1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/interrupt.h>
22 #include <linux/jiffies.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.h>
25 #include <linux/kasan.h>
26 #include <linux/kmsan.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/ratelimit.h>
30 #include <linux/oom.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/memory_hotplug.h>
36 #include <linux/nodemask.h>
37 #include <linux/vmstat.h>
38 #include <linux/fault-inject.h>
39 #include <linux/compaction.h>
40 #include <trace/events/kmem.h>
41 #include <trace/events/oom.h>
42 #include <linux/prefetch.h>
43 #include <linux/mm_inline.h>
44 #include <linux/mmu_notifier.h>
45 #include <linux/migrate.h>
46 #include <linux/sched/mm.h>
47 #include <linux/page_owner.h>
48 #include <linux/page_table_check.h>
49 #include <linux/memcontrol.h>
50 #include <linux/ftrace.h>
51 #include <linux/lockdep.h>
52 #include <linux/psi.h>
53 #include <linux/khugepaged.h>
54 #include <linux/delayacct.h>
55 #include <asm/div64.h>
58 #include "page_reporting.h"
60 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
61 typedef int __bitwise fpi_t;
63 /* No special request */
64 #define FPI_NONE ((__force fpi_t)0)
67 * Skip free page reporting notification for the (possibly merged) page.
68 * This does not hinder free page reporting from grabbing the page,
69 * reporting it and marking it "reported" - it only skips notifying
70 * the free page reporting infrastructure about a newly freed page. For
71 * example, used when temporarily pulling a page from a freelist and
72 * putting it back unmodified.
74 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
77 * Place the (possibly merged) page to the tail of the freelist. Will ignore
78 * page shuffling (relevant code - e.g., memory onlining - is expected to
79 * shuffle the whole zone).
81 * Note: No code should rely on this flag for correctness - it's purely
82 * to allow for optimizations when handing back either fresh pages
83 * (memory onlining) or untouched pages (page isolation, free page
86 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
88 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
89 static DEFINE_MUTEX(pcp_batch_high_lock);
90 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
92 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
94 * On SMP, spin_trylock is sufficient protection.
95 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
97 #define pcp_trylock_prepare(flags) do { } while (0)
98 #define pcp_trylock_finish(flag) do { } while (0)
101 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
102 #define pcp_trylock_prepare(flags) local_irq_save(flags)
103 #define pcp_trylock_finish(flags) local_irq_restore(flags)
107 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
108 * a migration causing the wrong PCP to be locked and remote memory being
109 * potentially allocated, pin the task to the CPU for the lookup+lock.
110 * preempt_disable is used on !RT because it is faster than migrate_disable.
111 * migrate_disable is used on RT because otherwise RT spinlock usage is
112 * interfered with and a high priority task cannot preempt the allocator.
114 #ifndef CONFIG_PREEMPT_RT
115 #define pcpu_task_pin() preempt_disable()
116 #define pcpu_task_unpin() preempt_enable()
118 #define pcpu_task_pin() migrate_disable()
119 #define pcpu_task_unpin() migrate_enable()
123 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
124 * Return value should be used with equivalent unlock helper.
126 #define pcpu_spin_lock(type, member, ptr) \
130 _ret = this_cpu_ptr(ptr); \
131 spin_lock(&_ret->member); \
135 #define pcpu_spin_trylock(type, member, ptr) \
139 _ret = this_cpu_ptr(ptr); \
140 if (!spin_trylock(&_ret->member)) { \
147 #define pcpu_spin_unlock(member, ptr) \
149 spin_unlock(&ptr->member); \
153 /* struct per_cpu_pages specific helpers. */
154 #define pcp_spin_lock(ptr) \
155 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
157 #define pcp_spin_trylock(ptr) \
158 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
160 #define pcp_spin_unlock(ptr) \
161 pcpu_spin_unlock(lock, ptr)
163 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
164 DEFINE_PER_CPU(int, numa_node);
165 EXPORT_PER_CPU_SYMBOL(numa_node);
168 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
170 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
172 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
173 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
174 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
175 * defined in <linux/topology.h>.
177 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
178 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
181 static DEFINE_MUTEX(pcpu_drain_mutex);
183 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
184 volatile unsigned long latent_entropy __latent_entropy;
185 EXPORT_SYMBOL(latent_entropy);
189 * Array of node states.
191 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
192 [N_POSSIBLE] = NODE_MASK_ALL,
193 [N_ONLINE] = { { [0] = 1UL } },
195 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
196 #ifdef CONFIG_HIGHMEM
197 [N_HIGH_MEMORY] = { { [0] = 1UL } },
199 [N_MEMORY] = { { [0] = 1UL } },
200 [N_CPU] = { { [0] = 1UL } },
203 EXPORT_SYMBOL(node_states);
205 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
207 #define ALLOC_IN_CMA_THRESHOLD_MAX 16
208 #define ALLOC_IN_CMA_THRESHOLD_DEFAULT 12
210 static unsigned long _alloc_in_cma_threshold __read_mostly
211 = ALLOC_IN_CMA_THRESHOLD_DEFAULT;
213 static int __init alloc_in_cma_threshold_setup(char *buf)
217 if (kstrtoul(buf, 10, &res) < 0 ||
218 res > ALLOC_IN_CMA_THRESHOLD_MAX) {
219 pr_err("Bad alloc_cma_threshold value\n");
222 _alloc_in_cma_threshold = res;
223 pr_info("Setting alloc_in_cma_threshold to %lu\n", res);
226 early_param("alloc_in_cma_threshold", alloc_in_cma_threshold_setup);
229 * A cached value of the page's pageblock's migratetype, used when the page is
230 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
231 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
232 * Also the migratetype set in the page does not necessarily match the pcplist
233 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
234 * other index - this ensures that it will be put on the correct CMA freelist.
236 static inline int get_pcppage_migratetype(struct page *page)
241 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
243 page->index = migratetype;
246 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
247 unsigned int pageblock_order __read_mostly;
250 static void __free_pages_ok(struct page *page, unsigned int order,
254 * results with 256, 32 in the lowmem_reserve sysctl:
255 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
256 * 1G machine -> (16M dma, 784M normal, 224M high)
257 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
258 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
259 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
261 * TBD: should special case ZONE_DMA32 machines here - in those we normally
262 * don't need any ZONE_NORMAL reservation
264 static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
265 #ifdef CONFIG_ZONE_DMA
268 #ifdef CONFIG_ZONE_DMA32
272 #ifdef CONFIG_HIGHMEM
278 char * const zone_names[MAX_NR_ZONES] = {
279 #ifdef CONFIG_ZONE_DMA
282 #ifdef CONFIG_ZONE_DMA32
286 #ifdef CONFIG_HIGHMEM
290 #ifdef CONFIG_ZONE_DEVICE
295 const char * const migratetype_names[MIGRATE_TYPES] = {
303 #ifdef CONFIG_MEMORY_ISOLATION
308 int min_free_kbytes = 1024;
309 int user_min_free_kbytes = -1;
310 static int watermark_boost_factor __read_mostly = 15000;
311 static int watermark_scale_factor = 10;
313 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
315 EXPORT_SYMBOL(movable_zone);
318 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
319 unsigned int nr_online_nodes __read_mostly = 1;
320 EXPORT_SYMBOL(nr_node_ids);
321 EXPORT_SYMBOL(nr_online_nodes);
324 static bool page_contains_unaccepted(struct page *page, unsigned int order);
325 static void accept_page(struct page *page, unsigned int order);
326 static bool try_to_accept_memory(struct zone *zone, unsigned int order);
327 static inline bool has_unaccepted_memory(void);
328 static bool __free_unaccepted(struct page *page);
330 int page_group_by_mobility_disabled __read_mostly;
332 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
334 * During boot we initialize deferred pages on-demand, as needed, but once
335 * page_alloc_init_late() has finished, the deferred pages are all initialized,
336 * and we can permanently disable that path.
338 DEFINE_STATIC_KEY_TRUE(deferred_pages);
340 static inline bool deferred_pages_enabled(void)
342 return static_branch_unlikely(&deferred_pages);
346 * deferred_grow_zone() is __init, but it is called from
347 * get_page_from_freelist() during early boot until deferred_pages permanently
348 * disables this call. This is why we have refdata wrapper to avoid warning,
349 * and to ensure that the function body gets unloaded.
352 _deferred_grow_zone(struct zone *zone, unsigned int order)
354 return deferred_grow_zone(zone, order);
357 static inline bool deferred_pages_enabled(void)
361 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
363 /* Return a pointer to the bitmap storing bits affecting a block of pages */
364 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
367 #ifdef CONFIG_SPARSEMEM
368 return section_to_usemap(__pfn_to_section(pfn));
370 return page_zone(page)->pageblock_flags;
371 #endif /* CONFIG_SPARSEMEM */
374 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
376 #ifdef CONFIG_SPARSEMEM
377 pfn &= (PAGES_PER_SECTION-1);
379 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
380 #endif /* CONFIG_SPARSEMEM */
381 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
385 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
386 * @page: The page within the block of interest
387 * @pfn: The target page frame number
388 * @mask: mask of bits that the caller is interested in
390 * Return: pageblock_bits flags
392 unsigned long get_pfnblock_flags_mask(const struct page *page,
393 unsigned long pfn, unsigned long mask)
395 unsigned long *bitmap;
396 unsigned long bitidx, word_bitidx;
399 bitmap = get_pageblock_bitmap(page, pfn);
400 bitidx = pfn_to_bitidx(page, pfn);
401 word_bitidx = bitidx / BITS_PER_LONG;
402 bitidx &= (BITS_PER_LONG-1);
404 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
405 * a consistent read of the memory array, so that results, even though
406 * racy, are not corrupted.
408 word = READ_ONCE(bitmap[word_bitidx]);
409 return (word >> bitidx) & mask;
412 static __always_inline int get_pfnblock_migratetype(const struct page *page,
415 return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
419 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
420 * @page: The page within the block of interest
421 * @flags: The flags to set
422 * @pfn: The target page frame number
423 * @mask: mask of bits that the caller is interested in
425 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
429 unsigned long *bitmap;
430 unsigned long bitidx, word_bitidx;
433 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
434 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
436 bitmap = get_pageblock_bitmap(page, pfn);
437 bitidx = pfn_to_bitidx(page, pfn);
438 word_bitidx = bitidx / BITS_PER_LONG;
439 bitidx &= (BITS_PER_LONG-1);
441 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
446 word = READ_ONCE(bitmap[word_bitidx]);
448 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
451 void set_pageblock_migratetype(struct page *page, int migratetype)
453 if (unlikely(page_group_by_mobility_disabled &&
454 migratetype < MIGRATE_PCPTYPES))
455 migratetype = MIGRATE_UNMOVABLE;
457 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
458 page_to_pfn(page), MIGRATETYPE_MASK);
461 #ifdef CONFIG_DEBUG_VM
462 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
466 unsigned long pfn = page_to_pfn(page);
467 unsigned long sp, start_pfn;
470 seq = zone_span_seqbegin(zone);
471 start_pfn = zone->zone_start_pfn;
472 sp = zone->spanned_pages;
473 ret = !zone_spans_pfn(zone, pfn);
474 } while (zone_span_seqretry(zone, seq));
477 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
478 pfn, zone_to_nid(zone), zone->name,
479 start_pfn, start_pfn + sp);
485 * Temporary debugging check for pages not lying within a given zone.
487 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
489 if (page_outside_zone_boundaries(zone, page))
491 if (zone != page_zone(page))
497 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
503 static void bad_page(struct page *page, const char *reason)
505 static unsigned long resume;
506 static unsigned long nr_shown;
507 static unsigned long nr_unshown;
510 * Allow a burst of 60 reports, then keep quiet for that minute;
511 * or allow a steady drip of one report per second.
513 if (nr_shown == 60) {
514 if (time_before(jiffies, resume)) {
520 "BUG: Bad page state: %lu messages suppressed\n",
527 resume = jiffies + 60 * HZ;
529 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
530 current->comm, page_to_pfn(page));
531 dump_page(page, reason);
536 /* Leave bad fields for debug, except PageBuddy could make trouble */
537 page_mapcount_reset(page); /* remove PageBuddy */
538 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
541 static inline unsigned int order_to_pindex(int migratetype, int order)
543 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
544 if (order > PAGE_ALLOC_COSTLY_ORDER) {
545 VM_BUG_ON(order != pageblock_order);
546 return NR_LOWORDER_PCP_LISTS;
549 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
552 return (MIGRATE_PCPTYPES * order) + migratetype;
555 static inline int pindex_to_order(unsigned int pindex)
557 int order = pindex / MIGRATE_PCPTYPES;
559 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
560 if (pindex == NR_LOWORDER_PCP_LISTS)
561 order = pageblock_order;
563 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
569 static inline bool pcp_allowed_order(unsigned int order)
571 if (order <= PAGE_ALLOC_COSTLY_ORDER)
573 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
574 if (order == pageblock_order)
580 static inline void free_the_page(struct page *page, unsigned int order)
582 if (pcp_allowed_order(order)) /* Via pcp? */
583 free_unref_page(page, order);
585 __free_pages_ok(page, order, FPI_NONE);
589 * Higher-order pages are called "compound pages". They are structured thusly:
591 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
593 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
594 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
596 * The first tail page's ->compound_order holds the order of allocation.
597 * This usage means that zero-order pages may not be compound.
600 void prep_compound_page(struct page *page, unsigned int order)
603 int nr_pages = 1 << order;
606 for (i = 1; i < nr_pages; i++)
607 prep_compound_tail(page, i);
609 prep_compound_head(page, order);
612 void destroy_large_folio(struct folio *folio)
614 if (folio_test_hugetlb(folio)) {
615 free_huge_folio(folio);
619 if (folio_test_large_rmappable(folio))
620 folio_undo_large_rmappable(folio);
622 mem_cgroup_uncharge(folio);
623 free_the_page(&folio->page, folio_order(folio));
626 static inline void set_buddy_order(struct page *page, unsigned int order)
628 set_page_private(page, order);
629 __SetPageBuddy(page);
632 #ifdef CONFIG_COMPACTION
633 static inline struct capture_control *task_capc(struct zone *zone)
635 struct capture_control *capc = current->capture_control;
637 return unlikely(capc) &&
638 !(current->flags & PF_KTHREAD) &&
640 capc->cc->zone == zone ? capc : NULL;
644 compaction_capture(struct capture_control *capc, struct page *page,
645 int order, int migratetype)
647 if (!capc || order != capc->cc->order)
650 /* Do not accidentally pollute CMA or isolated regions*/
651 if (is_migrate_cma(migratetype) ||
652 is_migrate_isolate(migratetype))
656 * Do not let lower order allocations pollute a movable pageblock.
657 * This might let an unmovable request use a reclaimable pageblock
658 * and vice-versa but no more than normal fallback logic which can
659 * have trouble finding a high-order free page.
661 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
669 static inline struct capture_control *task_capc(struct zone *zone)
675 compaction_capture(struct capture_control *capc, struct page *page,
676 int order, int migratetype)
680 #endif /* CONFIG_COMPACTION */
682 /* Used for pages not on another list */
683 static inline void add_to_free_list(struct page *page, struct zone *zone,
684 unsigned int order, int migratetype)
686 struct free_area *area = &zone->free_area[order];
688 list_add(&page->buddy_list, &area->free_list[migratetype]);
692 /* Used for pages not on another list */
693 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
694 unsigned int order, int migratetype)
696 struct free_area *area = &zone->free_area[order];
698 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
703 * Used for pages which are on another list. Move the pages to the tail
704 * of the list - so the moved pages won't immediately be considered for
705 * allocation again (e.g., optimization for memory onlining).
707 static inline void move_to_free_list(struct page *page, struct zone *zone,
708 unsigned int order, int migratetype)
710 struct free_area *area = &zone->free_area[order];
712 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
715 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
718 /* clear reported state and update reported page count */
719 if (page_reported(page))
720 __ClearPageReported(page);
722 list_del(&page->buddy_list);
723 __ClearPageBuddy(page);
724 set_page_private(page, 0);
725 zone->free_area[order].nr_free--;
728 static inline struct page *get_page_from_free_area(struct free_area *area,
731 return list_first_entry_or_null(&area->free_list[migratetype],
732 struct page, buddy_list);
736 * If this is not the largest possible page, check if the buddy
737 * of the next-highest order is free. If it is, it's possible
738 * that pages are being freed that will coalesce soon. In case,
739 * that is happening, add the free page to the tail of the list
740 * so it's less likely to be used soon and more likely to be merged
741 * as a higher order page
744 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
745 struct page *page, unsigned int order)
747 unsigned long higher_page_pfn;
748 struct page *higher_page;
750 if (order >= MAX_ORDER - 1)
753 higher_page_pfn = buddy_pfn & pfn;
754 higher_page = page + (higher_page_pfn - pfn);
756 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
761 * Freeing function for a buddy system allocator.
763 * The concept of a buddy system is to maintain direct-mapped table
764 * (containing bit values) for memory blocks of various "orders".
765 * The bottom level table contains the map for the smallest allocatable
766 * units of memory (here, pages), and each level above it describes
767 * pairs of units from the levels below, hence, "buddies".
768 * At a high level, all that happens here is marking the table entry
769 * at the bottom level available, and propagating the changes upward
770 * as necessary, plus some accounting needed to play nicely with other
771 * parts of the VM system.
772 * At each level, we keep a list of pages, which are heads of continuous
773 * free pages of length of (1 << order) and marked with PageBuddy.
774 * Page's order is recorded in page_private(page) field.
775 * So when we are allocating or freeing one, we can derive the state of the
776 * other. That is, if we allocate a small block, and both were
777 * free, the remainder of the region must be split into blocks.
778 * If a block is freed, and its buddy is also free, then this
779 * triggers coalescing into a block of larger size.
784 static inline void __free_one_page(struct page *page,
786 struct zone *zone, unsigned int order,
787 int migratetype, fpi_t fpi_flags)
789 struct capture_control *capc = task_capc(zone);
790 unsigned long buddy_pfn = 0;
791 unsigned long combined_pfn;
795 VM_BUG_ON(!zone_is_initialized(zone));
796 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
798 VM_BUG_ON(migratetype == -1);
799 if (likely(!is_migrate_isolate(migratetype)))
800 __mod_zone_freepage_state(zone, 1 << order, migratetype);
802 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
803 VM_BUG_ON_PAGE(bad_range(zone, page), page);
805 while (order < MAX_ORDER) {
806 if (compaction_capture(capc, page, order, migratetype)) {
807 __mod_zone_freepage_state(zone, -(1 << order),
812 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
816 if (unlikely(order >= pageblock_order)) {
818 * We want to prevent merge between freepages on pageblock
819 * without fallbacks and normal pageblock. Without this,
820 * pageblock isolation could cause incorrect freepage or CMA
821 * accounting or HIGHATOMIC accounting.
823 int buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
825 if (migratetype != buddy_mt
826 && (!migratetype_is_mergeable(migratetype) ||
827 !migratetype_is_mergeable(buddy_mt)))
832 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
833 * merge with it and move up one order.
