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 * This might let an unmovable request use a reclaimable pageblock
600 * and vice-versa but no more than normal fallback logic which can
601 * have trouble finding a high-order free page.
603 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
611 static inline struct capture_control *task_capc(struct zone *zone)
617 compaction_capture(struct capture_control *capc, struct page *page,
618 int order, int migratetype)
622 #endif /* CONFIG_COMPACTION */
624 static inline void account_freepages(struct zone *zone, int nr_pages,
627 if (is_migrate_isolate(migratetype))
630 __mod_zone_page_state(zone, NR_FREE_PAGES, nr_pages);
632 if (is_migrate_cma(migratetype))
633 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, nr_pages);
636 /* Used for pages not on another list */
637 static inline void __add_to_free_list(struct page *page, struct zone *zone,
638 unsigned int order, int migratetype,
641 struct free_area *area = &zone->free_area[order];
643 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
644 "page type is %lu, passed migratetype is %d (nr=%d)\n",
645 get_pageblock_migratetype(page), migratetype, 1 << order);
648 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
650 list_add(&page->buddy_list, &area->free_list[migratetype]);
655 * Used for pages which are on another list. Move the pages to the tail
656 * of the list - so the moved pages won't immediately be considered for
657 * allocation again (e.g., optimization for memory onlining).
659 static inline void move_to_free_list(struct page *page, struct zone *zone,
660 unsigned int order, int old_mt, int new_mt)
662 struct free_area *area = &zone->free_area[order];
664 /* Free page moving can fail, so it happens before the type update */
665 VM_WARN_ONCE(get_pageblock_migratetype(page) != old_mt,
666 "page type is %lu, passed migratetype is %d (nr=%d)\n",
667 get_pageblock_migratetype(page), old_mt, 1 << order);
669 list_move_tail(&page->buddy_list, &area->free_list[new_mt]);
671 account_freepages(zone, -(1 << order), old_mt);
672 account_freepages(zone, 1 << order, new_mt);
675 static inline void __del_page_from_free_list(struct page *page, struct zone *zone,
676 unsigned int order, int migratetype)
678 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
679 "page type is %lu, passed migratetype is %d (nr=%d)\n",
680 get_pageblock_migratetype(page), migratetype, 1 << order);
682 /* clear reported state and update reported page count */
683 if (page_reported(page))
684 __ClearPageReported(page);
686 list_del(&page->buddy_list);
687 __ClearPageBuddy(page);
688 set_page_private(page, 0);
689 zone->free_area[order].nr_free--;
692 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
693 unsigned int order, int migratetype)
695 __del_page_from_free_list(page, zone, order, migratetype);
696 account_freepages(zone, -(1 << order), migratetype);
699 static inline struct page *get_page_from_free_area(struct free_area *area,
702 return list_first_entry_or_null(&area->free_list[migratetype],
703 struct page, buddy_list);
707 * If this is not the largest possible page, check if the buddy
708 * of the next-highest order is free. If it is, it's possible
709 * that pages are being freed that will coalesce soon. In case,
710 * that is happening, add the free page to the tail of the list
711 * so it's less likely to be used soon and more likely to be merged
712 * as a higher order page
715 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
716 struct page *page, unsigned int order)
718 unsigned long higher_page_pfn;
719 struct page *higher_page;
721 if (order >= MAX_PAGE_ORDER - 1)
724 higher_page_pfn = buddy_pfn & pfn;
725 higher_page = page + (higher_page_pfn - pfn);
727 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
732 * Freeing function for a buddy system allocator.
734 * The concept of a buddy system is to maintain direct-mapped table
735 * (containing bit values) for memory blocks of various "orders".
736 * The bottom level table contains the map for the smallest allocatable
737 * units of memory (here, pages), and each level above it describes
738 * pairs of units from the levels below, hence, "buddies".
739 * At a high level, all that happens here is marking the table entry
740 * at the bottom level available, and propagating the changes upward
741 * as necessary, plus some accounting needed to play nicely with other
742 * parts of the VM system.
743 * At each level, we keep a list of pages, which are heads of continuous
744 * free pages of length of (1 << order) and marked with PageBuddy.
745 * Page's order is recorded in page_private(page) field.
746 * So when we are allocating or freeing one, we can derive the state of the
747 * other. That is, if we allocate a small block, and both were
748 * free, the remainder of the region must be split into blocks.
749 * If a block is freed, and its buddy is also free, then this
750 * triggers coalescing into a block of larger size.
755 static inline void __free_one_page(struct page *page,
757 struct zone *zone, unsigned int order,
758 int migratetype, fpi_t fpi_flags)
760 struct capture_control *capc = task_capc(zone);
761 unsigned long buddy_pfn = 0;
762 unsigned long combined_pfn;
766 VM_BUG_ON(!zone_is_initialized(zone));
767 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
769 VM_BUG_ON(migratetype == -1);
770 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
771 VM_BUG_ON_PAGE(bad_range(zone, page), page);
773 account_freepages(zone, 1 << order, migratetype);
775 while (order < MAX_PAGE_ORDER) {
776 int buddy_mt = migratetype;
778 if (compaction_capture(capc, page, order, migratetype)) {
779 account_freepages(zone, -(1 << order), migratetype);
783 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
787 if (unlikely(order >= pageblock_order)) {
789 * We want to prevent merge between freepages on pageblock
790 * without fallbacks and normal pageblock. Without this,
791 * pageblock isolation could cause incorrect freepage or CMA
792 * accounting or HIGHATOMIC accounting.
794 buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
796 if (migratetype != buddy_mt &&
797 (!migratetype_is_mergeable(migratetype) ||
798 !migratetype_is_mergeable(buddy_mt)))
803 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
804 * merge with it and move up one order.
806 if (page_is_guard(buddy))
807 clear_page_guard(zone, buddy, order);
809 __del_page_from_free_list(buddy, zone, order, buddy_mt);
811 if (unlikely(buddy_mt != migratetype)) {
813 * Match buddy type. This ensures that an
814 * expand() down the line puts the sub-blocks
815 * on the right freelists.
817 set_pageblock_migratetype(buddy, migratetype);
820 combined_pfn = buddy_pfn & pfn;
821 page = page + (combined_pfn - pfn);
827 set_buddy_order(page, order);
829 if (fpi_flags & FPI_TO_TAIL)
831 else if (is_shuffle_order(order))
832 to_tail = shuffle_pick_tail();
834 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
836 __add_to_free_list(page, zone, order, migratetype, to_tail);
838 /* Notify page reporting subsystem of freed page */
839 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
840 page_reporting_notify_free(order);
844 * A bad page could be due to a number of fields. Instead of multiple branches,
845 * try and check multiple fields with one check. The caller must do a detailed
846 * check if necessary.
848 static inline bool page_expected_state(struct page *page,
849 unsigned long check_flags)
851 if (unlikely(atomic_read(&page->_mapcount) != -1))
854 if (unlikely((unsigned long)page->mapping |
855 page_ref_count(page) |
859 #ifdef CONFIG_PAGE_POOL
860 ((page->pp_magic & ~0x3UL) == PP_SIGNATURE) |
862 (page->flags & check_flags)))
868 static const char *page_bad_reason(struct page *page, unsigned long flags)
870 const char *bad_reason = NULL;
872 if (unlikely(atomic_read(&page->_mapcount) != -1))
873 bad_reason = "nonzero mapcount";
874 if (unlikely(page->mapping != NULL))
875 bad_reason = "non-NULL mapping";
876 if (unlikely(page_ref_count(page) != 0))
877 bad_reason = "nonzero _refcount";
878 if (unlikely(page->flags & flags)) {
879 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
880 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
882 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
885 if (unlikely(page->memcg_data))
886 bad_reason = "page still charged to cgroup";
888 #ifdef CONFIG_PAGE_POOL
889 if (unlikely((page->pp_magic & ~0x3UL) == PP_SIGNATURE))
890 bad_reason = "page_pool leak";
895 static void free_page_is_bad_report(struct page *page)
898 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
901 static inline bool free_page_is_bad(struct page *page)
903 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
906 /* Something has gone sideways, find it */
907 free_page_is_bad_report(page);
911 static inline bool is_check_pages_enabled(void)
913 return static_branch_unlikely(&check_pages_enabled);
916 static int free_tail_page_prepare(struct page *head_page, struct page *page)
918 struct folio *folio = (struct folio *)head_page;
922 * We rely page->lru.next never has bit 0 set, unless the page
923 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
925 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
927 if (!is_check_pages_enabled()) {
931 switch (page - head_page) {
933 /* the first tail page: these may be in place of ->mapping */
934 if (unlikely(folio_entire_mapcount(folio))) {
935 bad_page(page, "nonzero entire_mapcount");
938 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
939 bad_page(page, "nonzero nr_pages_mapped");
942 if (unlikely(atomic_read(&folio->_pincount))) {
943 bad_page(page, "nonzero pincount");
948 /* the second tail page: deferred_list overlaps ->mapping */
949 if (unlikely(!list_empty(&folio->_deferred_list))) {
950 bad_page(page, "on deferred list");
955 if (page->mapping != TAIL_MAPPING) {
956 bad_page(page, "corrupted mapping in tail page");
961 if (unlikely(!PageTail(page))) {
962 bad_page(page, "PageTail not set");
965 if (unlikely(compound_head(page) != head_page)) {
966 bad_page(page, "compound_head not consistent");
971 page->mapping = NULL;
972 clear_compound_head(page);
977 * Skip KASAN memory poisoning when either:
979 * 1. For generic KASAN: deferred memory initialization has not yet completed.
980 * Tag-based KASAN modes skip pages freed via deferred memory initialization
981 * using page tags instead (see below).
982 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
983 * that error detection is disabled for accesses via the page address.
985 * Pages will have match-all tags in the following circumstances:
987 * 1. Pages are being initialized for the first time, including during deferred
988 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
989 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
990 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
991 * 3. The allocation was excluded from being checked due to sampling,
992 * see the call to kasan_unpoison_pages.
994 * Poisoning pages during deferred memory init will greatly lengthen the
995 * process and cause problem in large memory systems as the deferred pages
996 * initialization is done with interrupt disabled.
998 * Assuming that there will be no reference to those newly initialized
999 * pages before they are ever allocated, this should have no effect on
1000 * KASAN memory tracking as the poison will be properly inserted at page
1001 * allocation time. The only corner case is when pages are allocated by
1002 * on-demand allocation and then freed again before the deferred pages
1003 * initialization is done, but this is not likely to happen.
1005 static inline bool should_skip_kasan_poison(struct page *page)
1007 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1008 return deferred_pages_enabled();
1010 return page_kasan_tag(page) == KASAN_TAG_KERNEL;
1013 void kernel_init_pages(struct page *page, int numpages)
1017 /* s390's use of memset() could override KASAN redzones. */
1018 kasan_disable_current();
1019 for (i = 0; i < numpages; i++)
1020 clear_highpage_kasan_tagged(page + i);
1021 kasan_enable_current();
1024 __always_inline bool free_pages_prepare(struct page *page,
1028 bool skip_kasan_poison = should_skip_kasan_poison(page);
1029 bool init = want_init_on_free();
1030 bool compound = PageCompound(page);
1032 VM_BUG_ON_PAGE(PageTail(page), page);
1034 trace_mm_page_free(page, order);
1035 kmsan_free_page(page, order);
1037 if (memcg_kmem_online() && PageMemcgKmem(page))
1038 __memcg_kmem_uncharge_page(page, order);
1040 if (unlikely(PageHWPoison(page)) && !order) {
1041 /* Do not let hwpoison pages hit pcplists/buddy */
1042 reset_page_owner(page, order);
1043 page_table_check_free(page, order);
1044 pgalloc_tag_sub(page, 1 << order);
1048 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1051 * Check tail pages before head page information is cleared to
1052 * avoid checking PageCompound for order-0 pages.
1054 if (unlikely(order)) {
1058 page[1].flags &= ~PAGE_FLAGS_SECOND;
1059 for (i = 1; i < (1 << order); i++) {
1061 bad += free_tail_page_prepare(page, page + i);
1062 if (is_check_pages_enabled()) {
1063 if (free_page_is_bad(page + i)) {
1068 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1071 if (PageMappingFlags(page))
1072 page->mapping = NULL;
1073 if (is_check_pages_enabled()) {
1074 if (free_page_is_bad(page))
1080 page_cpupid_reset_last(page);
1081 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1082 reset_page_owner(page, order);
1083 page_table_check_free(page, order);
1084 pgalloc_tag_sub(page, 1 << order);
1086 if (!PageHighMem(page)) {
1087 debug_check_no_locks_freed(page_address(page),
1088 PAGE_SIZE << order);
1089 debug_check_no_obj_freed(page_address(page),
1090 PAGE_SIZE << order);
1093 kernel_poison_pages(page, 1 << order);
1096 * As memory initialization might be integrated into KASAN,
1097 * KASAN poisoning and memory initialization code must be
1098 * kept together to avoid discrepancies in behavior.
1100 * With hardware tag-based KASAN, memory tags must be set before the
1101 * page becomes unavailable via debug_pagealloc or arch_free_page.
1103 if (!skip_kasan_poison) {
1104 kasan_poison_pages(page, order, init);
1106 /* Memory is already initialized if KASAN did it internally. */
1107 if (kasan_has_integrated_init())
1111 kernel_init_pages(page, 1 << order);
1114 * arch_free_page() can make the page's contents inaccessible. s390
1115 * does this. So nothing which can access the page's contents should
1116 * happen after this.
1118 arch_free_page(page, order);
1120 debug_pagealloc_unmap_pages(page, 1 << order);
1126 * Frees a number of pages from the PCP lists
1127 * Assumes all pages on list are in same zone.
1128 * count is the number of pages to free.
1130 static void free_pcppages_bulk(struct zone *zone, int count,
1131 struct per_cpu_pages *pcp,
1134 unsigned long flags;
1139 * Ensure proper count is passed which otherwise would stuck in the
1140 * below while (list_empty(list)) loop.
1142 count = min(pcp->count, count);
1144 /* Ensure requested pindex is drained first. */
1145 pindex = pindex - 1;
1147 spin_lock_irqsave(&zone->lock, flags);
1150 struct list_head *list;
1153 /* Remove pages from lists in a round-robin fashion. */
1155 if (++pindex > NR_PCP_LISTS - 1)
1157 list = &pcp->lists[pindex];
1158 } while (list_empty(list));
1160 order = pindex_to_order(pindex);
1161 nr_pages = 1 << order;
1166 page = list_last_entry(list, struct page, pcp_list);
1167 pfn = page_to_pfn(page);
1168 mt = get_pfnblock_migratetype(page, pfn);
1170 /* must delete to avoid corrupting pcp list */
1171 list_del(&page->pcp_list);
1173 pcp->count -= nr_pages;
1175 __free_one_page(page, pfn, zone, order, mt, FPI_NONE);
1176 trace_mm_page_pcpu_drain(page, order, mt);
1177 } while (count > 0 && !list_empty(list));
1180 spin_unlock_irqrestore(&zone->lock, flags);
1183 static void free_one_page(struct zone *zone, struct page *page,
1184 unsigned long pfn, unsigned int order,
1187 unsigned long flags;
1190 spin_lock_irqsave(&zone->lock, flags);
1191 migratetype = get_pfnblock_migratetype(page, pfn);
1192 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1193 spin_unlock_irqrestore(&zone->lock, flags);
1196 static void __free_pages_ok(struct page *page, unsigned int order,
1199 unsigned long pfn = page_to_pfn(page);
1200 struct zone *zone = page_zone(page);
1202 if (!free_pages_prepare(page, order))
1205 free_one_page(zone, page, pfn, order, fpi_flags);
1207 __count_vm_events(PGFREE, 1 << order);
1210 void __free_pages_core(struct page *page, unsigned int order)
1212 unsigned int nr_pages = 1 << order;
1213 struct page *p = page;
1217 * When initializing the memmap, __init_single_page() sets the refcount
1218 * of all pages to 1 ("allocated"/"not free"). We have to set the
1219 * refcount of all involved pages to 0.
1222 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1224 __ClearPageReserved(p);
1225 set_page_count(p, 0);
1227 __ClearPageReserved(p);
1228 set_page_count(p, 0);
1230 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1232 if (page_contains_unaccepted(page, order)) {
1233 if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1236 accept_page(page, order);
1240 * Bypass PCP and place fresh pages right to the tail, primarily
1241 * relevant for memory onlining.
1243 __free_pages_ok(page, order, FPI_TO_TAIL);
1247 * Check that the whole (or subset of) a pageblock given by the interval of
1248 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1249 * with the migration of free compaction scanner.
1251 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1253 * It's possible on some configurations to have a setup like node0 node1 node0
1254 * i.e. it's possible that all pages within a zones range of pages do not
1255 * belong to a single zone. We assume that a border between node0 and node1
1256 * can occur within a single pageblock, but not a node0 node1 node0
1257 * interleaving within a single pageblock. It is therefore sufficient to check
1258 * the first and last page of a pageblock and avoid checking each individual
1259 * page in a pageblock.
1261 * Note: the function may return non-NULL struct page even for a page block
1262 * which contains a memory hole (i.e. there is no physical memory for a subset
1263 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1264 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1265 * even though the start pfn is online and valid. This should be safe most of
1266 * the time because struct pages are still initialized via init_unavailable_range()
1267 * and pfn walkers shouldn't touch any physical memory range for which they do
1268 * not recognize any specific metadata in struct pages.
1270 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1271 unsigned long end_pfn, struct zone *zone)
1273 struct page *start_page;
1274 struct page *end_page;
1276 /* end_pfn is one past the range we are checking */
1279 if (!pfn_valid(end_pfn))
1282 start_page = pfn_to_online_page(start_pfn);
1286 if (page_zone(start_page) != zone)
1289 end_page = pfn_to_page(end_pfn);
1291 /* This gives a shorter code than deriving page_zone(end_page) */
1292 if (page_zone_id(start_page) != page_zone_id(end_page))
1299 * The order of subdivision here is critical for the IO subsystem.
1300 * Please do not alter this order without good reasons and regression
1301 * testing. Specifically, as large blocks of memory are subdivided,
1302 * the order in which smaller blocks are delivered depends on the order
1303 * they're subdivided in this function. This is the primary factor
1304 * influencing the order in which pages are delivered to the IO
1305 * subsystem according to empirical testing, and this is also justified
1306 * by considering the behavior of a buddy system containing a single
1307 * large block of memory acted on by a series of small allocations.
1308 * This behavior is a critical factor in sglist merging's success.
1312 static inline void expand(struct zone *zone, struct page *page,
1313 int low, int high, int migratetype)
1315 unsigned long size = 1 << high;
1316 unsigned long nr_added = 0;
1318 while (high > low) {
1321 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1324 * Mark as guard pages (or page), that will allow to
1325 * merge back to allocator when buddy will be freed.
