1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/interrupt.h>
22 #include <linux/jiffies.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.h>
25 #include <linux/kasan.h>
26 #include <linux/kmsan.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/ratelimit.h>
30 #include <linux/oom.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/pagevec.h>
36 #include <linux/memory_hotplug.h>
37 #include <linux/nodemask.h>
38 #include <linux/vmstat.h>
39 #include <linux/fault-inject.h>
40 #include <linux/compaction.h>
41 #include <trace/events/kmem.h>
42 #include <trace/events/oom.h>
43 #include <linux/prefetch.h>
44 #include <linux/mm_inline.h>
45 #include <linux/mmu_notifier.h>
46 #include <linux/migrate.h>
47 #include <linux/sched/mm.h>
48 #include <linux/page_owner.h>
49 #include <linux/page_table_check.h>
50 #include <linux/memcontrol.h>
51 #include <linux/ftrace.h>
52 #include <linux/lockdep.h>
53 #include <linux/psi.h>
54 #include <linux/khugepaged.h>
55 #include <linux/delayacct.h>
56 #include <linux/cacheinfo.h>
57 #include <asm/div64.h>
60 #include "page_reporting.h"
62 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
63 typedef int __bitwise fpi_t;
65 /* No special request */
66 #define FPI_NONE ((__force fpi_t)0)
69 * Skip free page reporting notification for the (possibly merged) page.
70 * This does not hinder free page reporting from grabbing the page,
71 * reporting it and marking it "reported" - it only skips notifying
72 * the free page reporting infrastructure about a newly freed page. For
73 * example, used when temporarily pulling a page from a freelist and
74 * putting it back unmodified.
76 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
79 * Place the (possibly merged) page to the tail of the freelist. Will ignore
80 * page shuffling (relevant code - e.g., memory onlining - is expected to
81 * shuffle the whole zone).
83 * Note: No code should rely on this flag for correctness - it's purely
84 * to allow for optimizations when handing back either fresh pages
85 * (memory onlining) or untouched pages (page isolation, free page
88 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
90 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
91 static DEFINE_MUTEX(pcp_batch_high_lock);
92 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
94 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
96 * On SMP, spin_trylock is sufficient protection.
97 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
99 #define pcp_trylock_prepare(flags) do { } while (0)
100 #define pcp_trylock_finish(flag) do { } while (0)
103 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
104 #define pcp_trylock_prepare(flags) local_irq_save(flags)
105 #define pcp_trylock_finish(flags) local_irq_restore(flags)
109 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
110 * a migration causing the wrong PCP to be locked and remote memory being
111 * potentially allocated, pin the task to the CPU for the lookup+lock.
112 * preempt_disable is used on !RT because it is faster than migrate_disable.
113 * migrate_disable is used on RT because otherwise RT spinlock usage is
114 * interfered with and a high priority task cannot preempt the allocator.
116 #ifndef CONFIG_PREEMPT_RT
117 #define pcpu_task_pin() preempt_disable()
118 #define pcpu_task_unpin() preempt_enable()
120 #define pcpu_task_pin() migrate_disable()
121 #define pcpu_task_unpin() migrate_enable()
125 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
126 * Return value should be used with equivalent unlock helper.
128 #define pcpu_spin_lock(type, member, ptr) \
132 _ret = this_cpu_ptr(ptr); \
133 spin_lock(&_ret->member); \
137 #define pcpu_spin_trylock(type, member, ptr) \
141 _ret = this_cpu_ptr(ptr); \
142 if (!spin_trylock(&_ret->member)) { \
149 #define pcpu_spin_unlock(member, ptr) \
151 spin_unlock(&ptr->member); \
155 /* struct per_cpu_pages specific helpers. */
156 #define pcp_spin_lock(ptr) \
157 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
159 #define pcp_spin_trylock(ptr) \
160 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
162 #define pcp_spin_unlock(ptr) \
163 pcpu_spin_unlock(lock, ptr)
165 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
166 DEFINE_PER_CPU(int, numa_node);
167 EXPORT_PER_CPU_SYMBOL(numa_node);
170 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
172 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
174 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
175 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
176 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
177 * defined in <linux/topology.h>.
179 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
180 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
183 static DEFINE_MUTEX(pcpu_drain_mutex);
185 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
186 volatile unsigned long latent_entropy __latent_entropy;
187 EXPORT_SYMBOL(latent_entropy);
191 * Array of node states.
193 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
194 [N_POSSIBLE] = NODE_MASK_ALL,
195 [N_ONLINE] = { { [0] = 1UL } },
197 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
198 #ifdef CONFIG_HIGHMEM
199 [N_HIGH_MEMORY] = { { [0] = 1UL } },
201 [N_MEMORY] = { { [0] = 1UL } },
202 [N_CPU] = { { [0] = 1UL } },
205 EXPORT_SYMBOL(node_states);
207 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
210 * A cached value of the page's pageblock's migratetype, used when the page is
211 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
212 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
213 * Also the migratetype set in the page does not necessarily match the pcplist
214 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
215 * other index - this ensures that it will be put on the correct CMA freelist.
217 static inline int get_pcppage_migratetype(struct page *page)
222 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
224 page->index = migratetype;
227 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
228 unsigned int pageblock_order __read_mostly;
231 static void __free_pages_ok(struct page *page, unsigned int order,
235 * results with 256, 32 in the lowmem_reserve sysctl:
236 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
237 * 1G machine -> (16M dma, 784M normal, 224M high)
238 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
239 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
240 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
242 * TBD: should special case ZONE_DMA32 machines here - in those we normally
243 * don't need any ZONE_NORMAL reservation
245 static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
246 #ifdef CONFIG_ZONE_DMA
249 #ifdef CONFIG_ZONE_DMA32
253 #ifdef CONFIG_HIGHMEM
259 char * const zone_names[MAX_NR_ZONES] = {
260 #ifdef CONFIG_ZONE_DMA
263 #ifdef CONFIG_ZONE_DMA32
267 #ifdef CONFIG_HIGHMEM
271 #ifdef CONFIG_ZONE_DEVICE
276 const char * const migratetype_names[MIGRATE_TYPES] = {
284 #ifdef CONFIG_MEMORY_ISOLATION
289 int min_free_kbytes = 1024;
290 int user_min_free_kbytes = -1;
291 static int watermark_boost_factor __read_mostly = 15000;
292 static int watermark_scale_factor = 10;
294 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
296 EXPORT_SYMBOL(movable_zone);
299 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
300 unsigned int nr_online_nodes __read_mostly = 1;
301 EXPORT_SYMBOL(nr_node_ids);
302 EXPORT_SYMBOL(nr_online_nodes);
305 static bool page_contains_unaccepted(struct page *page, unsigned int order);
306 static void accept_page(struct page *page, unsigned int order);
307 static bool try_to_accept_memory(struct zone *zone, unsigned int order);
308 static inline bool has_unaccepted_memory(void);
309 static bool __free_unaccepted(struct page *page);
311 int page_group_by_mobility_disabled __read_mostly;
313 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
315 * During boot we initialize deferred pages on-demand, as needed, but once
316 * page_alloc_init_late() has finished, the deferred pages are all initialized,
317 * and we can permanently disable that path.
