1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
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)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/interrupt.h>
22 #include <linux/jiffies.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.h>
25 #include <linux/kasan.h>
26 #include <linux/kmsan.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/ratelimit.h>
30 #include <linux/oom.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/memory_hotplug.h>
36 #include <linux/nodemask.h>
37 #include <linux/vmstat.h>
38 #include <linux/fault-inject.h>
39 #include <linux/compaction.h>
40 #include <trace/events/kmem.h>
41 #include <trace/events/oom.h>
42 #include <linux/prefetch.h>
43 #include <linux/mm_inline.h>
44 #include <linux/mmu_notifier.h>
45 #include <linux/migrate.h>
46 #include <linux/sched/mm.h>
47 #include <linux/page_owner.h>
48 #include <linux/page_table_check.h>
49 #include <linux/memcontrol.h>
50 #include <linux/ftrace.h>
51 #include <linux/lockdep.h>
52 #include <linux/psi.h>
53 #include <linux/khugepaged.h>
54 #include <linux/delayacct.h>
55 #include <asm/div64.h>
58 #include "page_reporting.h"
60 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
61 typedef int __bitwise fpi_t;
63 /* No special request */
64 #define FPI_NONE ((__force fpi_t)0)
67 * Skip free page reporting notification for the (possibly merged) page.
68 * This does not hinder free page reporting from grabbing the page,
69 * reporting it and marking it "reported" - it only skips notifying
70 * the free page reporting infrastructure about a newly freed page. For
71 * example, used when temporarily pulling a page from a freelist and
72 * putting it back unmodified.
74 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
77 * Place the (possibly merged) page to the tail of the freelist. Will ignore
78 * page shuffling (relevant code - e.g., memory onlining - is expected to
79 * shuffle the whole zone).
81 * Note: No code should rely on this flag for correctness - it's purely
82 * to allow for optimizations when handing back either fresh pages
83 * (memory onlining) or untouched pages (page isolation, free page
86 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
88 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
89 static DEFINE_MUTEX(pcp_batch_high_lock);
90 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
92 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
94 * On SMP, spin_trylock is sufficient protection.
95 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
97 #define pcp_trylock_prepare(flags) do { } while (0)
98 #define pcp_trylock_finish(flag) do { } while (0)
101 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
102 #define pcp_trylock_prepare(flags) local_irq_save(flags)
103 #define pcp_trylock_finish(flags) local_irq_restore(flags)
107 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
108 * a migration causing the wrong PCP to be locked and remote memory being
109 * potentially allocated, pin the task to the CPU for the lookup+lock.
110 * preempt_disable is used on !RT because it is faster than migrate_disable.
111 * migrate_disable is used on RT because otherwise RT spinlock usage is
112 * interfered with and a high priority task cannot preempt the allocator.
114 #ifndef CONFIG_PREEMPT_RT
115 #define pcpu_task_pin() preempt_disable()
116 #define pcpu_task_unpin() preempt_enable()
118 #define pcpu_task_pin() migrate_disable()
119 #define pcpu_task_unpin() migrate_enable()
123 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
124 * Return value should be used with equivalent unlock helper.
126 #define pcpu_spin_lock(type, member, ptr) \
130 _ret = this_cpu_ptr(ptr); \
131 spin_lock(&_ret->member); \
135 #define pcpu_spin_trylock(type, member, ptr) \
139 _ret = this_cpu_ptr(ptr); \
140 if (!spin_trylock(&_ret->member)) { \
147 #define pcpu_spin_unlock(member, ptr) \
149 spin_unlock(&ptr->member); \
153 /* struct per_cpu_pages specific helpers. */
154 #define pcp_spin_lock(ptr) \
155 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
157 #define pcp_spin_trylock(ptr) \
158 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
160 #define pcp_spin_unlock(ptr) \
161 pcpu_spin_unlock(lock, ptr)
163 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
164 DEFINE_PER_CPU(int, numa_node);
165 EXPORT_PER_CPU_SYMBOL(numa_node);
168 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
170 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
172 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
173 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
174 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
175 * defined in <linux/topology.h>.
177 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
178 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
181 static DEFINE_MUTEX(pcpu_drain_mutex);
183 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
184 volatile unsigned long latent_entropy __latent_entropy;
185 EXPORT_SYMBOL(latent_entropy);
189 * Array of node states.
191 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
192 [N_POSSIBLE] = NODE_MASK_ALL,
193 [N_ONLINE] = { { [0] = 1UL } },
195 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
196 #ifdef CONFIG_HIGHMEM
197 [N_HIGH_MEMORY] = { { [0] = 1UL } },
199 [N_MEMORY] = { { [0] = 1UL } },
200 [N_CPU] = { { [0] = 1UL } },
203 EXPORT_SYMBOL(node_states);
205 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
208 * A cached value of the page's pageblock's migratetype, used when the page is
209 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
210 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
211 * Also the migratetype set in the page does not necessarily match the pcplist
212 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
213 * other index - this ensures that it will be put on the correct CMA freelist.
215 static inline int get_pcppage_migratetype(struct page *page)
220 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
222 page->index = migratetype;
225 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
226 unsigned int pageblock_order __read_mostly;
229 static void __free_pages_ok(struct page *page, unsigned int order,
233 * results with 256, 32 in the lowmem_reserve sysctl:
234 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
235 * 1G machine -> (16M dma, 784M normal, 224M high)
236 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
237 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
238 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
240 * TBD: should special case ZONE_DMA32 machines here - in those we normally
241 * don't need any ZONE_NORMAL reservation
243 static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
244 #ifdef CONFIG_ZONE_DMA
247 #ifdef CONFIG_ZONE_DMA32
251 #ifdef CONFIG_HIGHMEM
257 char * const zone_names[MAX_NR_ZONES] = {
258 #ifdef CONFIG_ZONE_DMA
261 #ifdef CONFIG_ZONE_DMA32
265 #ifdef CONFIG_HIGHMEM
269 #ifdef CONFIG_ZONE_DEVICE
274 const char * const migratetype_names[MIGRATE_TYPES] = {
282 #ifdef CONFIG_MEMORY_ISOLATION
287 static compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
288 [NULL_COMPOUND_DTOR] = NULL,
289 [COMPOUND_PAGE_DTOR] = free_compound_page,
290 #ifdef CONFIG_HUGETLB_PAGE
291 [HUGETLB_PAGE_DTOR] = free_huge_page,
293 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
294 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
298 int min_free_kbytes = 1024;
299 int user_min_free_kbytes = -1;
300 static int watermark_boost_factor __read_mostly = 15000;
301 static int watermark_scale_factor = 10;
303 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
305 EXPORT_SYMBOL(movable_zone);
308 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
309 unsigned int nr_online_nodes __read_mostly = 1;
310 EXPORT_SYMBOL(nr_node_ids);
311 EXPORT_SYMBOL(nr_online_nodes);
314 static bool page_contains_unaccepted(struct page *page, unsigned int order);
315 static void accept_page(struct page *page, unsigned int order);
316 static bool try_to_accept_memory(struct zone *zone, unsigned int order);
317 static inline bool has_unaccepted_memory(void);
318 static bool __free_unaccepted(struct page *page);
320 int page_group_by_mobility_disabled __read_mostly;
322 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
324 * During boot we initialize deferred pages on-demand, as needed, but once
325 * page_alloc_init_late() has finished, the deferred pages are all initialized,
326 * and we can permanently disable that path.
328 DEFINE_STATIC_KEY_TRUE(deferred_pages);
330 static inline bool deferred_pages_enabled(void)
332 return static_branch_unlikely(&deferred_pages);
336 * deferred_grow_zone() is __init, but it is called from
337 * get_page_from_freelist() during early boot until deferred_pages permanently
338 * disables this call. This is why we have refdata wrapper to avoid warning,
339 * and to ensure that the function body gets unloaded.
342 _deferred_grow_zone(struct zone *zone, unsigned int order)
344 return deferred_grow_zone(zone, order);
347 static inline bool deferred_pages_enabled(void)
351 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
353 /* Return a pointer to the bitmap storing bits affecting a block of pages */
354 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
357 #ifdef CONFIG_SPARSEMEM
358 return section_to_usemap(__pfn_to_section(pfn));
360 return page_zone(page)->pageblock_flags;
361 #endif /* CONFIG_SPARSEMEM */
364 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
366 #ifdef CONFIG_SPARSEMEM
367 pfn &= (PAGES_PER_SECTION-1);
369 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
370 #endif /* CONFIG_SPARSEMEM */
371 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
374 static __always_inline
375 unsigned long __get_pfnblock_flags_mask(const struct page *page,
379 unsigned long *bitmap;
380 unsigned long bitidx, word_bitidx;
383 bitmap = get_pageblock_bitmap(page, pfn);
384 bitidx = pfn_to_bitidx(page, pfn);
385 word_bitidx = bitidx / BITS_PER_LONG;
386 bitidx &= (BITS_PER_LONG-1);
388 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
389 * a consistent read of the memory array, so that results, even though
390 * racy, are not corrupted.
392 word = READ_ONCE(bitmap[word_bitidx]);
393 return (word >> bitidx) & mask;
397 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
398 * @page: The page within the block of interest
399 * @pfn: The target page frame number
400 * @mask: mask of bits that the caller is interested in
402 * Return: pageblock_bits flags
404 unsigned long get_pfnblock_flags_mask(const struct page *page,
405 unsigned long pfn, unsigned long mask)
407 return __get_pfnblock_flags_mask(page, pfn, mask);
410 static __always_inline int get_pfnblock_migratetype(const struct page *page,
413 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
417 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
418 * @page: The page within the block of interest
419 * @flags: The flags to set
420 * @pfn: The target page frame number
421 * @mask: mask of bits that the caller is interested in
423 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
427 unsigned long *bitmap;
428 unsigned long bitidx, word_bitidx;
431 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
432 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
434 bitmap = get_pageblock_bitmap(page, pfn);
435 bitidx = pfn_to_bitidx(page, pfn);
436 word_bitidx = bitidx / BITS_PER_LONG;
437 bitidx &= (BITS_PER_LONG-1);
439 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
444 word = READ_ONCE(bitmap[word_bitidx]);
446 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
449 void set_pageblock_migratetype(struct page *page, int migratetype)
451 if (unlikely(page_group_by_mobility_disabled &&
452 migratetype < MIGRATE_PCPTYPES))
453 migratetype = MIGRATE_UNMOVABLE;
455 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
456 page_to_pfn(page), MIGRATETYPE_MASK);
459 #ifdef CONFIG_DEBUG_VM
460 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
464 unsigned long pfn = page_to_pfn(page);
465 unsigned long sp, start_pfn;
468 seq = zone_span_seqbegin(zone);
469 start_pfn = zone->zone_start_pfn;
470 sp = zone->spanned_pages;
471 ret = !zone_spans_pfn(zone, pfn);
472 } while (zone_span_seqretry(zone, seq));
475 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
476 pfn, zone_to_nid(zone), zone->name,
477 start_pfn, start_pfn + sp);
483 * Temporary debugging check for pages not lying within a given zone.
485 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
487 if (page_outside_zone_boundaries(zone, page))
489 if (zone != page_zone(page))
495 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
501 static void bad_page(struct page *page, const char *reason)
503 static unsigned long resume;
504 static unsigned long nr_shown;
505 static unsigned long nr_unshown;
508 * Allow a burst of 60 reports, then keep quiet for that minute;
509 * or allow a steady drip of one report per second.
511 if (nr_shown == 60) {
512 if (time_before(jiffies, resume)) {
518 "BUG: Bad page state: %lu messages suppressed\n",
525 resume = jiffies + 60 * HZ;
527 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
528 current->comm, page_to_pfn(page));
529 dump_page(page, reason);
534 /* Leave bad fields for debug, except PageBuddy could make trouble */
535 page_mapcount_reset(page); /* remove PageBuddy */
536 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
539 static inline unsigned int order_to_pindex(int migratetype, int order)
543 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
544 if (order > PAGE_ALLOC_COSTLY_ORDER) {
545 VM_BUG_ON(order != pageblock_order);
546 return NR_LOWORDER_PCP_LISTS;
549 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
552 return (MIGRATE_PCPTYPES * base) + migratetype;
555 static inline int pindex_to_order(unsigned int pindex)
557 int order = pindex / MIGRATE_PCPTYPES;
559 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
560 if (pindex == NR_LOWORDER_PCP_LISTS)
561 order = pageblock_order;
563 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
569 static inline bool pcp_allowed_order(unsigned int order)
571 if (order <= PAGE_ALLOC_COSTLY_ORDER)
573 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
574 if (order == pageblock_order)
580 static inline void free_the_page(struct page *page, unsigned int order)
582 if (pcp_allowed_order(order)) /* Via pcp? */
583 free_unref_page(page, order);
585 __free_pages_ok(page, order, FPI_NONE);
589 * Higher-order pages are called "compound pages". They are structured thusly:
591 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
593 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
594 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
596 * The first tail page's ->compound_dtor holds the offset in array of compound
597 * page destructors. See compound_page_dtors.
599 * The first tail page's ->compound_order holds the order of allocation.
600 * This usage means that zero-order pages may not be compound.
603 void free_compound_page(struct page *page)
605 mem_cgroup_uncharge(page_folio(page));
606 free_the_page(page, compound_order(page));
609 void prep_compound_page(struct page *page, unsigned int order)
612 int nr_pages = 1 << order;
615 for (i = 1; i < nr_pages; i++)
616 prep_compound_tail(page, i);
618 prep_compound_head(page, order);
621 void destroy_large_folio(struct folio *folio)
623 enum compound_dtor_id dtor = folio->_folio_dtor;
625 VM_BUG_ON_FOLIO(dtor >= NR_COMPOUND_DTORS, folio);
626 compound_page_dtors[dtor](&folio->page);
629 static inline void set_buddy_order(struct page *page, unsigned int order)
631 set_page_private(page, order);
632 __SetPageBuddy(page);
635 #ifdef CONFIG_COMPACTION
636 static inline struct capture_control *task_capc(struct zone *zone)
638 struct capture_control *capc = current->capture_control;
640 return unlikely(capc) &&
641 !(current->flags & PF_KTHREAD) &&
643 capc->cc->zone == zone ? capc : NULL;
647 compaction_capture(struct capture_control *capc, struct page *page,
648 int order, int migratetype)
650 if (!capc || order != capc->cc->order)
653 /* Do not accidentally pollute CMA or isolated regions*/
654 if (is_migrate_cma(migratetype) ||
655 is_migrate_isolate(migratetype))
659 * Do not let lower order allocations pollute a movable pageblock.
660 * This might let an unmovable request use a reclaimable pageblock
661 * and vice-versa but no more than normal fallback logic which can
662 * have trouble finding a high-order free page.
664 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
672 static inline struct capture_control *task_capc(struct zone *zone)
678 compaction_capture(struct capture_control *capc, struct page *page,
679 int order, int migratetype)
683 #endif /* CONFIG_COMPACTION */
685 /* Used for pages not on another list */
686 static inline void add_to_free_list(struct page *page, struct zone *zone,
687 unsigned int order, int migratetype)
689 struct free_area *area = &zone->free_area[order];
691 list_add(&page->buddy_list, &area->free_list[migratetype]);
695 /* Used for pages not on another list */
696 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
697 unsigned int order, int migratetype)
699 struct free_area *area = &zone->free_area[order];
701 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
706 * Used for pages which are on another list. Move the pages to the tail
707 * of the list - so the moved pages won't immediately be considered for
708 * allocation again (e.g., optimization for memory onlining).
710 static inline void move_to_free_list(struct page *page, struct zone *zone,
711 unsigned int order, int migratetype)
713 struct free_area *area = &zone->free_area[order];
715 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
718 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
721 /* clear reported state and update reported page count */
722 if (page_reported(page))
723 __ClearPageReported(page);
725 list_del(&page->buddy_list);
726 __ClearPageBuddy(page);
727 set_page_private(page, 0);
728 zone->free_area[order].nr_free--;
731 static inline struct page *get_page_from_free_area(struct free_area *area,
734 return list_first_entry_or_null(&area->free_list[migratetype],
735 struct page, buddy_list);
739 * If this is not the largest possible page, check if the buddy
740 * of the next-highest order is free. If it is, it's possible
741 * that pages are being freed that will coalesce soon. In case,
742 * that is happening, add the free page to the tail of the list
743 * so it's less likely to be used soon and more likely to be merged
744 * as a higher order page
747 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
748 struct page *page, unsigned int order)
750 unsigned long higher_page_pfn;
751 struct page *higher_page;
753 if (order >= MAX_ORDER - 1)
756 higher_page_pfn = buddy_pfn & pfn;
757 higher_page = page + (higher_page_pfn - pfn);
759 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
764 * Freeing function for a buddy system allocator.
766 * The concept of a buddy system is to maintain direct-mapped table
767 * (containing bit values) for memory blocks of various "orders".
768 * The bottom level table contains the map for the smallest allocatable
769 * units of memory (here, pages), and each level above it describes
770 * pairs of units from the levels below, hence, "buddies".
771 * At a high level, all that happens here is marking the table entry
772 * at the bottom level available, and propagating the changes upward
773 * as necessary, plus some accounting needed to play nicely with other
774 * parts of the VM system.
775 * At each level, we keep a list of pages, which are heads of continuous
776 * free pages of length of (1 << order) and marked with PageBuddy.
777 * Page's order is recorded in page_private(page) field.
778 * So when we are allocating or freeing one, we can derive the state of the
779 * other. That is, if we allocate a small block, and both were
780 * free, the remainder of the region must be split into blocks.
781 * If a block is freed, and its buddy is also free, then this
782 * triggers coalescing into a block of larger size.
787 static inline void __free_one_page(struct page *page,
789 struct zone *zone, unsigned int order,
790 int migratetype, fpi_t fpi_flags)
792 struct capture_control *capc = task_capc(zone);
793 unsigned long buddy_pfn = 0;
794 unsigned long combined_pfn;
798 VM_BUG_ON(!zone_is_initialized(zone));
799 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
801 VM_BUG_ON(migratetype == -1);
802 if (likely(!is_migrate_isolate(migratetype)))
803 __mod_zone_freepage_state(zone, 1 << order, migratetype);
805 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
806 VM_BUG_ON_PAGE(bad_range(zone, page), page);
808 while (order < MAX_ORDER) {
809 if (compaction_capture(capc, page, order, migratetype)) {
810 __mod_zone_freepage_state(zone, -(1 << order),
815 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
819 if (unlikely(order >= pageblock_order)) {
821 * We want to prevent merge between freepages on pageblock
822 * without fallbacks and normal pageblock. Without this,
823 * pageblock isolation could cause incorrect freepage or CMA
824 * accounting or HIGHATOMIC accounting.
826 int buddy_mt = get_pageblock_migratetype(buddy);
828 if (migratetype != buddy_mt
829 && (!migratetype_is_mergeable(migratetype) ||
830 !migratetype_is_mergeable(buddy_mt)))
835 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
836 * merge with it and move up one order.