1326 * Corresponding page table entries will not be touched,
1327 * pages will stay not present in virtual address space
1329 if (set_page_guard(zone, &page[size], high))
1332 __add_to_free_list(&page[size], zone, high, migratetype, false);
1333 set_buddy_order(&page[size], high);
1336 account_freepages(zone, nr_added, migratetype);
1339 static void check_new_page_bad(struct page *page)
1341 if (unlikely(page->flags & __PG_HWPOISON)) {
1342 /* Don't complain about hwpoisoned pages */
1343 page_mapcount_reset(page); /* remove PageBuddy */
1348 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1352 * This page is about to be returned from the page allocator
1354 static bool check_new_page(struct page *page)
1356 if (likely(page_expected_state(page,
1357 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1360 check_new_page_bad(page);
1364 static inline bool check_new_pages(struct page *page, unsigned int order)
1366 if (is_check_pages_enabled()) {
1367 for (int i = 0; i < (1 << order); i++) {
1368 struct page *p = page + i;
1370 if (check_new_page(p))
1378 static inline bool should_skip_kasan_unpoison(gfp_t flags)
1380 /* Don't skip if a software KASAN mode is enabled. */
1381 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1382 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1385 /* Skip, if hardware tag-based KASAN is not enabled. */
1386 if (!kasan_hw_tags_enabled())
1390 * With hardware tag-based KASAN enabled, skip if this has been
1391 * requested via __GFP_SKIP_KASAN.
1393 return flags & __GFP_SKIP_KASAN;
1396 static inline bool should_skip_init(gfp_t flags)
1398 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1399 if (!kasan_hw_tags_enabled())
1402 /* For hardware tag-based KASAN, skip if requested. */
1403 return (flags & __GFP_SKIP_ZERO);
1406 inline void post_alloc_hook(struct page *page, unsigned int order,
1409 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1410 !should_skip_init(gfp_flags);
1411 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1414 set_page_private(page, 0);
1415 set_page_refcounted(page);
1417 arch_alloc_page(page, order);
1418 debug_pagealloc_map_pages(page, 1 << order);
1421 * Page unpoisoning must happen before memory initialization.
1422 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1423 * allocations and the page unpoisoning code will complain.
1425 kernel_unpoison_pages(page, 1 << order);
1428 * As memory initialization might be integrated into KASAN,
1429 * KASAN unpoisoning and memory initializion code must be
1430 * kept together to avoid discrepancies in behavior.
1434 * If memory tags should be zeroed
1435 * (which happens only when memory should be initialized as well).
1438 /* Initialize both memory and memory tags. */
1439 for (i = 0; i != 1 << order; ++i)
1440 tag_clear_highpage(page + i);
1442 /* Take note that memory was initialized by the loop above. */
1445 if (!should_skip_kasan_unpoison(gfp_flags) &&
1446 kasan_unpoison_pages(page, order, init)) {
1447 /* Take note that memory was initialized by KASAN. */
1448 if (kasan_has_integrated_init())
1452 * If memory tags have not been set by KASAN, reset the page
1453 * tags to ensure page_address() dereferencing does not fault.
1455 for (i = 0; i != 1 << order; ++i)
1456 page_kasan_tag_reset(page + i);
1458 /* If memory is still not initialized, initialize it now. */
1460 kernel_init_pages(page, 1 << order);
1462 set_page_owner(page, order, gfp_flags);
1463 page_table_check_alloc(page, order);
1464 pgalloc_tag_add(page, current, 1 << order);
1467 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1468 unsigned int alloc_flags)
1470 post_alloc_hook(page, order, gfp_flags);
1472 if (order && (gfp_flags & __GFP_COMP))
1473 prep_compound_page(page, order);
1476 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1477 * allocate the page. The expectation is that the caller is taking
1478 * steps that will free more memory. The caller should avoid the page
1479 * being used for !PFMEMALLOC purposes.
1481 if (alloc_flags & ALLOC_NO_WATERMARKS)
1482 set_page_pfmemalloc(page);
1484 clear_page_pfmemalloc(page);
1488 * Go through the free lists for the given migratetype and remove
1489 * the smallest available page from the freelists
1491 static __always_inline
1492 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1495 unsigned int current_order;
1496 struct free_area *area;
1499 /* Find a page of the appropriate size in the preferred list */
1500 for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
1501 area = &(zone->free_area[current_order]);
1502 page = get_page_from_free_area(area, migratetype);
1505 del_page_from_free_list(page, zone, current_order, migratetype);
1506 expand(zone, page, order, current_order, migratetype);
1507 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1508 pcp_allowed_order(order) &&
1509 migratetype < MIGRATE_PCPTYPES);
1518 * This array describes the order lists are fallen back to when
1519 * the free lists for the desirable migrate type are depleted
1521 * The other migratetypes do not have fallbacks.
1523 static int fallbacks[MIGRATE_PCPTYPES][MIGRATE_PCPTYPES - 1] = {
1524 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1525 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1526 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1530 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1533 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1536 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1537 unsigned int order) { return NULL; }
1541 * Change the type of a block and move all its free pages to that
1544 static int __move_freepages_block(struct zone *zone, unsigned long start_pfn,
1545 int old_mt, int new_mt)
1548 unsigned long pfn, end_pfn;
1550 int pages_moved = 0;
1552 VM_WARN_ON(start_pfn & (pageblock_nr_pages - 1));
1553 end_pfn = pageblock_end_pfn(start_pfn);
1555 for (pfn = start_pfn; pfn < end_pfn;) {
1556 page = pfn_to_page(pfn);
1557 if (!PageBuddy(page)) {
1562 /* Make sure we are not inadvertently changing nodes */
1563 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1564 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1566 order = buddy_order(page);
1568 move_to_free_list(page, zone, order, old_mt, new_mt);
1571 pages_moved += 1 << order;
1574 set_pageblock_migratetype(pfn_to_page(start_pfn), new_mt);
1579 static bool prep_move_freepages_block(struct zone *zone, struct page *page,
1580 unsigned long *start_pfn,
1581 int *num_free, int *num_movable)
1583 unsigned long pfn, start, end;
1585 pfn = page_to_pfn(page);
1586 start = pageblock_start_pfn(pfn);
1587 end = pageblock_end_pfn(pfn);
1590 * The caller only has the lock for @zone, don't touch ranges
1591 * that straddle into other zones. While we could move part of
1592 * the range that's inside the zone, this call is usually
1593 * accompanied by other operations such as migratetype updates
1594 * which also should be locked.
1596 if (!zone_spans_pfn(zone, start))
1598 if (!zone_spans_pfn(zone, end - 1))
1606 for (pfn = start; pfn < end;) {
1607 page = pfn_to_page(pfn);
1608 if (PageBuddy(page)) {
1609 int nr = 1 << buddy_order(page);
1616 * We assume that pages that could be isolated for
1617 * migration are movable. But we don't actually try
1618 * isolating, as that would be expensive.
1620 if (PageLRU(page) || __PageMovable(page))
1629 static int move_freepages_block(struct zone *zone, struct page *page,
1630 int old_mt, int new_mt)
1632 unsigned long start_pfn;
1634 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
1637 return __move_freepages_block(zone, start_pfn, old_mt, new_mt);
1640 #ifdef CONFIG_MEMORY_ISOLATION
1641 /* Look for a buddy that straddles start_pfn */
1642 static unsigned long find_large_buddy(unsigned long start_pfn)
1646 unsigned long pfn = start_pfn;
1648 while (!PageBuddy(page = pfn_to_page(pfn))) {
1650 if (++order > MAX_PAGE_ORDER)
1652 pfn &= ~0UL << order;
1656 * Found a preceding buddy, but does it straddle?
1658 if (pfn + (1 << buddy_order(page)) > start_pfn)
1665 /* Split a multi-block free page into its individual pageblocks */
1666 static void split_large_buddy(struct zone *zone, struct page *page,
1667 unsigned long pfn, int order)
1669 unsigned long end_pfn = pfn + (1 << order);
1671 VM_WARN_ON_ONCE(order <= pageblock_order);
1672 VM_WARN_ON_ONCE(pfn & (pageblock_nr_pages - 1));
1674 /* Caller removed page from freelist, buddy info cleared! */
1675 VM_WARN_ON_ONCE(PageBuddy(page));
1677 while (pfn != end_pfn) {
1678 int mt = get_pfnblock_migratetype(page, pfn);
1680 __free_one_page(page, pfn, zone, pageblock_order, mt, FPI_NONE);
1681 pfn += pageblock_nr_pages;
1682 page = pfn_to_page(pfn);
1687 * move_freepages_block_isolate - move free pages in block for page isolation
1689 * @page: the pageblock page
1690 * @migratetype: migratetype to set on the pageblock
1692 * This is similar to move_freepages_block(), but handles the special
1693 * case encountered in page isolation, where the block of interest
1694 * might be part of a larger buddy spanning multiple pageblocks.
1696 * Unlike the regular page allocator path, which moves pages while
1697 * stealing buddies off the freelist, page isolation is interested in
1698 * arbitrary pfn ranges that may have overlapping buddies on both ends.
1700 * This function handles that. Straddling buddies are split into
1701 * individual pageblocks. Only the block of interest is moved.
1703 * Returns %true if pages could be moved, %false otherwise.
1705 bool move_freepages_block_isolate(struct zone *zone, struct page *page,
1708 unsigned long start_pfn, pfn;
1710 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
1713 /* No splits needed if buddies can't span multiple blocks */
1714 if (pageblock_order == MAX_PAGE_ORDER)
1717 /* We're a tail block in a larger buddy */
1718 pfn = find_large_buddy(start_pfn);
1719 if (pfn != start_pfn) {
1720 struct page *buddy = pfn_to_page(pfn);
1721 int order = buddy_order(buddy);
1723 del_page_from_free_list(buddy, zone, order,
1724 get_pfnblock_migratetype(buddy, pfn));
1725 set_pageblock_migratetype(page, migratetype);
1726 split_large_buddy(zone, buddy, pfn, order);
1730 /* We're the starting block of a larger buddy */
1731 if (PageBuddy(page) && buddy_order(page) > pageblock_order) {
1732 int order = buddy_order(page);
1734 del_page_from_free_list(page, zone, order,
1735 get_pfnblock_migratetype(page, pfn));
1736 set_pageblock_migratetype(page, migratetype);
1737 split_large_buddy(zone, page, pfn, order);
1741 __move_freepages_block(zone, start_pfn,
1742 get_pfnblock_migratetype(page, start_pfn),
1746 #endif /* CONFIG_MEMORY_ISOLATION */
1748 static void change_pageblock_range(struct page *pageblock_page,
1749 int start_order, int migratetype)
1751 int nr_pageblocks = 1 << (start_order - pageblock_order);
1753 while (nr_pageblocks--) {
1754 set_pageblock_migratetype(pageblock_page, migratetype);
1755 pageblock_page += pageblock_nr_pages;
1760 * When we are falling back to another migratetype during allocation, try to
1761 * steal extra free pages from the same pageblocks to satisfy further
1762 * allocations, instead of polluting multiple pageblocks.
1764 * If we are stealing a relatively large buddy page, it is likely there will
1765 * be more free pages in the pageblock, so try to steal them all. For
1766 * reclaimable and unmovable allocations, we steal regardless of page size,
1767 * as fragmentation caused by those allocations polluting movable pageblocks
1768 * is worse than movable allocations stealing from unmovable and reclaimable
1771 static bool can_steal_fallback(unsigned int order, int start_mt)
1774 * Leaving this order check is intended, although there is
1775 * relaxed order check in next check. The reason is that
1776 * we can actually steal whole pageblock if this condition met,
1777 * but, below check doesn't guarantee it and that is just heuristic
1778 * so could be changed anytime.
1780 if (order >= pageblock_order)
1783 if (order >= pageblock_order / 2 ||
1784 start_mt == MIGRATE_RECLAIMABLE ||
1785 start_mt == MIGRATE_UNMOVABLE ||
1786 page_group_by_mobility_disabled)
1792 static inline bool boost_watermark(struct zone *zone)
1794 unsigned long max_boost;
1796 if (!watermark_boost_factor)
1799 * Don't bother in zones that are unlikely to produce results.
1800 * On small machines, including kdump capture kernels running
1801 * in a small area, boosting the watermark can cause an out of
1802 * memory situation immediately.
1804 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1807 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1808 watermark_boost_factor, 10000);
1811 * high watermark may be uninitialised if fragmentation occurs
1812 * very early in boot so do not boost. We do not fall
1813 * through and boost by pageblock_nr_pages as failing
1814 * allocations that early means that reclaim is not going
1815 * to help and it may even be impossible to reclaim the
1816 * boosted watermark resulting in a hang.
1821 max_boost = max(pageblock_nr_pages, max_boost);
1823 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1830 * This function implements actual steal behaviour. If order is large enough, we
1831 * can claim the whole pageblock for the requested migratetype. If not, we check
1832 * the pageblock for constituent pages; if at least half of the pages are free
1833 * or compatible, we can still claim the whole block, so pages freed in the
1834 * future will be put on the correct free list. Otherwise, we isolate exactly
1835 * the order we need from the fallback block and leave its migratetype alone.
1837 static struct page *
1838 steal_suitable_fallback(struct zone *zone, struct page *page,
1839 int current_order, int order, int start_type,
1840 unsigned int alloc_flags, bool whole_block)
1842 int free_pages, movable_pages, alike_pages;
1843 unsigned long start_pfn;
1846 block_type = get_pageblock_migratetype(page);
1849 * This can happen due to races and we want to prevent broken
1850 * highatomic accounting.
1852 if (is_migrate_highatomic(block_type))
1855 /* Take ownership for orders >= pageblock_order */
1856 if (current_order >= pageblock_order) {
1857 del_page_from_free_list(page, zone, current_order, block_type);
1858 change_pageblock_range(page, current_order, start_type);
1859 expand(zone, page, order, current_order, start_type);
1864 * Boost watermarks to increase reclaim pressure to reduce the
1865 * likelihood of future fallbacks. Wake kswapd now as the node
1866 * may be balanced overall and kswapd will not wake naturally.
1868 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1869 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1871 /* We are not allowed to try stealing from the whole block */
1875 /* moving whole block can fail due to zone boundary conditions */
1876 if (!prep_move_freepages_block(zone, page, &start_pfn, &free_pages,
1881 * Determine how many pages are compatible with our allocation.
1882 * For movable allocation, it's the number of movable pages which
1883 * we just obtained. For other types it's a bit more tricky.
1885 if (start_type == MIGRATE_MOVABLE) {
1886 alike_pages = movable_pages;
1889 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1890 * to MOVABLE pageblock, consider all non-movable pages as
1891 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1892 * vice versa, be conservative since we can't distinguish the
1893 * exact migratetype of non-movable pages.
1895 if (block_type == MIGRATE_MOVABLE)
1896 alike_pages = pageblock_nr_pages
1897 - (free_pages + movable_pages);
1902 * If a sufficient number of pages in the block are either free or of
1903 * compatible migratability as our allocation, claim the whole block.
1905 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1906 page_group_by_mobility_disabled) {
1907 __move_freepages_block(zone, start_pfn, block_type, start_type);
1908 return __rmqueue_smallest(zone, order, start_type);
1912 del_page_from_free_list(page, zone, current_order, block_type);
1913 expand(zone, page, order, current_order, block_type);
1918 * Check whether there is a suitable fallback freepage with requested order.
1919 * If only_stealable is true, this function returns fallback_mt only if
1920 * we can steal other freepages all together. This would help to reduce
1921 * fragmentation due to mixed migratetype pages in one pageblock.
1923 int find_suitable_fallback(struct free_area *area, unsigned int order,
1924 int migratetype, bool only_stealable, bool *can_steal)
1929 if (area->nr_free == 0)
1933 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
1934 fallback_mt = fallbacks[migratetype][i];
1935 if (free_area_empty(area, fallback_mt))
1938 if (can_steal_fallback(order, migratetype))
1941 if (!only_stealable)
1952 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1953 * there are no empty page blocks that contain a page with a suitable order
1955 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone)
1958 unsigned long max_managed, flags;
1961 * The number reserved as: minimum is 1 pageblock, maximum is
1962 * roughly 1% of a zone. But if 1% of a zone falls below a
1963 * pageblock size, then don't reserve any pageblocks.
1964 * Check is race-prone but harmless.
1966 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
1968 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
1969 if (zone->nr_reserved_highatomic >= max_managed)
1972 spin_lock_irqsave(&zone->lock, flags);
1974 /* Recheck the nr_reserved_highatomic limit under the lock */
1975 if (zone->nr_reserved_highatomic >= max_managed)
1979 mt = get_pageblock_migratetype(page);
1980 /* Only reserve normal pageblocks (i.e., they can merge with others) */
1981 if (migratetype_is_mergeable(mt))
1982 if (move_freepages_block(zone, page, mt,
1983 MIGRATE_HIGHATOMIC) != -1)
1984 zone->nr_reserved_highatomic += pageblock_nr_pages;
1987 spin_unlock_irqrestore(&zone->lock, flags);
1991 * Used when an allocation is about to fail under memory pressure. This
1992 * potentially hurts the reliability of high-order allocations when under
1993 * intense memory pressure but failed atomic allocations should be easier
1994 * to recover from than an OOM.
1996 * If @force is true, try to unreserve a pageblock even though highatomic
1997 * pageblock is exhausted.
1999 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2002 struct zonelist *zonelist = ac->zonelist;
2003 unsigned long flags;
2010 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2013 * Preserve at least one pageblock unless memory pressure
2016 if (!force && zone->nr_reserved_highatomic <=
2020 spin_lock_irqsave(&zone->lock, flags);
2021 for (order = 0; order < NR_PAGE_ORDERS; order++) {
2022 struct free_area *area = &(zone->free_area[order]);
2025 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2029 mt = get_pageblock_migratetype(page);
2031 * In page freeing path, migratetype change is racy so
2032 * we can counter several free pages in a pageblock
2033 * in this loop although we changed the pageblock type
2034 * from highatomic to ac->migratetype. So we should
2035 * adjust the count once.
2037 if (is_migrate_highatomic(mt)) {
2039 * It should never happen but changes to
2040 * locking could inadvertently allow a per-cpu
2041 * drain to add pages to MIGRATE_HIGHATOMIC
2042 * while unreserving so be safe and watch for
2045 zone->nr_reserved_highatomic -= min(
2047 zone->nr_reserved_highatomic);
2051 * Convert to ac->migratetype and avoid the normal
2052 * pageblock stealing heuristics. Minimally, the caller
2053 * is doing the work and needs the pages. More
2054 * importantly, if the block was always converted to
2055 * MIGRATE_UNMOVABLE or another type then the number
2056 * of pageblocks that cannot be completely freed
2059 ret = move_freepages_block(zone, page, mt,
2062 * Reserving this block already succeeded, so this should
2063 * not fail on zone boundaries.
2065 WARN_ON_ONCE(ret == -1);
2067 spin_unlock_irqrestore(&zone->lock, flags);
2071 spin_unlock_irqrestore(&zone->lock, flags);
2078 * Try finding a free buddy page on the fallback list and put it on the free
2079 * list of requested migratetype, possibly along with other pages from the same
2080 * block, depending on fragmentation avoidance heuristics. Returns true if
2081 * fallback was found so that __rmqueue_smallest() can grab it.
2083 * The use of signed ints for order and current_order is a deliberate
2084 * deviation from the rest of this file, to make the for loop
2085 * condition simpler.
