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/pagevec.h>
36 #include <linux/memory_hotplug.h>
37 #include <linux/nodemask.h>
38 #include <linux/vmstat.h>
39 #include <linux/fault-inject.h>
40 #include <linux/compaction.h>
41 #include <trace/events/kmem.h>
42 #include <trace/events/oom.h>
43 #include <linux/prefetch.h>
44 #include <linux/mm_inline.h>
45 #include <linux/mmu_notifier.h>
46 #include <linux/migrate.h>
47 #include <linux/sched/mm.h>
48 #include <linux/page_owner.h>
49 #include <linux/page_table_check.h>
50 #include <linux/memcontrol.h>
51 #include <linux/ftrace.h>
52 #include <linux/lockdep.h>
53 #include <linux/psi.h>
54 #include <linux/khugepaged.h>
55 #include <linux/delayacct.h>
56 #include <linux/cacheinfo.h>
57 #include <linux/pgalloc_tag.h>
58 #include <asm/div64.h>
61 #include "page_reporting.h"
63 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
64 typedef int __bitwise fpi_t;
66 /* No special request */
67 #define FPI_NONE ((__force fpi_t)0)
70 * Skip free page reporting notification for the (possibly merged) page.
71 * This does not hinder free page reporting from grabbing the page,
72 * reporting it and marking it "reported" - it only skips notifying
73 * the free page reporting infrastructure about a newly freed page. For
74 * example, used when temporarily pulling a page from a freelist and
75 * putting it back unmodified.
77 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
80 * Place the (possibly merged) page to the tail of the freelist. Will ignore
81 * page shuffling (relevant code - e.g., memory onlining - is expected to
82 * shuffle the whole zone).
84 * Note: No code should rely on this flag for correctness - it's purely
85 * to allow for optimizations when handing back either fresh pages
86 * (memory onlining) or untouched pages (page isolation, free page
89 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
91 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
92 static DEFINE_MUTEX(pcp_batch_high_lock);
93 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
95 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
97 * On SMP, spin_trylock is sufficient protection.
98 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
100 #define pcp_trylock_prepare(flags) do { } while (0)
101 #define pcp_trylock_finish(flag) do { } while (0)
104 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
105 #define pcp_trylock_prepare(flags) local_irq_save(flags)
106 #define pcp_trylock_finish(flags) local_irq_restore(flags)
110 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
111 * a migration causing the wrong PCP to be locked and remote memory being
112 * potentially allocated, pin the task to the CPU for the lookup+lock.
113 * preempt_disable is used on !RT because it is faster than migrate_disable.
114 * migrate_disable is used on RT because otherwise RT spinlock usage is
115 * interfered with and a high priority task cannot preempt the allocator.
117 #ifndef CONFIG_PREEMPT_RT
118 #define pcpu_task_pin() preempt_disable()
119 #define pcpu_task_unpin() preempt_enable()
121 #define pcpu_task_pin() migrate_disable()
122 #define pcpu_task_unpin() migrate_enable()
126 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
127 * Return value should be used with equivalent unlock helper.
129 #define pcpu_spin_lock(type, member, ptr) \
133 _ret = this_cpu_ptr(ptr); \
134 spin_lock(&_ret->member); \
138 #define pcpu_spin_trylock(type, member, ptr) \
142 _ret = this_cpu_ptr(ptr); \
143 if (!spin_trylock(&_ret->member)) { \
150 #define pcpu_spin_unlock(member, ptr) \
152 spin_unlock(&ptr->member); \
156 /* struct per_cpu_pages specific helpers. */
157 #define pcp_spin_lock(ptr) \
158 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
160 #define pcp_spin_trylock(ptr) \
161 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
163 #define pcp_spin_unlock(ptr) \
164 pcpu_spin_unlock(lock, ptr)
166 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
167 DEFINE_PER_CPU(int, numa_node);
168 EXPORT_PER_CPU_SYMBOL(numa_node);
171 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
173 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
175 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
176 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
177 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
178 * defined in <linux/topology.h>.
180 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
181 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
184 static DEFINE_MUTEX(pcpu_drain_mutex);
186 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
187 volatile unsigned long latent_entropy __latent_entropy;
188 EXPORT_SYMBOL(latent_entropy);
192 * Array of node states.
194 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
195 [N_POSSIBLE] = NODE_MASK_ALL,
196 [N_ONLINE] = { { [0] = 1UL } },
198 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
199 #ifdef CONFIG_HIGHMEM
200 [N_HIGH_MEMORY] = { { [0] = 1UL } },
202 [N_MEMORY] = { { [0] = 1UL } },
203 [N_CPU] = { { [0] = 1UL } },
206 EXPORT_SYMBOL(node_states);
208 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
210 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
211 unsigned int pageblock_order __read_mostly;
214 static void __free_pages_ok(struct page *page, unsigned int order,
218 * results with 256, 32 in the lowmem_reserve sysctl:
219 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
220 * 1G machine -> (16M dma, 784M normal, 224M high)
221 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
222 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
223 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
225 * TBD: should special case ZONE_DMA32 machines here - in those we normally
226 * don't need any ZONE_NORMAL reservation
228 static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
229 #ifdef CONFIG_ZONE_DMA
232 #ifdef CONFIG_ZONE_DMA32
236 #ifdef CONFIG_HIGHMEM
242 char * const zone_names[MAX_NR_ZONES] = {
243 #ifdef CONFIG_ZONE_DMA
246 #ifdef CONFIG_ZONE_DMA32
250 #ifdef CONFIG_HIGHMEM
254 #ifdef CONFIG_ZONE_DEVICE
259 const char * const migratetype_names[MIGRATE_TYPES] = {
267 #ifdef CONFIG_MEMORY_ISOLATION
272 int min_free_kbytes = 1024;
273 int user_min_free_kbytes = -1;
274 static int watermark_boost_factor __read_mostly = 15000;
275 static int watermark_scale_factor = 10;
277 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
279 EXPORT_SYMBOL(movable_zone);
282 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
283 unsigned int nr_online_nodes __read_mostly = 1;
284 EXPORT_SYMBOL(nr_node_ids);
285 EXPORT_SYMBOL(nr_online_nodes);
288 static bool page_contains_unaccepted(struct page *page, unsigned int order);
289 static void accept_page(struct page *page, unsigned int order);
290 static bool try_to_accept_memory(struct zone *zone, unsigned int order);
291 static inline bool has_unaccepted_memory(void);
292 static bool __free_unaccepted(struct page *page);
294 int page_group_by_mobility_disabled __read_mostly;
296 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
298 * During boot we initialize deferred pages on-demand, as needed, but once
299 * page_alloc_init_late() has finished, the deferred pages are all initialized,
300 * and we can permanently disable that path.
302 DEFINE_STATIC_KEY_TRUE(deferred_pages);
304 static inline bool deferred_pages_enabled(void)
306 return static_branch_unlikely(&deferred_pages);
310 * deferred_grow_zone() is __init, but it is called from
311 * get_page_from_freelist() during early boot until deferred_pages permanently
312 * disables this call. This is why we have refdata wrapper to avoid warning,
313 * and to ensure that the function body gets unloaded.
316 _deferred_grow_zone(struct zone *zone, unsigned int order)
318 return deferred_grow_zone(zone, order);
321 static inline bool deferred_pages_enabled(void)
325 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
327 /* Return a pointer to the bitmap storing bits affecting a block of pages */
328 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
331 #ifdef CONFIG_SPARSEMEM
332 return section_to_usemap(__pfn_to_section(pfn));
334 return page_zone(page)->pageblock_flags;
335 #endif /* CONFIG_SPARSEMEM */
338 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
340 #ifdef CONFIG_SPARSEMEM
341 pfn &= (PAGES_PER_SECTION-1);
343 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
344 #endif /* CONFIG_SPARSEMEM */
345 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
349 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
350 * @page: The page within the block of interest
351 * @pfn: The target page frame number
352 * @mask: mask of bits that the caller is interested in
354 * Return: pageblock_bits flags
356 unsigned long get_pfnblock_flags_mask(const struct page *page,
357 unsigned long pfn, unsigned long mask)
359 unsigned long *bitmap;
360 unsigned long bitidx, word_bitidx;
363 bitmap = get_pageblock_bitmap(page, pfn);
364 bitidx = pfn_to_bitidx(page, pfn);
365 word_bitidx = bitidx / BITS_PER_LONG;
366 bitidx &= (BITS_PER_LONG-1);
368 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
369 * a consistent read of the memory array, so that results, even though
370 * racy, are not corrupted.
372 word = READ_ONCE(bitmap[word_bitidx]);
373 return (word >> bitidx) & mask;
376 static __always_inline int get_pfnblock_migratetype(const struct page *page,
379 return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
383 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
384 * @page: The page within the block of interest
385 * @flags: The flags to set
386 * @pfn: The target page frame number
387 * @mask: mask of bits that the caller is interested in
389 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
393 unsigned long *bitmap;
394 unsigned long bitidx, word_bitidx;
397 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
398 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
400 bitmap = get_pageblock_bitmap(page, pfn);
401 bitidx = pfn_to_bitidx(page, pfn);
402 word_bitidx = bitidx / BITS_PER_LONG;
403 bitidx &= (BITS_PER_LONG-1);
405 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
410 word = READ_ONCE(bitmap[word_bitidx]);
412 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
415 void set_pageblock_migratetype(struct page *page, int migratetype)
417 if (unlikely(page_group_by_mobility_disabled &&
418 migratetype < MIGRATE_PCPTYPES))
419 migratetype = MIGRATE_UNMOVABLE;
421 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
422 page_to_pfn(page), MIGRATETYPE_MASK);
425 #ifdef CONFIG_DEBUG_VM
426 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
430 unsigned long pfn = page_to_pfn(page);
431 unsigned long sp, start_pfn;
434 seq = zone_span_seqbegin(zone);
435 start_pfn = zone->zone_start_pfn;
436 sp = zone->spanned_pages;
437 ret = !zone_spans_pfn(zone, pfn);
438 } while (zone_span_seqretry(zone, seq));
441 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
442 pfn, zone_to_nid(zone), zone->name,
443 start_pfn, start_pfn + sp);
449 * Temporary debugging check for pages not lying within a given zone.
451 static bool __maybe_unused bad_range(struct zone *zone, struct page *page)
453 if (page_outside_zone_boundaries(zone, page))
455 if (zone != page_zone(page))
461 static inline bool __maybe_unused bad_range(struct zone *zone, struct page *page)
467 static void bad_page(struct page *page, const char *reason)
469 static unsigned long resume;
470 static unsigned long nr_shown;
471 static unsigned long nr_unshown;
474 * Allow a burst of 60 reports, then keep quiet for that minute;
475 * or allow a steady drip of one report per second.
477 if (nr_shown == 60) {
478 if (time_before(jiffies, resume)) {
484 "BUG: Bad page state: %lu messages suppressed\n",
491 resume = jiffies + 60 * HZ;
493 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
494 current->comm, page_to_pfn(page));
495 dump_page(page, reason);
500 /* Leave bad fields for debug, except PageBuddy could make trouble */
501 page_mapcount_reset(page); /* remove PageBuddy */
502 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
505 static inline unsigned int order_to_pindex(int migratetype, int order)
507 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
508 if (order > PAGE_ALLOC_COSTLY_ORDER) {
509 VM_BUG_ON(order != HPAGE_PMD_ORDER);
510 return NR_LOWORDER_PCP_LISTS;
513 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
516 return (MIGRATE_PCPTYPES * order) + migratetype;
519 static inline int pindex_to_order(unsigned int pindex)
521 int order = pindex / MIGRATE_PCPTYPES;
523 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
524 if (pindex == NR_LOWORDER_PCP_LISTS)
525 order = HPAGE_PMD_ORDER;
527 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
533 static inline bool pcp_allowed_order(unsigned int order)
535 if (order <= PAGE_ALLOC_COSTLY_ORDER)
537 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
538 if (order == HPAGE_PMD_ORDER)
545 * Higher-order pages are called "compound pages". They are structured thusly:
547 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
549 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
550 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
552 * The first tail page's ->compound_order holds the order of allocation.
553 * This usage means that zero-order pages may not be compound.
556 void prep_compound_page(struct page *page, unsigned int order)
559 int nr_pages = 1 << order;
562 for (i = 1; i < nr_pages; i++)
563 prep_compound_tail(page, i);
565 prep_compound_head(page, order);
568 static inline void set_buddy_order(struct page *page, unsigned int order)
570 set_page_private(page, order);
571 __SetPageBuddy(page);
574 #ifdef CONFIG_COMPACTION
575 static inline struct capture_control *task_capc(struct zone *zone)
577 struct capture_control *capc = current->capture_control;
579 return unlikely(capc) &&
580 !(current->flags & PF_KTHREAD) &&
582 capc->cc->zone == zone ? capc : NULL;
586 compaction_capture(struct capture_control *capc, struct page *page,
587 int order, int migratetype)
589 if (!capc || order != capc->cc->order)
592 /* Do not accidentally pollute CMA or isolated regions*/
593 if (is_migrate_cma(migratetype) ||
594 is_migrate_isolate(migratetype))
598 * Do not let lower order allocations pollute a movable pageblock
599 * unless compaction is also requesting movable pages.
600 * This might let an unmovable request use a reclaimable pageblock
601 * and vice-versa but no more than normal fallback logic which can
602 * have trouble finding a high-order free page.
604 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE &&
605 capc->cc->migratetype != MIGRATE_MOVABLE)
613 static inline struct capture_control *task_capc(struct zone *zone)
619 compaction_capture(struct capture_control *capc, struct page *page,
620 int order, int migratetype)
624 #endif /* CONFIG_COMPACTION */
626 static inline void account_freepages(struct zone *zone, int nr_pages,
629 if (is_migrate_isolate(migratetype))
632 __mod_zone_page_state(zone, NR_FREE_PAGES, nr_pages);
634 if (is_migrate_cma(migratetype))
635 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, nr_pages);
638 /* Used for pages not on another list */
639 static inline void __add_to_free_list(struct page *page, struct zone *zone,
640 unsigned int order, int migratetype,
643 struct free_area *area = &zone->free_area[order];
645 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
646 "page type is %lu, passed migratetype is %d (nr=%d)\n",
647 get_pageblock_migratetype(page), migratetype, 1 << order);
650 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
652 list_add(&page->buddy_list, &area->free_list[migratetype]);
657 * Used for pages which are on another list. Move the pages to the tail
658 * of the list - so the moved pages won't immediately be considered for
659 * allocation again (e.g., optimization for memory onlining).
661 static inline void move_to_free_list(struct page *page, struct zone *zone,
662 unsigned int order, int old_mt, int new_mt)
664 struct free_area *area = &zone->free_area[order];
666 /* Free page moving can fail, so it happens before the type update */
667 VM_WARN_ONCE(get_pageblock_migratetype(page) != old_mt,
668 "page type is %lu, passed migratetype is %d (nr=%d)\n",
669 get_pageblock_migratetype(page), old_mt, 1 << order);
671 list_move_tail(&page->buddy_list, &area->free_list[new_mt]);
673 account_freepages(zone, -(1 << order), old_mt);
674 account_freepages(zone, 1 << order, new_mt);
677 static inline void __del_page_from_free_list(struct page *page, struct zone *zone,
678 unsigned int order, int migratetype)
680 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
681 "page type is %lu, passed migratetype is %d (nr=%d)\n",
682 get_pageblock_migratetype(page), migratetype, 1 << order);
684 /* clear reported state and update reported page count */
685 if (page_reported(page))
686 __ClearPageReported(page);
688 list_del(&page->buddy_list);
689 __ClearPageBuddy(page);
690 set_page_private(page, 0);
691 zone->free_area[order].nr_free--;
694 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
695 unsigned int order, int migratetype)
697 __del_page_from_free_list(page, zone, order, migratetype);
698 account_freepages(zone, -(1 << order), migratetype);
701 static inline struct page *get_page_from_free_area(struct free_area *area,
704 return list_first_entry_or_null(&area->free_list[migratetype],
705 struct page, buddy_list);
709 * If this is not the largest possible page, check if the buddy
710 * of the next-highest order is free. If it is, it's possible
711 * that pages are being freed that will coalesce soon. In case,
712 * that is happening, add the free page to the tail of the list
713 * so it's less likely to be used soon and more likely to be merged
714 * as a higher order page
717 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
718 struct page *page, unsigned int order)
720 unsigned long higher_page_pfn;
721 struct page *higher_page;
723 if (order >= MAX_PAGE_ORDER - 1)
726 higher_page_pfn = buddy_pfn & pfn;
727 higher_page = page + (higher_page_pfn - pfn);
729 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
734 * Freeing function for a buddy system allocator.
736 * The concept of a buddy system is to maintain direct-mapped table
737 * (containing bit values) for memory blocks of various "orders".
738 * The bottom level table contains the map for the smallest allocatable
739 * units of memory (here, pages), and each level above it describes
740 * pairs of units from the levels below, hence, "buddies".
741 * At a high level, all that happens here is marking the table entry
742 * at the bottom level available, and propagating the changes upward
743 * as necessary, plus some accounting needed to play nicely with other
744 * parts of the VM system.
745 * At each level, we keep a list of pages, which are heads of continuous
746 * free pages of length of (1 << order) and marked with PageBuddy.
747 * Page's order is recorded in page_private(page) field.
748 * So when we are allocating or freeing one, we can derive the state of the
749 * other. That is, if we allocate a small block, and both were
750 * free, the remainder of the region must be split into blocks.
751 * If a block is freed, and its buddy is also free, then this
752 * triggers coalescing into a block of larger size.
757 static inline void __free_one_page(struct page *page,
759 struct zone *zone, unsigned int order,
760 int migratetype, fpi_t fpi_flags)
762 struct capture_control *capc = task_capc(zone);
763 unsigned long buddy_pfn = 0;
764 unsigned long combined_pfn;
768 VM_BUG_ON(!zone_is_initialized(zone));
769 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
771 VM_BUG_ON(migratetype == -1);
772 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
773 VM_BUG_ON_PAGE(bad_range(zone, page), page);
775 account_freepages(zone, 1 << order, migratetype);
777 while (order < MAX_PAGE_ORDER) {
778 int buddy_mt = migratetype;
780 if (compaction_capture(capc, page, order, migratetype)) {
781 account_freepages(zone, -(1 << order), migratetype);
785 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
789 if (unlikely(order >= pageblock_order)) {
791 * We want to prevent merge between freepages on pageblock
792 * without fallbacks and normal pageblock. Without this,
793 * pageblock isolation could cause incorrect freepage or CMA
794 * accounting or HIGHATOMIC accounting.
796 buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
798 if (migratetype != buddy_mt &&
799 (!migratetype_is_mergeable(migratetype) ||
800 !migratetype_is_mergeable(buddy_mt)))
805 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
806 * merge with it and move up one order.
808 if (page_is_guard(buddy))
809 clear_page_guard(zone, buddy, order);
811 __del_page_from_free_list(buddy, zone, order, buddy_mt);
813 if (unlikely(buddy_mt != migratetype)) {
815 * Match buddy type. This ensures that an
816 * expand() down the line puts the sub-blocks
817 * on the right freelists.
819 set_pageblock_migratetype(buddy, migratetype);
822 combined_pfn = buddy_pfn & pfn;
823 page = page + (combined_pfn - pfn);
829 set_buddy_order(page, order);
831 if (fpi_flags & FPI_TO_TAIL)
833 else if (is_shuffle_order(order))
834 to_tail = shuffle_pick_tail();
836 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
838 __add_to_free_list(page, zone, order, migratetype, to_tail);
840 /* Notify page reporting subsystem of freed page */
841 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
842 page_reporting_notify_free(order);
846 * A bad page could be due to a number of fields. Instead of multiple branches,
847 * try and check multiple fields with one check. The caller must do a detailed
848 * check if necessary.
