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 <linux/cacheinfo.h>
56 #include <asm/div64.h>
59 #include "page_reporting.h"
61 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
62 typedef int __bitwise fpi_t;
64 /* No special request */
65 #define FPI_NONE ((__force fpi_t)0)
68 * Skip free page reporting notification for the (possibly merged) page.
69 * This does not hinder free page reporting from grabbing the page,
70 * reporting it and marking it "reported" - it only skips notifying
71 * the free page reporting infrastructure about a newly freed page. For
72 * example, used when temporarily pulling a page from a freelist and
73 * putting it back unmodified.
75 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
78 * Place the (possibly merged) page to the tail of the freelist. Will ignore
79 * page shuffling (relevant code - e.g., memory onlining - is expected to
80 * shuffle the whole zone).
82 * Note: No code should rely on this flag for correctness - it's purely
83 * to allow for optimizations when handing back either fresh pages
84 * (memory onlining) or untouched pages (page isolation, free page
87 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
89 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
90 static DEFINE_MUTEX(pcp_batch_high_lock);
91 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
93 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
95 * On SMP, spin_trylock is sufficient protection.
96 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
98 #define pcp_trylock_prepare(flags) do { } while (0)
99 #define pcp_trylock_finish(flag) do { } while (0)
102 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
103 #define pcp_trylock_prepare(flags) local_irq_save(flags)
104 #define pcp_trylock_finish(flags) local_irq_restore(flags)
108 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
109 * a migration causing the wrong PCP to be locked and remote memory being
110 * potentially allocated, pin the task to the CPU for the lookup+lock.
111 * preempt_disable is used on !RT because it is faster than migrate_disable.
112 * migrate_disable is used on RT because otherwise RT spinlock usage is
113 * interfered with and a high priority task cannot preempt the allocator.
115 #ifndef CONFIG_PREEMPT_RT
116 #define pcpu_task_pin() preempt_disable()
117 #define pcpu_task_unpin() preempt_enable()
119 #define pcpu_task_pin() migrate_disable()
120 #define pcpu_task_unpin() migrate_enable()
124 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
125 * Return value should be used with equivalent unlock helper.
127 #define pcpu_spin_lock(type, member, ptr) \
131 _ret = this_cpu_ptr(ptr); \
132 spin_lock(&_ret->member); \
136 #define pcpu_spin_trylock(type, member, ptr) \
140 _ret = this_cpu_ptr(ptr); \
141 if (!spin_trylock(&_ret->member)) { \
148 #define pcpu_spin_unlock(member, ptr) \
150 spin_unlock(&ptr->member); \
154 /* struct per_cpu_pages specific helpers. */
155 #define pcp_spin_lock(ptr) \
156 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
158 #define pcp_spin_trylock(ptr) \
159 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
161 #define pcp_spin_unlock(ptr) \
162 pcpu_spin_unlock(lock, ptr)
164 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
165 DEFINE_PER_CPU(int, numa_node);
166 EXPORT_PER_CPU_SYMBOL(numa_node);
169 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
171 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
173 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
174 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
175 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
176 * defined in <linux/topology.h>.
178 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
179 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
182 static DEFINE_MUTEX(pcpu_drain_mutex);
184 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
185 volatile unsigned long latent_entropy __latent_entropy;
186 EXPORT_SYMBOL(latent_entropy);
190 * Array of node states.
192 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
193 [N_POSSIBLE] = NODE_MASK_ALL,
194 [N_ONLINE] = { { [0] = 1UL } },
196 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
197 #ifdef CONFIG_HIGHMEM
198 [N_HIGH_MEMORY] = { { [0] = 1UL } },
200 [N_MEMORY] = { { [0] = 1UL } },
201 [N_CPU] = { { [0] = 1UL } },
204 EXPORT_SYMBOL(node_states);
206 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
209 * A cached value of the page's pageblock's migratetype, used when the page is
210 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
211 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
212 * Also the migratetype set in the page does not necessarily match the pcplist
213 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
214 * other index - this ensures that it will be put on the correct CMA freelist.
216 static inline int get_pcppage_migratetype(struct page *page)
221 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
223 page->index = migratetype;
226 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
227 unsigned int pageblock_order __read_mostly;
230 static void __free_pages_ok(struct page *page, unsigned int order,
234 * results with 256, 32 in the lowmem_reserve sysctl:
235 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
236 * 1G machine -> (16M dma, 784M normal, 224M high)
237 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
238 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
239 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
241 * TBD: should special case ZONE_DMA32 machines here - in those we normally
242 * don't need any ZONE_NORMAL reservation
244 static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
245 #ifdef CONFIG_ZONE_DMA
248 #ifdef CONFIG_ZONE_DMA32
252 #ifdef CONFIG_HIGHMEM
258 char * const zone_names[MAX_NR_ZONES] = {
259 #ifdef CONFIG_ZONE_DMA
262 #ifdef CONFIG_ZONE_DMA32
266 #ifdef CONFIG_HIGHMEM
270 #ifdef CONFIG_ZONE_DEVICE
275 const char * const migratetype_names[MIGRATE_TYPES] = {
283 #ifdef CONFIG_MEMORY_ISOLATION
288 int min_free_kbytes = 1024;
289 int user_min_free_kbytes = -1;
290 static int watermark_boost_factor __read_mostly = 15000;
291 static int watermark_scale_factor = 10;
293 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
295 EXPORT_SYMBOL(movable_zone);
298 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
299 unsigned int nr_online_nodes __read_mostly = 1;
300 EXPORT_SYMBOL(nr_node_ids);
301 EXPORT_SYMBOL(nr_online_nodes);
304 static bool page_contains_unaccepted(struct page *page, unsigned int order);
305 static void accept_page(struct page *page, unsigned int order);
306 static bool try_to_accept_memory(struct zone *zone, unsigned int order);
307 static inline bool has_unaccepted_memory(void);
308 static bool __free_unaccepted(struct page *page);
310 int page_group_by_mobility_disabled __read_mostly;
312 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
314 * During boot we initialize deferred pages on-demand, as needed, but once
315 * page_alloc_init_late() has finished, the deferred pages are all initialized,
316 * and we can permanently disable that path.
318 DEFINE_STATIC_KEY_TRUE(deferred_pages);
320 static inline bool deferred_pages_enabled(void)
322 return static_branch_unlikely(&deferred_pages);
326 * deferred_grow_zone() is __init, but it is called from
327 * get_page_from_freelist() during early boot until deferred_pages permanently
328 * disables this call. This is why we have refdata wrapper to avoid warning,
329 * and to ensure that the function body gets unloaded.
332 _deferred_grow_zone(struct zone *zone, unsigned int order)
334 return deferred_grow_zone(zone, order);
337 static inline bool deferred_pages_enabled(void)
341 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
343 /* Return a pointer to the bitmap storing bits affecting a block of pages */
344 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
347 #ifdef CONFIG_SPARSEMEM
348 return section_to_usemap(__pfn_to_section(pfn));
350 return page_zone(page)->pageblock_flags;
351 #endif /* CONFIG_SPARSEMEM */
354 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
356 #ifdef CONFIG_SPARSEMEM
357 pfn &= (PAGES_PER_SECTION-1);
359 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
360 #endif /* CONFIG_SPARSEMEM */
361 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
365 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
366 * @page: The page within the block of interest
367 * @pfn: The target page frame number
368 * @mask: mask of bits that the caller is interested in
370 * Return: pageblock_bits flags
372 unsigned long get_pfnblock_flags_mask(const struct page *page,
373 unsigned long pfn, unsigned long mask)
375 unsigned long *bitmap;
376 unsigned long bitidx, word_bitidx;
379 bitmap = get_pageblock_bitmap(page, pfn);
380 bitidx = pfn_to_bitidx(page, pfn);
381 word_bitidx = bitidx / BITS_PER_LONG;
382 bitidx &= (BITS_PER_LONG-1);
384 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
385 * a consistent read of the memory array, so that results, even though
386 * racy, are not corrupted.
388 word = READ_ONCE(bitmap[word_bitidx]);
389 return (word >> bitidx) & mask;
392 static __always_inline int get_pfnblock_migratetype(const struct page *page,
395 return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
399 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
400 * @page: The page within the block of interest
401 * @flags: The flags to set
402 * @pfn: The target page frame number
403 * @mask: mask of bits that the caller is interested in
405 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
409 unsigned long *bitmap;
410 unsigned long bitidx, word_bitidx;
413 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
414 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
416 bitmap = get_pageblock_bitmap(page, pfn);
417 bitidx = pfn_to_bitidx(page, pfn);
418 word_bitidx = bitidx / BITS_PER_LONG;
419 bitidx &= (BITS_PER_LONG-1);
421 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
426 word = READ_ONCE(bitmap[word_bitidx]);
428 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
431 void set_pageblock_migratetype(struct page *page, int migratetype)
433 if (unlikely(page_group_by_mobility_disabled &&
434 migratetype < MIGRATE_PCPTYPES))
435 migratetype = MIGRATE_UNMOVABLE;
437 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
438 page_to_pfn(page), MIGRATETYPE_MASK);
441 #ifdef CONFIG_DEBUG_VM
442 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
446 unsigned long pfn = page_to_pfn(page);
447 unsigned long sp, start_pfn;
450 seq = zone_span_seqbegin(zone);
451 start_pfn = zone->zone_start_pfn;
452 sp = zone->spanned_pages;
453 ret = !zone_spans_pfn(zone, pfn);
454 } while (zone_span_seqretry(zone, seq));
457 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
458 pfn, zone_to_nid(zone), zone->name,
459 start_pfn, start_pfn + sp);
465 * Temporary debugging check for pages not lying within a given zone.
467 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
469 if (page_outside_zone_boundaries(zone, page))
471 if (zone != page_zone(page))
477 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
483 static void bad_page(struct page *page, const char *reason)
485 static unsigned long resume;
486 static unsigned long nr_shown;
487 static unsigned long nr_unshown;
490 * Allow a burst of 60 reports, then keep quiet for that minute;
491 * or allow a steady drip of one report per second.
493 if (nr_shown == 60) {
494 if (time_before(jiffies, resume)) {
500 "BUG: Bad page state: %lu messages suppressed\n",
507 resume = jiffies + 60 * HZ;
509 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
510 current->comm, page_to_pfn(page));
511 dump_page(page, reason);
516 /* Leave bad fields for debug, except PageBuddy could make trouble */
517 page_mapcount_reset(page); /* remove PageBuddy */
518 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
521 static inline unsigned int order_to_pindex(int migratetype, int order)
523 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
524 if (order > PAGE_ALLOC_COSTLY_ORDER) {
525 VM_BUG_ON(order != pageblock_order);
526 return NR_LOWORDER_PCP_LISTS;
529 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
532 return (MIGRATE_PCPTYPES * order) + migratetype;
535 static inline int pindex_to_order(unsigned int pindex)
537 int order = pindex / MIGRATE_PCPTYPES;
539 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
540 if (pindex == NR_LOWORDER_PCP_LISTS)
541 order = pageblock_order;
543 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
549 static inline bool pcp_allowed_order(unsigned int order)
551 if (order <= PAGE_ALLOC_COSTLY_ORDER)
553 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
554 if (order == pageblock_order)
560 static inline void free_the_page(struct page *page, unsigned int order)
562 if (pcp_allowed_order(order)) /* Via pcp? */
563 free_unref_page(page, order);
565 __free_pages_ok(page, order, FPI_NONE);
569 * Higher-order pages are called "compound pages". They are structured thusly:
571 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
573 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
574 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
576 * The first tail page's ->compound_order holds the order of allocation.
577 * This usage means that zero-order pages may not be compound.
580 void prep_compound_page(struct page *page, unsigned int order)
583 int nr_pages = 1 << order;
586 for (i = 1; i < nr_pages; i++)
587 prep_compound_tail(page, i);
589 prep_compound_head(page, order);
592 void destroy_large_folio(struct folio *folio)
594 if (folio_test_hugetlb(folio)) {
595 free_huge_folio(folio);
599 if (folio_test_large_rmappable(folio))
600 folio_undo_large_rmappable(folio);
602 mem_cgroup_uncharge(folio);
603 free_the_page(&folio->page, folio_order(folio));
606 static inline void set_buddy_order(struct page *page, unsigned int order)
608 set_page_private(page, order);
609 __SetPageBuddy(page);
612 #ifdef CONFIG_COMPACTION
613 static inline struct capture_control *task_capc(struct zone *zone)
615 struct capture_control *capc = current->capture_control;
617 return unlikely(capc) &&
618 !(current->flags & PF_KTHREAD) &&
620 capc->cc->zone == zone ? capc : NULL;
624 compaction_capture(struct capture_control *capc, struct page *page,
625 int order, int migratetype)
627 if (!capc || order != capc->cc->order)
630 /* Do not accidentally pollute CMA or isolated regions*/
631 if (is_migrate_cma(migratetype) ||
632 is_migrate_isolate(migratetype))
636 * Do not let lower order allocations pollute a movable pageblock.
637 * This might let an unmovable request use a reclaimable pageblock
638 * and vice-versa but no more than normal fallback logic which can
639 * have trouble finding a high-order free page.
641 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
649 static inline struct capture_control *task_capc(struct zone *zone)
655 compaction_capture(struct capture_control *capc, struct page *page,
656 int order, int migratetype)
660 #endif /* CONFIG_COMPACTION */
662 /* Used for pages not on another list */
663 static inline void add_to_free_list(struct page *page, struct zone *zone,
664 unsigned int order, int migratetype)
666 struct free_area *area = &zone->free_area[order];
668 list_add(&page->buddy_list, &area->free_list[migratetype]);
672 /* Used for pages not on another list */
673 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
674 unsigned int order, int migratetype)
676 struct free_area *area = &zone->free_area[order];
678 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
683 * Used for pages which are on another list. Move the pages to the tail
684 * of the list - so the moved pages won't immediately be considered for
685 * allocation again (e.g., optimization for memory onlining).
687 static inline void move_to_free_list(struct page *page, struct zone *zone,
688 unsigned int order, int migratetype)
690 struct free_area *area = &zone->free_area[order];
692 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
695 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
698 /* clear reported state and update reported page count */
699 if (page_reported(page))
700 __ClearPageReported(page);
702 list_del(&page->buddy_list);
703 __ClearPageBuddy(page);
704 set_page_private(page, 0);
705 zone->free_area[order].nr_free--;
708 static inline struct page *get_page_from_free_area(struct free_area *area,
711 return list_first_entry_or_null(&area->free_list[migratetype],
712 struct page, buddy_list);
716 * If this is not the largest possible page, check if the buddy
717 * of the next-highest order is free. If it is, it's possible
718 * that pages are being freed that will coalesce soon. In case,
719 * that is happening, add the free page to the tail of the list
720 * so it's less likely to be used soon and more likely to be merged
721 * as a higher order page
724 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
725 struct page *page, unsigned int order)
727 unsigned long higher_page_pfn;
728 struct page *higher_page;
730 if (order >= MAX_ORDER - 1)
733 higher_page_pfn = buddy_pfn & pfn;
734 higher_page = page + (higher_page_pfn - pfn);
736 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
741 * Freeing function for a buddy system allocator.
743 * The concept of a buddy system is to maintain direct-mapped table
744 * (containing bit values) for memory blocks of various "orders".
745 * The bottom level table contains the map for the smallest allocatable
746 * units of memory (here, pages), and each level above it describes
747 * pairs of units from the levels below, hence, "buddies".
748 * At a high level, all that happens here is marking the table entry
749 * at the bottom level available, and propagating the changes upward
750 * as necessary, plus some accounting needed to play nicely with other
751 * parts of the VM system.
752 * At each level, we keep a list of pages, which are heads of continuous
753 * free pages of length of (1 << order) and marked with PageBuddy.
754 * Page's order is recorded in page_private(page) field.
755 * So when we are allocating or freeing one, we can derive the state of the
756 * other. That is, if we allocate a small block, and both were
757 * free, the remainder of the region must be split into blocks.
758 * If a block is freed, and its buddy is also free, then this
759 * triggers coalescing into a block of larger size.
764 static inline void __free_one_page(struct page *page,
766 struct zone *zone, unsigned int order,
767 int migratetype, fpi_t fpi_flags)
769 struct capture_control *capc = task_capc(zone);
770 unsigned long buddy_pfn = 0;
771 unsigned long combined_pfn;
775 VM_BUG_ON(!zone_is_initialized(zone));
776 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
778 VM_BUG_ON(migratetype == -1);
779 if (likely(!is_migrate_isolate(migratetype)))
780 __mod_zone_freepage_state(zone, 1 << order, migratetype);
782 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
783 VM_BUG_ON_PAGE(bad_range(zone, page), page);
785 while (order < MAX_ORDER) {
786 if (compaction_capture(capc, page, order, migratetype)) {
787 __mod_zone_freepage_state(zone, -(1 << order),
792 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
796 if (unlikely(order >= pageblock_order)) {
798 * We want to prevent merge between freepages on pageblock
799 * without fallbacks and normal pageblock. Without this,
800 * pageblock isolation could cause incorrect freepage or CMA
801 * accounting or HIGHATOMIC accounting.
803 int buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
805 if (migratetype != buddy_mt
806 && (!migratetype_is_mergeable(migratetype) ||
807 !migratetype_is_mergeable(buddy_mt)))
812 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
813 * merge with it and move up one order.
815 if (page_is_guard(buddy))
816 clear_page_guard(zone, buddy, order, migratetype);
818 del_page_from_free_list(buddy, zone, order);
819 combined_pfn = buddy_pfn & pfn;
820 page = page + (combined_pfn - pfn);
826 set_buddy_order(page, order);
828 if (fpi_flags & FPI_TO_TAIL)
830 else if (is_shuffle_order(order))
831 to_tail = shuffle_pick_tail();
833 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
836 add_to_free_list_tail(page, zone, order, migratetype);
838 add_to_free_list(page, zone, order, migratetype);
840 /* Notify page reporting subsystem of freed page */
841 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
842 page_reporting_notify_free(order);
846 * split_free_page() -- split a free page at split_pfn_offset
847 * @free_page: the original free page
848 * @order: the order of the page
849 * @split_pfn_offset: split offset within the page
851 * Return -ENOENT if the free page is changed, otherwise 0
853 * It is used when the free page crosses two pageblocks with different migratetypes
854 * at split_pfn_offset within the page. The split free page will be put into
855 * separate migratetype lists afterwards. Otherwise, the function achieves
858 int split_free_page(struct page *free_page,
859 unsigned int order, unsigned long split_pfn_offset)
861 struct zone *zone = page_zone(free_page);
862 unsigned long free_page_pfn = page_to_pfn(free_page);
869 if (split_pfn_offset == 0)
872 spin_lock_irqsave(&zone->lock, flags);
874 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
879 mt = get_pfnblock_migratetype(free_page, free_page_pfn);
880 if (likely(!is_migrate_isolate(mt)))
881 __mod_zone_freepage_state(zone, -(1UL << order), mt);
883 del_page_from_free_list(free_page, zone, order);
884 for (pfn = free_page_pfn;
885 pfn < free_page_pfn + (1UL << order);) {
886 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
888 free_page_order = min_t(unsigned int,
889 pfn ? __ffs(pfn) : order,
890 __fls(split_pfn_offset));
891 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
893 pfn += 1UL << free_page_order;
894 split_pfn_offset -= (1UL << free_page_order);
895 /* we have done the first part, now switch to second part */
896 if (split_pfn_offset == 0)
897 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
900 spin_unlock_irqrestore(&zone->lock, flags);
904 * A bad page could be due to a number of fields. Instead of multiple branches,
905 * try and check multiple fields with one check. The caller must do a detailed
906 * check if necessary.