2087 static __always_inline struct page *
2088 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2089 unsigned int alloc_flags)
2091 struct free_area *area;
2093 int min_order = order;
2099 * Do not steal pages from freelists belonging to other pageblocks
2100 * i.e. orders < pageblock_order. If there are no local zones free,
2101 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2103 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2104 min_order = pageblock_order;
2107 * Find the largest available free page in the other list. This roughly
2108 * approximates finding the pageblock with the most free pages, which
2109 * would be too costly to do exactly.
2111 for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2113 area = &(zone->free_area[current_order]);
2114 fallback_mt = find_suitable_fallback(area, current_order,
2115 start_migratetype, false, &can_steal);
2116 if (fallback_mt == -1)
2120 * We cannot steal all free pages from the pageblock and the
2121 * requested migratetype is movable. In that case it's better to
2122 * steal and split the smallest available page instead of the
2123 * largest available page, because even if the next movable
2124 * allocation falls back into a different pageblock than this
2125 * one, it won't cause permanent fragmentation.
2127 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2128 && current_order > order)
2137 for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
2138 area = &(zone->free_area[current_order]);
2139 fallback_mt = find_suitable_fallback(area, current_order,
2140 start_migratetype, false, &can_steal);
2141 if (fallback_mt != -1)
2146 * This should not happen - we already found a suitable fallback
2147 * when looking for the largest page.
2149 VM_BUG_ON(current_order > MAX_PAGE_ORDER);
2152 page = get_page_from_free_area(area, fallback_mt);
2154 /* take off list, maybe claim block, expand remainder */
2155 page = steal_suitable_fallback(zone, page, current_order, order,
2156 start_migratetype, alloc_flags, can_steal);
2158 trace_mm_page_alloc_extfrag(page, order, current_order,
2159 start_migratetype, fallback_mt);
2165 * Do the hard work of removing an element from the buddy allocator.
2166 * Call me with the zone->lock already held.
2168 static __always_inline struct page *
2169 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2170 unsigned int alloc_flags)
2174 if (IS_ENABLED(CONFIG_CMA)) {
2176 * Balance movable allocations between regular and CMA areas by
2177 * allocating from CMA when over half of the zone's free memory
2178 * is in the CMA area.
2180 if (alloc_flags & ALLOC_CMA &&
2181 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2182 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2183 page = __rmqueue_cma_fallback(zone, order);
2189 page = __rmqueue_smallest(zone, order, migratetype);
2190 if (unlikely(!page)) {
2191 if (alloc_flags & ALLOC_CMA)
2192 page = __rmqueue_cma_fallback(zone, order);
2195 page = __rmqueue_fallback(zone, order, migratetype,
2202 * Obtain a specified number of elements from the buddy allocator, all under
2203 * a single hold of the lock, for efficiency. Add them to the supplied list.
2204 * Returns the number of new pages which were placed at *list.
2206 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2207 unsigned long count, struct list_head *list,
2208 int migratetype, unsigned int alloc_flags)
2210 unsigned long flags;
2213 spin_lock_irqsave(&zone->lock, flags);
2214 for (i = 0; i < count; ++i) {
2215 struct page *page = __rmqueue(zone, order, migratetype,
2217 if (unlikely(page == NULL))
2221 * Split buddy pages returned by expand() are received here in
2222 * physical page order. The page is added to the tail of
2223 * caller's list. From the callers perspective, the linked list
2224 * is ordered by page number under some conditions. This is
2225 * useful for IO devices that can forward direction from the
2226 * head, thus also in the physical page order. This is useful
2227 * for IO devices that can merge IO requests if the physical
2228 * pages are ordered properly.
2230 list_add_tail(&page->pcp_list, list);
2232 spin_unlock_irqrestore(&zone->lock, flags);
2238 * Called from the vmstat counter updater to decay the PCP high.
2239 * Return whether there are addition works to do.
2241 int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2243 int high_min, to_drain, batch;
2246 high_min = READ_ONCE(pcp->high_min);
2247 batch = READ_ONCE(pcp->batch);
2249 * Decrease pcp->high periodically to try to free possible
2250 * idle PCP pages. And, avoid to free too many pages to
2251 * control latency. This caps pcp->high decrement too.
2253 if (pcp->high > high_min) {
2254 pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2255 pcp->high - (pcp->high >> 3), high_min);
2256 if (pcp->high > high_min)
2260 to_drain = pcp->count - pcp->high;
2262 spin_lock(&pcp->lock);
2263 free_pcppages_bulk(zone, to_drain, pcp, 0);
2264 spin_unlock(&pcp->lock);
2273 * Called from the vmstat counter updater to drain pagesets of this
2274 * currently executing processor on remote nodes after they have
2277 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2279 int to_drain, batch;
2281 batch = READ_ONCE(pcp->batch);
2282 to_drain = min(pcp->count, batch);
2284 spin_lock(&pcp->lock);
2285 free_pcppages_bulk(zone, to_drain, pcp, 0);
2286 spin_unlock(&pcp->lock);
2292 * Drain pcplists of the indicated processor and zone.
2294 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2296 struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2297 int count = READ_ONCE(pcp->count);
2300 int to_drain = min(count, pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
2303 spin_lock(&pcp->lock);
2304 free_pcppages_bulk(zone, to_drain, pcp, 0);
2305 spin_unlock(&pcp->lock);
2310 * Drain pcplists of all zones on the indicated processor.
2312 static void drain_pages(unsigned int cpu)
2316 for_each_populated_zone(zone) {
2317 drain_pages_zone(cpu, zone);
2322 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2324 void drain_local_pages(struct zone *zone)
2326 int cpu = smp_processor_id();
2329 drain_pages_zone(cpu, zone);
2335 * The implementation of drain_all_pages(), exposing an extra parameter to
2336 * drain on all cpus.
2338 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2339 * not empty. The check for non-emptiness can however race with a free to
2340 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2341 * that need the guarantee that every CPU has drained can disable the
2342 * optimizing racy check.
2344 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2349 * Allocate in the BSS so we won't require allocation in
2350 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2352 static cpumask_t cpus_with_pcps;
2355 * Do not drain if one is already in progress unless it's specific to
2356 * a zone. Such callers are primarily CMA and memory hotplug and need
2357 * the drain to be complete when the call returns.
2359 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2362 mutex_lock(&pcpu_drain_mutex);
2366 * We don't care about racing with CPU hotplug event
2367 * as offline notification will cause the notified
2368 * cpu to drain that CPU pcps and on_each_cpu_mask
2369 * disables preemption as part of its processing
2371 for_each_online_cpu(cpu) {
2372 struct per_cpu_pages *pcp;
2374 bool has_pcps = false;
2376 if (force_all_cpus) {
2378 * The pcp.count check is racy, some callers need a
2379 * guarantee that no cpu is missed.
2383 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2387 for_each_populated_zone(z) {
2388 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2397 cpumask_set_cpu(cpu, &cpus_with_pcps);
2399 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2402 for_each_cpu(cpu, &cpus_with_pcps) {
2404 drain_pages_zone(cpu, zone);
2409 mutex_unlock(&pcpu_drain_mutex);
2413 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2415 * When zone parameter is non-NULL, spill just the single zone's pages.
2417 void drain_all_pages(struct zone *zone)
2419 __drain_all_pages(zone, false);
2422 static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2424 int min_nr_free, max_nr_free;
2426 /* Free as much as possible if batch freeing high-order pages. */
2427 if (unlikely(free_high))
2428 return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2430 /* Check for PCP disabled or boot pageset */
2431 if (unlikely(high < batch))
2434 /* Leave at least pcp->batch pages on the list */
2435 min_nr_free = batch;
2436 max_nr_free = high - batch;
2439 * Increase the batch number to the number of the consecutive
2440 * freed pages to reduce zone lock contention.
2442 batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2447 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2448 int batch, bool free_high)
2450 int high, high_min, high_max;
2452 high_min = READ_ONCE(pcp->high_min);
2453 high_max = READ_ONCE(pcp->high_max);
2454 high = pcp->high = clamp(pcp->high, high_min, high_max);
2456 if (unlikely(!high))
2459 if (unlikely(free_high)) {
2460 pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2466 * If reclaim is active, limit the number of pages that can be
2467 * stored on pcp lists
2469 if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2470 int free_count = max_t(int, pcp->free_count, batch);
2472 pcp->high = max(high - free_count, high_min);
2473 return min(batch << 2, pcp->high);
2476 if (high_min == high_max)
2479 if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2480 int free_count = max_t(int, pcp->free_count, batch);
2482 pcp->high = max(high - free_count, high_min);
2483 high = max(pcp->count, high_min);
2484 } else if (pcp->count >= high) {
2485 int need_high = pcp->free_count + batch;
2487 /* pcp->high should be large enough to hold batch freed pages */
2488 if (pcp->high < need_high)
2489 pcp->high = clamp(need_high, high_min, high_max);
2495 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2496 struct page *page, int migratetype,
2501 bool free_high = false;
2504 * On freeing, reduce the number of pages that are batch allocated.
2505 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2508 pcp->alloc_factor >>= 1;
2509 __count_vm_events(PGFREE, 1 << order);
2510 pindex = order_to_pindex(migratetype, order);
2511 list_add(&page->pcp_list, &pcp->lists[pindex]);
2512 pcp->count += 1 << order;
2514 batch = READ_ONCE(pcp->batch);
2516 * As high-order pages other than THP's stored on PCP can contribute
2517 * to fragmentation, limit the number stored when PCP is heavily
2518 * freeing without allocation. The remainder after bulk freeing
2519 * stops will be drained from vmstat refresh context.
2521 if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2522 free_high = (pcp->free_count >= batch &&
2523 (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2524 (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2525 pcp->count >= READ_ONCE(batch)));
2526 pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2527 } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2528 pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2530 if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2531 pcp->free_count += (1 << order);
2532 high = nr_pcp_high(pcp, zone, batch, free_high);
2533 if (pcp->count >= high) {
2534 free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high),
2536 if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2537 zone_watermark_ok(zone, 0, high_wmark_pages(zone),
2539 clear_bit(ZONE_BELOW_HIGH, &zone->flags);
2546 void free_unref_page(struct page *page, unsigned int order)
2548 unsigned long __maybe_unused UP_flags;
2549 struct per_cpu_pages *pcp;
2551 unsigned long pfn = page_to_pfn(page);
2554 if (!pcp_allowed_order(order)) {
2555 __free_pages_ok(page, order, FPI_NONE);
2559 if (!free_pages_prepare(page, order))
2563 * We only track unmovable, reclaimable and movable on pcp lists.
2564 * Place ISOLATE pages on the isolated list because they are being
2565 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2566 * get those areas back if necessary. Otherwise, we may have to free
2567 * excessively into the page allocator
2569 migratetype = get_pfnblock_migratetype(page, pfn);
2570 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2571 if (unlikely(is_migrate_isolate(migratetype))) {
2572 free_one_page(page_zone(page), page, pfn, order, FPI_NONE);
2575 migratetype = MIGRATE_MOVABLE;
2578 zone = page_zone(page);
2579 pcp_trylock_prepare(UP_flags);
2580 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2582 free_unref_page_commit(zone, pcp, page, migratetype, order);
2583 pcp_spin_unlock(pcp);
2585 free_one_page(zone, page, pfn, order, FPI_NONE);
2587 pcp_trylock_finish(UP_flags);
2591 * Free a batch of folios
2593 void free_unref_folios(struct folio_batch *folios)
2595 unsigned long __maybe_unused UP_flags;
2596 struct per_cpu_pages *pcp = NULL;
2597 struct zone *locked_zone = NULL;
2600 /* Prepare folios for freeing */
2601 for (i = 0, j = 0; i < folios->nr; i++) {
2602 struct folio *folio = folios->folios[i];
2603 unsigned long pfn = folio_pfn(folio);
2604 unsigned int order = folio_order(folio);
2606 if (order > 0 && folio_test_large_rmappable(folio))
2607 folio_undo_large_rmappable(folio);
2608 if (!free_pages_prepare(&folio->page, order))
2611 * Free orders not handled on the PCP directly to the
2614 if (!pcp_allowed_order(order)) {
2615 free_one_page(folio_zone(folio), &folio->page,
2616 pfn, order, FPI_NONE);
2619 folio->private = (void *)(unsigned long)order;
2621 folios->folios[j] = folio;
2626 for (i = 0; i < folios->nr; i++) {
2627 struct folio *folio = folios->folios[i];
2628 struct zone *zone = folio_zone(folio);
2629 unsigned long pfn = folio_pfn(folio);
2630 unsigned int order = (unsigned long)folio->private;
2633 folio->private = NULL;
2634 migratetype = get_pfnblock_migratetype(&folio->page, pfn);
2636 /* Different zone requires a different pcp lock */
2637 if (zone != locked_zone ||
2638 is_migrate_isolate(migratetype)) {
2640 pcp_spin_unlock(pcp);
2641 pcp_trylock_finish(UP_flags);
2647 * Free isolated pages directly to the
2648 * allocator, see comment in free_unref_page.
2650 if (is_migrate_isolate(migratetype)) {
2651 free_one_page(zone, &folio->page, pfn,
2657 * trylock is necessary as folios may be getting freed
2658 * from IRQ or SoftIRQ context after an IO completion.
2660 pcp_trylock_prepare(UP_flags);
2661 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2662 if (unlikely(!pcp)) {
2663 pcp_trylock_finish(UP_flags);
2664 free_one_page(zone, &folio->page, pfn,
2672 * Non-isolated types over MIGRATE_PCPTYPES get added
2673 * to the MIGRATE_MOVABLE pcp list.
2675 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2676 migratetype = MIGRATE_MOVABLE;
2678 trace_mm_page_free_batched(&folio->page);
2679 free_unref_page_commit(zone, pcp, &folio->page, migratetype,
2684 pcp_spin_unlock(pcp);
2685 pcp_trylock_finish(UP_flags);
2687 folio_batch_reinit(folios);
2691 * split_page takes a non-compound higher-order page, and splits it into
2692 * n (1<<order) sub-pages: page[0..n]
2693 * Each sub-page must be freed individually.
2695 * Note: this is probably too low level an operation for use in drivers.
2696 * Please consult with lkml before using this in your driver.
2698 void split_page(struct page *page, unsigned int order)
2702 VM_BUG_ON_PAGE(PageCompound(page), page);
2703 VM_BUG_ON_PAGE(!page_count(page), page);
2705 for (i = 1; i < (1 << order); i++)
2706 set_page_refcounted(page + i);
2707 split_page_owner(page, order, 0);
2708 pgalloc_tag_split(page, 1 << order);
2709 split_page_memcg(page, order, 0);
2711 EXPORT_SYMBOL_GPL(split_page);
2713 int __isolate_free_page(struct page *page, unsigned int order)
2715 struct zone *zone = page_zone(page);
2716 int mt = get_pageblock_migratetype(page);
2718 if (!is_migrate_isolate(mt)) {
2719 unsigned long watermark;
2721 * Obey watermarks as if the page was being allocated. We can
2722 * emulate a high-order watermark check with a raised order-0
2723 * watermark, because we already know our high-order page
2726 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2727 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2731 del_page_from_free_list(page, zone, order, mt);
2734 * Set the pageblock if the isolated page is at least half of a
2737 if (order >= pageblock_order - 1) {
2738 struct page *endpage = page + (1 << order) - 1;
2739 for (; page < endpage; page += pageblock_nr_pages) {
2740 int mt = get_pageblock_migratetype(page);
2742 * Only change normal pageblocks (i.e., they can merge
2745 if (migratetype_is_mergeable(mt))
2746 move_freepages_block(zone, page, mt,
2751 return 1UL << order;
2755 * __putback_isolated_page - Return a now-isolated page back where we got it
2756 * @page: Page that was isolated
2757 * @order: Order of the isolated page
2758 * @mt: The page's pageblock's migratetype
2760 * This function is meant to return a page pulled from the free lists via
2761 * __isolate_free_page back to the free lists they were pulled from.
2763 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2765 struct zone *zone = page_zone(page);
2767 /* zone lock should be held when this function is called */
2768 lockdep_assert_held(&zone->lock);
2770 /* Return isolated page to tail of freelist. */
2771 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2772 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2776 * Update NUMA hit/miss statistics
2778 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2782 enum numa_stat_item local_stat = NUMA_LOCAL;
2784 /* skip numa counters update if numa stats is disabled */
2785 if (!static_branch_likely(&vm_numa_stat_key))
2788 if (zone_to_nid(z) != numa_node_id())
2789 local_stat = NUMA_OTHER;
2791 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2792 __count_numa_events(z, NUMA_HIT, nr_account);
2794 __count_numa_events(z, NUMA_MISS, nr_account);
2795 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2797 __count_numa_events(z, local_stat, nr_account);
2801 static __always_inline
2802 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2803 unsigned int order, unsigned int alloc_flags,
2807 unsigned long flags;
2811 spin_lock_irqsave(&zone->lock, flags);
2812 if (alloc_flags & ALLOC_HIGHATOMIC)
2813 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2815 page = __rmqueue(zone, order, migratetype, alloc_flags);
2818 * If the allocation fails, allow OOM handling access
2819 * to HIGHATOMIC reserves as failing now is worse than
2820 * failing a high-order atomic allocation in the
2823 if (!page && (alloc_flags & ALLOC_OOM))
2824 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2827 spin_unlock_irqrestore(&zone->lock, flags);
2831 spin_unlock_irqrestore(&zone->lock, flags);
2832 } while (check_new_pages(page, order));
2834 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2835 zone_statistics(preferred_zone, zone, 1);
2840 static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
2842 int high, base_batch, batch, max_nr_alloc;
2843 int high_max, high_min;
2845 base_batch = READ_ONCE(pcp->batch);
2846 high_min = READ_ONCE(pcp->high_min);
2847 high_max = READ_ONCE(pcp->high_max);
2848 high = pcp->high = clamp(pcp->high, high_min, high_max);
2850 /* Check for PCP disabled or boot pageset */
2851 if (unlikely(high < base_batch))
2857 batch = (base_batch << pcp->alloc_factor);
2860 * If we had larger pcp->high, we could avoid to allocate from
2863 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
2864 high = pcp->high = min(high + batch, high_max);
2867 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
2869 * Double the number of pages allocated each time there is
2870 * subsequent allocation of order-0 pages without any freeing.
2872 if (batch <= max_nr_alloc &&
2873 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
2874 pcp->alloc_factor++;
2875 batch = min(batch, max_nr_alloc);
2879 * Scale batch relative to order if batch implies free pages
2880 * can be stored on the PCP. Batch can be 1 for small zones or
2881 * for boot pagesets which should never store free pages as
2882 * the pages may belong to arbitrary zones.