319 DEFINE_STATIC_KEY_TRUE(deferred_pages);
321 static inline bool deferred_pages_enabled(void)
323 return static_branch_unlikely(&deferred_pages);
327 * deferred_grow_zone() is __init, but it is called from
328 * get_page_from_freelist() during early boot until deferred_pages permanently
329 * disables this call. This is why we have refdata wrapper to avoid warning,
330 * and to ensure that the function body gets unloaded.
333 _deferred_grow_zone(struct zone *zone, unsigned int order)
335 return deferred_grow_zone(zone, order);
338 static inline bool deferred_pages_enabled(void)
342 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
344 /* Return a pointer to the bitmap storing bits affecting a block of pages */
345 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
348 #ifdef CONFIG_SPARSEMEM
349 return section_to_usemap(__pfn_to_section(pfn));
351 return page_zone(page)->pageblock_flags;
352 #endif /* CONFIG_SPARSEMEM */
355 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
357 #ifdef CONFIG_SPARSEMEM
358 pfn &= (PAGES_PER_SECTION-1);
360 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
361 #endif /* CONFIG_SPARSEMEM */
362 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
366 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
367 * @page: The page within the block of interest
368 * @pfn: The target page frame number
369 * @mask: mask of bits that the caller is interested in
371 * Return: pageblock_bits flags
373 unsigned long get_pfnblock_flags_mask(const struct page *page,
374 unsigned long pfn, unsigned long mask)
376 unsigned long *bitmap;
377 unsigned long bitidx, word_bitidx;
380 bitmap = get_pageblock_bitmap(page, pfn);
381 bitidx = pfn_to_bitidx(page, pfn);
382 word_bitidx = bitidx / BITS_PER_LONG;
383 bitidx &= (BITS_PER_LONG-1);
385 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
386 * a consistent read of the memory array, so that results, even though
387 * racy, are not corrupted.
389 word = READ_ONCE(bitmap[word_bitidx]);
390 return (word >> bitidx) & mask;
393 static __always_inline int get_pfnblock_migratetype(const struct page *page,
396 return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
400 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
401 * @page: The page within the block of interest
402 * @flags: The flags to set
403 * @pfn: The target page frame number
404 * @mask: mask of bits that the caller is interested in
406 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
410 unsigned long *bitmap;
411 unsigned long bitidx, word_bitidx;
414 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
415 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
417 bitmap = get_pageblock_bitmap(page, pfn);
418 bitidx = pfn_to_bitidx(page, pfn);
419 word_bitidx = bitidx / BITS_PER_LONG;
420 bitidx &= (BITS_PER_LONG-1);
422 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
427 word = READ_ONCE(bitmap[word_bitidx]);
429 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
432 void set_pageblock_migratetype(struct page *page, int migratetype)
434 if (unlikely(page_group_by_mobility_disabled &&
435 migratetype < MIGRATE_PCPTYPES))
436 migratetype = MIGRATE_UNMOVABLE;
438 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
439 page_to_pfn(page), MIGRATETYPE_MASK);
442 #ifdef CONFIG_DEBUG_VM
443 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
447 unsigned long pfn = page_to_pfn(page);
448 unsigned long sp, start_pfn;
451 seq = zone_span_seqbegin(zone);
452 start_pfn = zone->zone_start_pfn;
453 sp = zone->spanned_pages;
454 ret = !zone_spans_pfn(zone, pfn);
455 } while (zone_span_seqretry(zone, seq));
458 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
459 pfn, zone_to_nid(zone), zone->name,
460 start_pfn, start_pfn + sp);
466 * Temporary debugging check for pages not lying within a given zone.
468 static bool __maybe_unused bad_range(struct zone *zone, struct page *page)
470 if (page_outside_zone_boundaries(zone, page))
472 if (zone != page_zone(page))
478 static inline bool __maybe_unused bad_range(struct zone *zone, struct page *page)
484 static void bad_page(struct page *page, const char *reason)
486 static unsigned long resume;
487 static unsigned long nr_shown;
488 static unsigned long nr_unshown;
491 * Allow a burst of 60 reports, then keep quiet for that minute;
492 * or allow a steady drip of one report per second.
494 if (nr_shown == 60) {
495 if (time_before(jiffies, resume)) {
501 "BUG: Bad page state: %lu messages suppressed\n",
508 resume = jiffies + 60 * HZ;
510 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
511 current->comm, page_to_pfn(page));
512 dump_page(page, reason);
517 /* Leave bad fields for debug, except PageBuddy could make trouble */
518 page_mapcount_reset(page); /* remove PageBuddy */
519 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
522 static inline unsigned int order_to_pindex(int migratetype, int order)
524 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
525 if (order > PAGE_ALLOC_COSTLY_ORDER) {
526 VM_BUG_ON(order != pageblock_order);
527 return NR_LOWORDER_PCP_LISTS;
530 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
533 return (MIGRATE_PCPTYPES * order) + migratetype;
536 static inline int pindex_to_order(unsigned int pindex)
538 int order = pindex / MIGRATE_PCPTYPES;
540 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
541 if (pindex == NR_LOWORDER_PCP_LISTS)
542 order = pageblock_order;
544 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
550 static inline bool pcp_allowed_order(unsigned int order)
552 if (order <= PAGE_ALLOC_COSTLY_ORDER)
554 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
555 if (order == pageblock_order)
561 static inline void free_the_page(struct page *page, unsigned int order)
563 if (pcp_allowed_order(order)) /* Via pcp? */
564 free_unref_page(page, order);
566 __free_pages_ok(page, order, FPI_NONE);
570 * Higher-order pages are called "compound pages". They are structured thusly:
572 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
574 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
575 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
577 * The first tail page's ->compound_order holds the order of allocation.