908 static inline bool page_expected_state(struct page *page,
909 unsigned long check_flags)
911 if (unlikely(atomic_read(&page->_mapcount) != -1))
914 if (unlikely((unsigned long)page->mapping |
915 page_ref_count(page) |
919 (page->flags & check_flags)))
925 static const char *page_bad_reason(struct page *page, unsigned long flags)
927 const char *bad_reason = NULL;
929 if (unlikely(atomic_read(&page->_mapcount) != -1))
930 bad_reason = "nonzero mapcount";
931 if (unlikely(page->mapping != NULL))
932 bad_reason = "non-NULL mapping";
933 if (unlikely(page_ref_count(page) != 0))
934 bad_reason = "nonzero _refcount";
935 if (unlikely(page->flags & flags)) {
936 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
937 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
939 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
942 if (unlikely(page->memcg_data))
943 bad_reason = "page still charged to cgroup";
948 static void free_page_is_bad_report(struct page *page)
951 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
954 static inline bool free_page_is_bad(struct page *page)
956 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
959 /* Something has gone sideways, find it */
960 free_page_is_bad_report(page);
964 static inline bool is_check_pages_enabled(void)
966 return static_branch_unlikely(&check_pages_enabled);
969 static int free_tail_page_prepare(struct page *head_page, struct page *page)
971 struct folio *folio = (struct folio *)head_page;
975 * We rely page->lru.next never has bit 0 set, unless the page
976 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
978 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
980 if (!is_check_pages_enabled()) {
984 switch (page - head_page) {
986 /* the first tail page: these may be in place of ->mapping */
987 if (unlikely(folio_entire_mapcount(folio))) {
988 bad_page(page, "nonzero entire_mapcount");
991 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
992 bad_page(page, "nonzero nr_pages_mapped");
995 if (unlikely(atomic_read(&folio->_pincount))) {
996 bad_page(page, "nonzero pincount");
1002 * the second tail page: ->mapping is
1003 * deferred_list.next -- ignore value.
1007 if (page->mapping != TAIL_MAPPING) {
1008 bad_page(page, "corrupted mapping in tail page");
1013 if (unlikely(!PageTail(page))) {
1014 bad_page(page, "PageTail not set");
1017 if (unlikely(compound_head(page) != head_page)) {
1018 bad_page(page, "compound_head not consistent");
1023 page->mapping = NULL;
1024 clear_compound_head(page);
1029 * Skip KASAN memory poisoning when either:
1031 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1032 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1033 * using page tags instead (see below).
1034 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1035 * that error detection is disabled for accesses via the page address.
1037 * Pages will have match-all tags in the following circumstances:
1039 * 1. Pages are being initialized for the first time, including during deferred
1040 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1041 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1042 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1043 * 3. The allocation was excluded from being checked due to sampling,
1044 * see the call to kasan_unpoison_pages.
1046 * Poisoning pages during deferred memory init will greatly lengthen the
1047 * process and cause problem in large memory systems as the deferred pages
1048 * initialization is done with interrupt disabled.
1050 * Assuming that there will be no reference to those newly initialized
1051 * pages before they are ever allocated, this should have no effect on
1052 * KASAN memory tracking as the poison will be properly inserted at page
1053 * allocation time. The only corner case is when pages are allocated by
1054 * on-demand allocation and then freed again before the deferred pages
1055 * initialization is done, but this is not likely to happen.
1057 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1059 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1060 return deferred_pages_enabled();
1062 return page_kasan_tag(page) == 0xff;
1065 static void kernel_init_pages(struct page *page, int numpages)
1069 /* s390's use of memset() could override KASAN redzones. */
1070 kasan_disable_current();
1071 for (i = 0; i < numpages; i++)
1072 clear_highpage_kasan_tagged(page + i);
1073 kasan_enable_current();
1076 static __always_inline bool free_pages_prepare(struct page *page,
1077 unsigned int order, fpi_t fpi_flags)
1080 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1081 bool init = want_init_on_free();
1082 bool compound = PageCompound(page);
1084 VM_BUG_ON_PAGE(PageTail(page), page);
1086 trace_mm_page_free(page, order);
1087 kmsan_free_page(page, order);
1089 if (memcg_kmem_online() && PageMemcgKmem(page))
1090 __memcg_kmem_uncharge_page(page, order);
1092 if (unlikely(PageHWPoison(page)) && !order) {
1093 /* Do not let hwpoison pages hit pcplists/buddy */
1094 reset_page_owner(page, order);
1095 page_table_check_free(page, order);
1099 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1102 * Check tail pages before head page information is cleared to
1103 * avoid checking PageCompound for order-0 pages.
1105 if (unlikely(order)) {
1109 page[1].flags &= ~PAGE_FLAGS_SECOND;
1110 for (i = 1; i < (1 << order); i++) {
1112 bad += free_tail_page_prepare(page, page + i);
1113 if (is_check_pages_enabled()) {
1114 if (free_page_is_bad(page + i)) {
1119 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1122 if (PageMappingFlags(page))
1123 page->mapping = NULL;
1124 if (is_check_pages_enabled()) {
1125 if (free_page_is_bad(page))
1131 page_cpupid_reset_last(page);
1132 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1133 reset_page_owner(page, order);
1134 page_table_check_free(page, order);
1136 if (!PageHighMem(page)) {
1137 debug_check_no_locks_freed(page_address(page),
1138 PAGE_SIZE << order);
1139 debug_check_no_obj_freed(page_address(page),
1140 PAGE_SIZE << order);
1143 kernel_poison_pages(page, 1 << order);
1146 * As memory initialization might be integrated into KASAN,
1147 * KASAN poisoning and memory initialization code must be
1148 * kept together to avoid discrepancies in behavior.
1150 * With hardware tag-based KASAN, memory tags must be set before the
1151 * page becomes unavailable via debug_pagealloc or arch_free_page.
1153 if (!skip_kasan_poison) {
1154 kasan_poison_pages(page, order, init);
1156 /* Memory is already initialized if KASAN did it internally. */
1157 if (kasan_has_integrated_init())
1161 kernel_init_pages(page, 1 << order);
1164 * arch_free_page() can make the page's contents inaccessible. s390
1165 * does this. So nothing which can access the page's contents should
1166 * happen after this.
1168 arch_free_page(page, order);
1170 debug_pagealloc_unmap_pages(page, 1 << order);
1176 * Frees a number of pages from the PCP lists
1177 * Assumes all pages on list are in same zone.
1178 * count is the number of pages to free.
1180 static void free_pcppages_bulk(struct zone *zone, int count,
1181 struct per_cpu_pages *pcp,
1184 unsigned long flags;
1186 bool isolated_pageblocks;
1190 * Ensure proper count is passed which otherwise would stuck in the
1191 * below while (list_empty(list)) loop.
1193 count = min(pcp->count, count);
1195 /* Ensure requested pindex is drained first. */
1196 pindex = pindex - 1;
1198 spin_lock_irqsave(&zone->lock, flags);
1199 isolated_pageblocks = has_isolate_pageblock(zone);
1202 struct list_head *list;
1205 /* Remove pages from lists in a round-robin fashion. */
1207 if (++pindex > NR_PCP_LISTS - 1)
1209 list = &pcp->lists[pindex];
1210 } while (list_empty(list));
1212 order = pindex_to_order(pindex);
1213 nr_pages = 1 << order;
1217 page = list_last_entry(list, struct page, pcp_list);
1218 mt = get_pcppage_migratetype(page);
1220 /* must delete to avoid corrupting pcp list */
1221 list_del(&page->pcp_list);
1223 pcp->count -= nr_pages;
1225 /* MIGRATE_ISOLATE page should not go to pcplists */
1226 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1227 /* Pageblock could have been isolated meanwhile */
1228 if (unlikely(isolated_pageblocks))
1229 mt = get_pageblock_migratetype(page);
1231 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1232 trace_mm_page_pcpu_drain(page, order, mt);
1233 } while (count > 0 && !list_empty(list));
1236 spin_unlock_irqrestore(&zone->lock, flags);
1239 static void free_one_page(struct zone *zone,
1240 struct page *page, unsigned long pfn,
1242 int migratetype, fpi_t fpi_flags)
1244 unsigned long flags;
1246 spin_lock_irqsave(&zone->lock, flags);
1247 if (unlikely(has_isolate_pageblock(zone) ||
1248 is_migrate_isolate(migratetype))) {
1249 migratetype = get_pfnblock_migratetype(page, pfn);
1251 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1252 spin_unlock_irqrestore(&zone->lock, flags);
1255 static void __free_pages_ok(struct page *page, unsigned int order,
1258 unsigned long flags;
1260 unsigned long pfn = page_to_pfn(page);
1261 struct zone *zone = page_zone(page);
1263 if (!free_pages_prepare(page, order, fpi_flags))
1267 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1268 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1269 * This will reduce the lock holding time.
1271 migratetype = get_pfnblock_migratetype(page, pfn);
1273 spin_lock_irqsave(&zone->lock, flags);
1274 if (unlikely(has_isolate_pageblock(zone) ||
1275 is_migrate_isolate(migratetype))) {
1276 migratetype = get_pfnblock_migratetype(page, pfn);
1278 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1279 spin_unlock_irqrestore(&zone->lock, flags);
1281 __count_vm_events(PGFREE, 1 << order);
1284 void __free_pages_core(struct page *page, unsigned int order)
1286 unsigned int nr_pages = 1 << order;
1287 struct page *p = page;
1291 * When initializing the memmap, __init_single_page() sets the refcount
1292 * of all pages to 1 ("allocated"/"not free"). We have to set the
1293 * refcount of all involved pages to 0.
1296 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1298 __ClearPageReserved(p);
1299 set_page_count(p, 0);
1301 __ClearPageReserved(p);
1302 set_page_count(p, 0);
1304 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1306 if (page_contains_unaccepted(page, order)) {
1307 if (order == MAX_ORDER && __free_unaccepted(page))
1310 accept_page(page, order);
1314 * Bypass PCP and place fresh pages right to the tail, primarily
1315 * relevant for memory onlining.
1317 __free_pages_ok(page, order, FPI_TO_TAIL);
1321 * Check that the whole (or subset of) a pageblock given by the interval of
1322 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1323 * with the migration of free compaction scanner.
1325 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1327 * It's possible on some configurations to have a setup like node0 node1 node0
1328 * i.e. it's possible that all pages within a zones range of pages do not
1329 * belong to a single zone. We assume that a border between node0 and node1
1330 * can occur within a single pageblock, but not a node0 node1 node0
1331 * interleaving within a single pageblock. It is therefore sufficient to check
1332 * the first and last page of a pageblock and avoid checking each individual
1333 * page in a pageblock.
1335 * Note: the function may return non-NULL struct page even for a page block
1336 * which contains a memory hole (i.e. there is no physical memory for a subset
1337 * of the pfn range). For example, if the pageblock order is MAX_ORDER, which
1338 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1339 * even though the start pfn is online and valid. This should be safe most of
1340 * the time because struct pages are still initialized via init_unavailable_range()
1341 * and pfn walkers shouldn't touch any physical memory range for which they do
1342 * not recognize any specific metadata in struct pages.
1344 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1345 unsigned long end_pfn, struct zone *zone)
1347 struct page *start_page;
1348 struct page *end_page;
1350 /* end_pfn is one past the range we are checking */
1353 if (!pfn_valid(end_pfn))
1356 start_page = pfn_to_online_page(start_pfn);
1360 if (page_zone(start_page) != zone)
1363 end_page = pfn_to_page(end_pfn);
1365 /* This gives a shorter code than deriving page_zone(end_page) */
1366 if (page_zone_id(start_page) != page_zone_id(end_page))
1373 * The order of subdivision here is critical for the IO subsystem.
1374 * Please do not alter this order without good reasons and regression
1375 * testing. Specifically, as large blocks of memory are subdivided,
1376 * the order in which smaller blocks are delivered depends on the order
1377 * they're subdivided in this function. This is the primary factor
1378 * influencing the order in which pages are delivered to the IO
1379 * subsystem according to empirical testing, and this is also justified
1380 * by considering the behavior of a buddy system containing a single
1381 * large block of memory acted on by a series of small allocations.
1382 * This behavior is a critical factor in sglist merging's success.
1386 static inline void expand(struct zone *zone, struct page *page,
1387 int low, int high, int migratetype)
1389 unsigned long size = 1 << high;
1391 while (high > low) {
1394 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1397 * Mark as guard pages (or page), that will allow to
1398 * merge back to allocator when buddy will be freed.
1399 * Corresponding page table entries will not be touched,
1400 * pages will stay not present in virtual address space
1402 if (set_page_guard(zone, &page[size], high, migratetype))
1405 add_to_free_list(&page[size], zone, high, migratetype);
1406 set_buddy_order(&page[size], high);
1410 static void check_new_page_bad(struct page *page)
1412 if (unlikely(page->flags & __PG_HWPOISON)) {
1413 /* Don't complain about hwpoisoned pages */
1414 page_mapcount_reset(page); /* remove PageBuddy */
1419 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1423 * This page is about to be returned from the page allocator
1425 static int check_new_page(struct page *page)
1427 if (likely(page_expected_state(page,
1428 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1431 check_new_page_bad(page);
1435 static inline bool check_new_pages(struct page *page, unsigned int order)
1437 if (is_check_pages_enabled()) {
1438 for (int i = 0; i < (1 << order); i++) {
1439 struct page *p = page + i;
1441 if (check_new_page(p))
1449 static inline bool should_skip_kasan_unpoison(gfp_t flags)
1451 /* Don't skip if a software KASAN mode is enabled. */
1452 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1453 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1456 /* Skip, if hardware tag-based KASAN is not enabled. */
1457 if (!kasan_hw_tags_enabled())
1461 * With hardware tag-based KASAN enabled, skip if this has been
1462 * requested via __GFP_SKIP_KASAN.
1464 return flags & __GFP_SKIP_KASAN;
1467 static inline bool should_skip_init(gfp_t flags)
1469 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1470 if (!kasan_hw_tags_enabled())
1473 /* For hardware tag-based KASAN, skip if requested. */
1474 return (flags & __GFP_SKIP_ZERO);
1477 inline void post_alloc_hook(struct page *page, unsigned int order,
1480 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1481 !should_skip_init(gfp_flags);
1482 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1485 set_page_private(page, 0);
1486 set_page_refcounted(page);
1488 arch_alloc_page(page, order);
1489 debug_pagealloc_map_pages(page, 1 << order);
1492 * Page unpoisoning must happen before memory initialization.
1493 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1494 * allocations and the page unpoisoning code will complain.
1496 kernel_unpoison_pages(page, 1 << order);
1499 * As memory initialization might be integrated into KASAN,
1500 * KASAN unpoisoning and memory initializion code must be
1501 * kept together to avoid discrepancies in behavior.
1505 * If memory tags should be zeroed
1506 * (which happens only when memory should be initialized as well).
1509 /* Initialize both memory and memory tags. */
1510 for (i = 0; i != 1 << order; ++i)
1511 tag_clear_highpage(page + i);
1513 /* Take note that memory was initialized by the loop above. */
1516 if (!should_skip_kasan_unpoison(gfp_flags) &&
1517 kasan_unpoison_pages(page, order, init)) {
1518 /* Take note that memory was initialized by KASAN. */
1519 if (kasan_has_integrated_init())
1523 * If memory tags have not been set by KASAN, reset the page
1524 * tags to ensure page_address() dereferencing does not fault.
1526 for (i = 0; i != 1 << order; ++i)
1527 page_kasan_tag_reset(page + i);
1529 /* If memory is still not initialized, initialize it now. */
1531 kernel_init_pages(page, 1 << order);
1533 set_page_owner(page, order, gfp_flags);
1534 page_table_check_alloc(page, order);
1537 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1538 unsigned int alloc_flags)
1540 post_alloc_hook(page, order, gfp_flags);
1542 if (order && (gfp_flags & __GFP_COMP))
1543 prep_compound_page(page, order);
1546 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1547 * allocate the page. The expectation is that the caller is taking
1548 * steps that will free more memory. The caller should avoid the page
1549 * being used for !PFMEMALLOC purposes.
1551 if (alloc_flags & ALLOC_NO_WATERMARKS)
1552 set_page_pfmemalloc(page);
1554 clear_page_pfmemalloc(page);
1558 * Go through the free lists for the given migratetype and remove
1559 * the smallest available page from the freelists
1561 static __always_inline
1562 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1565 unsigned int current_order;
1566 struct free_area *area;
1569 /* Find a page of the appropriate size in the preferred list */
1570 for (current_order = order; current_order <= MAX_ORDER; ++current_order) {
1571 area = &(zone->free_area[current_order]);
1572 page = get_page_from_free_area(area, migratetype);
1575 del_page_from_free_list(page, zone, current_order);
1576 expand(zone, page, order, current_order, migratetype);
1577 set_pcppage_migratetype(page, migratetype);
1578 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1579 pcp_allowed_order(order) &&
1580 migratetype < MIGRATE_PCPTYPES);
1589 * This array describes the order lists are fallen back to when
1590 * the free lists for the desirable migrate type are depleted
1592 * The other migratetypes do not have fallbacks.
1594 static int fallbacks[MIGRATE_TYPES][MIGRATE_PCPTYPES - 1] = {
1595 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1596 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1597 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1601 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1604 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1607 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1608 unsigned int order) { return NULL; }
1612 * Move the free pages in a range to the freelist tail of the requested type.
1613 * Note that start_page and end_pages are not aligned on a pageblock
1614 * boundary. If alignment is required, use move_freepages_block()
1616 static int move_freepages(struct zone *zone,
1617 unsigned long start_pfn, unsigned long end_pfn,
1618 int migratetype, int *num_movable)
1623 int pages_moved = 0;
1625 for (pfn = start_pfn; pfn <= end_pfn;) {
1626 page = pfn_to_page(pfn);
1627 if (!PageBuddy(page)) {
1629 * We assume that pages that could be isolated for
1630 * migration are movable. But we don't actually try
1631 * isolating, as that would be expensive.
1634 (PageLRU(page) || __PageMovable(page)))
1640 /* Make sure we are not inadvertently changing nodes */
1641 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1642 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1644 order = buddy_order(page);
1645 move_to_free_list(page, zone, order, migratetype);
1647 pages_moved += 1 << order;
1653 int move_freepages_block(struct zone *zone, struct page *page,
1654 int migratetype, int *num_movable)
1656 unsigned long start_pfn, end_pfn, pfn;
1661 pfn = page_to_pfn(page);
1662 start_pfn = pageblock_start_pfn(pfn);
1663 end_pfn = pageblock_end_pfn(pfn) - 1;
1665 /* Do not cross zone boundaries */
1666 if (!zone_spans_pfn(zone, start_pfn))
1668 if (!zone_spans_pfn(zone, end_pfn))
1671 return move_freepages(zone, start_pfn, end_pfn, migratetype,
1675 static void change_pageblock_range(struct page *pageblock_page,
1676 int start_order, int migratetype)
1678 int nr_pageblocks = 1 << (start_order - pageblock_order);
1680 while (nr_pageblocks--) {
1681 set_pageblock_migratetype(pageblock_page, migratetype);
1682 pageblock_page += pageblock_nr_pages;
1687 * When we are falling back to another migratetype during allocation, try to
1688 * steal extra free pages from the same pageblocks to satisfy further
1689 * allocations, instead of polluting multiple pageblocks.
1691 * If we are stealing a relatively large buddy page, it is likely there will
1692 * be more free pages in the pageblock, so try to steal them all. For
1693 * reclaimable and unmovable allocations, we steal regardless of page size,
1694 * as fragmentation caused by those allocations polluting movable pageblocks
1695 * is worse than movable allocations stealing from unmovable and reclaimable
1698 static bool can_steal_fallback(unsigned int order, int start_mt)
1701 * Leaving this order check is intended, although there is
1702 * relaxed order check in next check. The reason is that
1703 * we can actually steal whole pageblock if this condition met,
1704 * but, below check doesn't guarantee it and that is just heuristic
1705 * so could be changed anytime.
1707 if (order >= pageblock_order)
1710 if (order >= pageblock_order / 2 ||
1711 start_mt == MIGRATE_RECLAIMABLE ||
1712 start_mt == MIGRATE_UNMOVABLE ||
1713 page_group_by_mobility_disabled)
1719 static inline bool boost_watermark(struct zone *zone)
1721 unsigned long max_boost;
1723 if (!watermark_boost_factor)
1726 * Don't bother in zones that are unlikely to produce results.