2885 batch = max(batch >> order, 2);
2890 /* Remove page from the per-cpu list, caller must protect the list */
2892 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2894 unsigned int alloc_flags,
2895 struct per_cpu_pages *pcp,
2896 struct list_head *list)
2901 if (list_empty(list)) {
2902 int batch = nr_pcp_alloc(pcp, zone, order);
2905 alloced = rmqueue_bulk(zone, order,
2907 migratetype, alloc_flags);
2909 pcp->count += alloced << order;
2910 if (unlikely(list_empty(list)))
2914 page = list_first_entry(list, struct page, pcp_list);
2915 list_del(&page->pcp_list);
2916 pcp->count -= 1 << order;
2917 } while (check_new_pages(page, order));
2922 /* Lock and remove page from the per-cpu list */
2923 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2924 struct zone *zone, unsigned int order,
2925 int migratetype, unsigned int alloc_flags)
2927 struct per_cpu_pages *pcp;
2928 struct list_head *list;
2930 unsigned long __maybe_unused UP_flags;
2932 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2933 pcp_trylock_prepare(UP_flags);
2934 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2936 pcp_trylock_finish(UP_flags);
2941 * On allocation, reduce the number of pages that are batch freed.
2942 * See nr_pcp_free() where free_factor is increased for subsequent
2945 pcp->free_count >>= 1;
2946 list = &pcp->lists[order_to_pindex(migratetype, order)];
2947 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
2948 pcp_spin_unlock(pcp);
2949 pcp_trylock_finish(UP_flags);
2951 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2952 zone_statistics(preferred_zone, zone, 1);
2958 * Allocate a page from the given zone.
2959 * Use pcplists for THP or "cheap" high-order allocations.
2963 * Do not instrument rmqueue() with KMSAN. This function may call
2964 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
2965 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
2966 * may call rmqueue() again, which will result in a deadlock.
2968 __no_sanitize_memory
2970 struct page *rmqueue(struct zone *preferred_zone,
2971 struct zone *zone, unsigned int order,
2972 gfp_t gfp_flags, unsigned int alloc_flags,
2978 * We most definitely don't want callers attempting to
2979 * allocate greater than order-1 page units with __GFP_NOFAIL.
2981 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2983 if (likely(pcp_allowed_order(order))) {
2984 page = rmqueue_pcplist(preferred_zone, zone, order,
2985 migratetype, alloc_flags);
2990 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
2994 /* Separate test+clear to avoid unnecessary atomics */
2995 if ((alloc_flags & ALLOC_KSWAPD) &&
2996 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
2997 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2998 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3001 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3005 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3007 return __should_fail_alloc_page(gfp_mask, order);
3009 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3011 static inline long __zone_watermark_unusable_free(struct zone *z,
3012 unsigned int order, unsigned int alloc_flags)
3014 long unusable_free = (1 << order) - 1;
3017 * If the caller does not have rights to reserves below the min
3018 * watermark then subtract the high-atomic reserves. This will
3019 * over-estimate the size of the atomic reserve but it avoids a search.
3021 if (likely(!(alloc_flags & ALLOC_RESERVES)))
3022 unusable_free += z->nr_reserved_highatomic;
3025 /* If allocation can't use CMA areas don't use free CMA pages */
3026 if (!(alloc_flags & ALLOC_CMA))
3027 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3029 #ifdef CONFIG_UNACCEPTED_MEMORY
3030 unusable_free += zone_page_state(z, NR_UNACCEPTED);
3033 return unusable_free;
3037 * Return true if free base pages are above 'mark'. For high-order checks it
3038 * will return true of the order-0 watermark is reached and there is at least
3039 * one free page of a suitable size. Checking now avoids taking the zone lock
3040 * to check in the allocation paths if no pages are free.
3042 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3043 int highest_zoneidx, unsigned int alloc_flags,
3049 /* free_pages may go negative - that's OK */
3050 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3052 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
3054 * __GFP_HIGH allows access to 50% of the min reserve as well
3057 if (alloc_flags & ALLOC_MIN_RESERVE) {
3061 * Non-blocking allocations (e.g. GFP_ATOMIC) can
3062 * access more reserves than just __GFP_HIGH. Other
3063 * non-blocking allocations requests such as GFP_NOWAIT
3064 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
3065 * access to the min reserve.
3067 if (alloc_flags & ALLOC_NON_BLOCK)
3072 * OOM victims can try even harder than the normal reserve
3073 * users on the grounds that it's definitely going to be in
3074 * the exit path shortly and free memory. Any allocation it
3075 * makes during the free path will be small and short-lived.
3077 if (alloc_flags & ALLOC_OOM)
3082 * Check watermarks for an order-0 allocation request. If these
3083 * are not met, then a high-order request also cannot go ahead
3084 * even if a suitable page happened to be free.
3086 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3089 /* If this is an order-0 request then the watermark is fine */
3093 /* For a high-order request, check at least one suitable page is free */
3094 for (o = order; o < NR_PAGE_ORDERS; o++) {
3095 struct free_area *area = &z->free_area[o];
3101 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3102 if (!free_area_empty(area, mt))
3107 if ((alloc_flags & ALLOC_CMA) &&
3108 !free_area_empty(area, MIGRATE_CMA)) {
3112 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3113 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3120 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3121 int highest_zoneidx, unsigned int alloc_flags)
3123 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3124 zone_page_state(z, NR_FREE_PAGES));
3127 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3128 unsigned long mark, int highest_zoneidx,
3129 unsigned int alloc_flags, gfp_t gfp_mask)
3133 free_pages = zone_page_state(z, NR_FREE_PAGES);
3136 * Fast check for order-0 only. If this fails then the reserves
3137 * need to be calculated.
3143 usable_free = free_pages;
3144 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3146 /* reserved may over estimate high-atomic reserves. */
3147 usable_free -= min(usable_free, reserved);
3148 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3152 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3157 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3158 * when checking the min watermark. The min watermark is the
3159 * point where boosting is ignored so that kswapd is woken up
3160 * when below the low watermark.
3162 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3163 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3164 mark = z->_watermark[WMARK_MIN];
3165 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3166 alloc_flags, free_pages);
3172 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3173 unsigned long mark, int highest_zoneidx)
3175 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3177 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3178 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3180 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3185 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3187 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3189 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3190 node_reclaim_distance;
3192 #else /* CONFIG_NUMA */
3193 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3197 #endif /* CONFIG_NUMA */
3200 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3201 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3202 * premature use of a lower zone may cause lowmem pressure problems that
3203 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3204 * probably too small. It only makes sense to spread allocations to avoid
3205 * fragmentation between the Normal and DMA32 zones.
3207 static inline unsigned int
3208 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3210 unsigned int alloc_flags;
3213 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3216 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3218 #ifdef CONFIG_ZONE_DMA32
3222 if (zone_idx(zone) != ZONE_NORMAL)
3226 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3227 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3228 * on UMA that if Normal is populated then so is DMA32.
3230 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3231 if (nr_online_nodes > 1 && !populated_zone(--zone))
3234 alloc_flags |= ALLOC_NOFRAGMENT;
3235 #endif /* CONFIG_ZONE_DMA32 */
3239 /* Must be called after current_gfp_context() which can change gfp_mask */
3240 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3241 unsigned int alloc_flags)
3244 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3245 alloc_flags |= ALLOC_CMA;
3251 * get_page_from_freelist goes through the zonelist trying to allocate
3254 static struct page *
3255 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3256 const struct alloc_context *ac)
3260 struct pglist_data *last_pgdat = NULL;
3261 bool last_pgdat_dirty_ok = false;
3266 * Scan zonelist, looking for a zone with enough free.
3267 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3269 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3270 z = ac->preferred_zoneref;
3271 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3276 if (cpusets_enabled() &&
3277 (alloc_flags & ALLOC_CPUSET) &&
3278 !__cpuset_zone_allowed(zone, gfp_mask))
3281 * When allocating a page cache page for writing, we
3282 * want to get it from a node that is within its dirty
3283 * limit, such that no single node holds more than its
3284 * proportional share of globally allowed dirty pages.
3285 * The dirty limits take into account the node's
3286 * lowmem reserves and high watermark so that kswapd
3287 * should be able to balance it without having to
3288 * write pages from its LRU list.
3290 * XXX: For now, allow allocations to potentially
3291 * exceed the per-node dirty limit in the slowpath
3292 * (spread_dirty_pages unset) before going into reclaim,
3293 * which is important when on a NUMA setup the allowed
3294 * nodes are together not big enough to reach the
3295 * global limit. The proper fix for these situations
3296 * will require awareness of nodes in the
3297 * dirty-throttling and the flusher threads.
3299 if (ac->spread_dirty_pages) {
3300 if (last_pgdat != zone->zone_pgdat) {
3301 last_pgdat = zone->zone_pgdat;
3302 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3305 if (!last_pgdat_dirty_ok)
3309 if (no_fallback && nr_online_nodes > 1 &&
3310 zone != ac->preferred_zoneref->zone) {
3314 * If moving to a remote node, retry but allow
3315 * fragmenting fallbacks. Locality is more important
3316 * than fragmentation avoidance.
3318 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3319 if (zone_to_nid(zone) != local_nid) {
3320 alloc_flags &= ~ALLOC_NOFRAGMENT;
3326 * Detect whether the number of free pages is below high
3327 * watermark. If so, we will decrease pcp->high and free
3328 * PCP pages in free path to reduce the possibility of
3329 * premature page reclaiming. Detection is done here to
3330 * avoid to do that in hotter free path.
3332 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3333 goto check_alloc_wmark;
3335 mark = high_wmark_pages(zone);
3336 if (zone_watermark_fast(zone, order, mark,
3337 ac->highest_zoneidx, alloc_flags,
3341 set_bit(ZONE_BELOW_HIGH, &zone->flags);
3344 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3345 if (!zone_watermark_fast(zone, order, mark,
3346 ac->highest_zoneidx, alloc_flags,
3350 if (has_unaccepted_memory()) {
3351 if (try_to_accept_memory(zone, order))
3355 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3357 * Watermark failed for this zone, but see if we can
3358 * grow this zone if it contains deferred pages.
3360 if (deferred_pages_enabled()) {
3361 if (_deferred_grow_zone(zone, order))
3365 /* Checked here to keep the fast path fast */
3366 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3367 if (alloc_flags & ALLOC_NO_WATERMARKS)
3370 if (!node_reclaim_enabled() ||
3371 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3374 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3376 case NODE_RECLAIM_NOSCAN:
3379 case NODE_RECLAIM_FULL:
3380 /* scanned but unreclaimable */
3383 /* did we reclaim enough */
3384 if (zone_watermark_ok(zone, order, mark,
3385 ac->highest_zoneidx, alloc_flags))
3393 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3394 gfp_mask, alloc_flags, ac->migratetype);
3396 prep_new_page(page, order, gfp_mask, alloc_flags);
3399 * If this is a high-order atomic allocation then check
3400 * if the pageblock should be reserved for the future
3402 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3403 reserve_highatomic_pageblock(page, zone);
3407 if (has_unaccepted_memory()) {
3408 if (try_to_accept_memory(zone, order))
3412 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3413 /* Try again if zone has deferred pages */
3414 if (deferred_pages_enabled()) {
3415 if (_deferred_grow_zone(zone, order))
3423 * It's possible on a UMA machine to get through all zones that are
3424 * fragmented. If avoiding fragmentation, reset and try again.
3427 alloc_flags &= ~ALLOC_NOFRAGMENT;
3434 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3436 unsigned int filter = SHOW_MEM_FILTER_NODES;
3439 * This documents exceptions given to allocations in certain
3440 * contexts that are allowed to allocate outside current's set
3443 if (!(gfp_mask & __GFP_NOMEMALLOC))
3444 if (tsk_is_oom_victim(current) ||
3445 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3446 filter &= ~SHOW_MEM_FILTER_NODES;
3447 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3448 filter &= ~SHOW_MEM_FILTER_NODES;
3450 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3453 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3455 struct va_format vaf;
3457 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3459 if ((gfp_mask & __GFP_NOWARN) ||
3460 !__ratelimit(&nopage_rs) ||
3461 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3464 va_start(args, fmt);
3467 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3468 current->comm, &vaf, gfp_mask, &gfp_mask,
3469 nodemask_pr_args(nodemask));
3472 cpuset_print_current_mems_allowed();
3475 warn_alloc_show_mem(gfp_mask, nodemask);
3478 static inline struct page *
3479 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3480 unsigned int alloc_flags,
3481 const struct alloc_context *ac)
3485 page = get_page_from_freelist(gfp_mask, order,
3486 alloc_flags|ALLOC_CPUSET, ac);
3488 * fallback to ignore cpuset restriction if our nodes
3492 page = get_page_from_freelist(gfp_mask, order,
3498 static inline struct page *
3499 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3500 const struct alloc_context *ac, unsigned long *did_some_progress)
3502 struct oom_control oc = {
3503 .zonelist = ac->zonelist,
3504 .nodemask = ac->nodemask,
3506 .gfp_mask = gfp_mask,
3511 *did_some_progress = 0;
3514 * Acquire the oom lock. If that fails, somebody else is
3515 * making progress for us.
3517 if (!mutex_trylock(&oom_lock)) {
3518 *did_some_progress = 1;
3519 schedule_timeout_uninterruptible(1);
3524 * Go through the zonelist yet one more time, keep very high watermark
3525 * here, this is only to catch a parallel oom killing, we must fail if
3526 * we're still under heavy pressure. But make sure that this reclaim
3527 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3528 * allocation which will never fail due to oom_lock already held.
3530 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3531 ~__GFP_DIRECT_RECLAIM, order,
3532 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3536 /* Coredumps can quickly deplete all memory reserves */
3537 if (current->flags & PF_DUMPCORE)
3539 /* The OOM killer will not help higher order allocs */
3540 if (order > PAGE_ALLOC_COSTLY_ORDER)
3543 * We have already exhausted all our reclaim opportunities without any
3544 * success so it is time to admit defeat. We will skip the OOM killer
3545 * because it is very likely that the caller has a more reasonable
3546 * fallback than shooting a random task.
3548 * The OOM killer may not free memory on a specific node.
3550 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3552 /* The OOM killer does not needlessly kill tasks for lowmem */
3553 if (ac->highest_zoneidx < ZONE_NORMAL)
3555 if (pm_suspended_storage())
3558 * XXX: GFP_NOFS allocations should rather fail than rely on
3559 * other request to make a forward progress.
3560 * We are in an unfortunate situation where out_of_memory cannot
3561 * do much for this context but let's try it to at least get
3562 * access to memory reserved if the current task is killed (see
3563 * out_of_memory). Once filesystems are ready to handle allocation
3564 * failures more gracefully we should just bail out here.
3567 /* Exhausted what can be done so it's blame time */
3568 if (out_of_memory(&oc) ||
3569 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3570 *did_some_progress = 1;
3573 * Help non-failing allocations by giving them access to memory
3576 if (gfp_mask & __GFP_NOFAIL)
3577 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3578 ALLOC_NO_WATERMARKS, ac);
3581 mutex_unlock(&oom_lock);
3586 * Maximum number of compaction retries with a progress before OOM
3587 * killer is consider as the only way to move forward.
3589 #define MAX_COMPACT_RETRIES 16
3591 #ifdef CONFIG_COMPACTION
3592 /* Try memory compaction for high-order allocations before reclaim */
3593 static struct page *
3594 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3595 unsigned int alloc_flags, const struct alloc_context *ac,
3596 enum compact_priority prio, enum compact_result *compact_result)
3598 struct page *page = NULL;
3599 unsigned long pflags;
3600 unsigned int noreclaim_flag;
3605 psi_memstall_enter(&pflags);
3606 delayacct_compact_start();
3607 noreclaim_flag = memalloc_noreclaim_save();
3609 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3612 memalloc_noreclaim_restore(noreclaim_flag);
3613 psi_memstall_leave(&pflags);
3614 delayacct_compact_end();
3616 if (*compact_result == COMPACT_SKIPPED)
3619 * At least in one zone compaction wasn't deferred or skipped, so let's
3620 * count a compaction stall
3622 count_vm_event(COMPACTSTALL);
3624 /* Prep a captured page if available */
3626 prep_new_page(page, order, gfp_mask, alloc_flags);
3628 /* Try get a page from the freelist if available */
3630 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3633 struct zone *zone = page_zone(page);
3635 zone->compact_blockskip_flush = false;
3636 compaction_defer_reset(zone, order, true);
3637 count_vm_event(COMPACTSUCCESS);
3642 * It's bad if compaction run occurs and fails. The most likely reason
3643 * is that pages exist, but not enough to satisfy watermarks.
3645 count_vm_event(COMPACTFAIL);
3653 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3654 enum compact_result compact_result,
3655 enum compact_priority *compact_priority,
3656 int *compaction_retries)
3658 int max_retries = MAX_COMPACT_RETRIES;
3661 int retries = *compaction_retries;
3662 enum compact_priority priority = *compact_priority;
3667 if (fatal_signal_pending(current))
3671 * Compaction was skipped due to a lack of free order-0
3672 * migration targets. Continue if reclaim can help.
3674 if (compact_result == COMPACT_SKIPPED) {
3675 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3680 * Compaction managed to coalesce some page blocks, but the
3681 * allocation failed presumably due to a race. Retry some.
3683 if (compact_result == COMPACT_SUCCESS) {
3685 * !costly requests are much more important than
3686 * __GFP_RETRY_MAYFAIL costly ones because they are de
3687 * facto nofail and invoke OOM killer to move on while
3688 * costly can fail and users are ready to cope with
3689 * that. 1/4 retries is rather arbitrary but we would
3690 * need much more detailed feedback from compaction to
3691 * make a better decision.
3693 if (order > PAGE_ALLOC_COSTLY_ORDER)
3696 if (++(*compaction_retries) <= max_retries) {
3703 * Compaction failed. Retry with increasing priority.
3705 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3706 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3708 if (*compact_priority > min_priority) {
3709 (*compact_priority)--;
3710 *compaction_retries = 0;
3714 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3718 static inline struct page *
3719 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3720 unsigned int alloc_flags, const struct alloc_context *ac,
3721 enum compact_priority prio, enum compact_result *compact_result)
3723 *compact_result = COMPACT_SKIPPED;
3728 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3729 enum compact_result compact_result,
3730 enum compact_priority *compact_priority,
3731 int *compaction_retries)
3736 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3740 * There are setups with compaction disabled which would prefer to loop
3741 * inside the allocator rather than hit the oom killer prematurely.
3742 * Let's give them a good hope and keep retrying while the order-0
3743 * watermarks are OK.
3745 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3746 ac->highest_zoneidx, ac->nodemask) {
3747 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3748 ac->highest_zoneidx, alloc_flags))
3753 #endif /* CONFIG_COMPACTION */
3755 #ifdef CONFIG_LOCKDEP
3756 static struct lockdep_map __fs_reclaim_map =
3757 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3759 static bool __need_reclaim(gfp_t gfp_mask)
3761 /* no reclaim without waiting on it */
3762 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3765 /* this guy won't enter reclaim */
3766 if (current->flags & PF_MEMALLOC)
3769 if (gfp_mask & __GFP_NOLOCKDEP)
3775 void __fs_reclaim_acquire(unsigned long ip)
3777 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3780 void __fs_reclaim_release(unsigned long ip)
3782 lock_release(&__fs_reclaim_map, ip);
3785 void fs_reclaim_acquire(gfp_t gfp_mask)
3787 gfp_mask = current_gfp_context(gfp_mask);
3789 if (__need_reclaim(gfp_mask)) {
3790 if (gfp_mask & __GFP_FS)
3791 __fs_reclaim_acquire(_RET_IP_);
3793 #ifdef CONFIG_MMU_NOTIFIER
3794 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3795 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3800 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3802 void fs_reclaim_release(gfp_t gfp_mask)
3804 gfp_mask = current_gfp_context(gfp_mask);
3806 if (__need_reclaim(gfp_mask)) {
3807 if (gfp_mask & __GFP_FS)
3808 __fs_reclaim_release(_RET_IP_);
3811 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3815 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3816 * have been rebuilt so allocation retries. Reader side does not lock and
3817 * retries the allocation if zonelist changes. Writer side is protected by the
3818 * embedded spin_lock.