578 * This usage means that zero-order pages may not be compound.
581 void prep_compound_page(struct page *page, unsigned int order)
584 int nr_pages = 1 << order;
587 for (i = 1; i < nr_pages; i++)
588 prep_compound_tail(page, i);
590 prep_compound_head(page, order);
593 void destroy_large_folio(struct folio *folio)
595 if (folio_test_hugetlb(folio)) {
596 free_huge_folio(folio);
600 if (folio_test_large_rmappable(folio))
601 folio_undo_large_rmappable(folio);
603 mem_cgroup_uncharge(folio);
604 free_the_page(&folio->page, folio_order(folio));
607 static inline void set_buddy_order(struct page *page, unsigned int order)
609 set_page_private(page, order);
610 __SetPageBuddy(page);
613 #ifdef CONFIG_COMPACTION
614 static inline struct capture_control *task_capc(struct zone *zone)
616 struct capture_control *capc = current->capture_control;
618 return unlikely(capc) &&
619 !(current->flags & PF_KTHREAD) &&
621 capc->cc->zone == zone ? capc : NULL;
625 compaction_capture(struct capture_control *capc, struct page *page,
626 int order, int migratetype)
628 if (!capc || order != capc->cc->order)
631 /* Do not accidentally pollute CMA or isolated regions*/
632 if (is_migrate_cma(migratetype) ||
633 is_migrate_isolate(migratetype))
637 * Do not let lower order allocations pollute a movable pageblock.
638 * This might let an unmovable request use a reclaimable pageblock
639 * and vice-versa but no more than normal fallback logic which can
640 * have trouble finding a high-order free page.
642 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
650 static inline struct capture_control *task_capc(struct zone *zone)
656 compaction_capture(struct capture_control *capc, struct page *page,
657 int order, int migratetype)
661 #endif /* CONFIG_COMPACTION */
663 /* Used for pages not on another list */
664 static inline void add_to_free_list(struct page *page, struct zone *zone,
665 unsigned int order, int migratetype)
667 struct free_area *area = &zone->free_area[order];
669 list_add(&page->buddy_list, &area->free_list[migratetype]);
673 /* Used for pages not on another list */
674 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
675 unsigned int order, int migratetype)
677 struct free_area *area = &zone->free_area[order];
679 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
684 * Used for pages which are on another list. Move the pages to the tail
685 * of the list - so the moved pages won't immediately be considered for
686 * allocation again (e.g., optimization for memory onlining).
688 static inline void move_to_free_list(struct page *page, struct zone *zone,
689 unsigned int order, int migratetype)
691 struct free_area *area = &zone->free_area[order];
693 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
696 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
699 /* clear reported state and update reported page count */
700 if (page_reported(page))
701 __ClearPageReported(page);
703 list_del(&page->buddy_list);
704 __ClearPageBuddy(page);
705 set_page_private(page, 0);
706 zone->free_area[order].nr_free--;
709 static inline struct page *get_page_from_free_area(struct free_area *area,
712 return list_first_entry_or_null(&area->free_list[migratetype],
713 struct page, buddy_list);
717 * If this is not the largest possible page, check if the buddy
718 * of the next-highest order is free. If it is, it's possible
719 * that pages are being freed that will coalesce soon. In case,
720 * that is happening, add the free page to the tail of the list
721 * so it's less likely to be used soon and more likely to be merged
722 * as a higher order page
725 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
726 struct page *page, unsigned int order)
728 unsigned long higher_page_pfn;
729 struct page *higher_page;
731 if (order >= MAX_PAGE_ORDER - 1)
734 higher_page_pfn = buddy_pfn & pfn;
735 higher_page = page + (higher_page_pfn - pfn);
737 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
742 * Freeing function for a buddy system allocator.
744 * The concept of a buddy system is to maintain direct-mapped table
745 * (containing bit values) for memory blocks of various "orders".
746 * The bottom level table contains the map for the smallest allocatable
747 * units of memory (here, pages), and each level above it describes
748 * pairs of units from the levels below, hence, "buddies".
749 * At a high level, all that happens here is marking the table entry
750 * at the bottom level available, and propagating the changes upward
751 * as necessary, plus some accounting needed to play nicely with other
752 * parts of the VM system.
753 * At each level, we keep a list of pages, which are heads of continuous
754 * free pages of length of (1 << order) and marked with PageBuddy.
755 * Page's order is recorded in page_private(page) field.
756 * So when we are allocating or freeing one, we can derive the state of the
757 * other. That is, if we allocate a small block, and both were
758 * free, the remainder of the region must be split into blocks.
759 * If a block is freed, and its buddy is also free, then this
760 * triggers coalescing into a block of larger size.
765 static inline void __free_one_page(struct page *page,
767 struct zone *zone, unsigned int order,
768 int migratetype, fpi_t fpi_flags)
770 struct capture_control *capc = task_capc(zone);
771 unsigned long buddy_pfn = 0;
772 unsigned long combined_pfn;
776 VM_BUG_ON(!zone_is_initialized(zone));
777 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
779 VM_BUG_ON(migratetype == -1);
780 if (likely(!is_migrate_isolate(migratetype)))
781 __mod_zone_freepage_state(zone, 1 << order, migratetype);
783 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
784 VM_BUG_ON_PAGE(bad_range(zone, page), page);
786 while (order < MAX_PAGE_ORDER) {
787 if (compaction_capture(capc, page, order, migratetype)) {
788 __mod_zone_freepage_state(zone, -(1 << order),
793 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
797 if (unlikely(order >= pageblock_order)) {
799 * We want to prevent merge between freepages on pageblock
800 * without fallbacks and normal pageblock. Without this,
801 * pageblock isolation could cause incorrect freepage or CMA
802 * accounting or HIGHATOMIC accounting.
804 int buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
806 if (migratetype != buddy_mt
807 && (!migratetype_is_mergeable(migratetype) ||
808 !migratetype_is_mergeable(buddy_mt)))
813 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
814 * merge with it and move up one order.
