALLOC */
2935 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2940 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2943 * Return true if free base pages are above 'mark'. For high-order checks it
2944 * will return true of the order-0 watermark is reached and there is at least
2945 * one free page of a suitable size. Checking now avoids taking the zone lock
2946 * to check in the allocation paths if no pages are free.
2948 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2949 int classzone_idx, unsigned int alloc_flags,
2954 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
2956 /* free_pages may go negative - that's OK */
2957 free_pages -= (1 << order) - 1;
2959 if (alloc_flags & ALLOC_HIGH)
2963 * If the caller does not have rights to ALLOC_HARDER then subtract
2964 * the high-atomic reserves. This will over-estimate the size of the
2965 * atomic reserve but it avoids a search.
2967 if (likely(!alloc_harder)) {
2968 free_pages -= z->nr_reserved_highatomic;
2971 * OOM victims can try even harder than normal ALLOC_HARDER
2972 * users on the grounds that it's definitely going to be in
2973 * the exit path shortly and free memory. Any allocation it
2974 * makes during the free path will be small and short-lived.
2976 if (alloc_flags & ALLOC_OOM)
2984 /* If allocation can't use CMA areas don't use free CMA pages */
2985 if (!(alloc_flags & ALLOC_CMA))
2986 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2990 * Check watermarks for an order-0 allocation request. If these
2991 * are not met, then a high-order request also cannot go ahead
2992 * even if a suitable page happened to be free.
2994 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2997 /* If this is an order-0 request then the watermark is fine */
3001 /* For a high-order request, check at least one suitable page is free */
3002 for (o = order; o < MAX_ORDER; o++) {
3003 struct free_area *area = &z->free_area[o];
3012 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3013 if (!list_empty(&area->free_list[mt]))
3018 if ((alloc_flags & ALLOC_CMA) &&
3019 !list_empty(&area->free_list[MIGRATE_CMA])) {
3027 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3028 int classzone_idx, unsigned int alloc_flags)
3030 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3031 zone_page_state(z, NR_FREE_PAGES));
3034 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3035 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3037 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3041 /* If allocation can't use CMA areas don't use free CMA pages */
3042 if (!(alloc_flags & ALLOC_CMA))
3043 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3047 * Fast check for order-0 only. If this fails then the reserves
3048 * need to be calculated. There is a corner case where the check
3049 * passes but only the high-order atomic reserve are free. If
3050 * the caller is !atomic then it'll uselessly search the free
3051 * list. That corner case is then slower but it is harmless.
3053 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3056 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3060 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3061 unsigned long mark, int classzone_idx)
3063 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3065 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3066 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3068 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3073 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3075 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3078 #else /* CONFIG_NUMA */
3079 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3083 #endif /* CONFIG_NUMA */
3086 * get_page_from_freelist goes through the zonelist trying to allocate
3089 static struct page *
3090 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3091 const struct alloc_context *ac)
3093 struct zoneref *z = ac->preferred_zoneref;
3095 struct pglist_data *last_pgdat_dirty_limit = NULL;
3098 * Scan zonelist, looking for a zone with enough free.
3099 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3101 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3106 if (cpusets_enabled() &&
3107 (alloc_flags & ALLOC_CPUSET) &&
3108 !__cpuset_zone_allowed(zone, gfp_mask))
3111 * When allocating a page cache page for writing, we
3112 * want to get it from a node that is within its dirty
3113 * limit, such that no single node holds more than its
3114 * proportional share of globally allowed dirty pages.
3115 * The dirty limits take into account the node's
3116 * lowmem reserves and high watermark so that kswapd
3117 * should be able to balance it without having to
3118 * write pages from its LRU list.
3120 * XXX: For now, allow allocations to potentially
3121 * exceed the per-node dirty limit in the slowpath
3122 * (spread_dirty_pages unset) before going into reclaim,
3123 * which is important when on a NUMA setup the allowed
3124 * nodes are together not big enough to reach the
3125 * global limit. The proper fix for these situations
3126 * will require awareness of nodes in the
3127 * dirty-throttling and the flusher threads.