838 if (page_is_guard(buddy))
839 clear_page_guard(zone, buddy, order, migratetype);
841 del_page_from_free_list(buddy, zone, order);
842 combined_pfn = buddy_pfn & pfn;
843 page = page + (combined_pfn - pfn);
849 set_buddy_order(page, order);
851 if (fpi_flags & FPI_TO_TAIL)
853 else if (is_shuffle_order(order))
854 to_tail = shuffle_pick_tail();
856 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
859 add_to_free_list_tail(page, zone, order, migratetype);
861 add_to_free_list(page, zone, order, migratetype);
863 /* Notify page reporting subsystem of freed page */
864 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
865 page_reporting_notify_free(order);
869 * split_free_page() -- split a free page at split_pfn_offset
870 * @free_page: the original free page
871 * @order: the order of the page
872 * @split_pfn_offset: split offset within the page
874 * Return -ENOENT if the free page is changed, otherwise 0
876 * It is used when the free page crosses two pageblocks with different migratetypes
877 * at split_pfn_offset within the page. The split free page will be put into
878 * separate migratetype lists afterwards. Otherwise, the function achieves
881 int split_free_page(struct page *free_page,
882 unsigned int order, unsigned long split_pfn_offset)
884 struct zone *zone = page_zone(free_page);
885 unsigned long free_page_pfn = page_to_pfn(free_page);
892 if (split_pfn_offset == 0)
895 spin_lock_irqsave(&zone->lock, flags);
897 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
902 mt = get_pageblock_migratetype(free_page);
903 if (likely(!is_migrate_isolate(mt)))
904 __mod_zone_freepage_state(zone, -(1UL << order), mt);
906 del_page_from_free_list(free_page, zone, order);
907 for (pfn = free_page_pfn;
908 pfn < free_page_pfn + (1UL << order);) {
909 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
911 free_page_order = min_t(unsigned int,
912 pfn ? __ffs(pfn) : order,
913 __fls(split_pfn_offset));
914 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
916 pfn += 1UL << free_page_order;
917 split_pfn_offset -= (1UL << free_page_order);
918 /* we have done the first part, now switch to second part */
919 if (split_pfn_offset == 0)
920 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
923 spin_unlock_irqrestore(&zone->lock, flags);
927 * A bad page could be due to a number of fields. Instead of multiple branches,
928 * try and check multiple fields with one check. The caller must do a detailed
929 * check if necessary.
931 static inline bool page_expected_state(struct page *page,
932 unsigned long check_flags)
934 if (unlikely(atomic_read(&page->_mapcount) != -1))
937 if (unlikely((unsigned long)page->mapping |
938 page_ref_count(page) |
942 (page->flags & check_flags)))
948 static const char *page_bad_reason(struct page *page, unsigned long flags)
950 const char *bad_reason = NULL;
952 if (unlikely(atomic_read(&page->_mapcount) != -1))
953 bad_reason = "nonzero mapcount";
954 if (unlikely(page->mapping != NULL))
955 bad_reason = "non-NULL mapping";
956 if (unlikely(page_ref_count(page) != 0))
957 bad_reason = "nonzero _refcount";
958 if (unlikely(page->flags & flags)) {
959 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
960 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
962 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
965 if (unlikely(page->memcg_data))
966 bad_reason = "page still charged to cgroup";
971 static void free_page_is_bad_report(struct page *page)
974 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
977 static inline bool free_page_is_bad(struct page *page)
979 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
982 /* Something has gone sideways, find it */
983 free_page_is_bad_report(page);
987 static inline bool is_check_pages_enabled(void)
989 return static_branch_unlikely(&check_pages_enabled);
992 static int free_tail_page_prepare(struct page *head_page, struct page *page)
994 struct folio *folio = (struct folio *)head_page;
998 * We rely page->lru.next never has bit 0 set, unless the page
999 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1001 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1003 if (!is_check_pages_enabled()) {
1007 switch (page - head_page) {
1009 /* the first tail page: these may be in place of ->mapping */
1010 if (unlikely(folio_entire_mapcount(folio))) {
1011 bad_page(page, "nonzero entire_mapcount");
1014 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
1015 bad_page(page, "nonzero nr_pages_mapped");
1018 if (unlikely(atomic_read(&folio->_pincount))) {
1019 bad_page(page, "nonzero pincount");
1025 * the second tail page: ->mapping is
1026 * deferred_list.next -- ignore value.
1030 if (page->mapping != TAIL_MAPPING) {
1031 bad_page(page, "corrupted mapping in tail page");
1036 if (unlikely(!PageTail(page))) {
1037 bad_page(page, "PageTail not set");
1040 if (unlikely(compound_head(page) != head_page)) {
1041 bad_page(page, "compound_head not consistent");
1046 page->mapping = NULL;
1047 clear_compound_head(page);
1052 * Skip KASAN memory poisoning when either:
1054 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1055 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1056 * using page tags instead (see below).
1057 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1058 * that error detection is disabled for accesses via the page address.
1060 * Pages will have match-all tags in the following circumstances:
1062 * 1. Pages are being initialized for the first time, including during deferred
1063 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1064 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1065 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1066 * 3. The allocation was excluded from being checked due to sampling,
1067 * see the call to kasan_unpoison_pages.
1069 * Poisoning pages during deferred memory init will greatly lengthen the
1070 * process and cause problem in large memory systems as the deferred pages
1071 * initialization is done with interrupt disabled.
1073 * Assuming that there will be no reference to those newly initialized
1074 * pages before they are ever allocated, this should have no effect on
1075 * KASAN memory tracking as the poison will be properly inserted at page
1076 * allocation time. The only corner case is when pages are allocated by
1077 * on-demand allocation and then freed again before the deferred pages
1078 * initialization is done, but this is not likely to happen.
1080 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1082 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1083 return deferred_pages_enabled();
1085 return page_kasan_tag(page) == 0xff;
1088 static void kernel_init_pages(struct page *page, int numpages)
1092 /* s390's use of memset() could override KASAN redzones. */
1093 kasan_disable_current();
1094 for (i = 0; i < numpages; i++)
1095 clear_highpage_kasan_tagged(page + i);
1096 kasan_enable_current();
1099 static __always_inline bool free_pages_prepare(struct page *page,
1100 unsigned int order, fpi_t fpi_flags)
1103 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1104 bool init = want_init_on_free();
1106 VM_BUG_ON_PAGE(PageTail(page), page);
1108 trace_mm_page_free(page, order);
1109 kmsan_free_page(page, order);
1111 if (unlikely(PageHWPoison(page)) && !order) {
1113 * Do not let hwpoison pages hit pcplists/buddy
1114 * Untie memcg state and reset page's owner
1116 if (memcg_kmem_online() && PageMemcgKmem(page))
1117 __memcg_kmem_uncharge_page(page, order);
1118 reset_page_owner(page, order);
1119 page_table_check_free(page, order);
1124 * Check tail pages before head page information is cleared to
1125 * avoid checking PageCompound for order-0 pages.
1127 if (unlikely(order)) {
1128 bool compound = PageCompound(page);
1131 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1134 ClearPageHasHWPoisoned(page);
1135 for (i = 1; i < (1 << order); i++) {
1137 bad += free_tail_page_prepare(page, page + i);
1138 if (is_check_pages_enabled()) {
1139 if (free_page_is_bad(page + i)) {
1144 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1147 if (PageMappingFlags(page))
1148 page->mapping = NULL;
1149 if (memcg_kmem_online() && PageMemcgKmem(page))
1150 __memcg_kmem_uncharge_page(page, order);
1151 if (is_check_pages_enabled()) {
1152 if (free_page_is_bad(page))
1158 page_cpupid_reset_last(page);
1159 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1160 reset_page_owner(page, order);
1161 page_table_check_free(page, order);
1163 if (!PageHighMem(page)) {
1164 debug_check_no_locks_freed(page_address(page),
1165 PAGE_SIZE << order);
1166 debug_check_no_obj_freed(page_address(page),
1167 PAGE_SIZE << order);
1170 kernel_poison_pages(page, 1 << order);
1173 * As memory initialization might be integrated into KASAN,
1174 * KASAN poisoning and memory initialization code must be
1175 * kept together to avoid discrepancies in behavior.
1177 * With hardware tag-based KASAN, memory tags must be set before the
1178 * page becomes unavailable via debug_pagealloc or arch_free_page.
1180 if (!skip_kasan_poison) {
1181 kasan_poison_pages(page, order, init);
1183 /* Memory is already initialized if KASAN did it internally. */
1184 if (kasan_has_integrated_init())
1188 kernel_init_pages(page, 1 << order);
1191 * arch_free_page() can make the page's contents inaccessible. s390
1192 * does this. So nothing which can access the page's contents should
1193 * happen after this.
1195 arch_free_page(page, order);
1197 debug_pagealloc_unmap_pages(page, 1 << order);
1203 * Frees a number of pages from the PCP lists
1204 * Assumes all pages on list are in same zone.
1205 * count is the number of pages to free.
1207 static void free_pcppages_bulk(struct zone *zone, int count,
1208 struct per_cpu_pages *pcp,
1211 unsigned long flags;
1213 int max_pindex = NR_PCP_LISTS - 1;
1215 bool isolated_pageblocks;
1219 * Ensure proper count is passed which otherwise would stuck in the
1220 * below while (list_empty(list)) loop.
1222 count = min(pcp->count, count);
1224 /* Ensure requested pindex is drained first. */
1225 pindex = pindex - 1;
1227 spin_lock_irqsave(&zone->lock, flags);
1228 isolated_pageblocks = has_isolate_pageblock(zone);
1231 struct list_head *list;
1234 /* Remove pages from lists in a round-robin fashion. */
1236 if (++pindex > max_pindex)
1237 pindex = min_pindex;
1238 list = &pcp->lists[pindex];
1239 if (!list_empty(list))
1242 if (pindex == max_pindex)
1244 if (pindex == min_pindex)
1248 order = pindex_to_order(pindex);
1249 nr_pages = 1 << order;
1253 page = list_last_entry(list, struct page, pcp_list);
1254 mt = get_pcppage_migratetype(page);
1256 /* must delete to avoid corrupting pcp list */
1257 list_del(&page->pcp_list);
1259 pcp->count -= nr_pages;
1261 /* MIGRATE_ISOLATE page should not go to pcplists */
1262 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1263 /* Pageblock could have been isolated meanwhile */
1264 if (unlikely(isolated_pageblocks))
1265 mt = get_pageblock_migratetype(page);
1267 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1268 trace_mm_page_pcpu_drain(page, order, mt);
1269 } while (count > 0 && !list_empty(list));
1272 spin_unlock_irqrestore(&zone->lock, flags);
1275 static void free_one_page(struct zone *zone,
1276 struct page *page, unsigned long pfn,
1278 int migratetype, fpi_t fpi_flags)
1280 unsigned long flags;
1282 spin_lock_irqsave(&zone->lock, flags);
1283 if (unlikely(has_isolate_pageblock(zone) ||
1284 is_migrate_isolate(migratetype))) {
1285 migratetype = get_pfnblock_migratetype(page, pfn);
1287 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1288 spin_unlock_irqrestore(&zone->lock, flags);
1291 static void __free_pages_ok(struct page *page, unsigned int order,
1294 unsigned long flags;
1296 unsigned long pfn = page_to_pfn(page);
1297 struct zone *zone = page_zone(page);
1299 if (!free_pages_prepare(page, order, fpi_flags))
1303 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1304 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1305 * This will reduce the lock holding time.
1307 migratetype = get_pfnblock_migratetype(page, pfn);
1309 spin_lock_irqsave(&zone->lock, flags);
1310 if (unlikely(has_isolate_pageblock(zone) ||
1311 is_migrate_isolate(migratetype))) {
1312 migratetype = get_pfnblock_migratetype(page, pfn);
1314 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1315 spin_unlock_irqrestore(&zone->lock, flags);
1317 __count_vm_events(PGFREE, 1 << order);
1320 void __free_pages_core(struct page *page, unsigned int order)
1322 unsigned int nr_pages = 1 << order;
1323 struct page *p = page;
1327 * When initializing the memmap, __init_single_page() sets the refcount
1328 * of all pages to 1 ("allocated"/"not free"). We have to set the
1329 * refcount of all involved pages to 0.
1332 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1334 __ClearPageReserved(p);
1335 set_page_count(p, 0);
1337 __ClearPageReserved(p);
1338 set_page_count(p, 0);
1340 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1342 if (page_contains_unaccepted(page, order)) {
1343 if (order == MAX_ORDER && __free_unaccepted(page))
1346 accept_page(page, order);
1350 * Bypass PCP and place fresh pages right to the tail, primarily
1351 * relevant for memory onlining.
1353 __free_pages_ok(page, order, FPI_TO_TAIL);
1357 * Check that the whole (or subset of) a pageblock given by the interval of
1358 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1359 * with the migration of free compaction scanner.
1361 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1363 * It's possible on some configurations to have a setup like node0 node1 node0
1364 * i.e. it's possible that all pages within a zones range of pages do not
1365 * belong to a single zone. We assume that a border between node0 and node1
1366 * can occur within a single pageblock, but not a node0 node1 node0
1367 * interleaving within a single pageblock. It is therefore sufficient to check
1368 * the first and last page of a pageblock and avoid checking each individual
1369 * page in a pageblock.
1371 * Note: the function may return non-NULL struct page even for a page block
1372 * which contains a memory hole (i.e. there is no physical memory for a subset
1373 * of the pfn range). For example, if the pageblock order is MAX_ORDER, which
1374 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1375 * even though the start pfn is online and valid. This should be safe most of
1376 * the time because struct pages are still initialized via init_unavailable_range()
1377 * and pfn walkers shouldn't touch any physical memory range for which they do
1378 * not recognize any specific metadata in struct pages.
1380 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1381 unsigned long end_pfn, struct zone *zone)
1383 struct page *start_page;
1384 struct page *end_page;
1386 /* end_pfn is one past the range we are checking */
1389 if (!pfn_valid(end_pfn))
1392 start_page = pfn_to_online_page(start_pfn);
1396 if (page_zone(start_page) != zone)
1399 end_page = pfn_to_page(end_pfn);
1401 /* This gives a shorter code than deriving page_zone(end_page) */
1402 if (page_zone_id(start_page) != page_zone_id(end_page))
1409 * The order of subdivision here is critical for the IO subsystem.
1410 * Please do not alter this order without good reasons and regression
1411 * testing. Specifically, as large blocks of memory are subdivided,
1412 * the order in which smaller blocks are delivered depends on the order
1413 * they're subdivided in this function. This is the primary factor
1414 * influencing the order in which pages are delivered to the IO
1415 * subsystem according to empirical testing, and this is also justified
1416 * by considering the behavior of a buddy system containing a single
1417 * large block of memory acted on by a series of small allocations.
1418 * This behavior is a critical factor in sglist merging's success.
1422 static inline void expand(struct zone *zone, struct page *page,
1423 int low, int high, int migratetype)
1425 unsigned long size = 1 << high;
1427 while (high > low) {
1430 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1433 * Mark as guard pages (or page), that will allow to
1434 * merge back to allocator when buddy will be freed.
1435 * Corresponding page table entries will not be touched,
1436 * pages will stay not present in virtual address space
1438 if (set_page_guard(zone, &page[size], high, migratetype))
1441 add_to_free_list(&page[size], zone, high, migratetype);
1442 set_buddy_order(&page[size], high);
1446 static void check_new_page_bad(struct page *page)
1448 if (unlikely(page->flags & __PG_HWPOISON)) {
1449 /* Don't complain about hwpoisoned pages */
1450 page_mapcount_reset(page); /* remove PageBuddy */
1455 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1459 * This page is about to be returned from the page allocator
1461 static int check_new_page(struct page *page)
1463 if (likely(page_expected_state(page,
1464 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1467 check_new_page_bad(page);
1471 static inline bool check_new_pages(struct page *page, unsigned int order)
1473 if (is_check_pages_enabled()) {
1474 for (int i = 0; i < (1 << order); i++) {
1475 struct page *p = page + i;
1477 if (check_new_page(p))
1485 static inline bool should_skip_kasan_unpoison(gfp_t flags)
1487 /* Don't skip if a software KASAN mode is enabled. */
1488 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1489 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1492 /* Skip, if hardware tag-based KASAN is not enabled. */
1493 if (!kasan_hw_tags_enabled())
1497 * With hardware tag-based KASAN enabled, skip if this has been
1498 * requested via __GFP_SKIP_KASAN.
1500 return flags & __GFP_SKIP_KASAN;
1503 static inline bool should_skip_init(gfp_t flags)
1505 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1506 if (!kasan_hw_tags_enabled())
1509 /* For hardware tag-based KASAN, skip if requested. */
1510 return (flags & __GFP_SKIP_ZERO);
1513 inline void post_alloc_hook(struct page *page, unsigned int order,
1516 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1517 !should_skip_init(gfp_flags);
1518 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1521 set_page_private(page, 0);
1522 set_page_refcounted(page);
1524 arch_alloc_page(page, order);
1525 debug_pagealloc_map_pages(page, 1 << order);
1528 * Page unpoisoning must happen before memory initialization.
1529 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1530 * allocations and the page unpoisoning code will complain.
1532 kernel_unpoison_pages(page, 1 << order);
1535 * As memory initialization might be integrated into KASAN,
1536 * KASAN unpoisoning and memory initializion code must be
1537 * kept together to avoid discrepancies in behavior.
1541 * If memory tags should be zeroed
1542 * (which happens only when memory should be initialized as well).
1545 /* Initialize both memory and memory tags. */
1546 for (i = 0; i != 1 << order; ++i)
1547 tag_clear_highpage(page + i);
1549 /* Take note that memory was initialized by the loop above. */
1552 if (!should_skip_kasan_unpoison(gfp_flags) &&
1553 kasan_unpoison_pages(page, order, init)) {
1554 /* Take note that memory was initialized by KASAN. */
1555 if (kasan_has_integrated_init())
1559 * If memory tags have not been set by KASAN, reset the page
1560 * tags to ensure page_address() dereferencing does not fault.
1562 for (i = 0; i != 1 << order; ++i)
1563 page_kasan_tag_reset(page + i);
1565 /* If memory is still not initialized, initialize it now. */
1567 kernel_init_pages(page, 1 << order);
1569 set_page_owner(page, order, gfp_flags);
1570 page_table_check_alloc(page, order);
1573 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1574 unsigned int alloc_flags)
1576 post_alloc_hook(page, order, gfp_flags);
1578 if (order && (gfp_flags & __GFP_COMP))
1579 prep_compound_page(page, order);
1582 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1583 * allocate the page. The expectation is that the caller is taking
1584 * steps that will free more memory. The caller should avoid the page
1585 * being used for !PFMEMALLOC purposes.
1587 if (alloc_flags & ALLOC_NO_WATERMARKS)
1588 set_page_pfmemalloc(page);
1590 clear_page_pfmemalloc(page);
1594 * Go through the free lists for the given migratetype and remove
1595 * the smallest available page from the freelists
1597 static __always_inline
1598 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1601 unsigned int current_order;
1602 struct free_area *area;
1605 /* Find a page of the appropriate size in the preferred list */
1606 for (current_order = order; current_order <= MAX_ORDER; ++current_order) {
1607 area = &(zone->free_area[current_order]);
1608 page = get_page_from_free_area(area, migratetype);
1611 del_page_from_free_list(page, zone, current_order);
1612 expand(zone, page, order, current_order, migratetype);
1613 set_pcppage_migratetype(page, migratetype);
1614 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1615 pcp_allowed_order(order) &&
1616 migratetype < MIGRATE_PCPTYPES);
1625 * This array describes the order lists are fallen back to when
1626 * the free lists for the desirable migrate type are depleted
1628 * The other migratetypes do not have fallbacks.
1630 static int fallbacks[MIGRATE_TYPES][MIGRATE_PCPTYPES - 1] = {
1631 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1632 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1633 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1637 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1640 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1643 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1644 unsigned int order) { return NULL; }
1648 * Move the free pages in a range to the freelist tail of the requested type.
1649 * Note that start_page and end_pages are not aligned on a pageblock
1650 * boundary. If alignment is required, use move_freepages_block()
1652 static int move_freepages(struct zone *zone,
1653 unsigned long start_pfn, unsigned long end_pfn,
1654 int migratetype, int *num_movable)
1659 int pages_moved = 0;
1661 for (pfn = start_pfn; pfn <= end_pfn;) {
1662 page = pfn_to_page(pfn);
1663 if (!PageBuddy(page)) {
1665 * We assume that pages that could be isolated for
1666 * migration are movable. But we don't actually try
1667 * isolating, as that would be expensive.