850 static inline bool page_expected_state(struct page *page,
851 unsigned long check_flags)
853 if (unlikely(atomic_read(&page->_mapcount) != -1))
856 if (unlikely((unsigned long)page->mapping |
857 page_ref_count(page) |
861 #ifdef CONFIG_PAGE_POOL
862 ((page->pp_magic & ~0x3UL) == PP_SIGNATURE) |
864 (page->flags & check_flags)))
870 static const char *page_bad_reason(struct page *page, unsigned long flags)
872 const char *bad_reason = NULL;
874 if (unlikely(atomic_read(&page->_mapcount) != -1))
875 bad_reason = "nonzero mapcount";
876 if (unlikely(page->mapping != NULL))
877 bad_reason = "non-NULL mapping";
878 if (unlikely(page_ref_count(page) != 0))
879 bad_reason = "nonzero _refcount";
880 if (unlikely(page->flags & flags)) {
881 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
882 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
884 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
887 if (unlikely(page->memcg_data))
888 bad_reason = "page still charged to cgroup";
890 #ifdef CONFIG_PAGE_POOL
891 if (unlikely((page->pp_magic & ~0x3UL) == PP_SIGNATURE))
892 bad_reason = "page_pool leak";
897 static void free_page_is_bad_report(struct page *page)
900 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
903 static inline bool free_page_is_bad(struct page *page)
905 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
908 /* Something has gone sideways, find it */
909 free_page_is_bad_report(page);
913 static inline bool is_check_pages_enabled(void)
915 return static_branch_unlikely(&check_pages_enabled);
918 static int free_tail_page_prepare(struct page *head_page, struct page *page)
920 struct folio *folio = (struct folio *)head_page;
924 * We rely page->lru.next never has bit 0 set, unless the page
925 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
927 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
929 if (!is_check_pages_enabled()) {
933 switch (page - head_page) {
935 /* the first tail page: these may be in place of ->mapping */
936 if (unlikely(folio_entire_mapcount(folio))) {
937 bad_page(page, "nonzero entire_mapcount");
940 if (unlikely(folio_large_mapcount(folio))) {
941 bad_page(page, "nonzero large_mapcount");
944 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
945 bad_page(page, "nonzero nr_pages_mapped");
948 if (unlikely(atomic_read(&folio->_pincount))) {
949 bad_page(page, "nonzero pincount");
954 /* the second tail page: deferred_list overlaps ->mapping */
955 if (unlikely(!list_empty(&folio->_deferred_list))) {
956 bad_page(page, "on deferred list");
961 if (page->mapping != TAIL_MAPPING) {
962 bad_page(page, "corrupted mapping in tail page");
967 if (unlikely(!PageTail(page))) {
968 bad_page(page, "PageTail not set");
971 if (unlikely(compound_head(page) != head_page)) {
972 bad_page(page, "compound_head not consistent");
977 page->mapping = NULL;
978 clear_compound_head(page);
983 * Skip KASAN memory poisoning when either:
985 * 1. For generic KASAN: deferred memory initialization has not yet completed.
986 * Tag-based KASAN modes skip pages freed via deferred memory initialization
987 * using page tags instead (see below).
988 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
989 * that error detection is disabled for accesses via the page address.
991 * Pages will have match-all tags in the following circumstances:
993 * 1. Pages are being initialized for the first time, including during deferred
994 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
995 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
996 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
997 * 3. The allocation was excluded from being checked due to sampling,
998 * see the call to kasan_unpoison_pages.
1000 * Poisoning pages during deferred memory init will greatly lengthen the
1001 * process and cause problem in large memory systems as the deferred pages
1002 * initialization is done with interrupt disabled.
1004 * Assuming that there will be no reference to those newly initialized
1005 * pages before they are ever allocated, this should have no effect on
1006 * KASAN memory tracking as the poison will be properly inserted at page
1007 * allocation time. The only corner case is when pages are allocated by
1008 * on-demand allocation and then freed again before the deferred pages
1009 * initialization is done, but this is not likely to happen.
1011 static inline bool should_skip_kasan_poison(struct page *page)
1013 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1014 return deferred_pages_enabled();
1016 return page_kasan_tag(page) == KASAN_TAG_KERNEL;
1019 void kernel_init_pages(struct page *page, int numpages)
1023 /* s390's use of memset() could override KASAN redzones. */
1024 kasan_disable_current();
1025 for (i = 0; i < numpages; i++)
1026 clear_highpage_kasan_tagged(page + i);
1027 kasan_enable_current();
1030 __always_inline bool free_pages_prepare(struct page *page,
1034 bool skip_kasan_poison = should_skip_kasan_poison(page);
1035 bool init = want_init_on_free();
1036 bool compound = PageCompound(page);
1038 VM_BUG_ON_PAGE(PageTail(page), page);
1040 trace_mm_page_free(page, order);
1041 kmsan_free_page(page, order);
1043 if (memcg_kmem_online() && PageMemcgKmem(page))
1044 __memcg_kmem_uncharge_page(page, order);
1046 if (unlikely(PageHWPoison(page)) && !order) {
1047 /* Do not let hwpoison pages hit pcplists/buddy */
1048 reset_page_owner(page, order);
1049 page_table_check_free(page, order);
1050 pgalloc_tag_sub(page, 1 << order);
1054 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1057 * Check tail pages before head page information is cleared to
1058 * avoid checking PageCompound for order-0 pages.
1060 if (unlikely(order)) {
1064 page[1].flags &= ~PAGE_FLAGS_SECOND;
1065 for (i = 1; i < (1 << order); i++) {
1067 bad += free_tail_page_prepare(page, page + i);
1068 if (is_check_pages_enabled()) {
1069 if (free_page_is_bad(page + i)) {
1074 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1077 if (PageMappingFlags(page))
1078 page->mapping = NULL;
1079 if (is_check_pages_enabled()) {
1080 if (free_page_is_bad(page))
1086 page_cpupid_reset_last(page);
1087 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1088 reset_page_owner(page, order);
1089 page_table_check_free(page, order);
1090 pgalloc_tag_sub(page, 1 << order);
1092 if (!PageHighMem(page)) {
1093 debug_check_no_locks_freed(page_address(page),
1094 PAGE_SIZE << order);
1095 debug_check_no_obj_freed(page_address(page),
1096 PAGE_SIZE << order);
1099 kernel_poison_pages(page, 1 << order);
1102 * As memory initialization might be integrated into KASAN,
1103 * KASAN poisoning and memory initialization code must be
1104 * kept together to avoid discrepancies in behavior.
1106 * With hardware tag-based KASAN, memory tags must be set before the
1107 * page becomes unavailable via debug_pagealloc or arch_free_page.
1109 if (!skip_kasan_poison) {
1110 kasan_poison_pages(page, order, init);
1112 /* Memory is already initialized if KASAN did it internally. */
1113 if (kasan_has_integrated_init())
1117 kernel_init_pages(page, 1 << order);
1120 * arch_free_page() can make the page's contents inaccessible. s390
1121 * does this. So nothing which can access the page's contents should
1122 * happen after this.
1124 arch_free_page(page, order);
1126 debug_pagealloc_unmap_pages(page, 1 << order);
1132 * Frees a number of pages from the PCP lists
1133 * Assumes all pages on list are in same zone.
1134 * count is the number of pages to free.
1136 static void free_pcppages_bulk(struct zone *zone, int count,
1137 struct per_cpu_pages *pcp,
1140 unsigned long flags;
1145 * Ensure proper count is passed which otherwise would stuck in the
1146 * below while (list_empty(list)) loop.
1148 count = min(pcp->count, count);
1150 /* Ensure requested pindex is drained first. */
1151 pindex = pindex - 1;
1153 spin_lock_irqsave(&zone->lock, flags);
1156 struct list_head *list;
1159 /* Remove pages from lists in a round-robin fashion. */
1161 if (++pindex > NR_PCP_LISTS - 1)
1163 list = &pcp->lists[pindex];
1164 } while (list_empty(list));
1166 order = pindex_to_order(pindex);
1167 nr_pages = 1 << order;
1172 page = list_last_entry(list, struct page, pcp_list);
1173 pfn = page_to_pfn(page);
1174 mt = get_pfnblock_migratetype(page, pfn);
1176 /* must delete to avoid corrupting pcp list */
1177 list_del(&page->pcp_list);
1179 pcp->count -= nr_pages;
1181 __free_one_page(page, pfn, zone, order, mt, FPI_NONE);
1182 trace_mm_page_pcpu_drain(page, order, mt);
1183 } while (count > 0 && !list_empty(list));
1186 spin_unlock_irqrestore(&zone->lock, flags);
1189 static void free_one_page(struct zone *zone, struct page *page,
1190 unsigned long pfn, unsigned int order,
1193 unsigned long flags;
1196 spin_lock_irqsave(&zone->lock, flags);
1197 migratetype = get_pfnblock_migratetype(page, pfn);
1198 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1199 spin_unlock_irqrestore(&zone->lock, flags);
1202 static void __free_pages_ok(struct page *page, unsigned int order,
1205 unsigned long pfn = page_to_pfn(page);
1206 struct zone *zone = page_zone(page);
1208 if (!free_pages_prepare(page, order))
1211 free_one_page(zone, page, pfn, order, fpi_flags);
1213 __count_vm_events(PGFREE, 1 << order);
1216 void __free_pages_core(struct page *page, unsigned int order)
1218 unsigned int nr_pages = 1 << order;
1219 struct page *p = page;
1223 * When initializing the memmap, __init_single_page() sets the refcount
1224 * of all pages to 1 ("allocated"/"not free"). We have to set the
1225 * refcount of all involved pages to 0.
1228 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1230 __ClearPageReserved(p);
1231 set_page_count(p, 0);
1233 __ClearPageReserved(p);
1234 set_page_count(p, 0);
1236 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1238 if (page_contains_unaccepted(page, order)) {
1239 if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1242 accept_page(page, order);
1246 * Bypass PCP and place fresh pages right to the tail, primarily
1247 * relevant for memory onlining.
1249 __free_pages_ok(page, order, FPI_TO_TAIL);
1253 * Check that the whole (or subset of) a pageblock given by the interval of
1254 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1255 * with the migration of free compaction scanner.
1257 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1259 * It's possible on some configurations to have a setup like node0 node1 node0
1260 * i.e. it's possible that all pages within a zones range of pages do not
1261 * belong to a single zone. We assume that a border between node0 and node1
1262 * can occur within a single pageblock, but not a node0 node1 node0
1263 * interleaving within a single pageblock. It is therefore sufficient to check
1264 * the first and last page of a pageblock and avoid checking each individual
1265 * page in a pageblock.
1267 * Note: the function may return non-NULL struct page even for a page block
1268 * which contains a memory hole (i.e. there is no physical memory for a subset
1269 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1270 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1271 * even though the start pfn is online and valid. This should be safe most of
1272 * the time because struct pages are still initialized via init_unavailable_range()
1273 * and pfn walkers shouldn't touch any physical memory range for which they do
1274 * not recognize any specific metadata in struct pages.
1276 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1277 unsigned long end_pfn, struct zone *zone)
1279 struct page *start_page;
1280 struct page *end_page;
1282 /* end_pfn is one past the range we are checking */
1285 if (!pfn_valid(end_pfn))
1288 start_page = pfn_to_online_page(start_pfn);
1292 if (page_zone(start_page) != zone)
1295 end_page = pfn_to_page(end_pfn);
1297 /* This gives a shorter code than deriving page_zone(end_page) */
1298 if (page_zone_id(start_page) != page_zone_id(end_page))
1305 * The order of subdivision here is critical for the IO subsystem.
1306 * Please do not alter this order without good reasons and regression
1307 * testing. Specifically, as large blocks of memory are subdivided,
1308 * the order in which smaller blocks are delivered depends on the order
1309 * they're subdivided in this function. This is the primary factor
1310 * influencing the order in which pages are delivered to the IO
1311 * subsystem according to empirical testing, and this is also justified
1312 * by considering the behavior of a buddy system containing a single
1313 * large block of memory acted on by a series of small allocations.
1314 * This behavior is a critical factor in sglist merging's success.
1318 static inline void expand(struct zone *zone, struct page *page,
1319 int low, int high, int migratetype)
1321 unsigned long size = 1 << high;
1322 unsigned long nr_added = 0;
1324 while (high > low) {
1327 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1330 * Mark as guard pages (or page), that will allow to
1331 * merge back to allocator when buddy will be freed.
1332 * Corresponding page table entries will not be touched,
1333 * pages will stay not present in virtual address space
1335 if (set_page_guard(zone, &page[size], high))
1338 __add_to_free_list(&page[size], zone, high, migratetype, false);
1339 set_buddy_order(&page[size], high);
1342 account_freepages(zone, nr_added, migratetype);
1345 static void check_new_page_bad(struct page *page)
1347 if (unlikely(page->flags & __PG_HWPOISON)) {
1348 /* Don't complain about hwpoisoned pages */
1349 page_mapcount_reset(page); /* remove PageBuddy */
1354 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1358 * This page is about to be returned from the page allocator
1360 static bool check_new_page(struct page *page)
1362 if (likely(page_expected_state(page,
1363 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1366 check_new_page_bad(page);
1370 static inline bool check_new_pages(struct page *page, unsigned int order)
1372 if (is_check_pages_enabled()) {
1373 for (int i = 0; i < (1 << order); i++) {
1374 struct page *p = page + i;
1376 if (check_new_page(p))
1384 static inline bool should_skip_kasan_unpoison(gfp_t flags)
1386 /* Don't skip if a software KASAN mode is enabled. */
1387 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1388 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1391 /* Skip, if hardware tag-based KASAN is not enabled. */
1392 if (!kasan_hw_tags_enabled())
1396 * With hardware tag-based KASAN enabled, skip if this has been
1397 * requested via __GFP_SKIP_KASAN.
1399 return flags & __GFP_SKIP_KASAN;
1402 static inline bool should_skip_init(gfp_t flags)
1404 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1405 if (!kasan_hw_tags_enabled())
1408 /* For hardware tag-based KASAN, skip if requested. */
1409 return (flags & __GFP_SKIP_ZERO);
1412 inline void post_alloc_hook(struct page *page, unsigned int order,
1415 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1416 !should_skip_init(gfp_flags);
1417 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1420 set_page_private(page, 0);
1421 set_page_refcounted(page);
1423 arch_alloc_page(page, order);
1424 debug_pagealloc_map_pages(page, 1 << order);
1427 * Page unpoisoning must happen before memory initialization.
1428 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1429 * allocations and the page unpoisoning code will complain.
1431 kernel_unpoison_pages(page, 1 << order);
1434 * As memory initialization might be integrated into KASAN,
1435 * KASAN unpoisoning and memory initializion code must be
1436 * kept together to avoid discrepancies in behavior.
1440 * If memory tags should be zeroed
1441 * (which happens only when memory should be initialized as well).
1444 /* Initialize both memory and memory tags. */
1445 for (i = 0; i != 1 << order; ++i)
1446 tag_clear_highpage(page + i);
1448 /* Take note that memory was initialized by the loop above. */
1451 if (!should_skip_kasan_unpoison(gfp_flags) &&
1452 kasan_unpoison_pages(page, order, init)) {
1453 /* Take note that memory was initialized by KASAN. */
1454 if (kasan_has_integrated_init())
1458 * If memory tags have not been set by KASAN, reset the page
1459 * tags to ensure page_address() dereferencing does not fault.
1461 for (i = 0; i != 1 << order; ++i)
1462 page_kasan_tag_reset(page + i);
1464 /* If memory is still not initialized, initialize it now. */
1466 kernel_init_pages(page, 1 << order);
1468 set_page_owner(page, order, gfp_flags);
1469 page_table_check_alloc(page, order);
1470 pgalloc_tag_add(page, current, 1 << order);
1473 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1474 unsigned int alloc_flags)
1476 post_alloc_hook(page, order, gfp_flags);
1478 if (order && (gfp_flags & __GFP_COMP))
1479 prep_compound_page(page, order);
1482 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1483 * allocate the page. The expectation is that the caller is taking
1484 * steps that will free more memory. The caller should avoid the page
1485 * being used for !PFMEMALLOC purposes.
1487 if (alloc_flags & ALLOC_NO_WATERMARKS)
1488 set_page_pfmemalloc(page);
1490 clear_page_pfmemalloc(page);
1494 * Go through the free lists for the given migratetype and remove
1495 * the smallest available page from the freelists
1497 static __always_inline
1498 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1501 unsigned int current_order;
1502 struct free_area *area;
1505 /* Find a page of the appropriate size in the preferred list */
1506 for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
1507 area = &(zone->free_area[current_order]);
1508 page = get_page_from_free_area(area, migratetype);
1511 del_page_from_free_list(page, zone, current_order, migratetype);
1512 expand(zone, page, order, current_order, migratetype);
1513 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1514 pcp_allowed_order(order) &&
1515 migratetype < MIGRATE_PCPTYPES);
1524 * This array describes the order lists are fallen back to when
1525 * the free lists for the desirable migrate type are depleted
1527 * The other migratetypes do not have fallbacks.
1529 static int fallbacks[MIGRATE_PCPTYPES][MIGRATE_PCPTYPES - 1] = {
1530 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1531 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1532 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1536 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1539 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1542 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1543 unsigned int order) { return NULL; }
1547 * Change the type of a block and move all its free pages to that
1550 static int __move_freepages_block(struct zone *zone, unsigned long start_pfn,
1551 int old_mt, int new_mt)
1554 unsigned long pfn, end_pfn;
1556 int pages_moved = 0;
1558 VM_WARN_ON(start_pfn & (pageblock_nr_pages - 1));
1559 end_pfn = pageblock_end_pfn(start_pfn);
1561 for (pfn = start_pfn; pfn < end_pfn;) {
1562 page = pfn_to_page(pfn);
1563 if (!PageBuddy(page)) {
1568 /* Make sure we are not inadvertently changing nodes */
1569 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1570 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1572 order = buddy_order(page);
1574 move_to_free_list(page, zone, order, old_mt, new_mt);
1577 pages_moved += 1 << order;
1580 set_pageblock_migratetype(pfn_to_page(start_pfn), new_mt);
1585 static bool prep_move_freepages_block(struct zone *zone, struct page *page,
1586 unsigned long *start_pfn,
1587 int *num_free, int *num_movable)
1589 unsigned long pfn, start, end;
1591 pfn = page_to_pfn(page);
1592 start = pageblock_start_pfn(pfn);
1593 end = pageblock_end_pfn(pfn);
1596 * The caller only has the lock for @zone, don't touch ranges
1597 * that straddle into other zones. While we could move part of
1598 * the range that's inside the zone, this call is usually
1599 * accompanied by other operations such as migratetype updates
1600 * which also should be locked.
1602 if (!zone_spans_pfn(zone, start))
1604 if (!zone_spans_pfn(zone, end - 1))
1612 for (pfn = start; pfn < end;) {
1613 page = pfn_to_page(pfn);
1614 if (PageBuddy(page)) {
1615 int nr = 1 << buddy_order(page);
1622 * We assume that pages that could be isolated for
1623 * migration are movable. But we don't actually try
1624 * isolating, as that would be expensive.
1626 if (PageLRU(page) || __PageMovable(page))
1635 static int move_freepages_block(struct zone *zone, struct page *page,
1636 int old_mt, int new_mt)
1638 unsigned long start_pfn;
1640 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
1643 return __move_freepages_block(zone, start_pfn, old_mt, new_mt);
1646 #ifdef CONFIG_MEMORY_ISOLATION
1647 /* Look for a buddy that straddles start_pfn */
1648 static unsigned long find_large_buddy(unsigned long start_pfn)
1652 unsigned long pfn = start_pfn;
1654 while (!PageBuddy(page = pfn_to_page(pfn))) {
1656 if (++order > MAX_PAGE_ORDER)
1658 pfn &= ~0UL << order;
1662 * Found a preceding buddy, but does it straddle?
1664 if (pfn + (1 << buddy_order(page)) > start_pfn)
1671 /* Split a multi-block free page into its individual pageblocks */
1672 static void split_large_buddy(struct zone *zone, struct page *page,
1673 unsigned long pfn, int order)
1675 unsigned long end_pfn = pfn + (1 << order);
1677 VM_WARN_ON_ONCE(order <= pageblock_order);
1678 VM_WARN_ON_ONCE(pfn & (pageblock_nr_pages - 1));
1680 /* Caller removed page from freelist, buddy info cleared! */
1681 VM_WARN_ON_ONCE(PageBuddy(page));
1683 while (pfn != end_pfn) {
1684 int mt = get_pfnblock_migratetype(page, pfn);
1686 __free_one_page(page, pfn, zone, pageblock_order, mt, FPI_NONE);
1687 pfn += pageblock_nr_pages;
1688 page = pfn_to_page(pfn);
1693 * move_freepages_block_isolate - move free pages in block for page isolation
1695 * @page: the pageblock page
1696 * @migratetype: migratetype to set on the pageblock
1698 * This is similar to move_freepages_block(), but handles the special
1699 * case encountered in page isolation, where the block of interest
1700 * might be part of a larger buddy spanning multiple pageblocks.
1702 * Unlike the regular page allocator path, which moves pages while
1703 * stealing buddies off the freelist, page isolation is interested in
1704 * arbitrary pfn ranges that may have overlapping buddies on both ends.
1706 * This function handles that. Straddling buddies are split into
1707 * individual pageblocks. Only the block of interest is moved.
1709 * Returns %true if pages could be moved, %false otherwise.