1727 * On small machines, including kdump capture kernels running
1728 * in a small area, boosting the watermark can cause an out of
1729 * memory situation immediately.
1731 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1734 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1735 watermark_boost_factor, 10000);
1738 * high watermark may be uninitialised if fragmentation occurs
1739 * very early in boot so do not boost. We do not fall
1740 * through and boost by pageblock_nr_pages as failing
1741 * allocations that early means that reclaim is not going
1742 * to help and it may even be impossible to reclaim the
1743 * boosted watermark resulting in a hang.
1748 max_boost = max(pageblock_nr_pages, max_boost);
1750 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1757 * This function implements actual steal behaviour. If order is large enough,
1758 * we can steal whole pageblock. If not, we first move freepages in this
1759 * pageblock to our migratetype and determine how many already-allocated pages
1760 * are there in the pageblock with a compatible migratetype. If at least half
1761 * of pages are free or compatible, we can change migratetype of the pageblock
1762 * itself, so pages freed in the future will be put on the correct free list.
1764 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1765 unsigned int alloc_flags, int start_type, bool whole_block)
1767 unsigned int current_order = buddy_order(page);
1768 int free_pages, movable_pages, alike_pages;
1771 old_block_type = get_pageblock_migratetype(page);
1774 * This can happen due to races and we want to prevent broken
1775 * highatomic accounting.
1777 if (is_migrate_highatomic(old_block_type))
1780 /* Take ownership for orders >= pageblock_order */
1781 if (current_order >= pageblock_order) {
1782 change_pageblock_range(page, current_order, start_type);
1787 * Boost watermarks to increase reclaim pressure to reduce the
1788 * likelihood of future fallbacks. Wake kswapd now as the node
1789 * may be balanced overall and kswapd will not wake naturally.
1791 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1792 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1794 /* We are not allowed to try stealing from the whole block */
1798 free_pages = move_freepages_block(zone, page, start_type,
1800 /* moving whole block can fail due to zone boundary conditions */
1805 * Determine how many pages are compatible with our allocation.
1806 * For movable allocation, it's the number of movable pages which
1807 * we just obtained. For other types it's a bit more tricky.
1809 if (start_type == MIGRATE_MOVABLE) {
1810 alike_pages = movable_pages;
1813 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1814 * to MOVABLE pageblock, consider all non-movable pages as
1815 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1816 * vice versa, be conservative since we can't distinguish the
1817 * exact migratetype of non-movable pages.
1819 if (old_block_type == MIGRATE_MOVABLE)
1820 alike_pages = pageblock_nr_pages
1821 - (free_pages + movable_pages);
1826 * If a sufficient number of pages in the block are either free or of
1827 * compatible migratability as our allocation, claim the whole block.
1829 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1830 page_group_by_mobility_disabled)
1831 set_pageblock_migratetype(page, start_type);
1836 move_to_free_list(page, zone, current_order, start_type);
1840 * Check whether there is a suitable fallback freepage with requested order.
1841 * If only_stealable is true, this function returns fallback_mt only if
1842 * we can steal other freepages all together. This would help to reduce
1843 * fragmentation due to mixed migratetype pages in one pageblock.
1845 int find_suitable_fallback(struct free_area *area, unsigned int order,
1846 int migratetype, bool only_stealable, bool *can_steal)
1851 if (area->nr_free == 0)
1855 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
1856 fallback_mt = fallbacks[migratetype][i];
1857 if (free_area_empty(area, fallback_mt))
1860 if (can_steal_fallback(order, migratetype))
1863 if (!only_stealable)
1874 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1875 * there are no empty page blocks that contain a page with a suitable order
1877 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone)
1880 unsigned long max_managed, flags;
1883 * The number reserved as: minimum is 1 pageblock, maximum is
1884 * roughly 1% of a zone. But if 1% of a zone falls below a
1885 * pageblock size, then don't reserve any pageblocks.
1886 * Check is race-prone but harmless.
1888 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
1890 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
1891 if (zone->nr_reserved_highatomic >= max_managed)
1894 spin_lock_irqsave(&zone->lock, flags);
1896 /* Recheck the nr_reserved_highatomic limit under the lock */
1897 if (zone->nr_reserved_highatomic >= max_managed)
1901 mt = get_pageblock_migratetype(page);
1902 /* Only reserve normal pageblocks (i.e., they can merge with others) */
1903 if (migratetype_is_mergeable(mt)) {
1904 zone->nr_reserved_highatomic += pageblock_nr_pages;
1905 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1906 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
1910 spin_unlock_irqrestore(&zone->lock, flags);
1914 * Used when an allocation is about to fail under memory pressure. This
1915 * potentially hurts the reliability of high-order allocations when under
1916 * intense memory pressure but failed atomic allocations should be easier
1917 * to recover from than an OOM.
1919 * If @force is true, try to unreserve a pageblock even though highatomic
1920 * pageblock is exhausted.
1922 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
1925 struct zonelist *zonelist = ac->zonelist;
1926 unsigned long flags;
1933 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
1936 * Preserve at least one pageblock unless memory pressure
1939 if (!force && zone->nr_reserved_highatomic <=
1943 spin_lock_irqsave(&zone->lock, flags);
1944 for (order = 0; order <= MAX_ORDER; order++) {
1945 struct free_area *area = &(zone->free_area[order]);
1947 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
1952 * In page freeing path, migratetype change is racy so
1953 * we can counter several free pages in a pageblock
1954 * in this loop although we changed the pageblock type
1955 * from highatomic to ac->migratetype. So we should
1956 * adjust the count once.
1958 if (is_migrate_highatomic_page(page)) {
1960 * It should never happen but changes to
1961 * locking could inadvertently allow a per-cpu
1962 * drain to add pages to MIGRATE_HIGHATOMIC
1963 * while unreserving so be safe and watch for
1966 zone->nr_reserved_highatomic -= min(
1968 zone->nr_reserved_highatomic);
1972 * Convert to ac->migratetype and avoid the normal
1973 * pageblock stealing heuristics. Minimally, the caller
1974 * is doing the work and needs the pages. More
1975 * importantly, if the block was always converted to
1976 * MIGRATE_UNMOVABLE or another type then the number
1977 * of pageblocks that cannot be completely freed
1980 set_pageblock_migratetype(page, ac->migratetype);
1981 ret = move_freepages_block(zone, page, ac->migratetype,
1984 spin_unlock_irqrestore(&zone->lock, flags);
1988 spin_unlock_irqrestore(&zone->lock, flags);
1995 * Try finding a free buddy page on the fallback list and put it on the free
1996 * list of requested migratetype, possibly along with other pages from the same
1997 * block, depending on fragmentation avoidance heuristics. Returns true if
1998 * fallback was found so that __rmqueue_smallest() can grab it.
2000 * The use of signed ints for order and current_order is a deliberate
2001 * deviation from the rest of this file, to make the for loop
2002 * condition simpler.
2004 static __always_inline bool
2005 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2006 unsigned int alloc_flags)
2008 struct free_area *area;
2010 int min_order = order;
2016 * Do not steal pages from freelists belonging to other pageblocks
2017 * i.e. orders < pageblock_order. If there are no local zones free,
2018 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2020 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2021 min_order = pageblock_order;
2024 * Find the largest available free page in the other list. This roughly
2025 * approximates finding the pageblock with the most free pages, which
2026 * would be too costly to do exactly.
2028 for (current_order = MAX_ORDER; current_order >= min_order;
2030 area = &(zone->free_area[current_order]);
2031 fallback_mt = find_suitable_fallback(area, current_order,
2032 start_migratetype, false, &can_steal);
2033 if (fallback_mt == -1)
2037 * We cannot steal all free pages from the pageblock and the
2038 * requested migratetype is movable. In that case it's better to
2039 * steal and split the smallest available page instead of the
2040 * largest available page, because even if the next movable
2041 * allocation falls back into a different pageblock than this
2042 * one, it won't cause permanent fragmentation.
2044 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2045 && current_order > order)
2054 for (current_order = order; current_order <= MAX_ORDER;
2056 area = &(zone->free_area[current_order]);
2057 fallback_mt = find_suitable_fallback(area, current_order,
2058 start_migratetype, false, &can_steal);
2059 if (fallback_mt != -1)
2064 * This should not happen - we already found a suitable fallback
2065 * when looking for the largest page.
2067 VM_BUG_ON(current_order > MAX_ORDER);
2070 page = get_page_from_free_area(area, fallback_mt);
2072 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2075 trace_mm_page_alloc_extfrag(page, order, current_order,
2076 start_migratetype, fallback_mt);
2083 * Do the hard work of removing an element from the buddy allocator.
2084 * Call me with the zone->lock already held.
2086 static __always_inline struct page *
2087 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2088 unsigned int alloc_flags)
2092 if (IS_ENABLED(CONFIG_CMA)) {
2094 * Balance movable allocations between regular and CMA areas by
2095 * allocating from CMA when over half of the zone's free memory
2096 * is in the CMA area.
2098 if (alloc_flags & ALLOC_CMA &&
2099 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2100 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2101 page = __rmqueue_cma_fallback(zone, order);
2107 page = __rmqueue_smallest(zone, order, migratetype);
2108 if (unlikely(!page)) {
2109 if (alloc_flags & ALLOC_CMA)
2110 page = __rmqueue_cma_fallback(zone, order);
2112 if (!page && __rmqueue_fallback(zone, order, migratetype,
2120 * Obtain a specified number of elements from the buddy allocator, all under
2121 * a single hold of the lock, for efficiency. Add them to the supplied list.
2122 * Returns the number of new pages which were placed at *list.
2124 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2125 unsigned long count, struct list_head *list,
2126 int migratetype, unsigned int alloc_flags)
2128 unsigned long flags;
2131 spin_lock_irqsave(&zone->lock, flags);
2132 for (i = 0; i < count; ++i) {
2133 struct page *page = __rmqueue(zone, order, migratetype,
2135 if (unlikely(page == NULL))
2139 * Split buddy pages returned by expand() are received here in
2140 * physical page order. The page is added to the tail of
2141 * caller's list. From the callers perspective, the linked list
2142 * is ordered by page number under some conditions. This is
2143 * useful for IO devices that can forward direction from the
2144 * head, thus also in the physical page order. This is useful
2145 * for IO devices that can merge IO requests if the physical
2146 * pages are ordered properly.
2148 list_add_tail(&page->pcp_list, list);
2149 if (is_migrate_cma(get_pcppage_migratetype(page)))
2150 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2154 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2155 spin_unlock_irqrestore(&zone->lock, flags);
2161 * Called from the vmstat counter updater to decay the PCP high.
2162 * Return whether there are addition works to do.
2164 int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2166 int high_min, to_drain, batch;
2169 high_min = READ_ONCE(pcp->high_min);
2170 batch = READ_ONCE(pcp->batch);
2172 * Decrease pcp->high periodically to try to free possible
2173 * idle PCP pages. And, avoid to free too many pages to
2174 * control latency. This caps pcp->high decrement too.
2176 if (pcp->high > high_min) {
2177 pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2178 pcp->high - (pcp->high >> 3), high_min);
2179 if (pcp->high > high_min)
2183 to_drain = pcp->count - pcp->high;
2185 spin_lock(&pcp->lock);
2186 free_pcppages_bulk(zone, to_drain, pcp, 0);
2187 spin_unlock(&pcp->lock);
2196 * Called from the vmstat counter updater to drain pagesets of this
2197 * currently executing processor on remote nodes after they have
2200 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2202 int to_drain, batch;
2204 batch = READ_ONCE(pcp->batch);
2205 to_drain = min(pcp->count, batch);
2207 spin_lock(&pcp->lock);
2208 free_pcppages_bulk(zone, to_drain, pcp, 0);
2209 spin_unlock(&pcp->lock);
2215 * Drain pcplists of the indicated processor and zone.
2217 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2219 struct per_cpu_pages *pcp;
2221 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2223 spin_lock(&pcp->lock);
2224 free_pcppages_bulk(zone, pcp->count, pcp, 0);
2225 spin_unlock(&pcp->lock);
2230 * Drain pcplists of all zones on the indicated processor.
2232 static void drain_pages(unsigned int cpu)
2236 for_each_populated_zone(zone) {
2237 drain_pages_zone(cpu, zone);
2242 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2244 void drain_local_pages(struct zone *zone)
2246 int cpu = smp_processor_id();
2249 drain_pages_zone(cpu, zone);
2255 * The implementation of drain_all_pages(), exposing an extra parameter to
2256 * drain on all cpus.
2258 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2259 * not empty. The check for non-emptiness can however race with a free to
2260 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2261 * that need the guarantee that every CPU has drained can disable the
2262 * optimizing racy check.
2264 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2269 * Allocate in the BSS so we won't require allocation in
2270 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2272 static cpumask_t cpus_with_pcps;
2275 * Do not drain if one is already in progress unless it's specific to
2276 * a zone. Such callers are primarily CMA and memory hotplug and need
2277 * the drain to be complete when the call returns.
2279 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2282 mutex_lock(&pcpu_drain_mutex);
2286 * We don't care about racing with CPU hotplug event
2287 * as offline notification will cause the notified
2288 * cpu to drain that CPU pcps and on_each_cpu_mask
2289 * disables preemption as part of its processing
2291 for_each_online_cpu(cpu) {
2292 struct per_cpu_pages *pcp;
2294 bool has_pcps = false;
2296 if (force_all_cpus) {
2298 * The pcp.count check is racy, some callers need a
2299 * guarantee that no cpu is missed.
2303 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2307 for_each_populated_zone(z) {
2308 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2317 cpumask_set_cpu(cpu, &cpus_with_pcps);
2319 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2322 for_each_cpu(cpu, &cpus_with_pcps) {
2324 drain_pages_zone(cpu, zone);
2329 mutex_unlock(&pcpu_drain_mutex);
2333 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2335 * When zone parameter is non-NULL, spill just the single zone's pages.
2337 void drain_all_pages(struct zone *zone)
2339 __drain_all_pages(zone, false);
2342 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
2347 if (!free_pages_prepare(page, order, FPI_NONE))
2350 migratetype = get_pfnblock_migratetype(page, pfn);
2351 set_pcppage_migratetype(page, migratetype);
2355 static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2357 int min_nr_free, max_nr_free;
2359 /* Free as much as possible if batch freeing high-order pages. */
2360 if (unlikely(free_high))
2361 return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2363 /* Check for PCP disabled or boot pageset */
2364 if (unlikely(high < batch))
2367 /* Leave at least pcp->batch pages on the list */
2368 min_nr_free = batch;
2369 max_nr_free = high - batch;
2372 * Increase the batch number to the number of the consecutive
2373 * freed pages to reduce zone lock contention.
2375 batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2380 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2381 int batch, bool free_high)
2383 int high, high_min, high_max;
2385 high_min = READ_ONCE(pcp->high_min);
2386 high_max = READ_ONCE(pcp->high_max);
2387 high = pcp->high = clamp(pcp->high, high_min, high_max);
2389 if (unlikely(!high))
2392 if (unlikely(free_high)) {
2393 pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2399 * If reclaim is active, limit the number of pages that can be
2400 * stored on pcp lists
2402 if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2403 int free_count = max_t(int, pcp->free_count, batch);
2405 pcp->high = max(high - free_count, high_min);
2406 return min(batch << 2, pcp->high);
2409 if (high_min == high_max)
2412 if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2413 int free_count = max_t(int, pcp->free_count, batch);
2415 pcp->high = max(high - free_count, high_min);
2416 high = max(pcp->count, high_min);
2417 } else if (pcp->count >= high) {
2418 int need_high = pcp->free_count + batch;
2420 /* pcp->high should be large enough to hold batch freed pages */
2421 if (pcp->high < need_high)
2422 pcp->high = clamp(need_high, high_min, high_max);
2428 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2429 struct page *page, int migratetype,
2434 bool free_high = false;
2437 * On freeing, reduce the number of pages that are batch allocated.
2438 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2441 pcp->alloc_factor >>= 1;
2442 __count_vm_events(PGFREE, 1 << order);
2443 pindex = order_to_pindex(migratetype, order);
2444 list_add(&page->pcp_list, &pcp->lists[pindex]);
2445 pcp->count += 1 << order;
2447 batch = READ_ONCE(pcp->batch);
2449 * As high-order pages other than THP's stored on PCP can contribute
2450 * to fragmentation, limit the number stored when PCP is heavily
2451 * freeing without allocation. The remainder after bulk freeing
2452 * stops will be drained from vmstat refresh context.
2454 if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2455 free_high = (pcp->free_count >= batch &&
2456 (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2457 (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2458 pcp->count >= READ_ONCE(batch)));
2459 pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2460 } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2461 pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2463 if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2464 pcp->free_count += (1 << order);
2465 high = nr_pcp_high(pcp, zone, batch, free_high);
2466 if (pcp->count >= high) {
2467 free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high),
2469 if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2470 zone_watermark_ok(zone, 0, high_wmark_pages(zone),
2472 clear_bit(ZONE_BELOW_HIGH, &zone->flags);
2479 void free_unref_page(struct page *page, unsigned int order)
2481 unsigned long __maybe_unused UP_flags;
2482 struct per_cpu_pages *pcp;
2484 unsigned long pfn = page_to_pfn(page);
2485 int migratetype, pcpmigratetype;
2487 if (!free_unref_page_prepare(page, pfn, order))
2491 * We only track unmovable, reclaimable and movable on pcp lists.
2492 * Place ISOLATE pages on the isolated list because they are being
2493 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2494 * get those areas back if necessary. Otherwise, we may have to free
2495 * excessively into the page allocator
2497 migratetype = pcpmigratetype = get_pcppage_migratetype(page);
2498 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2499 if (unlikely(is_migrate_isolate(migratetype))) {
2500 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
2503 pcpmigratetype = MIGRATE_MOVABLE;
2506 zone = page_zone(page);
2507 pcp_trylock_prepare(UP_flags);
2508 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2510 free_unref_page_commit(zone, pcp, page, pcpmigratetype, order);
2511 pcp_spin_unlock(pcp);
2513 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
2515 pcp_trylock_finish(UP_flags);
2519 * Free a list of 0-order pages
2521 void free_unref_page_list(struct list_head *list)
2523 unsigned long __maybe_unused UP_flags;
2524 struct page *page, *next;
2525 struct per_cpu_pages *pcp = NULL;
2526 struct zone *locked_zone = NULL;
2527 int batch_count = 0;
2530 /* Prepare pages for freeing */
2531 list_for_each_entry_safe(page, next, list, lru) {
2532 unsigned long pfn = page_to_pfn(page);
2533 if (!free_unref_page_prepare(page, pfn, 0)) {
2534 list_del(&page->lru);
2539 * Free isolated pages directly to the allocator, see
2540 * comment in free_unref_page.
2542 migratetype = get_pcppage_migratetype(page);
2543 if (unlikely(is_migrate_isolate(migratetype))) {
2544 list_del(&page->lru);
2545 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
2550 list_for_each_entry_safe(page, next, list, lru) {
2551 struct zone *zone = page_zone(page);
2553 list_del(&page->lru);
2554 migratetype = get_pcppage_migratetype(page);
2557 * Either different zone requiring a different pcp lock or
2558 * excessive lock hold times when freeing a large list of
2561 if (zone != locked_zone || batch_count == SWAP_CLUSTER_MAX) {
2563 pcp_spin_unlock(pcp);
2564 pcp_trylock_finish(UP_flags);
2570 * trylock is necessary as pages may be getting freed
2571 * from IRQ or SoftIRQ context after an IO completion.
2573 pcp_trylock_prepare(UP_flags);
2574 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2575 if (unlikely(!pcp)) {
2576 pcp_trylock_finish(UP_flags);
2577 free_one_page(zone, page, page_to_pfn(page),
2578 0, migratetype, FPI_NONE);
2586 * Non-isolated types over MIGRATE_PCPTYPES get added
2587 * to the MIGRATE_MOVABLE pcp list.