835 if (page_is_guard(buddy))
836 clear_page_guard(zone, buddy, order, migratetype);
838 del_page_from_free_list(buddy, zone, order);
839 combined_pfn = buddy_pfn & pfn;
840 page = page + (combined_pfn - pfn);
846 set_buddy_order(page, order);
848 if (fpi_flags & FPI_TO_TAIL)
850 else if (is_shuffle_order(order))
851 to_tail = shuffle_pick_tail();
853 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
856 add_to_free_list_tail(page, zone, order, migratetype);
858 add_to_free_list(page, zone, order, migratetype);
860 /* Notify page reporting subsystem of freed page */
861 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
862 page_reporting_notify_free(order);
866 * split_free_page() -- split a free page at split_pfn_offset
867 * @free_page: the original free page
868 * @order: the order of the page
869 * @split_pfn_offset: split offset within the page
871 * Return -ENOENT if the free page is changed, otherwise 0
873 * It is used when the free page crosses two pageblocks with different migratetypes
874 * at split_pfn_offset within the page. The split free page will be put into
875 * separate migratetype lists afterwards. Otherwise, the function achieves
878 int split_free_page(struct page *free_page,
879 unsigned int order, unsigned long split_pfn_offset)
881 struct zone *zone = page_zone(free_page);
882 unsigned long free_page_pfn = page_to_pfn(free_page);
889 if (split_pfn_offset == 0)
892 spin_lock_irqsave(&zone->lock, flags);
894 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
899 mt = get_pfnblock_migratetype(free_page, free_page_pfn);
900 if (likely(!is_migrate_isolate(mt)))
901 __mod_zone_freepage_state(zone, -(1UL << order), mt);
903 del_page_from_free_list(free_page, zone, order);
904 for (pfn = free_page_pfn;
905 pfn < free_page_pfn + (1UL << order);) {
906 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
908 free_page_order = min_t(unsigned int,
909 pfn ? __ffs(pfn) : order,
910 __fls(split_pfn_offset));
911 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
913 pfn += 1UL << free_page_order;
914 split_pfn_offset -= (1UL << free_page_order);
915 /* we have done the first part, now switch to second part */
916 if (split_pfn_offset == 0)
917 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
920 spin_unlock_irqrestore(&zone->lock, flags);
924 * A bad page could be due to a number of fields. Instead of multiple branches,
925 * try and check multiple fields with one check. The caller must do a detailed
926 * check if necessary.
928 static inline bool page_expected_state(struct page *page,
929 unsigned long check_flags)
931 if (unlikely(atomic_read(&page->_mapcount) != -1))
934 if (unlikely((unsigned long)page->mapping |
935 page_ref_count(page) |
939 (page->flags & check_flags)))
945 static const char *page_bad_reason(struct page *page, unsigned long flags)
947 const char *bad_reason = NULL;
949 if (unlikely(atomic_read(&page->_mapcount) != -1))
950 bad_reason = "nonzero mapcount";
951 if (unlikely(page->mapping != NULL))
952 bad_reason = "non-NULL mapping";
953 if (unlikely(page_ref_count(page) != 0))
954 bad_reason = "nonzero _refcount";
955 if (unlikely(page->flags & flags)) {
956 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
957 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
959 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
962 if (unlikely(page->memcg_data))
963 bad_reason = "page still charged to cgroup";
968 static void free_page_is_bad_report(struct page *page)
971 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
974 static inline bool free_page_is_bad(struct page *page)
976 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
979 /* Something has gone sideways, find it */
980 free_page_is_bad_report(page);
984 static inline bool is_check_pages_enabled(void)
986 return static_branch_unlikely(&check_pages_enabled);
989 static int free_tail_page_prepare(struct page *head_page, struct page *page)
991 struct folio *folio = (struct folio *)head_page;
995 * We rely page->lru.next never has bit 0 set, unless the page
996 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
998 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1000 if (!is_check_pages_enabled()) {
1004 switch (page - head_page) {
1006 /* the first tail page: these may be in place of ->mapping */
1007 if (unlikely(folio_entire_mapcount(folio))) {
1008 bad_page(page, "nonzero entire_mapcount");
1011 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
1012 bad_page(page, "nonzero nr_pages_mapped");
1015 if (unlikely(atomic_read(&folio->_pincount))) {
1016 bad_page(page, "nonzero pincount");
1022 * the second tail page: ->mapping is
1023 * deferred_list.next -- ignore value.
1027 if (page->mapping != TAIL_MAPPING) {
1028 bad_page(page, "corrupted mapping in tail page");
1033 if (unlikely(!PageTail(page))) {
1034 bad_page(page, "PageTail not set");
1037 if (unlikely(compound_head(page) != head_page)) {
1038 bad_page(page, "compound_head not consistent");
1043 page->mapping = NULL;
1044 clear_compound_head(page);
1049 * Skip KASAN memory poisoning when either:
1051 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1052 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1053 * using page tags instead (see below).
1054 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1055 * that error detection is disabled for accesses via the page address.
1057 * Pages will have match-all tags in the following circumstances:
1059 * 1. Pages are being initialized for the first time, including during deferred
1060 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1061 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1062 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1063 * 3. The allocation was excluded from being checked due to sampling,
1064 * see the call to kasan_unpoison_pages.
1066 * Poisoning pages during deferred memory init will greatly lengthen the
1067 * process and cause problem in large memory systems as the deferred pages
1068 * initialization is done with interrupt disabled.
1070 * Assuming that there will be no reference to those newly initialized
1071 * pages before they are ever allocated, this should have no effect on
1072 * KASAN memory tracking as the poison will be properly inserted at page
1073 * allocation time. The only corner case is when pages are allocated by
1074 * on-demand allocation and then freed again before the deferred pages
1075 * initialization is done, but this is not likely to happen.
1077 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1079 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1080 return deferred_pages_enabled();
1082 return page_kasan_tag(page) == 0xff;
1085 static void kernel_init_pages(struct page *page, int numpages)
1089 /* s390's use of memset() could override KASAN redzones. */
1090 kasan_disable_current();
1091 for (i = 0; i < numpages; i++)
1092 clear_highpage_kasan_tagged(page + i);
1093 kasan_enable_current();
1096 static __always_inline bool free_pages_prepare(struct page *page,
1097 unsigned int order, fpi_t fpi_flags)
1100 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1101 bool init = want_init_on_free();
1103 VM_BUG_ON_PAGE(PageTail(page), page);
1105 trace_mm_page_free(page, order);
1106 kmsan_free_page(page, order);
1108 if (unlikely(PageHWPoison(page)) && !order) {
1110 * Do not let hwpoison pages hit pcplists/buddy
1111 * Untie memcg state and reset page's owner
1113 if (memcg_kmem_online() && PageMemcgKmem(page))
1114 __memcg_kmem_uncharge_page(page, order);
1115 reset_page_owner(page, order);
1116 page_table_check_free(page, order);
1121 * Check tail pages before head page information is cleared to
1122 * avoid checking PageCompound for order-0 pages.
1124 if (unlikely(order)) {
1125 bool compound = PageCompound(page);
1128 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1131 page[1].flags &= ~PAGE_FLAGS_SECOND;
1132 for (i = 1; i < (1 << order); i++) {
1134 bad += free_tail_page_prepare(page, page + i);
1135 if (is_check_pages_enabled()) {
1136 if (free_page_is_bad(page + i)) {
1141 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1144 if (PageMappingFlags(page))
1145 page->mapping = NULL;
1146 if (memcg_kmem_online() && PageMemcgKmem(page))
1147 __memcg_kmem_uncharge_page(page, order);
1148 if (is_check_pages_enabled()) {
1149 if (free_page_is_bad(page))
1155 page_cpupid_reset_last(page);
1156 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1157 reset_page_owner(page, order);
1158 page_table_check_free(page, order);
1160 if (!PageHighMem(page)) {
1161 debug_check_no_locks_freed(page_address(page),
1162 PAGE_SIZE << order);
1163 debug_check_no_obj_freed(page_address(page),
1164 PAGE_SIZE << order);
1167 kernel_poison_pages(page, 1 << order);
1170 * As memory initialization might be integrated into KASAN,
1171 * KASAN poisoning and memory initialization code must be
1172 * kept together to avoid discrepancies in behavior.
1174 * With hardware tag-based KASAN, memory tags must be set before the
1175 * page becomes unavailable via debug_pagealloc or arch_free_page.
1177 if (!skip_kasan_poison) {
1178 kasan_poison_pages(page, order, init);
1180 /* Memory is already initialized if KASAN did it internally. */
1181 if (kasan_has_integrated_init())
1185 kernel_init_pages(page, 1 << order);
1188 * arch_free_page() can make the page's contents inaccessible. s390
1189 * does this. So nothing which can access the page's contents should
1190 * happen after this.
1192 arch_free_page(page, order);
1194 debug_pagealloc_unmap_pages(page, 1 << order);
1200 * Frees a number of pages from the PCP lists
1201 * Assumes all pages on list are in same zone.
1202 * count is the number of pages to free.
1204 static void free_pcppages_bulk(struct zone *zone, int count,
1205 struct per_cpu_pages *pcp,
1208 unsigned long flags;
1210 bool isolated_pageblocks;
1214 * Ensure proper count is passed which otherwise would stuck in the
1215 * below while (list_empty(list)) loop.
1217 count = min(pcp->count, count);
1219 /* Ensure requested pindex is drained first. */
1220 pindex = pindex - 1;
1222 spin_lock_irqsave(&zone->lock, flags);
1223 isolated_pageblocks = has_isolate_pageblock(zone);
1226 struct list_head *list;
1229 /* Remove pages from lists in a round-robin fashion. */
1231 if (++pindex > NR_PCP_LISTS - 1)
1233 list = &pcp->lists[pindex];
1234 } while (list_empty(list));
1236 order = pindex_to_order(pindex);
1237 nr_pages = 1 << order;
1241 page = list_last_entry(list, struct page, pcp_list);
1242 mt = get_pcppage_migratetype(page);
1244 /* must delete to avoid corrupting pcp list */
1245 list_del(&page->pcp_list);
1247 pcp->count -= nr_pages;
1249 /* MIGRATE_ISOLATE page should not go to pcplists */
1250 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1251 /* Pageblock could have been isolated meanwhile */
1252 if (unlikely(isolated_pageblocks))
1253 mt = get_pageblock_migratetype(page);
1255 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1256 trace_mm_page_pcpu_drain(page, order, mt);
1257 } while (count > 0 && !list_empty(list));
1260 spin_unlock_irqrestore(&zone->lock, flags);
1263 static void free_one_page(struct zone *zone,
1264 struct page *page, unsigned long pfn,
1266 int migratetype, fpi_t fpi_flags)
1268 unsigned long flags;
1270 spin_lock_irqsave(&zone->lock, flags);
1271 if (unlikely(has_isolate_pageblock(zone) ||
1272 is_migrate_isolate(migratetype))) {
1273 migratetype = get_pfnblock_migratetype(page, pfn);
1275 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1276 spin_unlock_irqrestore(&zone->lock, flags);
1279 static void __free_pages_ok(struct page *page, unsigned int order,
1282 unsigned long flags;
1284 unsigned long pfn = page_to_pfn(page);
1285 struct zone *zone = page_zone(page);
1287 if (!free_pages_prepare(page, order, fpi_flags))
1291 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1292 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1293 * This will reduce the lock holding time.
1295 migratetype = get_pfnblock_migratetype(page, pfn);
1297 spin_lock_irqsave(&zone->lock, flags);
1298 if (unlikely(has_isolate_pageblock(zone) ||
1299 is_migrate_isolate(migratetype))) {
1300 migratetype = get_pfnblock_migratetype(page, pfn);
1302 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1303 spin_unlock_irqrestore(&zone->lock, flags);
1305 __count_vm_events(PGFREE, 1 << order);
1308 void __free_pages_core(struct page *page, unsigned int order)
1310 unsigned int nr_pages = 1 << order;
1311 struct page *p = page;
1315 * When initializing the memmap, __init_single_page() sets the refcount
1316 * of all pages to 1 ("allocated"/"not free"). We have to set the
1317 * refcount of all involved pages to 0.
1320 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1322 __ClearPageReserved(p);
1323 set_page_count(p, 0);
1325 __ClearPageReserved(p);
1326 set_page_count(p, 0);
1328 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1330 if (page_contains_unaccepted(page, order)) {
1331 if (order == MAX_ORDER && __free_unaccepted(page))
1334 accept_page(page, order);
1338 * Bypass PCP and place fresh pages right to the tail, primarily
1339 * relevant for memory onlining.
1341 __free_pages_ok(page, order, FPI_TO_TAIL);
1345 * Check that the whole (or subset of) a pageblock given by the interval of
1346 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1347 * with the migration of free compaction scanner.
1349 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1351 * It's possible on some configurations to have a setup like node0 node1 node0
1352 * i.e. it's possible that all pages within a zones range of pages do not
1353 * belong to a single zone. We assume that a border between node0 and node1
1354 * can occur within a single pageblock, but not a node0 node1 node0
1355 * interleaving within a single pageblock. It is therefore sufficient to check
1356 * the first and last page of a pageblock and avoid checking each individual
1357 * page in a pageblock.
1359 * Note: the function may return non-NULL struct page even for a page block
1360 * which contains a memory hole (i.e. there is no physical memory for a subset
1361 * of the pfn range). For example, if the pageblock order is MAX_ORDER, which
1362 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1363 * even though the start pfn is online and valid. This should be safe most of
1364 * the time because struct pages are still initialized via init_unavailable_range()
1365 * and pfn walkers shouldn't touch any physical memory range for which they do
1366 * not recognize any specific metadata in struct pages.
1368 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1369 unsigned long end_pfn, struct zone *zone)
1371 struct page *start_page;
1372 struct page *end_page;
1374 /* end_pfn is one past the range we are checking */
1377 if (!pfn_valid(end_pfn))
1380 start_page = pfn_to_online_page(start_pfn);
1384 if (page_zone(start_page) != zone)
1387 end_page = pfn_to_page(end_pfn);
1389 /* This gives a shorter code than deriving page_zone(end_page) */
1390 if (page_zone_id(start_page) != page_zone_id(end_page))
1397 * The order of subdivision here is critical for the IO subsystem.
1398 * Please do not alter this order without good reasons and regression
1399 * testing. Specifically, as large blocks of memory are subdivided,
1400 * the order in which smaller blocks are delivered depends on the order
1401 * they're subdivided in this function. This is the primary factor
1402 * influencing the order in which pages are delivered to the IO
1403 * subsystem according to empirical testing, and this is also justified
1404 * by considering the behavior of a buddy system containing a single
1405 * large block of memory acted on by a series of small allocations.
1406 * This behavior is a critical factor in sglist merging's success.
1410 static inline void expand(struct zone *zone, struct page *page,
1411 int low, int high, int migratetype)
1413 unsigned long size = 1 << high;
1415 while (high > low) {
1418 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1421 * Mark as guard pages (or page), that will allow to
1422 * merge back to allocator when buddy will be freed.
1423 * Corresponding page table entries will not be touched,
1424 * pages will stay not present in virtual address space
1426 if (set_page_guard(zone, &page[size], high, migratetype))
1429 add_to_free_list(&page[size], zone, high, migratetype);
1430 set_buddy_order(&page[size], high);
1434 static void check_new_page_bad(struct page *page)
1436 if (unlikely(page->flags & __PG_HWPOISON)) {
1437 /* Don't complain about hwpoisoned pages */
1438 page_mapcount_reset(page); /* remove PageBuddy */
1443 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1447 * This page is about to be returned from the page allocator
1449 static int check_new_page(struct page *page)
1451 if (likely(page_expected_state(page,
1452 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1455 check_new_page_bad(page);
1459 static inline bool check_new_pages(struct page *page, unsigned int order)
1461 if (is_check_pages_enabled()) {
1462 for (int i = 0; i < (1 << order); i++) {
1463 struct page *p = page + i;
1465 if (check_new_page(p))
1473 static inline bool should_skip_kasan_unpoison(gfp_t flags)
1475 /* Don't skip if a software KASAN mode is enabled. */
1476 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1477 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1480 /* Skip, if hardware tag-based KASAN is not enabled. */
1481 if (!kasan_hw_tags_enabled())
1485 * With hardware tag-based KASAN enabled, skip if this has been
1486 * requested via __GFP_SKIP_KASAN.
1488 return flags & __GFP_SKIP_KASAN;
1491 static inline bool should_skip_init(gfp_t flags)
1493 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1494 if (!kasan_hw_tags_enabled())
1497 /* For hardware tag-based KASAN, skip if requested. */
1498 return (flags & __GFP_SKIP_ZERO);
1501 inline void post_alloc_hook(struct page *page, unsigned int order,
1504 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1505 !should_skip_init(gfp_flags);
1506 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1509 set_page_private(page, 0);
1510 set_page_refcounted(page);
1512 arch_alloc_page(page, order);
1513 debug_pagealloc_map_pages(page, 1 << order);
1516 * Page unpoisoning must happen before memory initialization.
1517 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1518 * allocations and the page unpoisoning code will complain.
1520 kernel_unpoison_pages(page, 1 << order);
1523 * As memory initialization might be integrated into KASAN,
1524 * KASAN unpoisoning and memory initializion code must be
1525 * kept together to avoid discrepancies in behavior.
1529 * If memory tags should be zeroed
1530 * (which happens only when memory should be initialized as well).
1533 /* Initialize both memory and memory tags. */
1534 for (i = 0; i != 1 << order; ++i)
1535 tag_clear_highpage(page + i);
1537 /* Take note that memory was initialized by the loop above. */
1540 if (!should_skip_kasan_unpoison(gfp_flags) &&
1541 kasan_unpoison_pages(page, order, init)) {
1542 /* Take note that memory was initialized by KASAN. */
1543 if (kasan_has_integrated_init())
1547 * If memory tags have not been set by KASAN, reset the page
1548 * tags to ensure page_address() dereferencing does not fault.
1550 for (i = 0; i != 1 << order; ++i)
1551 page_kasan_tag_reset(page + i);
1553 /* If memory is still not initialized, initialize it now. */
1555 kernel_init_pages(page, 1 << order);
1557 set_page_owner(page, order, gfp_flags);
1558 page_table_check_alloc(page, order);
1561 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1562 unsigned int alloc_flags)
1564 post_alloc_hook(page, order, gfp_flags);
1566 if (order && (gfp_flags & __GFP_COMP))
1567 prep_compound_page(page, order);
1570 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1571 * allocate the page. The expectation is that the caller is taking
1572 * steps that will free more memory. The caller should avoid the page
1573 * being used for !PFMEMALLOC purposes.
1575 if (alloc_flags & ALLOC_NO_WATERMARKS)
1576 set_page_pfmemalloc(page);
1578 clear_page_pfmemalloc(page);
1582 * Go through the free lists for the given migratetype and remove
1583 * the smallest available page from the freelists
1585 static __always_inline
1586 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1589 unsigned int current_order;
1590 struct free_area *area;
1593 /* Find a page of the appropriate size in the preferred list */
1594 for (current_order = order; current_order <= MAX_ORDER; ++current_order) {
1595 area = &(zone->free_area[current_order]);
1596 page = get_page_from_free_area(area, migratetype);
1599 del_page_from_free_list(page, zone, current_order);
1600 expand(zone, page, order, current_order, migratetype);
1601 set_pcppage_migratetype(page, migratetype);
1602 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1603 pcp_allowed_order(order) &&
1604 migratetype < MIGRATE_PCPTYPES);
1613 * This array describes the order lists are fallen back to when
1614 * the free lists for the desirable migrate type are depleted
1616 * The other migratetypes do not have fallbacks.
1618 static int fallbacks[MIGRATE_TYPES][MIGRATE_PCPTYPES - 1] = {
1619 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1620 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1621 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1625 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1628 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1631 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1632 unsigned int order) { return NULL; }
1636 * Move the free pages in a range to the freelist tail of the requested type.
1637 * Note that start_page and end_pages are not aligned on a pageblock
1638 * boundary. If alignment is required, use move_freepages_block()
1640 static int move_freepages(struct zone *zone,
1641 unsigned long start_pfn, unsigned long end_pfn,
1642 int migratetype, int *num_movable)
1647 int pages_moved = 0;
1649 for (pfn = start_pfn; pfn <= end_pfn;) {
1650 page = pfn_to_page(pfn);
1651 if (!PageBuddy(page)) {
1653 * We assume that pages that could be isolated for
1654 * migration are movable. But we don't actually try
1655 * isolating, as that would be expensive.