3820 static DEFINE_SEQLOCK(zonelist_update_seq);
3822 static unsigned int zonelist_iter_begin(void)
3824 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3825 return read_seqbegin(&zonelist_update_seq);
3830 static unsigned int check_retry_zonelist(unsigned int seq)
3832 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3833 return read_seqretry(&zonelist_update_seq, seq);
3838 /* Perform direct synchronous page reclaim */
3839 static unsigned long
3840 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3841 const struct alloc_context *ac)
3843 unsigned int noreclaim_flag;
3844 unsigned long progress;
3848 /* We now go into synchronous reclaim */
3849 cpuset_memory_pressure_bump();
3850 fs_reclaim_acquire(gfp_mask);
3851 noreclaim_flag = memalloc_noreclaim_save();
3853 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3856 memalloc_noreclaim_restore(noreclaim_flag);
3857 fs_reclaim_release(gfp_mask);
3864 /* The really slow allocator path where we enter direct reclaim */
3865 static inline struct page *
3866 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3867 unsigned int alloc_flags, const struct alloc_context *ac,
3868 unsigned long *did_some_progress)
3870 struct page *page = NULL;
3871 unsigned long pflags;
3872 bool drained = false;
3874 psi_memstall_enter(&pflags);
3875 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3876 if (unlikely(!(*did_some_progress)))
3880 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3883 * If an allocation failed after direct reclaim, it could be because
3884 * pages are pinned on the per-cpu lists or in high alloc reserves.
3885 * Shrink them and try again
3887 if (!page && !drained) {
3888 unreserve_highatomic_pageblock(ac, false);
3889 drain_all_pages(NULL);
3894 psi_memstall_leave(&pflags);
3899 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3900 const struct alloc_context *ac)
3904 pg_data_t *last_pgdat = NULL;
3905 enum zone_type highest_zoneidx = ac->highest_zoneidx;
3907 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3909 if (!managed_zone(zone))
3911 if (last_pgdat != zone->zone_pgdat) {
3912 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3913 last_pgdat = zone->zone_pgdat;
3918 static inline unsigned int
3919 gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3921 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3924 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3925 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3926 * to save two branches.
3928 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
3929 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
3932 * The caller may dip into page reserves a bit more if the caller
3933 * cannot run direct reclaim, or if the caller has realtime scheduling
3934 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3935 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3937 alloc_flags |= (__force int)
3938 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
3940 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
3942 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3943 * if it can't schedule.
3945 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
3946 alloc_flags |= ALLOC_NON_BLOCK;
3949 alloc_flags |= ALLOC_HIGHATOMIC;
3953 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
3954 * GFP_ATOMIC) rather than fail, see the comment for
3955 * cpuset_node_allowed().
3957 if (alloc_flags & ALLOC_MIN_RESERVE)
3958 alloc_flags &= ~ALLOC_CPUSET;
3959 } else if (unlikely(rt_task(current)) && in_task())
3960 alloc_flags |= ALLOC_MIN_RESERVE;
3962 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
3967 static bool oom_reserves_allowed(struct task_struct *tsk)
3969 if (!tsk_is_oom_victim(tsk))
3973 * !MMU doesn't have oom reaper so give access to memory reserves
3974 * only to the thread with TIF_MEMDIE set
3976 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3983 * Distinguish requests which really need access to full memory
3984 * reserves from oom victims which can live with a portion of it
3986 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3988 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3990 if (gfp_mask & __GFP_MEMALLOC)
3991 return ALLOC_NO_WATERMARKS;
3992 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3993 return ALLOC_NO_WATERMARKS;
3994 if (!in_interrupt()) {
3995 if (current->flags & PF_MEMALLOC)
3996 return ALLOC_NO_WATERMARKS;
3997 else if (oom_reserves_allowed(current))
4004 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4006 return !!__gfp_pfmemalloc_flags(gfp_mask);
4010 * Checks whether it makes sense to retry the reclaim to make a forward progress
4011 * for the given allocation request.
4013 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4014 * without success, or when we couldn't even meet the watermark if we
4015 * reclaimed all remaining pages on the LRU lists.
4017 * Returns true if a retry is viable or false to enter the oom path.
4020 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4021 struct alloc_context *ac, int alloc_flags,
4022 bool did_some_progress, int *no_progress_loops)
4029 * Costly allocations might have made a progress but this doesn't mean
4030 * their order will become available due to high fragmentation so
4031 * always increment the no progress counter for them
4033 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4034 *no_progress_loops = 0;
4036 (*no_progress_loops)++;
4038 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
4043 * Keep reclaiming pages while there is a chance this will lead
4044 * somewhere. If none of the target zones can satisfy our allocation
4045 * request even if all reclaimable pages are considered then we are
4046 * screwed and have to go OOM.
4048 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4049 ac->highest_zoneidx, ac->nodemask) {
4050 unsigned long available;
4051 unsigned long reclaimable;
4052 unsigned long min_wmark = min_wmark_pages(zone);
4055 available = reclaimable = zone_reclaimable_pages(zone);
4056 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4059 * Would the allocation succeed if we reclaimed all
4060 * reclaimable pages?
4062 wmark = __zone_watermark_ok(zone, order, min_wmark,
4063 ac->highest_zoneidx, alloc_flags, available);
4064 trace_reclaim_retry_zone(z, order, reclaimable,
4065 available, min_wmark, *no_progress_loops, wmark);
4073 * Memory allocation/reclaim might be called from a WQ context and the
4074 * current implementation of the WQ concurrency control doesn't
4075 * recognize that a particular WQ is congested if the worker thread is
4076 * looping without ever sleeping. Therefore we have to do a short sleep
4077 * here rather than calling cond_resched().
4079 if (current->flags & PF_WQ_WORKER)
4080 schedule_timeout_uninterruptible(1);
4084 /* Before OOM, exhaust highatomic_reserve */
4086 return unreserve_highatomic_pageblock(ac, true);
4092 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4095 * It's possible that cpuset's mems_allowed and the nodemask from
4096 * mempolicy don't intersect. This should be normally dealt with by
4097 * policy_nodemask(), but it's possible to race with cpuset update in
4098 * such a way the check therein was true, and then it became false
4099 * before we got our cpuset_mems_cookie here.
4100 * This assumes that for all allocations, ac->nodemask can come only
4101 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4102 * when it does not intersect with the cpuset restrictions) or the
4103 * caller can deal with a violated nodemask.
4105 if (cpusets_enabled() && ac->nodemask &&
4106 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4107 ac->nodemask = NULL;
4112 * When updating a task's mems_allowed or mempolicy nodemask, it is
4113 * possible to race with parallel threads in such a way that our
4114 * allocation can fail while the mask is being updated. If we are about
4115 * to fail, check if the cpuset changed during allocation and if so,
4118 if (read_mems_allowed_retry(cpuset_mems_cookie))
4124 static inline struct page *
4125 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4126 struct alloc_context *ac)
4128 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4129 bool can_compact = gfp_compaction_allowed(gfp_mask);
4130 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4131 struct page *page = NULL;
4132 unsigned int alloc_flags;
4133 unsigned long did_some_progress;
4134 enum compact_priority compact_priority;
4135 enum compact_result compact_result;
4136 int compaction_retries;
4137 int no_progress_loops;
4138 unsigned int cpuset_mems_cookie;
4139 unsigned int zonelist_iter_cookie;
4143 compaction_retries = 0;
4144 no_progress_loops = 0;
4145 compact_priority = DEF_COMPACT_PRIORITY;
4146 cpuset_mems_cookie = read_mems_allowed_begin();
4147 zonelist_iter_cookie = zonelist_iter_begin();
4150 * The fast path uses conservative alloc_flags to succeed only until
4151 * kswapd needs to be woken up, and to avoid the cost of setting up
4152 * alloc_flags precisely. So we do that now.
4154 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4157 * We need to recalculate the starting point for the zonelist iterator
4158 * because we might have used different nodemask in the fast path, or
4159 * there was a cpuset modification and we are retrying - otherwise we
4160 * could end up iterating over non-eligible zones endlessly.
4162 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4163 ac->highest_zoneidx, ac->nodemask);
4164 if (!ac->preferred_zoneref->zone)
4168 * Check for insane configurations where the cpuset doesn't contain
4169 * any suitable zone to satisfy the request - e.g. non-movable
4170 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4172 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4173 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4174 ac->highest_zoneidx,
4175 &cpuset_current_mems_allowed);
4180 if (alloc_flags & ALLOC_KSWAPD)
4181 wake_all_kswapds(order, gfp_mask, ac);
4184 * The adjusted alloc_flags might result in immediate success, so try
4187 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4192 * For costly allocations, try direct compaction first, as it's likely
4193 * that we have enough base pages and don't need to reclaim. For non-
4194 * movable high-order allocations, do that as well, as compaction will
4195 * try prevent permanent fragmentation by migrating from blocks of the
4197 * Don't try this for allocations that are allowed to ignore
4198 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4200 if (can_direct_reclaim && can_compact &&
4202 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4203 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4204 page = __alloc_pages_direct_compact(gfp_mask, order,
4206 INIT_COMPACT_PRIORITY,
4212 * Checks for costly allocations with __GFP_NORETRY, which
4213 * includes some THP page fault allocations
4215 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4217 * If allocating entire pageblock(s) and compaction
4218 * failed because all zones are below low watermarks
4219 * or is prohibited because it recently failed at this
4220 * order, fail immediately unless the allocator has
4221 * requested compaction and reclaim retry.
4224 * - potentially very expensive because zones are far
4225 * below their low watermarks or this is part of very
4226 * bursty high order allocations,
4227 * - not guaranteed to help because isolate_freepages()
4228 * may not iterate over freed pages as part of its
4230 * - unlikely to make entire pageblocks free on its
4233 if (compact_result == COMPACT_SKIPPED ||
4234 compact_result == COMPACT_DEFERRED)
4238 * Looks like reclaim/compaction is worth trying, but
4239 * sync compaction could be very expensive, so keep
4240 * using async compaction.
4242 compact_priority = INIT_COMPACT_PRIORITY;
4247 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4248 if (alloc_flags & ALLOC_KSWAPD)
4249 wake_all_kswapds(order, gfp_mask, ac);
4251 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4253 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4254 (alloc_flags & ALLOC_KSWAPD);
4257 * Reset the nodemask and zonelist iterators if memory policies can be
4258 * ignored. These allocations are high priority and system rather than
4261 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4262 ac->nodemask = NULL;
4263 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4264 ac->highest_zoneidx, ac->nodemask);
4267 /* Attempt with potentially adjusted zonelist and alloc_flags */
4268 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4272 /* Caller is not willing to reclaim, we can't balance anything */
4273 if (!can_direct_reclaim)
4276 /* Avoid recursion of direct reclaim */
4277 if (current->flags & PF_MEMALLOC)
4280 /* Try direct reclaim and then allocating */
4281 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4282 &did_some_progress);
4286 /* Try direct compaction and then allocating */
4287 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4288 compact_priority, &compact_result);
4292 /* Do not loop if specifically requested */
4293 if (gfp_mask & __GFP_NORETRY)
4297 * Do not retry costly high order allocations unless they are
4298 * __GFP_RETRY_MAYFAIL and we can compact
4300 if (costly_order && (!can_compact ||
4301 !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4304 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4305 did_some_progress > 0, &no_progress_loops))
4309 * It doesn't make any sense to retry for the compaction if the order-0
4310 * reclaim is not able to make any progress because the current
4311 * implementation of the compaction depends on the sufficient amount
4312 * of free memory (see __compaction_suitable)
4314 if (did_some_progress > 0 && can_compact &&
4315 should_compact_retry(ac, order, alloc_flags,
4316 compact_result, &compact_priority,
4317 &compaction_retries))
4322 * Deal with possible cpuset update races or zonelist updates to avoid
4323 * a unnecessary OOM kill.
4325 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4326 check_retry_zonelist(zonelist_iter_cookie))
4329 /* Reclaim has failed us, start killing things */
4330 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4334 /* Avoid allocations with no watermarks from looping endlessly */
4335 if (tsk_is_oom_victim(current) &&
4336 (alloc_flags & ALLOC_OOM ||
4337 (gfp_mask & __GFP_NOMEMALLOC)))
4340 /* Retry as long as the OOM killer is making progress */
4341 if (did_some_progress) {
4342 no_progress_loops = 0;
4348 * Deal with possible cpuset update races or zonelist updates to avoid
4349 * a unnecessary OOM kill.
4351 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4352 check_retry_zonelist(zonelist_iter_cookie))
4356 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4359 if (gfp_mask & __GFP_NOFAIL) {
4361 * All existing users of the __GFP_NOFAIL are blockable, so warn
4362 * of any new users that actually require GFP_NOWAIT
4364 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4368 * PF_MEMALLOC request from this context is rather bizarre
4369 * because we cannot reclaim anything and only can loop waiting
4370 * for somebody to do a work for us
4372 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4375 * non failing costly orders are a hard requirement which we
4376 * are not prepared for much so let's warn about these users
4377 * so that we can identify them and convert them to something
4380 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4383 * Help non-failing allocations by giving some access to memory
4384 * reserves normally used for high priority non-blocking
4385 * allocations but do not use ALLOC_NO_WATERMARKS because this
4386 * could deplete whole memory reserves which would just make
4387 * the situation worse.
4389 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4397 warn_alloc(gfp_mask, ac->nodemask,
4398 "page allocation failure: order:%u", order);
4403 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4404 int preferred_nid, nodemask_t *nodemask,
4405 struct alloc_context *ac, gfp_t *alloc_gfp,
4406 unsigned int *alloc_flags)
4408 ac->highest_zoneidx = gfp_zone(gfp_mask);
4409 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4410 ac->nodemask = nodemask;
4411 ac->migratetype = gfp_migratetype(gfp_mask);
4413 if (cpusets_enabled()) {
4414 *alloc_gfp |= __GFP_HARDWALL;
4416 * When we are in the interrupt context, it is irrelevant
4417 * to the current task context. It means that any node ok.
4419 if (in_task() && !ac->nodemask)
4420 ac->nodemask = &cpuset_current_mems_allowed;
4422 *alloc_flags |= ALLOC_CPUSET;
4425 might_alloc(gfp_mask);
4427 if (should_fail_alloc_page(gfp_mask, order))
4430 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4432 /* Dirty zone balancing only done in the fast path */
4433 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4436 * The preferred zone is used for statistics but crucially it is
4437 * also used as the starting point for the zonelist iterator. It
4438 * may get reset for allocations that ignore memory policies.
4440 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4441 ac->highest_zoneidx, ac->nodemask);
4447 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4448 * @gfp: GFP flags for the allocation
4449 * @preferred_nid: The preferred NUMA node ID to allocate from
4450 * @nodemask: Set of nodes to allocate from, may be NULL
4451 * @nr_pages: The number of pages desired on the list or array
4452 * @page_list: Optional list to store the allocated pages
4453 * @page_array: Optional array to store the pages
4455 * This is a batched version of the page allocator that attempts to
4456 * allocate nr_pages quickly. Pages are added to page_list if page_list
4457 * is not NULL, otherwise it is assumed that the page_array is valid.
4459 * For lists, nr_pages is the number of pages that should be allocated.
4461 * For arrays, only NULL elements are populated with pages and nr_pages
4462 * is the maximum number of pages that will be stored in the array.
4464 * Returns the number of pages on the list or array.
4466 unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
4467 nodemask_t *nodemask, int nr_pages,
4468 struct list_head *page_list,
4469 struct page **page_array)
4472 unsigned long __maybe_unused UP_flags;
4475 struct per_cpu_pages *pcp;
4476 struct list_head *pcp_list;
4477 struct alloc_context ac;
4479 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4480 int nr_populated = 0, nr_account = 0;
4483 * Skip populated array elements to determine if any pages need
4484 * to be allocated before disabling IRQs.
4486 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4489 /* No pages requested? */
4490 if (unlikely(nr_pages <= 0))
4493 /* Already populated array? */
4494 if (unlikely(page_array && nr_pages - nr_populated == 0))
4497 /* Bulk allocator does not support memcg accounting. */
4498 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4501 /* Use the single page allocator for one page. */
4502 if (nr_pages - nr_populated == 1)
4505 #ifdef CONFIG_PAGE_OWNER
4507 * PAGE_OWNER may recurse into the allocator to allocate space to
4508 * save the stack with pagesets.lock held. Releasing/reacquiring
4509 * removes much of the performance benefit of bulk allocation so
4510 * force the caller to allocate one page at a time as it'll have
4511 * similar performance to added complexity to the bulk allocator.
4513 if (static_branch_unlikely(&page_owner_inited))
4517 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4518 gfp &= gfp_allowed_mask;
4520 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4524 /* Find an allowed local zone that meets the low watermark. */
4525 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4528 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4529 !__cpuset_zone_allowed(zone, gfp)) {
4533 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4534 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4538 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4539 if (zone_watermark_fast(zone, 0, mark,
4540 zonelist_zone_idx(ac.preferred_zoneref),
4541 alloc_flags, gfp)) {
4547 * If there are no allowed local zones that meets the watermarks then
4548 * try to allocate a single page and reclaim if necessary.
4550 if (unlikely(!zone))
4553 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4554 pcp_trylock_prepare(UP_flags);
4555 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4559 /* Attempt the batch allocation */
4560 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4561 while (nr_populated < nr_pages) {
4563 /* Skip existing pages */
4564 if (page_array && page_array[nr_populated]) {
4569 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4571 if (unlikely(!page)) {
4572 /* Try and allocate at least one page */
4574 pcp_spin_unlock(pcp);
4581 prep_new_page(page, 0, gfp, 0);
4583 list_add(&page->lru, page_list);
4585 page_array[nr_populated] = page;
4589 pcp_spin_unlock(pcp);
4590 pcp_trylock_finish(UP_flags);
4592 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4593 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4596 return nr_populated;
4599 pcp_trylock_finish(UP_flags);
4602 page = __alloc_pages_noprof(gfp, 0, preferred_nid, nodemask);
4605 list_add(&page->lru, page_list);
4607 page_array[nr_populated] = page;
4613 EXPORT_SYMBOL_GPL(alloc_pages_bulk_noprof);
4616 * This is the 'heart' of the zoned buddy allocator.
4618 struct page *__alloc_pages_noprof(gfp_t gfp, unsigned int order,
4619 int preferred_nid, nodemask_t *nodemask)
4622 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4623 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4624 struct alloc_context ac = { };
4627 * There are several places where we assume that the order value is sane
4628 * so bail out early if the request is out of bound.
4630 if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
4633 gfp &= gfp_allowed_mask;
4635 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4636 * resp. GFP_NOIO which has to be inherited for all allocation requests
4637 * from a particular context which has been marked by
4638 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4639 * movable zones are not used during allocation.