2589 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2590 migratetype = MIGRATE_MOVABLE;
2592 trace_mm_page_free_batched(page);
2593 free_unref_page_commit(zone, pcp, page, migratetype, 0);
2598 pcp_spin_unlock(pcp);
2599 pcp_trylock_finish(UP_flags);
2604 * split_page takes a non-compound higher-order page, and splits it into
2605 * n (1<<order) sub-pages: page[0..n]
2606 * Each sub-page must be freed individually.
2608 * Note: this is probably too low level an operation for use in drivers.
2609 * Please consult with lkml before using this in your driver.
2611 void split_page(struct page *page, unsigned int order)
2615 VM_BUG_ON_PAGE(PageCompound(page), page);
2616 VM_BUG_ON_PAGE(!page_count(page), page);
2618 for (i = 1; i < (1 << order); i++)
2619 set_page_refcounted(page + i);
2620 split_page_owner(page, 1 << order);
2621 split_page_memcg(page, 1 << order);
2623 EXPORT_SYMBOL_GPL(split_page);
2625 int __isolate_free_page(struct page *page, unsigned int order)
2627 struct zone *zone = page_zone(page);
2628 int mt = get_pageblock_migratetype(page);
2630 if (!is_migrate_isolate(mt)) {
2631 unsigned long watermark;
2633 * Obey watermarks as if the page was being allocated. We can
2634 * emulate a high-order watermark check with a raised order-0
2635 * watermark, because we already know our high-order page
2638 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2639 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2642 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2645 del_page_from_free_list(page, zone, order);
2648 * Set the pageblock if the isolated page is at least half of a
2651 if (order >= pageblock_order - 1) {
2652 struct page *endpage = page + (1 << order) - 1;
2653 for (; page < endpage; page += pageblock_nr_pages) {
2654 int mt = get_pageblock_migratetype(page);
2656 * Only change normal pageblocks (i.e., they can merge
2659 if (migratetype_is_mergeable(mt))
2660 set_pageblock_migratetype(page,
2665 return 1UL << order;
2669 * __putback_isolated_page - Return a now-isolated page back where we got it
2670 * @page: Page that was isolated
2671 * @order: Order of the isolated page
2672 * @mt: The page's pageblock's migratetype
2674 * This function is meant to return a page pulled from the free lists via
2675 * __isolate_free_page back to the free lists they were pulled from.
2677 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2679 struct zone *zone = page_zone(page);
2681 /* zone lock should be held when this function is called */
2682 lockdep_assert_held(&zone->lock);
2684 /* Return isolated page to tail of freelist. */
2685 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2686 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2690 * Update NUMA hit/miss statistics
2692 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2696 enum numa_stat_item local_stat = NUMA_LOCAL;
2698 /* skip numa counters update if numa stats is disabled */
2699 if (!static_branch_likely(&vm_numa_stat_key))
2702 if (zone_to_nid(z) != numa_node_id())
2703 local_stat = NUMA_OTHER;
2705 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2706 __count_numa_events(z, NUMA_HIT, nr_account);
2708 __count_numa_events(z, NUMA_MISS, nr_account);
2709 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2711 __count_numa_events(z, local_stat, nr_account);
2715 static __always_inline
2716 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2717 unsigned int order, unsigned int alloc_flags,
2721 unsigned long flags;
2725 spin_lock_irqsave(&zone->lock, flags);
2726 if (alloc_flags & ALLOC_HIGHATOMIC)
2727 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2729 page = __rmqueue(zone, order, migratetype, alloc_flags);
2732 * If the allocation fails, allow OOM handling access
2733 * to HIGHATOMIC reserves as failing now is worse than
2734 * failing a high-order atomic allocation in the
2737 if (!page && (alloc_flags & ALLOC_OOM))
2738 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2741 spin_unlock_irqrestore(&zone->lock, flags);
2745 __mod_zone_freepage_state(zone, -(1 << order),
2746 get_pcppage_migratetype(page));
2747 spin_unlock_irqrestore(&zone->lock, flags);
2748 } while (check_new_pages(page, order));
2750 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2751 zone_statistics(preferred_zone, zone, 1);
2756 static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
2758 int high, base_batch, batch, max_nr_alloc;
2759 int high_max, high_min;
2761 base_batch = READ_ONCE(pcp->batch);
2762 high_min = READ_ONCE(pcp->high_min);
2763 high_max = READ_ONCE(pcp->high_max);
2764 high = pcp->high = clamp(pcp->high, high_min, high_max);
2766 /* Check for PCP disabled or boot pageset */
2767 if (unlikely(high < base_batch))
2773 batch = (base_batch << pcp->alloc_factor);
2776 * If we had larger pcp->high, we could avoid to allocate from
2779 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
2780 high = pcp->high = min(high + batch, high_max);
2783 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
2785 * Double the number of pages allocated each time there is
2786 * subsequent allocation of order-0 pages without any freeing.
2788 if (batch <= max_nr_alloc &&
2789 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
2790 pcp->alloc_factor++;
2791 batch = min(batch, max_nr_alloc);
2795 * Scale batch relative to order if batch implies free pages
2796 * can be stored on the PCP. Batch can be 1 for small zones or
2797 * for boot pagesets which should never store free pages as
2798 * the pages may belong to arbitrary zones.
2801 batch = max(batch >> order, 2);
2806 /* Remove page from the per-cpu list, caller must protect the list */
2808 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2810 unsigned int alloc_flags,
2811 struct per_cpu_pages *pcp,
2812 struct list_head *list)
2817 if (list_empty(list)) {
2818 int batch = nr_pcp_alloc(pcp, zone, order);
2821 alloced = rmqueue_bulk(zone, order,
2823 migratetype, alloc_flags);
2825 pcp->count += alloced << order;
2826 if (unlikely(list_empty(list)))
2830 page = list_first_entry(list, struct page, pcp_list);
2831 list_del(&page->pcp_list);
2832 pcp->count -= 1 << order;
2833 } while (check_new_pages(page, order));
2838 /* Lock and remove page from the per-cpu list */
2839 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2840 struct zone *zone, unsigned int order,
2841 int migratetype, unsigned int alloc_flags)
2843 struct per_cpu_pages *pcp;
2844 struct list_head *list;
2846 unsigned long __maybe_unused UP_flags;
2848 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2849 pcp_trylock_prepare(UP_flags);
2850 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2852 pcp_trylock_finish(UP_flags);
2857 * On allocation, reduce the number of pages that are batch freed.
2858 * See nr_pcp_free() where free_factor is increased for subsequent
2861 pcp->free_count >>= 1;
2862 list = &pcp->lists[order_to_pindex(migratetype, order)];
2863 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
2864 pcp_spin_unlock(pcp);
2865 pcp_trylock_finish(UP_flags);
2867 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2868 zone_statistics(preferred_zone, zone, 1);
2874 * Allocate a page from the given zone.
2875 * Use pcplists for THP or "cheap" high-order allocations.
2879 * Do not instrument rmqueue() with KMSAN. This function may call
2880 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
2881 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
2882 * may call rmqueue() again, which will result in a deadlock.
2884 __no_sanitize_memory
2886 struct page *rmqueue(struct zone *preferred_zone,
2887 struct zone *zone, unsigned int order,
2888 gfp_t gfp_flags, unsigned int alloc_flags,
2894 * We most definitely don't want callers attempting to
2895 * allocate greater than order-1 page units with __GFP_NOFAIL.
2897 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2899 if (likely(pcp_allowed_order(order))) {
2900 page = rmqueue_pcplist(preferred_zone, zone, order,
2901 migratetype, alloc_flags);
2906 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
2910 /* Separate test+clear to avoid unnecessary atomics */
2911 if ((alloc_flags & ALLOC_KSWAPD) &&
2912 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
2913 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2914 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
2917 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2921 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2923 return __should_fail_alloc_page(gfp_mask, order);
2925 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
2927 static inline long __zone_watermark_unusable_free(struct zone *z,
2928 unsigned int order, unsigned int alloc_flags)
2930 long unusable_free = (1 << order) - 1;
2933 * If the caller does not have rights to reserves below the min
2934 * watermark then subtract the high-atomic reserves. This will
2935 * over-estimate the size of the atomic reserve but it avoids a search.
2937 if (likely(!(alloc_flags & ALLOC_RESERVES)))
2938 unusable_free += z->nr_reserved_highatomic;
2941 /* If allocation can't use CMA areas don't use free CMA pages */
2942 if (!(alloc_flags & ALLOC_CMA))
2943 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
2945 #ifdef CONFIG_UNACCEPTED_MEMORY
2946 unusable_free += zone_page_state(z, NR_UNACCEPTED);
2949 return unusable_free;
2953 * Return true if free base pages are above 'mark'. For high-order checks it
2954 * will return true of the order-0 watermark is reached and there is at least
2955 * one free page of a suitable size. Checking now avoids taking the zone lock
2956 * to check in the allocation paths if no pages are free.
2958 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2959 int highest_zoneidx, unsigned int alloc_flags,
2965 /* free_pages may go negative - that's OK */
2966 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
2968 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
2970 * __GFP_HIGH allows access to 50% of the min reserve as well
2973 if (alloc_flags & ALLOC_MIN_RESERVE) {
2977 * Non-blocking allocations (e.g. GFP_ATOMIC) can
2978 * access more reserves than just __GFP_HIGH. Other
2979 * non-blocking allocations requests such as GFP_NOWAIT
2980 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
2981 * access to the min reserve.
2983 if (alloc_flags & ALLOC_NON_BLOCK)
2988 * OOM victims can try even harder than the normal reserve
2989 * users on the grounds that it's definitely going to be in
2990 * the exit path shortly and free memory. Any allocation it
2991 * makes during the free path will be small and short-lived.
2993 if (alloc_flags & ALLOC_OOM)
2998 * Check watermarks for an order-0 allocation request. If these
2999 * are not met, then a high-order request also cannot go ahead
3000 * even if a suitable page happened to be free.
3002 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3005 /* If this is an order-0 request then the watermark is fine */
3009 /* For a high-order request, check at least one suitable page is free */
3010 for (o = order; o <= MAX_ORDER; o++) {
3011 struct free_area *area = &z->free_area[o];
3017 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3018 if (!free_area_empty(area, mt))
3023 if ((alloc_flags & ALLOC_CMA) &&
3024 !free_area_empty(area, MIGRATE_CMA)) {
3028 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3029 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3036 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3037 int highest_zoneidx, unsigned int alloc_flags)
3039 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3040 zone_page_state(z, NR_FREE_PAGES));
3043 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3044 unsigned long mark, int highest_zoneidx,
3045 unsigned int alloc_flags, gfp_t gfp_mask)
3049 free_pages = zone_page_state(z, NR_FREE_PAGES);
3052 * Fast check for order-0 only. If this fails then the reserves
3053 * need to be calculated.
3059 usable_free = free_pages;
3060 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3062 /* reserved may over estimate high-atomic reserves. */
3063 usable_free -= min(usable_free, reserved);
3064 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3068 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3073 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3074 * when checking the min watermark. The min watermark is the
3075 * point where boosting is ignored so that kswapd is woken up
3076 * when below the low watermark.
3078 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3079 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3080 mark = z->_watermark[WMARK_MIN];
3081 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3082 alloc_flags, free_pages);
3088 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3089 unsigned long mark, int highest_zoneidx)
3091 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3093 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3094 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3096 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3101 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3103 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3105 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3106 node_reclaim_distance;
3108 #else /* CONFIG_NUMA */
3109 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3113 #endif /* CONFIG_NUMA */
3116 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3117 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3118 * premature use of a lower zone may cause lowmem pressure problems that
3119 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3120 * probably too small. It only makes sense to spread allocations to avoid
3121 * fragmentation between the Normal and DMA32 zones.
3123 static inline unsigned int
3124 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3126 unsigned int alloc_flags;
3129 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3132 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3134 #ifdef CONFIG_ZONE_DMA32
3138 if (zone_idx(zone) != ZONE_NORMAL)
3142 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3143 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3144 * on UMA that if Normal is populated then so is DMA32.
3146 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3147 if (nr_online_nodes > 1 && !populated_zone(--zone))
3150 alloc_flags |= ALLOC_NOFRAGMENT;
3151 #endif /* CONFIG_ZONE_DMA32 */
3155 /* Must be called after current_gfp_context() which can change gfp_mask */
3156 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3157 unsigned int alloc_flags)
3160 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3161 alloc_flags |= ALLOC_CMA;
3167 * get_page_from_freelist goes through the zonelist trying to allocate
3170 static struct page *
3171 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3172 const struct alloc_context *ac)
3176 struct pglist_data *last_pgdat = NULL;
3177 bool last_pgdat_dirty_ok = false;
3182 * Scan zonelist, looking for a zone with enough free.
3183 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3185 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3186 z = ac->preferred_zoneref;
3187 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3192 if (cpusets_enabled() &&
3193 (alloc_flags & ALLOC_CPUSET) &&
3194 !__cpuset_zone_allowed(zone, gfp_mask))
3197 * When allocating a page cache page for writing, we
3198 * want to get it from a node that is within its dirty
3199 * limit, such that no single node holds more than its
3200 * proportional share of globally allowed dirty pages.
3201 * The dirty limits take into account the node's
3202 * lowmem reserves and high watermark so that kswapd
3203 * should be able to balance it without having to
3204 * write pages from its LRU list.
3206 * XXX: For now, allow allocations to potentially
3207 * exceed the per-node dirty limit in the slowpath
3208 * (spread_dirty_pages unset) before going into reclaim,
3209 * which is important when on a NUMA setup the allowed
3210 * nodes are together not big enough to reach the
3211 * global limit. The proper fix for these situations
3212 * will require awareness of nodes in the
3213 * dirty-throttling and the flusher threads.
3215 if (ac->spread_dirty_pages) {
3216 if (last_pgdat != zone->zone_pgdat) {
3217 last_pgdat = zone->zone_pgdat;
3218 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3221 if (!last_pgdat_dirty_ok)
3225 if (no_fallback && nr_online_nodes > 1 &&
3226 zone != ac->preferred_zoneref->zone) {
3230 * If moving to a remote node, retry but allow
3231 * fragmenting fallbacks. Locality is more important
3232 * than fragmentation avoidance.
3234 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3235 if (zone_to_nid(zone) != local_nid) {
3236 alloc_flags &= ~ALLOC_NOFRAGMENT;
3242 * Detect whether the number of free pages is below high
3243 * watermark. If so, we will decrease pcp->high and free
3244 * PCP pages in free path to reduce the possibility of
3245 * premature page reclaiming. Detection is done here to
3246 * avoid to do that in hotter free path.
3248 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3249 goto check_alloc_wmark;
3251 mark = high_wmark_pages(zone);
3252 if (zone_watermark_fast(zone, order, mark,
3253 ac->highest_zoneidx, alloc_flags,
3257 set_bit(ZONE_BELOW_HIGH, &zone->flags);
3260 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3261 if (!zone_watermark_fast(zone, order, mark,
3262 ac->highest_zoneidx, alloc_flags,
3266 if (has_unaccepted_memory()) {
3267 if (try_to_accept_memory(zone, order))
3271 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3273 * Watermark failed for this zone, but see if we can
3274 * grow this zone if it contains deferred pages.
3276 if (deferred_pages_enabled()) {
3277 if (_deferred_grow_zone(zone, order))
3281 /* Checked here to keep the fast path fast */
3282 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3283 if (alloc_flags & ALLOC_NO_WATERMARKS)
3286 if (!node_reclaim_enabled() ||
3287 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3290 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3292 case NODE_RECLAIM_NOSCAN:
3295 case NODE_RECLAIM_FULL:
3296 /* scanned but unreclaimable */
3299 /* did we reclaim enough */
3300 if (zone_watermark_ok(zone, order, mark,
3301 ac->highest_zoneidx, alloc_flags))
3309 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3310 gfp_mask, alloc_flags, ac->migratetype);
3312 prep_new_page(page, order, gfp_mask, alloc_flags);
3315 * If this is a high-order atomic allocation then check
3316 * if the pageblock should be reserved for the future
3318 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3319 reserve_highatomic_pageblock(page, zone);
3323 if (has_unaccepted_memory()) {
3324 if (try_to_accept_memory(zone, order))
3328 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3329 /* Try again if zone has deferred pages */
3330 if (deferred_pages_enabled()) {
3331 if (_deferred_grow_zone(zone, order))
3339 * It's possible on a UMA machine to get through all zones that are
3340 * fragmented. If avoiding fragmentation, reset and try again.
3343 alloc_flags &= ~ALLOC_NOFRAGMENT;
3350 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3352 unsigned int filter = SHOW_MEM_FILTER_NODES;
3355 * This documents exceptions given to allocations in certain
3356 * contexts that are allowed to allocate outside current's set
3359 if (!(gfp_mask & __GFP_NOMEMALLOC))
3360 if (tsk_is_oom_victim(current) ||
3361 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3362 filter &= ~SHOW_MEM_FILTER_NODES;
3363 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3364 filter &= ~SHOW_MEM_FILTER_NODES;
3366 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3369 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3371 struct va_format vaf;
3373 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3375 if ((gfp_mask & __GFP_NOWARN) ||
3376 !__ratelimit(&nopage_rs) ||
3377 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3380 va_start(args, fmt);
3383 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3384 current->comm, &vaf, gfp_mask, &gfp_mask,
3385 nodemask_pr_args(nodemask));
3388 cpuset_print_current_mems_allowed();
3391 warn_alloc_show_mem(gfp_mask, nodemask);
3394 static inline struct page *
3395 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3396 unsigned int alloc_flags,
3397 const struct alloc_context *ac)
3401 page = get_page_from_freelist(gfp_mask, order,
3402 alloc_flags|ALLOC_CPUSET, ac);
3404 * fallback to ignore cpuset restriction if our nodes
3408 page = get_page_from_freelist(gfp_mask, order,
3414 static inline struct page *
3415 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3416 const struct alloc_context *ac, unsigned long *did_some_progress)
3418 struct oom_control oc = {
3419 .zonelist = ac->zonelist,
3420 .nodemask = ac->nodemask,
3422 .gfp_mask = gfp_mask,
3427 *did_some_progress = 0;
3430 * Acquire the oom lock. If that fails, somebody else is
3431 * making progress for us.
3433 if (!mutex_trylock(&oom_lock)) {
3434 *did_some_progress = 1;
3435 schedule_timeout_uninterruptible(1);
3440 * Go through the zonelist yet one more time, keep very high watermark
3441 * here, this is only to catch a parallel oom killing, we must fail if
3442 * we're still under heavy pressure. But make sure that this reclaim
3443 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3444 * allocation which will never fail due to oom_lock already held.
3446 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3447 ~__GFP_DIRECT_RECLAIM, order,
3448 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3452 /* Coredumps can quickly deplete all memory reserves */
3453 if (current->flags & PF_DUMPCORE)
3455 /* The OOM killer will not help higher order allocs */
3456 if (order > PAGE_ALLOC_COSTLY_ORDER)
3459 * We have already exhausted all our reclaim opportunities without any
3460 * success so it is time to admit defeat. We will skip the OOM killer
3461 * because it is very likely that the caller has a more reasonable
3462 * fallback than shooting a random task.
3464 * The OOM killer may not free memory on a specific node.
3466 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3468 /* The OOM killer does not needlessly kill tasks for lowmem */
3469 if (ac->highest_zoneidx < ZONE_NORMAL)
3471 if (pm_suspended_storage())
3474 * XXX: GFP_NOFS allocations should rather fail than rely on
3475 * other request to make a forward progress.
3476 * We are in an unfortunate situation where out_of_memory cannot
3477 * do much for this context but let's try it to at least get
3478 * access to memory reserved if the current task is killed (see
3479 * out_of_memory). Once filesystems are ready to handle allocation
3480 * failures more gracefully we should just bail out here.