1658 (PageLRU(page) || __PageMovable(page)))
1664 /* Make sure we are not inadvertently changing nodes */
1665 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1666 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1668 order = buddy_order(page);
1669 move_to_free_list(page, zone, order, migratetype);
1671 pages_moved += 1 << order;
1677 int move_freepages_block(struct zone *zone, struct page *page,
1678 int migratetype, int *num_movable)
1680 unsigned long start_pfn, end_pfn, pfn;
1685 pfn = page_to_pfn(page);
1686 start_pfn = pageblock_start_pfn(pfn);
1687 end_pfn = pageblock_end_pfn(pfn) - 1;
1689 /* Do not cross zone boundaries */
1690 if (!zone_spans_pfn(zone, start_pfn))
1692 if (!zone_spans_pfn(zone, end_pfn))
1695 return move_freepages(zone, start_pfn, end_pfn, migratetype,
1699 static void change_pageblock_range(struct page *pageblock_page,
1700 int start_order, int migratetype)
1702 int nr_pageblocks = 1 << (start_order - pageblock_order);
1704 while (nr_pageblocks--) {
1705 set_pageblock_migratetype(pageblock_page, migratetype);
1706 pageblock_page += pageblock_nr_pages;
1711 * When we are falling back to another migratetype during allocation, try to
1712 * steal extra free pages from the same pageblocks to satisfy further
1713 * allocations, instead of polluting multiple pageblocks.
1715 * If we are stealing a relatively large buddy page, it is likely there will
1716 * be more free pages in the pageblock, so try to steal them all. For
1717 * reclaimable and unmovable allocations, we steal regardless of page size,
1718 * as fragmentation caused by those allocations polluting movable pageblocks
1719 * is worse than movable allocations stealing from unmovable and reclaimable
1722 static bool can_steal_fallback(unsigned int order, int start_mt)
1725 * Leaving this order check is intended, although there is
1726 * relaxed order check in next check. The reason is that
1727 * we can actually steal whole pageblock if this condition met,
1728 * but, below check doesn't guarantee it and that is just heuristic
1729 * so could be changed anytime.
1731 if (order >= pageblock_order)
1734 if (order >= pageblock_order / 2 ||
1735 start_mt == MIGRATE_RECLAIMABLE ||
1736 start_mt == MIGRATE_UNMOVABLE ||
1737 page_group_by_mobility_disabled)
1743 static inline bool boost_watermark(struct zone *zone)
1745 unsigned long max_boost;
1747 if (!watermark_boost_factor)
1750 * Don't bother in zones that are unlikely to produce results.
1751 * On small machines, including kdump capture kernels running
1752 * in a small area, boosting the watermark can cause an out of
1753 * memory situation immediately.
1755 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1758 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1759 watermark_boost_factor, 10000);
1762 * high watermark may be uninitialised if fragmentation occurs
1763 * very early in boot so do not boost. We do not fall
1764 * through and boost by pageblock_nr_pages as failing
1765 * allocations that early means that reclaim is not going
1766 * to help and it may even be impossible to reclaim the
1767 * boosted watermark resulting in a hang.
1772 max_boost = max(pageblock_nr_pages, max_boost);
1774 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1781 * This function implements actual steal behaviour. If order is large enough,
1782 * we can steal whole pageblock. If not, we first move freepages in this
1783 * pageblock to our migratetype and determine how many already-allocated pages
1784 * are there in the pageblock with a compatible migratetype. If at least half
1785 * of pages are free or compatible, we can change migratetype of the pageblock
1786 * itself, so pages freed in the future will be put on the correct free list.
1788 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1789 unsigned int alloc_flags, int start_type, bool whole_block)
1791 unsigned int current_order = buddy_order(page);
1792 int free_pages, movable_pages, alike_pages;
1795 old_block_type = get_pageblock_migratetype(page);
1798 * This can happen due to races and we want to prevent broken
1799 * highatomic accounting.
1801 if (is_migrate_highatomic(old_block_type))
1804 /* Take ownership for orders >= pageblock_order */
1805 if (current_order >= pageblock_order) {
1806 change_pageblock_range(page, current_order, start_type);
1811 * Boost watermarks to increase reclaim pressure to reduce the
1812 * likelihood of future fallbacks. Wake kswapd now as the node
1813 * may be balanced overall and kswapd will not wake naturally.
1815 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1816 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1818 /* We are not allowed to try stealing from the whole block */
1822 free_pages = move_freepages_block(zone, page, start_type,
1824 /* moving whole block can fail due to zone boundary conditions */
1829 * Determine how many pages are compatible with our allocation.
1830 * For movable allocation, it's the number of movable pages which
1831 * we just obtained. For other types it's a bit more tricky.
1833 if (start_type == MIGRATE_MOVABLE) {
1834 alike_pages = movable_pages;
1837 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1838 * to MOVABLE pageblock, consider all non-movable pages as
1839 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1840 * vice versa, be conservative since we can't distinguish the
1841 * exact migratetype of non-movable pages.
1843 if (old_block_type == MIGRATE_MOVABLE)
1844 alike_pages = pageblock_nr_pages
1845 - (free_pages + movable_pages);
1850 * If a sufficient number of pages in the block are either free or of
1851 * compatible migratability as our allocation, claim the whole block.
1853 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1854 page_group_by_mobility_disabled)
1855 set_pageblock_migratetype(page, start_type);
1860 move_to_free_list(page, zone, current_order, start_type);
1864 * Check whether there is a suitable fallback freepage with requested order.
1865 * If only_stealable is true, this function returns fallback_mt only if
1866 * we can steal other freepages all together. This would help to reduce
1867 * fragmentation due to mixed migratetype pages in one pageblock.
1869 int find_suitable_fallback(struct free_area *area, unsigned int order,
1870 int migratetype, bool only_stealable, bool *can_steal)
1875 if (area->nr_free == 0)
1879 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
1880 fallback_mt = fallbacks[migratetype][i];
1881 if (free_area_empty(area, fallback_mt))
1884 if (can_steal_fallback(order, migratetype))
1887 if (!only_stealable)
1898 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1899 * there are no empty page blocks that contain a page with a suitable order
1901 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone)
1904 unsigned long max_managed, flags;
1907 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1908 * Check is race-prone but harmless.
1910 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
1911 if (zone->nr_reserved_highatomic >= max_managed)
1914 spin_lock_irqsave(&zone->lock, flags);
1916 /* Recheck the nr_reserved_highatomic limit under the lock */
1917 if (zone->nr_reserved_highatomic >= max_managed)
1921 mt = get_pageblock_migratetype(page);
1922 /* Only reserve normal pageblocks (i.e., they can merge with others) */
1923 if (migratetype_is_mergeable(mt)) {
1924 zone->nr_reserved_highatomic += pageblock_nr_pages;
1925 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1926 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
1930 spin_unlock_irqrestore(&zone->lock, flags);
1934 * Used when an allocation is about to fail under memory pressure. This
1935 * potentially hurts the reliability of high-order allocations when under
1936 * intense memory pressure but failed atomic allocations should be easier
1937 * to recover from than an OOM.
1939 * If @force is true, try to unreserve a pageblock even though highatomic
1940 * pageblock is exhausted.
1942 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
1945 struct zonelist *zonelist = ac->zonelist;
1946 unsigned long flags;
1953 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
1956 * Preserve at least one pageblock unless memory pressure
1959 if (!force && zone->nr_reserved_highatomic <=
1963 spin_lock_irqsave(&zone->lock, flags);
1964 for (order = 0; order <= MAX_ORDER; order++) {
1965 struct free_area *area = &(zone->free_area[order]);
1967 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
1972 * In page freeing path, migratetype change is racy so
1973 * we can counter several free pages in a pageblock
1974 * in this loop although we changed the pageblock type
1975 * from highatomic to ac->migratetype. So we should
1976 * adjust the count once.
1978 if (is_migrate_highatomic_page(page)) {
1980 * It should never happen but changes to
1981 * locking could inadvertently allow a per-cpu
1982 * drain to add pages to MIGRATE_HIGHATOMIC
1983 * while unreserving so be safe and watch for
1986 zone->nr_reserved_highatomic -= min(
1988 zone->nr_reserved_highatomic);
1992 * Convert to ac->migratetype and avoid the normal
1993 * pageblock stealing heuristics. Minimally, the caller
1994 * is doing the work and needs the pages. More
1995 * importantly, if the block was always converted to
1996 * MIGRATE_UNMOVABLE or another type then the number
1997 * of pageblocks that cannot be completely freed
2000 set_pageblock_migratetype(page, ac->migratetype);
2001 ret = move_freepages_block(zone, page, ac->migratetype,
2004 spin_unlock_irqrestore(&zone->lock, flags);
2008 spin_unlock_irqrestore(&zone->lock, flags);
2015 * Try finding a free buddy page on the fallback list and put it on the free
2016 * list of requested migratetype, possibly along with other pages from the same
2017 * block, depending on fragmentation avoidance heuristics. Returns true if
2018 * fallback was found so that __rmqueue_smallest() can grab it.
2020 * The use of signed ints for order and current_order is a deliberate
2021 * deviation from the rest of this file, to make the for loop
2022 * condition simpler.
2024 static __always_inline bool
2025 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2026 unsigned int alloc_flags)
2028 struct free_area *area;
2030 int min_order = order;
2036 * Do not steal pages from freelists belonging to other pageblocks
2037 * i.e. orders < pageblock_order. If there are no local zones free,
2038 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2040 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2041 min_order = pageblock_order;
2044 * Find the largest available free page in the other list. This roughly
2045 * approximates finding the pageblock with the most free pages, which
2046 * would be too costly to do exactly.
2048 for (current_order = MAX_ORDER; current_order >= min_order;
2050 area = &(zone->free_area[current_order]);
2051 fallback_mt = find_suitable_fallback(area, current_order,
2052 start_migratetype, false, &can_steal);
2053 if (fallback_mt == -1)
2057 * We cannot steal all free pages from the pageblock and the
2058 * requested migratetype is movable. In that case it's better to
2059 * steal and split the smallest available page instead of the
2060 * largest available page, because even if the next movable
2061 * allocation falls back into a different pageblock than this
2062 * one, it won't cause permanent fragmentation.
2064 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2065 && current_order > order)
2074 for (current_order = order; current_order <= MAX_ORDER;
2076 area = &(zone->free_area[current_order]);
2077 fallback_mt = find_suitable_fallback(area, current_order,
2078 start_migratetype, false, &can_steal);
2079 if (fallback_mt != -1)
2084 * This should not happen - we already found a suitable fallback
2085 * when looking for the largest page.
2087 VM_BUG_ON(current_order > MAX_ORDER);
2090 page = get_page_from_free_area(area, fallback_mt);
2092 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2095 trace_mm_page_alloc_extfrag(page, order, current_order,
2096 start_migratetype, fallback_mt);
2103 * Do the hard work of removing an element from the buddy allocator.
2104 * Call me with the zone->lock already held.
2106 static __always_inline struct page *
2107 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2108 unsigned int alloc_flags)
2112 if (IS_ENABLED(CONFIG_CMA)) {
2114 * Balance movable allocations between regular and CMA areas by
2115 * allocating from CMA when over more than a given proportion of
2116 * the zone's free memory is in the CMA area.
2118 if (alloc_flags & ALLOC_CMA &&
2119 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2120 zone_page_state(zone, NR_FREE_PAGES) / ALLOC_IN_CMA_THRESHOLD_MAX
2121 * _alloc_in_cma_threshold) {
2122 page = __rmqueue_cma_fallback(zone, order);
2128 page = __rmqueue_smallest(zone, order, migratetype);
2129 if (unlikely(!page)) {
2130 if (alloc_flags & ALLOC_CMA)
2131 page = __rmqueue_cma_fallback(zone, order);
2133 if (!page && __rmqueue_fallback(zone, order, migratetype,
2141 * Obtain a specified number of elements from the buddy allocator, all under
2142 * a single hold of the lock, for efficiency. Add them to the supplied list.
2143 * Returns the number of new pages which were placed at *list.
2145 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2146 unsigned long count, struct list_head *list,
2147 int migratetype, unsigned int alloc_flags)
2149 unsigned long flags;
2152 spin_lock_irqsave(&zone->lock, flags);
2153 for (i = 0; i < count; ++i) {
2154 struct page *page = __rmqueue(zone, order, migratetype,
2156 if (unlikely(page == NULL))
2160 * Split buddy pages returned by expand() are received here in
2161 * physical page order. The page is added to the tail of
2162 * caller's list. From the callers perspective, the linked list
2163 * is ordered by page number under some conditions. This is
2164 * useful for IO devices that can forward direction from the
2165 * head, thus also in the physical page order. This is useful
2166 * for IO devices that can merge IO requests if the physical
2167 * pages are ordered properly.
2169 list_add_tail(&page->pcp_list, list);
2170 if (is_migrate_cma(get_pcppage_migratetype(page)))
2171 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2175 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2176 spin_unlock_irqrestore(&zone->lock, flags);
2183 * Called from the vmstat counter updater to drain pagesets of this
2184 * currently executing processor on remote nodes after they have
2187 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2189 int to_drain, batch;
2191 batch = READ_ONCE(pcp->batch);
2192 to_drain = min(pcp->count, batch);
2194 spin_lock(&pcp->lock);
2195 free_pcppages_bulk(zone, to_drain, pcp, 0);
2196 spin_unlock(&pcp->lock);
2202 * Drain pcplists of the indicated processor and zone.
2204 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2206 struct per_cpu_pages *pcp;
2208 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2210 spin_lock(&pcp->lock);
2211 free_pcppages_bulk(zone, pcp->count, pcp, 0);
2212 spin_unlock(&pcp->lock);
2217 * Drain pcplists of all zones on the indicated processor.
2219 static void drain_pages(unsigned int cpu)
2223 for_each_populated_zone(zone) {
2224 drain_pages_zone(cpu, zone);
2229 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2231 void drain_local_pages(struct zone *zone)
2233 int cpu = smp_processor_id();
2236 drain_pages_zone(cpu, zone);
2242 * The implementation of drain_all_pages(), exposing an extra parameter to
2243 * drain on all cpus.
2245 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2246 * not empty. The check for non-emptiness can however race with a free to
2247 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2248 * that need the guarantee that every CPU has drained can disable the
2249 * optimizing racy check.
2251 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2256 * Allocate in the BSS so we won't require allocation in
2257 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2259 static cpumask_t cpus_with_pcps;
2262 * Do not drain if one is already in progress unless it's specific to
2263 * a zone. Such callers are primarily CMA and memory hotplug and need
2264 * the drain to be complete when the call returns.
2266 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2269 mutex_lock(&pcpu_drain_mutex);
2273 * We don't care about racing with CPU hotplug event
2274 * as offline notification will cause the notified
2275 * cpu to drain that CPU pcps and on_each_cpu_mask
2276 * disables preemption as part of its processing
2278 for_each_online_cpu(cpu) {
2279 struct per_cpu_pages *pcp;
2281 bool has_pcps = false;
2283 if (force_all_cpus) {
2285 * The pcp.count check is racy, some callers need a
2286 * guarantee that no cpu is missed.
2290 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2294 for_each_populated_zone(z) {
2295 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2304 cpumask_set_cpu(cpu, &cpus_with_pcps);
2306 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2309 for_each_cpu(cpu, &cpus_with_pcps) {
2311 drain_pages_zone(cpu, zone);
2316 mutex_unlock(&pcpu_drain_mutex);
2320 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2322 * When zone parameter is non-NULL, spill just the single zone's pages.
2324 void drain_all_pages(struct zone *zone)
2326 __drain_all_pages(zone, false);
2329 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
2334 if (!free_pages_prepare(page, order, FPI_NONE))
2337 migratetype = get_pfnblock_migratetype(page, pfn);
2338 set_pcppage_migratetype(page, migratetype);
2342 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, bool free_high)
2344 int min_nr_free, max_nr_free;
2345 int batch = READ_ONCE(pcp->batch);
2347 /* Free everything if batch freeing high-order pages. */
2348 if (unlikely(free_high))
2351 /* Check for PCP disabled or boot pageset */
2352 if (unlikely(high < batch))
2355 /* Leave at least pcp->batch pages on the list */
2356 min_nr_free = batch;
2357 max_nr_free = high - batch;
2360 * Double the number of pages freed each time there is subsequent
2361 * freeing of pages without any allocation.
2363 batch <<= pcp->free_factor;
2364 if (batch < max_nr_free)
2366 batch = clamp(batch, min_nr_free, max_nr_free);
2371 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2374 int high = READ_ONCE(pcp->high);
2376 if (unlikely(!high || free_high))
2379 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
2383 * If reclaim is active, limit the number of pages that can be
2384 * stored on pcp lists
2386 return min(READ_ONCE(pcp->batch) << 2, high);
2389 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2390 struct page *page, int migratetype,
2397 __count_vm_events(PGFREE, 1 << order);
2398 pindex = order_to_pindex(migratetype, order);
2399 list_add(&page->pcp_list, &pcp->lists[pindex]);
2400 pcp->count += 1 << order;
2403 * As high-order pages other than THP's stored on PCP can contribute
2404 * to fragmentation, limit the number stored when PCP is heavily
2405 * freeing without allocation. The remainder after bulk freeing
2406 * stops will be drained from vmstat refresh context.
2408 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
2410 high = nr_pcp_high(pcp, zone, free_high);
2411 if (pcp->count >= high) {
2412 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, free_high), pcp, pindex);
2419 void free_unref_page(struct page *page, unsigned int order)
2421 unsigned long __maybe_unused UP_flags;
2422 struct per_cpu_pages *pcp;
2424 unsigned long pfn = page_to_pfn(page);
2425 int migratetype, pcpmigratetype;
2427 if (!free_unref_page_prepare(page, pfn, order))
2431 * We only track unmovable, reclaimable and movable on pcp lists.
2432 * Place ISOLATE pages on the isolated list because they are being
2433 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2434 * get those areas back if necessary. Otherwise, we may have to free
2435 * excessively into the page allocator
2437 migratetype = pcpmigratetype = get_pcppage_migratetype(page);
2438 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2439 if (unlikely(is_migrate_isolate(migratetype))) {
2440 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
2443 pcpmigratetype = MIGRATE_MOVABLE;
2446 zone = page_zone(page);
2447 pcp_trylock_prepare(UP_flags);
2448 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2450 free_unref_page_commit(zone, pcp, page, pcpmigratetype, order);
2451 pcp_spin_unlock(pcp);
2453 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
2455 pcp_trylock_finish(UP_flags);
2459 * Free a list of 0-order pages
2461 void free_unref_page_list(struct list_head *list)
2463 unsigned long __maybe_unused UP_flags;
2464 struct page *page, *next;
2465 struct per_cpu_pages *pcp = NULL;
2466 struct zone *locked_zone = NULL;
2467 int batch_count = 0;
2470 /* Prepare pages for freeing */
2471 list_for_each_entry_safe(page, next, list, lru) {
2472 unsigned long pfn = page_to_pfn(page);
2473 if (!free_unref_page_prepare(page, pfn, 0)) {
2474 list_del(&page->lru);
2479 * Free isolated pages directly to the allocator, see
2480 * comment in free_unref_page.