816 if (page_is_guard(buddy))
817 clear_page_guard(zone, buddy, order, migratetype);
819 del_page_from_free_list(buddy, zone, order);
820 combined_pfn = buddy_pfn & pfn;
821 page = page + (combined_pfn - pfn);
827 set_buddy_order(page, order);
829 if (fpi_flags & FPI_TO_TAIL)
831 else if (is_shuffle_order(order))
832 to_tail = shuffle_pick_tail();
834 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
837 add_to_free_list_tail(page, zone, order, migratetype);
839 add_to_free_list(page, zone, order, migratetype);
841 /* Notify page reporting subsystem of freed page */
842 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
843 page_reporting_notify_free(order);
847 * split_free_page() -- split a free page at split_pfn_offset
848 * @free_page: the original free page
849 * @order: the order of the page
850 * @split_pfn_offset: split offset within the page
852 * Return -ENOENT if the free page is changed, otherwise 0
854 * It is used when the free page crosses two pageblocks with different migratetypes
855 * at split_pfn_offset within the page. The split free page will be put into
856 * separate migratetype lists afterwards. Otherwise, the function achieves
859 int split_free_page(struct page *free_page,
860 unsigned int order, unsigned long split_pfn_offset)
862 struct zone *zone = page_zone(free_page);
863 unsigned long free_page_pfn = page_to_pfn(free_page);
870 if (split_pfn_offset == 0)
873 spin_lock_irqsave(&zone->lock, flags);
875 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
880 mt = get_pfnblock_migratetype(free_page, free_page_pfn);
881 if (likely(!is_migrate_isolate(mt)))
882 __mod_zone_freepage_state(zone, -(1UL << order), mt);
884 del_page_from_free_list(free_page, zone, order);
885 for (pfn = free_page_pfn;
886 pfn < free_page_pfn + (1UL << order);) {
887 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
889 free_page_order = min_t(unsigned int,
890 pfn ? __ffs(pfn) : order,
891 __fls(split_pfn_offset));
892 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
894 pfn += 1UL << free_page_order;
895 split_pfn_offset -= (1UL << free_page_order);
896 /* we have done the first part, now switch to second part */
897 if (split_pfn_offset == 0)
898 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
901 spin_unlock_irqrestore(&zone->lock, flags);
905 * A bad page could be due to a number of fields. Instead of multiple branches,
906 * try and check multiple fields with one check. The caller must do a detailed
907 * check if necessary.
909 static inline bool page_expected_state(struct page *page,
910 unsigned long check_flags)
912 if (unlikely(atomic_read(&page->_mapcount) != -1))
915 if (unlikely((unsigned long)page->mapping |
916 page_ref_count(page) |
920 #ifdef CONFIG_PAGE_POOL
921 ((page->pp_magic & ~0x3UL) == PP_SIGNATURE) |
923 (page->flags & check_flags)))
929 static const char *page_bad_reason(struct page *page, unsigned long flags)
931 const char *bad_reason = NULL;
933 if (unlikely(atomic_read(&page->_mapcount) != -1))
934 bad_reason = "nonzero mapcount";
935 if (unlikely(page->mapping != NULL))
936 bad_reason = "non-NULL mapping";
937 if (unlikely(page_ref_count(page) != 0))
938 bad_reason = "nonzero _refcount";
939 if (unlikely(page->flags & flags)) {
940 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
941 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
943 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
946 if (unlikely(page->memcg_data))
947 bad_reason = "page still charged to cgroup";
949 #ifdef CONFIG_PAGE_POOL
950 if (unlikely((page->pp_magic & ~0x3UL) == PP_SIGNATURE))
951 bad_reason = "page_pool leak";
956 static void free_page_is_bad_report(struct page *page)
959 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
962 static inline bool free_page_is_bad(struct page *page)
964 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
967 /* Something has gone sideways, find it */
968 free_page_is_bad_report(page);
972 static inline bool is_check_pages_enabled(void)
974 return static_branch_unlikely(&check_pages_enabled);
977 static int free_tail_page_prepare(struct page *head_page, struct page *page)
979 struct folio *folio = (struct folio *)head_page;
983 * We rely page->lru.next never has bit 0 set, unless the page
984 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
986 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
988 if (!is_check_pages_enabled()) {
992 switch (page - head_page) {
994 /* the first tail page: these may be in place of ->mapping */
995 if (unlikely(folio_entire_mapcount(folio))) {
996 bad_page(page, "nonzero entire_mapcount");
999 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
1000 bad_page(page, "nonzero nr_pages_mapped");
1003 if (unlikely(atomic_read(&folio->_pincount))) {
1004 bad_page(page, "nonzero pincount");
1010 * the second tail page: ->mapping is
1011 * deferred_list.next -- ignore value.
1015 if (page->mapping != TAIL_MAPPING) {
1016 bad_page(page, "corrupted mapping in tail page");
1021 if (unlikely(!PageTail(page))) {
1022 bad_page(page, "PageTail not set");
1025 if (unlikely(compound_head(page) != head_page)) {
1026 bad_page(page, "compound_head not consistent");
1031 page->mapping = NULL;
1032 clear_compound_head(page);
1037 * Skip KASAN memory poisoning when either:
1039 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1040 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1041 * using page tags instead (see below).
1042 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1043 * that error detection is disabled for accesses via the page address.
1045 * Pages will have match-all tags in the following circumstances:
1047 * 1. Pages are being initialized for the first time, including during deferred
1048 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1049 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1050 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1051 * 3. The allocation was excluded from being checked due to sampling,
1052 * see the call to kasan_unpoison_pages.
1054 * Poisoning pages during deferred memory init will greatly lengthen the
1055 * process and cause problem in large memory systems as the deferred pages
1056 * initialization is done with interrupt disabled.
1058 * Assuming that there will be no reference to those newly initialized
1059 * pages before they are ever allocated, this should have no effect on
1060 * KASAN memory tracking as the poison will be properly inserted at page
1061 * allocation time. The only corner case is when pages are allocated by
1062 * on-demand allocation and then freed again before the deferred pages
1063 * initialization is done, but this is not likely to happen.