1711 bool move_freepages_block_isolate(struct zone *zone, struct page *page,
1714 unsigned long start_pfn, pfn;
1716 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
1719 /* No splits needed if buddies can't span multiple blocks */
1720 if (pageblock_order == MAX_PAGE_ORDER)
1723 /* We're a tail block in a larger buddy */
1724 pfn = find_large_buddy(start_pfn);
1725 if (pfn != start_pfn) {
1726 struct page *buddy = pfn_to_page(pfn);
1727 int order = buddy_order(buddy);
1729 del_page_from_free_list(buddy, zone, order,
1730 get_pfnblock_migratetype(buddy, pfn));
1731 set_pageblock_migratetype(page, migratetype);
1732 split_large_buddy(zone, buddy, pfn, order);
1736 /* We're the starting block of a larger buddy */
1737 if (PageBuddy(page) && buddy_order(page) > pageblock_order) {
1738 int order = buddy_order(page);
1740 del_page_from_free_list(page, zone, order,
1741 get_pfnblock_migratetype(page, pfn));
1742 set_pageblock_migratetype(page, migratetype);
1743 split_large_buddy(zone, page, pfn, order);
1747 __move_freepages_block(zone, start_pfn,
1748 get_pfnblock_migratetype(page, start_pfn),
1752 #endif /* CONFIG_MEMORY_ISOLATION */
1754 static void change_pageblock_range(struct page *pageblock_page,
1755 int start_order, int migratetype)
1757 int nr_pageblocks = 1 << (start_order - pageblock_order);
1759 while (nr_pageblocks--) {
1760 set_pageblock_migratetype(pageblock_page, migratetype);
1761 pageblock_page += pageblock_nr_pages;
1766 * When we are falling back to another migratetype during allocation, try to
1767 * steal extra free pages from the same pageblocks to satisfy further
1768 * allocations, instead of polluting multiple pageblocks.
1770 * If we are stealing a relatively large buddy page, it is likely there will
1771 * be more free pages in the pageblock, so try to steal them all. For
1772 * reclaimable and unmovable allocations, we steal regardless of page size,
1773 * as fragmentation caused by those allocations polluting movable pageblocks
1774 * is worse than movable allocations stealing from unmovable and reclaimable
1777 static bool can_steal_fallback(unsigned int order, int start_mt)
1780 * Leaving this order check is intended, although there is
1781 * relaxed order check in next check. The reason is that
1782 * we can actually steal whole pageblock if this condition met,
1783 * but, below check doesn't guarantee it and that is just heuristic
1784 * so could be changed anytime.
1786 if (order >= pageblock_order)
1789 if (order >= pageblock_order / 2 ||
1790 start_mt == MIGRATE_RECLAIMABLE ||
1791 start_mt == MIGRATE_UNMOVABLE ||
1792 page_group_by_mobility_disabled)
1798 static inline bool boost_watermark(struct zone *zone)
1800 unsigned long max_boost;
1802 if (!watermark_boost_factor)
1805 * Don't bother in zones that are unlikely to produce results.
1806 * On small machines, including kdump capture kernels running
1807 * in a small area, boosting the watermark can cause an out of
1808 * memory situation immediately.
1810 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1813 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1814 watermark_boost_factor, 10000);
1817 * high watermark may be uninitialised if fragmentation occurs
1818 * very early in boot so do not boost. We do not fall
1819 * through and boost by pageblock_nr_pages as failing
1820 * allocations that early means that reclaim is not going
1821 * to help and it may even be impossible to reclaim the
1822 * boosted watermark resulting in a hang.
1827 max_boost = max(pageblock_nr_pages, max_boost);
1829 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1836 * This function implements actual steal behaviour. If order is large enough, we
1837 * can claim the whole pageblock for the requested migratetype. If not, we check
1838 * the pageblock for constituent pages; if at least half of the pages are free
1839 * or compatible, we can still claim the whole block, so pages freed in the
1840 * future will be put on the correct free list. Otherwise, we isolate exactly
1841 * the order we need from the fallback block and leave its migratetype alone.
1843 static struct page *
1844 steal_suitable_fallback(struct zone *zone, struct page *page,
1845 int current_order, int order, int start_type,
1846 unsigned int alloc_flags, bool whole_block)
1848 int free_pages, movable_pages, alike_pages;
1849 unsigned long start_pfn;
1852 block_type = get_pageblock_migratetype(page);
1855 * This can happen due to races and we want to prevent broken
1856 * highatomic accounting.
1858 if (is_migrate_highatomic(block_type))
1861 /* Take ownership for orders >= pageblock_order */
1862 if (current_order >= pageblock_order) {
1863 del_page_from_free_list(page, zone, current_order, block_type);
1864 change_pageblock_range(page, current_order, start_type);
1865 expand(zone, page, order, current_order, start_type);
1870 * Boost watermarks to increase reclaim pressure to reduce the
1871 * likelihood of future fallbacks. Wake kswapd now as the node
1872 * may be balanced overall and kswapd will not wake naturally.
1874 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1875 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1877 /* We are not allowed to try stealing from the whole block */
1881 /* moving whole block can fail due to zone boundary conditions */
1882 if (!prep_move_freepages_block(zone, page, &start_pfn, &free_pages,
1887 * Determine how many pages are compatible with our allocation.
1888 * For movable allocation, it's the number of movable pages which
1889 * we just obtained. For other types it's a bit more tricky.
1891 if (start_type == MIGRATE_MOVABLE) {
1892 alike_pages = movable_pages;
1895 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1896 * to MOVABLE pageblock, consider all non-movable pages as
1897 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1898 * vice versa, be conservative since we can't distinguish the
1899 * exact migratetype of non-movable pages.
1901 if (block_type == MIGRATE_MOVABLE)
1902 alike_pages = pageblock_nr_pages
1903 - (free_pages + movable_pages);
1908 * If a sufficient number of pages in the block are either free or of
1909 * compatible migratability as our allocation, claim the whole block.
1911 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1912 page_group_by_mobility_disabled) {
1913 __move_freepages_block(zone, start_pfn, block_type, start_type);
1914 return __rmqueue_smallest(zone, order, start_type);
1918 del_page_from_free_list(page, zone, current_order, block_type);
1919 expand(zone, page, order, current_order, block_type);
1924 * Check whether there is a suitable fallback freepage with requested order.
1925 * If only_stealable is true, this function returns fallback_mt only if
1926 * we can steal other freepages all together. This would help to reduce
1927 * fragmentation due to mixed migratetype pages in one pageblock.
1929 int find_suitable_fallback(struct free_area *area, unsigned int order,
1930 int migratetype, bool only_stealable, bool *can_steal)
1935 if (area->nr_free == 0)
1939 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
1940 fallback_mt = fallbacks[migratetype][i];
1941 if (free_area_empty(area, fallback_mt))
1944 if (can_steal_fallback(order, migratetype))
1947 if (!only_stealable)
1958 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1959 * there are no empty page blocks that contain a page with a suitable order
1961 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone)
1964 unsigned long max_managed, flags;
1967 * The number reserved as: minimum is 1 pageblock, maximum is
1968 * roughly 1% of a zone. But if 1% of a zone falls below a
1969 * pageblock size, then don't reserve any pageblocks.
1970 * Check is race-prone but harmless.
1972 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
1974 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
1975 if (zone->nr_reserved_highatomic >= max_managed)
1978 spin_lock_irqsave(&zone->lock, flags);
1980 /* Recheck the nr_reserved_highatomic limit under the lock */
1981 if (zone->nr_reserved_highatomic >= max_managed)
1985 mt = get_pageblock_migratetype(page);
1986 /* Only reserve normal pageblocks (i.e., they can merge with others) */
1987 if (migratetype_is_mergeable(mt))
1988 if (move_freepages_block(zone, page, mt,
1989 MIGRATE_HIGHATOMIC) != -1)
1990 zone->nr_reserved_highatomic += pageblock_nr_pages;
1993 spin_unlock_irqrestore(&zone->lock, flags);
1997 * Used when an allocation is about to fail under memory pressure. This
1998 * potentially hurts the reliability of high-order allocations when under
1999 * intense memory pressure but failed atomic allocations should be easier
2000 * to recover from than an OOM.
2002 * If @force is true, try to unreserve a pageblock even though highatomic
2003 * pageblock is exhausted.
2005 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2008 struct zonelist *zonelist = ac->zonelist;
2009 unsigned long flags;
2016 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2019 * Preserve at least one pageblock unless memory pressure
2022 if (!force && zone->nr_reserved_highatomic <=
2026 spin_lock_irqsave(&zone->lock, flags);
2027 for (order = 0; order < NR_PAGE_ORDERS; order++) {
2028 struct free_area *area = &(zone->free_area[order]);
2031 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2035 mt = get_pageblock_migratetype(page);
2037 * In page freeing path, migratetype change is racy so
2038 * we can counter several free pages in a pageblock
2039 * in this loop although we changed the pageblock type
2040 * from highatomic to ac->migratetype. So we should
2041 * adjust the count once.
2043 if (is_migrate_highatomic(mt)) {
2045 * It should never happen but changes to
2046 * locking could inadvertently allow a per-cpu
2047 * drain to add pages to MIGRATE_HIGHATOMIC
2048 * while unreserving so be safe and watch for
2051 zone->nr_reserved_highatomic -= min(
2053 zone->nr_reserved_highatomic);
2057 * Convert to ac->migratetype and avoid the normal
2058 * pageblock stealing heuristics. Minimally, the caller
2059 * is doing the work and needs the pages. More
2060 * importantly, if the block was always converted to
2061 * MIGRATE_UNMOVABLE or another type then the number
2062 * of pageblocks that cannot be completely freed
2065 ret = move_freepages_block(zone, page, mt,
2068 * Reserving this block already succeeded, so this should
2069 * not fail on zone boundaries.
2071 WARN_ON_ONCE(ret == -1);
2073 spin_unlock_irqrestore(&zone->lock, flags);
2077 spin_unlock_irqrestore(&zone->lock, flags);
2084 * Try finding a free buddy page on the fallback list and put it on the free
2085 * list of requested migratetype, possibly along with other pages from the same
2086 * block, depending on fragmentation avoidance heuristics. Returns true if
2087 * fallback was found so that __rmqueue_smallest() can grab it.
2089 * The use of signed ints for order and current_order is a deliberate
2090 * deviation from the rest of this file, to make the for loop
2091 * condition simpler.
2093 static __always_inline struct page *
2094 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2095 unsigned int alloc_flags)
2097 struct free_area *area;
2099 int min_order = order;
2105 * Do not steal pages from freelists belonging to other pageblocks
2106 * i.e. orders < pageblock_order. If there are no local zones free,
2107 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2109 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2110 min_order = pageblock_order;
2113 * Find the largest available free page in the other list. This roughly
2114 * approximates finding the pageblock with the most free pages, which
2115 * would be too costly to do exactly.
2117 for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2119 area = &(zone->free_area[current_order]);
2120 fallback_mt = find_suitable_fallback(area, current_order,
2121 start_migratetype, false, &can_steal);
2122 if (fallback_mt == -1)
2126 * We cannot steal all free pages from the pageblock and the
2127 * requested migratetype is movable. In that case it's better to
2128 * steal and split the smallest available page instead of the
2129 * largest available page, because even if the next movable
2130 * allocation falls back into a different pageblock than this
2131 * one, it won't cause permanent fragmentation.
2133 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2134 && current_order > order)
2143 for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
2144 area = &(zone->free_area[current_order]);
2145 fallback_mt = find_suitable_fallback(area, current_order,
2146 start_migratetype, false, &can_steal);
2147 if (fallback_mt != -1)
2152 * This should not happen - we already found a suitable fallback
2153 * when looking for the largest page.
2155 VM_BUG_ON(current_order > MAX_PAGE_ORDER);
2158 page = get_page_from_free_area(area, fallback_mt);
2160 /* take off list, maybe claim block, expand remainder */
2161 page = steal_suitable_fallback(zone, page, current_order, order,
2162 start_migratetype, alloc_flags, can_steal);
2164 trace_mm_page_alloc_extfrag(page, order, current_order,
2165 start_migratetype, fallback_mt);
2171 * Do the hard work of removing an element from the buddy allocator.
2172 * Call me with the zone->lock already held.
2174 static __always_inline struct page *
2175 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2176 unsigned int alloc_flags)
2180 if (IS_ENABLED(CONFIG_CMA)) {
2182 * Balance movable allocations between regular and CMA areas by
2183 * allocating from CMA when over half of the zone's free memory
2184 * is in the CMA area.
2186 if (alloc_flags & ALLOC_CMA &&
2187 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2188 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2189 page = __rmqueue_cma_fallback(zone, order);
2195 page = __rmqueue_smallest(zone, order, migratetype);
2196 if (unlikely(!page)) {
2197 if (alloc_flags & ALLOC_CMA)
2198 page = __rmqueue_cma_fallback(zone, order);
2201 page = __rmqueue_fallback(zone, order, migratetype,
2208 * Obtain a specified number of elements from the buddy allocator, all under
2209 * a single hold of the lock, for efficiency. Add them to the supplied list.
2210 * Returns the number of new pages which were placed at *list.
2212 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2213 unsigned long count, struct list_head *list,
2214 int migratetype, unsigned int alloc_flags)
2216 unsigned long flags;
2219 spin_lock_irqsave(&zone->lock, flags);
2220 for (i = 0; i < count; ++i) {
2221 struct page *page = __rmqueue(zone, order, migratetype,
2223 if (unlikely(page == NULL))
2227 * Split buddy pages returned by expand() are received here in
2228 * physical page order. The page is added to the tail of
2229 * caller's list. From the callers perspective, the linked list
2230 * is ordered by page number under some conditions. This is
2231 * useful for IO devices that can forward direction from the
2232 * head, thus also in the physical page order. This is useful
2233 * for IO devices that can merge IO requests if the physical
2234 * pages are ordered properly.
2236 list_add_tail(&page->pcp_list, list);
2238 spin_unlock_irqrestore(&zone->lock, flags);
2244 * Called from the vmstat counter updater to decay the PCP high.
2245 * Return whether there are addition works to do.
2247 int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2249 int high_min, to_drain, batch;
2252 high_min = READ_ONCE(pcp->high_min);
2253 batch = READ_ONCE(pcp->batch);
2255 * Decrease pcp->high periodically to try to free possible
2256 * idle PCP pages. And, avoid to free too many pages to
2257 * control latency. This caps pcp->high decrement too.
2259 if (pcp->high > high_min) {
2260 pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2261 pcp->high - (pcp->high >> 3), high_min);
2262 if (pcp->high > high_min)
2266 to_drain = pcp->count - pcp->high;
2268 spin_lock(&pcp->lock);
2269 free_pcppages_bulk(zone, to_drain, pcp, 0);
2270 spin_unlock(&pcp->lock);
2279 * Called from the vmstat counter updater to drain pagesets of this
2280 * currently executing processor on remote nodes after they have
2283 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2285 int to_drain, batch;
2287 batch = READ_ONCE(pcp->batch);
2288 to_drain = min(pcp->count, batch);
2290 spin_lock(&pcp->lock);
2291 free_pcppages_bulk(zone, to_drain, pcp, 0);
2292 spin_unlock(&pcp->lock);
2298 * Drain pcplists of the indicated processor and zone.
2300 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2302 struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2303 int count = READ_ONCE(pcp->count);
2306 int to_drain = min(count, pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
2309 spin_lock(&pcp->lock);
2310 free_pcppages_bulk(zone, to_drain, pcp, 0);
2311 spin_unlock(&pcp->lock);
2316 * Drain pcplists of all zones on the indicated processor.
2318 static void drain_pages(unsigned int cpu)
2322 for_each_populated_zone(zone) {
2323 drain_pages_zone(cpu, zone);
2328 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2330 void drain_local_pages(struct zone *zone)
2332 int cpu = smp_processor_id();
2335 drain_pages_zone(cpu, zone);
2341 * The implementation of drain_all_pages(), exposing an extra parameter to
2342 * drain on all cpus.
2344 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2345 * not empty. The check for non-emptiness can however race with a free to
2346 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2347 * that need the guarantee that every CPU has drained can disable the
2348 * optimizing racy check.
2350 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2355 * Allocate in the BSS so we won't require allocation in
2356 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2358 static cpumask_t cpus_with_pcps;
2361 * Do not drain if one is already in progress unless it's specific to
2362 * a zone. Such callers are primarily CMA and memory hotplug and need
2363 * the drain to be complete when the call returns.
2365 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2368 mutex_lock(&pcpu_drain_mutex);
2372 * We don't care about racing with CPU hotplug event
2373 * as offline notification will cause the notified
2374 * cpu to drain that CPU pcps and on_each_cpu_mask
2375 * disables preemption as part of its processing
2377 for_each_online_cpu(cpu) {
2378 struct per_cpu_pages *pcp;
2380 bool has_pcps = false;
2382 if (force_all_cpus) {
2384 * The pcp.count check is racy, some callers need a
2385 * guarantee that no cpu is missed.
2389 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2393 for_each_populated_zone(z) {
2394 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2403 cpumask_set_cpu(cpu, &cpus_with_pcps);
2405 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2408 for_each_cpu(cpu, &cpus_with_pcps) {
2410 drain_pages_zone(cpu, zone);
2415 mutex_unlock(&pcpu_drain_mutex);
2419 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2421 * When zone parameter is non-NULL, spill just the single zone's pages.
2423 void drain_all_pages(struct zone *zone)
2425 __drain_all_pages(zone, false);
2428 static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2430 int min_nr_free, max_nr_free;
2432 /* Free as much as possible if batch freeing high-order pages. */
2433 if (unlikely(free_high))
2434 return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2436 /* Check for PCP disabled or boot pageset */
2437 if (unlikely(high < batch))
2440 /* Leave at least pcp->batch pages on the list */
2441 min_nr_free = batch;
2442 max_nr_free = high - batch;
2445 * Increase the batch number to the number of the consecutive
2446 * freed pages to reduce zone lock contention.
2448 batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2453 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2454 int batch, bool free_high)
2456 int high, high_min, high_max;
2458 high_min = READ_ONCE(pcp->high_min);
2459 high_max = READ_ONCE(pcp->high_max);
2460 high = pcp->high = clamp(pcp->high, high_min, high_max);
2462 if (unlikely(!high))
2465 if (unlikely(free_high)) {
2466 pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2472 * If reclaim is active, limit the number of pages that can be
2473 * stored on pcp lists
2475 if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2476 int free_count = max_t(int, pcp->free_count, batch);
2478 pcp->high = max(high - free_count, high_min);
2479 return min(batch << 2, pcp->high);
2482 if (high_min == high_max)
2485 if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2486 int free_count = max_t(int, pcp->free_count, batch);
2488 pcp->high = max(high - free_count, high_min);
2489 high = max(pcp->count, high_min);
2490 } else if (pcp->count >= high) {
2491 int need_high = pcp->free_count + batch;
2493 /* pcp->high should be large enough to hold batch freed pages */
2494 if (pcp->high < need_high)
2495 pcp->high = clamp(need_high, high_min, high_max);
2501 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2502 struct page *page, int migratetype,
2507 bool free_high = false;
2510 * On freeing, reduce the number of pages that are batch allocated.
2511 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2514 pcp->alloc_factor >>= 1;
2515 __count_vm_events(PGFREE, 1 << order);
2516 pindex = order_to_pindex(migratetype, order);
2517 list_add(&page->pcp_list, &pcp->lists[pindex]);
2518 pcp->count += 1 << order;
2520 batch = READ_ONCE(pcp->batch);
2522 * As high-order pages other than THP's stored on PCP can contribute
2523 * to fragmentation, limit the number stored when PCP is heavily
2524 * freeing without allocation. The remainder after bulk freeing
2525 * stops will be drained from vmstat refresh context.
2527 if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2528 free_high = (pcp->free_count >= batch &&
2529 (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2530 (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2531 pcp->count >= READ_ONCE(batch)));
2532 pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2533 } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2534 pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2536 if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2537 pcp->free_count += (1 << order);
2538 high = nr_pcp_high(pcp, zone, batch, free_high);
2539 if (pcp->count >= high) {
2540 free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high),
2542 if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2543 zone_watermark_ok(zone, 0, high_wmark_pages(zone),
2545 clear_bit(ZONE_BELOW_HIGH, &zone->flags);
2552 void free_unref_page(struct page *page, unsigned int order)
2554 unsigned long __maybe_unused UP_flags;
2555 struct per_cpu_pages *pcp;
2557 unsigned long pfn = page_to_pfn(page);
2560 if (!pcp_allowed_order(order)) {
2561 __free_pages_ok(page, order, FPI_NONE);
2565 if (!free_pages_prepare(page, order))
2569 * We only track unmovable, reclaimable and movable on pcp lists.