2482 migratetype = get_pcppage_migratetype(page);
2483 if (unlikely(is_migrate_isolate(migratetype))) {
2484 list_del(&page->lru);
2485 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
2490 list_for_each_entry_safe(page, next, list, lru) {
2491 struct zone *zone = page_zone(page);
2493 list_del(&page->lru);
2494 migratetype = get_pcppage_migratetype(page);
2497 * Either different zone requiring a different pcp lock or
2498 * excessive lock hold times when freeing a large list of
2501 if (zone != locked_zone || batch_count == SWAP_CLUSTER_MAX) {
2503 pcp_spin_unlock(pcp);
2504 pcp_trylock_finish(UP_flags);
2510 * trylock is necessary as pages may be getting freed
2511 * from IRQ or SoftIRQ context after an IO completion.
2513 pcp_trylock_prepare(UP_flags);
2514 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2515 if (unlikely(!pcp)) {
2516 pcp_trylock_finish(UP_flags);
2517 free_one_page(zone, page, page_to_pfn(page),
2518 0, migratetype, FPI_NONE);
2526 * Non-isolated types over MIGRATE_PCPTYPES get added
2527 * to the MIGRATE_MOVABLE pcp list.
2529 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2530 migratetype = MIGRATE_MOVABLE;
2532 trace_mm_page_free_batched(page);
2533 free_unref_page_commit(zone, pcp, page, migratetype, 0);
2538 pcp_spin_unlock(pcp);
2539 pcp_trylock_finish(UP_flags);
2544 * split_page takes a non-compound higher-order page, and splits it into
2545 * n (1<<order) sub-pages: page[0..n]
2546 * Each sub-page must be freed individually.
2548 * Note: this is probably too low level an operation for use in drivers.
2549 * Please consult with lkml before using this in your driver.
2551 void split_page(struct page *page, unsigned int order)
2555 VM_BUG_ON_PAGE(PageCompound(page), page);
2556 VM_BUG_ON_PAGE(!page_count(page), page);
2558 for (i = 1; i < (1 << order); i++)
2559 set_page_refcounted(page + i);
2560 split_page_owner(page, 1 << order);
2561 split_page_memcg(page, 1 << order);
2563 EXPORT_SYMBOL_GPL(split_page);
2565 int __isolate_free_page(struct page *page, unsigned int order)
2567 struct zone *zone = page_zone(page);
2568 int mt = get_pageblock_migratetype(page);
2570 if (!is_migrate_isolate(mt)) {
2571 unsigned long watermark;
2573 * Obey watermarks as if the page was being allocated. We can
2574 * emulate a high-order watermark check with a raised order-0
2575 * watermark, because we already know our high-order page
2578 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2579 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2582 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2585 del_page_from_free_list(page, zone, order);
2588 * Set the pageblock if the isolated page is at least half of a
2591 if (order >= pageblock_order - 1) {
2592 struct page *endpage = page + (1 << order) - 1;
2593 for (; page < endpage; page += pageblock_nr_pages) {
2594 int mt = get_pageblock_migratetype(page);
2596 * Only change normal pageblocks (i.e., they can merge
2599 if (migratetype_is_mergeable(mt))
2600 set_pageblock_migratetype(page,
2605 return 1UL << order;
2609 * __putback_isolated_page - Return a now-isolated page back where we got it
2610 * @page: Page that was isolated
2611 * @order: Order of the isolated page
2612 * @mt: The page's pageblock's migratetype
2614 * This function is meant to return a page pulled from the free lists via
2615 * __isolate_free_page back to the free lists they were pulled from.
2617 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2619 struct zone *zone = page_zone(page);
2621 /* zone lock should be held when this function is called */
2622 lockdep_assert_held(&zone->lock);
2624 /* Return isolated page to tail of freelist. */
2625 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2626 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2630 * Update NUMA hit/miss statistics
2632 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2636 enum numa_stat_item local_stat = NUMA_LOCAL;
2638 /* skip numa counters update if numa stats is disabled */
2639 if (!static_branch_likely(&vm_numa_stat_key))
2642 if (zone_to_nid(z) != numa_node_id())
2643 local_stat = NUMA_OTHER;
2645 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2646 __count_numa_events(z, NUMA_HIT, nr_account);
2648 __count_numa_events(z, NUMA_MISS, nr_account);
2649 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2651 __count_numa_events(z, local_stat, nr_account);
2655 static __always_inline
2656 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2657 unsigned int order, unsigned int alloc_flags,
2661 unsigned long flags;
2665 spin_lock_irqsave(&zone->lock, flags);
2666 if (alloc_flags & ALLOC_HIGHATOMIC)
2667 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2669 page = __rmqueue(zone, order, migratetype, alloc_flags);
2672 * If the allocation fails, allow OOM handling access
2673 * to HIGHATOMIC reserves as failing now is worse than
2674 * failing a high-order atomic allocation in the
2677 if (!page && (alloc_flags & ALLOC_OOM))
2678 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2681 spin_unlock_irqrestore(&zone->lock, flags);
2685 __mod_zone_freepage_state(zone, -(1 << order),
2686 get_pcppage_migratetype(page));
2687 spin_unlock_irqrestore(&zone->lock, flags);
2688 } while (check_new_pages(page, order));
2690 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2691 zone_statistics(preferred_zone, zone, 1);
2696 /* Remove page from the per-cpu list, caller must protect the list */
2698 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2700 unsigned int alloc_flags,
2701 struct per_cpu_pages *pcp,
2702 struct list_head *list)
2707 if (list_empty(list)) {
2708 int batch = READ_ONCE(pcp->batch);
2712 * Scale batch relative to order if batch implies
2713 * free pages can be stored on the PCP. Batch can
2714 * be 1 for small zones or for boot pagesets which
2715 * should never store free pages as the pages may
2716 * belong to arbitrary zones.
2719 batch = max(batch >> order, 2);
2720 alloced = rmqueue_bulk(zone, order,
2722 migratetype, alloc_flags);
2724 pcp->count += alloced << order;
2725 if (unlikely(list_empty(list)))
2729 page = list_first_entry(list, struct page, pcp_list);
2730 list_del(&page->pcp_list);
2731 pcp->count -= 1 << order;
2732 } while (check_new_pages(page, order));
2737 /* Lock and remove page from the per-cpu list */
2738 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2739 struct zone *zone, unsigned int order,
2740 int migratetype, unsigned int alloc_flags)
2742 struct per_cpu_pages *pcp;
2743 struct list_head *list;
2745 unsigned long __maybe_unused UP_flags;
2747 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2748 pcp_trylock_prepare(UP_flags);
2749 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2751 pcp_trylock_finish(UP_flags);
2756 * On allocation, reduce the number of pages that are batch freed.
2757 * See nr_pcp_free() where free_factor is increased for subsequent
2760 pcp->free_factor >>= 1;
2761 list = &pcp->lists[order_to_pindex(migratetype, order)];
2762 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
2763 pcp_spin_unlock(pcp);
2764 pcp_trylock_finish(UP_flags);
2766 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2767 zone_statistics(preferred_zone, zone, 1);
2773 * Allocate a page from the given zone.
2774 * Use pcplists for THP or "cheap" high-order allocations.
2778 * Do not instrument rmqueue() with KMSAN. This function may call
2779 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
2780 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
2781 * may call rmqueue() again, which will result in a deadlock.
2783 __no_sanitize_memory
2785 struct page *rmqueue(struct zone *preferred_zone,
2786 struct zone *zone, unsigned int order,
2787 gfp_t gfp_flags, unsigned int alloc_flags,
2793 * We most definitely don't want callers attempting to
2794 * allocate greater than order-1 page units with __GFP_NOFAIL.
2796 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2798 if (likely(pcp_allowed_order(order))) {
2799 page = rmqueue_pcplist(preferred_zone, zone, order,
2800 migratetype, alloc_flags);
2805 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
2809 /* Separate test+clear to avoid unnecessary atomics */
2810 if ((alloc_flags & ALLOC_KSWAPD) &&
2811 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
2812 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2813 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
2816 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2820 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2822 return __should_fail_alloc_page(gfp_mask, order);
2824 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
2826 static inline long __zone_watermark_unusable_free(struct zone *z,
2827 unsigned int order, unsigned int alloc_flags)
2829 long unusable_free = (1 << order) - 1;
2832 * If the caller does not have rights to reserves below the min
2833 * watermark then subtract the high-atomic reserves. This will
2834 * over-estimate the size of the atomic reserve but it avoids a search.
2836 if (likely(!(alloc_flags & ALLOC_RESERVES)))
2837 unusable_free += z->nr_reserved_highatomic;
2840 /* If allocation can't use CMA areas don't use free CMA pages */
2841 if (!(alloc_flags & ALLOC_CMA))
2842 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
2844 #ifdef CONFIG_UNACCEPTED_MEMORY
2845 unusable_free += zone_page_state(z, NR_UNACCEPTED);
2848 return unusable_free;
2852 * Return true if free base pages are above 'mark'. For high-order checks it
2853 * will return true of the order-0 watermark is reached and there is at least
2854 * one free page of a suitable size. Checking now avoids taking the zone lock
2855 * to check in the allocation paths if no pages are free.
2857 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2858 int highest_zoneidx, unsigned int alloc_flags,
2864 /* free_pages may go negative - that's OK */
2865 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
2867 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
2869 * __GFP_HIGH allows access to 50% of the min reserve as well
2872 if (alloc_flags & ALLOC_MIN_RESERVE) {
2876 * Non-blocking allocations (e.g. GFP_ATOMIC) can
2877 * access more reserves than just __GFP_HIGH. Other
2878 * non-blocking allocations requests such as GFP_NOWAIT
2879 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
2880 * access to the min reserve.
2882 if (alloc_flags & ALLOC_NON_BLOCK)
2887 * OOM victims can try even harder than the normal reserve
2888 * users on the grounds that it's definitely going to be in
2889 * the exit path shortly and free memory. Any allocation it
2890 * makes during the free path will be small and short-lived.
2892 if (alloc_flags & ALLOC_OOM)
2897 * Check watermarks for an order-0 allocation request. If these
2898 * are not met, then a high-order request also cannot go ahead
2899 * even if a suitable page happened to be free.
2901 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
2904 /* If this is an order-0 request then the watermark is fine */
2908 /* For a high-order request, check at least one suitable page is free */
2909 for (o = order; o <= MAX_ORDER; o++) {
2910 struct free_area *area = &z->free_area[o];
2916 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2917 if (!free_area_empty(area, mt))
2922 if ((alloc_flags & ALLOC_CMA) &&
2923 !free_area_empty(area, MIGRATE_CMA)) {
2927 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
2928 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
2935 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2936 int highest_zoneidx, unsigned int alloc_flags)
2938 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
2939 zone_page_state(z, NR_FREE_PAGES));
2942 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2943 unsigned long mark, int highest_zoneidx,
2944 unsigned int alloc_flags, gfp_t gfp_mask)
2948 free_pages = zone_page_state(z, NR_FREE_PAGES);
2951 * Fast check for order-0 only. If this fails then the reserves
2952 * need to be calculated.
2958 usable_free = free_pages;
2959 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
2961 /* reserved may over estimate high-atomic reserves. */
2962 usable_free -= min(usable_free, reserved);
2963 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
2967 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
2972 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
2973 * when checking the min watermark. The min watermark is the
2974 * point where boosting is ignored so that kswapd is woken up
2975 * when below the low watermark.
2977 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
2978 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
2979 mark = z->_watermark[WMARK_MIN];
2980 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
2981 alloc_flags, free_pages);
2987 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2988 unsigned long mark, int highest_zoneidx)
2990 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2992 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2993 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2995 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3000 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3002 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3004 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3005 node_reclaim_distance;
3007 #else /* CONFIG_NUMA */
3008 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3012 #endif /* CONFIG_NUMA */
3015 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3016 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3017 * premature use of a lower zone may cause lowmem pressure problems that
3018 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3019 * probably too small. It only makes sense to spread allocations to avoid
3020 * fragmentation between the Normal and DMA32 zones.
3022 static inline unsigned int
3023 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3025 unsigned int alloc_flags;
3028 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3031 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3033 #ifdef CONFIG_ZONE_DMA32
3037 if (zone_idx(zone) != ZONE_NORMAL)
3041 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3042 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3043 * on UMA that if Normal is populated then so is DMA32.
3045 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3046 if (nr_online_nodes > 1 && !populated_zone(--zone))
3049 alloc_flags |= ALLOC_NOFRAGMENT;
3050 #endif /* CONFIG_ZONE_DMA32 */
3054 /* Must be called after current_gfp_context() which can change gfp_mask */
3055 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3056 unsigned int alloc_flags)
3059 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3060 alloc_flags |= ALLOC_CMA;
3066 * get_page_from_freelist goes through the zonelist trying to allocate
3069 static struct page *
3070 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3071 const struct alloc_context *ac)
3075 struct pglist_data *last_pgdat = NULL;
3076 bool last_pgdat_dirty_ok = false;
3081 * Scan zonelist, looking for a zone with enough free.
3082 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3084 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3085 z = ac->preferred_zoneref;
3086 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3091 if (cpusets_enabled() &&
3092 (alloc_flags & ALLOC_CPUSET) &&
3093 !__cpuset_zone_allowed(zone, gfp_mask))
3096 * When allocating a page cache page for writing, we
3097 * want to get it from a node that is within its dirty
3098 * limit, such that no single node holds more than its
3099 * proportional share of globally allowed dirty pages.
3100 * The dirty limits take into account the node's
3101 * lowmem reserves and high watermark so that kswapd
3102 * should be able to balance it without having to
3103 * write pages from its LRU list.
3105 * XXX: For now, allow allocations to potentially
3106 * exceed the per-node dirty limit in the slowpath
3107 * (spread_dirty_pages unset) before going into reclaim,
3108 * which is important when on a NUMA setup the allowed
3109 * nodes are together not big enough to reach the
3110 * global limit. The proper fix for these situations
3111 * will require awareness of nodes in the
3112 * dirty-throttling and the flusher threads.
3114 if (ac->spread_dirty_pages) {
3115 if (last_pgdat != zone->zone_pgdat) {
3116 last_pgdat = zone->zone_pgdat;
3117 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3120 if (!last_pgdat_dirty_ok)
3124 if (no_fallback && nr_online_nodes > 1 &&
3125 zone != ac->preferred_zoneref->zone) {
3129 * If moving to a remote node, retry but allow
3130 * fragmenting fallbacks. Locality is more important
3131 * than fragmentation avoidance.
3133 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3134 if (zone_to_nid(zone) != local_nid) {
3135 alloc_flags &= ~ALLOC_NOFRAGMENT;
3140 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3141 if (!zone_watermark_fast(zone, order, mark,
3142 ac->highest_zoneidx, alloc_flags,
3146 if (has_unaccepted_memory()) {
3147 if (try_to_accept_memory(zone, order))
3151 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3153 * Watermark failed for this zone, but see if we can
3154 * grow this zone if it contains deferred pages.
3156 if (deferred_pages_enabled()) {
3157 if (_deferred_grow_zone(zone, order))
3161 /* Checked here to keep the fast path fast */
3162 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3163 if (alloc_flags & ALLOC_NO_WATERMARKS)
3166 if (!node_reclaim_enabled() ||
3167 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3170 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3172 case NODE_RECLAIM_NOSCAN:
3175 case NODE_RECLAIM_FULL:
3176 /* scanned but unreclaimable */
3179 /* did we reclaim enough */
3180 if (zone_watermark_ok(zone, order, mark,
3181 ac->highest_zoneidx, alloc_flags))
3189 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3190 gfp_mask, alloc_flags, ac->migratetype);
3192 prep_new_page(page, order, gfp_mask, alloc_flags);
3195 * If this is a high-order atomic allocation then check
3196 * if the pageblock should be reserved for the future
3198 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3199 reserve_highatomic_pageblock(page, zone);
3203 if (has_unaccepted_memory()) {
3204 if (try_to_accept_memory(zone, order))
3208 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3209 /* Try again if zone has deferred pages */
3210 if (deferred_pages_enabled()) {
3211 if (_deferred_grow_zone(zone, order))
3219 * It's possible on a UMA machine to get through all zones that are
3220 * fragmented. If avoiding fragmentation, reset and try again.
3223 alloc_flags &= ~ALLOC_NOFRAGMENT;
3230 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3232 unsigned int filter = SHOW_MEM_FILTER_NODES;
3235 * This documents exceptions given to allocations in certain
3236 * contexts that are allowed to allocate outside current's set
3239 if (!(gfp_mask & __GFP_NOMEMALLOC))
3240 if (tsk_is_oom_victim(current) ||
3241 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3242 filter &= ~SHOW_MEM_FILTER_NODES;
3243 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3244 filter &= ~SHOW_MEM_FILTER_NODES;
3246 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3249 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3251 struct va_format vaf;
3253 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3255 if ((gfp_mask & __GFP_NOWARN) ||
3256 !__ratelimit(&nopage_rs) ||
3257 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3260 va_start(args, fmt);
3263 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3264 current->comm, &vaf, gfp_mask, &gfp_mask,
3265 nodemask_pr_args(nodemask));
3268 cpuset_print_current_mems_allowed();
3271 warn_alloc_show_mem(gfp_mask, nodemask);
3274 static inline struct page *
3275 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3276 unsigned int alloc_flags,
3277 const struct alloc_context *ac)
3281 page = get_page_from_freelist(gfp_mask, order,
3282 alloc_flags|ALLOC_CPUSET, ac);
3284 * fallback to ignore cpuset restriction if our nodes
3288 page = get_page_from_freelist(gfp_mask, order,
3294 static inline struct page *
3295 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3296 const struct alloc_context *ac, unsigned long *did_some_progress)
3298 struct oom_control oc = {
3299 .zonelist = ac->zonelist,
3300 .nodemask = ac->nodemask,
3302 .gfp_mask = gfp_mask,
3307 *did_some_progress = 0;
3310 * Acquire the oom lock. If that fails, somebody else is
3311 * making progress for us.
3313 if (!mutex_trylock(&oom_lock)) {
3314 *did_some_progress = 1;
3315 schedule_timeout_uninterruptible(1);
3320 * Go through the zonelist yet one more time, keep very high watermark
3321 * here, this is only to catch a parallel oom killing, we must fail if
3322 * we're still under heavy pressure. But make sure that this reclaim
3323 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3324 * allocation which will never fail due to oom_lock already held.
3326 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3327 ~__GFP_DIRECT_RECLAIM, order,
3328 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3332 /* Coredumps can quickly deplete all memory reserves */
3333 if (current->flags & PF_DUMPCORE)
3335 /* The OOM killer will not help higher order allocs */
3336 if (order > PAGE_ALLOC_COSTLY_ORDER)
3339 * We have already exhausted all our reclaim opportunities without any
3340 * success so it is time to admit defeat. We will skip the OOM killer
3341 * because it is very likely that the caller has a more reasonable
3342 * fallback than shooting a random task.
3344 * The OOM killer may not free memory on a specific node.
3346 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3348 /* The OOM killer does not needlessly kill tasks for lowmem */
3349 if (ac->highest_zoneidx < ZONE_NORMAL)
3351 if (pm_suspended_storage())
3354 * XXX: GFP_NOFS allocations should rather fail than rely on
3355 * other request to make a forward progress.
3356 * We are in an unfortunate situation where out_of_memory cannot
3357 * do much for this context but let's try it to at least get
3358 * access to memory reserved if the current task is killed (see
3359 * out_of_memory). Once filesystems are ready to handle allocation
3360 * failures more gracefully we should just bail out here.
3363 /* Exhausted what can be done so it's blame time */
3364 if (out_of_memory(&oc) ||
3365 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3366 *did_some_progress = 1;
3369 * Help non-failing allocations by giving them access to memory
3372 if (gfp_mask & __GFP_NOFAIL)
3373 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3374 ALLOC_NO_WATERMARKS, ac);
3377 mutex_unlock(&oom_lock);
3382 * Maximum number of compaction retries with a progress before OOM
3383 * killer is consider as the only way to move forward.