1065 static inline bool should_skip_kasan_poison(struct page *page)
1067 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1068 return deferred_pages_enabled();
1070 return page_kasan_tag(page) == KASAN_TAG_KERNEL;
1073 static void kernel_init_pages(struct page *page, int numpages)
1077 /* s390's use of memset() could override KASAN redzones. */
1078 kasan_disable_current();
1079 for (i = 0; i < numpages; i++)
1080 clear_highpage_kasan_tagged(page + i);
1081 kasan_enable_current();
1084 __always_inline bool free_pages_prepare(struct page *page,
1088 bool skip_kasan_poison = should_skip_kasan_poison(page);
1089 bool init = want_init_on_free();
1090 bool compound = PageCompound(page);
1092 VM_BUG_ON_PAGE(PageTail(page), page);
1094 trace_mm_page_free(page, order);
1095 kmsan_free_page(page, order);
1097 if (memcg_kmem_online() && PageMemcgKmem(page))
1098 __memcg_kmem_uncharge_page(page, order);
1100 if (unlikely(PageHWPoison(page)) && !order) {
1101 /* Do not let hwpoison pages hit pcplists/buddy */
1102 reset_page_owner(page, order);
1103 page_table_check_free(page, order);
1107 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1110 * Check tail pages before head page information is cleared to
1111 * avoid checking PageCompound for order-0 pages.
1113 if (unlikely(order)) {
1117 page[1].flags &= ~PAGE_FLAGS_SECOND;
1118 for (i = 1; i < (1 << order); i++) {
1120 bad += free_tail_page_prepare(page, page + i);
1121 if (is_check_pages_enabled()) {
1122 if (free_page_is_bad(page + i)) {
1127 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1130 if (PageMappingFlags(page))
1131 page->mapping = NULL;
1132 if (is_check_pages_enabled()) {
1133 if (free_page_is_bad(page))
1139 page_cpupid_reset_last(page);
1140 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1141 reset_page_owner(page, order);
1142 page_table_check_free(page, order);
1144 if (!PageHighMem(page)) {
1145 debug_check_no_locks_freed(page_address(page),
1146 PAGE_SIZE << order);
1147 debug_check_no_obj_freed(page_address(page),
1148 PAGE_SIZE << order);
1151 kernel_poison_pages(page, 1 << order);
1154 * As memory initialization might be integrated into KASAN,
1155 * KASAN poisoning and memory initialization code must be
1156 * kept together to avoid discrepancies in behavior.
1158 * With hardware tag-based KASAN, memory tags must be set before the
1159 * page becomes unavailable via debug_pagealloc or arch_free_page.
1161 if (!skip_kasan_poison) {
1162 kasan_poison_pages(page, order, init);
1164 /* Memory is already initialized if KASAN did it internally. */
1165 if (kasan_has_integrated_init())
1169 kernel_init_pages(page, 1 << order);
1172 * arch_free_page() can make the page's contents inaccessible. s390
1173 * does this. So nothing which can access the page's contents should
1174 * happen after this.
1176 arch_free_page(page, order);
1178 debug_pagealloc_unmap_pages(page, 1 << order);
1184 * Frees a number of pages from the PCP lists
1185 * Assumes all pages on list are in same zone.
1186 * count is the number of pages to free.
1188 static void free_pcppages_bulk(struct zone *zone, int count,
1189 struct per_cpu_pages *pcp,
1192 unsigned long flags;
1194 bool isolated_pageblocks;
1198 * Ensure proper count is passed which otherwise would stuck in the
1199 * below while (list_empty(list)) loop.
1201 count = min(pcp->count, count);
1203 /* Ensure requested pindex is drained first. */
1204 pindex = pindex - 1;
1206 spin_lock_irqsave(&zone->lock, flags);
1207 isolated_pageblocks = has_isolate_pageblock(zone);
1210 struct list_head *list;
1213 /* Remove pages from lists in a round-robin fashion. */
1215 if (++pindex > NR_PCP_LISTS - 1)
1217 list = &pcp->lists[pindex];
1218 } while (list_empty(list));
1220 order = pindex_to_order(pindex);
1221 nr_pages = 1 << order;
1225 page = list_last_entry(list, struct page, pcp_list);
1226 mt = get_pcppage_migratetype(page);
1228 /* must delete to avoid corrupting pcp list */
1229 list_del(&page->pcp_list);
1231 pcp->count -= nr_pages;
1233 /* MIGRATE_ISOLATE page should not go to pcplists */
1234 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1235 /* Pageblock could have been isolated meanwhile */
1236 if (unlikely(isolated_pageblocks))
1237 mt = get_pageblock_migratetype(page);
1239 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1240 trace_mm_page_pcpu_drain(page, order, mt);
1241 } while (count > 0 && !list_empty(list));
1244 spin_unlock_irqrestore(&zone->lock, flags);
1247 static void free_one_page(struct zone *zone,
1248 struct page *page, unsigned long pfn,
1250 int migratetype, fpi_t fpi_flags)
1252 unsigned long flags;
1254 spin_lock_irqsave(&zone->lock, flags);
1255 if (unlikely(has_isolate_pageblock(zone) ||
1256 is_migrate_isolate(migratetype))) {
1257 migratetype = get_pfnblock_migratetype(page, pfn);
1259 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1260 spin_unlock_irqrestore(&zone->lock, flags);
1263 static void __free_pages_ok(struct page *page, unsigned int order,
1267 unsigned long pfn = page_to_pfn(page);
1268 struct zone *zone = page_zone(page);
1270 if (!free_pages_prepare(page, order))
1274 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1275 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1276 * This will reduce the lock holding time.
1278 migratetype = get_pfnblock_migratetype(page, pfn);
1280 free_one_page(zone, page, pfn, order, migratetype, fpi_flags);
1282 __count_vm_events(PGFREE, 1 << order);
1285 void __free_pages_core(struct page *page, unsigned int order)
1287 unsigned int nr_pages = 1 << order;
1288 struct page *p = page;
1292 * When initializing the memmap, __init_single_page() sets the refcount
1293 * of all pages to 1 ("allocated"/"not free"). We have to set the
1294 * refcount of all involved pages to 0.
1297 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1299 __ClearPageReserved(p);
1300 set_page_count(p, 0);
1302 __ClearPageReserved(p);
1303 set_page_count(p, 0);
1305 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1307 if (page_contains_unaccepted(page, order)) {
1308 if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1311 accept_page(page, order);
1315 * Bypass PCP and place fresh pages right to the tail, primarily
1316 * relevant for memory onlining.