1670 (PageLRU(page) || __PageMovable(page)))
1676 /* Make sure we are not inadvertently changing nodes */
1677 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1678 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1680 order = buddy_order(page);
1681 move_to_free_list(page, zone, order, migratetype);
1683 pages_moved += 1 << order;
1689 int move_freepages_block(struct zone *zone, struct page *page,
1690 int migratetype, int *num_movable)
1692 unsigned long start_pfn, end_pfn, pfn;
1697 pfn = page_to_pfn(page);
1698 start_pfn = pageblock_start_pfn(pfn);
1699 end_pfn = pageblock_end_pfn(pfn) - 1;
1701 /* Do not cross zone boundaries */
1702 if (!zone_spans_pfn(zone, start_pfn))
1704 if (!zone_spans_pfn(zone, end_pfn))
1707 return move_freepages(zone, start_pfn, end_pfn, migratetype,
1711 static void change_pageblock_range(struct page *pageblock_page,
1712 int start_order, int migratetype)
1714 int nr_pageblocks = 1 << (start_order - pageblock_order);
1716 while (nr_pageblocks--) {
1717 set_pageblock_migratetype(pageblock_page, migratetype);
1718 pageblock_page += pageblock_nr_pages;
1723 * When we are falling back to another migratetype during allocation, try to
1724 * steal extra free pages from the same pageblocks to satisfy further
1725 * allocations, instead of polluting multiple pageblocks.
1727 * If we are stealing a relatively large buddy page, it is likely there will
1728 * be more free pages in the pageblock, so try to steal them all. For
1729 * reclaimable and unmovable allocations, we steal regardless of page size,
1730 * as fragmentation caused by those allocations polluting movable pageblocks
1731 * is worse than movable allocations stealing from unmovable and reclaimable
1734 static bool can_steal_fallback(unsigned int order, int start_mt)
1737 * Leaving this order check is intended, although there is
1738 * relaxed order check in next check. The reason is that
1739 * we can actually steal whole pageblock if this condition met,
1740 * but, below check doesn't guarantee it and that is just heuristic
1741 * so could be changed anytime.
1743 if (order >= pageblock_order)
1746 if (order >= pageblock_order / 2 ||
1747 start_mt == MIGRATE_RECLAIMABLE ||
1748 start_mt == MIGRATE_UNMOVABLE ||
1749 page_group_by_mobility_disabled)
1755 static inline bool boost_watermark(struct zone *zone)
1757 unsigned long max_boost;
1759 if (!watermark_boost_factor)
1762 * Don't bother in zones that are unlikely to produce results.
1763 * On small machines, including kdump capture kernels running
1764 * in a small area, boosting the watermark can cause an out of
1765 * memory situation immediately.
1767 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1770 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1771 watermark_boost_factor, 10000);
1774 * high watermark may be uninitialised if fragmentation occurs
1775 * very early in boot so do not boost. We do not fall
1776 * through and boost by pageblock_nr_pages as failing
1777 * allocations that early means that reclaim is not going
1778 * to help and it may even be impossible to reclaim the
1779 * boosted watermark resulting in a hang.
1784 max_boost = max(pageblock_nr_pages, max_boost);
1786 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1793 * This function implements actual steal behaviour. If order is large enough,
1794 * we can steal whole pageblock. If not, we first move freepages in this
1795 * pageblock to our migratetype and determine how many already-allocated pages
1796 * are there in the pageblock with a compatible migratetype. If at least half
1797 * of pages are free or compatible, we can change migratetype of the pageblock
1798 * itself, so pages freed in the future will be put on the correct free list.
1800 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1801 unsigned int alloc_flags, int start_type, bool whole_block)
1803 unsigned int current_order = buddy_order(page);
1804 int free_pages, movable_pages, alike_pages;
1807 old_block_type = get_pageblock_migratetype(page);
1810 * This can happen due to races and we want to prevent broken
1811 * highatomic accounting.
1813 if (is_migrate_highatomic(old_block_type))
1816 /* Take ownership for orders >= pageblock_order */
1817 if (current_order >= pageblock_order) {
1818 change_pageblock_range(page, current_order, start_type);
1823 * Boost watermarks to increase reclaim pressure to reduce the
1824 * likelihood of future fallbacks. Wake kswapd now as the node
1825 * may be balanced overall and kswapd will not wake naturally.
1827 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1828 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1830 /* We are not allowed to try stealing from the whole block */
1834 free_pages = move_freepages_block(zone, page, start_type,
1837 * Determine how many pages are compatible with our allocation.
1838 * For movable allocation, it's the number of movable pages which
1839 * we just obtained. For other types it's a bit more tricky.
1841 if (start_type == MIGRATE_MOVABLE) {
1842 alike_pages = movable_pages;
1845 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1846 * to MOVABLE pageblock, consider all non-movable pages as
1847 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1848 * vice versa, be conservative since we can't distinguish the
1849 * exact migratetype of non-movable pages.
1851 if (old_block_type == MIGRATE_MOVABLE)
1852 alike_pages = pageblock_nr_pages
1853 - (free_pages + movable_pages);
1858 /* moving whole block can fail due to zone boundary conditions */
1863 * If a sufficient number of pages in the block are either free or of
1864 * comparable migratability as our allocation, claim the whole block.
1866 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1867 page_group_by_mobility_disabled)
1868 set_pageblock_migratetype(page, start_type);
1873 move_to_free_list(page, zone, current_order, start_type);
1877 * Check whether there is a suitable fallback freepage with requested order.
1878 * If only_stealable is true, this function returns fallback_mt only if
1879 * we can steal other freepages all together. This would help to reduce
1880 * fragmentation due to mixed migratetype pages in one pageblock.
1882 int find_suitable_fallback(struct free_area *area, unsigned int order,
1883 int migratetype, bool only_stealable, bool *can_steal)
1888 if (area->nr_free == 0)
1892 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
1893 fallback_mt = fallbacks[migratetype][i];
1894 if (free_area_empty(area, fallback_mt))
1897 if (can_steal_fallback(order, migratetype))
1900 if (!only_stealable)
1911 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1912 * there are no empty page blocks that contain a page with a suitable order
1914 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1915 unsigned int alloc_order)
1918 unsigned long max_managed, flags;
1921 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1922 * Check is race-prone but harmless.
1924 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
1925 if (zone->nr_reserved_highatomic >= max_managed)
1928 spin_lock_irqsave(&zone->lock, flags);
1930 /* Recheck the nr_reserved_highatomic limit under the lock */
1931 if (zone->nr_reserved_highatomic >= max_managed)
1935 mt = get_pageblock_migratetype(page);
1936 /* Only reserve normal pageblocks (i.e., they can merge with others) */
1937 if (migratetype_is_mergeable(mt)) {
1938 zone->nr_reserved_highatomic += pageblock_nr_pages;
1939 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1940 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
1944 spin_unlock_irqrestore(&zone->lock, flags);
1948 * Used when an allocation is about to fail under memory pressure. This
1949 * potentially hurts the reliability of high-order allocations when under
1950 * intense memory pressure but failed atomic allocations should be easier
1951 * to recover from than an OOM.
1953 * If @force is true, try to unreserve a pageblock even though highatomic
1954 * pageblock is exhausted.
1956 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
1959 struct zonelist *zonelist = ac->zonelist;
1960 unsigned long flags;
1967 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
1970 * Preserve at least one pageblock unless memory pressure
1973 if (!force && zone->nr_reserved_highatomic <=
1977 spin_lock_irqsave(&zone->lock, flags);
1978 for (order = 0; order <= MAX_ORDER; order++) {
1979 struct free_area *area = &(zone->free_area[order]);
1981 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
1986 * In page freeing path, migratetype change is racy so
1987 * we can counter several free pages in a pageblock
1988 * in this loop although we changed the pageblock type
1989 * from highatomic to ac->migratetype. So we should
1990 * adjust the count once.
1992 if (is_migrate_highatomic_page(page)) {
1994 * It should never happen but changes to
1995 * locking could inadvertently allow a per-cpu
1996 * drain to add pages to MIGRATE_HIGHATOMIC
1997 * while unreserving so be safe and watch for
2000 zone->nr_reserved_highatomic -= min(
2002 zone->nr_reserved_highatomic);
2006 * Convert to ac->migratetype and avoid the normal
2007 * pageblock stealing heuristics. Minimally, the caller
2008 * is doing the work and needs the pages. More
2009 * importantly, if the block was always converted to
2010 * MIGRATE_UNMOVABLE or another type then the number
2011 * of pageblocks that cannot be completely freed
2014 set_pageblock_migratetype(page, ac->migratetype);
2015 ret = move_freepages_block(zone, page, ac->migratetype,
2018 spin_unlock_irqrestore(&zone->lock, flags);
2022 spin_unlock_irqrestore(&zone->lock, flags);
2029 * Try finding a free buddy page on the fallback list and put it on the free
2030 * list of requested migratetype, possibly along with other pages from the same
2031 * block, depending on fragmentation avoidance heuristics. Returns true if
2032 * fallback was found so that __rmqueue_smallest() can grab it.
2034 * The use of signed ints for order and current_order is a deliberate
2035 * deviation from the rest of this file, to make the for loop
2036 * condition simpler.
2038 static __always_inline bool
2039 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2040 unsigned int alloc_flags)
2042 struct free_area *area;
2044 int min_order = order;
2050 * Do not steal pages from freelists belonging to other pageblocks
2051 * i.e. orders < pageblock_order. If there are no local zones free,
2052 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2054 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2055 min_order = pageblock_order;
2058 * Find the largest available free page in the other list. This roughly
2059 * approximates finding the pageblock with the most free pages, which
2060 * would be too costly to do exactly.
2062 for (current_order = MAX_ORDER; current_order >= min_order;
2064 area = &(zone->free_area[current_order]);
2065 fallback_mt = find_suitable_fallback(area, current_order,
2066 start_migratetype, false, &can_steal);
2067 if (fallback_mt == -1)
2071 * We cannot steal all free pages from the pageblock and the
2072 * requested migratetype is movable. In that case it's better to
2073 * steal and split the smallest available page instead of the
2074 * largest available page, because even if the next movable
2075 * allocation falls back into a different pageblock than this
2076 * one, it won't cause permanent fragmentation.
2078 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2079 && current_order > order)
2088 for (current_order = order; current_order <= MAX_ORDER;
2090 area = &(zone->free_area[current_order]);
2091 fallback_mt = find_suitable_fallback(area, current_order,
2092 start_migratetype, false, &can_steal);
2093 if (fallback_mt != -1)
2098 * This should not happen - we already found a suitable fallback
2099 * when looking for the largest page.
2101 VM_BUG_ON(current_order > MAX_ORDER);
2104 page = get_page_from_free_area(area, fallback_mt);
2106 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2109 trace_mm_page_alloc_extfrag(page, order, current_order,
2110 start_migratetype, fallback_mt);
2117 * Do the hard work of removing an element from the buddy allocator.
2118 * Call me with the zone->lock already held.
2120 static __always_inline struct page *
2121 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2122 unsigned int alloc_flags)
2126 if (IS_ENABLED(CONFIG_CMA)) {
2128 * Balance movable allocations between regular and CMA areas by
2129 * allocating from CMA when over half of the zone's free memory
2130 * is in the CMA area.
2132 if (alloc_flags & ALLOC_CMA &&
2133 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2134 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2135 page = __rmqueue_cma_fallback(zone, order);
2141 page = __rmqueue_smallest(zone, order, migratetype);
2142 if (unlikely(!page)) {
2143 if (alloc_flags & ALLOC_CMA)
2144 page = __rmqueue_cma_fallback(zone, order);
2146 if (!page && __rmqueue_fallback(zone, order, migratetype,
2154 * Obtain a specified number of elements from the buddy allocator, all under
2155 * a single hold of the lock, for efficiency. Add them to the supplied list.
2156 * Returns the number of new pages which were placed at *list.
2158 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2159 unsigned long count, struct list_head *list,
2160 int migratetype, unsigned int alloc_flags)
2162 unsigned long flags;
2165 spin_lock_irqsave(&zone->lock, flags);
2166 for (i = 0; i < count; ++i) {
2167 struct page *page = __rmqueue(zone, order, migratetype,
2169 if (unlikely(page == NULL))
2173 * Split buddy pages returned by expand() are received here in
2174 * physical page order. The page is added to the tail of
2175 * caller's list. From the callers perspective, the linked list
2176 * is ordered by page number under some conditions. This is
2177 * useful for IO devices that can forward direction from the
2178 * head, thus also in the physical page order. This is useful
2179 * for IO devices that can merge IO requests if the physical
2180 * pages are ordered properly.
2182 list_add_tail(&page->pcp_list, list);
2183 if (is_migrate_cma(get_pcppage_migratetype(page)))
2184 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2188 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2189 spin_unlock_irqrestore(&zone->lock, flags);
2196 * Called from the vmstat counter updater to drain pagesets of this
2197 * currently executing processor on remote nodes after they have
2200 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2202 int to_drain, batch;
2204 batch = READ_ONCE(pcp->batch);
2205 to_drain = min(pcp->count, batch);
2207 spin_lock(&pcp->lock);
2208 free_pcppages_bulk(zone, to_drain, pcp, 0);
2209 spin_unlock(&pcp->lock);
2215 * Drain pcplists of the indicated processor and zone.
2217 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2219 struct per_cpu_pages *pcp;
2221 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2223 spin_lock(&pcp->lock);
2224 free_pcppages_bulk(zone, pcp->count, pcp, 0);
2225 spin_unlock(&pcp->lock);
2230 * Drain pcplists of all zones on the indicated processor.
2232 static void drain_pages(unsigned int cpu)
2236 for_each_populated_zone(zone) {
2237 drain_pages_zone(cpu, zone);
2242 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2244 void drain_local_pages(struct zone *zone)
2246 int cpu = smp_processor_id();
2249 drain_pages_zone(cpu, zone);
2255 * The implementation of drain_all_pages(), exposing an extra parameter to
2256 * drain on all cpus.
2258 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2259 * not empty. The check for non-emptiness can however race with a free to
2260 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2261 * that need the guarantee that every CPU has drained can disable the
2262 * optimizing racy check.
2264 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2269 * Allocate in the BSS so we won't require allocation in
2270 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2272 static cpumask_t cpus_with_pcps;
2275 * Do not drain if one is already in progress unless it's specific to
2276 * a zone. Such callers are primarily CMA and memory hotplug and need
2277 * the drain to be complete when the call returns.
2279 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2282 mutex_lock(&pcpu_drain_mutex);
2286 * We don't care about racing with CPU hotplug event
2287 * as offline notification will cause the notified
2288 * cpu to drain that CPU pcps and on_each_cpu_mask
2289 * disables preemption as part of its processing
2291 for_each_online_cpu(cpu) {
2292 struct per_cpu_pages *pcp;
2294 bool has_pcps = false;
2296 if (force_all_cpus) {
2298 * The pcp.count check is racy, some callers need a
2299 * guarantee that no cpu is missed.
2303 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2307 for_each_populated_zone(z) {
2308 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2317 cpumask_set_cpu(cpu, &cpus_with_pcps);
2319 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2322 for_each_cpu(cpu, &cpus_with_pcps) {
2324 drain_pages_zone(cpu, zone);
2329 mutex_unlock(&pcpu_drain_mutex);
2333 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2335 * When zone parameter is non-NULL, spill just the single zone's pages.
2337 void drain_all_pages(struct zone *zone)
2339 __drain_all_pages(zone, false);
2342 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
2347 if (!free_pages_prepare(page, order, FPI_NONE))
2350 migratetype = get_pfnblock_migratetype(page, pfn);
2351 set_pcppage_migratetype(page, migratetype);
2355 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
2358 int min_nr_free, max_nr_free;
2360 /* Free everything if batch freeing high-order pages. */
2361 if (unlikely(free_high))
2364 /* Check for PCP disabled or boot pageset */
2365 if (unlikely(high < batch))
2368 /* Leave at least pcp->batch pages on the list */
2369 min_nr_free = batch;
2370 max_nr_free = high - batch;
2373 * Double the number of pages freed each time there is subsequent
2374 * freeing of pages without any allocation.
2376 batch <<= pcp->free_factor;
2377 if (batch < max_nr_free)
2379 batch = clamp(batch, min_nr_free, max_nr_free);
2384 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2387 int high = READ_ONCE(pcp->high);
2389 if (unlikely(!high || free_high))
2392 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
2396 * If reclaim is active, limit the number of pages that can be
2397 * stored on pcp lists
2399 return min(READ_ONCE(pcp->batch) << 2, high);
2402 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2403 struct page *page, int migratetype,
2410 __count_vm_events(PGFREE, 1 << order);
2411 pindex = order_to_pindex(migratetype, order);
2412 list_add(&page->pcp_list, &pcp->lists[pindex]);
2413 pcp->count += 1 << order;
2416 * As high-order pages other than THP's stored on PCP can contribute
2417 * to fragmentation, limit the number stored when PCP is heavily
2418 * freeing without allocation. The remainder after bulk freeing
2419 * stops will be drained from vmstat refresh context.
2421 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
2423 high = nr_pcp_high(pcp, zone, free_high);
2424 if (pcp->count >= high) {
2425 int batch = READ_ONCE(pcp->batch);
2427 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
2434 void free_unref_page(struct page *page, unsigned int order)
2436 unsigned long __maybe_unused UP_flags;
2437 struct per_cpu_pages *pcp;
2439 unsigned long pfn = page_to_pfn(page);
2442 if (!free_unref_page_prepare(page, pfn, order))
2446 * We only track unmovable, reclaimable and movable on pcp lists.
2447 * Place ISOLATE pages on the isolated list because they are being
2448 * offlined but treat HIGHATOMIC as movable pages so we can get those
2449 * areas back if necessary. Otherwise, we may have to free
2450 * excessively into the page allocator
2452 migratetype = get_pcppage_migratetype(page);
2453 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2454 if (unlikely(is_migrate_isolate(migratetype))) {
2455 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
2458 migratetype = MIGRATE_MOVABLE;
2461 zone = page_zone(page);
2462 pcp_trylock_prepare(UP_flags);
2463 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2465 free_unref_page_commit(zone, pcp, page, migratetype, order);
2466 pcp_spin_unlock(pcp);
2468 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
2470 pcp_trylock_finish(UP_flags);
2474 * Free a list of 0-order pages
2476 void free_unref_page_list(struct list_head *list)
2478 unsigned long __maybe_unused UP_flags;
2479 struct page *page, *next;
2480 struct per_cpu_pages *pcp = NULL;
2481 struct zone *locked_zone = NULL;
2482 int batch_count = 0;
2485 /* Prepare pages for freeing */
2486 list_for_each_entry_safe(page, next, list, lru) {
2487 unsigned long pfn = page_to_pfn(page);
2488 if (!free_unref_page_prepare(page, pfn, 0)) {
2489 list_del(&page->lru);
2494 * Free isolated pages directly to the allocator, see
2495 * comment in free_unref_page.
2497 migratetype = get_pcppage_migratetype(page);
2498 if (unlikely(is_migrate_isolate(migratetype))) {
2499 list_del(&page->lru);
2500 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
2505 list_for_each_entry_safe(page, next, list, lru) {
2506 struct zone *zone = page_zone(page);
2508 list_del(&page->lru);
2509 migratetype = get_pcppage_migratetype(page);
2512 * Either different zone requiring a different pcp lock or
2513 * excessive lock hold times when freeing a large list of
2516 if (zone != locked_zone || batch_count == SWAP_CLUSTER_MAX) {
2518 pcp_spin_unlock(pcp);
2519 pcp_trylock_finish(UP_flags);
2525 * trylock is necessary as pages may be getting freed
2526 * from IRQ or SoftIRQ context after an IO completion.