2570 * Place ISOLATE pages on the isolated list because they are being
2571 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2572 * get those areas back if necessary. Otherwise, we may have to free
2573 * excessively into the page allocator
2575 migratetype = get_pfnblock_migratetype(page, pfn);
2576 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2577 if (unlikely(is_migrate_isolate(migratetype))) {
2578 free_one_page(page_zone(page), page, pfn, order, FPI_NONE);
2581 migratetype = MIGRATE_MOVABLE;
2584 zone = page_zone(page);
2585 pcp_trylock_prepare(UP_flags);
2586 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2588 free_unref_page_commit(zone, pcp, page, migratetype, order);
2589 pcp_spin_unlock(pcp);
2591 free_one_page(zone, page, pfn, order, FPI_NONE);
2593 pcp_trylock_finish(UP_flags);
2597 * Free a batch of folios
2599 void free_unref_folios(struct folio_batch *folios)
2601 unsigned long __maybe_unused UP_flags;
2602 struct per_cpu_pages *pcp = NULL;
2603 struct zone *locked_zone = NULL;
2606 /* Prepare folios for freeing */
2607 for (i = 0, j = 0; i < folios->nr; i++) {
2608 struct folio *folio = folios->folios[i];
2609 unsigned long pfn = folio_pfn(folio);
2610 unsigned int order = folio_order(folio);
2612 if (order > 0 && folio_test_large_rmappable(folio))
2613 folio_undo_large_rmappable(folio);
2614 if (!free_pages_prepare(&folio->page, order))
2617 * Free orders not handled on the PCP directly to the
2620 if (!pcp_allowed_order(order)) {
2621 free_one_page(folio_zone(folio), &folio->page,
2622 pfn, order, FPI_NONE);
2625 folio->private = (void *)(unsigned long)order;
2627 folios->folios[j] = folio;
2632 for (i = 0; i < folios->nr; i++) {
2633 struct folio *folio = folios->folios[i];
2634 struct zone *zone = folio_zone(folio);
2635 unsigned long pfn = folio_pfn(folio);
2636 unsigned int order = (unsigned long)folio->private;
2639 folio->private = NULL;
2640 migratetype = get_pfnblock_migratetype(&folio->page, pfn);
2642 /* Different zone requires a different pcp lock */
2643 if (zone != locked_zone ||
2644 is_migrate_isolate(migratetype)) {
2646 pcp_spin_unlock(pcp);
2647 pcp_trylock_finish(UP_flags);
2653 * Free isolated pages directly to the
2654 * allocator, see comment in free_unref_page.
2656 if (is_migrate_isolate(migratetype)) {
2657 free_one_page(zone, &folio->page, pfn,
2663 * trylock is necessary as folios may be getting freed
2664 * from IRQ or SoftIRQ context after an IO completion.
2666 pcp_trylock_prepare(UP_flags);
2667 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2668 if (unlikely(!pcp)) {
2669 pcp_trylock_finish(UP_flags);
2670 free_one_page(zone, &folio->page, pfn,
2678 * Non-isolated types over MIGRATE_PCPTYPES get added
2679 * to the MIGRATE_MOVABLE pcp list.
2681 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2682 migratetype = MIGRATE_MOVABLE;
2684 trace_mm_page_free_batched(&folio->page);
2685 free_unref_page_commit(zone, pcp, &folio->page, migratetype,
2690 pcp_spin_unlock(pcp);
2691 pcp_trylock_finish(UP_flags);
2693 folio_batch_reinit(folios);
2697 * split_page takes a non-compound higher-order page, and splits it into
2698 * n (1<<order) sub-pages: page[0..n]
2699 * Each sub-page must be freed individually.
2701 * Note: this is probably too low level an operation for use in drivers.
2702 * Please consult with lkml before using this in your driver.
2704 void split_page(struct page *page, unsigned int order)
2708 VM_BUG_ON_PAGE(PageCompound(page), page);
2709 VM_BUG_ON_PAGE(!page_count(page), page);
2711 for (i = 1; i < (1 << order); i++)
2712 set_page_refcounted(page + i);
2713 split_page_owner(page, order, 0);
2714 pgalloc_tag_split(page, 1 << order);
2715 split_page_memcg(page, order, 0);
2717 EXPORT_SYMBOL_GPL(split_page);
2719 int __isolate_free_page(struct page *page, unsigned int order)
2721 struct zone *zone = page_zone(page);
2722 int mt = get_pageblock_migratetype(page);
2724 if (!is_migrate_isolate(mt)) {
2725 unsigned long watermark;
2727 * Obey watermarks as if the page was being allocated. We can
2728 * emulate a high-order watermark check with a raised order-0
2729 * watermark, because we already know our high-order page
2732 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2733 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2737 del_page_from_free_list(page, zone, order, mt);
2740 * Set the pageblock if the isolated page is at least half of a
2743 if (order >= pageblock_order - 1) {
2744 struct page *endpage = page + (1 << order) - 1;
2745 for (; page < endpage; page += pageblock_nr_pages) {
2746 int mt = get_pageblock_migratetype(page);
2748 * Only change normal pageblocks (i.e., they can merge
2751 if (migratetype_is_mergeable(mt))
2752 move_freepages_block(zone, page, mt,
2757 return 1UL << order;
2761 * __putback_isolated_page - Return a now-isolated page back where we got it
2762 * @page: Page that was isolated
2763 * @order: Order of the isolated page
2764 * @mt: The page's pageblock's migratetype
2766 * This function is meant to return a page pulled from the free lists via
2767 * __isolate_free_page back to the free lists they were pulled from.
2769 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2771 struct zone *zone = page_zone(page);
2773 /* zone lock should be held when this function is called */
2774 lockdep_assert_held(&zone->lock);
2776 /* Return isolated page to tail of freelist. */
2777 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2778 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2782 * Update NUMA hit/miss statistics
2784 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2788 enum numa_stat_item local_stat = NUMA_LOCAL;
2790 /* skip numa counters update if numa stats is disabled */
2791 if (!static_branch_likely(&vm_numa_stat_key))
2794 if (zone_to_nid(z) != numa_node_id())
2795 local_stat = NUMA_OTHER;
2797 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2798 __count_numa_events(z, NUMA_HIT, nr_account);
2800 __count_numa_events(z, NUMA_MISS, nr_account);
2801 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2803 __count_numa_events(z, local_stat, nr_account);
2807 static __always_inline
2808 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2809 unsigned int order, unsigned int alloc_flags,
2813 unsigned long flags;
2817 spin_lock_irqsave(&zone->lock, flags);
2818 if (alloc_flags & ALLOC_HIGHATOMIC)
2819 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2821 page = __rmqueue(zone, order, migratetype, alloc_flags);
2824 * If the allocation fails, allow OOM handling access
2825 * to HIGHATOMIC reserves as failing now is worse than
2826 * failing a high-order atomic allocation in the
2829 if (!page && (alloc_flags & ALLOC_OOM))
2830 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2833 spin_unlock_irqrestore(&zone->lock, flags);
2837 spin_unlock_irqrestore(&zone->lock, flags);
2838 } while (check_new_pages(page, order));
2840 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2841 zone_statistics(preferred_zone, zone, 1);
2846 static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
2848 int high, base_batch, batch, max_nr_alloc;
2849 int high_max, high_min;
2851 base_batch = READ_ONCE(pcp->batch);
2852 high_min = READ_ONCE(pcp->high_min);
2853 high_max = READ_ONCE(pcp->high_max);
2854 high = pcp->high = clamp(pcp->high, high_min, high_max);
2856 /* Check for PCP disabled or boot pageset */
2857 if (unlikely(high < base_batch))
2863 batch = (base_batch << pcp->alloc_factor);
2866 * If we had larger pcp->high, we could avoid to allocate from
2869 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
2870 high = pcp->high = min(high + batch, high_max);
2873 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
2875 * Double the number of pages allocated each time there is
2876 * subsequent allocation of order-0 pages without any freeing.
2878 if (batch <= max_nr_alloc &&
2879 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
2880 pcp->alloc_factor++;
2881 batch = min(batch, max_nr_alloc);
2885 * Scale batch relative to order if batch implies free pages
2886 * can be stored on the PCP. Batch can be 1 for small zones or
2887 * for boot pagesets which should never store free pages as
2888 * the pages may belong to arbitrary zones.
2891 batch = max(batch >> order, 2);
2896 /* Remove page from the per-cpu list, caller must protect the list */
2898 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2900 unsigned int alloc_flags,
2901 struct per_cpu_pages *pcp,
2902 struct list_head *list)
2907 if (list_empty(list)) {
2908 int batch = nr_pcp_alloc(pcp, zone, order);
2911 alloced = rmqueue_bulk(zone, order,
2913 migratetype, alloc_flags);
2915 pcp->count += alloced << order;
2916 if (unlikely(list_empty(list)))
2920 page = list_first_entry(list, struct page, pcp_list);
2921 list_del(&page->pcp_list);
2922 pcp->count -= 1 << order;
2923 } while (check_new_pages(page, order));
2928 /* Lock and remove page from the per-cpu list */
2929 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2930 struct zone *zone, unsigned int order,
2931 int migratetype, unsigned int alloc_flags)
2933 struct per_cpu_pages *pcp;
2934 struct list_head *list;
2936 unsigned long __maybe_unused UP_flags;
2938 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2939 pcp_trylock_prepare(UP_flags);
2940 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2942 pcp_trylock_finish(UP_flags);
2947 * On allocation, reduce the number of pages that are batch freed.
2948 * See nr_pcp_free() where free_factor is increased for subsequent
2951 pcp->free_count >>= 1;
2952 list = &pcp->lists[order_to_pindex(migratetype, order)];
2953 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
2954 pcp_spin_unlock(pcp);
2955 pcp_trylock_finish(UP_flags);
2957 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2958 zone_statistics(preferred_zone, zone, 1);
2964 * Allocate a page from the given zone.
2965 * Use pcplists for THP or "cheap" high-order allocations.
2969 * Do not instrument rmqueue() with KMSAN. This function may call
2970 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
2971 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
2972 * may call rmqueue() again, which will result in a deadlock.
2974 __no_sanitize_memory
2976 struct page *rmqueue(struct zone *preferred_zone,
2977 struct zone *zone, unsigned int order,
2978 gfp_t gfp_flags, unsigned int alloc_flags,
2984 * We most definitely don't want callers attempting to
2985 * allocate greater than order-1 page units with __GFP_NOFAIL.
2987 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2989 if (likely(pcp_allowed_order(order))) {
2990 page = rmqueue_pcplist(preferred_zone, zone, order,
2991 migratetype, alloc_flags);
2996 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3000 /* Separate test+clear to avoid unnecessary atomics */
3001 if ((alloc_flags & ALLOC_KSWAPD) &&
3002 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3003 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3004 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3007 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3011 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3013 return __should_fail_alloc_page(gfp_mask, order);
3015 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3017 static inline long __zone_watermark_unusable_free(struct zone *z,
3018 unsigned int order, unsigned int alloc_flags)
3020 long unusable_free = (1 << order) - 1;
3023 * If the caller does not have rights to reserves below the min
3024 * watermark then subtract the high-atomic reserves. This will
3025 * over-estimate the size of the atomic reserve but it avoids a search.
3027 if (likely(!(alloc_flags & ALLOC_RESERVES)))
3028 unusable_free += z->nr_reserved_highatomic;
3031 /* If allocation can't use CMA areas don't use free CMA pages */
3032 if (!(alloc_flags & ALLOC_CMA))
3033 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3035 #ifdef CONFIG_UNACCEPTED_MEMORY
3036 unusable_free += zone_page_state(z, NR_UNACCEPTED);
3039 return unusable_free;
3043 * Return true if free base pages are above 'mark'. For high-order checks it
3044 * will return true of the order-0 watermark is reached and there is at least
3045 * one free page of a suitable size. Checking now avoids taking the zone lock
3046 * to check in the allocation paths if no pages are free.
3048 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3049 int highest_zoneidx, unsigned int alloc_flags,
3055 /* free_pages may go negative - that's OK */
3056 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3058 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
3060 * __GFP_HIGH allows access to 50% of the min reserve as well
3063 if (alloc_flags & ALLOC_MIN_RESERVE) {
3067 * Non-blocking allocations (e.g. GFP_ATOMIC) can
3068 * access more reserves than just __GFP_HIGH. Other
3069 * non-blocking allocations requests such as GFP_NOWAIT
3070 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
3071 * access to the min reserve.
3073 if (alloc_flags & ALLOC_NON_BLOCK)
3078 * OOM victims can try even harder than the normal reserve
3079 * users on the grounds that it's definitely going to be in
3080 * the exit path shortly and free memory. Any allocation it
3081 * makes during the free path will be small and short-lived.
3083 if (alloc_flags & ALLOC_OOM)
3088 * Check watermarks for an order-0 allocation request. If these
3089 * are not met, then a high-order request also cannot go ahead
3090 * even if a suitable page happened to be free.
3092 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3095 /* If this is an order-0 request then the watermark is fine */
3099 /* For a high-order request, check at least one suitable page is free */
3100 for (o = order; o < NR_PAGE_ORDERS; o++) {
3101 struct free_area *area = &z->free_area[o];
3107 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3108 if (!free_area_empty(area, mt))
3113 if ((alloc_flags & ALLOC_CMA) &&
3114 !free_area_empty(area, MIGRATE_CMA)) {
3118 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3119 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3126 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3127 int highest_zoneidx, unsigned int alloc_flags)
3129 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3130 zone_page_state(z, NR_FREE_PAGES));
3133 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3134 unsigned long mark, int highest_zoneidx,
3135 unsigned int alloc_flags, gfp_t gfp_mask)
3139 free_pages = zone_page_state(z, NR_FREE_PAGES);
3142 * Fast check for order-0 only. If this fails then the reserves
3143 * need to be calculated.
3149 usable_free = free_pages;
3150 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3152 /* reserved may over estimate high-atomic reserves. */
3153 usable_free -= min(usable_free, reserved);
3154 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3158 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3163 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3164 * when checking the min watermark. The min watermark is the
3165 * point where boosting is ignored so that kswapd is woken up
3166 * when below the low watermark.
3168 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3169 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3170 mark = z->_watermark[WMARK_MIN];
3171 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3172 alloc_flags, free_pages);
3178 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3179 unsigned long mark, int highest_zoneidx)
3181 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3183 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3184 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3186 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3191 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3193 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3195 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3196 node_reclaim_distance;
3198 #else /* CONFIG_NUMA */
3199 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3203 #endif /* CONFIG_NUMA */
3206 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3207 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3208 * premature use of a lower zone may cause lowmem pressure problems that
3209 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3210 * probably too small. It only makes sense to spread allocations to avoid
3211 * fragmentation between the Normal and DMA32 zones.
3213 static inline unsigned int
3214 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3216 unsigned int alloc_flags;
3219 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3222 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3224 #ifdef CONFIG_ZONE_DMA32
3228 if (zone_idx(zone) != ZONE_NORMAL)
3232 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3233 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3234 * on UMA that if Normal is populated then so is DMA32.
3236 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3237 if (nr_online_nodes > 1 && !populated_zone(--zone))
3240 alloc_flags |= ALLOC_NOFRAGMENT;
3241 #endif /* CONFIG_ZONE_DMA32 */
3245 /* Must be called after current_gfp_context() which can change gfp_mask */
3246 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3247 unsigned int alloc_flags)
3250 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3251 alloc_flags |= ALLOC_CMA;
3257 * get_page_from_freelist goes through the zonelist trying to allocate
3260 static struct page *
3261 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3262 const struct alloc_context *ac)
3266 struct pglist_data *last_pgdat = NULL;
3267 bool last_pgdat_dirty_ok = false;
3272 * Scan zonelist, looking for a zone with enough free.
3273 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3275 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3276 z = ac->preferred_zoneref;
3277 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3282 if (cpusets_enabled() &&
3283 (alloc_flags & ALLOC_CPUSET) &&
3284 !__cpuset_zone_allowed(zone, gfp_mask))
3287 * When allocating a page cache page for writing, we
3288 * want to get it from a node that is within its dirty
3289 * limit, such that no single node holds more than its
3290 * proportional share of globally allowed dirty pages.
3291 * The dirty limits take into account the node's
3292 * lowmem reserves and high watermark so that kswapd
3293 * should be able to balance it without having to
3294 * write pages from its LRU list.
3296 * XXX: For now, allow allocations to potentially
3297 * exceed the per-node dirty limit in the slowpath
3298 * (spread_dirty_pages unset) before going into reclaim,
3299 * which is important when on a NUMA setup the allowed
3300 * nodes are together not big enough to reach the
3301 * global limit. The proper fix for these situations
3302 * will require awareness of nodes in the
3303 * dirty-throttling and the flusher threads.
3305 if (ac->spread_dirty_pages) {
3306 if (last_pgdat != zone->zone_pgdat) {
3307 last_pgdat = zone->zone_pgdat;
3308 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3311 if (!last_pgdat_dirty_ok)
3315 if (no_fallback && nr_online_nodes > 1 &&
3316 zone != ac->preferred_zoneref->zone) {
3320 * If moving to a remote node, retry but allow
3321 * fragmenting fallbacks. Locality is more important
3322 * than fragmentation avoidance.
3324 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3325 if (zone_to_nid(zone) != local_nid) {
3326 alloc_flags &= ~ALLOC_NOFRAGMENT;
3332 * Detect whether the number of free pages is below high
3333 * watermark. If so, we will decrease pcp->high and free
3334 * PCP pages in free path to reduce the possibility of
3335 * premature page reclaiming. Detection is done here to
3336 * avoid to do that in hotter free path.
3338 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3339 goto check_alloc_wmark;
3341 mark = high_wmark_pages(zone);
3342 if (zone_watermark_fast(zone, order, mark,
3343 ac->highest_zoneidx, alloc_flags,
3347 set_bit(ZONE_BELOW_HIGH, &zone->flags);
3350 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3351 if (!zone_watermark_fast(zone, order, mark,
3352 ac->highest_zoneidx, alloc_flags,
3356 if (has_unaccepted_memory()) {
3357 if (try_to_accept_memory(zone, order))
3361 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3363 * Watermark failed for this zone, but see if we can
3364 * grow this zone if it contains deferred pages.
3366 if (deferred_pages_enabled()) {
3367 if (_deferred_grow_zone(zone, order))
3371 /* Checked here to keep the fast path fast */
3372 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3373 if (alloc_flags & ALLOC_NO_WATERMARKS)
3376 if (!node_reclaim_enabled() ||
3377 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3380 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3382 case NODE_RECLAIM_NOSCAN:
3385 case NODE_RECLAIM_FULL:
3386 /* scanned but unreclaimable */
3389 /* did we reclaim enough */
3390 if (zone_watermark_ok(zone, order, mark,
3391 ac->highest_zoneidx, alloc_flags))
3399 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3400 gfp_mask, alloc_flags, ac->migratetype);
3402 prep_new_page(page, order, gfp_mask, alloc_flags);
3405 * If this is a high-order atomic allocation then check
3406 * if the pageblock should be reserved for the future
3408 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3409 reserve_highatomic_pageblock(page, zone);
3413 if (has_unaccepted_memory()) {
3414 if (try_to_accept_memory(zone, order))
3418 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3419 /* Try again if zone has deferred pages */
3420 if (deferred_pages_enabled()) {
3421 if (_deferred_grow_zone(zone, order))
3429 * It's possible on a UMA machine to get through all zones that are
3430 * fragmented. If avoiding fragmentation, reset and try again.
3433 alloc_flags &= ~ALLOC_NOFRAGMENT;
3440 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3442 unsigned int filter = SHOW_MEM_FILTER_NODES;
3445 * This documents exceptions given to allocations in certain
3446 * contexts that are allowed to allocate outside current's set
3449 if (!(gfp_mask & __GFP_NOMEMALLOC))
3450 if (tsk_is_oom_victim(current) ||
3451 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3452 filter &= ~SHOW_MEM_FILTER_NODES;
3453 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3454 filter &= ~SHOW_MEM_FILTER_NODES;
3456 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3459 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3461 struct va_format vaf;
3463 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3465 if ((gfp_mask & __GFP_NOWARN) ||
3466 !__ratelimit(&nopage_rs) ||
3467 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3470 va_start(args, fmt);
3473 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3474 current->comm, &vaf, gfp_mask, &gfp_mask,
3475 nodemask_pr_args(nodemask));
3478 cpuset_print_current_mems_allowed();
3481 warn_alloc_show_mem(gfp_mask, nodemask);
3484 static inline struct page *
3485 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3486 unsigned int alloc_flags,
3487 const struct alloc_context *ac)
3491 page = get_page_from_freelist(gfp_mask, order,
3492 alloc_flags|ALLOC_CPUSET, ac);
3494 * fallback to ignore cpuset restriction if our nodes
3498 page = get_page_from_freelist(gfp_mask, order,
3504 static inline struct page *
3505 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3506 const struct alloc_context *ac, unsigned long *did_some_progress)
3508 struct oom_control oc = {
3509 .zonelist = ac->zonelist,
3510 .nodemask = ac->nodemask,
3512 .gfp_mask = gfp_mask,
3517 *did_some_progress = 0;
3520 * Acquire the oom lock. If that fails, somebody else is
3521 * making progress for us.