1318 __free_pages_ok(page, order, FPI_TO_TAIL);
1322 * Check that the whole (or subset of) a pageblock given by the interval of
1323 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1324 * with the migration of free compaction scanner.
1326 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1328 * It's possible on some configurations to have a setup like node0 node1 node0
1329 * i.e. it's possible that all pages within a zones range of pages do not
1330 * belong to a single zone. We assume that a border between node0 and node1
1331 * can occur within a single pageblock, but not a node0 node1 node0
1332 * interleaving within a single pageblock. It is therefore sufficient to check
1333 * the first and last page of a pageblock and avoid checking each individual
1334 * page in a pageblock.
1336 * Note: the function may return non-NULL struct page even for a page block
1337 * which contains a memory hole (i.e. there is no physical memory for a subset
1338 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1339 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1340 * even though the start pfn is online and valid. This should be safe most of
1341 * the time because struct pages are still initialized via init_unavailable_range()
1342 * and pfn walkers shouldn't touch any physical memory range for which they do
1343 * not recognize any specific metadata in struct pages.
1345 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1346 unsigned long end_pfn, struct zone *zone)
1348 struct page *start_page;
1349 struct page *end_page;
1351 /* end_pfn is one past the range we are checking */
1354 if (!pfn_valid(end_pfn))
1357 start_page = pfn_to_online_page(start_pfn);
1361 if (page_zone(start_page) != zone)
1364 end_page = pfn_to_page(end_pfn);
1366 /* This gives a shorter code than deriving page_zone(end_page) */
1367 if (page_zone_id(start_page) != page_zone_id(end_page))
1374 * The order of subdivision here is critical for the IO subsystem.
1375 * Please do not alter this order without good reasons and regression
1376 * testing. Specifically, as large blocks of memory are subdivided,
1377 * the order in which smaller blocks are delivered depends on the order
1378 * they're subdivided in this function. This is the primary factor
1379 * influencing the order in which pages are delivered to the IO
1380 * subsystem according to empirical testing, and this is also justified
1381 * by considering the behavior of a buddy system containing a single
1382 * large block of memory acted on by a series of small allocations.
1383 * This behavior is a critical factor in sglist merging's success.
1387 static inline void expand(struct zone *zone, struct page *page,
1388 int low, int high, int migratetype)
1390 unsigned long size = 1 << high;
1392 while (high > low) {
1395 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1398 * Mark as guard pages (or page), that will allow to
1399 * merge back to allocator when buddy will be freed.
1400 * Corresponding page table entries will not be touched,
1401 * pages will stay not present in virtual address space
1403 if (set_page_guard(zone, &page[size], high, migratetype))
1406 add_to_free_list(&page[size], zone, high, migratetype);
1407 set_buddy_order(&page[size], high);
1411 static void check_new_page_bad(struct page *page)
1413 if (unlikely(page->flags & __PG_HWPOISON)) {
1414 /* Don't complain about hwpoisoned pages */
1415 page_mapcount_reset(page); /* remove PageBuddy */
1420 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1424 * This page is about to be returned from the page allocator
1426 static bool check_new_page(struct page *page)
1428 if (likely(page_expected_state(page,
1429 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1432 check_new_page_bad(page);
1436 static inline bool check_new_pages(struct page *page, unsigned int order)
1438 if (is_check_pages_enabled()) {
1439 for (int i = 0; i < (1 << order); i++) {
1440 struct page *p = page + i;
1442 if (check_new_page(p))
1450 static inline bool should_skip_kasan_unpoison(gfp_t flags)
1452 /* Don't skip if a software KASAN mode is enabled. */
1453 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1454 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1457 /* Skip, if hardware tag-based KASAN is not enabled. */
1458 if (!kasan_hw_tags_enabled())
1462 * With hardware tag-based KASAN enabled, skip if this has been
1463 * requested via __GFP_SKIP_KASAN.
1465 return flags & __GFP_SKIP_KASAN;
1468 static inline bool should_skip_init(gfp_t flags)
1470 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1471 if (!kasan_hw_tags_enabled())
1474 /* For hardware tag-based KASAN, skip if requested. */
1475 return (flags & __GFP_SKIP_ZERO);
1478 inline void post_alloc_hook(struct page *page, unsigned int order,
1481 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1482 !should_skip_init(gfp_flags);
1483 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1486 set_page_private(page, 0);
1487 set_page_refcounted(page);
1489 arch_alloc_page(page, order);
1490 debug_pagealloc_map_pages(page, 1 << order);
1493 * Page unpoisoning must happen before memory initialization.
1494 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1495 * allocations and the page unpoisoning code will complain.
1497 kernel_unpoison_pages(page, 1 << order);
1500 * As memory initialization might be integrated into KASAN,
1501 * KASAN unpoisoning and memory initializion code must be
1502 * kept together to avoid discrepancies in behavior.
1506 * If memory tags should be zeroed
1507 * (which happens only when memory should be initialized as well).
1510 /* Initialize both memory and memory tags. */
1511 for (i = 0; i != 1 << order; ++i)
1512 tag_clear_highpage(page + i);
1514 /* Take note that memory was initialized by the loop above. */
1517 if (!should_skip_kasan_unpoison(gfp_flags) &&
1518 kasan_unpoison_pages(page, order, init)) {
1519 /* Take note that memory was initialized by KASAN. */
1520 if (kasan_has_integrated_init())
1524 * If memory tags have not been set by KASAN, reset the page
1525 * tags to ensure page_address() dereferencing does not fault.
1527 for (i = 0; i != 1 << order; ++i)
1528 page_kasan_tag_reset(page + i);
1530 /* If memory is still not initialized, initialize it now. */
1532 kernel_init_pages(page, 1 << order);
1534 set_page_owner(page, order, gfp_flags);
1535 page_table_check_alloc(page, order);
1538 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1539 unsigned int alloc_flags)
1541 post_alloc_hook(page, order, gfp_flags);
1543 if (order && (gfp_flags & __GFP_COMP))
1544 prep_compound_page(page, order);
1547 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1548 * allocate the page. The expectation is that the caller is taking
1549 * steps that will free more memory. The caller should avoid the page
1550 * being used for !PFMEMALLOC purposes.