2528 pcp_trylock_prepare(UP_flags);
2529 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2530 if (unlikely(!pcp)) {
2531 pcp_trylock_finish(UP_flags);
2532 free_one_page(zone, page, page_to_pfn(page),
2533 0, migratetype, FPI_NONE);
2541 * Non-isolated types over MIGRATE_PCPTYPES get added
2542 * to the MIGRATE_MOVABLE pcp list.
2544 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2545 migratetype = MIGRATE_MOVABLE;
2547 trace_mm_page_free_batched(page);
2548 free_unref_page_commit(zone, pcp, page, migratetype, 0);
2553 pcp_spin_unlock(pcp);
2554 pcp_trylock_finish(UP_flags);
2559 * split_page takes a non-compound higher-order page, and splits it into
2560 * n (1<<order) sub-pages: page[0..n]
2561 * Each sub-page must be freed individually.
2563 * Note: this is probably too low level an operation for use in drivers.
2564 * Please consult with lkml before using this in your driver.
2566 void split_page(struct page *page, unsigned int order)
2570 VM_BUG_ON_PAGE(PageCompound(page), page);
2571 VM_BUG_ON_PAGE(!page_count(page), page);
2573 for (i = 1; i < (1 << order); i++)
2574 set_page_refcounted(page + i);
2575 split_page_owner(page, 1 << order);
2576 split_page_memcg(page, 1 << order);
2578 EXPORT_SYMBOL_GPL(split_page);
2580 int __isolate_free_page(struct page *page, unsigned int order)
2582 struct zone *zone = page_zone(page);
2583 int mt = get_pageblock_migratetype(page);
2585 if (!is_migrate_isolate(mt)) {
2586 unsigned long watermark;
2588 * Obey watermarks as if the page was being allocated. We can
2589 * emulate a high-order watermark check with a raised order-0
2590 * watermark, because we already know our high-order page
2593 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2594 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2597 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2600 del_page_from_free_list(page, zone, order);
2603 * Set the pageblock if the isolated page is at least half of a
2606 if (order >= pageblock_order - 1) {
2607 struct page *endpage = page + (1 << order) - 1;
2608 for (; page < endpage; page += pageblock_nr_pages) {
2609 int mt = get_pageblock_migratetype(page);
2611 * Only change normal pageblocks (i.e., they can merge
2614 if (migratetype_is_mergeable(mt))
2615 set_pageblock_migratetype(page,
2620 return 1UL << order;
2624 * __putback_isolated_page - Return a now-isolated page back where we got it
2625 * @page: Page that was isolated
2626 * @order: Order of the isolated page
2627 * @mt: The page's pageblock's migratetype
2629 * This function is meant to return a page pulled from the free lists via
2630 * __isolate_free_page back to the free lists they were pulled from.
2632 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2634 struct zone *zone = page_zone(page);
2636 /* zone lock should be held when this function is called */
2637 lockdep_assert_held(&zone->lock);
2639 /* Return isolated page to tail of freelist. */
2640 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2641 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2645 * Update NUMA hit/miss statistics
2647 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2651 enum numa_stat_item local_stat = NUMA_LOCAL;
2653 /* skip numa counters update if numa stats is disabled */
2654 if (!static_branch_likely(&vm_numa_stat_key))
2657 if (zone_to_nid(z) != numa_node_id())
2658 local_stat = NUMA_OTHER;
2660 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2661 __count_numa_events(z, NUMA_HIT, nr_account);
2663 __count_numa_events(z, NUMA_MISS, nr_account);
2664 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2666 __count_numa_events(z, local_stat, nr_account);
2670 static __always_inline
2671 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2672 unsigned int order, unsigned int alloc_flags,
2676 unsigned long flags;
2680 spin_lock_irqsave(&zone->lock, flags);
2682 * order-0 request can reach here when the pcplist is skipped
2683 * due to non-CMA allocation context. HIGHATOMIC area is
2684 * reserved for high-order atomic allocation, so order-0
2685 * request should skip it.
2687 if (alloc_flags & ALLOC_HIGHATOMIC)
2688 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2690 page = __rmqueue(zone, order, migratetype, alloc_flags);
2693 * If the allocation fails, allow OOM handling access
2694 * to HIGHATOMIC reserves as failing now is worse than
2695 * failing a high-order atomic allocation in the
2698 if (!page && (alloc_flags & ALLOC_OOM))
2699 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2702 spin_unlock_irqrestore(&zone->lock, flags);
2706 __mod_zone_freepage_state(zone, -(1 << order),
2707 get_pcppage_migratetype(page));
2708 spin_unlock_irqrestore(&zone->lock, flags);
2709 } while (check_new_pages(page, order));
2711 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2712 zone_statistics(preferred_zone, zone, 1);
2717 /* Remove page from the per-cpu list, caller must protect the list */
2719 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2721 unsigned int alloc_flags,
2722 struct per_cpu_pages *pcp,
2723 struct list_head *list)
2728 if (list_empty(list)) {
2729 int batch = READ_ONCE(pcp->batch);
2733 * Scale batch relative to order if batch implies
2734 * free pages can be stored on the PCP. Batch can
2735 * be 1 for small zones or for boot pagesets which
2736 * should never store free pages as the pages may
2737 * belong to arbitrary zones.
2740 batch = max(batch >> order, 2);
2741 alloced = rmqueue_bulk(zone, order,
2743 migratetype, alloc_flags);
2745 pcp->count += alloced << order;
2746 if (unlikely(list_empty(list)))
2750 page = list_first_entry(list, struct page, pcp_list);
2751 list_del(&page->pcp_list);
2752 pcp->count -= 1 << order;
2753 } while (check_new_pages(page, order));
2758 /* Lock and remove page from the per-cpu list */
2759 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2760 struct zone *zone, unsigned int order,
2761 int migratetype, unsigned int alloc_flags)
2763 struct per_cpu_pages *pcp;
2764 struct list_head *list;
2766 unsigned long __maybe_unused UP_flags;
2768 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2769 pcp_trylock_prepare(UP_flags);
2770 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2772 pcp_trylock_finish(UP_flags);
2777 * On allocation, reduce the number of pages that are batch freed.
2778 * See nr_pcp_free() where free_factor is increased for subsequent
2781 pcp->free_factor >>= 1;
2782 list = &pcp->lists[order_to_pindex(migratetype, order)];
2783 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
2784 pcp_spin_unlock(pcp);
2785 pcp_trylock_finish(UP_flags);
2787 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2788 zone_statistics(preferred_zone, zone, 1);
2794 * Allocate a page from the given zone.
2795 * Use pcplists for THP or "cheap" high-order allocations.
2799 * Do not instrument rmqueue() with KMSAN. This function may call
2800 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
2801 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
2802 * may call rmqueue() again, which will result in a deadlock.
2804 __no_sanitize_memory
2806 struct page *rmqueue(struct zone *preferred_zone,
2807 struct zone *zone, unsigned int order,
2808 gfp_t gfp_flags, unsigned int alloc_flags,
2814 * We most definitely don't want callers attempting to
2815 * allocate greater than order-1 page units with __GFP_NOFAIL.
2817 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2819 if (likely(pcp_allowed_order(order))) {
2821 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
2822 * we need to skip it when CMA area isn't allowed.
2824 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
2825 migratetype != MIGRATE_MOVABLE) {
2826 page = rmqueue_pcplist(preferred_zone, zone, order,
2827 migratetype, alloc_flags);
2833 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
2837 /* Separate test+clear to avoid unnecessary atomics */
2838 if ((alloc_flags & ALLOC_KSWAPD) &&
2839 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
2840 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2841 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
2844 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2848 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2850 return __should_fail_alloc_page(gfp_mask, order);
2852 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
2854 static inline long __zone_watermark_unusable_free(struct zone *z,
2855 unsigned int order, unsigned int alloc_flags)
2857 long unusable_free = (1 << order) - 1;
2860 * If the caller does not have rights to reserves below the min
2861 * watermark then subtract the high-atomic reserves. This will
2862 * over-estimate the size of the atomic reserve but it avoids a search.
2864 if (likely(!(alloc_flags & ALLOC_RESERVES)))
2865 unusable_free += z->nr_reserved_highatomic;
2868 /* If allocation can't use CMA areas don't use free CMA pages */
2869 if (!(alloc_flags & ALLOC_CMA))
2870 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
2872 #ifdef CONFIG_UNACCEPTED_MEMORY
2873 unusable_free += zone_page_state(z, NR_UNACCEPTED);
2876 return unusable_free;
2880 * Return true if free base pages are above 'mark'. For high-order checks it
2881 * will return true of the order-0 watermark is reached and there is at least
2882 * one free page of a suitable size. Checking now avoids taking the zone lock
2883 * to check in the allocation paths if no pages are free.
2885 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2886 int highest_zoneidx, unsigned int alloc_flags,
2892 /* free_pages may go negative - that's OK */
2893 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
2895 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
2897 * __GFP_HIGH allows access to 50% of the min reserve as well
2900 if (alloc_flags & ALLOC_MIN_RESERVE) {
2904 * Non-blocking allocations (e.g. GFP_ATOMIC) can
2905 * access more reserves than just __GFP_HIGH. Other
2906 * non-blocking allocations requests such as GFP_NOWAIT
2907 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
2908 * access to the min reserve.
2910 if (alloc_flags & ALLOC_NON_BLOCK)
2915 * OOM victims can try even harder than the normal reserve
2916 * users on the grounds that it's definitely going to be in
2917 * the exit path shortly and free memory. Any allocation it
2918 * makes during the free path will be small and short-lived.
2920 if (alloc_flags & ALLOC_OOM)
2925 * Check watermarks for an order-0 allocation request. If these
2926 * are not met, then a high-order request also cannot go ahead
2927 * even if a suitable page happened to be free.
2929 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
2932 /* If this is an order-0 request then the watermark is fine */
2936 /* For a high-order request, check at least one suitable page is free */
2937 for (o = order; o <= MAX_ORDER; o++) {
2938 struct free_area *area = &z->free_area[o];
2944 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2945 if (!free_area_empty(area, mt))
2950 if ((alloc_flags & ALLOC_CMA) &&
2951 !free_area_empty(area, MIGRATE_CMA)) {
2955 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
2956 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
2963 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2964 int highest_zoneidx, unsigned int alloc_flags)
2966 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
2967 zone_page_state(z, NR_FREE_PAGES));
2970 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2971 unsigned long mark, int highest_zoneidx,
2972 unsigned int alloc_flags, gfp_t gfp_mask)
2976 free_pages = zone_page_state(z, NR_FREE_PAGES);
2979 * Fast check for order-0 only. If this fails then the reserves
2980 * need to be calculated.
2986 usable_free = free_pages;
2987 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
2989 /* reserved may over estimate high-atomic reserves. */
2990 usable_free -= min(usable_free, reserved);
2991 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
2995 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3000 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3001 * when checking the min watermark. The min watermark is the
3002 * point where boosting is ignored so that kswapd is woken up
3003 * when below the low watermark.
3005 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3006 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3007 mark = z->_watermark[WMARK_MIN];
3008 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3009 alloc_flags, free_pages);
3015 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3016 unsigned long mark, int highest_zoneidx)
3018 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3020 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3021 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3023 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3028 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3030 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3032 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3033 node_reclaim_distance;
3035 #else /* CONFIG_NUMA */
3036 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3040 #endif /* CONFIG_NUMA */
3043 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3044 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3045 * premature use of a lower zone may cause lowmem pressure problems that
3046 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3047 * probably too small. It only makes sense to spread allocations to avoid
3048 * fragmentation between the Normal and DMA32 zones.
3050 static inline unsigned int
3051 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3053 unsigned int alloc_flags;
3056 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3059 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3061 #ifdef CONFIG_ZONE_DMA32
3065 if (zone_idx(zone) != ZONE_NORMAL)
3069 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3070 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3071 * on UMA that if Normal is populated then so is DMA32.
3073 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3074 if (nr_online_nodes > 1 && !populated_zone(--zone))
3077 alloc_flags |= ALLOC_NOFRAGMENT;
3078 #endif /* CONFIG_ZONE_DMA32 */
3082 /* Must be called after current_gfp_context() which can change gfp_mask */
3083 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3084 unsigned int alloc_flags)
3087 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3088 alloc_flags |= ALLOC_CMA;
3094 * get_page_from_freelist goes through the zonelist trying to allocate
3097 static struct page *
3098 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3099 const struct alloc_context *ac)
3103 struct pglist_data *last_pgdat = NULL;
3104 bool last_pgdat_dirty_ok = false;
3109 * Scan zonelist, looking for a zone with enough free.
3110 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3112 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3113 z = ac->preferred_zoneref;
3114 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3119 if (cpusets_enabled() &&
3120 (alloc_flags & ALLOC_CPUSET) &&
3121 !__cpuset_zone_allowed(zone, gfp_mask))
3124 * When allocating a page cache page for writing, we
3125 * want to get it from a node that is within its dirty
3126 * limit, such that no single node holds more than its
3127 * proportional share of globally allowed dirty pages.
3128 * The dirty limits take into account the node's
3129 * lowmem reserves and high watermark so that kswapd
3130 * should be able to balance it without having to
3131 * write pages from its LRU list.
3133 * XXX: For now, allow allocations to potentially
3134 * exceed the per-node dirty limit in the slowpath
3135 * (spread_dirty_pages unset) before going into reclaim,
3136 * which is important when on a NUMA setup the allowed
3137 * nodes are together not big enough to reach the
3138 * global limit. The proper fix for these situations
3139 * will require awareness of nodes in the
3140 * dirty-throttling and the flusher threads.
3142 if (ac->spread_dirty_pages) {
3143 if (last_pgdat != zone->zone_pgdat) {
3144 last_pgdat = zone->zone_pgdat;
3145 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3148 if (!last_pgdat_dirty_ok)
3152 if (no_fallback && nr_online_nodes > 1 &&
3153 zone != ac->preferred_zoneref->zone) {
3157 * If moving to a remote node, retry but allow
3158 * fragmenting fallbacks. Locality is more important
3159 * than fragmentation avoidance.
3161 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3162 if (zone_to_nid(zone) != local_nid) {
3163 alloc_flags &= ~ALLOC_NOFRAGMENT;
3168 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3169 if (!zone_watermark_fast(zone, order, mark,
3170 ac->highest_zoneidx, alloc_flags,
3174 if (has_unaccepted_memory()) {
3175 if (try_to_accept_memory(zone, order))
3179 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3181 * Watermark failed for this zone, but see if we can
3182 * grow this zone if it contains deferred pages.
3184 if (deferred_pages_enabled()) {
3185 if (_deferred_grow_zone(zone, order))
3189 /* Checked here to keep the fast path fast */
3190 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3191 if (alloc_flags & ALLOC_NO_WATERMARKS)
3194 if (!node_reclaim_enabled() ||
3195 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3198 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3200 case NODE_RECLAIM_NOSCAN:
3203 case NODE_RECLAIM_FULL:
3204 /* scanned but unreclaimable */
3207 /* did we reclaim enough */
3208 if (zone_watermark_ok(zone, order, mark,
3209 ac->highest_zoneidx, alloc_flags))
3217 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3218 gfp_mask, alloc_flags, ac->migratetype);
3220 prep_new_page(page, order, gfp_mask, alloc_flags);
3223 * If this is a high-order atomic allocation then check
3224 * if the pageblock should be reserved for the future
3226 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3227 reserve_highatomic_pageblock(page, zone, order);
3231 if (has_unaccepted_memory()) {
3232 if (try_to_accept_memory(zone, order))
3236 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3237 /* Try again if zone has deferred pages */
3238 if (deferred_pages_enabled()) {
3239 if (_deferred_grow_zone(zone, order))
3247 * It's possible on a UMA machine to get through all zones that are
3248 * fragmented. If avoiding fragmentation, reset and try again.
3251 alloc_flags &= ~ALLOC_NOFRAGMENT;
3258 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3260 unsigned int filter = SHOW_MEM_FILTER_NODES;
3263 * This documents exceptions given to allocations in certain
3264 * contexts that are allowed to allocate outside current's set
3267 if (!(gfp_mask & __GFP_NOMEMALLOC))
3268 if (tsk_is_oom_victim(current) ||
3269 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3270 filter &= ~SHOW_MEM_FILTER_NODES;
3271 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3272 filter &= ~SHOW_MEM_FILTER_NODES;
3274 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3277 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3279 struct va_format vaf;
3281 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3283 if ((gfp_mask & __GFP_NOWARN) ||
3284 !__ratelimit(&nopage_rs) ||
3285 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3288 va_start(args, fmt);
3291 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3292 current->comm, &vaf, gfp_mask, &gfp_mask,
3293 nodemask_pr_args(nodemask));
3296 cpuset_print_current_mems_allowed();
3299 warn_alloc_show_mem(gfp_mask, nodemask);
3302 static inline struct page *
3303 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3304 unsigned int alloc_flags,
3305 const struct alloc_context *ac)
3309 page = get_page_from_freelist(gfp_mask, order,
3310 alloc_flags|ALLOC_CPUSET, ac);
3312 * fallback to ignore cpuset restriction if our nodes
3316 page = get_page_from_freelist(gfp_mask, order,
3322 static inline struct page *
3323 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3324 const struct alloc_context *ac, unsigned long *did_some_progress)
3326 struct oom_control oc = {
3327 .zonelist = ac->zonelist,
3328 .nodemask = ac->nodemask,
3330 .gfp_mask = gfp_mask,
3335 *did_some_progress = 0;
3338 * Acquire the oom lock. If that fails, somebody else is
3339 * making progress for us.
3341 if (!mutex_trylock(&oom_lock)) {
3342 *did_some_progress = 1;
3343 schedule_timeout_uninterruptible(1);
3348 * Go through the zonelist yet one more time, keep very high watermark
3349 * here, this is only to catch a parallel oom killing, we must fail if
3350 * we're still under heavy pressure. But make sure that this reclaim
3351 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3352 * allocation which will never fail due to oom_lock already held.
3354 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3355 ~__GFP_DIRECT_RECLAIM, order,
3356 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3360 /* Coredumps can quickly deplete all memory reserves */
3361 if (current->flags & PF_DUMPCORE)
3363 /* The OOM killer will not help higher order allocs */
3364 if (order > PAGE_ALLOC_COSTLY_ORDER)
3367 * We have already exhausted all our reclaim opportunities without any
3368 * success so it is time to admit defeat. We will skip the OOM killer
3369 * because it is very likely that the caller has a more reasonable
3370 * fallback than shooting a random task.
3372 * The OOM killer may not free memory on a specific node.
3374 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3376 /* The OOM killer does not needlessly kill tasks for lowmem */
3377 if (ac->highest_zoneidx < ZONE_NORMAL)
3379 if (pm_suspended_storage())
3382 * XXX: GFP_NOFS allocations should rather fail than rely on
3383 * other request to make a forward progress.
3384 * We are in an unfortunate situation where out_of_memory cannot
3385 * do much for this context but let's try it to at least get
3386 * access to memory reserved if the current task is killed (see
3387 * out_of_memory). Once filesystems are ready to handle allocation
3388 * failures more gracefully we should just bail out here.