3483 /* Exhausted what can be done so it's blame time */
3484 if (out_of_memory(&oc) ||
3485 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3486 *did_some_progress = 1;
3489 * Help non-failing allocations by giving them access to memory
3492 if (gfp_mask & __GFP_NOFAIL)
3493 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3494 ALLOC_NO_WATERMARKS, ac);
3497 mutex_unlock(&oom_lock);
3502 * Maximum number of compaction retries with a progress before OOM
3503 * killer is consider as the only way to move forward.
3505 #define MAX_COMPACT_RETRIES 16
3507 #ifdef CONFIG_COMPACTION
3508 /* Try memory compaction for high-order allocations before reclaim */
3509 static struct page *
3510 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3511 unsigned int alloc_flags, const struct alloc_context *ac,
3512 enum compact_priority prio, enum compact_result *compact_result)
3514 struct page *page = NULL;
3515 unsigned long pflags;
3516 unsigned int noreclaim_flag;
3521 psi_memstall_enter(&pflags);
3522 delayacct_compact_start();
3523 noreclaim_flag = memalloc_noreclaim_save();
3525 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3528 memalloc_noreclaim_restore(noreclaim_flag);
3529 psi_memstall_leave(&pflags);
3530 delayacct_compact_end();
3532 if (*compact_result == COMPACT_SKIPPED)
3535 * At least in one zone compaction wasn't deferred or skipped, so let's
3536 * count a compaction stall
3538 count_vm_event(COMPACTSTALL);
3540 /* Prep a captured page if available */
3542 prep_new_page(page, order, gfp_mask, alloc_flags);
3544 /* Try get a page from the freelist if available */
3546 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3549 struct zone *zone = page_zone(page);
3551 zone->compact_blockskip_flush = false;
3552 compaction_defer_reset(zone, order, true);
3553 count_vm_event(COMPACTSUCCESS);
3558 * It's bad if compaction run occurs and fails. The most likely reason
3559 * is that pages exist, but not enough to satisfy watermarks.
3561 count_vm_event(COMPACTFAIL);
3569 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3570 enum compact_result compact_result,
3571 enum compact_priority *compact_priority,
3572 int *compaction_retries)
3574 int max_retries = MAX_COMPACT_RETRIES;
3577 int retries = *compaction_retries;
3578 enum compact_priority priority = *compact_priority;
3583 if (fatal_signal_pending(current))
3587 * Compaction was skipped due to a lack of free order-0
3588 * migration targets. Continue if reclaim can help.
3590 if (compact_result == COMPACT_SKIPPED) {
3591 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3596 * Compaction managed to coalesce some page blocks, but the
3597 * allocation failed presumably due to a race. Retry some.
3599 if (compact_result == COMPACT_SUCCESS) {
3601 * !costly requests are much more important than
3602 * __GFP_RETRY_MAYFAIL costly ones because they are de
3603 * facto nofail and invoke OOM killer to move on while
3604 * costly can fail and users are ready to cope with
3605 * that. 1/4 retries is rather arbitrary but we would
3606 * need much more detailed feedback from compaction to
3607 * make a better decision.
3609 if (order > PAGE_ALLOC_COSTLY_ORDER)
3612 if (++(*compaction_retries) <= max_retries) {
3619 * Compaction failed. Retry with increasing priority.
3621 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3622 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3624 if (*compact_priority > min_priority) {
3625 (*compact_priority)--;
3626 *compaction_retries = 0;
3630 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3634 static inline struct page *
3635 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3636 unsigned int alloc_flags, const struct alloc_context *ac,
3637 enum compact_priority prio, enum compact_result *compact_result)
3639 *compact_result = COMPACT_SKIPPED;
3644 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3645 enum compact_result compact_result,
3646 enum compact_priority *compact_priority,
3647 int *compaction_retries)
3652 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3656 * There are setups with compaction disabled which would prefer to loop
3657 * inside the allocator rather than hit the oom killer prematurely.
3658 * Let's give them a good hope and keep retrying while the order-0
3659 * watermarks are OK.
3661 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3662 ac->highest_zoneidx, ac->nodemask) {
3663 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3664 ac->highest_zoneidx, alloc_flags))
3669 #endif /* CONFIG_COMPACTION */
3671 #ifdef CONFIG_LOCKDEP
3672 static struct lockdep_map __fs_reclaim_map =
3673 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3675 static bool __need_reclaim(gfp_t gfp_mask)
3677 /* no reclaim without waiting on it */
3678 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3681 /* this guy won't enter reclaim */
3682 if (current->flags & PF_MEMALLOC)
3685 if (gfp_mask & __GFP_NOLOCKDEP)
3691 void __fs_reclaim_acquire(unsigned long ip)
3693 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3696 void __fs_reclaim_release(unsigned long ip)
3698 lock_release(&__fs_reclaim_map, ip);
3701 void fs_reclaim_acquire(gfp_t gfp_mask)
3703 gfp_mask = current_gfp_context(gfp_mask);
3705 if (__need_reclaim(gfp_mask)) {
3706 if (gfp_mask & __GFP_FS)
3707 __fs_reclaim_acquire(_RET_IP_);
3709 #ifdef CONFIG_MMU_NOTIFIER
3710 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3711 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3716 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3718 void fs_reclaim_release(gfp_t gfp_mask)
3720 gfp_mask = current_gfp_context(gfp_mask);
3722 if (__need_reclaim(gfp_mask)) {
3723 if (gfp_mask & __GFP_FS)
3724 __fs_reclaim_release(_RET_IP_);
3727 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3731 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3732 * have been rebuilt so allocation retries. Reader side does not lock and
3733 * retries the allocation if zonelist changes. Writer side is protected by the
3734 * embedded spin_lock.
3736 static DEFINE_SEQLOCK(zonelist_update_seq);
3738 static unsigned int zonelist_iter_begin(void)
3740 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3741 return read_seqbegin(&zonelist_update_seq);
3746 static unsigned int check_retry_zonelist(unsigned int seq)
3748 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3749 return read_seqretry(&zonelist_update_seq, seq);
3754 /* Perform direct synchronous page reclaim */
3755 static unsigned long
3756 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3757 const struct alloc_context *ac)
3759 unsigned int noreclaim_flag;
3760 unsigned long progress;
3764 /* We now go into synchronous reclaim */
3765 cpuset_memory_pressure_bump();
3766 fs_reclaim_acquire(gfp_mask);
3767 noreclaim_flag = memalloc_noreclaim_save();
3769 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3772 memalloc_noreclaim_restore(noreclaim_flag);
3773 fs_reclaim_release(gfp_mask);
3780 /* The really slow allocator path where we enter direct reclaim */
3781 static inline struct page *
3782 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3783 unsigned int alloc_flags, const struct alloc_context *ac,
3784 unsigned long *did_some_progress)
3786 struct page *page = NULL;
3787 unsigned long pflags;
3788 bool drained = false;
3790 psi_memstall_enter(&pflags);
3791 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3792 if (unlikely(!(*did_some_progress)))
3796 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3799 * If an allocation failed after direct reclaim, it could be because
3800 * pages are pinned on the per-cpu lists or in high alloc reserves.
3801 * Shrink them and try again
3803 if (!page && !drained) {
3804 unreserve_highatomic_pageblock(ac, false);
3805 drain_all_pages(NULL);
3810 psi_memstall_leave(&pflags);
3815 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3816 const struct alloc_context *ac)
3820 pg_data_t *last_pgdat = NULL;
3821 enum zone_type highest_zoneidx = ac->highest_zoneidx;
3823 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3825 if (!managed_zone(zone))
3827 if (last_pgdat != zone->zone_pgdat) {
3828 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3829 last_pgdat = zone->zone_pgdat;
3834 static inline unsigned int
3835 gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3837 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3840 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3841 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3842 * to save two branches.
3844 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
3845 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
3848 * The caller may dip into page reserves a bit more if the caller
3849 * cannot run direct reclaim, or if the caller has realtime scheduling
3850 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3851 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3853 alloc_flags |= (__force int)
3854 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
3856 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
3858 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3859 * if it can't schedule.
3861 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
3862 alloc_flags |= ALLOC_NON_BLOCK;
3865 alloc_flags |= ALLOC_HIGHATOMIC;
3869 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
3870 * GFP_ATOMIC) rather than fail, see the comment for
3871 * cpuset_node_allowed().
3873 if (alloc_flags & ALLOC_MIN_RESERVE)
3874 alloc_flags &= ~ALLOC_CPUSET;
3875 } else if (unlikely(rt_task(current)) && in_task())
3876 alloc_flags |= ALLOC_MIN_RESERVE;
3878 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
3883 static bool oom_reserves_allowed(struct task_struct *tsk)
3885 if (!tsk_is_oom_victim(tsk))
3889 * !MMU doesn't have oom reaper so give access to memory reserves
3890 * only to the thread with TIF_MEMDIE set
3892 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3899 * Distinguish requests which really need access to full memory
3900 * reserves from oom victims which can live with a portion of it
3902 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3904 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3906 if (gfp_mask & __GFP_MEMALLOC)
3907 return ALLOC_NO_WATERMARKS;
3908 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3909 return ALLOC_NO_WATERMARKS;
3910 if (!in_interrupt()) {
3911 if (current->flags & PF_MEMALLOC)
3912 return ALLOC_NO_WATERMARKS;
3913 else if (oom_reserves_allowed(current))
3920 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3922 return !!__gfp_pfmemalloc_flags(gfp_mask);
3926 * Checks whether it makes sense to retry the reclaim to make a forward progress
3927 * for the given allocation request.
3929 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3930 * without success, or when we couldn't even meet the watermark if we
3931 * reclaimed all remaining pages on the LRU lists.
3933 * Returns true if a retry is viable or false to enter the oom path.
3936 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3937 struct alloc_context *ac, int alloc_flags,
3938 bool did_some_progress, int *no_progress_loops)
3945 * Costly allocations might have made a progress but this doesn't mean
3946 * their order will become available due to high fragmentation so
3947 * always increment the no progress counter for them
3949 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3950 *no_progress_loops = 0;
3952 (*no_progress_loops)++;
3954 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
3959 * Keep reclaiming pages while there is a chance this will lead
3960 * somewhere. If none of the target zones can satisfy our allocation
3961 * request even if all reclaimable pages are considered then we are
3962 * screwed and have to go OOM.
3964 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3965 ac->highest_zoneidx, ac->nodemask) {
3966 unsigned long available;
3967 unsigned long reclaimable;
3968 unsigned long min_wmark = min_wmark_pages(zone);
3971 available = reclaimable = zone_reclaimable_pages(zone);
3972 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3975 * Would the allocation succeed if we reclaimed all
3976 * reclaimable pages?
3978 wmark = __zone_watermark_ok(zone, order, min_wmark,
3979 ac->highest_zoneidx, alloc_flags, available);
3980 trace_reclaim_retry_zone(z, order, reclaimable,
3981 available, min_wmark, *no_progress_loops, wmark);
3989 * Memory allocation/reclaim might be called from a WQ context and the
3990 * current implementation of the WQ concurrency control doesn't
3991 * recognize that a particular WQ is congested if the worker thread is
3992 * looping without ever sleeping. Therefore we have to do a short sleep
3993 * here rather than calling cond_resched().
3995 if (current->flags & PF_WQ_WORKER)
3996 schedule_timeout_uninterruptible(1);
4000 /* Before OOM, exhaust highatomic_reserve */
4002 return unreserve_highatomic_pageblock(ac, true);
4008 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4011 * It's possible that cpuset's mems_allowed and the nodemask from
4012 * mempolicy don't intersect. This should be normally dealt with by
4013 * policy_nodemask(), but it's possible to race with cpuset update in
4014 * such a way the check therein was true, and then it became false
4015 * before we got our cpuset_mems_cookie here.
4016 * This assumes that for all allocations, ac->nodemask can come only
4017 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4018 * when it does not intersect with the cpuset restrictions) or the
4019 * caller can deal with a violated nodemask.
4021 if (cpusets_enabled() && ac->nodemask &&
4022 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4023 ac->nodemask = NULL;
4028 * When updating a task's mems_allowed or mempolicy nodemask, it is
4029 * possible to race with parallel threads in such a way that our
4030 * allocation can fail while the mask is being updated. If we are about
4031 * to fail, check if the cpuset changed during allocation and if so,
4034 if (read_mems_allowed_retry(cpuset_mems_cookie))
4040 static inline struct page *
4041 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4042 struct alloc_context *ac)
4044 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4045 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4046 struct page *page = NULL;
4047 unsigned int alloc_flags;
4048 unsigned long did_some_progress;
4049 enum compact_priority compact_priority;
4050 enum compact_result compact_result;
4051 int compaction_retries;
4052 int no_progress_loops;
4053 unsigned int cpuset_mems_cookie;
4054 unsigned int zonelist_iter_cookie;
4058 compaction_retries = 0;
4059 no_progress_loops = 0;
4060 compact_priority = DEF_COMPACT_PRIORITY;
4061 cpuset_mems_cookie = read_mems_allowed_begin();
4062 zonelist_iter_cookie = zonelist_iter_begin();
4065 * The fast path uses conservative alloc_flags to succeed only until
4066 * kswapd needs to be woken up, and to avoid the cost of setting up
4067 * alloc_flags precisely. So we do that now.
4069 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4072 * We need to recalculate the starting point for the zonelist iterator
4073 * because we might have used different nodemask in the fast path, or
4074 * there was a cpuset modification and we are retrying - otherwise we
4075 * could end up iterating over non-eligible zones endlessly.
4077 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4078 ac->highest_zoneidx, ac->nodemask);
4079 if (!ac->preferred_zoneref->zone)
4083 * Check for insane configurations where the cpuset doesn't contain
4084 * any suitable zone to satisfy the request - e.g. non-movable
4085 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4087 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4088 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4089 ac->highest_zoneidx,
4090 &cpuset_current_mems_allowed);
4095 if (alloc_flags & ALLOC_KSWAPD)
4096 wake_all_kswapds(order, gfp_mask, ac);
4099 * The adjusted alloc_flags might result in immediate success, so try
4102 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4107 * For costly allocations, try direct compaction first, as it's likely
4108 * that we have enough base pages and don't need to reclaim. For non-
4109 * movable high-order allocations, do that as well, as compaction will
4110 * try prevent permanent fragmentation by migrating from blocks of the
4112 * Don't try this for allocations that are allowed to ignore
4113 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4115 if (can_direct_reclaim &&
4117 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4118 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4119 page = __alloc_pages_direct_compact(gfp_mask, order,
4121 INIT_COMPACT_PRIORITY,
4127 * Checks for costly allocations with __GFP_NORETRY, which
4128 * includes some THP page fault allocations
4130 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4132 * If allocating entire pageblock(s) and compaction
4133 * failed because all zones are below low watermarks
4134 * or is prohibited because it recently failed at this
4135 * order, fail immediately unless the allocator has
4136 * requested compaction and reclaim retry.
4139 * - potentially very expensive because zones are far
4140 * below their low watermarks or this is part of very
4141 * bursty high order allocations,
4142 * - not guaranteed to help because isolate_freepages()
4143 * may not iterate over freed pages as part of its
4145 * - unlikely to make entire pageblocks free on its
4148 if (compact_result == COMPACT_SKIPPED ||
4149 compact_result == COMPACT_DEFERRED)
4153 * Looks like reclaim/compaction is worth trying, but
4154 * sync compaction could be very expensive, so keep
4155 * using async compaction.
4157 compact_priority = INIT_COMPACT_PRIORITY;
4162 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4163 if (alloc_flags & ALLOC_KSWAPD)
4164 wake_all_kswapds(order, gfp_mask, ac);
4166 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4168 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4169 (alloc_flags & ALLOC_KSWAPD);
4172 * Reset the nodemask and zonelist iterators if memory policies can be
4173 * ignored. These allocations are high priority and system rather than
4176 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4177 ac->nodemask = NULL;
4178 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4179 ac->highest_zoneidx, ac->nodemask);
4182 /* Attempt with potentially adjusted zonelist and alloc_flags */
4183 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4187 /* Caller is not willing to reclaim, we can't balance anything */
4188 if (!can_direct_reclaim)
4191 /* Avoid recursion of direct reclaim */
4192 if (current->flags & PF_MEMALLOC)
4195 /* Try direct reclaim and then allocating */
4196 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4197 &did_some_progress);
4201 /* Try direct compaction and then allocating */
4202 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4203 compact_priority, &compact_result);
4207 /* Do not loop if specifically requested */
4208 if (gfp_mask & __GFP_NORETRY)
4212 * Do not retry costly high order allocations unless they are
4213 * __GFP_RETRY_MAYFAIL
4215 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4218 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4219 did_some_progress > 0, &no_progress_loops))
4223 * It doesn't make any sense to retry for the compaction if the order-0
4224 * reclaim is not able to make any progress because the current
4225 * implementation of the compaction depends on the sufficient amount
4226 * of free memory (see __compaction_suitable)
4228 if (did_some_progress > 0 &&
4229 should_compact_retry(ac, order, alloc_flags,
4230 compact_result, &compact_priority,
4231 &compaction_retries))
4236 * Deal with possible cpuset update races or zonelist updates to avoid
4237 * a unnecessary OOM kill.
4239 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4240 check_retry_zonelist(zonelist_iter_cookie))
4243 /* Reclaim has failed us, start killing things */
4244 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4248 /* Avoid allocations with no watermarks from looping endlessly */
4249 if (tsk_is_oom_victim(current) &&
4250 (alloc_flags & ALLOC_OOM ||
4251 (gfp_mask & __GFP_NOMEMALLOC)))
4254 /* Retry as long as the OOM killer is making progress */
4255 if (did_some_progress) {
4256 no_progress_loops = 0;
4262 * Deal with possible cpuset update races or zonelist updates to avoid
4263 * a unnecessary OOM kill.
4265 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4266 check_retry_zonelist(zonelist_iter_cookie))
4270 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4273 if (gfp_mask & __GFP_NOFAIL) {
4275 * All existing users of the __GFP_NOFAIL are blockable, so warn
4276 * of any new users that actually require GFP_NOWAIT
4278 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4282 * PF_MEMALLOC request from this context is rather bizarre
4283 * because we cannot reclaim anything and only can loop waiting
4284 * for somebody to do a work for us
4286 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4289 * non failing costly orders are a hard requirement which we
4290 * are not prepared for much so let's warn about these users
4291 * so that we can identify them and convert them to something
4294 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4297 * Help non-failing allocations by giving some access to memory
4298 * reserves normally used for high priority non-blocking
4299 * allocations but do not use ALLOC_NO_WATERMARKS because this
4300 * could deplete whole memory reserves which would just make
4301 * the situation worse.
4303 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4311 warn_alloc(gfp_mask, ac->nodemask,
4312 "page allocation failure: order:%u", order);
4317 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4318 int preferred_nid, nodemask_t *nodemask,
4319 struct alloc_context *ac, gfp_t *alloc_gfp,
4320 unsigned int *alloc_flags)
4322 ac->highest_zoneidx = gfp_zone(gfp_mask);
4323 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4324 ac->nodemask = nodemask;
4325 ac->migratetype = gfp_migratetype(gfp_mask);
4327 if (cpusets_enabled()) {
4328 *alloc_gfp |= __GFP_HARDWALL;
4330 * When we are in the interrupt context, it is irrelevant
4331 * to the current task context. It means that any node ok.
4333 if (in_task() && !ac->nodemask)
4334 ac->nodemask = &cpuset_current_mems_allowed;
4336 *alloc_flags |= ALLOC_CPUSET;
4339 might_alloc(gfp_mask);
4341 if (should_fail_alloc_page(gfp_mask, order))
4344 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4346 /* Dirty zone balancing only done in the fast path */
4347 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4350 * The preferred zone is used for statistics but crucially it is
4351 * also used as the starting point for the zonelist iterator. It
4352 * may get reset for allocations that ignore memory policies.
4354 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4355 ac->highest_zoneidx, ac->nodemask);
4361 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4362 * @gfp: GFP flags for the allocation
4363 * @preferred_nid: The preferred NUMA node ID to allocate from
4364 * @nodemask: Set of nodes to allocate from, may be NULL
4365 * @nr_pages: The number of pages desired on the list or array
4366 * @page_list: Optional list to store the allocated pages
4367 * @page_array: Optional array to store the pages
4369 * This is a batched version of the page allocator that attempts to
4370 * allocate nr_pages quickly. Pages are added to page_list if page_list
4371 * is not NULL, otherwise it is assumed that the page_array is valid.