3385 #define MAX_COMPACT_RETRIES 16
3387 #ifdef CONFIG_COMPACTION
3388 /* Try memory compaction for high-order allocations before reclaim */
3389 static struct page *
3390 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3391 unsigned int alloc_flags, const struct alloc_context *ac,
3392 enum compact_priority prio, enum compact_result *compact_result)
3394 struct page *page = NULL;
3395 unsigned long pflags;
3396 unsigned int noreclaim_flag;
3401 psi_memstall_enter(&pflags);
3402 delayacct_compact_start();
3403 noreclaim_flag = memalloc_noreclaim_save();
3405 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3408 memalloc_noreclaim_restore(noreclaim_flag);
3409 psi_memstall_leave(&pflags);
3410 delayacct_compact_end();
3412 if (*compact_result == COMPACT_SKIPPED)
3415 * At least in one zone compaction wasn't deferred or skipped, so let's
3416 * count a compaction stall
3418 count_vm_event(COMPACTSTALL);
3420 /* Prep a captured page if available */
3422 prep_new_page(page, order, gfp_mask, alloc_flags);
3424 /* Try get a page from the freelist if available */
3426 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3429 struct zone *zone = page_zone(page);
3431 zone->compact_blockskip_flush = false;
3432 compaction_defer_reset(zone, order, true);
3433 count_vm_event(COMPACTSUCCESS);
3438 * It's bad if compaction run occurs and fails. The most likely reason
3439 * is that pages exist, but not enough to satisfy watermarks.
3441 count_vm_event(COMPACTFAIL);
3449 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3450 enum compact_result compact_result,
3451 enum compact_priority *compact_priority,
3452 int *compaction_retries)
3454 int max_retries = MAX_COMPACT_RETRIES;
3457 int retries = *compaction_retries;
3458 enum compact_priority priority = *compact_priority;
3463 if (fatal_signal_pending(current))
3467 * Compaction was skipped due to a lack of free order-0
3468 * migration targets. Continue if reclaim can help.
3470 if (compact_result == COMPACT_SKIPPED) {
3471 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3476 * Compaction managed to coalesce some page blocks, but the
3477 * allocation failed presumably due to a race. Retry some.
3479 if (compact_result == COMPACT_SUCCESS) {
3481 * !costly requests are much more important than
3482 * __GFP_RETRY_MAYFAIL costly ones because they are de
3483 * facto nofail and invoke OOM killer to move on while
3484 * costly can fail and users are ready to cope with
3485 * that. 1/4 retries is rather arbitrary but we would
3486 * need much more detailed feedback from compaction to
3487 * make a better decision.
3489 if (order > PAGE_ALLOC_COSTLY_ORDER)
3492 if (++(*compaction_retries) <= max_retries) {
3499 * Compaction failed. Retry with increasing priority.
3501 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3502 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3504 if (*compact_priority > min_priority) {
3505 (*compact_priority)--;
3506 *compaction_retries = 0;
3510 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3514 static inline struct page *
3515 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3516 unsigned int alloc_flags, const struct alloc_context *ac,
3517 enum compact_priority prio, enum compact_result *compact_result)
3519 *compact_result = COMPACT_SKIPPED;
3524 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3525 enum compact_result compact_result,
3526 enum compact_priority *compact_priority,
3527 int *compaction_retries)
3532 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3536 * There are setups with compaction disabled which would prefer to loop
3537 * inside the allocator rather than hit the oom killer prematurely.
3538 * Let's give them a good hope and keep retrying while the order-0
3539 * watermarks are OK.
3541 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3542 ac->highest_zoneidx, ac->nodemask) {
3543 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3544 ac->highest_zoneidx, alloc_flags))
3549 #endif /* CONFIG_COMPACTION */
3551 #ifdef CONFIG_LOCKDEP
3552 static struct lockdep_map __fs_reclaim_map =
3553 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3555 static bool __need_reclaim(gfp_t gfp_mask)
3557 /* no reclaim without waiting on it */
3558 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3561 /* this guy won't enter reclaim */
3562 if (current->flags & PF_MEMALLOC)
3565 if (gfp_mask & __GFP_NOLOCKDEP)
3571 void __fs_reclaim_acquire(unsigned long ip)
3573 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3576 void __fs_reclaim_release(unsigned long ip)
3578 lock_release(&__fs_reclaim_map, ip);
3581 void fs_reclaim_acquire(gfp_t gfp_mask)
3583 gfp_mask = current_gfp_context(gfp_mask);
3585 if (__need_reclaim(gfp_mask)) {
3586 if (gfp_mask & __GFP_FS)
3587 __fs_reclaim_acquire(_RET_IP_);
3589 #ifdef CONFIG_MMU_NOTIFIER
3590 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3591 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3596 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3598 void fs_reclaim_release(gfp_t gfp_mask)
3600 gfp_mask = current_gfp_context(gfp_mask);
3602 if (__need_reclaim(gfp_mask)) {
3603 if (gfp_mask & __GFP_FS)
3604 __fs_reclaim_release(_RET_IP_);
3607 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3611 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3612 * have been rebuilt so allocation retries. Reader side does not lock and
3613 * retries the allocation if zonelist changes. Writer side is protected by the
3614 * embedded spin_lock.
3616 static DEFINE_SEQLOCK(zonelist_update_seq);
3618 static unsigned int zonelist_iter_begin(void)
3620 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3621 return read_seqbegin(&zonelist_update_seq);
3626 static unsigned int check_retry_zonelist(unsigned int seq)
3628 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3629 return read_seqretry(&zonelist_update_seq, seq);
3634 /* Perform direct synchronous page reclaim */
3635 static unsigned long
3636 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3637 const struct alloc_context *ac)
3639 unsigned int noreclaim_flag;
3640 unsigned long progress;
3644 /* We now go into synchronous reclaim */
3645 cpuset_memory_pressure_bump();
3646 fs_reclaim_acquire(gfp_mask);
3647 noreclaim_flag = memalloc_noreclaim_save();
3649 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3652 memalloc_noreclaim_restore(noreclaim_flag);
3653 fs_reclaim_release(gfp_mask);
3660 /* The really slow allocator path where we enter direct reclaim */
3661 static inline struct page *
3662 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3663 unsigned int alloc_flags, const struct alloc_context *ac,
3664 unsigned long *did_some_progress)
3666 struct page *page = NULL;
3667 unsigned long pflags;
3668 bool drained = false;
3670 psi_memstall_enter(&pflags);
3671 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3672 if (unlikely(!(*did_some_progress)))
3676 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3679 * If an allocation failed after direct reclaim, it could be because
3680 * pages are pinned on the per-cpu lists or in high alloc reserves.
3681 * Shrink them and try again
3683 if (!page && !drained) {
3684 unreserve_highatomic_pageblock(ac, false);
3685 drain_all_pages(NULL);
3690 psi_memstall_leave(&pflags);
3695 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3696 const struct alloc_context *ac)
3700 pg_data_t *last_pgdat = NULL;
3701 enum zone_type highest_zoneidx = ac->highest_zoneidx;
3703 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3705 if (!managed_zone(zone))
3707 if (last_pgdat != zone->zone_pgdat) {
3708 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3709 last_pgdat = zone->zone_pgdat;
3714 static inline unsigned int
3715 gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3717 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3720 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3721 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3722 * to save two branches.
3724 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
3725 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
3728 * The caller may dip into page reserves a bit more if the caller
3729 * cannot run direct reclaim, or if the caller has realtime scheduling
3730 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3731 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3733 alloc_flags |= (__force int)
3734 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
3736 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
3738 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3739 * if it can't schedule.
3741 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
3742 alloc_flags |= ALLOC_NON_BLOCK;
3745 alloc_flags |= ALLOC_HIGHATOMIC;
3749 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
3750 * GFP_ATOMIC) rather than fail, see the comment for
3751 * cpuset_node_allowed().
3753 if (alloc_flags & ALLOC_MIN_RESERVE)
3754 alloc_flags &= ~ALLOC_CPUSET;
3755 } else if (unlikely(rt_task(current)) && in_task())
3756 alloc_flags |= ALLOC_MIN_RESERVE;
3758 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
3763 static bool oom_reserves_allowed(struct task_struct *tsk)
3765 if (!tsk_is_oom_victim(tsk))
3769 * !MMU doesn't have oom reaper so give access to memory reserves
3770 * only to the thread with TIF_MEMDIE set
3772 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3779 * Distinguish requests which really need access to full memory
3780 * reserves from oom victims which can live with a portion of it
3782 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3784 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3786 if (gfp_mask & __GFP_MEMALLOC)
3787 return ALLOC_NO_WATERMARKS;
3788 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3789 return ALLOC_NO_WATERMARKS;
3790 if (!in_interrupt()) {
3791 if (current->flags & PF_MEMALLOC)
3792 return ALLOC_NO_WATERMARKS;
3793 else if (oom_reserves_allowed(current))
3800 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3802 return !!__gfp_pfmemalloc_flags(gfp_mask);
3806 * Checks whether it makes sense to retry the reclaim to make a forward progress
3807 * for the given allocation request.
3809 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3810 * without success, or when we couldn't even meet the watermark if we
3811 * reclaimed all remaining pages on the LRU lists.
3813 * Returns true if a retry is viable or false to enter the oom path.
3816 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3817 struct alloc_context *ac, int alloc_flags,
3818 bool did_some_progress, int *no_progress_loops)
3825 * Costly allocations might have made a progress but this doesn't mean
3826 * their order will become available due to high fragmentation so
3827 * always increment the no progress counter for them
3829 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3830 *no_progress_loops = 0;
3832 (*no_progress_loops)++;
3834 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
3839 * Keep reclaiming pages while there is a chance this will lead
3840 * somewhere. If none of the target zones can satisfy our allocation
3841 * request even if all reclaimable pages are considered then we are
3842 * screwed and have to go OOM.
3844 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3845 ac->highest_zoneidx, ac->nodemask) {
3846 unsigned long available;
3847 unsigned long reclaimable;
3848 unsigned long min_wmark = min_wmark_pages(zone);
3851 available = reclaimable = zone_reclaimable_pages(zone);
3852 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3855 * Would the allocation succeed if we reclaimed all
3856 * reclaimable pages?
3858 wmark = __zone_watermark_ok(zone, order, min_wmark,
3859 ac->highest_zoneidx, alloc_flags, available);
3860 trace_reclaim_retry_zone(z, order, reclaimable,
3861 available, min_wmark, *no_progress_loops, wmark);
3869 * Memory allocation/reclaim might be called from a WQ context and the
3870 * current implementation of the WQ concurrency control doesn't
3871 * recognize that a particular WQ is congested if the worker thread is
3872 * looping without ever sleeping. Therefore we have to do a short sleep
3873 * here rather than calling cond_resched().
3875 if (current->flags & PF_WQ_WORKER)
3876 schedule_timeout_uninterruptible(1);
3880 /* Before OOM, exhaust highatomic_reserve */
3882 return unreserve_highatomic_pageblock(ac, true);
3888 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3891 * It's possible that cpuset's mems_allowed and the nodemask from
3892 * mempolicy don't intersect. This should be normally dealt with by
3893 * policy_nodemask(), but it's possible to race with cpuset update in
3894 * such a way the check therein was true, and then it became false
3895 * before we got our cpuset_mems_cookie here.
3896 * This assumes that for all allocations, ac->nodemask can come only
3897 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3898 * when it does not intersect with the cpuset restrictions) or the
3899 * caller can deal with a violated nodemask.
3901 if (cpusets_enabled() && ac->nodemask &&
3902 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3903 ac->nodemask = NULL;
3908 * When updating a task's mems_allowed or mempolicy nodemask, it is
3909 * possible to race with parallel threads in such a way that our
3910 * allocation can fail while the mask is being updated. If we are about
3911 * to fail, check if the cpuset changed during allocation and if so,
3914 if (read_mems_allowed_retry(cpuset_mems_cookie))
3920 static inline struct page *
3921 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3922 struct alloc_context *ac)
3924 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3925 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3926 struct page *page = NULL;
3927 unsigned int alloc_flags;
3928 unsigned long did_some_progress;
3929 enum compact_priority compact_priority;
3930 enum compact_result compact_result;
3931 int compaction_retries;
3932 int no_progress_loops;
3933 unsigned int cpuset_mems_cookie;
3934 unsigned int zonelist_iter_cookie;
3938 compaction_retries = 0;
3939 no_progress_loops = 0;
3940 compact_priority = DEF_COMPACT_PRIORITY;
3941 cpuset_mems_cookie = read_mems_allowed_begin();
3942 zonelist_iter_cookie = zonelist_iter_begin();
3945 * The fast path uses conservative alloc_flags to succeed only until
3946 * kswapd needs to be woken up, and to avoid the cost of setting up
3947 * alloc_flags precisely. So we do that now.
3949 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
3952 * We need to recalculate the starting point for the zonelist iterator
3953 * because we might have used different nodemask in the fast path, or
3954 * there was a cpuset modification and we are retrying - otherwise we
3955 * could end up iterating over non-eligible zones endlessly.
3957 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3958 ac->highest_zoneidx, ac->nodemask);
3959 if (!ac->preferred_zoneref->zone)
3963 * Check for insane configurations where the cpuset doesn't contain
3964 * any suitable zone to satisfy the request - e.g. non-movable
3965 * GFP_HIGHUSER allocations from MOVABLE nodes only.
3967 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
3968 struct zoneref *z = first_zones_zonelist(ac->zonelist,
3969 ac->highest_zoneidx,
3970 &cpuset_current_mems_allowed);
3975 if (alloc_flags & ALLOC_KSWAPD)
3976 wake_all_kswapds(order, gfp_mask, ac);
3979 * The adjusted alloc_flags might result in immediate success, so try
3982 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3987 * For costly allocations, try direct compaction first, as it's likely
3988 * that we have enough base pages and don't need to reclaim. For non-
3989 * movable high-order allocations, do that as well, as compaction will
3990 * try prevent permanent fragmentation by migrating from blocks of the
3992 * Don't try this for allocations that are allowed to ignore
3993 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3995 if (can_direct_reclaim &&
3997 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3998 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3999 page = __alloc_pages_direct_compact(gfp_mask, order,
4001 INIT_COMPACT_PRIORITY,
4007 * Checks for costly allocations with __GFP_NORETRY, which
4008 * includes some THP page fault allocations
4010 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4012 * If allocating entire pageblock(s) and compaction
4013 * failed because all zones are below low watermarks
4014 * or is prohibited because it recently failed at this
4015 * order, fail immediately unless the allocator has
4016 * requested compaction and reclaim retry.
4019 * - potentially very expensive because zones are far
4020 * below their low watermarks or this is part of very
4021 * bursty high order allocations,
4022 * - not guaranteed to help because isolate_freepages()
4023 * may not iterate over freed pages as part of its
4025 * - unlikely to make entire pageblocks free on its
4028 if (compact_result == COMPACT_SKIPPED ||
4029 compact_result == COMPACT_DEFERRED)
4033 * Looks like reclaim/compaction is worth trying, but
4034 * sync compaction could be very expensive, so keep
4035 * using async compaction.
4037 compact_priority = INIT_COMPACT_PRIORITY;
4042 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4043 if (alloc_flags & ALLOC_KSWAPD)
4044 wake_all_kswapds(order, gfp_mask, ac);
4046 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4048 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4049 (alloc_flags & ALLOC_KSWAPD);
4052 * Reset the nodemask and zonelist iterators if memory policies can be
4053 * ignored. These allocations are high priority and system rather than
4056 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4057 ac->nodemask = NULL;
4058 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4059 ac->highest_zoneidx, ac->nodemask);
4062 /* Attempt with potentially adjusted zonelist and alloc_flags */
4063 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4067 /* Caller is not willing to reclaim, we can't balance anything */
4068 if (!can_direct_reclaim)
4071 /* Avoid recursion of direct reclaim */
4072 if (current->flags & PF_MEMALLOC)
4075 /* Try direct reclaim and then allocating */
4076 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4077 &did_some_progress);
4081 /* Try direct compaction and then allocating */
4082 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4083 compact_priority, &compact_result);
4087 /* Do not loop if specifically requested */
4088 if (gfp_mask & __GFP_NORETRY)
4092 * Do not retry costly high order allocations unless they are
4093 * __GFP_RETRY_MAYFAIL
4095 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4098 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4099 did_some_progress > 0, &no_progress_loops))
4103 * It doesn't make any sense to retry for the compaction if the order-0
4104 * reclaim is not able to make any progress because the current
4105 * implementation of the compaction depends on the sufficient amount
4106 * of free memory (see __compaction_suitable)
4108 if (did_some_progress > 0 &&
4109 should_compact_retry(ac, order, alloc_flags,
4110 compact_result, &compact_priority,
4111 &compaction_retries))
4116 * Deal with possible cpuset update races or zonelist updates to avoid
4117 * a unnecessary OOM kill.
4119 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4120 check_retry_zonelist(zonelist_iter_cookie))
4123 /* Reclaim has failed us, start killing things */
4124 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4128 /* Avoid allocations with no watermarks from looping endlessly */
4129 if (tsk_is_oom_victim(current) &&
4130 (alloc_flags & ALLOC_OOM ||
4131 (gfp_mask & __GFP_NOMEMALLOC)))
4134 /* Retry as long as the OOM killer is making progress */
4135 if (did_some_progress) {
4136 no_progress_loops = 0;
4142 * Deal with possible cpuset update races or zonelist updates to avoid
4143 * a unnecessary OOM kill.
4145 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4146 check_retry_zonelist(zonelist_iter_cookie))
4150 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4153 if (gfp_mask & __GFP_NOFAIL) {
4155 * All existing users of the __GFP_NOFAIL are blockable, so warn
4156 * of any new users that actually require GFP_NOWAIT
4158 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4162 * PF_MEMALLOC request from this context is rather bizarre
4163 * because we cannot reclaim anything and only can loop waiting
4164 * for somebody to do a work for us
4166 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4169 * non failing costly orders are a hard requirement which we
4170 * are not prepared for much so let's warn about these users
4171 * so that we can identify them and convert them to something
4174 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4177 * Help non-failing allocations by giving some access to memory
4178 * reserves normally used for high priority non-blocking
4179 * allocations but do not use ALLOC_NO_WATERMARKS because this
4180 * could deplete whole memory reserves which would just make
4181 * the situation worse.
4183 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4191 warn_alloc(gfp_mask, ac->nodemask,
4192 "page allocation failure: order:%u", order);
4197 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4198 int preferred_nid, nodemask_t *nodemask,
4199 struct alloc_context *ac, gfp_t *alloc_gfp,
4200 unsigned int *alloc_flags)
4202 ac->highest_zoneidx = gfp_zone(gfp_mask);
4203 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4204 ac->nodemask = nodemask;
4205 ac->migratetype = gfp_migratetype(gfp_mask);
4207 if (cpusets_enabled()) {
4208 *alloc_gfp |= __GFP_HARDWALL;
4210 * When we are in the interrupt context, it is irrelevant
4211 * to the current task context. It means that any node ok.
4213 if (in_task() && !ac->nodemask)
4214 ac->nodemask = &cpuset_current_mems_allowed;
4216 *alloc_flags |= ALLOC_CPUSET;
4219 might_alloc(gfp_mask);
4221 if (should_fail_alloc_page(gfp_mask, order))
4224 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4226 /* Dirty zone balancing only done in the fast path */
4227 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4230 * The preferred zone is used for statistics but crucially it is
4231 * also used as the starting point for the zonelist iterator. It
4232 * may get reset for allocations that ignore memory policies.