3391 /* Exhausted what can be done so it's blame time */
3392 if (out_of_memory(&oc) ||
3393 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3394 *did_some_progress = 1;
3397 * Help non-failing allocations by giving them access to memory
3400 if (gfp_mask & __GFP_NOFAIL)
3401 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3402 ALLOC_NO_WATERMARKS, ac);
3405 mutex_unlock(&oom_lock);
3410 * Maximum number of compaction retries with a progress before OOM
3411 * killer is consider as the only way to move forward.
3413 #define MAX_COMPACT_RETRIES 16
3415 #ifdef CONFIG_COMPACTION
3416 /* Try memory compaction for high-order allocations before reclaim */
3417 static struct page *
3418 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3419 unsigned int alloc_flags, const struct alloc_context *ac,
3420 enum compact_priority prio, enum compact_result *compact_result)
3422 struct page *page = NULL;
3423 unsigned long pflags;
3424 unsigned int noreclaim_flag;
3429 psi_memstall_enter(&pflags);
3430 delayacct_compact_start();
3431 noreclaim_flag = memalloc_noreclaim_save();
3433 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3436 memalloc_noreclaim_restore(noreclaim_flag);
3437 psi_memstall_leave(&pflags);
3438 delayacct_compact_end();
3440 if (*compact_result == COMPACT_SKIPPED)
3443 * At least in one zone compaction wasn't deferred or skipped, so let's
3444 * count a compaction stall
3446 count_vm_event(COMPACTSTALL);
3448 /* Prep a captured page if available */
3450 prep_new_page(page, order, gfp_mask, alloc_flags);
3452 /* Try get a page from the freelist if available */
3454 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3457 struct zone *zone = page_zone(page);
3459 zone->compact_blockskip_flush = false;
3460 compaction_defer_reset(zone, order, true);
3461 count_vm_event(COMPACTSUCCESS);
3466 * It's bad if compaction run occurs and fails. The most likely reason
3467 * is that pages exist, but not enough to satisfy watermarks.
3469 count_vm_event(COMPACTFAIL);
3477 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3478 enum compact_result compact_result,
3479 enum compact_priority *compact_priority,
3480 int *compaction_retries)
3482 int max_retries = MAX_COMPACT_RETRIES;
3485 int retries = *compaction_retries;
3486 enum compact_priority priority = *compact_priority;
3491 if (fatal_signal_pending(current))
3495 * Compaction was skipped due to a lack of free order-0
3496 * migration targets. Continue if reclaim can help.
3498 if (compact_result == COMPACT_SKIPPED) {
3499 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3504 * Compaction managed to coalesce some page blocks, but the
3505 * allocation failed presumably due to a race. Retry some.
3507 if (compact_result == COMPACT_SUCCESS) {
3509 * !costly requests are much more important than
3510 * __GFP_RETRY_MAYFAIL costly ones because they are de
3511 * facto nofail and invoke OOM killer to move on while
3512 * costly can fail and users are ready to cope with
3513 * that. 1/4 retries is rather arbitrary but we would
3514 * need much more detailed feedback from compaction to
3515 * make a better decision.
3517 if (order > PAGE_ALLOC_COSTLY_ORDER)
3520 if (++(*compaction_retries) <= max_retries) {
3527 * Compaction failed. Retry with increasing priority.
3529 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3530 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3532 if (*compact_priority > min_priority) {
3533 (*compact_priority)--;
3534 *compaction_retries = 0;
3538 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3542 static inline struct page *
3543 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3544 unsigned int alloc_flags, const struct alloc_context *ac,
3545 enum compact_priority prio, enum compact_result *compact_result)
3547 *compact_result = COMPACT_SKIPPED;
3552 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3553 enum compact_result compact_result,
3554 enum compact_priority *compact_priority,
3555 int *compaction_retries)
3560 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3564 * There are setups with compaction disabled which would prefer to loop
3565 * inside the allocator rather than hit the oom killer prematurely.
3566 * Let's give them a good hope and keep retrying while the order-0
3567 * watermarks are OK.
3569 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3570 ac->highest_zoneidx, ac->nodemask) {
3571 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3572 ac->highest_zoneidx, alloc_flags))
3577 #endif /* CONFIG_COMPACTION */
3579 #ifdef CONFIG_LOCKDEP
3580 static struct lockdep_map __fs_reclaim_map =
3581 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3583 static bool __need_reclaim(gfp_t gfp_mask)
3585 /* no reclaim without waiting on it */
3586 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3589 /* this guy won't enter reclaim */
3590 if (current->flags & PF_MEMALLOC)
3593 if (gfp_mask & __GFP_NOLOCKDEP)
3599 void __fs_reclaim_acquire(unsigned long ip)
3601 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3604 void __fs_reclaim_release(unsigned long ip)
3606 lock_release(&__fs_reclaim_map, ip);
3609 void fs_reclaim_acquire(gfp_t gfp_mask)
3611 gfp_mask = current_gfp_context(gfp_mask);
3613 if (__need_reclaim(gfp_mask)) {
3614 if (gfp_mask & __GFP_FS)
3615 __fs_reclaim_acquire(_RET_IP_);
3617 #ifdef CONFIG_MMU_NOTIFIER
3618 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3619 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3624 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3626 void fs_reclaim_release(gfp_t gfp_mask)
3628 gfp_mask = current_gfp_context(gfp_mask);
3630 if (__need_reclaim(gfp_mask)) {
3631 if (gfp_mask & __GFP_FS)
3632 __fs_reclaim_release(_RET_IP_);
3635 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3639 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3640 * have been rebuilt so allocation retries. Reader side does not lock and
3641 * retries the allocation if zonelist changes. Writer side is protected by the
3642 * embedded spin_lock.
3644 static DEFINE_SEQLOCK(zonelist_update_seq);
3646 static unsigned int zonelist_iter_begin(void)
3648 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3649 return read_seqbegin(&zonelist_update_seq);
3654 static unsigned int check_retry_zonelist(unsigned int seq)
3656 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3657 return read_seqretry(&zonelist_update_seq, seq);
3662 /* Perform direct synchronous page reclaim */
3663 static unsigned long
3664 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3665 const struct alloc_context *ac)
3667 unsigned int noreclaim_flag;
3668 unsigned long progress;
3672 /* We now go into synchronous reclaim */
3673 cpuset_memory_pressure_bump();
3674 fs_reclaim_acquire(gfp_mask);
3675 noreclaim_flag = memalloc_noreclaim_save();
3677 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3680 memalloc_noreclaim_restore(noreclaim_flag);
3681 fs_reclaim_release(gfp_mask);
3688 /* The really slow allocator path where we enter direct reclaim */
3689 static inline struct page *
3690 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3691 unsigned int alloc_flags, const struct alloc_context *ac,
3692 unsigned long *did_some_progress)
3694 struct page *page = NULL;
3695 unsigned long pflags;
3696 bool drained = false;
3698 psi_memstall_enter(&pflags);
3699 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3700 if (unlikely(!(*did_some_progress)))
3704 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3707 * If an allocation failed after direct reclaim, it could be because
3708 * pages are pinned on the per-cpu lists or in high alloc reserves.
3709 * Shrink them and try again
3711 if (!page && !drained) {
3712 unreserve_highatomic_pageblock(ac, false);
3713 drain_all_pages(NULL);
3718 psi_memstall_leave(&pflags);
3723 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3724 const struct alloc_context *ac)
3728 pg_data_t *last_pgdat = NULL;
3729 enum zone_type highest_zoneidx = ac->highest_zoneidx;
3731 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3733 if (!managed_zone(zone))
3735 if (last_pgdat != zone->zone_pgdat) {
3736 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3737 last_pgdat = zone->zone_pgdat;
3742 static inline unsigned int
3743 gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3745 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3748 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3749 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3750 * to save two branches.
3752 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
3753 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
3756 * The caller may dip into page reserves a bit more if the caller
3757 * cannot run direct reclaim, or if the caller has realtime scheduling
3758 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3759 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3761 alloc_flags |= (__force int)
3762 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
3764 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
3766 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3767 * if it can't schedule.
3769 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
3770 alloc_flags |= ALLOC_NON_BLOCK;
3773 alloc_flags |= ALLOC_HIGHATOMIC;
3777 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
3778 * GFP_ATOMIC) rather than fail, see the comment for
3779 * cpuset_node_allowed().
3781 if (alloc_flags & ALLOC_MIN_RESERVE)
3782 alloc_flags &= ~ALLOC_CPUSET;
3783 } else if (unlikely(rt_task(current)) && in_task())
3784 alloc_flags |= ALLOC_MIN_RESERVE;
3786 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
3791 static bool oom_reserves_allowed(struct task_struct *tsk)
3793 if (!tsk_is_oom_victim(tsk))
3797 * !MMU doesn't have oom reaper so give access to memory reserves
3798 * only to the thread with TIF_MEMDIE set
3800 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3807 * Distinguish requests which really need access to full memory
3808 * reserves from oom victims which can live with a portion of it
3810 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3812 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3814 if (gfp_mask & __GFP_MEMALLOC)
3815 return ALLOC_NO_WATERMARKS;
3816 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3817 return ALLOC_NO_WATERMARKS;
3818 if (!in_interrupt()) {
3819 if (current->flags & PF_MEMALLOC)
3820 return ALLOC_NO_WATERMARKS;
3821 else if (oom_reserves_allowed(current))
3828 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3830 return !!__gfp_pfmemalloc_flags(gfp_mask);
3834 * Checks whether it makes sense to retry the reclaim to make a forward progress
3835 * for the given allocation request.
3837 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3838 * without success, or when we couldn't even meet the watermark if we
3839 * reclaimed all remaining pages on the LRU lists.
3841 * Returns true if a retry is viable or false to enter the oom path.
3844 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3845 struct alloc_context *ac, int alloc_flags,
3846 bool did_some_progress, int *no_progress_loops)
3853 * Costly allocations might have made a progress but this doesn't mean
3854 * their order will become available due to high fragmentation so
3855 * always increment the no progress counter for them
3857 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3858 *no_progress_loops = 0;
3860 (*no_progress_loops)++;
3863 * Make sure we converge to OOM if we cannot make any progress
3864 * several times in the row.
3866 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3867 /* Before OOM, exhaust highatomic_reserve */
3868 return unreserve_highatomic_pageblock(ac, true);
3872 * Keep reclaiming pages while there is a chance this will lead
3873 * somewhere. If none of the target zones can satisfy our allocation
3874 * request even if all reclaimable pages are considered then we are
3875 * screwed and have to go OOM.
3877 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3878 ac->highest_zoneidx, ac->nodemask) {
3879 unsigned long available;
3880 unsigned long reclaimable;
3881 unsigned long min_wmark = min_wmark_pages(zone);
3884 available = reclaimable = zone_reclaimable_pages(zone);
3885 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3888 * Would the allocation succeed if we reclaimed all
3889 * reclaimable pages?
3891 wmark = __zone_watermark_ok(zone, order, min_wmark,
3892 ac->highest_zoneidx, alloc_flags, available);
3893 trace_reclaim_retry_zone(z, order, reclaimable,
3894 available, min_wmark, *no_progress_loops, wmark);
3902 * Memory allocation/reclaim might be called from a WQ context and the
3903 * current implementation of the WQ concurrency control doesn't
3904 * recognize that a particular WQ is congested if the worker thread is
3905 * looping without ever sleeping. Therefore we have to do a short sleep
3906 * here rather than calling cond_resched().
3908 if (current->flags & PF_WQ_WORKER)
3909 schedule_timeout_uninterruptible(1);
3916 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3919 * It's possible that cpuset's mems_allowed and the nodemask from
3920 * mempolicy don't intersect. This should be normally dealt with by
3921 * policy_nodemask(), but it's possible to race with cpuset update in
3922 * such a way the check therein was true, and then it became false
3923 * before we got our cpuset_mems_cookie here.
3924 * This assumes that for all allocations, ac->nodemask can come only
3925 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3926 * when it does not intersect with the cpuset restrictions) or the
3927 * caller can deal with a violated nodemask.
3929 if (cpusets_enabled() && ac->nodemask &&
3930 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3931 ac->nodemask = NULL;
3936 * When updating a task's mems_allowed or mempolicy nodemask, it is
3937 * possible to race with parallel threads in such a way that our
3938 * allocation can fail while the mask is being updated. If we are about
3939 * to fail, check if the cpuset changed during allocation and if so,
3942 if (read_mems_allowed_retry(cpuset_mems_cookie))
3948 static inline struct page *
3949 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3950 struct alloc_context *ac)
3952 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3953 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3954 struct page *page = NULL;
3955 unsigned int alloc_flags;
3956 unsigned long did_some_progress;
3957 enum compact_priority compact_priority;
3958 enum compact_result compact_result;
3959 int compaction_retries;
3960 int no_progress_loops;
3961 unsigned int cpuset_mems_cookie;
3962 unsigned int zonelist_iter_cookie;
3966 compaction_retries = 0;
3967 no_progress_loops = 0;
3968 compact_priority = DEF_COMPACT_PRIORITY;
3969 cpuset_mems_cookie = read_mems_allowed_begin();
3970 zonelist_iter_cookie = zonelist_iter_begin();
3973 * The fast path uses conservative alloc_flags to succeed only until
3974 * kswapd needs to be woken up, and to avoid the cost of setting up
3975 * alloc_flags precisely. So we do that now.
3977 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
3980 * We need to recalculate the starting point for the zonelist iterator
3981 * because we might have used different nodemask in the fast path, or
3982 * there was a cpuset modification and we are retrying - otherwise we
3983 * could end up iterating over non-eligible zones endlessly.
3985 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3986 ac->highest_zoneidx, ac->nodemask);
3987 if (!ac->preferred_zoneref->zone)
3991 * Check for insane configurations where the cpuset doesn't contain
3992 * any suitable zone to satisfy the request - e.g. non-movable
3993 * GFP_HIGHUSER allocations from MOVABLE nodes only.
3995 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
3996 struct zoneref *z = first_zones_zonelist(ac->zonelist,
3997 ac->highest_zoneidx,
3998 &cpuset_current_mems_allowed);
4003 if (alloc_flags & ALLOC_KSWAPD)
4004 wake_all_kswapds(order, gfp_mask, ac);
4007 * The adjusted alloc_flags might result in immediate success, so try
4010 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4015 * For costly allocations, try direct compaction first, as it's likely
4016 * that we have enough base pages and don't need to reclaim. For non-
4017 * movable high-order allocations, do that as well, as compaction will
4018 * try prevent permanent fragmentation by migrating from blocks of the
4020 * Don't try this for allocations that are allowed to ignore
4021 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4023 if (can_direct_reclaim &&
4025 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4026 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4027 page = __alloc_pages_direct_compact(gfp_mask, order,
4029 INIT_COMPACT_PRIORITY,
4035 * Checks for costly allocations with __GFP_NORETRY, which
4036 * includes some THP page fault allocations
4038 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4040 * If allocating entire pageblock(s) and compaction
4041 * failed because all zones are below low watermarks
4042 * or is prohibited because it recently failed at this
4043 * order, fail immediately unless the allocator has
4044 * requested compaction and reclaim retry.
4047 * - potentially very expensive because zones are far
4048 * below their low watermarks or this is part of very
4049 * bursty high order allocations,
4050 * - not guaranteed to help because isolate_freepages()
4051 * may not iterate over freed pages as part of its
4053 * - unlikely to make entire pageblocks free on its
4056 if (compact_result == COMPACT_SKIPPED ||
4057 compact_result == COMPACT_DEFERRED)
4061 * Looks like reclaim/compaction is worth trying, but
4062 * sync compaction could be very expensive, so keep
4063 * using async compaction.
4065 compact_priority = INIT_COMPACT_PRIORITY;
4070 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4071 if (alloc_flags & ALLOC_KSWAPD)
4072 wake_all_kswapds(order, gfp_mask, ac);
4074 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4076 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4077 (alloc_flags & ALLOC_KSWAPD);
4080 * Reset the nodemask and zonelist iterators if memory policies can be
4081 * ignored. These allocations are high priority and system rather than
4084 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4085 ac->nodemask = NULL;
4086 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4087 ac->highest_zoneidx, ac->nodemask);
4090 /* Attempt with potentially adjusted zonelist and alloc_flags */
4091 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4095 /* Caller is not willing to reclaim, we can't balance anything */
4096 if (!can_direct_reclaim)
4099 /* Avoid recursion of direct reclaim */
4100 if (current->flags & PF_MEMALLOC)
4103 /* Try direct reclaim and then allocating */
4104 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4105 &did_some_progress);
4109 /* Try direct compaction and then allocating */
4110 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4111 compact_priority, &compact_result);
4115 /* Do not loop if specifically requested */
4116 if (gfp_mask & __GFP_NORETRY)
4120 * Do not retry costly high order allocations unless they are
4121 * __GFP_RETRY_MAYFAIL
4123 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4126 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4127 did_some_progress > 0, &no_progress_loops))
4131 * It doesn't make any sense to retry for the compaction if the order-0
4132 * reclaim is not able to make any progress because the current
4133 * implementation of the compaction depends on the sufficient amount
4134 * of free memory (see __compaction_suitable)
4136 if (did_some_progress > 0 &&
4137 should_compact_retry(ac, order, alloc_flags,
4138 compact_result, &compact_priority,
4139 &compaction_retries))
4144 * Deal with possible cpuset update races or zonelist updates to avoid
4145 * a unnecessary OOM kill.
4147 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4148 check_retry_zonelist(zonelist_iter_cookie))
4151 /* Reclaim has failed us, start killing things */
4152 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4156 /* Avoid allocations with no watermarks from looping endlessly */
4157 if (tsk_is_oom_victim(current) &&
4158 (alloc_flags & ALLOC_OOM ||
4159 (gfp_mask & __GFP_NOMEMALLOC)))
4162 /* Retry as long as the OOM killer is making progress */
4163 if (did_some_progress) {
4164 no_progress_loops = 0;
4170 * Deal with possible cpuset update races or zonelist updates to avoid
4171 * a unnecessary OOM kill.
4173 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4174 check_retry_zonelist(zonelist_iter_cookie))
4178 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4181 if (gfp_mask & __GFP_NOFAIL) {
4183 * All existing users of the __GFP_NOFAIL are blockable, so warn
4184 * of any new users that actually require GFP_NOWAIT
4186 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4190 * PF_MEMALLOC request from this context is rather bizarre
4191 * because we cannot reclaim anything and only can loop waiting
4192 * for somebody to do a work for us
4194 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4197 * non failing costly orders are a hard requirement which we
4198 * are not prepared for much so let's warn about these users
4199 * so that we can identify them and convert them to something
4202 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4205 * Help non-failing allocations by giving some access to memory
4206 * reserves normally used for high priority non-blocking
4207 * allocations but do not use ALLOC_NO_WATERMARKS because this
4208 * could deplete whole memory reserves which would just make
4209 * the situation worse.
4211 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4219 warn_alloc(gfp_mask, ac->nodemask,
4220 "page allocation failure: order:%u", order);
4225 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4226 int preferred_nid, nodemask_t *nodemask,
4227 struct alloc_context *ac, gfp_t *alloc_gfp,
4228 unsigned int *alloc_flags)
4230 ac->highest_zoneidx = gfp_zone(gfp_mask);
4231 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4232 ac->nodemask = nodemask;
4233 ac->migratetype = gfp_migratetype(gfp_mask);
4235 if (cpusets_enabled()) {
4236 *alloc_gfp |= __GFP_HARDWALL;
4238 * When we are in the interrupt context, it is irrelevant
4239 * to the current task context. It means that any node ok.
4241 if (in_task() && !ac->nodemask)
4242 ac->nodemask = &cpuset_current_mems_allowed;
4244 *alloc_flags |= ALLOC_CPUSET;
4247 might_alloc(gfp_mask);
4249 if (should_fail_alloc_page(gfp_mask, order))
4252 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4254 /* Dirty zone balancing only done in the fast path */
4255 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4258 * The preferred zone is used for statistics but crucially it is
4259 * also used as the starting point for the zonelist iterator. It
4260 * may get reset for allocations that ignore memory policies.