4373 * For lists, nr_pages is the number of pages that should be allocated.
4375 * For arrays, only NULL elements are populated with pages and nr_pages
4376 * is the maximum number of pages that will be stored in the array.
4378 * Returns the number of pages on the list or array.
4380 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
4381 nodemask_t *nodemask, int nr_pages,
4382 struct list_head *page_list,
4383 struct page **page_array)
4386 unsigned long __maybe_unused UP_flags;
4389 struct per_cpu_pages *pcp;
4390 struct list_head *pcp_list;
4391 struct alloc_context ac;
4393 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4394 int nr_populated = 0, nr_account = 0;
4397 * Skip populated array elements to determine if any pages need
4398 * to be allocated before disabling IRQs.
4400 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4403 /* No pages requested? */
4404 if (unlikely(nr_pages <= 0))
4407 /* Already populated array? */
4408 if (unlikely(page_array && nr_pages - nr_populated == 0))
4411 /* Bulk allocator does not support memcg accounting. */
4412 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4415 /* Use the single page allocator for one page. */
4416 if (nr_pages - nr_populated == 1)
4419 #ifdef CONFIG_PAGE_OWNER
4421 * PAGE_OWNER may recurse into the allocator to allocate space to
4422 * save the stack with pagesets.lock held. Releasing/reacquiring
4423 * removes much of the performance benefit of bulk allocation so
4424 * force the caller to allocate one page at a time as it'll have
4425 * similar performance to added complexity to the bulk allocator.
4427 if (static_branch_unlikely(&page_owner_inited))
4431 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4432 gfp &= gfp_allowed_mask;
4434 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4438 /* Find an allowed local zone that meets the low watermark. */
4439 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4442 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4443 !__cpuset_zone_allowed(zone, gfp)) {
4447 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4448 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4452 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4453 if (zone_watermark_fast(zone, 0, mark,
4454 zonelist_zone_idx(ac.preferred_zoneref),
4455 alloc_flags, gfp)) {
4461 * If there are no allowed local zones that meets the watermarks then
4462 * try to allocate a single page and reclaim if necessary.
4464 if (unlikely(!zone))
4467 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4468 pcp_trylock_prepare(UP_flags);
4469 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4473 /* Attempt the batch allocation */
4474 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4475 while (nr_populated < nr_pages) {
4477 /* Skip existing pages */
4478 if (page_array && page_array[nr_populated]) {
4483 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4485 if (unlikely(!page)) {
4486 /* Try and allocate at least one page */
4488 pcp_spin_unlock(pcp);
4495 prep_new_page(page, 0, gfp, 0);
4497 list_add(&page->lru, page_list);
4499 page_array[nr_populated] = page;
4503 pcp_spin_unlock(pcp);
4504 pcp_trylock_finish(UP_flags);
4506 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4507 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4510 return nr_populated;
4513 pcp_trylock_finish(UP_flags);
4516 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
4519 list_add(&page->lru, page_list);
4521 page_array[nr_populated] = page;
4527 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
4530 * This is the 'heart' of the zoned buddy allocator.
4532 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
4533 nodemask_t *nodemask)
4536 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4537 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4538 struct alloc_context ac = { };
4541 * There are several places where we assume that the order value is sane
4542 * so bail out early if the request is out of bound.
4544 if (WARN_ON_ONCE_GFP(order > MAX_ORDER, gfp))
4547 gfp &= gfp_allowed_mask;
4549 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4550 * resp. GFP_NOIO which has to be inherited for all allocation requests
4551 * from a particular context which has been marked by
4552 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4553 * movable zones are not used during allocation.
4555 gfp = current_gfp_context(gfp);
4557 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4558 &alloc_gfp, &alloc_flags))
4562 * Forbid the first pass from falling back to types that fragment
4563 * memory until all local zones are considered.
4565 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4567 /* First allocation attempt */
4568 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4573 ac.spread_dirty_pages = false;
4576 * Restore the original nodemask if it was potentially replaced with
4577 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4579 ac.nodemask = nodemask;
4581 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4584 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4585 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4586 __free_pages(page, order);
4590 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4591 kmsan_alloc_page(page, order, alloc_gfp);
4595 EXPORT_SYMBOL(__alloc_pages);
4597 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
4598 nodemask_t *nodemask)
4600 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
4601 preferred_nid, nodemask);
4602 return page_rmappable_folio(page);
4604 EXPORT_SYMBOL(__folio_alloc);
4607 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4608 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4609 * you need to access high mem.
4611 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4615 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4618 return (unsigned long) page_address(page);
4620 EXPORT_SYMBOL(__get_free_pages);
4622 unsigned long get_zeroed_page(gfp_t gfp_mask)
4624 return __get_free_page(gfp_mask | __GFP_ZERO);
4626 EXPORT_SYMBOL(get_zeroed_page);
4629 * __free_pages - Free pages allocated with alloc_pages().
4630 * @page: The page pointer returned from alloc_pages().
4631 * @order: The order of the allocation.
4633 * This function can free multi-page allocations that are not compound
4634 * pages. It does not check that the @order passed in matches that of
4635 * the allocation, so it is easy to leak memory. Freeing more memory
4636 * than was allocated will probably emit a warning.
4638 * If the last reference to this page is speculative, it will be released
4639 * by put_page() which only frees the first page of a non-compound
4640 * allocation. To prevent the remaining pages from being leaked, we free
4641 * the subsequent pages here. If you want to use the page's reference
4642 * count to decide when to free the allocation, you should allocate a
4643 * compound page, and use put_page() instead of __free_pages().
4645 * Context: May be called in interrupt context or while holding a normal
4646 * spinlock, but not in NMI context or while holding a raw spinlock.
4648 void __free_pages(struct page *page, unsigned int order)
4650 /* get PageHead before we drop reference */
4651 int head = PageHead(page);
4653 if (put_page_testzero(page))
4654 free_the_page(page, order);
4657 free_the_page(page + (1 << order), order);
4659 EXPORT_SYMBOL(__free_pages);
4661 void free_pages(unsigned long addr, unsigned int order)
4664 VM_BUG_ON(!virt_addr_valid((void *)addr));
4665 __free_pages(virt_to_page((void *)addr), order);
4669 EXPORT_SYMBOL(free_pages);
4673 * An arbitrary-length arbitrary-offset area of memory which resides
4674 * within a 0 or higher order page. Multiple fragments within that page
4675 * are individually refcounted, in the page's reference counter.
4677 * The page_frag functions below provide a simple allocation framework for
4678 * page fragments. This is used by the network stack and network device
4679 * drivers to provide a backing region of memory for use as either an
4680 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4682 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4685 struct page *page = NULL;
4686 gfp_t gfp = gfp_mask;
4688 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4689 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4691 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4692 PAGE_FRAG_CACHE_MAX_ORDER);
4693 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4695 if (unlikely(!page))
4696 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4698 nc->va = page ? page_address(page) : NULL;
4703 void __page_frag_cache_drain(struct page *page, unsigned int count)
4705 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4707 if (page_ref_sub_and_test(page, count))
4708 free_the_page(page, compound_order(page));
4710 EXPORT_SYMBOL(__page_frag_cache_drain);
4712 void *page_frag_alloc_align(struct page_frag_cache *nc,
4713 unsigned int fragsz, gfp_t gfp_mask,
4714 unsigned int align_mask)
4716 unsigned int size = PAGE_SIZE;
4720 if (unlikely(!nc->va)) {
4722 page = __page_frag_cache_refill(nc, gfp_mask);
4726 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4727 /* if size can vary use size else just use PAGE_SIZE */
4730 /* Even if we own the page, we do not use atomic_set().
4731 * This would break get_page_unless_zero() users.
4733 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4735 /* reset page count bias and offset to start of new frag */
4736 nc->pfmemalloc = page_is_pfmemalloc(page);
4737 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4741 offset = nc->offset - fragsz;
4742 if (unlikely(offset < 0)) {
4743 page = virt_to_page(nc->va);
4745 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4748 if (unlikely(nc->pfmemalloc)) {
4749 free_the_page(page, compound_order(page));
4753 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4754 /* if size can vary use size else just use PAGE_SIZE */
4757 /* OK, page count is 0, we can safely set it */
4758 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4760 /* reset page count bias and offset to start of new frag */
4761 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4762 offset = size - fragsz;
4763 if (unlikely(offset < 0)) {
4765 * The caller is trying to allocate a fragment
4766 * with fragsz > PAGE_SIZE but the cache isn't big
4767 * enough to satisfy the request, this may
4768 * happen in low memory conditions.
4769 * We don't release the cache page because
4770 * it could make memory pressure worse
4771 * so we simply return NULL here.
4778 offset &= align_mask;
4779 nc->offset = offset;
4781 return nc->va + offset;
4783 EXPORT_SYMBOL(page_frag_alloc_align);
4786 * Frees a page fragment allocated out of either a compound or order 0 page.
4788 void page_frag_free(void *addr)
4790 struct page *page = virt_to_head_page(addr);
4792 if (unlikely(put_page_testzero(page)))
4793 free_the_page(page, compound_order(page));
4795 EXPORT_SYMBOL(page_frag_free);
4797 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4801 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4802 struct page *page = virt_to_page((void *)addr);
4803 struct page *last = page + nr;
4805 split_page_owner(page, 1 << order);
4806 split_page_memcg(page, 1 << order);
4807 while (page < --last)
4808 set_page_refcounted(last);
4810 last = page + (1UL << order);
4811 for (page += nr; page < last; page++)
4812 __free_pages_ok(page, 0, FPI_TO_TAIL);
4814 return (void *)addr;
4818 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4819 * @size: the number of bytes to allocate
4820 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4822 * This function is similar to alloc_pages(), except that it allocates the
4823 * minimum number of pages to satisfy the request. alloc_pages() can only
4824 * allocate memory in power-of-two pages.
4826 * This function is also limited by MAX_ORDER.
4828 * Memory allocated by this function must be released by free_pages_exact().
4830 * Return: pointer to the allocated area or %NULL in case of error.
4832 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4834 unsigned int order = get_order(size);
4837 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4838 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4840 addr = __get_free_pages(gfp_mask, order);
4841 return make_alloc_exact(addr, order, size);
4843 EXPORT_SYMBOL(alloc_pages_exact);
4846 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4848 * @nid: the preferred node ID where memory should be allocated
4849 * @size: the number of bytes to allocate
4850 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4852 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4855 * Return: pointer to the allocated area or %NULL in case of error.
4857 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4859 unsigned int order = get_order(size);
4862 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4863 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4865 p = alloc_pages_node(nid, gfp_mask, order);
4868 return make_alloc_exact((unsigned long)page_address(p), order, size);
4872 * free_pages_exact - release memory allocated via alloc_pages_exact()
4873 * @virt: the value returned by alloc_pages_exact.
4874 * @size: size of allocation, same value as passed to alloc_pages_exact().
4876 * Release the memory allocated by a previous call to alloc_pages_exact.
4878 void free_pages_exact(void *virt, size_t size)
4880 unsigned long addr = (unsigned long)virt;
4881 unsigned long end = addr + PAGE_ALIGN(size);
4883 while (addr < end) {
4888 EXPORT_SYMBOL(free_pages_exact);
4891 * nr_free_zone_pages - count number of pages beyond high watermark
4892 * @offset: The zone index of the highest zone
4894 * nr_free_zone_pages() counts the number of pages which are beyond the
4895 * high watermark within all zones at or below a given zone index. For each
4896 * zone, the number of pages is calculated as:
4898 * nr_free_zone_pages = managed_pages - high_pages
4900 * Return: number of pages beyond high watermark.
4902 static unsigned long nr_free_zone_pages(int offset)
4907 /* Just pick one node, since fallback list is circular */
4908 unsigned long sum = 0;
4910 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4912 for_each_zone_zonelist(zone, z, zonelist, offset) {
4913 unsigned long size = zone_managed_pages(zone);
4914 unsigned long high = high_wmark_pages(zone);
4923 * nr_free_buffer_pages - count number of pages beyond high watermark
4925 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4926 * watermark within ZONE_DMA and ZONE_NORMAL.
4928 * Return: number of pages beyond high watermark within ZONE_DMA and
4931 unsigned long nr_free_buffer_pages(void)
4933 return nr_free_zone_pages(gfp_zone(GFP_USER));
4935 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4937 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4939 zoneref->zone = zone;
4940 zoneref->zone_idx = zone_idx(zone);
4944 * Builds allocation fallback zone lists.
4946 * Add all populated zones of a node to the zonelist.
4948 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4951 enum zone_type zone_type = MAX_NR_ZONES;
4956 zone = pgdat->node_zones + zone_type;
4957 if (populated_zone(zone)) {
4958 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4959 check_highest_zone(zone_type);
4961 } while (zone_type);
4968 static int __parse_numa_zonelist_order(char *s)
4971 * We used to support different zonelists modes but they turned
4972 * out to be just not useful. Let's keep the warning in place
4973 * if somebody still use the cmd line parameter so that we do
4974 * not fail it silently
4976 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4977 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4983 static char numa_zonelist_order[] = "Node";
4984 #define NUMA_ZONELIST_ORDER_LEN 16
4986 * sysctl handler for numa_zonelist_order
4988 static int numa_zonelist_order_handler(struct ctl_table *table, int write,
4989 void *buffer, size_t *length, loff_t *ppos)
4992 return __parse_numa_zonelist_order(buffer);
4993 return proc_dostring(table, write, buffer, length, ppos);
4996 static int node_load[MAX_NUMNODES];
4999 * find_next_best_node - find the next node that should appear in a given node's fallback list
5000 * @node: node whose fallback list we're appending
5001 * @used_node_mask: nodemask_t of already used nodes
5003 * We use a number of factors to determine which is the next node that should
5004 * appear on a given node's fallback list. The node should not have appeared
5005 * already in @node's fallback list, and it should be the next closest node
5006 * according to the distance array (which contains arbitrary distance values
5007 * from each node to each node in the system), and should also prefer nodes
5008 * with no CPUs, since presumably they'll have very little allocation pressure
5009 * on them otherwise.
5011 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5013 int find_next_best_node(int node, nodemask_t *used_node_mask)
5016 int min_val = INT_MAX;
5017 int best_node = NUMA_NO_NODE;
5020 * Use the local node if we haven't already, but for memoryless local
5021 * node, we should skip it and fall back to other nodes.
5023 if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5024 node_set(node, *used_node_mask);
5028 for_each_node_state(n, N_MEMORY) {
5030 /* Don't want a node to appear more than once */
5031 if (node_isset(n, *used_node_mask))
5034 /* Use the distance array to find the distance */
5035 val = node_distance(node, n);
5037 /* Penalize nodes under us ("prefer the next node") */
5040 /* Give preference to headless and unused nodes */
5041 if (!cpumask_empty(cpumask_of_node(n)))
5042 val += PENALTY_FOR_NODE_WITH_CPUS;
5044 /* Slight preference for less loaded node */
5045 val *= MAX_NUMNODES;
5046 val += node_load[n];
5048 if (val < min_val) {
5055 node_set(best_node, *used_node_mask);
5062 * Build zonelists ordered by node and zones within node.
5063 * This results in maximum locality--normal zone overflows into local
5064 * DMA zone, if any--but risks exhausting DMA zone.
5066 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5069 struct zoneref *zonerefs;
5072 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5074 for (i = 0; i < nr_nodes; i++) {
5077 pg_data_t *node = NODE_DATA(node_order[i]);
5079 nr_zones = build_zonerefs_node(node, zonerefs);
5080 zonerefs += nr_zones;
5082 zonerefs->zone = NULL;
5083 zonerefs->zone_idx = 0;
5087 * Build gfp_thisnode zonelists
5089 static void build_thisnode_zonelists(pg_data_t *pgdat)
5091 struct zoneref *zonerefs;
5094 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5095 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5096 zonerefs += nr_zones;
5097 zonerefs->zone = NULL;
5098 zonerefs->zone_idx = 0;
5102 * Build zonelists ordered by zone and nodes within zones.
5103 * This results in conserving DMA zone[s] until all Normal memory is
5104 * exhausted, but results in overflowing to remote node while memory
5105 * may still exist in local DMA zone.
5108 static void build_zonelists(pg_data_t *pgdat)
5110 static int node_order[MAX_NUMNODES];
5111 int node, nr_nodes = 0;
5112 nodemask_t used_mask = NODE_MASK_NONE;
5113 int local_node, prev_node;
5115 /* NUMA-aware ordering of nodes */
5116 local_node = pgdat->node_id;
5117 prev_node = local_node;
5119 memset(node_order, 0, sizeof(node_order));
5120 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5122 * We don't want to pressure a particular node.
5123 * So adding penalty to the first node in same
5124 * distance group to make it round-robin.
5126 if (node_distance(local_node, node) !=
5127 node_distance(local_node, prev_node))
5128 node_load[node] += 1;
5130 node_order[nr_nodes++] = node;
5134 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5135 build_thisnode_zonelists(pgdat);
5136 pr_info("Fallback order for Node %d: ", local_node);
5137 for (node = 0; node < nr_nodes; node++)
5138 pr_cont("%d ", node_order[node]);
5142 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5144 * Return node id of node used for "local" allocations.
5145 * I.e., first node id of first zone in arg node's generic zonelist.
5146 * Used for initializing percpu 'numa_mem', which is used primarily
5147 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5149 int local_memory_node(int node)
5153 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5154 gfp_zone(GFP_KERNEL),
5156 return zone_to_nid(z->zone);
5160 static void setup_min_unmapped_ratio(void);
5161 static void setup_min_slab_ratio(void);
5162 #else /* CONFIG_NUMA */
5164 static void build_zonelists(pg_data_t *pgdat)
5166 int node, local_node;
5167 struct zoneref *zonerefs;
5170 local_node = pgdat->node_id;
5172 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5173 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5174 zonerefs += nr_zones;
5177 * Now we build the zonelist so that it contains the zones
5178 * of all the other nodes.
5179 * We don't want to pressure a particular node, so when
5180 * building the zones for node N, we make sure that the
5181 * zones coming right after the local ones are those from
5182 * node N+1 (modulo N)
5184 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5185 if (!node_online(node))
5187 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5188 zonerefs += nr_zones;
5190 for (node = 0; node < local_node; node++) {
5191 if (!node_online(node))
5193 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5194 zonerefs += nr_zones;
5197 zonerefs->zone = NULL;
5198 zonerefs->zone_idx = 0;
5201 #endif /* CONFIG_NUMA */
5204 * Boot pageset table. One per cpu which is going to be used for all
5205 * zones and all nodes. The parameters will be set in such a way
5206 * that an item put on a list will immediately be handed over to
5207 * the buddy list. This is safe since pageset manipulation is done
5208 * with interrupts disabled.
5210 * The boot_pagesets must be kept even after bootup is complete for
5211 * unused processors and/or zones. They do play a role for bootstrapping
5212 * hotplugged processors.
5214 * zoneinfo_show() and maybe other functions do
5215 * not check if the processor is online before following the pageset pointer.
5216 * Other parts of the kernel may not check if the zone is available.
5218 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5219 /* These effectively disable the pcplists in the boot pageset completely */
5220 #define BOOT_PAGESET_HIGH 0
5221 #define BOOT_PAGESET_BATCH 1
5222 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5223 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5225 static void __build_all_zonelists(void *data)
5228 int __maybe_unused cpu;
5229 pg_data_t *self = data;
5230 unsigned long flags;
5233 * The zonelist_update_seq must be acquired with irqsave because the
5234 * reader can be invoked from IRQ with GFP_ATOMIC.