1552 if (alloc_flags & ALLOC_NO_WATERMARKS)
1553 set_page_pfmemalloc(page);
1555 clear_page_pfmemalloc(page);
1559 * Go through the free lists for the given migratetype and remove
1560 * the smallest available page from the freelists
1562 static __always_inline
1563 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1566 unsigned int current_order;
1567 struct free_area *area;
1570 /* Find a page of the appropriate size in the preferred list */
1571 for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
1572 area = &(zone->free_area[current_order]);
1573 page = get_page_from_free_area(area, migratetype);
1576 del_page_from_free_list(page, zone, current_order);
1577 expand(zone, page, order, current_order, migratetype);
1578 set_pcppage_migratetype(page, migratetype);
1579 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1580 pcp_allowed_order(order) &&
1581 migratetype < MIGRATE_PCPTYPES);
1590 * This array describes the order lists are fallen back to when
1591 * the free lists for the desirable migrate type are depleted
1593 * The other migratetypes do not have fallbacks.
1595 static int fallbacks[MIGRATE_TYPES][MIGRATE_PCPTYPES - 1] = {
1596 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1597 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1598 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1602 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1605 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1608 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1609 unsigned int order) { return NULL; }
1613 * Move the free pages in a range to the freelist tail of the requested type.
1614 * Note that start_page and end_pages are not aligned on a pageblock
1615 * boundary. If alignment is required, use move_freepages_block()
1617 static int move_freepages(struct zone *zone,
1618 unsigned long start_pfn, unsigned long end_pfn,
1619 int migratetype, int *num_movable)
1624 int pages_moved = 0;
1626 for (pfn = start_pfn; pfn <= end_pfn;) {
1627 page = pfn_to_page(pfn);
1628 if (!PageBuddy(page)) {
1630 * We assume that pages that could be isolated for
1631 * migration are movable. But we don't actually try
1632 * isolating, as that would be expensive.
1635 (PageLRU(page) || __PageMovable(page)))
1641 /* Make sure we are not inadvertently changing nodes */
1642 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1643 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1645 order = buddy_order(page);
1646 move_to_free_list(page, zone, order, migratetype);
1648 pages_moved += 1 << order;
1654 int move_freepages_block(struct zone *zone, struct page *page,
1655 int migratetype, int *num_movable)
1657 unsigned long start_pfn, end_pfn, pfn;
1662 pfn = page_to_pfn(page);
1663 start_pfn = pageblock_start_pfn(pfn);
1664 end_pfn = pageblock_end_pfn(pfn) - 1;
1666 /* Do not cross zone boundaries */
1667 if (!zone_spans_pfn(zone, start_pfn))
1669 if (!zone_spans_pfn(zone, end_pfn))
1672 return move_freepages(zone, start_pfn, end_pfn, migratetype,
1676 static void change_pageblock_range(struct page *pageblock_page,
1677 int start_order, int migratetype)
1679 int nr_pageblocks = 1 << (start_order - pageblock_order);
1681 while (nr_pageblocks--) {
1682 set_pageblock_migratetype(pageblock_page, migratetype);
1683 pageblock_page += pageblock_nr_pages;
1688 * When we are falling back to another migratetype during allocation, try to
1689 * steal extra free pages from the same pageblocks to satisfy further
1690 * allocations, instead of polluting multiple pageblocks.
1692 * If we are stealing a relatively large buddy page, it is likely there will
1693 * be more free pages in the pageblock, so try to steal them all. For
1694 * reclaimable and unmovable allocations, we steal regardless of page size,
1695 * as fragmentation caused by those allocations polluting movable pageblocks
1696 * is worse than movable allocations stealing from unmovable and reclaimable
1699 static bool can_steal_fallback(unsigned int order, int start_mt)
1702 * Leaving this order check is intended, although there is
1703 * relaxed order check in next check. The reason is that
1704 * we can actually steal whole pageblock if this condition met,
1705 * but, below check doesn't guarantee it and that is just heuristic
1706 * so could be changed anytime.
1708 if (order >= pageblock_order)
1711 if (order >= pageblock_order / 2 ||
1712 start_mt == MIGRATE_RECLAIMABLE ||
1713 start_mt == MIGRATE_UNMOVABLE ||
1714 page_group_by_mobility_disabled)
1720 static inline bool boost_watermark(struct zone *zone)
1722 unsigned long max_boost;
1724 if (!watermark_boost_factor)
1727 * Don't bother in zones that are unlikely to produce results.
1728 * On small machines, including kdump capture kernels running
1729 * in a small area, boosting the watermark can cause an out of
1730 * memory situation immediately.
1732 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1735 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1736 watermark_boost_factor, 10000);
1739 * high watermark may be uninitialised if fragmentation occurs
1740 * very early in boot so do not boost. We do not fall
1741 * through and boost by pageblock_nr_pages as failing
1742 * allocations that early means that reclaim is not going
1743 * to help and it may even be impossible to reclaim the
1744 * boosted watermark resulting in a hang.
1749 max_boost = max(pageblock_nr_pages, max_boost);
1751 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1758 * This function implements actual steal behaviour. If order is large enough,
1759 * we can steal whole pageblock. If not, we first move freepages in this
1760 * pageblock to our migratetype and determine how many already-allocated pages
1761 * are there in the pageblock with a compatible migratetype. If at least half
1762 * of pages are free or compatible, we can change migratetype of the pageblock
1763 * itself, so pages freed in the future will be put on the correct free list.
1765 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1766 unsigned int alloc_flags, int start_type, bool whole_block)
1768 unsigned int current_order = buddy_order(page);
1769 int free_pages, movable_pages, alike_pages;
1772 old_block_type = get_pageblock_migratetype(page);
1775 * This can happen due to races and we want to prevent broken
1776 * highatomic accounting.
1778 if (is_migrate_highatomic(old_block_type))
1781 /* Take ownership for orders >= pageblock_order */
1782 if (current_order >= pageblock_order) {
1783 change_pageblock_range(page, current_order, start_type);
1788 * Boost watermarks to increase reclaim pressure to reduce the
1789 * likelihood of future fallbacks. Wake kswapd now as the node
1790 * may be balanced overall and kswapd will not wake naturally.
1792 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1793 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1795 /* We are not allowed to try stealing from the whole block */
1799 free_pages = move_freepages_block(zone, page, start_type,
1801 /* moving whole block can fail due to zone boundary conditions */
1806 * Determine how many pages are compatible with our allocation.
1807 * For movable allocation, it's the number of movable pages which
1808 * we just obtained. For other types it's a bit more tricky.