4262 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4263 ac->highest_zoneidx, ac->nodemask);
4269 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4270 * @gfp: GFP flags for the allocation
4271 * @preferred_nid: The preferred NUMA node ID to allocate from
4272 * @nodemask: Set of nodes to allocate from, may be NULL
4273 * @nr_pages: The number of pages desired on the list or array
4274 * @page_list: Optional list to store the allocated pages
4275 * @page_array: Optional array to store the pages
4277 * This is a batched version of the page allocator that attempts to
4278 * allocate nr_pages quickly. Pages are added to page_list if page_list
4279 * is not NULL, otherwise it is assumed that the page_array is valid.
4281 * For lists, nr_pages is the number of pages that should be allocated.
4283 * For arrays, only NULL elements are populated with pages and nr_pages
4284 * is the maximum number of pages that will be stored in the array.
4286 * Returns the number of pages on the list or array.
4288 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
4289 nodemask_t *nodemask, int nr_pages,
4290 struct list_head *page_list,
4291 struct page **page_array)
4294 unsigned long __maybe_unused UP_flags;
4297 struct per_cpu_pages *pcp;
4298 struct list_head *pcp_list;
4299 struct alloc_context ac;
4301 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4302 int nr_populated = 0, nr_account = 0;
4305 * Skip populated array elements to determine if any pages need
4306 * to be allocated before disabling IRQs.
4308 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4311 /* No pages requested? */
4312 if (unlikely(nr_pages <= 0))
4315 /* Already populated array? */
4316 if (unlikely(page_array && nr_pages - nr_populated == 0))
4319 /* Bulk allocator does not support memcg accounting. */
4320 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4323 /* Use the single page allocator for one page. */
4324 if (nr_pages - nr_populated == 1)
4327 #ifdef CONFIG_PAGE_OWNER
4329 * PAGE_OWNER may recurse into the allocator to allocate space to
4330 * save the stack with pagesets.lock held. Releasing/reacquiring
4331 * removes much of the performance benefit of bulk allocation so
4332 * force the caller to allocate one page at a time as it'll have
4333 * similar performance to added complexity to the bulk allocator.
4335 if (static_branch_unlikely(&page_owner_inited))
4339 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4340 gfp &= gfp_allowed_mask;
4342 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4346 /* Find an allowed local zone that meets the low watermark. */
4347 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4350 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4351 !__cpuset_zone_allowed(zone, gfp)) {
4355 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4356 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4360 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4361 if (zone_watermark_fast(zone, 0, mark,
4362 zonelist_zone_idx(ac.preferred_zoneref),
4363 alloc_flags, gfp)) {
4369 * If there are no allowed local zones that meets the watermarks then
4370 * try to allocate a single page and reclaim if necessary.
4372 if (unlikely(!zone))
4375 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4376 pcp_trylock_prepare(UP_flags);
4377 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4381 /* Attempt the batch allocation */
4382 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4383 while (nr_populated < nr_pages) {
4385 /* Skip existing pages */
4386 if (page_array && page_array[nr_populated]) {
4391 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4393 if (unlikely(!page)) {
4394 /* Try and allocate at least one page */
4396 pcp_spin_unlock(pcp);
4403 prep_new_page(page, 0, gfp, 0);
4405 list_add(&page->lru, page_list);
4407 page_array[nr_populated] = page;
4411 pcp_spin_unlock(pcp);
4412 pcp_trylock_finish(UP_flags);
4414 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4415 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4418 return nr_populated;
4421 pcp_trylock_finish(UP_flags);
4424 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
4427 list_add(&page->lru, page_list);
4429 page_array[nr_populated] = page;
4435 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
4438 * This is the 'heart' of the zoned buddy allocator.
4440 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
4441 nodemask_t *nodemask)
4444 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4445 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4446 struct alloc_context ac = { };
4449 * There are several places where we assume that the order value is sane
4450 * so bail out early if the request is out of bound.
4452 if (WARN_ON_ONCE_GFP(order > MAX_ORDER, gfp))
4455 gfp &= gfp_allowed_mask;
4457 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4458 * resp. GFP_NOIO which has to be inherited for all allocation requests
4459 * from a particular context which has been marked by
4460 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4461 * movable zones are not used during allocation.
4463 gfp = current_gfp_context(gfp);
4465 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4466 &alloc_gfp, &alloc_flags))
4470 * Forbid the first pass from falling back to types that fragment
4471 * memory until all local zones are considered.
4473 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4475 /* First allocation attempt */
4476 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4481 ac.spread_dirty_pages = false;
4484 * Restore the original nodemask if it was potentially replaced with
4485 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4487 ac.nodemask = nodemask;
4489 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4492 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4493 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4494 __free_pages(page, order);
4498 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4499 kmsan_alloc_page(page, order, alloc_gfp);
4503 EXPORT_SYMBOL(__alloc_pages);
4505 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
4506 nodemask_t *nodemask)
4508 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
4509 preferred_nid, nodemask);
4511 if (page && order > 1)
4512 prep_transhuge_page(page);
4513 return (struct folio *)page;
4515 EXPORT_SYMBOL(__folio_alloc);
4518 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4519 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4520 * you need to access high mem.
4522 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4526 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4529 return (unsigned long) page_address(page);
4531 EXPORT_SYMBOL(__get_free_pages);
4533 unsigned long get_zeroed_page(gfp_t gfp_mask)
4535 return __get_free_page(gfp_mask | __GFP_ZERO);
4537 EXPORT_SYMBOL(get_zeroed_page);
4540 * __free_pages - Free pages allocated with alloc_pages().
4541 * @page: The page pointer returned from alloc_pages().
4542 * @order: The order of the allocation.
4544 * This function can free multi-page allocations that are not compound
4545 * pages. It does not check that the @order passed in matches that of
4546 * the allocation, so it is easy to leak memory. Freeing more memory
4547 * than was allocated will probably emit a warning.
4549 * If the last reference to this page is speculative, it will be released
4550 * by put_page() which only frees the first page of a non-compound
4551 * allocation. To prevent the remaining pages from being leaked, we free
4552 * the subsequent pages here. If you want to use the page's reference
4553 * count to decide when to free the allocation, you should allocate a
4554 * compound page, and use put_page() instead of __free_pages().
4556 * Context: May be called in interrupt context or while holding a normal
4557 * spinlock, but not in NMI context or while holding a raw spinlock.
4559 void __free_pages(struct page *page, unsigned int order)
4561 /* get PageHead before we drop reference */
4562 int head = PageHead(page);
4564 if (put_page_testzero(page))
4565 free_the_page(page, order);
4568 free_the_page(page + (1 << order), order);
4570 EXPORT_SYMBOL(__free_pages);
4572 void free_pages(unsigned long addr, unsigned int order)
4575 VM_BUG_ON(!virt_addr_valid((void *)addr));
4576 __free_pages(virt_to_page((void *)addr), order);
4580 EXPORT_SYMBOL(free_pages);
4584 * An arbitrary-length arbitrary-offset area of memory which resides
4585 * within a 0 or higher order page. Multiple fragments within that page
4586 * are individually refcounted, in the page's reference counter.
4588 * The page_frag functions below provide a simple allocation framework for
4589 * page fragments. This is used by the network stack and network device
4590 * drivers to provide a backing region of memory for use as either an
4591 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4593 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4596 struct page *page = NULL;
4597 gfp_t gfp = gfp_mask;
4599 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4600 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4602 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4603 PAGE_FRAG_CACHE_MAX_ORDER);
4604 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4606 if (unlikely(!page))
4607 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4609 nc->va = page ? page_address(page) : NULL;
4614 void __page_frag_cache_drain(struct page *page, unsigned int count)
4616 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4618 if (page_ref_sub_and_test(page, count))
4619 free_the_page(page, compound_order(page));
4621 EXPORT_SYMBOL(__page_frag_cache_drain);
4623 void *page_frag_alloc_align(struct page_frag_cache *nc,
4624 unsigned int fragsz, gfp_t gfp_mask,
4625 unsigned int align_mask)
4627 unsigned int size = PAGE_SIZE;
4631 if (unlikely(!nc->va)) {
4633 page = __page_frag_cache_refill(nc, gfp_mask);
4637 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4638 /* if size can vary use size else just use PAGE_SIZE */
4641 /* Even if we own the page, we do not use atomic_set().
4642 * This would break get_page_unless_zero() users.
4644 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4646 /* reset page count bias and offset to start of new frag */
4647 nc->pfmemalloc = page_is_pfmemalloc(page);
4648 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4652 offset = nc->offset - fragsz;
4653 if (unlikely(offset < 0)) {
4654 page = virt_to_page(nc->va);
4656 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4659 if (unlikely(nc->pfmemalloc)) {
4660 free_the_page(page, compound_order(page));
4664 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4665 /* if size can vary use size else just use PAGE_SIZE */
4668 /* OK, page count is 0, we can safely set it */
4669 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4671 /* reset page count bias and offset to start of new frag */
4672 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4673 offset = size - fragsz;
4674 if (unlikely(offset < 0)) {
4676 * The caller is trying to allocate a fragment
4677 * with fragsz > PAGE_SIZE but the cache isn't big
4678 * enough to satisfy the request, this may
4679 * happen in low memory conditions.
4680 * We don't release the cache page because
4681 * it could make memory pressure worse
4682 * so we simply return NULL here.
4689 offset &= align_mask;
4690 nc->offset = offset;
4692 return nc->va + offset;
4694 EXPORT_SYMBOL(page_frag_alloc_align);
4697 * Frees a page fragment allocated out of either a compound or order 0 page.
4699 void page_frag_free(void *addr)
4701 struct page *page = virt_to_head_page(addr);
4703 if (unlikely(put_page_testzero(page)))
4704 free_the_page(page, compound_order(page));
4706 EXPORT_SYMBOL(page_frag_free);
4708 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4712 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4713 struct page *page = virt_to_page((void *)addr);
4714 struct page *last = page + nr;
4716 split_page_owner(page, 1 << order);
4717 split_page_memcg(page, 1 << order);
4718 while (page < --last)
4719 set_page_refcounted(last);
4721 last = page + (1UL << order);
4722 for (page += nr; page < last; page++)
4723 __free_pages_ok(page, 0, FPI_TO_TAIL);
4725 return (void *)addr;
4729 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4730 * @size: the number of bytes to allocate
4731 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4733 * This function is similar to alloc_pages(), except that it allocates the
4734 * minimum number of pages to satisfy the request. alloc_pages() can only
4735 * allocate memory in power-of-two pages.
4737 * This function is also limited by MAX_ORDER.
4739 * Memory allocated by this function must be released by free_pages_exact().
4741 * Return: pointer to the allocated area or %NULL in case of error.
4743 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4745 unsigned int order = get_order(size);
4748 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4749 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4751 addr = __get_free_pages(gfp_mask, order);
4752 return make_alloc_exact(addr, order, size);
4754 EXPORT_SYMBOL(alloc_pages_exact);
4757 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4759 * @nid: the preferred node ID where memory should be allocated
4760 * @size: the number of bytes to allocate
4761 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4763 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4766 * Return: pointer to the allocated area or %NULL in case of error.
4768 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4770 unsigned int order = get_order(size);
4773 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4774 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4776 p = alloc_pages_node(nid, gfp_mask, order);
4779 return make_alloc_exact((unsigned long)page_address(p), order, size);
4783 * free_pages_exact - release memory allocated via alloc_pages_exact()
4784 * @virt: the value returned by alloc_pages_exact.
4785 * @size: size of allocation, same value as passed to alloc_pages_exact().
4787 * Release the memory allocated by a previous call to alloc_pages_exact.
4789 void free_pages_exact(void *virt, size_t size)
4791 unsigned long addr = (unsigned long)virt;
4792 unsigned long end = addr + PAGE_ALIGN(size);
4794 while (addr < end) {
4799 EXPORT_SYMBOL(free_pages_exact);
4802 * nr_free_zone_pages - count number of pages beyond high watermark
4803 * @offset: The zone index of the highest zone
4805 * nr_free_zone_pages() counts the number of pages which are beyond the
4806 * high watermark within all zones at or below a given zone index. For each
4807 * zone, the number of pages is calculated as:
4809 * nr_free_zone_pages = managed_pages - high_pages
4811 * Return: number of pages beyond high watermark.
4813 static unsigned long nr_free_zone_pages(int offset)
4818 /* Just pick one node, since fallback list is circular */
4819 unsigned long sum = 0;
4821 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4823 for_each_zone_zonelist(zone, z, zonelist, offset) {
4824 unsigned long size = zone_managed_pages(zone);
4825 unsigned long high = high_wmark_pages(zone);
4834 * nr_free_buffer_pages - count number of pages beyond high watermark
4836 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4837 * watermark within ZONE_DMA and ZONE_NORMAL.
4839 * Return: number of pages beyond high watermark within ZONE_DMA and
4842 unsigned long nr_free_buffer_pages(void)
4844 return nr_free_zone_pages(gfp_zone(GFP_USER));
4846 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4848 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4850 zoneref->zone = zone;
4851 zoneref->zone_idx = zone_idx(zone);
4855 * Builds allocation fallback zone lists.
4857 * Add all populated zones of a node to the zonelist.
4859 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4862 enum zone_type zone_type = MAX_NR_ZONES;
4867 zone = pgdat->node_zones + zone_type;
4868 if (populated_zone(zone)) {
4869 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4870 check_highest_zone(zone_type);
4872 } while (zone_type);
4879 static int __parse_numa_zonelist_order(char *s)
4882 * We used to support different zonelists modes but they turned
4883 * out to be just not useful. Let's keep the warning in place
4884 * if somebody still use the cmd line parameter so that we do
4885 * not fail it silently
4887 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4888 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4894 static char numa_zonelist_order[] = "Node";
4895 #define NUMA_ZONELIST_ORDER_LEN 16
4897 * sysctl handler for numa_zonelist_order
4899 static int numa_zonelist_order_handler(struct ctl_table *table, int write,
4900 void *buffer, size_t *length, loff_t *ppos)
4903 return __parse_numa_zonelist_order(buffer);
4904 return proc_dostring(table, write, buffer, length, ppos);
4907 static int node_load[MAX_NUMNODES];
4910 * find_next_best_node - find the next node that should appear in a given node's fallback list
4911 * @node: node whose fallback list we're appending
4912 * @used_node_mask: nodemask_t of already used nodes
4914 * We use a number of factors to determine which is the next node that should
4915 * appear on a given node's fallback list. The node should not have appeared
4916 * already in @node's fallback list, and it should be the next closest node
4917 * according to the distance array (which contains arbitrary distance values
4918 * from each node to each node in the system), and should also prefer nodes
4919 * with no CPUs, since presumably they'll have very little allocation pressure
4920 * on them otherwise.
4922 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
4924 int find_next_best_node(int node, nodemask_t *used_node_mask)
4927 int min_val = INT_MAX;
4928 int best_node = NUMA_NO_NODE;
4930 /* Use the local node if we haven't already */
4931 if (!node_isset(node, *used_node_mask)) {
4932 node_set(node, *used_node_mask);
4936 for_each_node_state(n, N_MEMORY) {
4938 /* Don't want a node to appear more than once */
4939 if (node_isset(n, *used_node_mask))
4942 /* Use the distance array to find the distance */
4943 val = node_distance(node, n);
4945 /* Penalize nodes under us ("prefer the next node") */
4948 /* Give preference to headless and unused nodes */
4949 if (!cpumask_empty(cpumask_of_node(n)))
4950 val += PENALTY_FOR_NODE_WITH_CPUS;
4952 /* Slight preference for less loaded node */
4953 val *= MAX_NUMNODES;
4954 val += node_load[n];
4956 if (val < min_val) {
4963 node_set(best_node, *used_node_mask);
4970 * Build zonelists ordered by node and zones within node.
4971 * This results in maximum locality--normal zone overflows into local
4972 * DMA zone, if any--but risks exhausting DMA zone.
4974 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
4977 struct zoneref *zonerefs;
4980 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
4982 for (i = 0; i < nr_nodes; i++) {
4985 pg_data_t *node = NODE_DATA(node_order[i]);
4987 nr_zones = build_zonerefs_node(node, zonerefs);
4988 zonerefs += nr_zones;
4990 zonerefs->zone = NULL;
4991 zonerefs->zone_idx = 0;
4995 * Build gfp_thisnode zonelists
4997 static void build_thisnode_zonelists(pg_data_t *pgdat)
4999 struct zoneref *zonerefs;
5002 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5003 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5004 zonerefs += nr_zones;
5005 zonerefs->zone = NULL;
5006 zonerefs->zone_idx = 0;
5010 * Build zonelists ordered by zone and nodes within zones.
5011 * This results in conserving DMA zone[s] until all Normal memory is
5012 * exhausted, but results in overflowing to remote node while memory
5013 * may still exist in local DMA zone.
5016 static void build_zonelists(pg_data_t *pgdat)
5018 static int node_order[MAX_NUMNODES];
5019 int node, nr_nodes = 0;
5020 nodemask_t used_mask = NODE_MASK_NONE;
5021 int local_node, prev_node;
5023 /* NUMA-aware ordering of nodes */
5024 local_node = pgdat->node_id;
5025 prev_node = local_node;
5027 memset(node_order, 0, sizeof(node_order));
5028 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5030 * We don't want to pressure a particular node.
5031 * So adding penalty to the first node in same
5032 * distance group to make it round-robin.
5034 if (node_distance(local_node, node) !=
5035 node_distance(local_node, prev_node))
5036 node_load[node] += 1;
5038 node_order[nr_nodes++] = node;
5042 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5043 build_thisnode_zonelists(pgdat);
5044 pr_info("Fallback order for Node %d: ", local_node);
5045 for (node = 0; node < nr_nodes; node++)
5046 pr_cont("%d ", node_order[node]);
5050 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5052 * Return node id of node used for "local" allocations.
5053 * I.e., first node id of first zone in arg node's generic zonelist.
5054 * Used for initializing percpu 'numa_mem', which is used primarily
5055 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5057 int local_memory_node(int node)
5061 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5062 gfp_zone(GFP_KERNEL),
5064 return zone_to_nid(z->zone);
5068 static void setup_min_unmapped_ratio(void);
5069 static void setup_min_slab_ratio(void);
5070 #else /* CONFIG_NUMA */
5072 static void build_zonelists(pg_data_t *pgdat)
5074 int node, local_node;
5075 struct zoneref *zonerefs;
5078 local_node = pgdat->node_id;
5080 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5081 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5082 zonerefs += nr_zones;
5085 * Now we build the zonelist so that it contains the zones
5086 * of all the other nodes.
5087 * We don't want to pressure a particular node, so when
5088 * building the zones for node N, we make sure that the
5089 * zones coming right after the local ones are those from
5090 * node N+1 (modulo N)
5092 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5093 if (!node_online(node))
5095 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5096 zonerefs += nr_zones;
5098 for (node = 0; node < local_node; node++) {
5099 if (!node_online(node))
5101 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5102 zonerefs += nr_zones;
5105 zonerefs->zone = NULL;
5106 zonerefs->zone_idx = 0;
5109 #endif /* CONFIG_NUMA */
5112 * Boot pageset table. One per cpu which is going to be used for all
5113 * zones and all nodes. The parameters will be set in such a way
5114 * that an item put on a list will immediately be handed over to
5115 * the buddy list. This is safe since pageset manipulation is done
5116 * with interrupts disabled.