4641 gfp = current_gfp_context(gfp);
4643 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4644 &alloc_gfp, &alloc_flags))
4648 * Forbid the first pass from falling back to types that fragment
4649 * memory until all local zones are considered.
4651 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4653 /* First allocation attempt */
4654 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4659 ac.spread_dirty_pages = false;
4662 * Restore the original nodemask if it was potentially replaced with
4663 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4665 ac.nodemask = nodemask;
4667 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4670 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4671 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4672 __free_pages(page, order);
4676 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4677 kmsan_alloc_page(page, order, alloc_gfp);
4681 EXPORT_SYMBOL(__alloc_pages_noprof);
4683 struct folio *__folio_alloc_noprof(gfp_t gfp, unsigned int order, int preferred_nid,
4684 nodemask_t *nodemask)
4686 struct page *page = __alloc_pages_noprof(gfp | __GFP_COMP, order,
4687 preferred_nid, nodemask);
4688 return page_rmappable_folio(page);
4690 EXPORT_SYMBOL(__folio_alloc_noprof);
4693 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4694 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4695 * you need to access high mem.
4697 unsigned long get_free_pages_noprof(gfp_t gfp_mask, unsigned int order)
4701 page = alloc_pages_noprof(gfp_mask & ~__GFP_HIGHMEM, order);
4704 return (unsigned long) page_address(page);
4706 EXPORT_SYMBOL(get_free_pages_noprof);
4708 unsigned long get_zeroed_page_noprof(gfp_t gfp_mask)
4710 return get_free_pages_noprof(gfp_mask | __GFP_ZERO, 0);
4712 EXPORT_SYMBOL(get_zeroed_page_noprof);
4715 * __free_pages - Free pages allocated with alloc_pages().
4716 * @page: The page pointer returned from alloc_pages().
4717 * @order: The order of the allocation.
4719 * This function can free multi-page allocations that are not compound
4720 * pages. It does not check that the @order passed in matches that of
4721 * the allocation, so it is easy to leak memory. Freeing more memory
4722 * than was allocated will probably emit a warning.
4724 * If the last reference to this page is speculative, it will be released
4725 * by put_page() which only frees the first page of a non-compound
4726 * allocation. To prevent the remaining pages from being leaked, we free
4727 * the subsequent pages here. If you want to use the page's reference
4728 * count to decide when to free the allocation, you should allocate a
4729 * compound page, and use put_page() instead of __free_pages().
4731 * Context: May be called in interrupt context or while holding a normal
4732 * spinlock, but not in NMI context or while holding a raw spinlock.
4734 void __free_pages(struct page *page, unsigned int order)
4736 /* get PageHead before we drop reference */
4737 int head = PageHead(page);
4738 struct alloc_tag *tag = pgalloc_tag_get(page);
4740 if (put_page_testzero(page))
4741 free_unref_page(page, order);
4743 pgalloc_tag_sub_pages(tag, (1 << order) - 1);
4745 free_unref_page(page + (1 << order), order);
4748 EXPORT_SYMBOL(__free_pages);
4750 void free_pages(unsigned long addr, unsigned int order)
4753 VM_BUG_ON(!virt_addr_valid((void *)addr));
4754 __free_pages(virt_to_page((void *)addr), order);
4758 EXPORT_SYMBOL(free_pages);
4762 * An arbitrary-length arbitrary-offset area of memory which resides
4763 * within a 0 or higher order page. Multiple fragments within that page
4764 * are individually refcounted, in the page's reference counter.
4766 * The page_frag functions below provide a simple allocation framework for
4767 * page fragments. This is used by the network stack and network device
4768 * drivers to provide a backing region of memory for use as either an
4769 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4771 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4774 struct page *page = NULL;
4775 gfp_t gfp = gfp_mask;
4777 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4778 gfp_mask = (gfp_mask & ~__GFP_DIRECT_RECLAIM) | __GFP_COMP |
4779 __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC;
4780 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4781 PAGE_FRAG_CACHE_MAX_ORDER);
4782 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4784 if (unlikely(!page))
4785 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4787 nc->va = page ? page_address(page) : NULL;
4792 void page_frag_cache_drain(struct page_frag_cache *nc)
4797 __page_frag_cache_drain(virt_to_head_page(nc->va), nc->pagecnt_bias);
4800 EXPORT_SYMBOL(page_frag_cache_drain);
4802 void __page_frag_cache_drain(struct page *page, unsigned int count)
4804 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4806 if (page_ref_sub_and_test(page, count))
4807 free_unref_page(page, compound_order(page));
4809 EXPORT_SYMBOL(__page_frag_cache_drain);
4811 void *__page_frag_alloc_align(struct page_frag_cache *nc,
4812 unsigned int fragsz, gfp_t gfp_mask,
4813 unsigned int align_mask)
4815 unsigned int size = PAGE_SIZE;
4819 if (unlikely(!nc->va)) {
4821 page = __page_frag_cache_refill(nc, gfp_mask);
4825 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4826 /* if size can vary use size else just use PAGE_SIZE */
4829 /* Even if we own the page, we do not use atomic_set().
4830 * This would break get_page_unless_zero() users.
4832 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4834 /* reset page count bias and offset to start of new frag */
4835 nc->pfmemalloc = page_is_pfmemalloc(page);
4836 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4840 offset = nc->offset - fragsz;
4841 if (unlikely(offset < 0)) {
4842 page = virt_to_page(nc->va);
4844 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4847 if (unlikely(nc->pfmemalloc)) {
4848 free_unref_page(page, compound_order(page));
4852 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4853 /* if size can vary use size else just use PAGE_SIZE */
4856 /* OK, page count is 0, we can safely set it */
4857 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4859 /* reset page count bias and offset to start of new frag */
4860 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4861 offset = size - fragsz;
4862 if (unlikely(offset < 0)) {
4864 * The caller is trying to allocate a fragment
4865 * with fragsz > PAGE_SIZE but the cache isn't big
4866 * enough to satisfy the request, this may
4867 * happen in low memory conditions.
4868 * We don't release the cache page because
4869 * it could make memory pressure worse
4870 * so we simply return NULL here.
4877 offset &= align_mask;
4878 nc->offset = offset;
4880 return nc->va + offset;
4882 EXPORT_SYMBOL(__page_frag_alloc_align);
4885 * Frees a page fragment allocated out of either a compound or order 0 page.
4887 void page_frag_free(void *addr)
4889 struct page *page = virt_to_head_page(addr);
4891 if (unlikely(put_page_testzero(page)))
4892 free_unref_page(page, compound_order(page));
4894 EXPORT_SYMBOL(page_frag_free);
4896 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4900 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4901 struct page *page = virt_to_page((void *)addr);
4902 struct page *last = page + nr;
4904 split_page_owner(page, order, 0);
4905 pgalloc_tag_split(page, 1 << order);
4906 split_page_memcg(page, order, 0);
4907 while (page < --last)
4908 set_page_refcounted(last);
4910 last = page + (1UL << order);
4911 for (page += nr; page < last; page++)
4912 __free_pages_ok(page, 0, FPI_TO_TAIL);
4914 return (void *)addr;
4918 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4919 * @size: the number of bytes to allocate
4920 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4922 * This function is similar to alloc_pages(), except that it allocates the
4923 * minimum number of pages to satisfy the request. alloc_pages() can only
4924 * allocate memory in power-of-two pages.
4926 * This function is also limited by MAX_PAGE_ORDER.
4928 * Memory allocated by this function must be released by free_pages_exact().
4930 * Return: pointer to the allocated area or %NULL in case of error.
4932 void *alloc_pages_exact_noprof(size_t size, gfp_t gfp_mask)
4934 unsigned int order = get_order(size);
4937 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4938 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4940 addr = get_free_pages_noprof(gfp_mask, order);
4941 return make_alloc_exact(addr, order, size);
4943 EXPORT_SYMBOL(alloc_pages_exact_noprof);
4946 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4948 * @nid: the preferred node ID where memory should be allocated
4949 * @size: the number of bytes to allocate
4950 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4952 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4955 * Return: pointer to the allocated area or %NULL in case of error.
4957 void * __meminit alloc_pages_exact_nid_noprof(int nid, size_t size, gfp_t gfp_mask)
4959 unsigned int order = get_order(size);
4962 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4963 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4965 p = alloc_pages_node_noprof(nid, gfp_mask, order);
4968 return make_alloc_exact((unsigned long)page_address(p), order, size);
4972 * free_pages_exact - release memory allocated via alloc_pages_exact()
4973 * @virt: the value returned by alloc_pages_exact.
4974 * @size: size of allocation, same value as passed to alloc_pages_exact().
4976 * Release the memory allocated by a previous call to alloc_pages_exact.
4978 void free_pages_exact(void *virt, size_t size)
4980 unsigned long addr = (unsigned long)virt;
4981 unsigned long end = addr + PAGE_ALIGN(size);
4983 while (addr < end) {
4988 EXPORT_SYMBOL(free_pages_exact);
4991 * nr_free_zone_pages - count number of pages beyond high watermark
4992 * @offset: The zone index of the highest zone
4994 * nr_free_zone_pages() counts the number of pages which are beyond the
4995 * high watermark within all zones at or below a given zone index. For each
4996 * zone, the number of pages is calculated as:
4998 * nr_free_zone_pages = managed_pages - high_pages
5000 * Return: number of pages beyond high watermark.
5002 static unsigned long nr_free_zone_pages(int offset)
5007 /* Just pick one node, since fallback list is circular */
5008 unsigned long sum = 0;
5010 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5012 for_each_zone_zonelist(zone, z, zonelist, offset) {
5013 unsigned long size = zone_managed_pages(zone);
5014 unsigned long high = high_wmark_pages(zone);
5023 * nr_free_buffer_pages - count number of pages beyond high watermark
5025 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5026 * watermark within ZONE_DMA and ZONE_NORMAL.
5028 * Return: number of pages beyond high watermark within ZONE_DMA and
5031 unsigned long nr_free_buffer_pages(void)
5033 return nr_free_zone_pages(gfp_zone(GFP_USER));
5035 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5037 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5039 zoneref->zone = zone;
5040 zoneref->zone_idx = zone_idx(zone);
5044 * Builds allocation fallback zone lists.
5046 * Add all populated zones of a node to the zonelist.
5048 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5051 enum zone_type zone_type = MAX_NR_ZONES;
5056 zone = pgdat->node_zones + zone_type;
5057 if (populated_zone(zone)) {
5058 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5059 check_highest_zone(zone_type);
5061 } while (zone_type);
5068 static int __parse_numa_zonelist_order(char *s)
5071 * We used to support different zonelists modes but they turned
5072 * out to be just not useful. Let's keep the warning in place
5073 * if somebody still use the cmd line parameter so that we do
5074 * not fail it silently
5076 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5077 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5083 static char numa_zonelist_order[] = "Node";
5084 #define NUMA_ZONELIST_ORDER_LEN 16
5086 * sysctl handler for numa_zonelist_order
5088 static int numa_zonelist_order_handler(struct ctl_table *table, int write,
5089 void *buffer, size_t *length, loff_t *ppos)
5092 return __parse_numa_zonelist_order(buffer);
5093 return proc_dostring(table, write, buffer, length, ppos);
5096 static int node_load[MAX_NUMNODES];
5099 * find_next_best_node - find the next node that should appear in a given node's fallback list
5100 * @node: node whose fallback list we're appending
5101 * @used_node_mask: nodemask_t of already used nodes
5103 * We use a number of factors to determine which is the next node that should
5104 * appear on a given node's fallback list. The node should not have appeared
5105 * already in @node's fallback list, and it should be the next closest node
5106 * according to the distance array (which contains arbitrary distance values
5107 * from each node to each node in the system), and should also prefer nodes
5108 * with no CPUs, since presumably they'll have very little allocation pressure
5109 * on them otherwise.
5111 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5113 int find_next_best_node(int node, nodemask_t *used_node_mask)
5116 int min_val = INT_MAX;
5117 int best_node = NUMA_NO_NODE;
5120 * Use the local node if we haven't already, but for memoryless local
5121 * node, we should skip it and fall back to other nodes.
5123 if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5124 node_set(node, *used_node_mask);
5128 for_each_node_state(n, N_MEMORY) {
5130 /* Don't want a node to appear more than once */
5131 if (node_isset(n, *used_node_mask))
5134 /* Use the distance array to find the distance */
5135 val = node_distance(node, n);
5137 /* Penalize nodes under us ("prefer the next node") */
5140 /* Give preference to headless and unused nodes */
5141 if (!cpumask_empty(cpumask_of_node(n)))
5142 val += PENALTY_FOR_NODE_WITH_CPUS;
5144 /* Slight preference for less loaded node */
5145 val *= MAX_NUMNODES;
5146 val += node_load[n];
5148 if (val < min_val) {
5155 node_set(best_node, *used_node_mask);
5162 * Build zonelists ordered by node and zones within node.
5163 * This results in maximum locality--normal zone overflows into local
5164 * DMA zone, if any--but risks exhausting DMA zone.
5166 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5169 struct zoneref *zonerefs;
5172 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5174 for (i = 0; i < nr_nodes; i++) {
5177 pg_data_t *node = NODE_DATA(node_order[i]);
5179 nr_zones = build_zonerefs_node(node, zonerefs);
5180 zonerefs += nr_zones;
5182 zonerefs->zone = NULL;
5183 zonerefs->zone_idx = 0;
5187 * Build gfp_thisnode zonelists
5189 static void build_thisnode_zonelists(pg_data_t *pgdat)
5191 struct zoneref *zonerefs;
5194 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5195 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5196 zonerefs += nr_zones;
5197 zonerefs->zone = NULL;
5198 zonerefs->zone_idx = 0;
5202 * Build zonelists ordered by zone and nodes within zones.
5203 * This results in conserving DMA zone[s] until all Normal memory is
5204 * exhausted, but results in overflowing to remote node while memory
5205 * may still exist in local DMA zone.
5208 static void build_zonelists(pg_data_t *pgdat)
5210 static int node_order[MAX_NUMNODES];
5211 int node, nr_nodes = 0;
5212 nodemask_t used_mask = NODE_MASK_NONE;
5213 int local_node, prev_node;
5215 /* NUMA-aware ordering of nodes */
5216 local_node = pgdat->node_id;
5217 prev_node = local_node;
5219 memset(node_order, 0, sizeof(node_order));
5220 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5222 * We don't want to pressure a particular node.
5223 * So adding penalty to the first node in same
5224 * distance group to make it round-robin.
5226 if (node_distance(local_node, node) !=
5227 node_distance(local_node, prev_node))
5228 node_load[node] += 1;
5230 node_order[nr_nodes++] = node;
5234 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5235 build_thisnode_zonelists(pgdat);
5236 pr_info("Fallback order for Node %d: ", local_node);
5237 for (node = 0; node < nr_nodes; node++)
5238 pr_cont("%d ", node_order[node]);
5242 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5244 * Return node id of node used for "local" allocations.
5245 * I.e., first node id of first zone in arg node's generic zonelist.
5246 * Used for initializing percpu 'numa_mem', which is used primarily
5247 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5249 int local_memory_node(int node)
5253 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5254 gfp_zone(GFP_KERNEL),
5256 return zone_to_nid(z->zone);
5260 static void setup_min_unmapped_ratio(void);
5261 static void setup_min_slab_ratio(void);
5262 #else /* CONFIG_NUMA */
5264 static void build_zonelists(pg_data_t *pgdat)
5266 struct zoneref *zonerefs;
5269 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5270 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5271 zonerefs += nr_zones;
5273 zonerefs->zone = NULL;
5274 zonerefs->zone_idx = 0;
5277 #endif /* CONFIG_NUMA */
5280 * Boot pageset table. One per cpu which is going to be used for all
5281 * zones and all nodes. The parameters will be set in such a way
5282 * that an item put on a list will immediately be handed over to
5283 * the buddy list. This is safe since pageset manipulation is done
5284 * with interrupts disabled.
5286 * The boot_pagesets must be kept even after bootup is complete for
5287 * unused processors and/or zones. They do play a role for bootstrapping
5288 * hotplugged processors.
5290 * zoneinfo_show() and maybe other functions do
5291 * not check if the processor is online before following the pageset pointer.
5292 * Other parts of the kernel may not check if the zone is available.
5294 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5295 /* These effectively disable the pcplists in the boot pageset completely */
5296 #define BOOT_PAGESET_HIGH 0
5297 #define BOOT_PAGESET_BATCH 1
5298 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5299 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5301 static void __build_all_zonelists(void *data)
5304 int __maybe_unused cpu;
5305 pg_data_t *self = data;
5306 unsigned long flags;
5309 * The zonelist_update_seq must be acquired with irqsave because the
5310 * reader can be invoked from IRQ with GFP_ATOMIC.
5312 write_seqlock_irqsave(&zonelist_update_seq, flags);
5314 * Also disable synchronous printk() to prevent any printk() from
5315 * trying to hold port->lock, for
5316 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5317 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5319 printk_deferred_enter();
5322 memset(node_load, 0, sizeof(node_load));
5326 * This node is hotadded and no memory is yet present. So just
5327 * building zonelists is fine - no need to touch other nodes.
5329 if (self && !node_online(self->node_id)) {
5330 build_zonelists(self);
5333 * All possible nodes have pgdat preallocated
5336 for_each_node(nid) {
5337 pg_data_t *pgdat = NODE_DATA(nid);
5339 build_zonelists(pgdat);
5342 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5344 * We now know the "local memory node" for each node--
5345 * i.e., the node of the first zone in the generic zonelist.
5346 * Set up numa_mem percpu variable for on-line cpus. During
5347 * boot, only the boot cpu should be on-line; we'll init the
5348 * secondary cpus' numa_mem as they come on-line. During
5349 * node/memory hotplug, we'll fixup all on-line cpus.
5351 for_each_online_cpu(cpu)
5352 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5356 printk_deferred_exit();
5357 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5360 static noinline void __init
5361 build_all_zonelists_init(void)
5365 __build_all_zonelists(NULL);
5368 * Initialize the boot_pagesets that are going to be used
5369 * for bootstrapping processors. The real pagesets for
5370 * each zone will be allocated later when the per cpu
5371 * allocator is available.
5373 * boot_pagesets are used also for bootstrapping offline
5374 * cpus if the system is already booted because the pagesets
5375 * are needed to initialize allocators on a specific cpu too.
5376 * F.e. the percpu allocator needs the page allocator which
5377 * needs the percpu allocator in order to allocate its pagesets
5378 * (a chicken-egg dilemma).
5380 for_each_possible_cpu(cpu)
5381 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5383 mminit_verify_zonelist();
5384 cpuset_init_current_mems_allowed();
5388 * unless system_state == SYSTEM_BOOTING.
5390 * __ref due to call of __init annotated helper build_all_zonelists_init
5391 * [protected by SYSTEM_BOOTING].