3129 if (ac->spread_dirty_pages) {
3130 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3133 if (!node_dirty_ok(zone->zone_pgdat)) {
3134 last_pgdat_dirty_limit = zone->zone_pgdat;
3139 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3140 if (!zone_watermark_fast(zone, order, mark,
3141 ac_classzone_idx(ac), alloc_flags)) {
3144 /* Checked here to keep the fast path fast */
3145 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3146 if (alloc_flags & ALLOC_NO_WATERMARKS)
3149 if (node_reclaim_mode == 0 ||
3150 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3153 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3155 case NODE_RECLAIM_NOSCAN:
3158 case NODE_RECLAIM_FULL:
3159 /* scanned but unreclaimable */
3162 /* did we reclaim enough */
3163 if (zone_watermark_ok(zone, order, mark,
3164 ac_classzone_idx(ac), alloc_flags))
3172 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3173 gfp_mask, alloc_flags, ac->migratetype);
3175 prep_new_page(page, order, gfp_mask, alloc_flags);
3178 * If this is a high-order atomic allocation then check
3179 * if the pageblock should be reserved for the future
3181 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3182 reserve_highatomic_pageblock(page, zone, order);
3192 * Large machines with many possible nodes should not always dump per-node
3193 * meminfo in irq context.
3195 static inline bool should_suppress_show_mem(void)
3200 ret = in_interrupt();
3205 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3207 unsigned int filter = SHOW_MEM_FILTER_NODES;
3208 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3210 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3214 * This documents exceptions given to allocations in certain
3215 * contexts that are allowed to allocate outside current's set
3218 if (!(gfp_mask & __GFP_NOMEMALLOC))
3219 if (tsk_is_oom_victim(current) ||
3220 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3221 filter &= ~SHOW_MEM_FILTER_NODES;
3222 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3223 filter &= ~SHOW_MEM_FILTER_NODES;
3225 show_mem(filter, nodemask);
3228 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3230 struct va_format vaf;
3232 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3233 DEFAULT_RATELIMIT_BURST);
3235 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3238 pr_warn("%s: ", current->comm);
3240 va_start(args, fmt);
3243 pr_cont("%pV", &vaf);
3246 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3248 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3250 pr_cont("(null)\n");
3252 cpuset_print_current_mems_allowed();
3255 warn_alloc_show_mem(gfp_mask, nodemask);
3258 static inline struct page *
3259 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3260 unsigned int alloc_flags,
3261 const struct alloc_context *ac)
3265 page = get_page_from_freelist(gfp_mask, order,
3266 alloc_flags|ALLOC_CPUSET, ac);
3268 * fallback to ignore cpuset restriction if our nodes
3272 page = get_page_from_freelist(gfp_mask, order,
3278 static inline struct page *
3279 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3280 const struct alloc_context *ac, unsigned long *did_some_progress)
3282 struct oom_control oc = {
3283 .zonelist = ac->zonelist,
3284 .nodemask = ac->nodemask,
3286 .gfp_mask = gfp_mask,
3291 *did_some_progress = 0;
3294 * Acquire the oom lock. If that fails, somebody else is
3295 * making progress for us.
3297 if (!mutex_trylock(&oom_lock)) {
3298 *did_some_progress = 1;
3299 schedule_timeout_uninterruptible(1);
3304 * Go through the zonelist yet one more time, keep very high watermark
3305 * here, this is only to catch a parallel oom killing, we must fail if
3306 * we're still under heavy pressure. But make sure that this reclaim
3307 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3308 * allocation which will never fail due to oom_lock already held.
3310 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3311 ~__GFP_DIRECT_RECLAIM, order,
3312 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3316 /* Coredumps can quickly deplete all memory reserves */
3317 if (current->flags & PF_DUMPCORE)
3319 /* The OOM killer will not help higher order allocs */
3320 if (order > PAGE_ALLOC_COSTLY_ORDER)
3323 * We have already exhausted all our reclaim opportunities without any
3324 * success so it is time to admit defeat. We will skip the OOM killer
3325 * because it is very likely that the caller has a more reasonable
3326 * fallback than shooting a random task.
3328 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3330 /* The OOM killer does not needlessly kill tasks for lowmem */
3331 if (ac->high_zoneidx < ZONE_NORMAL)
3333 if (pm_suspended_storage())
3336 * XXX: GFP_NOFS allocations should rather fail than rely on
3337 * other request to make a forward progress.
3338 * We are in an unfortunate situation where out_of_memory cannot
3339 * do much for this context but let's try it to at least get
3340 * access to memory reserved if the current task is killed (see
3341 * out_of_memory). Once filesystems are ready to handle allocation
3342 * failures more gracefully we should just bail out here.
3345 /* The OOM killer may not free memory on a specific node */
3346 if (gfp_mask & __GFP_THISNODE)
3349 /* Exhausted what can be done so it's blamo time */
3350 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3351 *did_some_progress = 1;
3354 * Help non-failing allocations by giving them access to memory
3357 if (gfp_mask & __GFP_NOFAIL)
3358 page = __alloc_pages_cpu