5236 write_seqlock_irqsave(&zonelist_update_seq, flags);
5238 * Also disable synchronous printk() to prevent any printk() from
5239 * trying to hold port->lock, for
5240 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5241 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5243 printk_deferred_enter();
5246 memset(node_load, 0, sizeof(node_load));
5250 * This node is hotadded and no memory is yet present. So just
5251 * building zonelists is fine - no need to touch other nodes.
5253 if (self && !node_online(self->node_id)) {
5254 build_zonelists(self);
5257 * All possible nodes have pgdat preallocated
5260 for_each_node(nid) {
5261 pg_data_t *pgdat = NODE_DATA(nid);
5263 build_zonelists(pgdat);
5266 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5268 * We now know the "local memory node" for each node--
5269 * i.e., the node of the first zone in the generic zonelist.
5270 * Set up numa_mem percpu variable for on-line cpus. During
5271 * boot, only the boot cpu should be on-line; we'll init the
5272 * secondary cpus' numa_mem as they come on-line. During
5273 * node/memory hotplug, we'll fixup all on-line cpus.
5275 for_each_online_cpu(cpu)
5276 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5280 printk_deferred_exit();
5281 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5284 static noinline void __init
5285 build_all_zonelists_init(void)
5289 __build_all_zonelists(NULL);
5292 * Initialize the boot_pagesets that are going to be used
5293 * for bootstrapping processors. The real pagesets for
5294 * each zone will be allocated later when the per cpu
5295 * allocator is available.
5297 * boot_pagesets are used also for bootstrapping offline
5298 * cpus if the system is already booted because the pagesets
5299 * are needed to initialize allocators on a specific cpu too.
5300 * F.e. the percpu allocator needs the page allocator which
5301 * needs the percpu allocator in order to allocate its pagesets
5302 * (a chicken-egg dilemma).
5304 for_each_possible_cpu(cpu)
5305 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5307 mminit_verify_zonelist();
5308 cpuset_init_current_mems_allowed();
5312 * unless system_state == SYSTEM_BOOTING.
5314 * __ref due to call of __init annotated helper build_all_zonelists_init
5315 * [protected by SYSTEM_BOOTING].
5317 void __ref build_all_zonelists(pg_data_t *pgdat)
5319 unsigned long vm_total_pages;
5321 if (system_state == SYSTEM_BOOTING) {
5322 build_all_zonelists_init();
5324 __build_all_zonelists(pgdat);
5325 /* cpuset refresh routine should be here */
5327 /* Get the number of free pages beyond high watermark in all zones. */
5328 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5330 * Disable grouping by mobility if the number of pages in the
5331 * system is too low to allow the mechanism to work. It would be
5332 * more accurate, but expensive to check per-zone. This check is
5333 * made on memory-hotadd so a system can start with mobility
5334 * disabled and enable it later
5336 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5337 page_group_by_mobility_disabled = 1;
5339 page_group_by_mobility_disabled = 0;
5341 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5343 page_group_by_mobility_disabled ? "off" : "on",
5346 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5350 static int zone_batchsize(struct zone *zone)
5356 * The number of pages to batch allocate is either ~0.1%
5357 * of the zone or 1MB, whichever is smaller. The batch
5358 * size is striking a balance between allocation latency
5359 * and zone lock contention.
5361 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5362 batch /= 4; /* We effectively *= 4 below */
5367 * Clamp the batch to a 2^n - 1 value. Having a power
5368 * of 2 value was found to be more likely to have
5369 * suboptimal cache aliasing properties in some cases.
5371 * For example if 2 tasks are alternately allocating
5372 * batches of pages, one task can end up with a lot
5373 * of pages of one half of the possible page colors
5374 * and the other with pages of the other colors.
5376 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5381 /* The deferral and batching of frees should be suppressed under NOMMU
5384 * The problem is that NOMMU needs to be able to allocate large chunks
5385 * of contiguous memory as there's no hardware page translation to
5386 * assemble apparent contiguous memory from discontiguous pages.
5388 * Queueing large contiguous runs of pages for batching, however,
5389 * causes the pages to actually be freed in smaller chunks. As there
5390 * can be a significant delay between the individual batches being
5391 * recycled, this leads to the once large chunks of space being
5392 * fragmented and becoming unavailable for high-order allocations.
5398 static int percpu_pagelist_high_fraction;
5399 static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5405 unsigned long total_pages;
5407 if (!high_fraction) {
5409 * By default, the high value of the pcp is based on the zone
5410 * low watermark so that if they are full then background
5411 * reclaim will not be started prematurely.
5413 total_pages = low_wmark_pages(zone);
5416 * If percpu_pagelist_high_fraction is configured, the high
5417 * value is based on a fraction of the managed pages in the
5420 total_pages = zone_managed_pages(zone) / high_fraction;
5424 * Split the high value across all online CPUs local to the zone. Note
5425 * that early in boot that CPUs may not be online yet and that during
5426 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5427 * onlined. For memory nodes that have no CPUs, split the high value
5428 * across all online CPUs to mitigate the risk that reclaim is triggered
5429 * prematurely due to pages stored on pcp lists.
5431 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5433 nr_split_cpus = num_online_cpus();
5434 high = total_pages / nr_split_cpus;
5437 * Ensure high is at least batch*4. The multiple is based on the
5438 * historical relationship between high and batch.
5440 high = max(high, batch << 2);
5449 * pcp->high and pcp->batch values are related and generally batch is lower
5450 * than high. They are also related to pcp->count such that count is lower
5451 * than high, and as soon as it reaches high, the pcplist is flushed.
5453 * However, guaranteeing these relations at all times would require e.g. write
5454 * barriers here but also careful usage of read barriers at the read side, and
5455 * thus be prone to error and bad for performance. Thus the update only prevents
5456 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5457 * should ensure they can cope with those fields changing asynchronously, and
5458 * fully trust only the pcp->count field on the local CPU with interrupts
5461 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5462 * outside of boot time (or some other assurance that no concurrent updaters
5465 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5466 unsigned long high_max, unsigned long batch)
5468 WRITE_ONCE(pcp->batch, batch);
5469 WRITE_ONCE(pcp->high_min, high_min);
5470 WRITE_ONCE(pcp->high_max, high_max);
5473 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5477 memset(pcp, 0, sizeof(*pcp));
5478 memset(pzstats, 0, sizeof(*pzstats));
5480 spin_lock_init(&pcp->lock);
5481 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5482 INIT_LIST_HEAD(&pcp->lists[pindex]);
5485 * Set batch and high values safe for a boot pageset. A true percpu
5486 * pageset's initialization will update them subsequently. Here we don't
5487 * need to be as careful as pageset_update() as nobody can access the
5490 pcp->high_min = BOOT_PAGESET_HIGH;
5491 pcp->high_max = BOOT_PAGESET_HIGH;
5492 pcp->batch = BOOT_PAGESET_BATCH;
5493 pcp->free_count = 0;
5496 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
5497 unsigned long high_max, unsigned long batch)
5499 struct per_cpu_pages *pcp;
5502 for_each_possible_cpu(cpu) {
5503 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5504 pageset_update(pcp, high_min, high_max, batch);
5509 * Calculate and set new high and batch values for all per-cpu pagesets of a
5510 * zone based on the zone's size.
5512 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5514 int new_high_min, new_high_max, new_batch;
5516 new_batch = max(1, zone_batchsize(zone));
5517 if (percpu_pagelist_high_fraction) {
5518 new_high_min = zone_highsize(zone, new_batch, cpu_online,
5519 percpu_pagelist_high_fraction);
5521 * PCP high is tuned manually, disable auto-tuning via
5522 * setting high_min and high_max to the manual value.
5524 new_high_max = new_high_min;
5526 new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
5527 new_high_max = zone_highsize(zone, new_batch, cpu_online,
5528 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
5531 if (zone->pageset_high_min == new_high_min &&
5532 zone->pageset_high_max == new_high_max &&
5533 zone->pageset_batch == new_batch)
5536 zone->pageset_high_min = new_high_min;
5537 zone->pageset_high_max = new_high_max;
5538 zone->pageset_batch = new_batch;
5540 __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
5544 void __meminit setup_zone_pageset(struct zone *zone)
5548 /* Size may be 0 on !SMP && !NUMA */
5549 if (sizeof(struct per_cpu_zonestat) > 0)
5550 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5552 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5553 for_each_possible_cpu(cpu) {
5554 struct per_cpu_pages *pcp;
5555 struct per_cpu_zonestat *pzstats;
5557 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5558 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5559 per_cpu_pages_init(pcp, pzstats);
5562 zone_set_pageset_high_and_batch(zone, 0);
5566 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5567 * page high values need to be recalculated.
5569 static void zone_pcp_update(struct zone *zone, int cpu_online)
5571 mutex_lock(&pcp_batch_high_lock);
5572 zone_set_pageset_high_and_batch(zone, cpu_online);
5573 mutex_unlock(&pcp_batch_high_lock);
5576 static void zone_pcp_update_cacheinfo(struct zone *zone)
5579 struct per_cpu_pages *pcp;
5580 struct cpu_cacheinfo *cci;
5582 for_each_online_cpu(cpu) {
5583 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5584 cci = get_cpu_cacheinfo(cpu);
5586 * If data cache slice of CPU is large enough, "pcp->batch"
5587 * pages can be preserved in PCP before draining PCP for
5588 * consecutive high-order pages freeing without allocation.
5589 * This can reduce zone lock contention without hurting
5590 * cache-hot pages sharing.
5592 spin_lock(&pcp->lock);
5593 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
5594 pcp->flags |= PCPF_FREE_HIGH_BATCH;
5596 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
5597 spin_unlock(&pcp->lock);
5601 void setup_pcp_cacheinfo(void)
5605 for_each_populated_zone(zone)
5606 zone_pcp_update_cacheinfo(zone);
5610 * Allocate per cpu pagesets and initialize them.
5611 * Before this call only boot pagesets were available.
5613 void __init setup_per_cpu_pageset(void)
5615 struct pglist_data *pgdat;
5617 int __maybe_unused cpu;
5619 for_each_populated_zone(zone)
5620 setup_zone_pageset(zone);
5624 * Unpopulated zones continue using the boot pagesets.
5625 * The numa stats for these pagesets need to be reset.
5626 * Otherwise, they will end up skewing the stats of
5627 * the nodes these zones are associated with.
5629 for_each_possible_cpu(cpu) {
5630 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5631 memset(pzstats->vm_numa_event, 0,
5632 sizeof(pzstats->vm_numa_event));
5636 for_each_online_pgdat(pgdat)
5637 pgdat->per_cpu_nodestats =
5638 alloc_percpu(struct per_cpu_nodestat);
5641 __meminit void zone_pcp_init(struct zone *zone)
5644 * per cpu subsystem is not up at this point. The following code
5645 * relies on the ability of the linker to provide the
5646 * offset of a (static) per cpu variable into the per cpu area.
5648 zone->per_cpu_pageset = &boot_pageset;
5649 zone->per_cpu_zonestats = &boot_zonestats;
5650 zone->pageset_high_min = BOOT_PAGESET_HIGH;
5651 zone->pageset_high_max = BOOT_PAGESET_HIGH;
5652 zone->pageset_batch = BOOT_PAGESET_BATCH;
5654 if (populated_zone(zone))
5655 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5656 zone->present_pages, zone_batchsize(zone));
5659 void adjust_managed_page_count(struct page *page, long count)
5661 atomic_long_add(count, &page_zone(page)->managed_pages);
5662 totalram_pages_add(count);
5663 #ifdef CONFIG_HIGHMEM
5664 if (PageHighMem(page))
5665 totalhigh_pages_add(count);
5668 EXPORT_SYMBOL(adjust_managed_page_count);
5670 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5673 unsigned long pages = 0;
5675 start = (void *)PAGE_ALIGN((unsigned long)start);
5676 end = (void *)((unsigned long)end & PAGE_MASK);
5677 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5678 struct page *page = virt_to_page(pos);
5679 void *direct_map_addr;
5682 * 'direct_map_addr' might be different from 'pos'
5683 * because some architectures' virt_to_page()
5684 * work with aliases. Getting the direct map
5685 * address ensures that we get a _writeable_
5686 * alias for the memset().
5688 direct_map_addr = page_address(page);
5690 * Perform a kasan-unchecked memset() since this memory
5691 * has not been initialized.
5693 direct_map_addr = kasan_reset_tag(direct_map_addr);
5694 if ((unsigned int)poison <= 0xFF)
5695 memset(direct_map_addr, poison, PAGE_SIZE);
5697 free_reserved_page(page);
5701 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5706 static int page_alloc_cpu_dead(unsigned int cpu)
5710 lru_add_drain_cpu(cpu);
5711 mlock_drain_remote(cpu);
5715 * Spill the event counters of the dead processor
5716 * into the current processors event counters.
5717 * This artificially elevates the count of the current
5720 vm_events_fold_cpu(cpu);
5723 * Zero the differential counters of the dead processor
5724 * so that the vm statistics are consistent.
5726 * This is only okay since the processor is dead and cannot
5727 * race with what we are doing.
5729 cpu_vm_stats_fold(cpu);
5731 for_each_populated_zone(zone)
5732 zone_pcp_update(zone, 0);
5737 static int page_alloc_cpu_online(unsigned int cpu)
5741 for_each_populated_zone(zone)
5742 zone_pcp_update(zone, 1);
5746 void __init page_alloc_init_cpuhp(void)
5750 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5751 "mm/page_alloc:pcp",
5752 page_alloc_cpu_online,
5753 page_alloc_cpu_dead);
5758 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5759 * or min_free_kbytes changes.
5761 static void calculate_totalreserve_pages(void)
5763 struct pglist_data *pgdat;
5764 unsigned long reserve_pages = 0;
5765 enum zone_type i, j;
5767 for_each_online_pgdat(pgdat) {
5769 pgdat->totalreserve_pages = 0;
5771 for (i = 0; i < MAX_NR_ZONES; i++) {
5772 struct zone *zone = pgdat->node_zones + i;
5774 unsigned long managed_pages = zone_managed_pages(zone);
5776 /* Find valid and maximum lowmem_reserve in the zone */
5777 for (j = i; j < MAX_NR_ZONES; j++) {
5778 if (zone->lowmem_reserve[j] > max)
5779 max = zone->lowmem_reserve[j];
5782 /* we treat the high watermark as reserved pages. */
5783 max += high_wmark_pages(zone);
5785 if (max > managed_pages)
5786 max = managed_pages;
5788 pgdat->totalreserve_pages += max;
5790 reserve_pages += max;
5793 totalreserve_pages = reserve_pages;
5797 * setup_per_zone_lowmem_reserve - called whenever
5798 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5799 * has a correct pages reserved value, so an adequate number of
5800 * pages are left in the zone after a successful __alloc_pages().
5802 static void setup_per_zone_lowmem_reserve(void)
5804 struct pglist_data *pgdat;
5805 enum zone_type i, j;
5807 for_each_online_pgdat(pgdat) {
5808 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5809 struct zone *zone = &pgdat->node_zones[i];
5810 int ratio = sysctl_lowmem_reserve_ratio[i];
5811 bool clear = !ratio || !zone_managed_pages(zone);
5812 unsigned long managed_pages = 0;
5814 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5815 struct zone *upper_zone = &pgdat->node_zones[j];
5817 managed_pages += zone_managed_pages(upper_zone);
5820 zone->lowmem_reserve[j] = 0;
5822 zone->lowmem_reserve[j] = managed_pages / ratio;
5827 /* update totalreserve_pages */
5828 calculate_totalreserve_pages();
5831 static void __setup_per_zone_wmarks(void)
5833 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5834 unsigned long lowmem_pages = 0;
5836 unsigned long flags;
5838 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5839 for_each_zone(zone) {
5840 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5841 lowmem_pages += zone_managed_pages(zone);
5844 for_each_zone(zone) {
5847 spin_lock_irqsave(&zone->lock, flags);
5848 tmp = (u64)pages_min * zone_managed_pages(zone);
5849 do_div(tmp, lowmem_pages);
5850 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5852 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5853 * need highmem and movable zones pages, so cap pages_min
5854 * to a small value here.
5856 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5857 * deltas control async page reclaim, and so should
5858 * not be capped for highmem and movable zones.
5860 unsigned long min_pages;
5862 min_pages = zone_managed_pages(zone) / 1024;
5863 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5864 zone->_watermark[WMARK_MIN] = min_pages;
5867 * If it's a lowmem zone, reserve a number of pages
5868 * proportionate to the zone's size.
5870 zone->_watermark[WMARK_MIN] = tmp;
5874 * Set the kswapd watermarks distance according to the
5875 * scale factor in proportion to available memory, but
5876 * ensure a minimum size on small systems.
5878 tmp = max_t(u64, tmp >> 2,
5879 mult_frac(zone_managed_pages(zone),
5880 watermark_scale_factor, 10000));
5882 zone->watermark_boost = 0;
5883 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
5884 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5885 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5887 spin_unlock_irqrestore(&zone->lock, flags);
5890 /* update totalreserve_pages */
5891 calculate_totalreserve_pages();
5895 * setup_per_zone_wmarks - called when min_free_kbytes changes
5896 * or when memory is hot-{added|removed}
5898 * Ensures that the watermark[min,low,high] values for each zone are set
5899 * correctly with respect to min_free_kbytes.
5901 void setup_per_zone_wmarks(void)
5904 static DEFINE_SPINLOCK(lock);
5907 __setup_per_zone_wmarks();
5911 * The watermark size have changed so update the pcpu batch
5912 * and high limits or the limits may be inappropriate.
5915 zone_pcp_update(zone, 0);
5919 * Initialise min_free_kbytes.
5921 * For small machines we want it small (128k min). For large machines
5922 * we want it large (256MB max). But it is not linear, because network
5923 * bandwidth does not increase linearly with machine size. We use
5925 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5926 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5942 void calculate_min_free_kbytes(void)
5944 unsigned long lowmem_kbytes;
5945 int new_min_free_kbytes;
5947 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5948 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5950 if (new_min_free_kbytes > user_min_free_kbytes)
5951 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
5953 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5954 new_min_free_kbytes, user_min_free_kbytes);
5958 int __meminit init_per_zone_wmark_min(void)
5960 calculate_min_free_kbytes();
5961 setup_per_zone_wmarks();
5962 refresh_zone_stat_thresholds();
5963 setup_per_zone_lowmem_reserve();
5966 setup_min_unmapped_ratio();
5967 setup_min_slab_ratio();
5970 khugepaged_min_free_kbytes_update();
5974 postcore_initcall(init_per_zone_wmark_min)
5977 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5978 * that we can call two helper functions whenever min_free_kbytes
5981 static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5982 void *buffer, size_t *length, loff_t *ppos)
5986 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5991 user_min_free_kbytes = min_free_kbytes;
5992 setup_per_zone_wmarks();
5997 static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
5998 void *buffer, size_t *length, loff_t *ppos)
6002 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6007 setup_per_zone_wmarks();
6013 static void setup_min_unmapped_ratio(void)
6018 for_each_online_pgdat(pgdat)
6019 pgdat->min_unmapped_pages = 0;
6022 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6023 sysctl_min_unmapped_ratio) / 100;
6027 static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6028 void *buffer, size_t *length, loff_t *ppos)
6032 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6036 setup_min_unmapped_ratio();
6041 static void setup_min_slab_ratio(void)
6046 for_each_online_pgdat(pgdat)
6047 pgdat->min_slab_pages = 0;
6050 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6051 sysctl_min_slab_ratio) / 100;
6054 static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6055 void *buffer, size_t *length, loff_t *ppos)
6059 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6063 setup_min_slab_ratio();
6070 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6071 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6072 * whenever sysctl_lowmem_reserve_ratio changes.
6074 * The reserve ratio obviously has absolutely no relation with the
6075 * minimum watermarks. The lowmem reserve ratio can only make sense
6076 * if in function of the boot time zone sizes.
6078 static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table,
6079 int write, void *buffer, size_t *length, loff_t *ppos)
6083 proc_dointvec_minmax(table, write, buffer, length, ppos);
6085 for (i = 0; i < MAX_NR_ZONES; i++) {
6086 if (sysctl_lowmem_reserve_ratio[i] < 1)
6087 sysctl_lowmem_reserve_ratio[i] = 0;
6090 setup_per_zone_lowmem_reserve();
6095 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6096 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6097 * pagelist can have before it gets flushed back to buddy allocator.