5118 * The boot_pagesets must be kept even after bootup is complete for
5119 * unused processors and/or zones. They do play a role for bootstrapping
5120 * hotplugged processors.
5122 * zoneinfo_show() and maybe other functions do
5123 * not check if the processor is online before following the pageset pointer.
5124 * Other parts of the kernel may not check if the zone is available.
5126 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5127 /* These effectively disable the pcplists in the boot pageset completely */
5128 #define BOOT_PAGESET_HIGH 0
5129 #define BOOT_PAGESET_BATCH 1
5130 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5131 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5133 static void __build_all_zonelists(void *data)
5136 int __maybe_unused cpu;
5137 pg_data_t *self = data;
5138 unsigned long flags;
5141 * The zonelist_update_seq must be acquired with irqsave because the
5142 * reader can be invoked from IRQ with GFP_ATOMIC.
5144 write_seqlock_irqsave(&zonelist_update_seq, flags);
5146 * Also disable synchronous printk() to prevent any printk() from
5147 * trying to hold port->lock, for
5148 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5149 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5151 printk_deferred_enter();
5154 memset(node_load, 0, sizeof(node_load));
5158 * This node is hotadded and no memory is yet present. So just
5159 * building zonelists is fine - no need to touch other nodes.
5161 if (self && !node_online(self->node_id)) {
5162 build_zonelists(self);
5165 * All possible nodes have pgdat preallocated
5168 for_each_node(nid) {
5169 pg_data_t *pgdat = NODE_DATA(nid);
5171 build_zonelists(pgdat);
5174 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5176 * We now know the "local memory node" for each node--
5177 * i.e., the node of the first zone in the generic zonelist.
5178 * Set up numa_mem percpu variable for on-line cpus. During
5179 * boot, only the boot cpu should be on-line; we'll init the
5180 * secondary cpus' numa_mem as they come on-line. During
5181 * node/memory hotplug, we'll fixup all on-line cpus.
5183 for_each_online_cpu(cpu)
5184 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5188 printk_deferred_exit();
5189 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5192 static noinline void __init
5193 build_all_zonelists_init(void)
5197 __build_all_zonelists(NULL);
5200 * Initialize the boot_pagesets that are going to be used
5201 * for bootstrapping processors. The real pagesets for
5202 * each zone will be allocated later when the per cpu
5203 * allocator is available.
5205 * boot_pagesets are used also for bootstrapping offline
5206 * cpus if the system is already booted because the pagesets
5207 * are needed to initialize allocators on a specific cpu too.
5208 * F.e. the percpu allocator needs the page allocator which
5209 * needs the percpu allocator in order to allocate its pagesets
5210 * (a chicken-egg dilemma).
5212 for_each_possible_cpu(cpu)
5213 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5215 mminit_verify_zonelist();
5216 cpuset_init_current_mems_allowed();
5220 * unless system_state == SYSTEM_BOOTING.
5222 * __ref due to call of __init annotated helper build_all_zonelists_init
5223 * [protected by SYSTEM_BOOTING].
5225 void __ref build_all_zonelists(pg_data_t *pgdat)
5227 unsigned long vm_total_pages;
5229 if (system_state == SYSTEM_BOOTING) {
5230 build_all_zonelists_init();
5232 __build_all_zonelists(pgdat);
5233 /* cpuset refresh routine should be here */
5235 /* Get the number of free pages beyond high watermark in all zones. */
5236 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5238 * Disable grouping by mobility if the number of pages in the
5239 * system is too low to allow the mechanism to work. It would be
5240 * more accurate, but expensive to check per-zone. This check is
5241 * made on memory-hotadd so a system can start with mobility
5242 * disabled and enable it later
5244 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5245 page_group_by_mobility_disabled = 1;
5247 page_group_by_mobility_disabled = 0;
5249 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5251 page_group_by_mobility_disabled ? "off" : "on",
5254 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5258 static int zone_batchsize(struct zone *zone)
5264 * The number of pages to batch allocate is either ~0.1%
5265 * of the zone or 1MB, whichever is smaller. The batch
5266 * size is striking a balance between allocation latency
5267 * and zone lock contention.
5269 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5270 batch /= 4; /* We effectively *= 4 below */
5275 * Clamp the batch to a 2^n - 1 value. Having a power
5276 * of 2 value was found to be more likely to have
5277 * suboptimal cache aliasing properties in some cases.
5279 * For example if 2 tasks are alternately allocating
5280 * batches of pages, one task can end up with a lot
5281 * of pages of one half of the possible page colors
5282 * and the other with pages of the other colors.
5284 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5289 /* The deferral and batching of frees should be suppressed under NOMMU
5292 * The problem is that NOMMU needs to be able to allocate large chunks
5293 * of contiguous memory as there's no hardware page translation to
5294 * assemble apparent contiguous memory from discontiguous pages.
5296 * Queueing large contiguous runs of pages for batching, however,
5297 * causes the pages to actually be freed in smaller chunks. As there
5298 * can be a significant delay between the individual batches being
5299 * recycled, this leads to the once large chunks of space being
5300 * fragmented and becoming unavailable for high-order allocations.
5306 static int percpu_pagelist_high_fraction;
5307 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
5312 unsigned long total_pages;
5314 if (!percpu_pagelist_high_fraction) {
5316 * By default, the high value of the pcp is based on the zone
5317 * low watermark so that if they are full then background
5318 * reclaim will not be started prematurely.
5320 total_pages = low_wmark_pages(zone);
5323 * If percpu_pagelist_high_fraction is configured, the high
5324 * value is based on a fraction of the managed pages in the
5327 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
5331 * Split the high value across all online CPUs local to the zone. Note
5332 * that early in boot that CPUs may not be online yet and that during
5333 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5334 * onlined. For memory nodes that have no CPUs, split pcp->high across
5335 * all online CPUs to mitigate the risk that reclaim is triggered
5336 * prematurely due to pages stored on pcp lists.
5338 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5340 nr_split_cpus = num_online_cpus();
5341 high = total_pages / nr_split_cpus;
5344 * Ensure high is at least batch*4. The multiple is based on the
5345 * historical relationship between high and batch.
5347 high = max(high, batch << 2);
5356 * pcp->high and pcp->batch values are related and generally batch is lower
5357 * than high. They are also related to pcp->count such that count is lower
5358 * than high, and as soon as it reaches high, the pcplist is flushed.
5360 * However, guaranteeing these relations at all times would require e.g. write
5361 * barriers here but also careful usage of read barriers at the read side, and
5362 * thus be prone to error and bad for performance. Thus the update only prevents
5363 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
5364 * can cope with those fields changing asynchronously, and fully trust only the
5365 * pcp->count field on the local CPU with interrupts disabled.
5367 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5368 * outside of boot time (or some other assurance that no concurrent updaters
5371 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5372 unsigned long batch)
5374 WRITE_ONCE(pcp->batch, batch);
5375 WRITE_ONCE(pcp->high, high);
5378 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5382 memset(pcp, 0, sizeof(*pcp));
5383 memset(pzstats, 0, sizeof(*pzstats));
5385 spin_lock_init(&pcp->lock);
5386 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5387 INIT_LIST_HEAD(&pcp->lists[pindex]);
5390 * Set batch and high values safe for a boot pageset. A true percpu
5391 * pageset's initialization will update them subsequently. Here we don't
5392 * need to be as careful as pageset_update() as nobody can access the
5395 pcp->high = BOOT_PAGESET_HIGH;
5396 pcp->batch = BOOT_PAGESET_BATCH;
5397 pcp->free_factor = 0;
5400 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
5401 unsigned long batch)
5403 struct per_cpu_pages *pcp;
5406 for_each_possible_cpu(cpu) {
5407 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5408 pageset_update(pcp, high, batch);
5413 * Calculate and set new high and batch values for all per-cpu pagesets of a
5414 * zone based on the zone's size.
5416 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5418 int new_high, new_batch;
5420 new_batch = max(1, zone_batchsize(zone));
5421 new_high = zone_highsize(zone, new_batch, cpu_online);
5423 if (zone->pageset_high == new_high &&
5424 zone->pageset_batch == new_batch)
5427 zone->pageset_high = new_high;
5428 zone->pageset_batch = new_batch;
5430 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
5433 void __meminit setup_zone_pageset(struct zone *zone)
5437 /* Size may be 0 on !SMP && !NUMA */
5438 if (sizeof(struct per_cpu_zonestat) > 0)
5439 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5441 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5442 for_each_possible_cpu(cpu) {
5443 struct per_cpu_pages *pcp;
5444 struct per_cpu_zonestat *pzstats;
5446 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5447 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5448 per_cpu_pages_init(pcp, pzstats);
5451 zone_set_pageset_high_and_batch(zone, 0);
5455 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5456 * page high values need to be recalculated.
5458 static void zone_pcp_update(struct zone *zone, int cpu_online)
5460 mutex_lock(&pcp_batch_high_lock);
5461 zone_set_pageset_high_and_batch(zone, cpu_online);
5462 mutex_unlock(&pcp_batch_high_lock);
5466 * Allocate per cpu pagesets and initialize them.
5467 * Before this call only boot pagesets were available.
5469 void __init setup_per_cpu_pageset(void)
5471 struct pglist_data *pgdat;
5473 int __maybe_unused cpu;
5475 for_each_populated_zone(zone)
5476 setup_zone_pageset(zone);
5480 * Unpopulated zones continue using the boot pagesets.
5481 * The numa stats for these pagesets need to be reset.
5482 * Otherwise, they will end up skewing the stats of
5483 * the nodes these zones are associated with.
5485 for_each_possible_cpu(cpu) {
5486 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5487 memset(pzstats->vm_numa_event, 0,
5488 sizeof(pzstats->vm_numa_event));
5492 for_each_online_pgdat(pgdat)
5493 pgdat->per_cpu_nodestats =
5494 alloc_percpu(struct per_cpu_nodestat);
5497 __meminit void zone_pcp_init(struct zone *zone)
5500 * per cpu subsystem is not up at this point. The following code
5501 * relies on the ability of the linker to provide the
5502 * offset of a (static) per cpu variable into the per cpu area.
5504 zone->per_cpu_pageset = &boot_pageset;
5505 zone->per_cpu_zonestats = &boot_zonestats;
5506 zone->pageset_high = BOOT_PAGESET_HIGH;
5507 zone->pageset_batch = BOOT_PAGESET_BATCH;
5509 if (populated_zone(zone))
5510 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5511 zone->present_pages, zone_batchsize(zone));
5514 void adjust_managed_page_count(struct page *page, long count)
5516 atomic_long_add(count, &page_zone(page)->managed_pages);
5517 totalram_pages_add(count);
5518 #ifdef CONFIG_HIGHMEM
5519 if (PageHighMem(page))
5520 totalhigh_pages_add(count);
5523 EXPORT_SYMBOL(adjust_managed_page_count);
5525 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5528 unsigned long pages = 0;
5530 start = (void *)PAGE_ALIGN((unsigned long)start);
5531 end = (void *)((unsigned long)end & PAGE_MASK);
5532 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5533 struct page *page = virt_to_page(pos);
5534 void *direct_map_addr;
5537 * 'direct_map_addr' might be different from 'pos'
5538 * because some architectures' virt_to_page()
5539 * work with aliases. Getting the direct map
5540 * address ensures that we get a _writeable_
5541 * alias for the memset().
5543 direct_map_addr = page_address(page);
5545 * Perform a kasan-unchecked memset() since this memory
5546 * has not been initialized.
5548 direct_map_addr = kasan_reset_tag(direct_map_addr);
5549 if ((unsigned int)poison <= 0xFF)
5550 memset(direct_map_addr, poison, PAGE_SIZE);
5552 free_reserved_page(page);
5556 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5561 static int page_alloc_cpu_dead(unsigned int cpu)
5565 lru_add_drain_cpu(cpu);
5566 mlock_drain_remote(cpu);
5570 * Spill the event counters of the dead processor
5571 * into the current processors event counters.
5572 * This artificially elevates the count of the current
5575 vm_events_fold_cpu(cpu);
5578 * Zero the differential counters of the dead processor
5579 * so that the vm statistics are consistent.
5581 * This is only okay since the processor is dead and cannot
5582 * race with what we are doing.
5584 cpu_vm_stats_fold(cpu);
5586 for_each_populated_zone(zone)
5587 zone_pcp_update(zone, 0);
5592 static int page_alloc_cpu_online(unsigned int cpu)
5596 for_each_populated_zone(zone)
5597 zone_pcp_update(zone, 1);
5601 void __init page_alloc_init_cpuhp(void)
5605 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5606 "mm/page_alloc:pcp",
5607 page_alloc_cpu_online,
5608 page_alloc_cpu_dead);
5613 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5614 * or min_free_kbytes changes.
5616 static void calculate_totalreserve_pages(void)
5618 struct pglist_data *pgdat;
5619 unsigned long reserve_pages = 0;
5620 enum zone_type i, j;
5622 for_each_online_pgdat(pgdat) {
5624 pgdat->totalreserve_pages = 0;
5626 for (i = 0; i < MAX_NR_ZONES; i++) {
5627 struct zone *zone = pgdat->node_zones + i;
5629 unsigned long managed_pages = zone_managed_pages(zone);
5631 /* Find valid and maximum lowmem_reserve in the zone */
5632 for (j = i; j < MAX_NR_ZONES; j++) {
5633 if (zone->lowmem_reserve[j] > max)
5634 max = zone->lowmem_reserve[j];
5637 /* we treat the high watermark as reserved pages. */
5638 max += high_wmark_pages(zone);
5640 if (max > managed_pages)
5641 max = managed_pages;
5643 pgdat->totalreserve_pages += max;
5645 reserve_pages += max;
5648 totalreserve_pages = reserve_pages;
5652 * setup_per_zone_lowmem_reserve - called whenever
5653 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5654 * has a correct pages reserved value, so an adequate number of
5655 * pages are left in the zone after a successful __alloc_pages().
5657 static void setup_per_zone_lowmem_reserve(void)
5659 struct pglist_data *pgdat;
5660 enum zone_type i, j;
5662 for_each_online_pgdat(pgdat) {
5663 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5664 struct zone *zone = &pgdat->node_zones[i];
5665 int ratio = sysctl_lowmem_reserve_ratio[i];
5666 bool clear = !ratio || !zone_managed_pages(zone);
5667 unsigned long managed_pages = 0;
5669 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5670 struct zone *upper_zone = &pgdat->node_zones[j];
5672 managed_pages += zone_managed_pages(upper_zone);
5675 zone->lowmem_reserve[j] = 0;
5677 zone->lowmem_reserve[j] = managed_pages / ratio;
5682 /* update totalreserve_pages */
5683 calculate_totalreserve_pages();
5686 static void __setup_per_zone_wmarks(void)
5688 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5689 unsigned long lowmem_pages = 0;
5691 unsigned long flags;
5693 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5694 for_each_zone(zone) {
5695 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5696 lowmem_pages += zone_managed_pages(zone);
5699 for_each_zone(zone) {
5702 spin_lock_irqsave(&zone->lock, flags);
5703 tmp = (u64)pages_min * zone_managed_pages(zone);
5704 do_div(tmp, lowmem_pages);
5705 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5707 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5708 * need highmem and movable zones pages, so cap pages_min
5709 * to a small value here.
5711 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5712 * deltas control async page reclaim, and so should
5713 * not be capped for highmem and movable zones.
5715 unsigned long min_pages;
5717 min_pages = zone_managed_pages(zone) / 1024;
5718 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5719 zone->_watermark[WMARK_MIN] = min_pages;
5722 * If it's a lowmem zone, reserve a number of pages
5723 * proportionate to the zone's size.
5725 zone->_watermark[WMARK_MIN] = tmp;
5729 * Set the kswapd watermarks distance according to the
5730 * scale factor in proportion to available memory, but
5731 * ensure a minimum size on small systems.
5733 tmp = max_t(u64, tmp >> 2,
5734 mult_frac(zone_managed_pages(zone),
5735 watermark_scale_factor, 10000));
5737 zone->watermark_boost = 0;
5738 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
5739 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5740 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5742 spin_unlock_irqrestore(&zone->lock, flags);
5745 /* update totalreserve_pages */
5746 calculate_totalreserve_pages();
5750 * setup_per_zone_wmarks - called when min_free_kbytes changes
5751 * or when memory is hot-{added|removed}
5753 * Ensures that the watermark[min,low,high] values for each zone are set
5754 * correctly with respect to min_free_kbytes.
5756 void setup_per_zone_wmarks(void)
5759 static DEFINE_SPINLOCK(lock);
5762 __setup_per_zone_wmarks();
5766 * The watermark size have changed so update the pcpu batch
5767 * and high limits or the limits may be inappropriate.
5770 zone_pcp_update(zone, 0);
5774 * Initialise min_free_kbytes.
5776 * For small machines we want it small (128k min). For large machines
5777 * we want it large (256MB max). But it is not linear, because network
5778 * bandwidth does not increase linearly with machine size. We use
5780 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5781 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5797 void calculate_min_free_kbytes(void)
5799 unsigned long lowmem_kbytes;
5800 int new_min_free_kbytes;
5802 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5803 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5805 if (new_min_free_kbytes > user_min_free_kbytes)
5806 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
5808 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5809 new_min_free_kbytes, user_min_free_kbytes);
5813 int __meminit init_per_zone_wmark_min(void)
5815 calculate_min_free_kbytes();
5816 setup_per_zone_wmarks();
5817 refresh_zone_stat_thresholds();
5818 setup_per_zone_lowmem_reserve();
5821 setup_min_unmapped_ratio();
5822 setup_min_slab_ratio();
5825 khugepaged_min_free_kbytes_update();
5829 postcore_initcall(init_per_zone_wmark_min)
5832 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5833 * that we can call two helper functions whenever min_free_kbytes
5836 static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5837 void *buffer, size_t *length, loff_t *ppos)
5841 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5846 user_min_free_kbytes = min_free_kbytes;
5847 setup_per_zone_wmarks();
5852 static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
5853 void *buffer, size_t *length, loff_t *ppos)
5857 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5862 setup_per_zone_wmarks();
5868 static void setup_min_unmapped_ratio(void)
5873 for_each_online_pgdat(pgdat)
5874 pgdat->min_unmapped_pages = 0;
5877 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
5878 sysctl_min_unmapped_ratio) / 100;
5882 static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5883 void *buffer, size_t *length, loff_t *ppos)
5887 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5891 setup_min_unmapped_ratio();
5896 static void setup_min_slab_ratio(void)
5901 for_each_online_pgdat(pgdat)
5902 pgdat->min_slab_pages = 0;
5905 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
5906 sysctl_min_slab_ratio) / 100;
5909 static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5910 void *buffer, size_t *length, loff_t *ppos)
5914 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5918 setup_min_slab_ratio();
5925 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5926 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5927 * whenever sysctl_lowmem_reserve_ratio changes.
5929 * The reserve ratio obviously has absolutely no relation with the
5930 * minimum watermarks. The lowmem reserve ratio can only make sense
5931 * if in function of the boot time zone sizes.
5933 static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table,
5934 int write, void *buffer, size_t *length, loff_t *ppos)
5938 proc_dointvec_minmax(table, write, buffer, length, ppos);
5940 for (i = 0; i < MAX_NR_ZONES; i++) {
5941 if (sysctl_lowmem_reserve_ratio[i] < 1)
5942 sysctl_lowmem_reserve_ratio[i] = 0;
5945 setup_per_zone_lowmem_reserve();
5950 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
5951 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5952 * pagelist can have before it gets flushed back to buddy allocator.