4234 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4235 ac->highest_zoneidx, ac->nodemask);
4241 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4242 * @gfp: GFP flags for the allocation
4243 * @preferred_nid: The preferred NUMA node ID to allocate from
4244 * @nodemask: Set of nodes to allocate from, may be NULL
4245 * @nr_pages: The number of pages desired on the list or array
4246 * @page_list: Optional list to store the allocated pages
4247 * @page_array: Optional array to store the pages
4249 * This is a batched version of the page allocator that attempts to
4250 * allocate nr_pages quickly. Pages are added to page_list if page_list
4251 * is not NULL, otherwise it is assumed that the page_array is valid.
4253 * For lists, nr_pages is the number of pages that should be allocated.
4255 * For arrays, only NULL elements are populated with pages and nr_pages
4256 * is the maximum number of pages that will be stored in the array.
4258 * Returns the number of pages on the list or array.
4260 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
4261 nodemask_t *nodemask, int nr_pages,
4262 struct list_head *page_list,
4263 struct page **page_array)
4266 unsigned long __maybe_unused UP_flags;
4269 struct per_cpu_pages *pcp;
4270 struct list_head *pcp_list;
4271 struct alloc_context ac;
4273 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4274 int nr_populated = 0, nr_account = 0;
4277 * Skip populated array elements to determine if any pages need
4278 * to be allocated before disabling IRQs.
4280 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4283 /* No pages requested? */
4284 if (unlikely(nr_pages <= 0))
4287 /* Already populated array? */
4288 if (unlikely(page_array && nr_pages - nr_populated == 0))
4291 /* Bulk allocator does not support memcg accounting. */
4292 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4295 /* Use the single page allocator for one page. */
4296 if (nr_pages - nr_populated == 1)
4299 #ifdef CONFIG_PAGE_OWNER
4301 * PAGE_OWNER may recurse into the allocator to allocate space to
4302 * save the stack with pagesets.lock held. Releasing/reacquiring
4303 * removes much of the performance benefit of bulk allocation so
4304 * force the caller to allocate one page at a time as it'll have
4305 * similar performance to added complexity to the bulk allocator.
4307 if (static_branch_unlikely(&page_owner_inited))
4311 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4312 gfp &= gfp_allowed_mask;
4314 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4318 /* Find an allowed local zone that meets the low watermark. */
4319 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4322 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4323 !__cpuset_zone_allowed(zone, gfp)) {
4327 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4328 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4332 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4333 if (zone_watermark_fast(zone, 0, mark,
4334 zonelist_zone_idx(ac.preferred_zoneref),
4335 alloc_flags, gfp)) {
4341 * If there are no allowed local zones that meets the watermarks then
4342 * try to allocate a single page and reclaim if necessary.
4344 if (unlikely(!zone))
4347 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4348 pcp_trylock_prepare(UP_flags);
4349 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4353 /* Attempt the batch allocation */
4354 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4355 while (nr_populated < nr_pages) {
4357 /* Skip existing pages */
4358 if (page_array && page_array[nr_populated]) {
4363 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4365 if (unlikely(!page)) {
4366 /* Try and allocate at least one page */
4368 pcp_spin_unlock(pcp);
4375 prep_new_page(page, 0, gfp, 0);
4377 list_add(&page->lru, page_list);
4379 page_array[nr_populated] = page;
4383 pcp_spin_unlock(pcp);
4384 pcp_trylock_finish(UP_flags);
4386 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4387 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4390 return nr_populated;
4393 pcp_trylock_finish(UP_flags);
4396 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
4399 list_add(&page->lru, page_list);
4401 page_array[nr_populated] = page;
4407 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
4410 * This is the 'heart' of the zoned buddy allocator.
4412 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
4413 nodemask_t *nodemask)
4416 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4417 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4418 struct alloc_context ac = { };
4421 * There are several places where we assume that the order value is sane
4422 * so bail out early if the request is out of bound.
4424 if (WARN_ON_ONCE_GFP(order > MAX_ORDER, gfp))
4427 gfp &= gfp_allowed_mask;
4429 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4430 * resp. GFP_NOIO which has to be inherited for all allocation requests
4431 * from a particular context which has been marked by
4432 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4433 * movable zones are not used during allocation.
4435 gfp = current_gfp_context(gfp);
4437 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4438 &alloc_gfp, &alloc_flags))
4442 * Forbid the first pass from falling back to types that fragment
4443 * memory until all local zones are considered.
4445 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4447 /* First allocation attempt */
4448 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4453 ac.spread_dirty_pages = false;
4456 * Restore the original nodemask if it was potentially replaced with
4457 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4459 ac.nodemask = nodemask;
4461 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4464 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4465 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4466 __free_pages(page, order);
4470 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4471 kmsan_alloc_page(page, order, alloc_gfp);
4475 EXPORT_SYMBOL(__alloc_pages);
4477 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
4478 nodemask_t *nodemask)
4480 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
4481 preferred_nid, nodemask);
4482 struct folio *folio = (struct folio *)page;
4484 if (folio && order > 1)
4485 folio_prep_large_rmappable(folio);
4488 EXPORT_SYMBOL(__folio_alloc);
4491 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4492 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4493 * you need to access high mem.
4495 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4499 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4502 return (unsigned long) page_address(page);
4504 EXPORT_SYMBOL(__get_free_pages);
4506 unsigned long get_zeroed_page(gfp_t gfp_mask)
4508 return __get_free_page(gfp_mask | __GFP_ZERO);
4510 EXPORT_SYMBOL(get_zeroed_page);
4513 * __free_pages - Free pages allocated with alloc_pages().
4514 * @page: The page pointer returned from alloc_pages().
4515 * @order: The order of the allocation.
4517 * This function can free multi-page allocations that are not compound
4518 * pages. It does not check that the @order passed in matches that of
4519 * the allocation, so it is easy to leak memory. Freeing more memory
4520 * than was allocated will probably emit a warning.
4522 * If the last reference to this page is speculative, it will be released
4523 * by put_page() which only frees the first page of a non-compound
4524 * allocation. To prevent the remaining pages from being leaked, we free
4525 * the subsequent pages here. If you want to use the page's reference
4526 * count to decide when to free the allocation, you should allocate a
4527 * compound page, and use put_page() instead of __free_pages().
4529 * Context: May be called in interrupt context or while holding a normal
4530 * spinlock, but not in NMI context or while holding a raw spinlock.
4532 void __free_pages(struct page *page, unsigned int order)
4534 /* get PageHead before we drop reference */
4535 int head = PageHead(page);
4537 if (put_page_testzero(page))
4538 free_the_page(page, order);
4541 free_the_page(page + (1 << order), order);
4543 EXPORT_SYMBOL(__free_pages);
4545 void free_pages(unsigned long addr, unsigned int order)
4548 VM_BUG_ON(!virt_addr_valid((void *)addr));
4549 __free_pages(virt_to_page((void *)addr), order);
4553 EXPORT_SYMBOL(free_pages);
4557 * An arbitrary-length arbitrary-offset area of memory which resides
4558 * within a 0 or higher order page. Multiple fragments within that page
4559 * are individually refcounted, in the page's reference counter.
4561 * The page_frag functions below provide a simple allocation framework for
4562 * page fragments. This is used by the network stack and network device
4563 * drivers to provide a backing region of memory for use as either an
4564 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4566 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4569 struct page *page = NULL;
4570 gfp_t gfp = gfp_mask;
4572 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4573 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4575 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4576 PAGE_FRAG_CACHE_MAX_ORDER);
4577 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4579 if (unlikely(!page))
4580 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4582 nc->va = page ? page_address(page) : NULL;
4587 void __page_frag_cache_drain(struct page *page, unsigned int count)
4589 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4591 if (page_ref_sub_and_test(page, count))
4592 free_the_page(page, compound_order(page));
4594 EXPORT_SYMBOL(__page_frag_cache_drain);
4596 void *page_frag_alloc_align(struct page_frag_cache *nc,
4597 unsigned int fragsz, gfp_t gfp_mask,
4598 unsigned int align_mask)
4600 unsigned int size = PAGE_SIZE;
4604 if (unlikely(!nc->va)) {
4606 page = __page_frag_cache_refill(nc, gfp_mask);
4610 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4611 /* if size can vary use size else just use PAGE_SIZE */
4614 /* Even if we own the page, we do not use atomic_set().
4615 * This would break get_page_unless_zero() users.
4617 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4619 /* reset page count bias and offset to start of new frag */
4620 nc->pfmemalloc = page_is_pfmemalloc(page);
4621 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4625 offset = nc->offset - fragsz;
4626 if (unlikely(offset < 0)) {
4627 page = virt_to_page(nc->va);
4629 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4632 if (unlikely(nc->pfmemalloc)) {
4633 free_the_page(page, compound_order(page));
4637 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4638 /* if size can vary use size else just use PAGE_SIZE */
4641 /* OK, page count is 0, we can safely set it */
4642 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4644 /* reset page count bias and offset to start of new frag */
4645 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4646 offset = size - fragsz;
4647 if (unlikely(offset < 0)) {
4649 * The caller is trying to allocate a fragment
4650 * with fragsz > PAGE_SIZE but the cache isn't big
4651 * enough to satisfy the request, this may
4652 * happen in low memory conditions.
4653 * We don't release the cache page because
4654 * it could make memory pressure worse
4655 * so we simply return NULL here.
4662 offset &= align_mask;
4663 nc->offset = offset;
4665 return nc->va + offset;
4667 EXPORT_SYMBOL(page_frag_alloc_align);
4670 * Frees a page fragment allocated out of either a compound or order 0 page.
4672 void page_frag_free(void *addr)
4674 struct page *page = virt_to_head_page(addr);
4676 if (unlikely(put_page_testzero(page)))
4677 free_the_page(page, compound_order(page));
4679 EXPORT_SYMBOL(page_frag_free);
4681 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4685 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4686 struct page *page = virt_to_page((void *)addr);
4687 struct page *last = page + nr;
4689 split_page_owner(page, 1 << order);
4690 split_page_memcg(page, 1 << order);
4691 while (page < --last)
4692 set_page_refcounted(last);
4694 last = page + (1UL << order);
4695 for (page += nr; page < last; page++)
4696 __free_pages_ok(page, 0, FPI_TO_TAIL);
4698 return (void *)addr;
4702 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4703 * @size: the number of bytes to allocate
4704 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4706 * This function is similar to alloc_pages(), except that it allocates the
4707 * minimum number of pages to satisfy the request. alloc_pages() can only
4708 * allocate memory in power-of-two pages.
4710 * This function is also limited by MAX_ORDER.
4712 * Memory allocated by this function must be released by free_pages_exact().
4714 * Return: pointer to the allocated area or %NULL in case of error.
4716 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4718 unsigned int order = get_order(size);
4721 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4722 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4724 addr = __get_free_pages(gfp_mask, order);
4725 return make_alloc_exact(addr, order, size);
4727 EXPORT_SYMBOL(alloc_pages_exact);
4730 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4732 * @nid: the preferred node ID where memory should be allocated
4733 * @size: the number of bytes to allocate
4734 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4736 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4739 * Return: pointer to the allocated area or %NULL in case of error.
4741 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4743 unsigned int order = get_order(size);
4746 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4747 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4749 p = alloc_pages_node(nid, gfp_mask, order);
4752 return make_alloc_exact((unsigned long)page_address(p), order, size);
4756 * free_pages_exact - release memory allocated via alloc_pages_exact()
4757 * @virt: the value returned by alloc_pages_exact.
4758 * @size: size of allocation, same value as passed to alloc_pages_exact().
4760 * Release the memory allocated by a previous call to alloc_pages_exact.
4762 void free_pages_exact(void *virt, size_t size)
4764 unsigned long addr = (unsigned long)virt;
4765 unsigned long end = addr + PAGE_ALIGN(size);
4767 while (addr < end) {
4772 EXPORT_SYMBOL(free_pages_exact);
4775 * nr_free_zone_pages - count number of pages beyond high watermark
4776 * @offset: The zone index of the highest zone
4778 * nr_free_zone_pages() counts the number of pages which are beyond the
4779 * high watermark within all zones at or below a given zone index. For each
4780 * zone, the number of pages is calculated as:
4782 * nr_free_zone_pages = managed_pages - high_pages
4784 * Return: number of pages beyond high watermark.
4786 static unsigned long nr_free_zone_pages(int offset)
4791 /* Just pick one node, since fallback list is circular */
4792 unsigned long sum = 0;
4794 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4796 for_each_zone_zonelist(zone, z, zonelist, offset) {
4797 unsigned long size = zone_managed_pages(zone);
4798 unsigned long high = high_wmark_pages(zone);
4807 * nr_free_buffer_pages - count number of pages beyond high watermark
4809 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4810 * watermark within ZONE_DMA and ZONE_NORMAL.
4812 * Return: number of pages beyond high watermark within ZONE_DMA and
4815 unsigned long nr_free_buffer_pages(void)
4817 return nr_free_zone_pages(gfp_zone(GFP_USER));
4819 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4821 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4823 zoneref->zone = zone;
4824 zoneref->zone_idx = zone_idx(zone);
4828 * Builds allocation fallback zone lists.
4830 * Add all populated zones of a node to the zonelist.
4832 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4835 enum zone_type zone_type = MAX_NR_ZONES;
4840 zone = pgdat->node_zones + zone_type;
4841 if (populated_zone(zone)) {
4842 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4843 check_highest_zone(zone_type);
4845 } while (zone_type);
4852 static int __parse_numa_zonelist_order(char *s)
4855 * We used to support different zonelists modes but they turned
4856 * out to be just not useful. Let's keep the warning in place
4857 * if somebody still use the cmd line parameter so that we do
4858 * not fail it silently
4860 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4861 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4867 static char numa_zonelist_order[] = "Node";
4868 #define NUMA_ZONELIST_ORDER_LEN 16
4870 * sysctl handler for numa_zonelist_order
4872 static int numa_zonelist_order_handler(struct ctl_table *table, int write,
4873 void *buffer, size_t *length, loff_t *ppos)
4876 return __parse_numa_zonelist_order(buffer);
4877 return proc_dostring(table, write, buffer, length, ppos);
4880 static int node_load[MAX_NUMNODES];
4883 * find_next_best_node - find the next node that should appear in a given node's fallback list
4884 * @node: node whose fallback list we're appending
4885 * @used_node_mask: nodemask_t of already used nodes
4887 * We use a number of factors to determine which is the next node that should
4888 * appear on a given node's fallback list. The node should not have appeared
4889 * already in @node's fallback list, and it should be the next closest node
4890 * according to the distance array (which contains arbitrary distance values
4891 * from each node to each node in the system), and should also prefer nodes
4892 * with no CPUs, since presumably they'll have very little allocation pressure
4893 * on them otherwise.
4895 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
4897 int find_next_best_node(int node, nodemask_t *used_node_mask)
4900 int min_val = INT_MAX;
4901 int best_node = NUMA_NO_NODE;
4903 /* Use the local node if we haven't already */
4904 if (!node_isset(node, *used_node_mask)) {
4905 node_set(node, *used_node_mask);
4909 for_each_node_state(n, N_MEMORY) {
4911 /* Don't want a node to appear more than once */
4912 if (node_isset(n, *used_node_mask))
4915 /* Use the distance array to find the distance */
4916 val = node_distance(node, n);
4918 /* Penalize nodes under us ("prefer the next node") */
4921 /* Give preference to headless and unused nodes */
4922 if (!cpumask_empty(cpumask_of_node(n)))
4923 val += PENALTY_FOR_NODE_WITH_CPUS;
4925 /* Slight preference for less loaded node */
4926 val *= MAX_NUMNODES;
4927 val += node_load[n];
4929 if (val < min_val) {
4936 node_set(best_node, *used_node_mask);
4943 * Build zonelists ordered by node and zones within node.
4944 * This results in maximum locality--normal zone overflows into local
4945 * DMA zone, if any--but risks exhausting DMA zone.
4947 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
4950 struct zoneref *zonerefs;
4953 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
4955 for (i = 0; i < nr_nodes; i++) {
4958 pg_data_t *node = NODE_DATA(node_order[i]);
4960 nr_zones = build_zonerefs_node(node, zonerefs);
4961 zonerefs += nr_zones;
4963 zonerefs->zone = NULL;
4964 zonerefs->zone_idx = 0;
4968 * Build gfp_thisnode zonelists
4970 static void build_thisnode_zonelists(pg_data_t *pgdat)
4972 struct zoneref *zonerefs;
4975 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
4976 nr_zones = build_zonerefs_node(pgdat, zonerefs);
4977 zonerefs += nr_zones;
4978 zonerefs->zone = NULL;
4979 zonerefs->zone_idx = 0;
4983 * Build zonelists ordered by zone and nodes within zones.
4984 * This results in conserving DMA zone[s] until all Normal memory is
4985 * exhausted, but results in overflowing to remote node while memory
4986 * may still exist in local DMA zone.
4989 static void build_zonelists(pg_data_t *pgdat)
4991 static int node_order[MAX_NUMNODES];
4992 int node, nr_nodes = 0;
4993 nodemask_t used_mask = NODE_MASK_NONE;
4994 int local_node, prev_node;
4996 /* NUMA-aware ordering of nodes */
4997 local_node = pgdat->node_id;
4998 prev_node = local_node;
5000 memset(node_order, 0, sizeof(node_order));
5001 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5003 * We don't want to pressure a particular node.
5004 * So adding penalty to the first node in same
5005 * distance group to make it round-robin.
5007 if (node_distance(local_node, node) !=
5008 node_distance(local_node, prev_node))
5009 node_load[node] += 1;
5011 node_order[nr_nodes++] = node;
5015 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5016 build_thisnode_zonelists(pgdat);
5017 pr_info("Fallback order for Node %d: ", local_node);
5018 for (node = 0; node < nr_nodes; node++)
5019 pr_cont("%d ", node_order[node]);
5023 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5025 * Return node id of node used for "local" allocations.
5026 * I.e., first node id of first zone in arg node's generic zonelist.
5027 * Used for initializing percpu 'numa_mem', which is used primarily
5028 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5030 int local_memory_node(int node)
5034 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5035 gfp_zone(GFP_KERNEL),
5037 return zone_to_nid(z->zone);
5041 static void setup_min_unmapped_ratio(void);
5042 static void setup_min_slab_ratio(void);
5043 #else /* CONFIG_NUMA */
5045 static void build_zonelists(pg_data_t *pgdat)
5047 int node, local_node;
5048 struct zoneref *zonerefs;
5051 local_node = pgdat->node_id;
5053 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5054 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5055 zonerefs += nr_zones;
5058 * Now we build the zonelist so that it contains the zones
5059 * of all the other nodes.
5060 * We don't want to pressure a particular node, so when
5061 * building the zones for node N, we make sure that the
5062 * zones coming right after the local ones are those from
5063 * node N+1 (modulo N)
5065 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5066 if (!node_online(node))
5068 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5069 zonerefs += nr_zones;
5071 for (node = 0; node < local_node; node++) {
5072 if (!node_online(node))
5074 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5075 zonerefs += nr_zones;
5078 zonerefs->zone = NULL;
5079 zonerefs->zone_idx = 0;
5082 #endif /* CONFIG_NUMA */
5085 * Boot pageset table. One per cpu which is going to be used for all
5086 * zones and all nodes. The parameters will be set in such a way
5087 * that an item put on a list will immediately be handed over to
5088 * the buddy list. This is safe since pageset manipulation is done
5089 * with interrupts disabled.
5091 * The boot_pagesets must be kept even after bootup is complete for
5092 * unused processors and/or zones. They do play a role for bootstrapping
5093 * hotplugged processors.