3523 if (!mutex_trylock(&oom_lock)) {
3524 *did_some_progress = 1;
3525 schedule_timeout_uninterruptible(1);
3530 * Go through the zonelist yet one more time, keep very high watermark
3531 * here, this is only to catch a parallel oom killing, we must fail if
3532 * we're still under heavy pressure. But make sure that this reclaim
3533 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3534 * allocation which will never fail due to oom_lock already held.
3536 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3537 ~__GFP_DIRECT_RECLAIM, order,
3538 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3542 /* Coredumps can quickly deplete all memory reserves */
3543 if (current->flags & PF_DUMPCORE)
3545 /* The OOM killer will not help higher order allocs */
3546 if (order > PAGE_ALLOC_COSTLY_ORDER)
3549 * We have already exhausted all our reclaim opportunities without any
3550 * success so it is time to admit defeat. We will skip the OOM killer
3551 * because it is very likely that the caller has a more reasonable
3552 * fallback than shooting a random task.
3554 * The OOM killer may not free memory on a specific node.
3556 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3558 /* The OOM killer does not needlessly kill tasks for lowmem */
3559 if (ac->highest_zoneidx < ZONE_NORMAL)
3561 if (pm_suspended_storage())
3564 * XXX: GFP_NOFS allocations should rather fail than rely on
3565 * other request to make a forward progress.
3566 * We are in an unfortunate situation where out_of_memory cannot
3567 * do much for this context but let's try it to at least get
3568 * access to memory reserved if the current task is killed (see
3569 * out_of_memory). Once filesystems are ready to handle allocation
3570 * failures more gracefully we should just bail out here.
3573 /* Exhausted what can be done so it's blame time */
3574 if (out_of_memory(&oc) ||
3575 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3576 *did_some_progress = 1;
3579 * Help non-failing allocations by giving them access to memory
3582 if (gfp_mask & __GFP_NOFAIL)
3583 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3584 ALLOC_NO_WATERMARKS, ac);
3587 mutex_unlock(&oom_lock);
3592 * Maximum number of compaction retries with a progress before OOM
3593 * killer is consider as the only way to move forward.
3595 #define MAX_COMPACT_RETRIES 16
3597 #ifdef CONFIG_COMPACTION
3598 /* Try memory compaction for high-order allocations before reclaim */
3599 static struct page *
3600 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3601 unsigned int alloc_flags, const struct alloc_context *ac,
3602 enum compact_priority prio, enum compact_result *compact_result)
3604 struct page *page = NULL;
3605 unsigned long pflags;
3606 unsigned int noreclaim_flag;
3611 psi_memstall_enter(&pflags);
3612 delayacct_compact_start();
3613 noreclaim_flag = memalloc_noreclaim_save();
3615 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3618 memalloc_noreclaim_restore(noreclaim_flag);
3619 psi_memstall_leave(&pflags);
3620 delayacct_compact_end();
3622 if (*compact_result == COMPACT_SKIPPED)
3625 * At least in one zone compaction wasn't deferred or skipped, so let's
3626 * count a compaction stall
3628 count_vm_event(COMPACTSTALL);
3630 /* Prep a captured page if available */
3632 prep_new_page(page, order, gfp_mask, alloc_flags);
3634 /* Try get a page from the freelist if available */
3636 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3639 struct zone *zone = page_zone(page);
3641 zone->compact_blockskip_flush = false;
3642 compaction_defer_reset(zone, order, true);
3643 count_vm_event(COMPACTSUCCESS);
3648 * It's bad if compaction run occurs and fails. The most likely reason
3649 * is that pages exist, but not enough to satisfy watermarks.
3651 count_vm_event(COMPACTFAIL);
3659 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3660 enum compact_result compact_result,
3661 enum compact_priority *compact_priority,
3662 int *compaction_retries)
3664 int max_retries = MAX_COMPACT_RETRIES;
3667 int retries = *compaction_retries;
3668 enum compact_priority priority = *compact_priority;
3673 if (fatal_signal_pending(current))
3677 * Compaction was skipped due to a lack of free order-0
3678 * migration targets. Continue if reclaim can help.
3680 if (compact_result == COMPACT_SKIPPED) {
3681 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3686 * Compaction managed to coalesce some page blocks, but the
3687 * allocation failed presumably due to a race. Retry some.
3689 if (compact_result == COMPACT_SUCCESS) {
3691 * !costly requests are much more important than
3692 * __GFP_RETRY_MAYFAIL costly ones because they are de
3693 * facto nofail and invoke OOM killer to move on while
3694 * costly can fail and users are ready to cope with
3695 * that. 1/4 retries is rather arbitrary but we would
3696 * need much more detailed feedback from compaction to
3697 * make a better decision.
3699 if (order > PAGE_ALLOC_COSTLY_ORDER)
3702 if (++(*compaction_retries) <= max_retries) {
3709 * Compaction failed. Retry with increasing priority.
3711 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3712 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3714 if (*compact_priority > min_priority) {
3715 (*compact_priority)--;
3716 *compaction_retries = 0;
3720 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3724 static inline struct page *
3725 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3726 unsigned int alloc_flags, const struct alloc_context *ac,
3727 enum compact_priority prio, enum compact_result *compact_result)
3729 *compact_result = COMPACT_SKIPPED;
3734 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3735 enum compact_result compact_result,
3736 enum compact_priority *compact_priority,
3737 int *compaction_retries)
3742 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3746 * There are setups with compaction disabled which would prefer to loop
3747 * inside the allocator rather than hit the oom killer prematurely.
3748 * Let's give them a good hope and keep retrying while the order-0
3749 * watermarks are OK.
3751 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3752 ac->highest_zoneidx, ac->nodemask) {
3753 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3754 ac->highest_zoneidx, alloc_flags))
3759 #endif /* CONFIG_COMPACTION */
3761 #ifdef CONFIG_LOCKDEP
3762 static struct lockdep_map __fs_reclaim_map =
3763 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3765 static bool __need_reclaim(gfp_t gfp_mask)
3767 /* no reclaim without waiting on it */
3768 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3771 /* this guy won't enter reclaim */
3772 if (current->flags & PF_MEMALLOC)
3775 if (gfp_mask & __GFP_NOLOCKDEP)
3781 void __fs_reclaim_acquire(unsigned long ip)
3783 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3786 void __fs_reclaim_release(unsigned long ip)
3788 lock_release(&__fs_reclaim_map, ip);
3791 void fs_reclaim_acquire(gfp_t gfp_mask)
3793 gfp_mask = current_gfp_context(gfp_mask);
3795 if (__need_reclaim(gfp_mask)) {
3796 if (gfp_mask & __GFP_FS)
3797 __fs_reclaim_acquire(_RET_IP_);
3799 #ifdef CONFIG_MMU_NOTIFIER
3800 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3801 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3806 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3808 void fs_reclaim_release(gfp_t gfp_mask)
3810 gfp_mask = current_gfp_context(gfp_mask);
3812 if (__need_reclaim(gfp_mask)) {
3813 if (gfp_mask & __GFP_FS)
3814 __fs_reclaim_release(_RET_IP_);
3817 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3821 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3822 * have been rebuilt so allocation retries. Reader side does not lock and
3823 * retries the allocation if zonelist changes. Writer side is protected by the
3824 * embedded spin_lock.
3826 static DEFINE_SEQLOCK(zonelist_update_seq);
3828 static unsigned int zonelist_iter_begin(void)
3830 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3831 return read_seqbegin(&zonelist_update_seq);
3836 static unsigned int check_retry_zonelist(unsigned int seq)
3838 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3839 return read_seqretry(&zonelist_update_seq, seq);
3844 /* Perform direct synchronous page reclaim */
3845 static unsigned long
3846 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3847 const struct alloc_context *ac)
3849 unsigned int noreclaim_flag;
3850 unsigned long progress;
3854 /* We now go into synchronous reclaim */
3855 cpuset_memory_pressure_bump();
3856 fs_reclaim_acquire(gfp_mask);
3857 noreclaim_flag = memalloc_noreclaim_save();
3859 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3862 memalloc_noreclaim_restore(noreclaim_flag);
3863 fs_reclaim_release(gfp_mask);
3870 /* The really slow allocator path where we enter direct reclaim */
3871 static inline struct page *
3872 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3873 unsigned int alloc_flags, const struct alloc_context *ac,
3874 unsigned long *did_some_progress)
3876 struct page *page = NULL;
3877 unsigned long pflags;
3878 bool drained = false;
3880 psi_memstall_enter(&pflags);
3881 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3882 if (unlikely(!(*did_some_progress)))
3886 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3889 * If an allocation failed after direct reclaim, it could be because
3890 * pages are pinned on the per-cpu lists or in high alloc reserves.
3891 * Shrink them and try again
3893 if (!page && !drained) {
3894 unreserve_highatomic_pageblock(ac, false);
3895 drain_all_pages(NULL);
3900 psi_memstall_leave(&pflags);
3905 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3906 const struct alloc_context *ac)
3910 pg_data_t *last_pgdat = NULL;
3911 enum zone_type highest_zoneidx = ac->highest_zoneidx;
3913 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3915 if (!managed_zone(zone))
3917 if (last_pgdat != zone->zone_pgdat) {
3918 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3919 last_pgdat = zone->zone_pgdat;
3924 static inline unsigned int
3925 gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3927 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3930 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3931 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3932 * to save two branches.
3934 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
3935 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
3938 * The caller may dip into page reserves a bit more if the caller
3939 * cannot run direct reclaim, or if the caller has realtime scheduling
3940 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3941 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3943 alloc_flags |= (__force int)
3944 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
3946 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
3948 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3949 * if it can't schedule.
3951 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
3952 alloc_flags |= ALLOC_NON_BLOCK;
3955 alloc_flags |= ALLOC_HIGHATOMIC;
3959 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
3960 * GFP_ATOMIC) rather than fail, see the comment for
3961 * cpuset_node_allowed().
3963 if (alloc_flags & ALLOC_MIN_RESERVE)
3964 alloc_flags &= ~ALLOC_CPUSET;
3965 } else if (unlikely(rt_task(current)) && in_task())
3966 alloc_flags |= ALLOC_MIN_RESERVE;
3968 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
3973 static bool oom_reserves_allowed(struct task_struct *tsk)
3975 if (!tsk_is_oom_victim(tsk))
3979 * !MMU doesn't have oom reaper so give access to memory reserves
3980 * only to the thread with TIF_MEMDIE set
3982 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3989 * Distinguish requests which really need access to full memory
3990 * reserves from oom victims which can live with a portion of it
3992 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3994 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3996 if (gfp_mask & __GFP_MEMALLOC)
3997 return ALLOC_NO_WATERMARKS;
3998 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3999 return ALLOC_NO_WATERMARKS;
4000 if (!in_interrupt()) {
4001 if (current->flags & PF_MEMALLOC)
4002 return ALLOC_NO_WATERMARKS;
4003 else if (oom_reserves_allowed(current))
4010 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4012 return !!__gfp_pfmemalloc_flags(gfp_mask);
4016 * Checks whether it makes sense to retry the reclaim to make a forward progress
4017 * for the given allocation request.
4019 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4020 * without success, or when we couldn't even meet the watermark if we
4021 * reclaimed all remaining pages on the LRU lists.
4023 * Returns true if a retry is viable or false to enter the oom path.
4026 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4027 struct alloc_context *ac, int alloc_flags,
4028 bool did_some_progress, int *no_progress_loops)
4035 * Costly allocations might have made a progress but this doesn't mean
4036 * their order will become available due to high fragmentation so
4037 * always increment the no progress counter for them
4039 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4040 *no_progress_loops = 0;
4042 (*no_progress_loops)++;
4044 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
4049 * Keep reclaiming pages while there is a chance this will lead
4050 * somewhere. If none of the target zones can satisfy our allocation
4051 * request even if all reclaimable pages are considered then we are
4052 * screwed and have to go OOM.
4054 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4055 ac->highest_zoneidx, ac->nodemask) {
4056 unsigned long available;
4057 unsigned long reclaimable;
4058 unsigned long min_wmark = min_wmark_pages(zone);
4061 available = reclaimable = zone_reclaimable_pages(zone);
4062 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4065 * Would the allocation succeed if we reclaimed all
4066 * reclaimable pages?
4068 wmark = __zone_watermark_ok(zone, order, min_wmark,
4069 ac->highest_zoneidx, alloc_flags, available);
4070 trace_reclaim_retry_zone(z, order, reclaimable,
4071 available, min_wmark, *no_progress_loops, wmark);
4079 * Memory allocation/reclaim might be called from a WQ context and the
4080 * current implementation of the WQ concurrency control doesn't
4081 * recognize that a particular WQ is congested if the worker thread is
4082 * looping without ever sleeping. Therefore we have to do a short sleep
4083 * here rather than calling cond_resched().
4085 if (current->flags & PF_WQ_WORKER)
4086 schedule_timeout_uninterruptible(1);
4090 /* Before OOM, exhaust highatomic_reserve */
4092 return unreserve_highatomic_pageblock(ac, true);
4098 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4101 * It's possible that cpuset's mems_allowed and the nodemask from
4102 * mempolicy don't intersect. This should be normally dealt with by
4103 * policy_nodemask(), but it's possible to race with cpuset update in
4104 * such a way the check therein was true, and then it became false
4105 * before we got our cpuset_mems_cookie here.
4106 * This assumes that for all allocations, ac->nodemask can come only
4107 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4108 * when it does not intersect with the cpuset restrictions) or the
4109 * caller can deal with a violated nodemask.
4111 if (cpusets_enabled() && ac->nodemask &&
4112 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4113 ac->nodemask = NULL;
4118 * When updating a task's mems_allowed or mempolicy nodemask, it is
4119 * possible to race with parallel threads in such a way that our
4120 * allocation can fail while the mask is being updated. If we are about
4121 * to fail, check if the cpuset changed during allocation and if so,
4124 if (read_mems_allowed_retry(cpuset_mems_cookie))
4130 static inline struct page *
4131 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4132 struct alloc_context *ac)
4134 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4135 bool can_compact = gfp_compaction_allowed(gfp_mask);
4136 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4137 struct page *page = NULL;
4138 unsigned int alloc_flags;
4139 unsigned long did_some_progress;
4140 enum compact_priority compact_priority;
4141 enum compact_result compact_result;
4142 int compaction_retries;
4143 int no_progress_loops;
4144 unsigned int cpuset_mems_cookie;
4145 unsigned int zonelist_iter_cookie;
4149 compaction_retries = 0;
4150 no_progress_loops = 0;
4151 compact_priority = DEF_COMPACT_PRIORITY;
4152 cpuset_mems_cookie = read_mems_allowed_begin();
4153 zonelist_iter_cookie = zonelist_iter_begin();
4156 * The fast path uses conservative alloc_flags to succeed only until
4157 * kswapd needs to be woken up, and to avoid the cost of setting up
4158 * alloc_flags precisely. So we do that now.
4160 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4163 * We need to recalculate the starting point for the zonelist iterator
4164 * because we might have used different nodemask in the fast path, or
4165 * there was a cpuset modification and we are retrying - otherwise we
4166 * could end up iterating over non-eligible zones endlessly.
4168 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4169 ac->highest_zoneidx, ac->nodemask);
4170 if (!ac->preferred_zoneref->zone)
4174 * Check for insane configurations where the cpuset doesn't contain
4175 * any suitable zone to satisfy the request - e.g. non-movable
4176 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4178 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4179 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4180 ac->highest_zoneidx,
4181 &cpuset_current_mems_allowed);
4186 if (alloc_flags & ALLOC_KSWAPD)
4187 wake_all_kswapds(order, gfp_mask, ac);
4190 * The adjusted alloc_flags might result in immediate success, so try
4193 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4198 * For costly allocations, try direct compaction first, as it's likely
4199 * that we have enough base pages and don't need to reclaim. For non-
4200 * movable high-order allocations, do that as well, as compaction will
4201 * try prevent permanent fragmentation by migrating from blocks of the
4203 * Don't try this for allocations that are allowed to ignore
4204 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4206 if (can_direct_reclaim && can_compact &&
4208 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4209 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4210 page = __alloc_pages_direct_compact(gfp_mask, order,
4212 INIT_COMPACT_PRIORITY,
4218 * Checks for costly allocations with __GFP_NORETRY, which
4219 * includes some THP page fault allocations
4221 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4223 * If allocating entire pageblock(s) and compaction
4224 * failed because all zones are below low watermarks
4225 * or is prohibited because it recently failed at this
4226 * order, fail immediately unless the allocator has
4227 * requested compaction and reclaim retry.
4230 * - potentially very expensive because zones are far
4231 * below their low watermarks or this is part of very
4232 * bursty high order allocations,
4233 * - not guaranteed to help because isolate_freepages()
4234 * may not iterate over freed pages as part of its
4236 * - unlikely to make entire pageblocks free on its
4239 if (compact_result == COMPACT_SKIPPED ||
4240 compact_result == COMPACT_DEFERRED)
4244 * Looks like reclaim/compaction is worth trying, but
4245 * sync compaction could be very expensive, so keep
4246 * using async compaction.
4248 compact_priority = INIT_COMPACT_PRIORITY;
4253 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4254 if (alloc_flags & ALLOC_KSWAPD)
4255 wake_all_kswapds(order, gfp_mask, ac);
4257 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4259 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4260 (alloc_flags & ALLOC_KSWAPD);
4263 * Reset the nodemask and zonelist iterators if memory policies can be
4264 * ignored. These allocations are high priority and system rather than
4267 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4268 ac->nodemask = NULL;
4269 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4270 ac->highest_zoneidx, ac->nodemask);
4273 /* Attempt with potentially adjusted zonelist and alloc_flags */
4274 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4278 /* Caller is not willing to reclaim, we can't balance anything */
4279 if (!can_direct_reclaim)
4282 /* Avoid recursion of direct reclaim */
4283 if (current->flags & PF_MEMALLOC)
4286 /* Try direct reclaim and then allocating */
4287 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4288 &did_some_progress);
4292 /* Try direct compaction and then allocating */
4293 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4294 compact_priority, &compact_result);
4298 /* Do not loop if specifically requested */
4299 if (gfp_mask & __GFP_NORETRY)
4303 * Do not retry costly high order allocations unless they are
4304 * __GFP_RETRY_MAYFAIL and we can compact
4306 if (costly_order && (!can_compact ||
4307 !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4310 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4311 did_some_progress > 0, &no_progress_loops))
4315 * It doesn't make any sense to retry for the compaction if the order-0
4316 * reclaim is not able to make any progress because the current
4317 * implementation of the compaction depends on the sufficient amount
4318 * of free memory (see __compaction_suitable)
4320 if (did_some_progress > 0 && can_compact &&
4321 should_compact_retry(ac, order, alloc_flags,
4322 compact_result, &compact_priority,
4323 &compaction_retries))
4328 * Deal with possible cpuset update races or zonelist updates to avoid
4329 * a unnecessary OOM kill.
4331 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4332 check_retry_zonelist(zonelist_iter_cookie))
4335 /* Reclaim has failed us, start killing things */
4336 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4340 /* Avoid allocations with no watermarks from looping endlessly */
4341 if (tsk_is_oom_victim(current) &&
4342 (alloc_flags & ALLOC_OOM ||
4343 (gfp_mask & __GFP_NOMEMALLOC)))
4346 /* Retry as long as the OOM killer is making progress */
4347 if (did_some_progress) {
4348 no_progress_loops = 0;
4354 * Deal with possible cpuset update races or zonelist updates to avoid
4355 * a unnecessary OOM kill.
4357 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4358 check_retry_zonelist(zonelist_iter_cookie))
4362 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4365 if (gfp_mask & __GFP_NOFAIL) {
4367 * All existing users of the __GFP_NOFAIL are blockable, so warn
4368 * of any new users that actually require GFP_NOWAIT
4370 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4374 * PF_MEMALLOC request from this context is rather bizarre
4375 * because we cannot reclaim anything and only can loop waiting
4376 * for somebody to do a work for us
4378 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4381 * non failing costly orders are a hard requirement which we
4382 * are not prepared for much so let's warn about these users
4383 * so that we can identify them and convert them to something
4386 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4389 * Help non-failing allocations by giving some access to memory
4390 * reserves normally used for high priority non-blocking
4391 * allocations but do not use ALLOC_NO_WATERMARKS because this
4392 * could deplete whole memory reserves which would just make
4393 * the situation worse.
4395 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4403 warn_alloc(gfp_mask, ac->nodemask,
4404 "page allocation failure: order:%u", order);
4409 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4410 int preferred_nid, nodemask_t *nodemask,
4411 struct alloc_context *ac, gfp_t *alloc_gfp,
4412 unsigned int *alloc_flags)
4414 ac->highest_zoneidx = gfp_zone(gfp_mask);
4415 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4416 ac->nodemask = nodemask;
4417 ac->migratetype = gfp_migratetype(gfp_mask);
4419 if (cpusets_enabled()) {
4420 *alloc_gfp |= __GFP_HARDWALL;
4422 * When we are in the interrupt context, it is irrelevant
4423 * to the current task context. It means that any node ok.