6099 static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
6100 int write, void *buffer, size_t *length, loff_t *ppos)
6103 int old_percpu_pagelist_high_fraction;
6106 mutex_lock(&pcp_batch_high_lock);
6107 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6109 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6110 if (!write || ret < 0)
6113 /* Sanity checking to avoid pcp imbalance */
6114 if (percpu_pagelist_high_fraction &&
6115 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6116 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6122 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6125 for_each_populated_zone(zone)
6126 zone_set_pageset_high_and_batch(zone, 0);
6128 mutex_unlock(&pcp_batch_high_lock);
6132 static struct ctl_table page_alloc_sysctl_table[] = {
6134 .procname = "min_free_kbytes",
6135 .data = &min_free_kbytes,
6136 .maxlen = sizeof(min_free_kbytes),
6138 .proc_handler = min_free_kbytes_sysctl_handler,
6139 .extra1 = SYSCTL_ZERO,
6142 .procname = "watermark_boost_factor",
6143 .data = &watermark_boost_factor,
6144 .maxlen = sizeof(watermark_boost_factor),
6146 .proc_handler = proc_dointvec_minmax,
6147 .extra1 = SYSCTL_ZERO,
6150 .procname = "watermark_scale_factor",
6151 .data = &watermark_scale_factor,
6152 .maxlen = sizeof(watermark_scale_factor),
6154 .proc_handler = watermark_scale_factor_sysctl_handler,
6155 .extra1 = SYSCTL_ONE,
6156 .extra2 = SYSCTL_THREE_THOUSAND,
6159 .procname = "percpu_pagelist_high_fraction",
6160 .data = &percpu_pagelist_high_fraction,
6161 .maxlen = sizeof(percpu_pagelist_high_fraction),
6163 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6164 .extra1 = SYSCTL_ZERO,
6167 .procname = "lowmem_reserve_ratio",
6168 .data = &sysctl_lowmem_reserve_ratio,
6169 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6171 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6175 .procname = "numa_zonelist_order",
6176 .data = &numa_zonelist_order,
6177 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6179 .proc_handler = numa_zonelist_order_handler,
6182 .procname = "min_unmapped_ratio",
6183 .data = &sysctl_min_unmapped_ratio,
6184 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6186 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6187 .extra1 = SYSCTL_ZERO,
6188 .extra2 = SYSCTL_ONE_HUNDRED,
6191 .procname = "min_slab_ratio",
6192 .data = &sysctl_min_slab_ratio,
6193 .maxlen = sizeof(sysctl_min_slab_ratio),
6195 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6196 .extra1 = SYSCTL_ZERO,
6197 .extra2 = SYSCTL_ONE_HUNDRED,
6203 void __init page_alloc_sysctl_init(void)
6205 register_sysctl_init("vm", page_alloc_sysctl_table);
6208 #ifdef CONFIG_CONTIG_ALLOC
6209 /* Usage: See admin-guide/dynamic-debug-howto.rst */
6210 static void alloc_contig_dump_pages(struct list_head *page_list)
6212 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6214 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6218 list_for_each_entry(page, page_list, lru)
6219 dump_page(page, "migration failure");
6223 /* [start, end) must belong to a single zone. */
6224 int __alloc_contig_migrate_range(struct compact_control *cc,
6225 unsigned long start, unsigned long end)
6227 /* This function is based on compact_zone() from compaction.c. */
6228 unsigned int nr_reclaimed;
6229 unsigned long pfn = start;
6230 unsigned int tries = 0;
6232 struct migration_target_control mtc = {
6233 .nid = zone_to_nid(cc->zone),
6234 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6237 lru_cache_disable();
6239 while (pfn < end || !list_empty(&cc->migratepages)) {
6240 if (fatal_signal_pending(current)) {
6245 if (list_empty(&cc->migratepages)) {
6246 cc->nr_migratepages = 0;
6247 ret = isolate_migratepages_range(cc, pfn, end);
6248 if (ret && ret != -EAGAIN)
6250 pfn = cc->migrate_pfn;
6252 } else if (++tries == 5) {
6257 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6259 cc->nr_migratepages -= nr_reclaimed;
6261 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6262 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6265 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6266 * to retry again over this error, so do the same here.
6274 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6275 alloc_contig_dump_pages(&cc->migratepages);
6276 putback_movable_pages(&cc->migratepages);
6283 * alloc_contig_range() -- tries to allocate given range of pages
6284 * @start: start PFN to allocate
6285 * @end: one-past-the-last PFN to allocate
6286 * @migratetype: migratetype of the underlying pageblocks (either
6287 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6288 * in range must have the same migratetype and it must
6289 * be either of the two.
6290 * @gfp_mask: GFP mask to use during compaction
6292 * The PFN range does not have to be pageblock aligned. The PFN range must
6293 * belong to a single zone.
6295 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6296 * pageblocks in the range. Once isolated, the pageblocks should not
6297 * be modified by others.
6299 * Return: zero on success or negative error code. On success all
6300 * pages which PFN is in [start, end) are allocated for the caller and
6301 * need to be freed with free_contig_range().
6303 int alloc_contig_range(unsigned long start, unsigned long end,
6304 unsigned migratetype, gfp_t gfp_mask)
6306 unsigned long outer_start, outer_end;
6310 struct compact_control cc = {
6311 .nr_migratepages = 0,
6313 .zone = page_zone(pfn_to_page(start)),
6314 .mode = MIGRATE_SYNC,
6315 .ignore_skip_hint = true,
6316 .no_set_skip_hint = true,
6317 .gfp_mask = current_gfp_context(gfp_mask),
6318 .alloc_contig = true,
6320 INIT_LIST_HEAD(&cc.migratepages);
6323 * What we do here is we mark all pageblocks in range as
6324 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6325 * have different sizes, and due to the way page allocator
6326 * work, start_isolate_page_range() has special handlings for this.
6328 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6329 * migrate the pages from an unaligned range (ie. pages that
6330 * we are interested in). This will put all the pages in
6331 * range back to page allocator as MIGRATE_ISOLATE.
6333 * When this is done, we take the pages in range from page
6334 * allocator removing them from the buddy system. This way
6335 * page allocator will never consider using them.
6337 * This lets us mark the pageblocks back as
6338 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6339 * aligned range but not in the unaligned, original range are
6340 * put back to page allocator so that buddy can use them.
6343 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6347 drain_all_pages(cc.zone);
6350 * In case of -EBUSY, we'd like to know which page causes problem.
6351 * So, just fall through. test_pages_isolated() has a tracepoint
6352 * which will report the busy page.
6354 * It is possible that busy pages could become available before
6355 * the call to test_pages_isolated, and the range will actually be
6356 * allocated. So, if we fall through be sure to clear ret so that
6357 * -EBUSY is not accidentally used or returned to caller.
6359 ret = __alloc_contig_migrate_range(&cc, start, end);
6360 if (ret && ret != -EBUSY)
6365 * Pages from [start, end) are within a pageblock_nr_pages
6366 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6367 * more, all pages in [start, end) are free in page allocator.
6368 * What we are going to do is to allocate all pages from
6369 * [start, end) (that is remove them from page allocator).
6371 * The only problem is that pages at the beginning and at the
6372 * end of interesting range may be not aligned with pages that
6373 * page allocator holds, ie. they can be part of higher order
6374 * pages. Because of this, we reserve the bigger range and
6375 * once this is done free the pages we are not interested in.
6377 * We don't have to hold zone->lock here because the pages are
6378 * isolated thus they won't get removed from buddy.
6382 outer_start = start;
6383 while (!PageBuddy(pfn_to_page(outer_start))) {
6384 if (++order > MAX_ORDER) {
6385 outer_start = start;
6388 outer_start &= ~0UL << order;
6391 if (outer_start != start) {
6392 order = buddy_order(pfn_to_page(outer_start));
6395 * outer_start page could be small order buddy page and
6396 * it doesn't include start page. Adjust outer_start
6397 * in this case to report failed page properly
6398 * on tracepoint in test_pages_isolated()
6400 if (outer_start + (1UL << order) <= start)
6401 outer_start = start;
6404 /* Make sure the range is really isolated. */
6405 if (test_pages_isolated(outer_start, end, 0)) {
6410 /* Grab isolated pages from freelists. */
6411 outer_end = isolate_freepages_range(&cc, outer_start, end);
6417 /* Free head and tail (if any) */
6418 if (start != outer_start)
6419 free_contig_range(outer_start, start - outer_start);
6420 if (end != outer_end)
6421 free_contig_range(end, outer_end - end);
6424 undo_isolate_page_range(start, end, migratetype);
6427 EXPORT_SYMBOL(alloc_contig_range);
6429 static int __alloc_contig_pages(unsigned long start_pfn,
6430 unsigned long nr_pages, gfp_t gfp_mask)
6432 unsigned long end_pfn = start_pfn + nr_pages;
6434 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
6438 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6439 unsigned long nr_pages)
6441 unsigned long i, end_pfn = start_pfn + nr_pages;
6444 for (i = start_pfn; i < end_pfn; i++) {
6445 page = pfn_to_online_page(i);
6449 if (page_zone(page) != z)
6452 if (PageReserved(page))
6461 static bool zone_spans_last_pfn(const struct zone *zone,
6462 unsigned long start_pfn, unsigned long nr_pages)
6464 unsigned long last_pfn = start_pfn + nr_pages - 1;
6466 return zone_spans_pfn(zone, last_pfn);
6470 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6471 * @nr_pages: Number of contiguous pages to allocate
6472 * @gfp_mask: GFP mask to limit search and used during compaction
6474 * @nodemask: Mask for other possible nodes
6476 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6477 * on an applicable zonelist to find a contiguous pfn range which can then be
6478 * tried for allocation with alloc_contig_range(). This routine is intended
6479 * for allocation requests which can not be fulfilled with the buddy allocator.
6481 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6482 * power of two, then allocated range is also guaranteed to be aligned to same
6483 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6485 * Allocated pages can be freed with free_contig_range() or by manually calling
6486 * __free_page() on each allocated page.
6488 * Return: pointer to contiguous pages on success, or NULL if not successful.
6490 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
6491 int nid, nodemask_t *nodemask)
6493 unsigned long ret, pfn, flags;
6494 struct zonelist *zonelist;
6498 zonelist = node_zonelist(nid, gfp_mask);
6499 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6500 gfp_zone(gfp_mask), nodemask) {
6501 spin_lock_irqsave(&zone->lock, flags);
6503 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6504 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6505 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6507 * We release the zone lock here because
6508 * alloc_contig_range() will also lock the zone
6509 * at some point. If there's an allocation
6510 * spinning on this lock, it may win the race
6511 * and cause alloc_contig_range() to fail...
6513 spin_unlock_irqrestore(&zone->lock, flags);
6514 ret = __alloc_contig_pages(pfn, nr_pages,
6517 return pfn_to_page(pfn);
6518 spin_lock_irqsave(&zone->lock, flags);
6522 spin_unlock_irqrestore(&zone->lock, flags);
6526 #endif /* CONFIG_CONTIG_ALLOC */
6528 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6530 unsigned long count = 0;
6532 for (; nr_pages--; pfn++) {
6533 struct page *page = pfn_to_page(pfn);
6535 count += page_count(page) != 1;
6538 WARN(count != 0, "%lu pages are still in use!\n", count);
6540 EXPORT_SYMBOL(free_contig_range);
6543 * Effectively disable pcplists for the zone by setting the high limit to 0
6544 * and draining all cpus. A concurrent page freeing on another CPU that's about
6545 * to put the page on pcplist will either finish before the drain and the page
6546 * will be drained, or observe the new high limit and skip the pcplist.
6548 * Must be paired with a call to zone_pcp_enable().
6550 void zone_pcp_disable(struct zone *zone)
6552 mutex_lock(&pcp_batch_high_lock);
6553 __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
6554 __drain_all_pages(zone, true);
6557 void zone_pcp_enable(struct zone *zone)
6559 __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
6560 zone->pageset_high_max, zone->pageset_batch);
6561 mutex_unlock(&pcp_batch_high_lock);
6564 void zone_pcp_reset(struct zone *zone)
6567 struct per_cpu_zonestat *pzstats;
6569 if (zone->per_cpu_pageset != &boot_pageset) {
6570 for_each_online_cpu(cpu) {
6571 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6572 drain_zonestat(zone, pzstats);
6574 free_percpu(zone->per_cpu_pageset);
6575 zone->per_cpu_pageset = &boot_pageset;
6576 if (zone->per_cpu_zonestats != &boot_zonestats) {
6577 free_percpu(zone->per_cpu_zonestats);
6578 zone->per_cpu_zonestats = &boot_zonestats;
6583 #ifdef CONFIG_MEMORY_HOTREMOVE
6585 * All pages in the range must be in a single zone, must not contain holes,
6586 * must span full sections, and must be isolated before calling this function.
6588 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6590 unsigned long pfn = start_pfn;
6594 unsigned long flags;
6596 offline_mem_sections(pfn, end_pfn);
6597 zone = page_zone(pfn_to_page(pfn));
6598 spin_lock_irqsave(&zone->lock, flags);
6599 while (pfn < end_pfn) {
6600 page = pfn_to_page(pfn);
6602 * The HWPoisoned page may be not in buddy system, and
6603 * page_count() is not 0.
6605 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6610 * At this point all remaining PageOffline() pages have a
6611 * reference count of 0 and can simply be skipped.
6613 if (PageOffline(page)) {
6614 BUG_ON(page_count(page));
6615 BUG_ON(PageBuddy(page));
6620 BUG_ON(page_count(page));
6621 BUG_ON(!PageBuddy(page));
6622 order = buddy_order(page);
6623 del_page_from_free_list(page, zone, order);
6624 pfn += (1 << order);
6626 spin_unlock_irqrestore(&zone->lock, flags);
6631 * This function returns a stable result only if called under zone lock.
6633 bool is_free_buddy_page(struct page *page)
6635 unsigned long pfn = page_to_pfn(page);
6638 for (order = 0; order <= MAX_ORDER; order++) {
6639 struct page *page_head = page - (pfn & ((1 << order) - 1));
6641 if (PageBuddy(page_head) &&
6642 buddy_order_unsafe(page_head) >= order)
6646 return order <= MAX_ORDER;
6648 EXPORT_SYMBOL(is_free_buddy_page);
6650 #ifdef CONFIG_MEMORY_FAILURE
6652 * Break down a higher-order page in sub-pages, and keep our target out of
6655 static void break_down_buddy_pages(struct zone *zone, struct page *page,
6656 struct page *target, int low, int high,
6659 unsigned long size = 1 << high;
6660 struct page *current_buddy;
6662 while (high > low) {
6666 if (target >= &page[size]) {
6667 current_buddy = page;
6670 current_buddy = page + size;
6673 if (set_page_guard(zone, current_buddy, high, migratetype))
6676 add_to_free_list(current_buddy, zone, high, migratetype);
6677 set_buddy_order(current_buddy, high);
6682 * Take a page that will be marked as poisoned off the buddy allocator.
6684 bool take_page_off_buddy(struct page *page)
6686 struct zone *zone = page_zone(page);
6687 unsigned long pfn = page_to_pfn(page);
6688 unsigned long flags;
6692 spin_lock_irqsave(&zone->lock, flags);
6693 for (order = 0; order <= MAX_ORDER; order++) {
6694 struct page *page_head = page - (pfn & ((1 << order) - 1));
6695 int page_order = buddy_order(page_head);
6697 if (PageBuddy(page_head) && page_order >= order) {
6698 unsigned long pfn_head = page_to_pfn(page_head);
6699 int migratetype = get_pfnblock_migratetype(page_head,
6702 del_page_from_free_list(page_head, zone, page_order);
6703 break_down_buddy_pages(zone, page_head, page, 0,
6704 page_order, migratetype);
6705 SetPageHWPoisonTakenOff(page);
6706 if (!is_migrate_isolate(migratetype))
6707 __mod_zone_freepage_state(zone, -1, migratetype);
6711 if (page_count(page_head) > 0)
6714 spin_unlock_irqrestore(&zone->lock, flags);
6719 * Cancel takeoff done by take_page_off_buddy().
6721 bool put_page_back_buddy(struct page *page)
6723 struct zone *zone = page_zone(page);
6724 unsigned long pfn = page_to_pfn(page);
6725 unsigned long flags;
6726 int migratetype = get_pfnblock_migratetype(page, pfn);
6729 spin_lock_irqsave(&zone->lock, flags);
6730 if (put_page_testzero(page)) {
6731 ClearPageHWPoisonTakenOff(page);
6732 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6733 if (TestClearPageHWPoison(page)) {
6737 spin_unlock_irqrestore(&zone->lock, flags);
6743 #ifdef CONFIG_ZONE_DMA
6744 bool has_managed_dma(void)
6746 struct pglist_data *pgdat;
6748 for_each_online_pgdat(pgdat) {
6749 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6751 if (managed_zone(zone))
6756 #endif /* CONFIG_ZONE_DMA */
6758 #ifdef CONFIG_UNACCEPTED_MEMORY
6760 /* Counts number of zones with unaccepted pages. */
6761 static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6763 static bool lazy_accept = true;
6765 static int __init accept_memory_parse(char *p)
6767 if (!strcmp(p, "lazy")) {
6770 } else if (!strcmp(p, "eager")) {
6771 lazy_accept = false;
6777 early_param("accept_memory", accept_memory_parse);
6779 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6781 phys_addr_t start = page_to_phys(page);
6782 phys_addr_t end = start + (PAGE_SIZE << order);
6784 return range_contains_unaccepted_memory(start, end);
6787 static void accept_page(struct page *page, unsigned int order)
6789 phys_addr_t start = page_to_phys(page);
6791 accept_memory(start, start + (PAGE_SIZE << order));
6794 static bool try_to_accept_memory_one(struct zone *zone)
6796 unsigned long flags;
6800 if (list_empty(&zone->unaccepted_pages))
6803 spin_lock_irqsave(&zone->lock, flags);
6804 page = list_first_entry_or_null(&zone->unaccepted_pages,
6807 spin_unlock_irqrestore(&zone->lock, flags);
6811 list_del(&page->lru);
6812 last = list_empty(&zone->unaccepted_pages);
6814 __mod_zone_freepage_state(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6815 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6816 spin_unlock_irqrestore(&zone->lock, flags);
6818 accept_page(page, MAX_ORDER);
6820 __free_pages_ok(page, MAX_ORDER, FPI_TO_TAIL);
6823 static_branch_dec(&zones_with_unaccepted_pages);
6828 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6833 /* How much to accept to get to high watermark? */
6834 to_accept = high_wmark_pages(zone) -
6835 (zone_page_state(zone, NR_FREE_PAGES) -
6836 __zone_watermark_unusable_free(zone, order, 0));
6838 /* Accept at least one page */
6840 if (!try_to_accept_memory_one(zone))
6843 to_accept -= MAX_ORDER_NR_PAGES;
6844 } while (to_accept > 0);
6849 static inline bool has_unaccepted_memory(void)
6851 return static_branch_unlikely(&zones_with_unaccepted_pages);
6854 static bool __free_unaccepted(struct page *page)
6856 struct zone *zone = page_zone(page);
6857 unsigned long flags;
6863 spin_lock_irqsave(&zone->lock, flags);
6864 first = list_empty(&zone->unaccepted_pages);
6865 list_add_tail(&page->lru, &zone->unaccepted_pages);
6866 __mod_zone_freepage_state(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6867 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
6868 spin_unlock_irqrestore(&zone->lock, flags);
6871 static_branch_inc(&zones_with_unaccepted_pages);
6878 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6883 static void accept_page(struct page *page, unsigned int order)
6887 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6892 static inline bool has_unaccepted_memory(void)
6897 static bool __free_unaccepted(struct page *page)
6903 #endif /* CONFIG_UNACCEPTED_MEMORY */