5393 void __ref build_all_zonelists(pg_data_t *pgdat)
5395 unsigned long vm_total_pages;
5397 if (system_state == SYSTEM_BOOTING) {
5398 build_all_zonelists_init();
5400 __build_all_zonelists(pgdat);
5401 /* cpuset refresh routine should be here */
5403 /* Get the number of free pages beyond high watermark in all zones. */
5404 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5406 * Disable grouping by mobility if the number of pages in the
5407 * system is too low to allow the mechanism to work. It would be
5408 * more accurate, but expensive to check per-zone. This check is
5409 * made on memory-hotadd so a system can start with mobility
5410 * disabled and enable it later
5412 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5413 page_group_by_mobility_disabled = 1;
5415 page_group_by_mobility_disabled = 0;
5417 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5419 page_group_by_mobility_disabled ? "off" : "on",
5422 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5426 static int zone_batchsize(struct zone *zone)
5432 * The number of pages to batch allocate is either ~0.1%
5433 * of the zone or 1MB, whichever is smaller. The batch
5434 * size is striking a balance between allocation latency
5435 * and zone lock contention.
5437 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5438 batch /= 4; /* We effectively *= 4 below */
5443 * Clamp the batch to a 2^n - 1 value. Having a power
5444 * of 2 value was found to be more likely to have
5445 * suboptimal cache aliasing properties in some cases.
5447 * For example if 2 tasks are alternately allocating
5448 * batches of pages, one task can end up with a lot
5449 * of pages of one half of the possible page colors
5450 * and the other with pages of the other colors.
5452 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5457 /* The deferral and batching of frees should be suppressed under NOMMU
5460 * The problem is that NOMMU needs to be able to allocate large chunks
5461 * of contiguous memory as there's no hardware page translation to
5462 * assemble apparent contiguous memory from discontiguous pages.
5464 * Queueing large contiguous runs of pages for batching, however,
5465 * causes the pages to actually be freed in smaller chunks. As there
5466 * can be a significant delay between the individual batches being
5467 * recycled, this leads to the once large chunks of space being
5468 * fragmented and becoming unavailable for high-order allocations.
5474 static int percpu_pagelist_high_fraction;
5475 static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5481 unsigned long total_pages;
5483 if (!high_fraction) {
5485 * By default, the high value of the pcp is based on the zone
5486 * low watermark so that if they are full then background
5487 * reclaim will not be started prematurely.
5489 total_pages = low_wmark_pages(zone);
5492 * If percpu_pagelist_high_fraction is configured, the high
5493 * value is based on a fraction of the managed pages in the
5496 total_pages = zone_managed_pages(zone) / high_fraction;
5500 * Split the high value across all online CPUs local to the zone. Note
5501 * that early in boot that CPUs may not be online yet and that during
5502 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5503 * onlined. For memory nodes that have no CPUs, split the high value
5504 * across all online CPUs to mitigate the risk that reclaim is triggered
5505 * prematurely due to pages stored on pcp lists.
5507 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5509 nr_split_cpus = num_online_cpus();
5510 high = total_pages / nr_split_cpus;
5513 * Ensure high is at least batch*4. The multiple is based on the
5514 * historical relationship between high and batch.
5516 high = max(high, batch << 2);
5525 * pcp->high and pcp->batch values are related and generally batch is lower
5526 * than high. They are also related to pcp->count such that count is lower
5527 * than high, and as soon as it reaches high, the pcplist is flushed.
5529 * However, guaranteeing these relations at all times would require e.g. write
5530 * barriers here but also careful usage of read barriers at the read side, and
5531 * thus be prone to error and bad for performance. Thus the update only prevents
5532 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5533 * should ensure they can cope with those fields changing asynchronously, and
5534 * fully trust only the pcp->count field on the local CPU with interrupts
5537 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5538 * outside of boot time (or some other assurance that no concurrent updaters
5541 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5542 unsigned long high_max, unsigned long batch)
5544 WRITE_ONCE(pcp->batch, batch);
5545 WRITE_ONCE(pcp->high_min, high_min);
5546 WRITE_ONCE(pcp->high_max, high_max);
5549 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5553 memset(pcp, 0, sizeof(*pcp));
5554 memset(pzstats, 0, sizeof(*pzstats));
5556 spin_lock_init(&pcp->lock);
5557 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5558 INIT_LIST_HEAD(&pcp->lists[pindex]);
5561 * Set batch and high values safe for a boot pageset. A true percpu
5562 * pageset's initialization will update them subsequently. Here we don't
5563 * need to be as careful as pageset_update() as nobody can access the
5566 pcp->high_min = BOOT_PAGESET_HIGH;
5567 pcp->high_max = BOOT_PAGESET_HIGH;
5568 pcp->batch = BOOT_PAGESET_BATCH;
5569 pcp->free_count = 0;
5572 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
5573 unsigned long high_max, unsigned long batch)
5575 struct per_cpu_pages *pcp;
5578 for_each_possible_cpu(cpu) {
5579 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5580 pageset_update(pcp, high_min, high_max, batch);
5585 * Calculate and set new high and batch values for all per-cpu pagesets of a
5586 * zone based on the zone's size.
5588 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5590 int new_high_min, new_high_max, new_batch;
5592 new_batch = max(1, zone_batchsize(zone));
5593 if (percpu_pagelist_high_fraction) {
5594 new_high_min = zone_highsize(zone, new_batch, cpu_online,
5595 percpu_pagelist_high_fraction);
5597 * PCP high is tuned manually, disable auto-tuning via
5598 * setting high_min and high_max to the manual value.
5600 new_high_max = new_high_min;
5602 new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
5603 new_high_max = zone_highsize(zone, new_batch, cpu_online,
5604 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
5607 if (zone->pageset_high_min == new_high_min &&
5608 zone->pageset_high_max == new_high_max &&
5609 zone->pageset_batch == new_batch)
5612 zone->pageset_high_min = new_high_min;
5613 zone->pageset_high_max = new_high_max;
5614 zone->pageset_batch = new_batch;
5616 __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
5620 void __meminit setup_zone_pageset(struct zone *zone)
5624 /* Size may be 0 on !SMP && !NUMA */
5625 if (sizeof(struct per_cpu_zonestat) > 0)
5626 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5628 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5629 for_each_possible_cpu(cpu) {
5630 struct per_cpu_pages *pcp;
5631 struct per_cpu_zonestat *pzstats;
5633 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5634 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5635 per_cpu_pages_init(pcp, pzstats);
5638 zone_set_pageset_high_and_batch(zone, 0);
5642 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5643 * page high values need to be recalculated.
5645 static void zone_pcp_update(struct zone *zone, int cpu_online)
5647 mutex_lock(&pcp_batch_high_lock);
5648 zone_set_pageset_high_and_batch(zone, cpu_online);
5649 mutex_unlock(&pcp_batch_high_lock);
5652 static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu)
5654 struct per_cpu_pages *pcp;
5655 struct cpu_cacheinfo *cci;
5657 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5658 cci = get_cpu_cacheinfo(cpu);
5660 * If data cache slice of CPU is large enough, "pcp->batch"
5661 * pages can be preserved in PCP before draining PCP for
5662 * consecutive high-order pages freeing without allocation.
5663 * This can reduce zone lock contention without hurting
5664 * cache-hot pages sharing.
5666 spin_lock(&pcp->lock);
5667 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
5668 pcp->flags |= PCPF_FREE_HIGH_BATCH;
5670 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
5671 spin_unlock(&pcp->lock);
5674 void setup_pcp_cacheinfo(unsigned int cpu)
5678 for_each_populated_zone(zone)
5679 zone_pcp_update_cacheinfo(zone, cpu);
5683 * Allocate per cpu pagesets and initialize them.
5684 * Before this call only boot pagesets were available.
5686 void __init setup_per_cpu_pageset(void)
5688 struct pglist_data *pgdat;
5690 int __maybe_unused cpu;
5692 for_each_populated_zone(zone)
5693 setup_zone_pageset(zone);
5697 * Unpopulated zones continue using the boot pagesets.
5698 * The numa stats for these pagesets need to be reset.
5699 * Otherwise, they will end up skewing the stats of
5700 * the nodes these zones are associated with.
5702 for_each_possible_cpu(cpu) {
5703 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5704 memset(pzstats->vm_numa_event, 0,
5705 sizeof(pzstats->vm_numa_event));
5709 for_each_online_pgdat(pgdat)
5710 pgdat->per_cpu_nodestats =
5711 alloc_percpu(struct per_cpu_nodestat);
5714 __meminit void zone_pcp_init(struct zone *zone)
5717 * per cpu subsystem is not up at this point. The following code
5718 * relies on the ability of the linker to provide the
5719 * offset of a (static) per cpu variable into the per cpu area.
5721 zone->per_cpu_pageset = &boot_pageset;
5722 zone->per_cpu_zonestats = &boot_zonestats;
5723 zone->pageset_high_min = BOOT_PAGESET_HIGH;
5724 zone->pageset_high_max = BOOT_PAGESET_HIGH;
5725 zone->pageset_batch = BOOT_PAGESET_BATCH;
5727 if (populated_zone(zone))
5728 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5729 zone->present_pages, zone_batchsize(zone));
5732 void adjust_managed_page_count(struct page *page, long count)
5734 atomic_long_add(count, &page_zone(page)->managed_pages);
5735 totalram_pages_add(count);
5736 #ifdef CONFIG_HIGHMEM
5737 if (PageHighMem(page))
5738 totalhigh_pages_add(count);
5741 EXPORT_SYMBOL(adjust_managed_page_count);
5743 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5746 unsigned long pages = 0;
5748 start = (void *)PAGE_ALIGN((unsigned long)start);
5749 end = (void *)((unsigned long)end & PAGE_MASK);
5750 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5751 struct page *page = virt_to_page(pos);
5752 void *direct_map_addr;
5755 * 'direct_map_addr' might be different from 'pos'
5756 * because some architectures' virt_to_page()
5757 * work with aliases. Getting the direct map
5758 * address ensures that we get a _writeable_
5759 * alias for the memset().
5761 direct_map_addr = page_address(page);
5763 * Perform a kasan-unchecked memset() since this memory
5764 * has not been initialized.
5766 direct_map_addr = kasan_reset_tag(direct_map_addr);
5767 if ((unsigned int)poison <= 0xFF)
5768 memset(direct_map_addr, poison, PAGE_SIZE);
5770 free_reserved_page(page);
5774 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5779 static int page_alloc_cpu_dead(unsigned int cpu)
5783 lru_add_drain_cpu(cpu);
5784 mlock_drain_remote(cpu);
5788 * Spill the event counters of the dead processor
5789 * into the current processors event counters.
5790 * This artificially elevates the count of the current
5793 vm_events_fold_cpu(cpu);
5796 * Zero the differential counters of the dead processor
5797 * so that the vm statistics are consistent.
5799 * This is only okay since the processor is dead and cannot
5800 * race with what we are doing.
5802 cpu_vm_stats_fold(cpu);
5804 for_each_populated_zone(zone)
5805 zone_pcp_update(zone, 0);
5810 static int page_alloc_cpu_online(unsigned int cpu)
5814 for_each_populated_zone(zone)
5815 zone_pcp_update(zone, 1);
5819 void __init page_alloc_init_cpuhp(void)
5823 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5824 "mm/page_alloc:pcp",
5825 page_alloc_cpu_online,
5826 page_alloc_cpu_dead);
5831 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5832 * or min_free_kbytes changes.
5834 static void calculate_totalreserve_pages(void)
5836 struct pglist_data *pgdat;
5837 unsigned long reserve_pages = 0;
5838 enum zone_type i, j;
5840 for_each_online_pgdat(pgdat) {
5842 pgdat->totalreserve_pages = 0;
5844 for (i = 0; i < MAX_NR_ZONES; i++) {
5845 struct zone *zone = pgdat->node_zones + i;
5847 unsigned long managed_pages = zone_managed_pages(zone);
5849 /* Find valid and maximum lowmem_reserve in the zone */
5850 for (j = i; j < MAX_NR_ZONES; j++) {
5851 if (zone->lowmem_reserve[j] > max)
5852 max = zone->lowmem_reserve[j];
5855 /* we treat the high watermark as reserved pages. */
5856 max += high_wmark_pages(zone);
5858 if (max > managed_pages)
5859 max = managed_pages;
5861 pgdat->totalreserve_pages += max;
5863 reserve_pages += max;
5866 totalreserve_pages = reserve_pages;
5870 * setup_per_zone_lowmem_reserve - called whenever
5871 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5872 * has a correct pages reserved value, so an adequate number of
5873 * pages are left in the zone after a successful __alloc_pages().
5875 static void setup_per_zone_lowmem_reserve(void)
5877 struct pglist_data *pgdat;
5878 enum zone_type i, j;
5880 for_each_online_pgdat(pgdat) {
5881 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5882 struct zone *zone = &pgdat->node_zones[i];
5883 int ratio = sysctl_lowmem_reserve_ratio[i];
5884 bool clear = !ratio || !zone_managed_pages(zone);
5885 unsigned long managed_pages = 0;
5887 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5888 struct zone *upper_zone = &pgdat->node_zones[j];
5889 bool empty = !zone_managed_pages(upper_zone);
5891 managed_pages += zone_managed_pages(upper_zone);
5894 zone->lowmem_reserve[j] = 0;
5896 zone->lowmem_reserve[j] = managed_pages / ratio;
5901 /* update totalreserve_pages */
5902 calculate_totalreserve_pages();
5905 static void __setup_per_zone_wmarks(void)
5907 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5908 unsigned long lowmem_pages = 0;
5910 unsigned long flags;
5912 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5913 for_each_zone(zone) {
5914 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5915 lowmem_pages += zone_managed_pages(zone);
5918 for_each_zone(zone) {
5921 spin_lock_irqsave(&zone->lock, flags);
5922 tmp = (u64)pages_min * zone_managed_pages(zone);
5923 tmp = div64_ul(tmp, lowmem_pages);
5924 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5926 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5927 * need highmem and movable zones pages, so cap pages_min
5928 * to a small value here.
5930 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5931 * deltas control async page reclaim, and so should
5932 * not be capped for highmem and movable zones.
5934 unsigned long min_pages;
5936 min_pages = zone_managed_pages(zone) / 1024;
5937 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5938 zone->_watermark[WMARK_MIN] = min_pages;
5941 * If it's a lowmem zone, reserve a number of pages
5942 * proportionate to the zone's size.
5944 zone->_watermark[WMARK_MIN] = tmp;
5948 * Set the kswapd watermarks distance according to the
5949 * scale factor in proportion to available memory, but
5950 * ensure a minimum size on small systems.
5952 tmp = max_t(u64, tmp >> 2,
5953 mult_frac(zone_managed_pages(zone),
5954 watermark_scale_factor, 10000));
5956 zone->watermark_boost = 0;
5957 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
5958 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5959 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5961 spin_unlock_irqrestore(&zone->lock, flags);
5964 /* update totalreserve_pages */
5965 calculate_totalreserve_pages();
5969 * setup_per_zone_wmarks - called when min_free_kbytes changes
5970 * or when memory is hot-{added|removed}
5972 * Ensures that the watermark[min,low,high] values for each zone are set
5973 * correctly with respect to min_free_kbytes.
5975 void setup_per_zone_wmarks(void)
5978 static DEFINE_SPINLOCK(lock);
5981 __setup_per_zone_wmarks();
5985 * The watermark size have changed so update the pcpu batch
5986 * and high limits or the limits may be inappropriate.
5989 zone_pcp_update(zone, 0);
5993 * Initialise min_free_kbytes.
5995 * For small machines we want it small (128k min). For large machines
5996 * we want it large (256MB max). But it is not linear, because network
5997 * bandwidth does not increase linearly with machine size. We use
5999 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6000 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6016 void calculate_min_free_kbytes(void)
6018 unsigned long lowmem_kbytes;
6019 int new_min_free_kbytes;
6021 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6022 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6024 if (new_min_free_kbytes > user_min_free_kbytes)
6025 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
6027 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6028 new_min_free_kbytes, user_min_free_kbytes);
6032 int __meminit init_per_zone_wmark_min(void)
6034 calculate_min_free_kbytes();
6035 setup_per_zone_wmarks();
6036 refresh_zone_stat_thresholds();
6037 setup_per_zone_lowmem_reserve();
6040 setup_min_unmapped_ratio();
6041 setup_min_slab_ratio();
6044 khugepaged_min_free_kbytes_update();
6048 postcore_initcall(init_per_zone_wmark_min)
6051 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6052 * that we can call two helper functions whenever min_free_kbytes
6055 static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6056 void *buffer, size_t *length, loff_t *ppos)
6060 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6065 user_min_free_kbytes = min_free_kbytes;
6066 setup_per_zone_wmarks();
6071 static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6072 void *buffer, size_t *length, loff_t *ppos)
6076 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6081 setup_per_zone_wmarks();
6087 static void setup_min_unmapped_ratio(void)
6092 for_each_online_pgdat(pgdat)
6093 pgdat->min_unmapped_pages = 0;
6096 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6097 sysctl_min_unmapped_ratio) / 100;
6101 static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6102 void *buffer, size_t *length, loff_t *ppos)
6106 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6110 setup_min_unmapped_ratio();
6115 static void setup_min_slab_ratio(void)
6120 for_each_online_pgdat(pgdat)
6121 pgdat->min_slab_pages = 0;
6124 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6125 sysctl_min_slab_ratio) / 100;
6128 static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6129 void *buffer, size_t *length, loff_t *ppos)
6133 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6137 setup_min_slab_ratio();
6144 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6145 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6146 * whenever sysctl_lowmem_reserve_ratio changes.
6148 * The reserve ratio obviously has absolutely no relation with the
6149 * minimum watermarks. The lowmem reserve ratio can only make sense
6150 * if in function of the boot time zone sizes.
6152 static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table,
6153 int write, void *buffer, size_t *length, loff_t *ppos)
6157 proc_dointvec_minmax(table, write, buffer, length, ppos);
6159 for (i = 0; i < MAX_NR_ZONES; i++) {
6160 if (sysctl_lowmem_reserve_ratio[i] < 1)
6161 sysctl_lowmem_reserve_ratio[i] = 0;
6164 setup_per_zone_lowmem_reserve();
6169 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6170 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6171 * pagelist can have before it gets flushed back to buddy allocator.