4425 if (in_task() && !ac->nodemask)
4426 ac->nodemask = &cpuset_current_mems_allowed;
4428 *alloc_flags |= ALLOC_CPUSET;
4431 might_alloc(gfp_mask);
4433 if (should_fail_alloc_page(gfp_mask, order))
4436 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4438 /* Dirty zone balancing only done in the fast path */
4439 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4442 * The preferred zone is used for statistics but crucially it is
4443 * also used as the starting point for the zonelist iterator. It
4444 * may get reset for allocations that ignore memory policies.
4446 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4447 ac->highest_zoneidx, ac->nodemask);
4453 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4454 * @gfp: GFP flags for the allocation
4455 * @preferred_nid: The preferred NUMA node ID to allocate from
4456 * @nodemask: Set of nodes to allocate from, may be NULL
4457 * @nr_pages: The number of pages desired on the list or array
4458 * @page_list: Optional list to store the allocated pages
4459 * @page_array: Optional array to store the pages
4461 * This is a batched version of the page allocator that attempts to
4462 * allocate nr_pages quickly. Pages are added to page_list if page_list
4463 * is not NULL, otherwise it is assumed that the page_array is valid.
4465 * For lists, nr_pages is the number of pages that should be allocated.
4467 * For arrays, only NULL elements are populated with pages and nr_pages
4468 * is the maximum number of pages that will be stored in the array.
4470 * Returns the number of pages on the list or array.
4472 unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
4473 nodemask_t *nodemask, int nr_pages,
4474 struct list_head *page_list,
4475 struct page **page_array)
4478 unsigned long __maybe_unused UP_flags;
4481 struct per_cpu_pages *pcp;
4482 struct list_head *pcp_list;
4483 struct alloc_context ac;
4485 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4486 int nr_populated = 0, nr_account = 0;
4489 * Skip populated array elements to determine if any pages need
4490 * to be allocated before disabling IRQs.
4492 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4495 /* No pages requested? */
4496 if (unlikely(nr_pages <= 0))
4499 /* Already populated array? */
4500 if (unlikely(page_array && nr_pages - nr_populated == 0))
4503 /* Bulk allocator does not support memcg accounting. */
4504 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4507 /* Use the single page allocator for one page. */
4508 if (nr_pages - nr_populated == 1)
4511 #ifdef CONFIG_PAGE_OWNER
4513 * PAGE_OWNER may recurse into the allocator to allocate space to
4514 * save the stack with pagesets.lock held. Releasing/reacquiring
4515 * removes much of the performance benefit of bulk allocation so
4516 * force the caller to allocate one page at a time as it'll have
4517 * similar performance to added complexity to the bulk allocator.
4519 if (static_branch_unlikely(&page_owner_inited))
4523 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4524 gfp &= gfp_allowed_mask;
4526 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4530 /* Find an allowed local zone that meets the low watermark. */
4531 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4534 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4535 !__cpuset_zone_allowed(zone, gfp)) {
4539 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4540 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4544 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4545 if (zone_watermark_fast(zone, 0, mark,
4546 zonelist_zone_idx(ac.preferred_zoneref),
4547 alloc_flags, gfp)) {
4553 * If there are no allowed local zones that meets the watermarks then
4554 * try to allocate a single page and reclaim if necessary.
4556 if (unlikely(!zone))
4559 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4560 pcp_trylock_prepare(UP_flags);
4561 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4565 /* Attempt the batch allocation */
4566 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4567 while (nr_populated < nr_pages) {
4569 /* Skip existing pages */
4570 if (page_array && page_array[nr_populated]) {
4575 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4577 if (unlikely(!page)) {
4578 /* Try and allocate at least one page */
4580 pcp_spin_unlock(pcp);
4587 prep_new_page(page, 0, gfp, 0);
4589 list_add(&page->lru, page_list);
4591 page_array[nr_populated] = page;
4595 pcp_spin_unlock(pcp);
4596 pcp_trylock_finish(UP_flags);
4598 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4599 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4602 return nr_populated;
4605 pcp_trylock_finish(UP_flags);
4608 page = __alloc_pages_noprof(gfp, 0, preferred_nid, nodemask);
4611 list_add(&page->lru, page_list);
4613 page_array[nr_populated] = page;
4619 EXPORT_SYMBOL_GPL(alloc_pages_bulk_noprof);
4622 * This is the 'heart' of the zoned buddy allocator.
4624 struct page *__alloc_pages_noprof(gfp_t gfp, unsigned int order,
4625 int preferred_nid, nodemask_t *nodemask)
4628 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4629 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4630 struct alloc_context ac = { };
4633 * There are several places where we assume that the order value is sane
4634 * so bail out early if the request is out of bound.
4636 if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
4639 gfp &= gfp_allowed_mask;
4641 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4642 * resp. GFP_NOIO which has to be inherited for all allocation requests
4643 * from a particular context which has been marked by
4644 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4645 * movable zones are not used during allocation.
4647 gfp = current_gfp_context(gfp);
4649 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4650 &alloc_gfp, &alloc_flags))
4654 * Forbid the first pass from falling back to types that fragment
4655 * memory until all local zones are considered.
4657 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4659 /* First allocation attempt */
4660 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4665 ac.spread_dirty_pages = false;
4668 * Restore the original nodemask if it was potentially replaced with
4669 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4671 ac.nodemask = nodemask;
4673 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4676 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4677 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4678 __free_pages(page, order);
4682 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4683 kmsan_alloc_page(page, order, alloc_gfp);
4687 EXPORT_SYMBOL(__alloc_pages_noprof);
4689 struct folio *__folio_alloc_noprof(gfp_t gfp, unsigned int order, int preferred_nid,
4690 nodemask_t *nodemask)
4692 struct page *page = __alloc_pages_noprof(gfp | __GFP_COMP, order,
4693 preferred_nid, nodemask);
4694 return page_rmappable_folio(page);
4696 EXPORT_SYMBOL(__folio_alloc_noprof);
4699 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4700 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4701 * you need to access high mem.
4703 unsigned long get_free_pages_noprof(gfp_t gfp_mask, unsigned int order)
4707 page = alloc_pages_noprof(gfp_mask & ~__GFP_HIGHMEM, order);
4710 return (unsigned long) page_address(page);
4712 EXPORT_SYMBOL(get_free_pages_noprof);
4714 unsigned long get_zeroed_page_noprof(gfp_t gfp_mask)
4716 return get_free_pages_noprof(gfp_mask | __GFP_ZERO, 0);
4718 EXPORT_SYMBOL(get_zeroed_page_noprof);
4721 * __free_pages - Free pages allocated with alloc_pages().
4722 * @page: The page pointer returned from alloc_pages().
4723 * @order: The order of the allocation.
4725 * This function can free multi-page allocations that are not compound
4726 * pages. It does not check that the @order passed in matches that of
4727 * the allocation, so it is easy to leak memory. Freeing more memory
4728 * than was allocated will probably emit a warning.
4730 * If the last reference to this page is speculative, it will be released
4731 * by put_page() which only frees the first page of a non-compound
4732 * allocation. To prevent the remaining pages from being leaked, we free
4733 * the subsequent pages here. If you want to use the page's reference
4734 * count to decide when to free the allocation, you should allocate a
4735 * compound page, and use put_page() instead of __free_pages().
4737 * Context: May be called in interrupt context or while holding a normal
4738 * spinlock, but not in NMI context or while holding a raw spinlock.
4740 void __free_pages(struct page *page, unsigned int order)
4742 /* get PageHead before we drop reference */
4743 int head = PageHead(page);
4744 struct alloc_tag *tag = pgalloc_tag_get(page);
4746 if (put_page_testzero(page))
4747 free_unref_page(page, order);
4749 pgalloc_tag_sub_pages(tag, (1 << order) - 1);
4751 free_unref_page(page + (1 << order), order);
4754 EXPORT_SYMBOL(__free_pages);
4756 void free_pages(unsigned long addr, unsigned int order)
4759 VM_BUG_ON(!virt_addr_valid((void *)addr));
4760 __free_pages(virt_to_page((void *)addr), order);
4764 EXPORT_SYMBOL(free_pages);
4768 * An arbitrary-length arbitrary-offset area of memory which resides
4769 * within a 0 or higher order page. Multiple fragments within that page
4770 * are individually refcounted, in the page's reference counter.
4772 * The page_frag functions below provide a simple allocation framework for
4773 * page fragments. This is used by the network stack and network device
4774 * drivers to provide a backing region of memory for use as either an
4775 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4777 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4780 struct page *page = NULL;
4781 gfp_t gfp = gfp_mask;
4783 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4784 gfp_mask = (gfp_mask & ~__GFP_DIRECT_RECLAIM) | __GFP_COMP |
4785 __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC;
4786 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4787 PAGE_FRAG_CACHE_MAX_ORDER);
4788 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4790 if (unlikely(!page))
4791 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4793 nc->va = page ? page_address(page) : NULL;
4798 void page_frag_cache_drain(struct page_frag_cache *nc)
4803 __page_frag_cache_drain(virt_to_head_page(nc->va), nc->pagecnt_bias);
4806 EXPORT_SYMBOL(page_frag_cache_drain);
4808 void __page_frag_cache_drain(struct page *page, unsigned int count)
4810 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4812 if (page_ref_sub_and_test(page, count))
4813 free_unref_page(page, compound_order(page));
4815 EXPORT_SYMBOL(__page_frag_cache_drain);
4817 void *__page_frag_alloc_align(struct page_frag_cache *nc,
4818 unsigned int fragsz, gfp_t gfp_mask,
4819 unsigned int align_mask)
4821 unsigned int size = PAGE_SIZE;
4825 if (unlikely(!nc->va)) {
4827 page = __page_frag_cache_refill(nc, gfp_mask);
4831 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4832 /* if size can vary use size else just use PAGE_SIZE */
4835 /* Even if we own the page, we do not use atomic_set().
4836 * This would break get_page_unless_zero() users.
4838 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4840 /* reset page count bias and offset to start of new frag */
4841 nc->pfmemalloc = page_is_pfmemalloc(page);
4842 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4846 offset = nc->offset - fragsz;
4847 if (unlikely(offset < 0)) {
4848 page = virt_to_page(nc->va);
4850 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4853 if (unlikely(nc->pfmemalloc)) {
4854 free_unref_page(page, compound_order(page));
4858 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4859 /* if size can vary use size else just use PAGE_SIZE */
4862 /* OK, page count is 0, we can safely set it */
4863 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4865 /* reset page count bias and offset to start of new frag */
4866 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4867 offset = size - fragsz;
4868 if (unlikely(offset < 0)) {
4870 * The caller is trying to allocate a fragment
4871 * with fragsz > PAGE_SIZE but the cache isn't big
4872 * enough to satisfy the request, this may
4873 * happen in low memory conditions.
4874 * We don't release the cache page because
4875 * it could make memory pressure worse
4876 * so we simply return NULL here.
4883 offset &= align_mask;
4884 nc->offset = offset;
4886 return nc->va + offset;
4888 EXPORT_SYMBOL(__page_frag_alloc_align);
4891 * Frees a page fragment allocated out of either a compound or order 0 page.
4893 void page_frag_free(void *addr)
4895 struct page *page = virt_to_head_page(addr);
4897 if (unlikely(put_page_testzero(page)))
4898 free_unref_page(page, compound_order(page));
4900 EXPORT_SYMBOL(page_frag_free);
4902 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4906 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4907 struct page *page = virt_to_page((void *)addr);
4908 struct page *last = page + nr;
4910 split_page_owner(page, order, 0);
4911 pgalloc_tag_split(page, 1 << order);
4912 split_page_memcg(page, order, 0);
4913 while (page < --last)
4914 set_page_refcounted(last);
4916 last = page + (1UL << order);
4917 for (page += nr; page < last; page++)
4918 __free_pages_ok(page, 0, FPI_TO_TAIL);
4920 return (void *)addr;
4924 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4925 * @size: the number of bytes to allocate
4926 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4928 * This function is similar to alloc_pages(), except that it allocates the
4929 * minimum number of pages to satisfy the request. alloc_pages() can only
4930 * allocate memory in power-of-two pages.
4932 * This function is also limited by MAX_PAGE_ORDER.
4934 * Memory allocated by this function must be released by free_pages_exact().
4936 * Return: pointer to the allocated area or %NULL in case of error.
4938 void *alloc_pages_exact_noprof(size_t size, gfp_t gfp_mask)
4940 unsigned int order = get_order(size);
4943 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4944 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4946 addr = get_free_pages_noprof(gfp_mask, order);
4947 return make_alloc_exact(addr, order, size);
4949 EXPORT_SYMBOL(alloc_pages_exact_noprof);
4952 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4954 * @nid: the preferred node ID where memory should be allocated
4955 * @size: the number of bytes to allocate
4956 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4958 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4961 * Return: pointer to the allocated area or %NULL in case of error.
4963 void * __meminit alloc_pages_exact_nid_noprof(int nid, size_t size, gfp_t gfp_mask)
4965 unsigned int order = get_order(size);
4968 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4969 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4971 p = alloc_pages_node_noprof(nid, gfp_mask, order);
4974 return make_alloc_exact((unsigned long)page_address(p), order, size);
4978 * free_pages_exact - release memory allocated via alloc_pages_exact()
4979 * @virt: the value returned by alloc_pages_exact.
4980 * @size: size of allocation, same value as passed to alloc_pages_exact().
4982 * Release the memory allocated by a previous call to alloc_pages_exact.
4984 void free_pages_exact(void *virt, size_t size)
4986 unsigned long addr = (unsigned long)virt;
4987 unsigned long end = addr + PAGE_ALIGN(size);
4989 while (addr < end) {
4994 EXPORT_SYMBOL(free_pages_exact);
4997 * nr_free_zone_pages - count number of pages beyond high watermark
4998 * @offset: The zone index of the highest zone
5000 * nr_free_zone_pages() counts the number of pages which are beyond the
5001 * high watermark within all zones at or below a given zone index. For each
5002 * zone, the number of pages is calculated as:
5004 * nr_free_zone_pages = managed_pages - high_pages
5006 * Return: number of pages beyond high watermark.
5008 static unsigned long nr_free_zone_pages(int offset)
5013 /* Just pick one node, since fallback list is circular */
5014 unsigned long sum = 0;
5016 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5018 for_each_zone_zonelist(zone, z, zonelist, offset) {
5019 unsigned long size = zone_managed_pages(zone);
5020 unsigned long high = high_wmark_pages(zone);
5029 * nr_free_buffer_pages - count number of pages beyond high watermark
5031 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5032 * watermark within ZONE_DMA and ZONE_NORMAL.
5034 * Return: number of pages beyond high watermark within ZONE_DMA and
5037 unsigned long nr_free_buffer_pages(void)
5039 return nr_free_zone_pages(gfp_zone(GFP_USER));
5041 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5043 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5045 zoneref->zone = zone;
5046 zoneref->zone_idx = zone_idx(zone);
5050 * Builds allocation fallback zone lists.
5052 * Add all populated zones of a node to the zonelist.
5054 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5057 enum zone_type zone_type = MAX_NR_ZONES;
5062 zone = pgdat->node_zones + zone_type;
5063 if (populated_zone(zone)) {
5064 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5065 check_highest_zone(zone_type);
5067 } while (zone_type);
5074 static int __parse_numa_zonelist_order(char *s)
5077 * We used to support different zonelists modes but they turned
5078 * out to be just not useful. Let's keep the warning in place
5079 * if somebody still use the cmd line parameter so that we do
5080 * not fail it silently
5082 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5083 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5089 static char numa_zonelist_order[] = "Node";
5090 #define NUMA_ZONELIST_ORDER_LEN 16
5092 * sysctl handler for numa_zonelist_order
5094 static int numa_zonelist_order_handler(struct ctl_table *table, int write,
5095 void *buffer, size_t *length, loff_t *ppos)
5098 return __parse_numa_zonelist_order(buffer);
5099 return proc_dostring(table, write, buffer, length, ppos);
5102 static int node_load[MAX_NUMNODES];
5105 * find_next_best_node - find the next node that should appear in a given node's fallback list
5106 * @node: node whose fallback list we're appending
5107 * @used_node_mask: nodemask_t of already used nodes
5109 * We use a number of factors to determine which is the next node that should
5110 * appear on a given node's fallback list. The node should not have appeared
5111 * already in @node's fallback list, and it should be the next closest node
5112 * according to the distance array (which contains arbitrary distance values
5113 * from each node to each node in the system), and should also prefer nodes
5114 * with no CPUs, since presumably they'll have very little allocation pressure
5115 * on them otherwise.
5117 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5119 int find_next_best_node(int node, nodemask_t *used_node_mask)
5122 int min_val = INT_MAX;
5123 int best_node = NUMA_NO_NODE;
5126 * Use the local node if we haven't already, but for memoryless local
5127 * node, we should skip it and fall back to other nodes.
5129 if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5130 node_set(node, *used_node_mask);
5134 for_each_node_state(n, N_MEMORY) {
5136 /* Don't want a node to appear more than once */
5137 if (node_isset(n, *used_node_mask))
5140 /* Use the distance array to find the distance */
5141 val = node_distance(node, n);
5143 /* Penalize nodes under us ("prefer the next node") */
5146 /* Give preference to headless and unused nodes */
5147 if (!cpumask_empty(cpumask_of_node(n)))
5148 val += PENALTY_FOR_NODE_WITH_CPUS;
5150 /* Slight preference for less loaded node */
5151 val *= MAX_NUMNODES;
5152 val += node_load[n];
5154 if (val < min_val) {
5161 node_set(best_node, *used_node_mask);
5168 * Build zonelists ordered by node and zones within node.
5169 * This results in maximum locality--normal zone overflows into local
5170 * DMA zone, if any--but risks exhausting DMA zone.
5172 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5175 struct zoneref *zonerefs;
5178 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5180 for (i = 0; i < nr_nodes; i++) {
5183 pg_data_t *node = NODE_DATA(node_order[i]);
5185 nr_zones = build_zonerefs_node(node, zonerefs);
5186 zonerefs += nr_zones;
5188 zonerefs->zone = NULL;
5189 zonerefs->zone_idx = 0;
5193 * Build gfp_thisnode zonelists
5195 static void build_thisnode_zonelists(pg_data_t *pgdat)
5197 struct zoneref *zonerefs;
5200 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5201 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5202 zonerefs += nr_zones;
5203 zonerefs->zone = NULL;
5204 zonerefs->zone_idx = 0;
5208 * Build zonelists ordered by zone and nodes within zones.
5209 * This results in conserving DMA zone[s] until all Normal memory is
5210 * exhausted, but results in overflowing to remote node while memory
5211 * may still exist in local DMA zone.
5214 static void build_zonelists(pg_data_t *pgdat)
5216 static int node_order[MAX_NUMNODES];
5217 int node, nr_nodes = 0;
5218 nodemask_t used_mask = NODE_MASK_NONE;
5219 int local_node, prev_node;
5221 /* NUMA-aware ordering of nodes */
5222 local_node = pgdat->node_id;
5223 prev_node = local_node;
5225 memset(node_order, 0, sizeof(node_order));
5226 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5228 * We don't want to pressure a particular node.
5229 * So adding penalty to the first node in same
5230 * distance group to make it round-robin.
5232 if (node_distance(local_node, node) !=
5233 node_distance(local_node, prev_node))
5234 node_load[node] += 1;
5236 node_order[nr_nodes++] = node;
5240 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5241 build_thisnode_zonelists(pgdat);
5242 pr_info("Fallback order for Node %d: ", local_node);
5243 for (node = 0; node < nr_nodes; node++)
5244 pr_cont("%d ", node_order[node]);
5248 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5250 * Return node id of node used for "local" allocations.
5251 * I.e., first node id of first zone in arg node's generic zonelist.
5252 * Used for initializing percpu 'numa_mem', which is used primarily
5253 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5255 int local_memory_node(int node)
5259 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5260 gfp_zone(GFP_KERNEL),
5262 return zone_to_nid(z->zone);
5266 static void setup_min_unmapped_ratio(void);
5267 static void setup_min_slab_ratio(void);
5268 #else /* CONFIG_NUMA */
5270 static void build_zonelists(pg_data_t *pgdat)
5272 struct zoneref *zonerefs;
5275 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5276 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5277 zonerefs += nr_zones;
5279 zonerefs->zone = NULL;
5280 zonerefs->zone_idx = 0;
5283 #endif /* CONFIG_NUMA */
5286 * Boot pageset table. One per cpu which is going to be used for all
5287 * zones and all nodes. The parameters will be set in such a way
5288 * that an item put on a list will immediately be handed over to
5289 * the buddy list. This is safe since pageset manipulation is done
5290 * with interrupts disabled.