5095 * zoneinfo_show() and maybe other functions do
5096 * not check if the processor is online before following the pageset pointer.
5097 * Other parts of the kernel may not check if the zone is available.
5099 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5100 /* These effectively disable the pcplists in the boot pageset completely */
5101 #define BOOT_PAGESET_HIGH 0
5102 #define BOOT_PAGESET_BATCH 1
5103 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5104 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5106 static void __build_all_zonelists(void *data)
5109 int __maybe_unused cpu;
5110 pg_data_t *self = data;
5111 unsigned long flags;
5114 * The zonelist_update_seq must be acquired with irqsave because the
5115 * reader can be invoked from IRQ with GFP_ATOMIC.
5117 write_seqlock_irqsave(&zonelist_update_seq, flags);
5119 * Also disable synchronous printk() to prevent any printk() from
5120 * trying to hold port->lock, for
5121 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5122 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5124 printk_deferred_enter();
5127 memset(node_load, 0, sizeof(node_load));
5131 * This node is hotadded and no memory is yet present. So just
5132 * building zonelists is fine - no need to touch other nodes.
5134 if (self && !node_online(self->node_id)) {
5135 build_zonelists(self);
5138 * All possible nodes have pgdat preallocated
5141 for_each_node(nid) {
5142 pg_data_t *pgdat = NODE_DATA(nid);
5144 build_zonelists(pgdat);
5147 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5149 * We now know the "local memory node" for each node--
5150 * i.e., the node of the first zone in the generic zonelist.
5151 * Set up numa_mem percpu variable for on-line cpus. During
5152 * boot, only the boot cpu should be on-line; we'll init the
5153 * secondary cpus' numa_mem as they come on-line. During
5154 * node/memory hotplug, we'll fixup all on-line cpus.
5156 for_each_online_cpu(cpu)
5157 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5161 printk_deferred_exit();
5162 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5165 static noinline void __init
5166 build_all_zonelists_init(void)
5170 __build_all_zonelists(NULL);
5173 * Initialize the boot_pagesets that are going to be used
5174 * for bootstrapping processors. The real pagesets for
5175 * each zone will be allocated later when the per cpu
5176 * allocator is available.
5178 * boot_pagesets are used also for bootstrapping offline
5179 * cpus if the system is already booted because the pagesets
5180 * are needed to initialize allocators on a specific cpu too.
5181 * F.e. the percpu allocator needs the page allocator which
5182 * needs the percpu allocator in order to allocate its pagesets
5183 * (a chicken-egg dilemma).
5185 for_each_possible_cpu(cpu)
5186 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5188 mminit_verify_zonelist();
5189 cpuset_init_current_mems_allowed();
5193 * unless system_state == SYSTEM_BOOTING.
5195 * __ref due to call of __init annotated helper build_all_zonelists_init
5196 * [protected by SYSTEM_BOOTING].
5198 void __ref build_all_zonelists(pg_data_t *pgdat)
5200 unsigned long vm_total_pages;
5202 if (system_state == SYSTEM_BOOTING) {
5203 build_all_zonelists_init();
5205 __build_all_zonelists(pgdat);
5206 /* cpuset refresh routine should be here */
5208 /* Get the number of free pages beyond high watermark in all zones. */
5209 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5211 * Disable grouping by mobility if the number of pages in the
5212 * system is too low to allow the mechanism to work. It would be
5213 * more accurate, but expensive to check per-zone. This check is
5214 * made on memory-hotadd so a system can start with mobility
5215 * disabled and enable it later
5217 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5218 page_group_by_mobility_disabled = 1;
5220 page_group_by_mobility_disabled = 0;
5222 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5224 page_group_by_mobility_disabled ? "off" : "on",
5227 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5231 static int zone_batchsize(struct zone *zone)
5237 * The number of pages to batch allocate is either ~0.1%
5238 * of the zone or 1MB, whichever is smaller. The batch
5239 * size is striking a balance between allocation latency
5240 * and zone lock contention.
5242 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5243 batch /= 4; /* We effectively *= 4 below */
5248 * Clamp the batch to a 2^n - 1 value. Having a power
5249 * of 2 value was found to be more likely to have
5250 * suboptimal cache aliasing properties in some cases.
5252 * For example if 2 tasks are alternately allocating
5253 * batches of pages, one task can end up with a lot
5254 * of pages of one half of the possible page colors
5255 * and the other with pages of the other colors.
5257 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5262 /* The deferral and batching of frees should be suppressed under NOMMU
5265 * The problem is that NOMMU needs to be able to allocate large chunks
5266 * of contiguous memory as there's no hardware page translation to
5267 * assemble apparent contiguous memory from discontiguous pages.
5269 * Queueing large contiguous runs of pages for batching, however,
5270 * causes the pages to actually be freed in smaller chunks. As there
5271 * can be a significant delay between the individual batches being
5272 * recycled, this leads to the once large chunks of space being
5273 * fragmented and becoming unavailable for high-order allocations.
5279 static int percpu_pagelist_high_fraction;
5280 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
5285 unsigned long total_pages;
5287 if (!percpu_pagelist_high_fraction) {
5289 * By default, the high value of the pcp is based on the zone
5290 * low watermark so that if they are full then background
5291 * reclaim will not be started prematurely.
5293 total_pages = low_wmark_pages(zone);
5296 * If percpu_pagelist_high_fraction is configured, the high
5297 * value is based on a fraction of the managed pages in the
5300 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
5304 * Split the high value across all online CPUs local to the zone. Note
5305 * that early in boot that CPUs may not be online yet and that during
5306 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5307 * onlined. For memory nodes that have no CPUs, split pcp->high across
5308 * all online CPUs to mitigate the risk that reclaim is triggered
5309 * prematurely due to pages stored on pcp lists.
5311 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5313 nr_split_cpus = num_online_cpus();
5314 high = total_pages / nr_split_cpus;
5317 * Ensure high is at least batch*4. The multiple is based on the
5318 * historical relationship between high and batch.
5320 high = max(high, batch << 2);
5329 * pcp->high and pcp->batch values are related and generally batch is lower
5330 * than high. They are also related to pcp->count such that count is lower
5331 * than high, and as soon as it reaches high, the pcplist is flushed.
5333 * However, guaranteeing these relations at all times would require e.g. write
5334 * barriers here but also careful usage of read barriers at the read side, and
5335 * thus be prone to error and bad for performance. Thus the update only prevents
5336 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
5337 * can cope with those fields changing asynchronously, and fully trust only the
5338 * pcp->count field on the local CPU with interrupts disabled.
5340 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5341 * outside of boot time (or some other assurance that no concurrent updaters
5344 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5345 unsigned long batch)
5347 WRITE_ONCE(pcp->batch, batch);
5348 WRITE_ONCE(pcp->high, high);
5351 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5355 memset(pcp, 0, sizeof(*pcp));
5356 memset(pzstats, 0, sizeof(*pzstats));
5358 spin_lock_init(&pcp->lock);
5359 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5360 INIT_LIST_HEAD(&pcp->lists[pindex]);
5363 * Set batch and high values safe for a boot pageset. A true percpu
5364 * pageset's initialization will update them subsequently. Here we don't
5365 * need to be as careful as pageset_update() as nobody can access the
5368 pcp->high = BOOT_PAGESET_HIGH;
5369 pcp->batch = BOOT_PAGESET_BATCH;
5370 pcp->free_factor = 0;
5373 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
5374 unsigned long batch)
5376 struct per_cpu_pages *pcp;
5379 for_each_possible_cpu(cpu) {
5380 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5381 pageset_update(pcp, high, batch);
5386 * Calculate and set new high and batch values for all per-cpu pagesets of a
5387 * zone based on the zone's size.
5389 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5391 int new_high, new_batch;
5393 new_batch = max(1, zone_batchsize(zone));
5394 new_high = zone_highsize(zone, new_batch, cpu_online);
5396 if (zone->pageset_high == new_high &&
5397 zone->pageset_batch == new_batch)
5400 zone->pageset_high = new_high;
5401 zone->pageset_batch = new_batch;
5403 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
5406 void __meminit setup_zone_pageset(struct zone *zone)
5410 /* Size may be 0 on !SMP && !NUMA */
5411 if (sizeof(struct per_cpu_zonestat) > 0)
5412 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5414 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5415 for_each_possible_cpu(cpu) {
5416 struct per_cpu_pages *pcp;
5417 struct per_cpu_zonestat *pzstats;
5419 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5420 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5421 per_cpu_pages_init(pcp, pzstats);
5424 zone_set_pageset_high_and_batch(zone, 0);
5428 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5429 * page high values need to be recalculated.
5431 static void zone_pcp_update(struct zone *zone, int cpu_online)
5433 mutex_lock(&pcp_batch_high_lock);
5434 zone_set_pageset_high_and_batch(zone, cpu_online);
5435 mutex_unlock(&pcp_batch_high_lock);
5439 * Allocate per cpu pagesets and initialize them.
5440 * Before this call only boot pagesets were available.
5442 void __init setup_per_cpu_pageset(void)
5444 struct pglist_data *pgdat;
5446 int __maybe_unused cpu;
5448 for_each_populated_zone(zone)
5449 setup_zone_pageset(zone);
5453 * Unpopulated zones continue using the boot pagesets.
5454 * The numa stats for these pagesets need to be reset.
5455 * Otherwise, they will end up skewing the stats of
5456 * the nodes these zones are associated with.
5458 for_each_possible_cpu(cpu) {
5459 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5460 memset(pzstats->vm_numa_event, 0,
5461 sizeof(pzstats->vm_numa_event));
5465 for_each_online_pgdat(pgdat)
5466 pgdat->per_cpu_nodestats =
5467 alloc_percpu(struct per_cpu_nodestat);
5470 __meminit void zone_pcp_init(struct zone *zone)
5473 * per cpu subsystem is not up at this point. The following code
5474 * relies on the ability of the linker to provide the
5475 * offset of a (static) per cpu variable into the per cpu area.
5477 zone->per_cpu_pageset = &boot_pageset;
5478 zone->per_cpu_zonestats = &boot_zonestats;
5479 zone->pageset_high = BOOT_PAGESET_HIGH;
5480 zone->pageset_batch = BOOT_PAGESET_BATCH;
5482 if (populated_zone(zone))
5483 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5484 zone->present_pages, zone_batchsize(zone));
5487 void adjust_managed_page_count(struct page *page, long count)
5489 atomic_long_add(count, &page_zone(page)->managed_pages);
5490 totalram_pages_add(count);
5491 #ifdef CONFIG_HIGHMEM
5492 if (PageHighMem(page))
5493 totalhigh_pages_add(count);
5496 EXPORT_SYMBOL(adjust_managed_page_count);
5498 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5501 unsigned long pages = 0;
5503 start = (void *)PAGE_ALIGN((unsigned long)start);
5504 end = (void *)((unsigned long)end & PAGE_MASK);
5505 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5506 struct page *page = virt_to_page(pos);
5507 void *direct_map_addr;
5510 * 'direct_map_addr' might be different from 'pos'
5511 * because some architectures' virt_to_page()
5512 * work with aliases. Getting the direct map
5513 * address ensures that we get a _writeable_
5514 * alias for the memset().
5516 direct_map_addr = page_address(page);
5518 * Perform a kasan-unchecked memset() since this memory
5519 * has not been initialized.
5521 direct_map_addr = kasan_reset_tag(direct_map_addr);
5522 if ((unsigned int)poison <= 0xFF)
5523 memset(direct_map_addr, poison, PAGE_SIZE);
5525 free_reserved_page(page);
5529 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5534 static int page_alloc_cpu_dead(unsigned int cpu)
5538 lru_add_drain_cpu(cpu);
5539 mlock_drain_remote(cpu);
5543 * Spill the event counters of the dead processor
5544 * into the current processors event counters.
5545 * This artificially elevates the count of the current
5548 vm_events_fold_cpu(cpu);
5551 * Zero the differential counters of the dead processor
5552 * so that the vm statistics are consistent.
5554 * This is only okay since the processor is dead and cannot
5555 * race with what we are doing.
5557 cpu_vm_stats_fold(cpu);
5559 for_each_populated_zone(zone)
5560 zone_pcp_update(zone, 0);
5565 static int page_alloc_cpu_online(unsigned int cpu)
5569 for_each_populated_zone(zone)
5570 zone_pcp_update(zone, 1);
5574 void __init page_alloc_init_cpuhp(void)
5578 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5579 "mm/page_alloc:pcp",
5580 page_alloc_cpu_online,
5581 page_alloc_cpu_dead);
5586 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5587 * or min_free_kbytes changes.
5589 static void calculate_totalreserve_pages(void)
5591 struct pglist_data *pgdat;
5592 unsigned long reserve_pages = 0;
5593 enum zone_type i, j;
5595 for_each_online_pgdat(pgdat) {
5597 pgdat->totalreserve_pages = 0;
5599 for (i = 0; i < MAX_NR_ZONES; i++) {
5600 struct zone *zone = pgdat->node_zones + i;
5602 unsigned long managed_pages = zone_managed_pages(zone);
5604 /* Find valid and maximum lowmem_reserve in the zone */
5605 for (j = i; j < MAX_NR_ZONES; j++) {
5606 if (zone->lowmem_reserve[j] > max)
5607 max = zone->lowmem_reserve[j];
5610 /* we treat the high watermark as reserved pages. */
5611 max += high_wmark_pages(zone);
5613 if (max > managed_pages)
5614 max = managed_pages;
5616 pgdat->totalreserve_pages += max;
5618 reserve_pages += max;
5621 totalreserve_pages = reserve_pages;
5625 * setup_per_zone_lowmem_reserve - called whenever
5626 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5627 * has a correct pages reserved value, so an adequate number of
5628 * pages are left in the zone after a successful __alloc_pages().
5630 static void setup_per_zone_lowmem_reserve(void)
5632 struct pglist_data *pgdat;
5633 enum zone_type i, j;
5635 for_each_online_pgdat(pgdat) {
5636 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5637 struct zone *zone = &pgdat->node_zones[i];
5638 int ratio = sysctl_lowmem_reserve_ratio[i];
5639 bool clear = !ratio || !zone_managed_pages(zone);
5640 unsigned long managed_pages = 0;
5642 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5643 struct zone *upper_zone = &pgdat->node_zones[j];
5645 managed_pages += zone_managed_pages(upper_zone);
5648 zone->lowmem_reserve[j] = 0;
5650 zone->lowmem_reserve[j] = managed_pages / ratio;
5655 /* update totalreserve_pages */
5656 calculate_totalreserve_pages();
5659 static void __setup_per_zone_wmarks(void)
5661 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5662 unsigned long lowmem_pages = 0;
5664 unsigned long flags;
5666 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5667 for_each_zone(zone) {
5668 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5669 lowmem_pages += zone_managed_pages(zone);
5672 for_each_zone(zone) {
5675 spin_lock_irqsave(&zone->lock, flags);
5676 tmp = (u64)pages_min * zone_managed_pages(zone);
5677 do_div(tmp, lowmem_pages);
5678 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5680 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5681 * need highmem and movable zones pages, so cap pages_min
5682 * to a small value here.
5684 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5685 * deltas control async page reclaim, and so should
5686 * not be capped for highmem and movable zones.
5688 unsigned long min_pages;
5690 min_pages = zone_managed_pages(zone) / 1024;
5691 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5692 zone->_watermark[WMARK_MIN] = min_pages;
5695 * If it's a lowmem zone, reserve a number of pages
5696 * proportionate to the zone's size.
5698 zone->_watermark[WMARK_MIN] = tmp;
5702 * Set the kswapd watermarks distance according to the
5703 * scale factor in proportion to available memory, but
5704 * ensure a minimum size on small systems.
5706 tmp = max_t(u64, tmp >> 2,
5707 mult_frac(zone_managed_pages(zone),
5708 watermark_scale_factor, 10000));
5710 zone->watermark_boost = 0;
5711 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
5712 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5713 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5715 spin_unlock_irqrestore(&zone->lock, flags);
5718 /* update totalreserve_pages */
5719 calculate_totalreserve_pages();
5723 * setup_per_zone_wmarks - called when min_free_kbytes changes
5724 * or when memory is hot-{added|removed}
5726 * Ensures that the watermark[min,low,high] values for each zone are set
5727 * correctly with respect to min_free_kbytes.
5729 void setup_per_zone_wmarks(void)
5732 static DEFINE_SPINLOCK(lock);
5735 __setup_per_zone_wmarks();
5739 * The watermark size have changed so update the pcpu batch
5740 * and high limits or the limits may be inappropriate.
5743 zone_pcp_update(zone, 0);
5747 * Initialise min_free_kbytes.
5749 * For small machines we want it small (128k min). For large machines
5750 * we want it large (256MB max). But it is not linear, because network
5751 * bandwidth does not increase linearly with machine size. We use
5753 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5754 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5770 void calculate_min_free_kbytes(void)
5772 unsigned long lowmem_kbytes;
5773 int new_min_free_kbytes;
5775 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5776 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5778 if (new_min_free_kbytes > user_min_free_kbytes)
5779 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
5781 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5782 new_min_free_kbytes, user_min_free_kbytes);
5786 int __meminit init_per_zone_wmark_min(void)
5788 calculate_min_free_kbytes();
5789 setup_per_zone_wmarks();
5790 refresh_zone_stat_thresholds();
5791 setup_per_zone_lowmem_reserve();
5794 setup_min_unmapped_ratio();
5795 setup_min_slab_ratio();
5798 khugepaged_min_free_kbytes_update();
5802 postcore_initcall(init_per_zone_wmark_min)
5805 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5806 * that we can call two helper functions whenever min_free_kbytes
5809 static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5810 void *buffer, size_t *length, loff_t *ppos)
5814 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5819 user_min_free_kbytes = min_free_kbytes;
5820 setup_per_zone_wmarks();
5825 static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
5826 void *buffer, size_t *length, loff_t *ppos)
5830 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5835 setup_per_zone_wmarks();
5841 static void setup_min_unmapped_ratio(void)
5846 for_each_online_pgdat(pgdat)
5847 pgdat->min_unmapped_pages = 0;
5850 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
5851 sysctl_min_unmapped_ratio) / 100;
5855 static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5856 void *buffer, size_t *length, loff_t *ppos)
5860 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5864 setup_min_unmapped_ratio();
5869 static void setup_min_slab_ratio(void)
5874 for_each_online_pgdat(pgdat)
5875 pgdat->min_slab_pages = 0;
5878 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
5879 sysctl_min_slab_ratio) / 100;
5882 static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5883 void *buffer, size_t *length, loff_t *ppos)
5887 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5891 setup_min_slab_ratio();
5898 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5899 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5900 * whenever sysctl_lowmem_reserve_ratio changes.
5902 * The reserve ratio obviously has absolutely no relation with the
5903 * minimum watermarks. The lowmem reserve ratio can only make sense
5904 * if in function of the boot time zone sizes.
5906 static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table,
5907 int write, void *buffer, size_t *length, loff_t *ppos)
5911 proc_dointvec_minmax(table, write, buffer, length, ppos);
5913 for (i = 0; i < MAX_NR_ZONES; i++) {
5914 if (sysctl_lowmem_reserve_ratio[i] < 1)
5915 sysctl_lowmem_reserve_ratio[i] = 0;
5918 setup_per_zone_lowmem_reserve();
5923 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
5924 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5925 * pagelist can have before it gets flushed back to buddy allocator.