6173 static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
6174 int write, void *buffer, size_t *length, loff_t *ppos)
6177 int old_percpu_pagelist_high_fraction;
6180 mutex_lock(&pcp_batch_high_lock);
6181 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6183 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6184 if (!write || ret < 0)
6187 /* Sanity checking to avoid pcp imbalance */
6188 if (percpu_pagelist_high_fraction &&
6189 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6190 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6196 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6199 for_each_populated_zone(zone)
6200 zone_set_pageset_high_and_batch(zone, 0);
6202 mutex_unlock(&pcp_batch_high_lock);
6206 static struct ctl_table page_alloc_sysctl_table[] = {
6208 .procname = "min_free_kbytes",
6209 .data = &min_free_kbytes,
6210 .maxlen = sizeof(min_free_kbytes),
6212 .proc_handler = min_free_kbytes_sysctl_handler,
6213 .extra1 = SYSCTL_ZERO,
6216 .procname = "watermark_boost_factor",
6217 .data = &watermark_boost_factor,
6218 .maxlen = sizeof(watermark_boost_factor),
6220 .proc_handler = proc_dointvec_minmax,
6221 .extra1 = SYSCTL_ZERO,
6224 .procname = "watermark_scale_factor",
6225 .data = &watermark_scale_factor,
6226 .maxlen = sizeof(watermark_scale_factor),
6228 .proc_handler = watermark_scale_factor_sysctl_handler,
6229 .extra1 = SYSCTL_ONE,
6230 .extra2 = SYSCTL_THREE_THOUSAND,
6233 .procname = "percpu_pagelist_high_fraction",
6234 .data = &percpu_pagelist_high_fraction,
6235 .maxlen = sizeof(percpu_pagelist_high_fraction),
6237 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6238 .extra1 = SYSCTL_ZERO,
6241 .procname = "lowmem_reserve_ratio",
6242 .data = &sysctl_lowmem_reserve_ratio,
6243 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6245 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6249 .procname = "numa_zonelist_order",
6250 .data = &numa_zonelist_order,
6251 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6253 .proc_handler = numa_zonelist_order_handler,
6256 .procname = "min_unmapped_ratio",
6257 .data = &sysctl_min_unmapped_ratio,
6258 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6260 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6261 .extra1 = SYSCTL_ZERO,
6262 .extra2 = SYSCTL_ONE_HUNDRED,
6265 .procname = "min_slab_ratio",
6266 .data = &sysctl_min_slab_ratio,
6267 .maxlen = sizeof(sysctl_min_slab_ratio),
6269 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6270 .extra1 = SYSCTL_ZERO,
6271 .extra2 = SYSCTL_ONE_HUNDRED,
6276 void __init page_alloc_sysctl_init(void)
6278 register_sysctl_init("vm", page_alloc_sysctl_table);
6281 #ifdef CONFIG_CONTIG_ALLOC
6282 /* Usage: See admin-guide/dynamic-debug-howto.rst */
6283 static void alloc_contig_dump_pages(struct list_head *page_list)
6285 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6287 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6291 list_for_each_entry(page, page_list, lru)
6292 dump_page(page, "migration failure");
6297 * [start, end) must belong to a single zone.
6298 * @migratetype: using migratetype to filter the type of migration in
6299 * trace_mm_alloc_contig_migrate_range_info.
6301 int __alloc_contig_migrate_range(struct compact_control *cc,
6302 unsigned long start, unsigned long end,
6305 /* This function is based on compact_zone() from compaction.c. */
6306 unsigned int nr_reclaimed;
6307 unsigned long pfn = start;
6308 unsigned int tries = 0;
6310 struct migration_target_control mtc = {
6311 .nid = zone_to_nid(cc->zone),
6312 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6313 .reason = MR_CONTIG_RANGE,
6316 unsigned long total_mapped = 0;
6317 unsigned long total_migrated = 0;
6318 unsigned long total_reclaimed = 0;
6320 lru_cache_disable();
6322 while (pfn < end || !list_empty(&cc->migratepages)) {
6323 if (fatal_signal_pending(current)) {
6328 if (list_empty(&cc->migratepages)) {
6329 cc->nr_migratepages = 0;
6330 ret = isolate_migratepages_range(cc, pfn, end);
6331 if (ret && ret != -EAGAIN)
6333 pfn = cc->migrate_pfn;
6335 } else if (++tries == 5) {
6340 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6342 cc->nr_migratepages -= nr_reclaimed;
6344 if (trace_mm_alloc_contig_migrate_range_info_enabled()) {
6345 total_reclaimed += nr_reclaimed;
6346 list_for_each_entry(page, &cc->migratepages, lru)
6347 total_mapped += page_mapcount(page);
6350 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6351 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6353 if (trace_mm_alloc_contig_migrate_range_info_enabled() && !ret)
6354 total_migrated += cc->nr_migratepages;
6357 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6358 * to retry again over this error, so do the same here.
6366 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6367 alloc_contig_dump_pages(&cc->migratepages);
6368 putback_movable_pages(&cc->migratepages);
6371 trace_mm_alloc_contig_migrate_range_info(start, end, migratetype,
6375 return (ret < 0) ? ret : 0;
6379 * alloc_contig_range() -- tries to allocate given range of pages
6380 * @start: start PFN to allocate
6381 * @end: one-past-the-last PFN to allocate
6382 * @migratetype: migratetype of the underlying pageblocks (either
6383 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6384 * in range must have the same migratetype and it must
6385 * be either of the two.
6386 * @gfp_mask: GFP mask to use during compaction
6388 * The PFN range does not have to be pageblock aligned. The PFN range must
6389 * belong to a single zone.
6391 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6392 * pageblocks in the range. Once isolated, the pageblocks should not
6393 * be modified by others.
6395 * Return: zero on success or negative error code. On success all
6396 * pages which PFN is in [start, end) are allocated for the caller and
6397 * need to be freed with free_contig_range().
6399 int alloc_contig_range_noprof(unsigned long start, unsigned long end,
6400 unsigned migratetype, gfp_t gfp_mask)
6402 unsigned long outer_start, outer_end;
6405 struct compact_control cc = {
6406 .nr_migratepages = 0,
6408 .zone = page_zone(pfn_to_page(start)),
6409 .mode = MIGRATE_SYNC,
6410 .ignore_skip_hint = true,
6411 .no_set_skip_hint = true,
6412 .gfp_mask = current_gfp_context(gfp_mask),
6413 .alloc_contig = true,
6415 INIT_LIST_HEAD(&cc.migratepages);
6418 * What we do here is we mark all pageblocks in range as
6419 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6420 * have different sizes, and due to the way page allocator
6421 * work, start_isolate_page_range() has special handlings for this.
6423 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6424 * migrate the pages from an unaligned range (ie. pages that
6425 * we are interested in). This will put all the pages in
6426 * range back to page allocator as MIGRATE_ISOLATE.
6428 * When this is done, we take the pages in range from page
6429 * allocator removing them from the buddy system. This way
6430 * page allocator will never consider using them.
6432 * This lets us mark the pageblocks back as
6433 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6434 * aligned range but not in the unaligned, original range are
6435 * put back to page allocator so that buddy can use them.
6438 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6442 drain_all_pages(cc.zone);
6445 * In case of -EBUSY, we'd like to know which page causes problem.
6446 * So, just fall through. test_pages_isolated() has a tracepoint
6447 * which will report the busy page.
6449 * It is possible that busy pages could become available before
6450 * the call to test_pages_isolated, and the range will actually be
6451 * allocated. So, if we fall through be sure to clear ret so that
6452 * -EBUSY is not accidentally used or returned to caller.
6454 ret = __alloc_contig_migrate_range(&cc, start, end, migratetype);
6455 if (ret && ret != -EBUSY)
6460 * Pages from [start, end) are within a pageblock_nr_pages
6461 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6462 * more, all pages in [start, end) are free in page allocator.
6463 * What we are going to do is to allocate all pages from
6464 * [start, end) (that is remove them from page allocator).
6466 * The only problem is that pages at the beginning and at the
6467 * end of interesting range may be not aligned with pages that
6468 * page allocator holds, ie. they can be part of higher order
6469 * pages. Because of this, we reserve the bigger range and
6470 * once this is done free the pages we are not interested in.
6472 * We don't have to hold zone->lock here because the pages are
6473 * isolated thus they won't get removed from buddy.
6475 outer_start = find_large_buddy(start);
6477 /* Make sure the range is really isolated. */
6478 if (test_pages_isolated(outer_start, end, 0)) {
6483 /* Grab isolated pages from freelists. */
6484 outer_end = isolate_freepages_range(&cc, outer_start, end);
6490 /* Free head and tail (if any) */
6491 if (start != outer_start)
6492 free_contig_range(outer_start, start - outer_start);
6493 if (end != outer_end)
6494 free_contig_range(end, outer_end - end);
6497 undo_isolate_page_range(start, end, migratetype);
6500 EXPORT_SYMBOL(alloc_contig_range_noprof);
6502 static int __alloc_contig_pages(unsigned long start_pfn,
6503 unsigned long nr_pages, gfp_t gfp_mask)
6505 unsigned long end_pfn = start_pfn + nr_pages;
6507 return alloc_contig_range_noprof(start_pfn, end_pfn, MIGRATE_MOVABLE,
6511 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6512 unsigned long nr_pages)
6514 unsigned long i, end_pfn = start_pfn + nr_pages;
6517 for (i = start_pfn; i < end_pfn; i++) {
6518 page = pfn_to_online_page(i);
6522 if (page_zone(page) != z)
6525 if (PageReserved(page))
6534 static bool zone_spans_last_pfn(const struct zone *zone,
6535 unsigned long start_pfn, unsigned long nr_pages)
6537 unsigned long last_pfn = start_pfn + nr_pages - 1;
6539 return zone_spans_pfn(zone, last_pfn);
6543 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6544 * @nr_pages: Number of contiguous pages to allocate
6545 * @gfp_mask: GFP mask to limit search and used during compaction
6547 * @nodemask: Mask for other possible nodes
6549 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6550 * on an applicable zonelist to find a contiguous pfn range which can then be
6551 * tried for allocation with alloc_contig_range(). This routine is intended
6552 * for allocation requests which can not be fulfilled with the buddy allocator.
6554 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6555 * power of two, then allocated range is also guaranteed to be aligned to same
6556 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6558 * Allocated pages can be freed with free_contig_range() or by manually calling
6559 * __free_page() on each allocated page.
6561 * Return: pointer to contiguous pages on success, or NULL if not successful.
6563 struct page *alloc_contig_pages_noprof(unsigned long nr_pages, gfp_t gfp_mask,
6564 int nid, nodemask_t *nodemask)
6566 unsigned long ret, pfn, flags;
6567 struct zonelist *zonelist;
6571 zonelist = node_zonelist(nid, gfp_mask);
6572 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6573 gfp_zone(gfp_mask), nodemask) {
6574 spin_lock_irqsave(&zone->lock, flags);
6576 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6577 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6578 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6580 * We release the zone lock here because
6581 * alloc_contig_range() will also lock the zone
6582 * at some point. If there's an allocation
6583 * spinning on this lock, it may win the race
6584 * and cause alloc_contig_range() to fail...
6586 spin_unlock_irqrestore(&zone->lock, flags);
6587 ret = __alloc_contig_pages(pfn, nr_pages,
6590 return pfn_to_page(pfn);
6591 spin_lock_irqsave(&zone->lock, flags);
6595 spin_unlock_irqrestore(&zone->lock, flags);
6599 #endif /* CONFIG_CONTIG_ALLOC */
6601 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6603 unsigned long count = 0;
6605 for (; nr_pages--; pfn++) {
6606 struct page *page = pfn_to_page(pfn);
6608 count += page_count(page) != 1;
6611 WARN(count != 0, "%lu pages are still in use!\n", count);
6613 EXPORT_SYMBOL(free_contig_range);
6616 * Effectively disable pcplists for the zone by setting the high limit to 0
6617 * and draining all cpus. A concurrent page freeing on another CPU that's about
6618 * to put the page on pcplist will either finish before the drain and the page
6619 * will be drained, or observe the new high limit and skip the pcplist.
6621 * Must be paired with a call to zone_pcp_enable().
6623 void zone_pcp_disable(struct zone *zone)
6625 mutex_lock(&pcp_batch_high_lock);
6626 __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
6627 __drain_all_pages(zone, true);
6630 void zone_pcp_enable(struct zone *zone)
6632 __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
6633 zone->pageset_high_max, zone->pageset_batch);
6634 mutex_unlock(&pcp_batch_high_lock);
6637 void zone_pcp_reset(struct zone *zone)
6640 struct per_cpu_zonestat *pzstats;
6642 if (zone->per_cpu_pageset != &boot_pageset) {
6643 for_each_online_cpu(cpu) {
6644 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6645 drain_zonestat(zone, pzstats);
6647 free_percpu(zone->per_cpu_pageset);
6648 zone->per_cpu_pageset = &boot_pageset;
6649 if (zone->per_cpu_zonestats != &boot_zonestats) {
6650 free_percpu(zone->per_cpu_zonestats);
6651 zone->per_cpu_zonestats = &boot_zonestats;
6656 #ifdef CONFIG_MEMORY_HOTREMOVE
6658 * All pages in the range must be in a single zone, must not contain holes,
6659 * must span full sections, and must be isolated before calling this function.
6661 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6663 unsigned long pfn = start_pfn;
6667 unsigned long flags;
6669 offline_mem_sections(pfn, end_pfn);
6670 zone = page_zone(pfn_to_page(pfn));
6671 spin_lock_irqsave(&zone->lock, flags);
6672 while (pfn < end_pfn) {
6673 page = pfn_to_page(pfn);
6675 * The HWPoisoned page may be not in buddy system, and
6676 * page_count() is not 0.
6678 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6683 * At this point all remaining PageOffline() pages have a
6684 * reference count of 0 and can simply be skipped.
6686 if (PageOffline(page)) {
6687 BUG_ON(page_count(page));
6688 BUG_ON(PageBuddy(page));
6693 BUG_ON(page_count(page));
6694 BUG_ON(!PageBuddy(page));
6695 VM_WARN_ON(get_pageblock_migratetype(page) != MIGRATE_ISOLATE);
6696 order = buddy_order(page);
6697 del_page_from_free_list(page, zone, order, MIGRATE_ISOLATE);
6698 pfn += (1 << order);
6700 spin_unlock_irqrestore(&zone->lock, flags);
6705 * This function returns a stable result only if called under zone lock.
6707 bool is_free_buddy_page(const struct page *page)
6709 unsigned long pfn = page_to_pfn(page);
6712 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6713 const struct page *head = page - (pfn & ((1 << order) - 1));
6715 if (PageBuddy(head) &&
6716 buddy_order_unsafe(head) >= order)
6720 return order <= MAX_PAGE_ORDER;
6722 EXPORT_SYMBOL(is_free_buddy_page);
6724 #ifdef CONFIG_MEMORY_FAILURE
6725 static inline void add_to_free_list(struct page *page, struct zone *zone,
6726 unsigned int order, int migratetype,
6729 __add_to_free_list(page, zone, order, migratetype, tail);
6730 account_freepages(zone, 1 << order, migratetype);
6734 * Break down a higher-order page in sub-pages, and keep our target out of
6737 static void break_down_buddy_pages(struct zone *zone, struct page *page,
6738 struct page *target, int low, int high,
6741 unsigned long size = 1 << high;
6742 struct page *current_buddy;
6744 while (high > low) {
6748 if (target >= &page[size]) {
6749 current_buddy = page;
6752 current_buddy = page + size;
6755 if (set_page_guard(zone, current_buddy, high))
6758 add_to_free_list(current_buddy, zone, high, migratetype, false);
6759 set_buddy_order(current_buddy, high);
6764 * Take a page that will be marked as poisoned off the buddy allocator.
6766 bool take_page_off_buddy(struct page *page)
6768 struct zone *zone = page_zone(page);
6769 unsigned long pfn = page_to_pfn(page);
6770 unsigned long flags;
6774 spin_lock_irqsave(&zone->lock, flags);
6775 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6776 struct page *page_head = page - (pfn & ((1 << order) - 1));
6777 int page_order = buddy_order(page_head);
6779 if (PageBuddy(page_head) && page_order >= order) {
6780 unsigned long pfn_head = page_to_pfn(page_head);
6781 int migratetype = get_pfnblock_migratetype(page_head,
6784 del_page_from_free_list(page_head, zone, page_order,
6786 break_down_buddy_pages(zone, page_head, page, 0,
6787 page_order, migratetype);
6788 SetPageHWPoisonTakenOff(page);
6792 if (page_count(page_head) > 0)
6795 spin_unlock_irqrestore(&zone->lock, flags);
6800 * Cancel takeoff done by take_page_off_buddy().
6802 bool put_page_back_buddy(struct page *page)
6804 struct zone *zone = page_zone(page);
6805 unsigned long flags;
6808 spin_lock_irqsave(&zone->lock, flags);
6809 if (put_page_testzero(page)) {
6810 unsigned long pfn = page_to_pfn(page);
6811 int migratetype = get_pfnblock_migratetype(page, pfn);
6813 ClearPageHWPoisonTakenOff(page);
6814 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6815 if (TestClearPageHWPoison(page)) {
6819 spin_unlock_irqrestore(&zone->lock, flags);
6825 #ifdef CONFIG_ZONE_DMA
6826 bool has_managed_dma(void)
6828 struct pglist_data *pgdat;
6830 for_each_online_pgdat(pgdat) {
6831 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6833 if (managed_zone(zone))
6838 #endif /* CONFIG_ZONE_DMA */
6840 #ifdef CONFIG_UNACCEPTED_MEMORY
6842 /* Counts number of zones with unaccepted pages. */
6843 static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6845 static bool lazy_accept = true;
6847 static int __init accept_memory_parse(char *p)
6849 if (!strcmp(p, "lazy")) {
6852 } else if (!strcmp(p, "eager")) {
6853 lazy_accept = false;
6859 early_param("accept_memory", accept_memory_parse);
6861 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6863 phys_addr_t start = page_to_phys(page);
6864 phys_addr_t end = start + (PAGE_SIZE << order);
6866 return range_contains_unaccepted_memory(start, end);
6869 static void accept_page(struct page *page, unsigned int order)
6871 phys_addr_t start = page_to_phys(page);
6873 accept_memory(start, start + (PAGE_SIZE << order));
6876 static bool try_to_accept_memory_one(struct zone *zone)
6878 unsigned long flags;
6882 if (list_empty(&zone->unaccepted_pages))
6885 spin_lock_irqsave(&zone->lock, flags);
6886 page = list_first_entry_or_null(&zone->unaccepted_pages,
6889 spin_unlock_irqrestore(&zone->lock, flags);
6893 list_del(&page->lru);
6894 last = list_empty(&zone->unaccepted_pages);
6896 account_freepages(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6897 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6898 spin_unlock_irqrestore(&zone->lock, flags);
6900 accept_page(page, MAX_PAGE_ORDER);
6902 __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
6905 static_branch_dec(&zones_with_unaccepted_pages);
6910 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6915 /* How much to accept to get to high watermark? */
6916 to_accept = high_wmark_pages(zone) -
6917 (zone_page_state(zone, NR_FREE_PAGES) -
6918 __zone_watermark_unusable_free(zone, order, 0));
6920 /* Accept at least one page */
6922 if (!try_to_accept_memory_one(zone))
6925 to_accept -= MAX_ORDER_NR_PAGES;
6926 } while (to_accept > 0);
6931 static inline bool has_unaccepted_memory(void)
6933 return static_branch_unlikely(&zones_with_unaccepted_pages);
6936 static bool __free_unaccepted(struct page *page)
6938 struct zone *zone = page_zone(page);
6939 unsigned long flags;
6945 spin_lock_irqsave(&zone->lock, flags);
6946 first = list_empty(&zone->unaccepted_pages);
6947 list_add_tail(&page->lru, &zone->unaccepted_pages);
6948 account_freepages(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6949 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
6950 spin_unlock_irqrestore(&zone->lock, flags);
6953 static_branch_inc(&zones_with_unaccepted_pages);
6960 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6965 static void accept_page(struct page *page, unsigned int order)
6969 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6974 static inline bool has_unaccepted_memory(void)
6979 static bool __free_unaccepted(struct page *page)
6985 #endif /* CONFIG_UNACCEPTED_MEMORY */