5292 * The boot_pagesets must be kept even after bootup is complete for
5293 * unused processors and/or zones. They do play a role for bootstrapping
5294 * hotplugged processors.
5296 * zoneinfo_show() and maybe other functions do
5297 * not check if the processor is online before following the pageset pointer.
5298 * Other parts of the kernel may not check if the zone is available.
5300 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5301 /* These effectively disable the pcplists in the boot pageset completely */
5302 #define BOOT_PAGESET_HIGH 0
5303 #define BOOT_PAGESET_BATCH 1
5304 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5305 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5307 static void __build_all_zonelists(void *data)
5310 int __maybe_unused cpu;
5311 pg_data_t *self = data;
5312 unsigned long flags;
5315 * The zonelist_update_seq must be acquired with irqsave because the
5316 * reader can be invoked from IRQ with GFP_ATOMIC.
5318 write_seqlock_irqsave(&zonelist_update_seq, flags);
5320 * Also disable synchronous printk() to prevent any printk() from
5321 * trying to hold port->lock, for
5322 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5323 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5325 printk_deferred_enter();
5328 memset(node_load, 0, sizeof(node_load));
5332 * This node is hotadded and no memory is yet present. So just
5333 * building zonelists is fine - no need to touch other nodes.
5335 if (self && !node_online(self->node_id)) {
5336 build_zonelists(self);
5339 * All possible nodes have pgdat preallocated
5342 for_each_node(nid) {
5343 pg_data_t *pgdat = NODE_DATA(nid);
5345 build_zonelists(pgdat);
5348 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5350 * We now know the "local memory node" for each node--
5351 * i.e., the node of the first zone in the generic zonelist.
5352 * Set up numa_mem percpu variable for on-line cpus. During
5353 * boot, only the boot cpu should be on-line; we'll init the
5354 * secondary cpus' numa_mem as they come on-line. During
5355 * node/memory hotplug, we'll fixup all on-line cpus.
5357 for_each_online_cpu(cpu)
5358 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5362 printk_deferred_exit();
5363 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5366 static noinline void __init
5367 build_all_zonelists_init(void)
5371 __build_all_zonelists(NULL);
5374 * Initialize the boot_pagesets that are going to be used
5375 * for bootstrapping processors. The real pagesets for
5376 * each zone will be allocated later when the per cpu
5377 * allocator is available.
5379 * boot_pagesets are used also for bootstrapping offline
5380 * cpus if the system is already booted because the pagesets
5381 * are needed to initialize allocators on a specific cpu too.
5382 * F.e. the percpu allocator needs the page allocator which
5383 * needs the percpu allocator in order to allocate its pagesets
5384 * (a chicken-egg dilemma).
5386 for_each_possible_cpu(cpu)
5387 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5389 mminit_verify_zonelist();
5390 cpuset_init_current_mems_allowed();
5394 * unless system_state == SYSTEM_BOOTING.
5396 * __ref due to call of __init annotated helper build_all_zonelists_init
5397 * [protected by SYSTEM_BOOTING].
5399 void __ref build_all_zonelists(pg_data_t *pgdat)
5401 unsigned long vm_total_pages;
5403 if (system_state == SYSTEM_BOOTING) {
5404 build_all_zonelists_init();
5406 __build_all_zonelists(pgdat);
5407 /* cpuset refresh routine should be here */
5409 /* Get the number of free pages beyond high watermark in all zones. */
5410 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5412 * Disable grouping by mobility if the number of pages in the
5413 * system is too low to allow the mechanism to work. It would be
5414 * more accurate, but expensive to check per-zone. This check is
5415 * made on memory-hotadd so a system can start with mobility
5416 * disabled and enable it later
5418 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5419 page_group_by_mobility_disabled = 1;
5421 page_group_by_mobility_disabled = 0;
5423 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5425 page_group_by_mobility_disabled ? "off" : "on",
5428 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5432 static int zone_batchsize(struct zone *zone)
5438 * The number of pages to batch allocate is either ~0.1%
5439 * of the zone or 1MB, whichever is smaller. The batch
5440 * size is striking a balance between allocation latency
5441 * and zone lock contention.
5443 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5444 batch /= 4; /* We effectively *= 4 below */
5449 * Clamp the batch to a 2^n - 1 value. Having a power
5450 * of 2 value was found to be more likely to have
5451 * suboptimal cache aliasing properties in some cases.
5453 * For example if 2 tasks are alternately allocating
5454 * batches of pages, one task can end up with a lot
5455 * of pages of one half of the possible page colors
5456 * and the other with pages of the other colors.
5458 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5463 /* The deferral and batching of frees should be suppressed under NOMMU
5466 * The problem is that NOMMU needs to be able to allocate large chunks
5467 * of contiguous memory as there's no hardware page translation to
5468 * assemble apparent contiguous memory from discontiguous pages.
5470 * Queueing large contiguous runs of pages for batching, however,
5471 * causes the pages to actually be freed in smaller chunks. As there
5472 * can be a significant delay between the individual batches being
5473 * recycled, this leads to the once large chunks of space being
5474 * fragmented and becoming unavailable for high-order allocations.
5480 static int percpu_pagelist_high_fraction;
5481 static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5487 unsigned long total_pages;
5489 if (!high_fraction) {
5491 * By default, the high value of the pcp is based on the zone
5492 * low watermark so that if they are full then background
5493 * reclaim will not be started prematurely.
5495 total_pages = low_wmark_pages(zone);
5498 * If percpu_pagelist_high_fraction is configured, the high
5499 * value is based on a fraction of the managed pages in the
5502 total_pages = zone_managed_pages(zone) / high_fraction;
5506 * Split the high value across all online CPUs local to the zone. Note
5507 * that early in boot that CPUs may not be online yet and that during
5508 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5509 * onlined. For memory nodes that have no CPUs, split the high value
5510 * across all online CPUs to mitigate the risk that reclaim is triggered
5511 * prematurely due to pages stored on pcp lists.
5513 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5515 nr_split_cpus = num_online_cpus();
5516 high = total_pages / nr_split_cpus;
5519 * Ensure high is at least batch*4. The multiple is based on the
5520 * historical relationship between high and batch.
5522 high = max(high, batch << 2);
5531 * pcp->high and pcp->batch values are related and generally batch is lower
5532 * than high. They are also related to pcp->count such that count is lower
5533 * than high, and as soon as it reaches high, the pcplist is flushed.
5535 * However, guaranteeing these relations at all times would require e.g. write
5536 * barriers here but also careful usage of read barriers at the read side, and
5537 * thus be prone to error and bad for performance. Thus the update only prevents
5538 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5539 * should ensure they can cope with those fields changing asynchronously, and
5540 * fully trust only the pcp->count field on the local CPU with interrupts
5543 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5544 * outside of boot time (or some other assurance that no concurrent updaters
5547 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5548 unsigned long high_max, unsigned long batch)
5550 WRITE_ONCE(pcp->batch, batch);
5551 WRITE_ONCE(pcp->high_min, high_min);
5552 WRITE_ONCE(pcp->high_max, high_max);
5555 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5559 memset(pcp, 0, sizeof(*pcp));
5560 memset(pzstats, 0, sizeof(*pzstats));
5562 spin_lock_init(&pcp->lock);
5563 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5564 INIT_LIST_HEAD(&pcp->lists[pindex]);
5567 * Set batch and high values safe for a boot pageset. A true percpu
5568 * pageset's initialization will update them subsequently. Here we don't
5569 * need to be as careful as pageset_update() as nobody can access the
5572 pcp->high_min = BOOT_PAGESET_HIGH;
5573 pcp->high_max = BOOT_PAGESET_HIGH;
5574 pcp->batch = BOOT_PAGESET_BATCH;
5575 pcp->free_count = 0;
5578 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
5579 unsigned long high_max, unsigned long batch)
5581 struct per_cpu_pages *pcp;
5584 for_each_possible_cpu(cpu) {
5585 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5586 pageset_update(pcp, high_min, high_max, batch);
5591 * Calculate and set new high and batch values for all per-cpu pagesets of a
5592 * zone based on the zone's size.
5594 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5596 int new_high_min, new_high_max, new_batch;
5598 new_batch = max(1, zone_batchsize(zone));
5599 if (percpu_pagelist_high_fraction) {
5600 new_high_min = zone_highsize(zone, new_batch, cpu_online,
5601 percpu_pagelist_high_fraction);
5603 * PCP high is tuned manually, disable auto-tuning via
5604 * setting high_min and high_max to the manual value.
5606 new_high_max = new_high_min;
5608 new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
5609 new_high_max = zone_highsize(zone, new_batch, cpu_online,
5610 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
5613 if (zone->pageset_high_min == new_high_min &&
5614 zone->pageset_high_max == new_high_max &&
5615 zone->pageset_batch == new_batch)
5618 zone->pageset_high_min = new_high_min;
5619 zone->pageset_high_max = new_high_max;
5620 zone->pageset_batch = new_batch;
5622 __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
5626 void __meminit setup_zone_pageset(struct zone *zone)
5630 /* Size may be 0 on !SMP && !NUMA */
5631 if (sizeof(struct per_cpu_zonestat) > 0)
5632 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5634 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5635 for_each_possible_cpu(cpu) {
5636 struct per_cpu_pages *pcp;
5637 struct per_cpu_zonestat *pzstats;
5639 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5640 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5641 per_cpu_pages_init(pcp, pzstats);
5644 zone_set_pageset_high_and_batch(zone, 0);
5648 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5649 * page high values need to be recalculated.
5651 static void zone_pcp_update(struct zone *zone, int cpu_online)
5653 mutex_lock(&pcp_batch_high_lock);
5654 zone_set_pageset_high_and_batch(zone, cpu_online);
5655 mutex_unlock(&pcp_batch_high_lock);
5658 static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu)
5660 struct per_cpu_pages *pcp;
5661 struct cpu_cacheinfo *cci;
5663 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5664 cci = get_cpu_cacheinfo(cpu);
5666 * If data cache slice of CPU is large enough, "pcp->batch"
5667 * pages can be preserved in PCP before draining PCP for
5668 * consecutive high-order pages freeing without allocation.
5669 * This can reduce zone lock contention without hurting
5670 * cache-hot pages sharing.
5672 spin_lock(&pcp->lock);
5673 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
5674 pcp->flags |= PCPF_FREE_HIGH_BATCH;
5676 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
5677 spin_unlock(&pcp->lock);
5680 void setup_pcp_cacheinfo(unsigned int cpu)
5684 for_each_populated_zone(zone)
5685 zone_pcp_update_cacheinfo(zone, cpu);
5689 * Allocate per cpu pagesets and initialize them.
5690 * Before this call only boot pagesets were available.
5692 void __init setup_per_cpu_pageset(void)
5694 struct pglist_data *pgdat;
5696 int __maybe_unused cpu;
5698 for_each_populated_zone(zone)
5699 setup_zone_pageset(zone);
5703 * Unpopulated zones continue using the boot pagesets.
5704 * The numa stats for these pagesets need to be reset.
5705 * Otherwise, they will end up skewing the stats of
5706 * the nodes these zones are associated with.
5708 for_each_possible_cpu(cpu) {
5709 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5710 memset(pzstats->vm_numa_event, 0,
5711 sizeof(pzstats->vm_numa_event));
5715 for_each_online_pgdat(pgdat)
5716 pgdat->per_cpu_nodestats =
5717 alloc_percpu(struct per_cpu_nodestat);
5720 __meminit void zone_pcp_init(struct zone *zone)
5723 * per cpu subsystem is not up at this point. The following code
5724 * relies on the ability of the linker to provide the
5725 * offset of a (static) per cpu variable into the per cpu area.
5727 zone->per_cpu_pageset = &boot_pageset;
5728 zone->per_cpu_zonestats = &boot_zonestats;
5729 zone->pageset_high_min = BOOT_PAGESET_HIGH;
5730 zone->pageset_high_max = BOOT_PAGESET_HIGH;
5731 zone->pageset_batch = BOOT_PAGESET_BATCH;
5733 if (populated_zone(zone))
5734 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5735 zone->present_pages, zone_batchsize(zone));
5738 void adjust_managed_page_count(struct page *page, long count)
5740 atomic_long_add(count, &page_zone(page)->managed_pages);
5741 totalram_pages_add(count);
5742 #ifdef CONFIG_HIGHMEM
5743 if (PageHighMem(page))
5744 totalhigh_pages_add(count);
5747 EXPORT_SYMBOL(adjust_managed_page_count);
5749 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5752 unsigned long pages = 0;
5754 start = (void *)PAGE_ALIGN((unsigned long)start);
5755 end = (void *)((unsigned long)end & PAGE_MASK);
5756 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5757 struct page *page = virt_to_page(pos);
5758 void *direct_map_addr;
5761 * 'direct_map_addr' might be different from 'pos'
5762 * because some architectures' virt_to_page()
5763 * work with aliases. Getting the direct map
5764 * address ensures that we get a _writeable_
5765 * alias for the memset().
5767 direct_map_addr = page_address(page);
5769 * Perform a kasan-unchecked memset() since this memory
5770 * has not been initialized.
5772 direct_map_addr = kasan_reset_tag(direct_map_addr);
5773 if ((unsigned int)poison <= 0xFF)
5774 memset(direct_map_addr, poison, PAGE_SIZE);
5776 free_reserved_page(page);
5780 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5785 static int page_alloc_cpu_dead(unsigned int cpu)
5789 lru_add_drain_cpu(cpu);
5790 mlock_drain_remote(cpu);
5794 * Spill the event counters of the dead processor
5795 * into the current processors event counters.
5796 * This artificially elevates the count of the current
5799 vm_events_fold_cpu(cpu);
5802 * Zero the differential counters of the dead processor
5803 * so that the vm statistics are consistent.
5805 * This is only okay since the processor is dead and cannot
5806 * race with what we are doing.
5808 cpu_vm_stats_fold(cpu);
5810 for_each_populated_zone(zone)
5811 zone_pcp_update(zone, 0);
5816 static int page_alloc_cpu_online(unsigned int cpu)
5820 for_each_populated_zone(zone)
5821 zone_pcp_update(zone, 1);
5825 void __init page_alloc_init_cpuhp(void)
5829 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5830 "mm/page_alloc:pcp",
5831 page_alloc_cpu_online,
5832 page_alloc_cpu_dead);
5837 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5838 * or min_free_kbytes changes.
5840 static void calculate_totalreserve_pages(void)
5842 struct pglist_data *pgdat;
5843 unsigned long reserve_pages = 0;
5844 enum zone_type i, j;
5846 for_each_online_pgdat(pgdat) {
5848 pgdat->totalreserve_pages = 0;
5850 for (i = 0; i < MAX_NR_ZONES; i++) {
5851 struct zone *zone = pgdat->node_zones + i;
5853 unsigned long managed_pages = zone_managed_pages(zone);
5855 /* Find valid and maximum lowmem_reserve in the zone */
5856 for (j = i; j < MAX_NR_ZONES; j++) {
5857 if (zone->lowmem_reserve[j] > max)
5858 max = zone->lowmem_reserve[j];
5861 /* we treat the high watermark as reserved pages. */
5862 max += high_wmark_pages(zone);
5864 if (max > managed_pages)
5865 max = managed_pages;
5867 pgdat->totalreserve_pages += max;
5869 reserve_pages += max;
5872 totalreserve_pages = reserve_pages;
5876 * setup_per_zone_lowmem_reserve - called whenever
5877 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5878 * has a correct pages reserved value, so an adequate number of
5879 * pages are left in the zone after a successful __alloc_pages().
5881 static void setup_per_zone_lowmem_reserve(void)
5883 struct pglist_data *pgdat;
5884 enum zone_type i, j;
5886 for_each_online_pgdat(pgdat) {
5887 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5888 struct zone *zone = &pgdat->node_zones[i];
5889 int ratio = sysctl_lowmem_reserve_ratio[i];
5890 bool clear = !ratio || !zone_managed_pages(zone);
5891 unsigned long managed_pages = 0;
5893 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5894 struct zone *upper_zone = &pgdat->node_zones[j];
5895 bool empty = !zone_managed_pages(upper_zone);
5897 managed_pages += zone_managed_pages(upper_zone);
5900 zone->lowmem_reserve[j] = 0;
5902 zone->lowmem_reserve[j] = managed_pages / ratio;
5907 /* update totalreserve_pages */
5908 calculate_totalreserve_pages();
5911 static void __setup_per_zone_wmarks(void)
5913 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5914 unsigned long lowmem_pages = 0;
5916 unsigned long flags;
5918 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5919 for_each_zone(zone) {
5920 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5921 lowmem_pages += zone_managed_pages(zone);
5924 for_each_zone(zone) {
5927 spin_lock_irqsave(&zone->lock, flags);
5928 tmp = (u64)pages_min * zone_managed_pages(zone);
5929 tmp = div64_ul(tmp, lowmem_pages);
5930 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5932 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5933 * need highmem and movable zones pages, so cap pages_min
5934 * to a small value here.
5936 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5937 * deltas control async page reclaim, and so should
5938 * not be capped for highmem and movable zones.
5940 unsigned long min_pages;
5942 min_pages = zone_managed_pages(zone) / 1024;
5943 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5944 zone->_watermark[WMARK_MIN] = min_pages;
5947 * If it's a lowmem zone, reserve a number of pages
5948 * proportionate to the zone's size.
5950 zone->_watermark[WMARK_MIN] = tmp;
5954 * Set the kswapd watermarks distance according to the
5955 * scale factor in proportion to available memory, but
5956 * ensure a minimum size on small systems.
5958 tmp = max_t(u64, tmp >> 2,
5959 mult_frac(zone_managed_pages(zone),
5960 watermark_scale_factor, 10000));
5962 zone->watermark_boost = 0;
5963 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
5964 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5965 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5967 spin_unlock_irqrestore(&zone->lock, flags);
5970 /* update totalreserve_pages */
5971 calculate_totalreserve_pages();
5975 * setup_per_zone_wmarks - called when min_free_kbytes changes
5976 * or when memory is hot-{added|removed}
5978 * Ensures that the watermark[min,low,high] values for each zone are set
5979 * correctly with respect to min_free_kbytes.
5981 void setup_per_zone_wmarks(void)
5984 static DEFINE_SPINLOCK(lock);
5987 __setup_per_zone_wmarks();
5991 * The watermark size have changed so update the pcpu batch
5992 * and high limits or the limits may be inappropriate.
5995 zone_pcp_update(zone, 0);
5999 * Initialise min_free_kbytes.
6001 * For small machines we want it small (128k min). For large machines
6002 * we want it large (256MB max). But it is not linear, because network
6003 * bandwidth does not increase linearly with machine size. We use
6005 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6006 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6022 void calculate_min_free_kbytes(void)
6024 unsigned long lowmem_kbytes;
6025 int new_min_free_kbytes;
6027 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6028 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6030 if (new_min_free_kbytes > user_min_free_kbytes)
6031 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
6033 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6034 new_min_free_kbytes, user_min_free_kbytes);
6038 int __meminit init_per_zone_wmark_min(void)
6040 calculate_min_free_kbytes();
6041 setup_per_zone_wmarks();
6042 refresh_zone_stat_thresholds();
6043 setup_per_zone_lowmem_reserve();
6046 setup_min_unmapped_ratio();
6047 setup_min_slab_ratio();
6050 khugepaged_min_free_kbytes_update();
6054 postcore_initcall(init_per_zone_wmark_min)
6057 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6058 * that we can call two helper functions whenever min_free_kbytes
6061 static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6062 void *buffer, size_t *length, loff_t *ppos)
6066 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6071 user_min_free_kbytes = min_free_kbytes;
6072 setup_per_zone_wmarks();
6077 static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6078 void *buffer, size_t *length, loff_t *ppos)
6082 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6087 setup_per_zone_wmarks();
6093 static void setup_min_unmapped_ratio(void)
6098 for_each_online_pgdat(pgdat)
6099 pgdat->min_unmapped_pages = 0;
6102 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6103 sysctl_min_unmapped_ratio) / 100;
6107 static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6108 void *buffer, size_t *length, loff_t *ppos)
6112 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6116 setup_min_unmapped_ratio();
6121 static void setup_min_slab_ratio(void)
6126 for_each_online_pgdat(pgdat)
6127 pgdat->min_slab_pages = 0;
6130 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6131 sysctl_min_slab_ratio) / 100;
6134 static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6135 void *buffer, size_t *length, loff_t *ppos)
6139 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6143 setup_min_slab_ratio();
6150 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6151 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6152 * whenever sysctl_lowmem_reserve_ratio changes.
6154 * The reserve ratio obviously has absolutely no relation with the
6155 * minimum watermarks. The lowmem reserve ratio can only make sense
6156 * if in function of the boot time zone sizes.