5954 static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
5955 int write, void *buffer, size_t *length, loff_t *ppos)
5958 int old_percpu_pagelist_high_fraction;
5961 mutex_lock(&pcp_batch_high_lock);
5962 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
5964 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5965 if (!write || ret < 0)
5968 /* Sanity checking to avoid pcp imbalance */
5969 if (percpu_pagelist_high_fraction &&
5970 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
5971 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
5977 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
5980 for_each_populated_zone(zone)
5981 zone_set_pageset_high_and_batch(zone, 0);
5983 mutex_unlock(&pcp_batch_high_lock);
5987 static struct ctl_table page_alloc_sysctl_table[] = {
5989 .procname = "min_free_kbytes",
5990 .data = &min_free_kbytes,
5991 .maxlen = sizeof(min_free_kbytes),
5993 .proc_handler = min_free_kbytes_sysctl_handler,
5994 .extra1 = SYSCTL_ZERO,
5997 .procname = "watermark_boost_factor",
5998 .data = &watermark_boost_factor,
5999 .maxlen = sizeof(watermark_boost_factor),
6001 .proc_handler = proc_dointvec_minmax,
6002 .extra1 = SYSCTL_ZERO,
6005 .procname = "watermark_scale_factor",
6006 .data = &watermark_scale_factor,
6007 .maxlen = sizeof(watermark_scale_factor),
6009 .proc_handler = watermark_scale_factor_sysctl_handler,
6010 .extra1 = SYSCTL_ONE,
6011 .extra2 = SYSCTL_THREE_THOUSAND,
6014 .procname = "percpu_pagelist_high_fraction",
6015 .data = &percpu_pagelist_high_fraction,
6016 .maxlen = sizeof(percpu_pagelist_high_fraction),
6018 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6019 .extra1 = SYSCTL_ZERO,
6022 .procname = "lowmem_reserve_ratio",
6023 .data = &sysctl_lowmem_reserve_ratio,
6024 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6026 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6030 .procname = "numa_zonelist_order",
6031 .data = &numa_zonelist_order,
6032 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6034 .proc_handler = numa_zonelist_order_handler,
6037 .procname = "min_unmapped_ratio",
6038 .data = &sysctl_min_unmapped_ratio,
6039 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6041 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6042 .extra1 = SYSCTL_ZERO,
6043 .extra2 = SYSCTL_ONE_HUNDRED,
6046 .procname = "min_slab_ratio",
6047 .data = &sysctl_min_slab_ratio,
6048 .maxlen = sizeof(sysctl_min_slab_ratio),
6050 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6051 .extra1 = SYSCTL_ZERO,
6052 .extra2 = SYSCTL_ONE_HUNDRED,
6058 void __init page_alloc_sysctl_init(void)
6060 register_sysctl_init("vm", page_alloc_sysctl_table);
6063 #ifdef CONFIG_CONTIG_ALLOC
6064 /* Usage: See admin-guide/dynamic-debug-howto.rst */
6065 static void alloc_contig_dump_pages(struct list_head *page_list)
6067 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6069 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6073 list_for_each_entry(page, page_list, lru)
6074 dump_page(page, "migration failure");
6078 /* [start, end) must belong to a single zone. */
6079 int __alloc_contig_migrate_range(struct compact_control *cc,
6080 unsigned long start, unsigned long end)
6082 /* This function is based on compact_zone() from compaction.c. */
6083 unsigned int nr_reclaimed;
6084 unsigned long pfn = start;
6085 unsigned int tries = 0;
6087 struct migration_target_control mtc = {
6088 .nid = zone_to_nid(cc->zone),
6089 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6092 lru_cache_disable();
6094 while (pfn < end || !list_empty(&cc->migratepages)) {
6095 if (fatal_signal_pending(current)) {
6100 if (list_empty(&cc->migratepages)) {
6101 cc->nr_migratepages = 0;
6102 ret = isolate_migratepages_range(cc, pfn, end);
6103 if (ret && ret != -EAGAIN)
6105 pfn = cc->migrate_pfn;
6107 } else if (++tries == 5) {
6112 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6114 cc->nr_migratepages -= nr_reclaimed;
6116 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6117 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6120 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6121 * to retry again over this error, so do the same here.
6129 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6130 alloc_contig_dump_pages(&cc->migratepages);
6131 putback_movable_pages(&cc->migratepages);
6138 * alloc_contig_range() -- tries to allocate given range of pages
6139 * @start: start PFN to allocate
6140 * @end: one-past-the-last PFN to allocate
6141 * @migratetype: migratetype of the underlying pageblocks (either
6142 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6143 * in range must have the same migratetype and it must
6144 * be either of the two.
6145 * @gfp_mask: GFP mask to use during compaction
6147 * The PFN range does not have to be pageblock aligned. The PFN range must
6148 * belong to a single zone.
6150 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6151 * pageblocks in the range. Once isolated, the pageblocks should not
6152 * be modified by others.
6154 * Return: zero on success or negative error code. On success all
6155 * pages which PFN is in [start, end) are allocated for the caller and
6156 * need to be freed with free_contig_range().
6158 int alloc_contig_range(unsigned long start, unsigned long end,
6159 unsigned migratetype, gfp_t gfp_mask)
6161 unsigned long outer_start, outer_end;
6165 struct compact_control cc = {
6166 .nr_migratepages = 0,
6168 .zone = page_zone(pfn_to_page(start)),
6169 .mode = MIGRATE_SYNC,
6170 .ignore_skip_hint = true,
6171 .no_set_skip_hint = true,
6172 .gfp_mask = current_gfp_context(gfp_mask),
6173 .alloc_contig = true,
6175 INIT_LIST_HEAD(&cc.migratepages);
6178 * What we do here is we mark all pageblocks in range as
6179 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6180 * have different sizes, and due to the way page allocator
6181 * work, start_isolate_page_range() has special handlings for this.
6183 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6184 * migrate the pages from an unaligned range (ie. pages that
6185 * we are interested in). This will put all the pages in
6186 * range back to page allocator as MIGRATE_ISOLATE.
6188 * When this is done, we take the pages in range from page
6189 * allocator removing them from the buddy system. This way
6190 * page allocator will never consider using them.
6192 * This lets us mark the pageblocks back as
6193 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6194 * aligned range but not in the unaligned, original range are
6195 * put back to page allocator so that buddy can use them.
6198 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6202 drain_all_pages(cc.zone);
6205 * In case of -EBUSY, we'd like to know which page causes problem.
6206 * So, just fall through. test_pages_isolated() has a tracepoint
6207 * which will report the busy page.
6209 * It is possible that busy pages could become available before
6210 * the call to test_pages_isolated, and the range will actually be
6211 * allocated. So, if we fall through be sure to clear ret so that
6212 * -EBUSY is not accidentally used or returned to caller.
6214 ret = __alloc_contig_migrate_range(&cc, start, end);
6215 if (ret && ret != -EBUSY)
6220 * Pages from [start, end) are within a pageblock_nr_pages
6221 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6222 * more, all pages in [start, end) are free in page allocator.
6223 * What we are going to do is to allocate all pages from
6224 * [start, end) (that is remove them from page allocator).
6226 * The only problem is that pages at the beginning and at the
6227 * end of interesting range may be not aligned with pages that
6228 * page allocator holds, ie. they can be part of higher order
6229 * pages. Because of this, we reserve the bigger range and
6230 * once this is done free the pages we are not interested in.
6232 * We don't have to hold zone->lock here because the pages are
6233 * isolated thus they won't get removed from buddy.
6237 outer_start = start;
6238 while (!PageBuddy(pfn_to_page(outer_start))) {
6239 if (++order > MAX_ORDER) {
6240 outer_start = start;
6243 outer_start &= ~0UL << order;
6246 if (outer_start != start) {
6247 order = buddy_order(pfn_to_page(outer_start));
6250 * outer_start page could be small order buddy page and
6251 * it doesn't include start page. Adjust outer_start
6252 * in this case to report failed page properly
6253 * on tracepoint in test_pages_isolated()
6255 if (outer_start + (1UL << order) <= start)
6256 outer_start = start;
6259 /* Make sure the range is really isolated. */
6260 if (test_pages_isolated(outer_start, end, 0)) {
6265 /* Grab isolated pages from freelists. */
6266 outer_end = isolate_freepages_range(&cc, outer_start, end);
6272 /* Free head and tail (if any) */
6273 if (start != outer_start)
6274 free_contig_range(outer_start, start - outer_start);
6275 if (end != outer_end)
6276 free_contig_range(end, outer_end - end);
6279 undo_isolate_page_range(start, end, migratetype);
6282 EXPORT_SYMBOL(alloc_contig_range);
6284 static int __alloc_contig_pages(unsigned long start_pfn,
6285 unsigned long nr_pages, gfp_t gfp_mask)
6287 unsigned long end_pfn = start_pfn + nr_pages;
6289 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
6293 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6294 unsigned long nr_pages)
6296 unsigned long i, end_pfn = start_pfn + nr_pages;
6299 for (i = start_pfn; i < end_pfn; i++) {
6300 page = pfn_to_online_page(i);
6304 if (page_zone(page) != z)
6307 if (PageReserved(page))
6316 static bool zone_spans_last_pfn(const struct zone *zone,
6317 unsigned long start_pfn, unsigned long nr_pages)
6319 unsigned long last_pfn = start_pfn + nr_pages - 1;
6321 return zone_spans_pfn(zone, last_pfn);
6325 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6326 * @nr_pages: Number of contiguous pages to allocate
6327 * @gfp_mask: GFP mask to limit search and used during compaction
6329 * @nodemask: Mask for other possible nodes
6331 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6332 * on an applicable zonelist to find a contiguous pfn range which can then be
6333 * tried for allocation with alloc_contig_range(). This routine is intended
6334 * for allocation requests which can not be fulfilled with the buddy allocator.
6336 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6337 * power of two, then allocated range is also guaranteed to be aligned to same
6338 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6340 * Allocated pages can be freed with free_contig_range() or by manually calling
6341 * __free_page() on each allocated page.
6343 * Return: pointer to contiguous pages on success, or NULL if not successful.
6345 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
6346 int nid, nodemask_t *nodemask)
6348 unsigned long ret, pfn, flags;
6349 struct zonelist *zonelist;
6353 zonelist = node_zonelist(nid, gfp_mask);
6354 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6355 gfp_zone(gfp_mask), nodemask) {
6356 spin_lock_irqsave(&zone->lock, flags);
6358 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6359 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6360 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6362 * We release the zone lock here because
6363 * alloc_contig_range() will also lock the zone
6364 * at some point. If there's an allocation
6365 * spinning on this lock, it may win the race
6366 * and cause alloc_contig_range() to fail...
6368 spin_unlock_irqrestore(&zone->lock, flags);
6369 ret = __alloc_contig_pages(pfn, nr_pages,
6372 return pfn_to_page(pfn);
6373 spin_lock_irqsave(&zone->lock, flags);
6377 spin_unlock_irqrestore(&zone->lock, flags);
6381 #endif /* CONFIG_CONTIG_ALLOC */
6383 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6385 unsigned long count = 0;
6387 for (; nr_pages--; pfn++) {
6388 struct page *page = pfn_to_page(pfn);
6390 count += page_count(page) != 1;
6393 WARN(count != 0, "%lu pages are still in use!\n", count);
6395 EXPORT_SYMBOL(free_contig_range);
6398 * Effectively disable pcplists for the zone by setting the high limit to 0
6399 * and draining all cpus. A concurrent page freeing on another CPU that's about
6400 * to put the page on pcplist will either finish before the drain and the page
6401 * will be drained, or observe the new high limit and skip the pcplist.
6403 * Must be paired with a call to zone_pcp_enable().
6405 void zone_pcp_disable(struct zone *zone)
6407 mutex_lock(&pcp_batch_high_lock);
6408 __zone_set_pageset_high_and_batch(zone, 0, 1);
6409 __drain_all_pages(zone, true);
6412 void zone_pcp_enable(struct zone *zone)
6414 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
6415 mutex_unlock(&pcp_batch_high_lock);
6418 void zone_pcp_reset(struct zone *zone)
6421 struct per_cpu_zonestat *pzstats;
6423 if (zone->per_cpu_pageset != &boot_pageset) {
6424 for_each_online_cpu(cpu) {
6425 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6426 drain_zonestat(zone, pzstats);
6428 free_percpu(zone->per_cpu_pageset);
6429 zone->per_cpu_pageset = &boot_pageset;
6430 if (zone->per_cpu_zonestats != &boot_zonestats) {
6431 free_percpu(zone->per_cpu_zonestats);
6432 zone->per_cpu_zonestats = &boot_zonestats;
6437 #ifdef CONFIG_MEMORY_HOTREMOVE
6439 * All pages in the range must be in a single zone, must not contain holes,
6440 * must span full sections, and must be isolated before calling this function.
6442 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6444 unsigned long pfn = start_pfn;
6448 unsigned long flags;
6450 offline_mem_sections(pfn, end_pfn);
6451 zone = page_zone(pfn_to_page(pfn));
6452 spin_lock_irqsave(&zone->lock, flags);
6453 while (pfn < end_pfn) {
6454 page = pfn_to_page(pfn);
6456 * The HWPoisoned page may be not in buddy system, and
6457 * page_count() is not 0.
6459 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6464 * At this point all remaining PageOffline() pages have a
6465 * reference count of 0 and can simply be skipped.
6467 if (PageOffline(page)) {
6468 BUG_ON(page_count(page));
6469 BUG_ON(PageBuddy(page));
6474 BUG_ON(page_count(page));
6475 BUG_ON(!PageBuddy(page));
6476 order = buddy_order(page);
6477 del_page_from_free_list(page, zone, order);
6478 pfn += (1 << order);
6480 spin_unlock_irqrestore(&zone->lock, flags);
6485 * This function returns a stable result only if called under zone lock.
6487 bool is_free_buddy_page(struct page *page)
6489 unsigned long pfn = page_to_pfn(page);
6492 for (order = 0; order <= MAX_ORDER; order++) {
6493 struct page *page_head = page - (pfn & ((1 << order) - 1));
6495 if (PageBuddy(page_head) &&
6496 buddy_order_unsafe(page_head) >= order)
6500 return order <= MAX_ORDER;
6502 EXPORT_SYMBOL(is_free_buddy_page);
6504 #ifdef CONFIG_MEMORY_FAILURE
6506 * Break down a higher-order page in sub-pages, and keep our target out of
6509 static void break_down_buddy_pages(struct zone *zone, struct page *page,
6510 struct page *target, int low, int high,
6513 unsigned long size = 1 << high;
6514 struct page *current_buddy, *next_page;
6516 while (high > low) {
6520 if (target >= &page[size]) {
6521 next_page = page + size;
6522 current_buddy = page;
6525 current_buddy = page + size;
6528 if (set_page_guard(zone, current_buddy, high, migratetype))
6531 if (current_buddy != target) {
6532 add_to_free_list(current_buddy, zone, high, migratetype);
6533 set_buddy_order(current_buddy, high);
6540 * Take a page that will be marked as poisoned off the buddy allocator.
6542 bool take_page_off_buddy(struct page *page)
6544 struct zone *zone = page_zone(page);
6545 unsigned long pfn = page_to_pfn(page);
6546 unsigned long flags;
6550 spin_lock_irqsave(&zone->lock, flags);
6551 for (order = 0; order <= MAX_ORDER; order++) {
6552 struct page *page_head = page - (pfn & ((1 << order) - 1));
6553 int page_order = buddy_order(page_head);
6555 if (PageBuddy(page_head) && page_order >= order) {
6556 unsigned long pfn_head = page_to_pfn(page_head);
6557 int migratetype = get_pfnblock_migratetype(page_head,
6560 del_page_from_free_list(page_head, zone, page_order);
6561 break_down_buddy_pages(zone, page_head, page, 0,
6562 page_order, migratetype);
6563 SetPageHWPoisonTakenOff(page);
6564 if (!is_migrate_isolate(migratetype))
6565 __mod_zone_freepage_state(zone, -1, migratetype);
6569 if (page_count(page_head) > 0)
6572 spin_unlock_irqrestore(&zone->lock, flags);
6577 * Cancel takeoff done by take_page_off_buddy().
6579 bool put_page_back_buddy(struct page *page)
6581 struct zone *zone = page_zone(page);
6582 unsigned long pfn = page_to_pfn(page);
6583 unsigned long flags;
6584 int migratetype = get_pfnblock_migratetype(page, pfn);
6587 spin_lock_irqsave(&zone->lock, flags);
6588 if (put_page_testzero(page)) {
6589 ClearPageHWPoisonTakenOff(page);
6590 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6591 if (TestClearPageHWPoison(page)) {
6595 spin_unlock_irqrestore(&zone->lock, flags);
6601 #ifdef CONFIG_ZONE_DMA
6602 bool has_managed_dma(void)
6604 struct pglist_data *pgdat;
6606 for_each_online_pgdat(pgdat) {
6607 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6609 if (managed_zone(zone))
6614 #endif /* CONFIG_ZONE_DMA */
6616 #ifdef CONFIG_UNACCEPTED_MEMORY
6618 /* Counts number of zones with unaccepted pages. */
6619 static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6621 static bool lazy_accept = true;
6623 static int __init accept_memory_parse(char *p)
6625 if (!strcmp(p, "lazy")) {
6628 } else if (!strcmp(p, "eager")) {
6629 lazy_accept = false;
6635 early_param("accept_memory", accept_memory_parse);
6637 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6639 phys_addr_t start = page_to_phys(page);
6640 phys_addr_t end = start + (PAGE_SIZE << order);
6642 return range_contains_unaccepted_memory(start, end);
6645 static void accept_page(struct page *page, unsigned int order)
6647 phys_addr_t start = page_to_phys(page);
6649 accept_memory(start, start + (PAGE_SIZE << order));
6652 static bool try_to_accept_memory_one(struct zone *zone)
6654 unsigned long flags;
6658 if (list_empty(&zone->unaccepted_pages))
6661 spin_lock_irqsave(&zone->lock, flags);
6662 page = list_first_entry_or_null(&zone->unaccepted_pages,
6665 spin_unlock_irqrestore(&zone->lock, flags);
6669 list_del(&page->lru);
6670 last = list_empty(&zone->unaccepted_pages);
6672 __mod_zone_freepage_state(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6673 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6674 spin_unlock_irqrestore(&zone->lock, flags);
6676 accept_page(page, MAX_ORDER);
6678 __free_pages_ok(page, MAX_ORDER, FPI_TO_TAIL);
6681 static_branch_dec(&zones_with_unaccepted_pages);
6686 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6691 /* How much to accept to get to high watermark? */
6692 to_accept = high_wmark_pages(zone) -
6693 (zone_page_state(zone, NR_FREE_PAGES) -
6694 __zone_watermark_unusable_free(zone, order, 0));
6696 /* Accept at least one page */
6698 if (!try_to_accept_memory_one(zone))
6701 to_accept -= MAX_ORDER_NR_PAGES;
6702 } while (to_accept > 0);
6707 static inline bool has_unaccepted_memory(void)
6709 return static_branch_unlikely(&zones_with_unaccepted_pages);
6712 static bool __free_unaccepted(struct page *page)
6714 struct zone *zone = page_zone(page);
6715 unsigned long flags;
6721 spin_lock_irqsave(&zone->lock, flags);
6722 first = list_empty(&zone->unaccepted_pages);
6723 list_add_tail(&page->lru, &zone->unaccepted_pages);
6724 __mod_zone_freepage_state(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6725 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
6726 spin_unlock_irqrestore(&zone->lock, flags);
6729 static_branch_inc(&zones_with_unaccepted_pages);
6736 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6741 static void accept_page(struct page *page, unsigned int order)
6745 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6750 static inline bool has_unaccepted_memory(void)
6755 static bool __free_unaccepted(struct page *page)
6761 #endif /* CONFIG_UNACCEPTED_MEMORY */