5927 static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
5928 int write, void *buffer, size_t *length, loff_t *ppos)
5931 int old_percpu_pagelist_high_fraction;
5934 mutex_lock(&pcp_batch_high_lock);
5935 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
5937 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5938 if (!write || ret < 0)
5941 /* Sanity checking to avoid pcp imbalance */
5942 if (percpu_pagelist_high_fraction &&
5943 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
5944 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
5950 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
5953 for_each_populated_zone(zone)
5954 zone_set_pageset_high_and_batch(zone, 0);
5956 mutex_unlock(&pcp_batch_high_lock);
5960 static struct ctl_table page_alloc_sysctl_table[] = {
5962 .procname = "min_free_kbytes",
5963 .data = &min_free_kbytes,
5964 .maxlen = sizeof(min_free_kbytes),
5966 .proc_handler = min_free_kbytes_sysctl_handler,
5967 .extra1 = SYSCTL_ZERO,
5970 .procname = "watermark_boost_factor",
5971 .data = &watermark_boost_factor,
5972 .maxlen = sizeof(watermark_boost_factor),
5974 .proc_handler = proc_dointvec_minmax,
5975 .extra1 = SYSCTL_ZERO,
5978 .procname = "watermark_scale_factor",
5979 .data = &watermark_scale_factor,
5980 .maxlen = sizeof(watermark_scale_factor),
5982 .proc_handler = watermark_scale_factor_sysctl_handler,
5983 .extra1 = SYSCTL_ONE,
5984 .extra2 = SYSCTL_THREE_THOUSAND,
5987 .procname = "percpu_pagelist_high_fraction",
5988 .data = &percpu_pagelist_high_fraction,
5989 .maxlen = sizeof(percpu_pagelist_high_fraction),
5991 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
5992 .extra1 = SYSCTL_ZERO,
5995 .procname = "lowmem_reserve_ratio",
5996 .data = &sysctl_lowmem_reserve_ratio,
5997 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
5999 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6003 .procname = "numa_zonelist_order",
6004 .data = &numa_zonelist_order,
6005 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6007 .proc_handler = numa_zonelist_order_handler,
6010 .procname = "min_unmapped_ratio",
6011 .data = &sysctl_min_unmapped_ratio,
6012 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6014 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6015 .extra1 = SYSCTL_ZERO,
6016 .extra2 = SYSCTL_ONE_HUNDRED,
6019 .procname = "min_slab_ratio",
6020 .data = &sysctl_min_slab_ratio,
6021 .maxlen = sizeof(sysctl_min_slab_ratio),
6023 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6024 .extra1 = SYSCTL_ZERO,
6025 .extra2 = SYSCTL_ONE_HUNDRED,
6031 void __init page_alloc_sysctl_init(void)
6033 register_sysctl_init("vm", page_alloc_sysctl_table);
6036 #ifdef CONFIG_CONTIG_ALLOC
6037 /* Usage: See admin-guide/dynamic-debug-howto.rst */
6038 static void alloc_contig_dump_pages(struct list_head *page_list)
6040 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6042 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6046 list_for_each_entry(page, page_list, lru)
6047 dump_page(page, "migration failure");
6051 /* [start, end) must belong to a single zone. */
6052 int __alloc_contig_migrate_range(struct compact_control *cc,
6053 unsigned long start, unsigned long end)
6055 /* This function is based on compact_zone() from compaction.c. */
6056 unsigned int nr_reclaimed;
6057 unsigned long pfn = start;
6058 unsigned int tries = 0;
6060 struct migration_target_control mtc = {
6061 .nid = zone_to_nid(cc->zone),
6062 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6065 lru_cache_disable();
6067 while (pfn < end || !list_empty(&cc->migratepages)) {
6068 if (fatal_signal_pending(current)) {
6073 if (list_empty(&cc->migratepages)) {
6074 cc->nr_migratepages = 0;
6075 ret = isolate_migratepages_range(cc, pfn, end);
6076 if (ret && ret != -EAGAIN)
6078 pfn = cc->migrate_pfn;
6080 } else if (++tries == 5) {
6085 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6087 cc->nr_migratepages -= nr_reclaimed;
6089 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6090 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6093 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6094 * to retry again over this error, so do the same here.
6102 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6103 alloc_contig_dump_pages(&cc->migratepages);
6104 putback_movable_pages(&cc->migratepages);
6111 * alloc_contig_range() -- tries to allocate given range of pages
6112 * @start: start PFN to allocate
6113 * @end: one-past-the-last PFN to allocate
6114 * @migratetype: migratetype of the underlying pageblocks (either
6115 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6116 * in range must have the same migratetype and it must
6117 * be either of the two.
6118 * @gfp_mask: GFP mask to use during compaction
6120 * The PFN range does not have to be pageblock aligned. The PFN range must
6121 * belong to a single zone.
6123 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6124 * pageblocks in the range. Once isolated, the pageblocks should not
6125 * be modified by others.
6127 * Return: zero on success or negative error code. On success all
6128 * pages which PFN is in [start, end) are allocated for the caller and
6129 * need to be freed with free_contig_range().
6131 int alloc_contig_range(unsigned long start, unsigned long end,
6132 unsigned migratetype, gfp_t gfp_mask)
6134 unsigned long outer_start, outer_end;
6138 struct compact_control cc = {
6139 .nr_migratepages = 0,
6141 .zone = page_zone(pfn_to_page(start)),
6142 .mode = MIGRATE_SYNC,
6143 .ignore_skip_hint = true,
6144 .no_set_skip_hint = true,
6145 .gfp_mask = current_gfp_context(gfp_mask),
6146 .alloc_contig = true,
6148 INIT_LIST_HEAD(&cc.migratepages);
6151 * What we do here is we mark all pageblocks in range as
6152 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6153 * have different sizes, and due to the way page allocator
6154 * work, start_isolate_page_range() has special handlings for this.
6156 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6157 * migrate the pages from an unaligned range (ie. pages that
6158 * we are interested in). This will put all the pages in
6159 * range back to page allocator as MIGRATE_ISOLATE.
6161 * When this is done, we take the pages in range from page
6162 * allocator removing them from the buddy system. This way
6163 * page allocator will never consider using them.
6165 * This lets us mark the pageblocks back as
6166 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6167 * aligned range but not in the unaligned, original range are
6168 * put back to page allocator so that buddy can use them.
6171 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6175 drain_all_pages(cc.zone);
6178 * In case of -EBUSY, we'd like to know which page causes problem.
6179 * So, just fall through. test_pages_isolated() has a tracepoint
6180 * which will report the busy page.
6182 * It is possible that busy pages could become available before
6183 * the call to test_pages_isolated, and the range will actually be
6184 * allocated. So, if we fall through be sure to clear ret so that
6185 * -EBUSY is not accidentally used or returned to caller.
6187 ret = __alloc_contig_migrate_range(&cc, start, end);
6188 if (ret && ret != -EBUSY)
6193 * Pages from [start, end) are within a pageblock_nr_pages
6194 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6195 * more, all pages in [start, end) are free in page allocator.
6196 * What we are going to do is to allocate all pages from
6197 * [start, end) (that is remove them from page allocator).
6199 * The only problem is that pages at the beginning and at the
6200 * end of interesting range may be not aligned with pages that
6201 * page allocator holds, ie. they can be part of higher order
6202 * pages. Because of this, we reserve the bigger range and
6203 * once this is done free the pages we are not interested in.
6205 * We don't have to hold zone->lock here because the pages are
6206 * isolated thus they won't get removed from buddy.
6210 outer_start = start;
6211 while (!PageBuddy(pfn_to_page(outer_start))) {
6212 if (++order > MAX_ORDER) {
6213 outer_start = start;
6216 outer_start &= ~0UL << order;
6219 if (outer_start != start) {
6220 order = buddy_order(pfn_to_page(outer_start));
6223 * outer_start page could be small order buddy page and
6224 * it doesn't include start page. Adjust outer_start
6225 * in this case to report failed page properly
6226 * on tracepoint in test_pages_isolated()
6228 if (outer_start + (1UL << order) <= start)
6229 outer_start = start;
6232 /* Make sure the range is really isolated. */
6233 if (test_pages_isolated(outer_start, end, 0)) {
6238 /* Grab isolated pages from freelists. */
6239 outer_end = isolate_freepages_range(&cc, outer_start, end);
6245 /* Free head and tail (if any) */
6246 if (start != outer_start)
6247 free_contig_range(outer_start, start - outer_start);
6248 if (end != outer_end)
6249 free_contig_range(end, outer_end - end);
6252 undo_isolate_page_range(start, end, migratetype);
6255 EXPORT_SYMBOL(alloc_contig_range);
6257 static int __alloc_contig_pages(unsigned long start_pfn,
6258 unsigned long nr_pages, gfp_t gfp_mask)
6260 unsigned long end_pfn = start_pfn + nr_pages;
6262 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
6266 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6267 unsigned long nr_pages)
6269 unsigned long i, end_pfn = start_pfn + nr_pages;
6272 for (i = start_pfn; i < end_pfn; i++) {
6273 page = pfn_to_online_page(i);
6277 if (page_zone(page) != z)
6280 if (PageReserved(page))
6289 static bool zone_spans_last_pfn(const struct zone *zone,
6290 unsigned long start_pfn, unsigned long nr_pages)
6292 unsigned long last_pfn = start_pfn + nr_pages - 1;
6294 return zone_spans_pfn(zone, last_pfn);
6298 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6299 * @nr_pages: Number of contiguous pages to allocate
6300 * @gfp_mask: GFP mask to limit search and used during compaction
6302 * @nodemask: Mask for other possible nodes
6304 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6305 * on an applicable zonelist to find a contiguous pfn range which can then be
6306 * tried for allocation with alloc_contig_range(). This routine is intended
6307 * for allocation requests which can not be fulfilled with the buddy allocator.
6309 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6310 * power of two, then allocated range is also guaranteed to be aligned to same
6311 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6313 * Allocated pages can be freed with free_contig_range() or by manually calling
6314 * __free_page() on each allocated page.
6316 * Return: pointer to contiguous pages on success, or NULL if not successful.
6318 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
6319 int nid, nodemask_t *nodemask)
6321 unsigned long ret, pfn, flags;
6322 struct zonelist *zonelist;
6326 zonelist = node_zonelist(nid, gfp_mask);
6327 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6328 gfp_zone(gfp_mask), nodemask) {
6329 spin_lock_irqsave(&zone->lock, flags);
6331 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6332 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6333 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6335 * We release the zone lock here because
6336 * alloc_contig_range() will also lock the zone
6337 * at some point. If there's an allocation
6338 * spinning on this lock, it may win the race
6339 * and cause alloc_contig_range() to fail...
6341 spin_unlock_irqrestore(&zone->lock, flags);
6342 ret = __alloc_contig_pages(pfn, nr_pages,
6345 return pfn_to_page(pfn);
6346 spin_lock_irqsave(&zone->lock, flags);
6350 spin_unlock_irqrestore(&zone->lock, flags);
6354 #endif /* CONFIG_CONTIG_ALLOC */
6356 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6358 unsigned long count = 0;
6360 for (; nr_pages--; pfn++) {
6361 struct page *page = pfn_to_page(pfn);
6363 count += page_count(page) != 1;
6366 WARN(count != 0, "%lu pages are still in use!\n", count);
6368 EXPORT_SYMBOL(free_contig_range);
6371 * Effectively disable pcplists for the zone by setting the high limit to 0
6372 * and draining all cpus. A concurrent page freeing on another CPU that's about
6373 * to put the page on pcplist will either finish before the drain and the page
6374 * will be drained, or observe the new high limit and skip the pcplist.
6376 * Must be paired with a call to zone_pcp_enable().
6378 void zone_pcp_disable(struct zone *zone)
6380 mutex_lock(&pcp_batch_high_lock);
6381 __zone_set_pageset_high_and_batch(zone, 0, 1);
6382 __drain_all_pages(zone, true);
6385 void zone_pcp_enable(struct zone *zone)
6387 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
6388 mutex_unlock(&pcp_batch_high_lock);
6391 void zone_pcp_reset(struct zone *zone)
6394 struct per_cpu_zonestat *pzstats;
6396 if (zone->per_cpu_pageset != &boot_pageset) {
6397 for_each_online_cpu(cpu) {
6398 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6399 drain_zonestat(zone, pzstats);
6401 free_percpu(zone->per_cpu_pageset);
6402 zone->per_cpu_pageset = &boot_pageset;
6403 if (zone->per_cpu_zonestats != &boot_zonestats) {
6404 free_percpu(zone->per_cpu_zonestats);
6405 zone->per_cpu_zonestats = &boot_zonestats;
6410 #ifdef CONFIG_MEMORY_HOTREMOVE
6412 * All pages in the range must be in a single zone, must not contain holes,
6413 * must span full sections, and must be isolated before calling this function.
6415 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6417 unsigned long pfn = start_pfn;
6421 unsigned long flags;
6423 offline_mem_sections(pfn, end_pfn);
6424 zone = page_zone(pfn_to_page(pfn));
6425 spin_lock_irqsave(&zone->lock, flags);
6426 while (pfn < end_pfn) {
6427 page = pfn_to_page(pfn);
6429 * The HWPoisoned page may be not in buddy system, and
6430 * page_count() is not 0.
6432 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6437 * At this point all remaining PageOffline() pages have a
6438 * reference count of 0 and can simply be skipped.
6440 if (PageOffline(page)) {
6441 BUG_ON(page_count(page));
6442 BUG_ON(PageBuddy(page));
6447 BUG_ON(page_count(page));
6448 BUG_ON(!PageBuddy(page));
6449 order = buddy_order(page);
6450 del_page_from_free_list(page, zone, order);
6451 pfn += (1 << order);
6453 spin_unlock_irqrestore(&zone->lock, flags);
6458 * This function returns a stable result only if called under zone lock.
6460 bool is_free_buddy_page(struct page *page)
6462 unsigned long pfn = page_to_pfn(page);
6465 for (order = 0; order <= MAX_ORDER; order++) {
6466 struct page *page_head = page - (pfn & ((1 << order) - 1));
6468 if (PageBuddy(page_head) &&
6469 buddy_order_unsafe(page_head) >= order)
6473 return order <= MAX_ORDER;
6475 EXPORT_SYMBOL(is_free_buddy_page);
6477 #ifdef CONFIG_MEMORY_FAILURE
6479 * Break down a higher-order page in sub-pages, and keep our target out of
6482 static void break_down_buddy_pages(struct zone *zone, struct page *page,
6483 struct page *target, int low, int high,
6486 unsigned long size = 1 << high;
6487 struct page *current_buddy, *next_page;
6489 while (high > low) {
6493 if (target >= &page[size]) {
6494 next_page = page + size;
6495 current_buddy = page;
6498 current_buddy = page + size;
6502 if (set_page_guard(zone, current_buddy, high, migratetype))
6505 if (current_buddy != target) {
6506 add_to_free_list(current_buddy, zone, high, migratetype);
6507 set_buddy_order(current_buddy, high);
6513 * Take a page that will be marked as poisoned off the buddy allocator.
6515 bool take_page_off_buddy(struct page *page)
6517 struct zone *zone = page_zone(page);
6518 unsigned long pfn = page_to_pfn(page);
6519 unsigned long flags;
6523 spin_lock_irqsave(&zone->lock, flags);
6524 for (order = 0; order <= MAX_ORDER; order++) {
6525 struct page *page_head = page - (pfn & ((1 << order) - 1));
6526 int page_order = buddy_order(page_head);
6528 if (PageBuddy(page_head) && page_order >= order) {
6529 unsigned long pfn_head = page_to_pfn(page_head);
6530 int migratetype = get_pfnblock_migratetype(page_head,
6533 del_page_from_free_list(page_head, zone, page_order);
6534 break_down_buddy_pages(zone, page_head, page, 0,
6535 page_order, migratetype);
6536 SetPageHWPoisonTakenOff(page);
6537 if (!is_migrate_isolate(migratetype))
6538 __mod_zone_freepage_state(zone, -1, migratetype);
6542 if (page_count(page_head) > 0)
6545 spin_unlock_irqrestore(&zone->lock, flags);
6550 * Cancel takeoff done by take_page_off_buddy().
6552 bool put_page_back_buddy(struct page *page)
6554 struct zone *zone = page_zone(page);
6555 unsigned long pfn = page_to_pfn(page);
6556 unsigned long flags;
6557 int migratetype = get_pfnblock_migratetype(page, pfn);
6560 spin_lock_irqsave(&zone->lock, flags);
6561 if (put_page_testzero(page)) {
6562 ClearPageHWPoisonTakenOff(page);
6563 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6564 if (TestClearPageHWPoison(page)) {
6568 spin_unlock_irqrestore(&zone->lock, flags);
6574 #ifdef CONFIG_ZONE_DMA
6575 bool has_managed_dma(void)
6577 struct pglist_data *pgdat;
6579 for_each_online_pgdat(pgdat) {
6580 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6582 if (managed_zone(zone))
6587 #endif /* CONFIG_ZONE_DMA */
6589 #ifdef CONFIG_UNACCEPTED_MEMORY
6591 /* Counts number of zones with unaccepted pages. */
6592 static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6594 static bool lazy_accept = true;
6596 static int __init accept_memory_parse(char *p)
6598 if (!strcmp(p, "lazy")) {
6601 } else if (!strcmp(p, "eager")) {
6602 lazy_accept = false;
6608 early_param("accept_memory", accept_memory_parse);
6610 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6612 phys_addr_t start = page_to_phys(page);
6613 phys_addr_t end = start + (PAGE_SIZE << order);
6615 return range_contains_unaccepted_memory(start, end);
6618 static void accept_page(struct page *page, unsigned int order)
6620 phys_addr_t start = page_to_phys(page);
6622 accept_memory(start, start + (PAGE_SIZE << order));
6625 static bool try_to_accept_memory_one(struct zone *zone)
6627 unsigned long flags;
6631 if (list_empty(&zone->unaccepted_pages))
6634 spin_lock_irqsave(&zone->lock, flags);
6635 page = list_first_entry_or_null(&zone->unaccepted_pages,
6638 spin_unlock_irqrestore(&zone->lock, flags);
6642 list_del(&page->lru);
6643 last = list_empty(&zone->unaccepted_pages);
6645 __mod_zone_freepage_state(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6646 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6647 spin_unlock_irqrestore(&zone->lock, flags);
6649 accept_page(page, MAX_ORDER);
6651 __free_pages_ok(page, MAX_ORDER, FPI_TO_TAIL);
6654 static_branch_dec(&zones_with_unaccepted_pages);
6659 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6664 /* How much to accept to get to high watermark? */
6665 to_accept = high_wmark_pages(zone) -
6666 (zone_page_state(zone, NR_FREE_PAGES) -
6667 __zone_watermark_unusable_free(zone, order, 0));
6669 /* Accept at least one page */
6671 if (!try_to_accept_memory_one(zone))
6674 to_accept -= MAX_ORDER_NR_PAGES;
6675 } while (to_accept > 0);
6680 static inline bool has_unaccepted_memory(void)
6682 return static_branch_unlikely(&zones_with_unaccepted_pages);
6685 static bool __free_unaccepted(struct page *page)
6687 struct zone *zone = page_zone(page);
6688 unsigned long flags;
6694 spin_lock_irqsave(&zone->lock, flags);
6695 first = list_empty(&zone->unaccepted_pages);
6696 list_add_tail(&page->lru, &zone->unaccepted_pages);
6697 __mod_zone_freepage_state(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6698 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
6699 spin_unlock_irqrestore(&zone->lock, flags);
6702 static_branch_inc(&zones_with_unaccepted_pages);
6709 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6714 static void accept_page(struct page *page, unsigned int order)
6718 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6723 static inline bool has_unaccepted_memory(void)
6728 static bool __free_unaccepted(struct page *page)
6734 #endif /* CONFIG_UNACCEPTED_MEMORY */