6158 static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table,
6159 int write, void *buffer, size_t *length, loff_t *ppos)
6163 proc_dointvec_minmax(table, write, buffer, length, ppos);
6165 for (i = 0; i < MAX_NR_ZONES; i++) {
6166 if (sysctl_lowmem_reserve_ratio[i] < 1)
6167 sysctl_lowmem_reserve_ratio[i] = 0;
6170 setup_per_zone_lowmem_reserve();
6175 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6176 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6177 * pagelist can have before it gets flushed back to buddy allocator.
6179 static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
6180 int write, void *buffer, size_t *length, loff_t *ppos)
6183 int old_percpu_pagelist_high_fraction;
6186 mutex_lock(&pcp_batch_high_lock);
6187 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6189 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6190 if (!write || ret < 0)
6193 /* Sanity checking to avoid pcp imbalance */
6194 if (percpu_pagelist_high_fraction &&
6195 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6196 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6202 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6205 for_each_populated_zone(zone)
6206 zone_set_pageset_high_and_batch(zone, 0);
6208 mutex_unlock(&pcp_batch_high_lock);
6212 static struct ctl_table page_alloc_sysctl_table[] = {
6214 .procname = "min_free_kbytes",
6215 .data = &min_free_kbytes,
6216 .maxlen = sizeof(min_free_kbytes),
6218 .proc_handler = min_free_kbytes_sysctl_handler,
6219 .extra1 = SYSCTL_ZERO,
6222 .procname = "watermark_boost_factor",
6223 .data = &watermark_boost_factor,
6224 .maxlen = sizeof(watermark_boost_factor),
6226 .proc_handler = proc_dointvec_minmax,
6227 .extra1 = SYSCTL_ZERO,
6230 .procname = "watermark_scale_factor",
6231 .data = &watermark_scale_factor,
6232 .maxlen = sizeof(watermark_scale_factor),
6234 .proc_handler = watermark_scale_factor_sysctl_handler,
6235 .extra1 = SYSCTL_ONE,
6236 .extra2 = SYSCTL_THREE_THOUSAND,
6239 .procname = "percpu_pagelist_high_fraction",
6240 .data = &percpu_pagelist_high_fraction,
6241 .maxlen = sizeof(percpu_pagelist_high_fraction),
6243 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6244 .extra1 = SYSCTL_ZERO,
6247 .procname = "lowmem_reserve_ratio",
6248 .data = &sysctl_lowmem_reserve_ratio,
6249 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6251 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6255 .procname = "numa_zonelist_order",
6256 .data = &numa_zonelist_order,
6257 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6259 .proc_handler = numa_zonelist_order_handler,
6262 .procname = "min_unmapped_ratio",
6263 .data = &sysctl_min_unmapped_ratio,
6264 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6266 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6267 .extra1 = SYSCTL_ZERO,
6268 .extra2 = SYSCTL_ONE_HUNDRED,
6271 .procname = "min_slab_ratio",
6272 .data = &sysctl_min_slab_ratio,
6273 .maxlen = sizeof(sysctl_min_slab_ratio),
6275 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6276 .extra1 = SYSCTL_ZERO,
6277 .extra2 = SYSCTL_ONE_HUNDRED,
6282 void __init page_alloc_sysctl_init(void)
6284 register_sysctl_init("vm", page_alloc_sysctl_table);
6287 #ifdef CONFIG_CONTIG_ALLOC
6288 /* Usage: See admin-guide/dynamic-debug-howto.rst */
6289 static void alloc_contig_dump_pages(struct list_head *page_list)
6291 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6293 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6297 list_for_each_entry(page, page_list, lru)
6298 dump_page(page, "migration failure");
6303 * [start, end) must belong to a single zone.
6304 * @migratetype: using migratetype to filter the type of migration in
6305 * trace_mm_alloc_contig_migrate_range_info.
6307 int __alloc_contig_migrate_range(struct compact_control *cc,
6308 unsigned long start, unsigned long end,
6311 /* This function is based on compact_zone() from compaction.c. */
6312 unsigned int nr_reclaimed;
6313 unsigned long pfn = start;
6314 unsigned int tries = 0;
6316 struct migration_target_control mtc = {
6317 .nid = zone_to_nid(cc->zone),
6318 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6319 .reason = MR_CONTIG_RANGE,
6322 unsigned long total_mapped = 0;
6323 unsigned long total_migrated = 0;
6324 unsigned long total_reclaimed = 0;
6326 lru_cache_disable();
6328 while (pfn < end || !list_empty(&cc->migratepages)) {
6329 if (fatal_signal_pending(current)) {
6334 if (list_empty(&cc->migratepages)) {
6335 cc->nr_migratepages = 0;
6336 ret = isolate_migratepages_range(cc, pfn, end);
6337 if (ret && ret != -EAGAIN)
6339 pfn = cc->migrate_pfn;
6341 } else if (++tries == 5) {
6346 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6348 cc->nr_migratepages -= nr_reclaimed;
6350 if (trace_mm_alloc_contig_migrate_range_info_enabled()) {
6351 total_reclaimed += nr_reclaimed;
6352 list_for_each_entry(page, &cc->migratepages, lru) {
6353 struct folio *folio = page_folio(page);
6355 total_mapped += folio_mapped(folio) *
6356 folio_nr_pages(folio);
6360 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6361 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6363 if (trace_mm_alloc_contig_migrate_range_info_enabled() && !ret)
6364 total_migrated += cc->nr_migratepages;
6367 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6368 * to retry again over this error, so do the same here.
6376 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6377 alloc_contig_dump_pages(&cc->migratepages);
6378 putback_movable_pages(&cc->migratepages);
6381 trace_mm_alloc_contig_migrate_range_info(start, end, migratetype,
6385 return (ret < 0) ? ret : 0;
6389 * alloc_contig_range() -- tries to allocate given range of pages
6390 * @start: start PFN to allocate
6391 * @end: one-past-the-last PFN to allocate
6392 * @migratetype: migratetype of the underlying pageblocks (either
6393 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6394 * in range must have the same migratetype and it must
6395 * be either of the two.
6396 * @gfp_mask: GFP mask to use during compaction
6398 * The PFN range does not have to be pageblock aligned. The PFN range must
6399 * belong to a single zone.
6401 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6402 * pageblocks in the range. Once isolated, the pageblocks should not
6403 * be modified by others.
6405 * Return: zero on success or negative error code. On success all
6406 * pages which PFN is in [start, end) are allocated for the caller and
6407 * need to be freed with free_contig_range().
6409 int alloc_contig_range_noprof(unsigned long start, unsigned long end,
6410 unsigned migratetype, gfp_t gfp_mask)
6412 unsigned long outer_start, outer_end;
6415 struct compact_control cc = {
6416 .nr_migratepages = 0,
6418 .zone = page_zone(pfn_to_page(start)),
6419 .mode = MIGRATE_SYNC,
6420 .ignore_skip_hint = true,
6421 .no_set_skip_hint = true,
6422 .gfp_mask = current_gfp_context(gfp_mask),
6423 .alloc_contig = true,
6425 INIT_LIST_HEAD(&cc.migratepages);
6428 * What we do here is we mark all pageblocks in range as
6429 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6430 * have different sizes, and due to the way page allocator
6431 * work, start_isolate_page_range() has special handlings for this.
6433 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6434 * migrate the pages from an unaligned range (ie. pages that
6435 * we are interested in). This will put all the pages in
6436 * range back to page allocator as MIGRATE_ISOLATE.
6438 * When this is done, we take the pages in range from page
6439 * allocator removing them from the buddy system. This way
6440 * page allocator will never consider using them.
6442 * This lets us mark the pageblocks back as
6443 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6444 * aligned range but not in the unaligned, original range are
6445 * put back to page allocator so that buddy can use them.
6448 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6452 drain_all_pages(cc.zone);
6455 * In case of -EBUSY, we'd like to know which page causes problem.
6456 * So, just fall through. test_pages_isolated() has a tracepoint
6457 * which will report the busy page.
6459 * It is possible that busy pages could become available before
6460 * the call to test_pages_isolated, and the range will actually be
6461 * allocated. So, if we fall through be sure to clear ret so that
6462 * -EBUSY is not accidentally used or returned to caller.
6464 ret = __alloc_contig_migrate_range(&cc, start, end, migratetype);
6465 if (ret && ret != -EBUSY)
6470 * Pages from [start, end) are within a pageblock_nr_pages
6471 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6472 * more, all pages in [start, end) are free in page allocator.
6473 * What we are going to do is to allocate all pages from
6474 * [start, end) (that is remove them from page allocator).
6476 * The only problem is that pages at the beginning and at the
6477 * end of interesting range may be not aligned with pages that
6478 * page allocator holds, ie. they can be part of higher order
6479 * pages. Because of this, we reserve the bigger range and
6480 * once this is done free the pages we are not interested in.
6482 * We don't have to hold zone->lock here because the pages are
6483 * isolated thus they won't get removed from buddy.
6485 outer_start = find_large_buddy(start);
6487 /* Make sure the range is really isolated. */
6488 if (test_pages_isolated(outer_start, end, 0)) {
6493 /* Grab isolated pages from freelists. */
6494 outer_end = isolate_freepages_range(&cc, outer_start, end);
6500 /* Free head and tail (if any) */
6501 if (start != outer_start)
6502 free_contig_range(outer_start, start - outer_start);
6503 if (end != outer_end)
6504 free_contig_range(end, outer_end - end);
6507 undo_isolate_page_range(start, end, migratetype);
6510 EXPORT_SYMBOL(alloc_contig_range_noprof);
6512 static int __alloc_contig_pages(unsigned long start_pfn,
6513 unsigned long nr_pages, gfp_t gfp_mask)
6515 unsigned long end_pfn = start_pfn + nr_pages;
6517 return alloc_contig_range_noprof(start_pfn, end_pfn, MIGRATE_MOVABLE,
6521 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6522 unsigned long nr_pages)
6524 unsigned long i, end_pfn = start_pfn + nr_pages;
6527 for (i = start_pfn; i < end_pfn; i++) {
6528 page = pfn_to_online_page(i);
6532 if (page_zone(page) != z)
6535 if (PageReserved(page))
6544 static bool zone_spans_last_pfn(const struct zone *zone,
6545 unsigned long start_pfn, unsigned long nr_pages)
6547 unsigned long last_pfn = start_pfn + nr_pages - 1;
6549 return zone_spans_pfn(zone, last_pfn);
6553 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6554 * @nr_pages: Number of contiguous pages to allocate
6555 * @gfp_mask: GFP mask to limit search and used during compaction
6557 * @nodemask: Mask for other possible nodes
6559 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6560 * on an applicable zonelist to find a contiguous pfn range which can then be
6561 * tried for allocation with alloc_contig_range(). This routine is intended
6562 * for allocation requests which can not be fulfilled with the buddy allocator.
6564 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6565 * power of two, then allocated range is also guaranteed to be aligned to same
6566 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6568 * Allocated pages can be freed with free_contig_range() or by manually calling
6569 * __free_page() on each allocated page.
6571 * Return: pointer to contiguous pages on success, or NULL if not successful.
6573 struct page *alloc_contig_pages_noprof(unsigned long nr_pages, gfp_t gfp_mask,
6574 int nid, nodemask_t *nodemask)
6576 unsigned long ret, pfn, flags;
6577 struct zonelist *zonelist;
6581 zonelist = node_zonelist(nid, gfp_mask);
6582 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6583 gfp_zone(gfp_mask), nodemask) {
6584 spin_lock_irqsave(&zone->lock, flags);
6586 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6587 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6588 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6590 * We release the zone lock here because
6591 * alloc_contig_range() will also lock the zone
6592 * at some point. If there's an allocation
6593 * spinning on this lock, it may win the race
6594 * and cause alloc_contig_range() to fail...
6596 spin_unlock_irqrestore(&zone->lock, flags);
6597 ret = __alloc_contig_pages(pfn, nr_pages,
6600 return pfn_to_page(pfn);
6601 spin_lock_irqsave(&zone->lock, flags);
6605 spin_unlock_irqrestore(&zone->lock, flags);
6609 #endif /* CONFIG_CONTIG_ALLOC */
6611 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6613 unsigned long count = 0;
6615 for (; nr_pages--; pfn++) {
6616 struct page *page = pfn_to_page(pfn);
6618 count += page_count(page) != 1;
6621 WARN(count != 0, "%lu pages are still in use!\n", count);
6623 EXPORT_SYMBOL(free_contig_range);
6626 * Effectively disable pcplists for the zone by setting the high limit to 0
6627 * and draining all cpus. A concurrent page freeing on another CPU that's about
6628 * to put the page on pcplist will either finish before the drain and the page
6629 * will be drained, or observe the new high limit and skip the pcplist.
6631 * Must be paired with a call to zone_pcp_enable().
6633 void zone_pcp_disable(struct zone *zone)
6635 mutex_lock(&pcp_batch_high_lock);
6636 __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
6637 __drain_all_pages(zone, true);
6640 void zone_pcp_enable(struct zone *zone)
6642 __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
6643 zone->pageset_high_max, zone->pageset_batch);
6644 mutex_unlock(&pcp_batch_high_lock);
6647 void zone_pcp_reset(struct zone *zone)
6650 struct per_cpu_zonestat *pzstats;
6652 if (zone->per_cpu_pageset != &boot_pageset) {
6653 for_each_online_cpu(cpu) {
6654 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6655 drain_zonestat(zone, pzstats);
6657 free_percpu(zone->per_cpu_pageset);
6658 zone->per_cpu_pageset = &boot_pageset;
6659 if (zone->per_cpu_zonestats != &boot_zonestats) {
6660 free_percpu(zone->per_cpu_zonestats);
6661 zone->per_cpu_zonestats = &boot_zonestats;
6666 #ifdef CONFIG_MEMORY_HOTREMOVE
6668 * All pages in the range must be in a single zone, must not contain holes,
6669 * must span full sections, and must be isolated before calling this function.
6671 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6673 unsigned long pfn = start_pfn;
6677 unsigned long flags;
6679 offline_mem_sections(pfn, end_pfn);
6680 zone = page_zone(pfn_to_page(pfn));
6681 spin_lock_irqsave(&zone->lock, flags);
6682 while (pfn < end_pfn) {
6683 page = pfn_to_page(pfn);
6685 * The HWPoisoned page may be not in buddy system, and
6686 * page_count() is not 0.
6688 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6693 * At this point all remaining PageOffline() pages have a
6694 * reference count of 0 and can simply be skipped.
6696 if (PageOffline(page)) {
6697 BUG_ON(page_count(page));
6698 BUG_ON(PageBuddy(page));
6703 BUG_ON(page_count(page));
6704 BUG_ON(!PageBuddy(page));
6705 VM_WARN_ON(get_pageblock_migratetype(page) != MIGRATE_ISOLATE);
6706 order = buddy_order(page);
6707 del_page_from_free_list(page, zone, order, MIGRATE_ISOLATE);
6708 pfn += (1 << order);
6710 spin_unlock_irqrestore(&zone->lock, flags);
6715 * This function returns a stable result only if called under zone lock.
6717 bool is_free_buddy_page(const struct page *page)
6719 unsigned long pfn = page_to_pfn(page);
6722 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6723 const struct page *head = page - (pfn & ((1 << order) - 1));
6725 if (PageBuddy(head) &&
6726 buddy_order_unsafe(head) >= order)
6730 return order <= MAX_PAGE_ORDER;
6732 EXPORT_SYMBOL(is_free_buddy_page);
6734 #ifdef CONFIG_MEMORY_FAILURE
6735 static inline void add_to_free_list(struct page *page, struct zone *zone,
6736 unsigned int order, int migratetype,
6739 __add_to_free_list(page, zone, order, migratetype, tail);
6740 account_freepages(zone, 1 << order, migratetype);
6744 * Break down a higher-order page in sub-pages, and keep our target out of
6747 static void break_down_buddy_pages(struct zone *zone, struct page *page,
6748 struct page *target, int low, int high,
6751 unsigned long size = 1 << high;
6752 struct page *current_buddy;
6754 while (high > low) {
6758 if (target >= &page[size]) {
6759 current_buddy = page;
6762 current_buddy = page + size;
6765 if (set_page_guard(zone, current_buddy, high))
6768 add_to_free_list(current_buddy, zone, high, migratetype, false);
6769 set_buddy_order(current_buddy, high);
6774 * Take a page that will be marked as poisoned off the buddy allocator.
6776 bool take_page_off_buddy(struct page *page)
6778 struct zone *zone = page_zone(page);
6779 unsigned long pfn = page_to_pfn(page);
6780 unsigned long flags;
6784 spin_lock_irqsave(&zone->lock, flags);
6785 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6786 struct page *page_head = page - (pfn & ((1 << order) - 1));
6787 int page_order = buddy_order(page_head);
6789 if (PageBuddy(page_head) && page_order >= order) {
6790 unsigned long pfn_head = page_to_pfn(page_head);
6791 int migratetype = get_pfnblock_migratetype(page_head,
6794 del_page_from_free_list(page_head, zone, page_order,
6796 break_down_buddy_pages(zone, page_head, page, 0,
6797 page_order, migratetype);
6798 SetPageHWPoisonTakenOff(page);
6802 if (page_count(page_head) > 0)
6805 spin_unlock_irqrestore(&zone->lock, flags);
6810 * Cancel takeoff done by take_page_off_buddy().
6812 bool put_page_back_buddy(struct page *page)
6814 struct zone *zone = page_zone(page);
6815 unsigned long flags;
6818 spin_lock_irqsave(&zone->lock, flags);
6819 if (put_page_testzero(page)) {
6820 unsigned long pfn = page_to_pfn(page);
6821 int migratetype = get_pfnblock_migratetype(page, pfn);
6823 ClearPageHWPoisonTakenOff(page);
6824 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6825 if (TestClearPageHWPoison(page)) {
6829 spin_unlock_irqrestore(&zone->lock, flags);
6835 #ifdef CONFIG_ZONE_DMA
6836 bool has_managed_dma(void)
6838 struct pglist_data *pgdat;
6840 for_each_online_pgdat(pgdat) {
6841 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6843 if (managed_zone(zone))
6848 #endif /* CONFIG_ZONE_DMA */
6850 #ifdef CONFIG_UNACCEPTED_MEMORY
6852 /* Counts number of zones with unaccepted pages. */
6853 static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6855 static bool lazy_accept = true;
6857 static int __init accept_memory_parse(char *p)
6859 if (!strcmp(p, "lazy")) {
6862 } else if (!strcmp(p, "eager")) {
6863 lazy_accept = false;
6869 early_param("accept_memory", accept_memory_parse);
6871 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6873 phys_addr_t start = page_to_phys(page);
6874 phys_addr_t end = start + (PAGE_SIZE << order);
6876 return range_contains_unaccepted_memory(start, end);
6879 static void accept_page(struct page *page, unsigned int order)
6881 phys_addr_t start = page_to_phys(page);
6883 accept_memory(start, start + (PAGE_SIZE << order));
6886 static bool try_to_accept_memory_one(struct zone *zone)
6888 unsigned long flags;
6892 if (list_empty(&zone->unaccepted_pages))
6895 spin_lock_irqsave(&zone->lock, flags);
6896 page = list_first_entry_or_null(&zone->unaccepted_pages,
6899 spin_unlock_irqrestore(&zone->lock, flags);
6903 list_del(&page->lru);
6904 last = list_empty(&zone->unaccepted_pages);
6906 account_freepages(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6907 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6908 spin_unlock_irqrestore(&zone->lock, flags);
6910 accept_page(page, MAX_PAGE_ORDER);
6912 __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
6915 static_branch_dec(&zones_with_unaccepted_pages);
6920 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6925 /* How much to accept to get to high watermark? */
6926 to_accept = high_wmark_pages(zone) -
6927 (zone_page_state(zone, NR_FREE_PAGES) -
6928 __zone_watermark_unusable_free(zone, order, 0));
6930 /* Accept at least one page */
6932 if (!try_to_accept_memory_one(zone))
6935 to_accept -= MAX_ORDER_NR_PAGES;
6936 } while (to_accept > 0);
6941 static inline bool has_unaccepted_memory(void)
6943 return static_branch_unlikely(&zones_with_unaccepted_pages);
6946 static bool __free_unaccepted(struct page *page)
6948 struct zone *zone = page_zone(page);
6949 unsigned long flags;
6955 spin_lock_irqsave(&zone->lock, flags);
6956 first = list_empty(&zone->unaccepted_pages);
6957 list_add_tail(&page->lru, &zone->unaccepted_pages);
6958 account_freepages(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6959 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
6960 spin_unlock_irqrestore(&zone->lock, flags);
6963 static_branch_inc(&zones_with_unaccepted_pages);
6970 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6975 static void accept_page(struct page *page, unsigned int order)
6979 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6984 static inline bool has_unaccepted_memory(void)
6989 static bool __free_unaccepted(struct page *page)
6995 #endif /* CONFIG_UNACCEPTED_MEMORY */