1810 if (start_type == MIGRATE_MOVABLE) {
1811 alike_pages = movable_pages;
1814 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1815 * to MOVABLE pageblock, consider all non-movable pages as
1816 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1817 * vice versa, be conservative since we can't distinguish the
1818 * exact migratetype of non-movable pages.
1820 if (old_block_type == MIGRATE_MOVABLE)
1821 alike_pages = pageblock_nr_pages
1822 - (free_pages + movable_pages);
1827 * If a sufficient number of pages in the block are either free or of
1828 * compatible migratability as our allocation, claim the whole block.
1830 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1831 page_group_by_mobility_disabled)
1832 set_pageblock_migratetype(page, start_type);
1837 move_to_free_list(page, zone, current_order, start_type);
1841 * Check whether there is a suitable fallback freepage with requested order.
1842 * If only_stealable is true, this function returns fallback_mt only if
1843 * we can steal other freepages all together. This would help to reduce
1844 * fragmentation due to mixed migratetype pages in one pageblock.
1846 int find_suitable_fallback(struct free_area *area, unsigned int order,
1847 int migratetype, bool only_stealable, bool *can_steal)
1852 if (area->nr_free == 0)
1856 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
1857 fallback_mt = fallbacks[migratetype][i];
1858 if (free_area_empty(area, fallback_mt))
1861 if (can_steal_fallback(order, migratetype))
1864 if (!only_stealable)
1875 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1876 * there are no empty page blocks that contain a page with a suitable order
1878 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone)
1881 unsigned long max_managed, flags;
1884 * The number reserved as: minimum is 1 pageblock, maximum is
1885 * roughly 1% of a zone. But if 1% of a zone falls below a
1886 * pageblock size, then don't reserve any pageblocks.
1887 * Check is race-prone but harmless.
1889 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
1891 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
1892 if (zone->nr_reserved_highatomic >= max_managed)
1895 spin_lock_irqsave(&zone->lock, flags);
1897 /* Recheck the nr_reserved_highatomic limit under the lock */
1898 if (zone->nr_reserved_highatomic >= max_managed)
1902 mt = get_pageblock_migratetype(page);
1903 /* Only reserve normal pageblocks (i.e., they can merge with others) */
1904 if (migratetype_is_mergeable(mt)) {
1905 zone->nr_reserved_highatomic += pageblock_nr_pages;
1906 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1907 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
1911 spin_unlock_irqrestore(&zone->lock, flags);
1915 * Used when an allocation is about to fail under memory pressure. This
1916 * potentially hurts the reliability of high-order allocations when under
1917 * intense memory pressure but failed atomic allocations should be easier
1918 * to recover from than an OOM.
1920 * If @force is true, try to unreserve a pageblock even though highatomic
1921 * pageblock is exhausted.
1923 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
1926 struct zonelist *zonelist = ac->zonelist;
1927 unsigned long flags;
1934 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
1937 * Preserve at least one pageblock unless memory pressure
1940 if (!force && zone->nr_reserved_highatomic <=
1944 spin_lock_irqsave(&zone->lock, flags);
1945 for (order = 0; order < NR_PAGE_ORDERS; order++) {
1946 struct free_area *area = &(zone->free_area[order]);
1948 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
1953 * In page freeing path, migratetype change is racy so
1954 * we can counter several free pages in a pageblock
1955 * in this loop although we changed the pageblock type
1956 * from highatomic to ac->migratetype. So we should
1957 * adjust the count once.
1959 if (is_migrate_highatomic_page(page)) {
1961 * It should never happen but changes to
1962 * locking could inadvertently allow a per-cpu
1963 * drain to add pages to MIGRATE_HIGHATOMIC
1964 * while unreserving so be safe and watch for
1967 zone->nr_reserved_highatomic -= min(
1969 zone->nr_reserved_highatomic);
1973 * Convert to ac->migratetype and avoid the normal
1974 * pageblock stealing heuristics. Minimally, the caller
1975 * is doing the work and needs the pages. More
1976 * importantly, if the block was always converted to
1977 * MIGRATE_UNMOVABLE or another type then the number
1978 * of pageblocks that cannot be completely freed
1981 set_pageblock_migratetype(page, ac->migratetype);
1982 ret = move_freepages_block(zone, page, ac->migratetype,
1985 spin_unlock_irqrestore(&zone->lock, flags);
1989 spin_unlock_irqrestore(&zone->lock, flags);
1996 * Try finding a free buddy page on the fallback list and put it on the free
1997 * list of requested migratetype, possibly along with other pages from the same
1998 * block, depending on fragmentation avoidance heuristics. Returns true if
1999 * fallback was found so that __rmqueue_smallest() can grab it.
2001 * The use of signed ints for order and current_order is a deliberate
2002 * deviation from the rest of this file, to make the for loop
2003 * condition simpler.
2005 static __always_inline bool
2006 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2007 unsigned int alloc_flags)
2009 struct free_area *area;
2011 int min_order = order;
2017 * Do not steal pages from freelists belonging to other pageblocks
2018 * i.e. orders < pageblock_order. If there are no local zones free,
2019 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2021 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2022 min_order = pageblock_order;
2025 * Find the largest available free page in the other list. This roughly
2026 * approximates finding the pageblock with the most free pages, which
2027 * would be too costly to do exactly.
2029 for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2031 area = &(zone->free_area[current_order]);
2032 fallback_mt = find_suitable_fallback(area, current_order,
2033 start_migratetype, false, &can_steal);
2034 if (fallback_mt == -1)
2038 * We cannot steal all free pages from the pageblock and the
2039 * requested migratetype is movable. In that case it's better to
2040 * steal and split the smallest available page instead of the
2041 * largest available page, because even if the next movable
2042 * allocation falls back into a different pageblock than this
2043 * one, it won't cause permanent fragmentation.
2045 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2046 && current_order > order)
2055 for (current_order = order; current_order < NR_PAGE_ORDERS; current_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_PAGE_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))
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 batch of folios
2521 void free_unref_folios(struct folio_batch *folios)
2523 unsigned long __maybe_unused UP_flags;
2524 struct per_cpu_pages *pcp = NULL;
2525 struct zone *locked_zone = NULL;
2526 int i, j, migratetype;
2528 /* Prepare folios for freeing */
2529 for (i = 0, j = 0; i < folios->nr; i++) {
2530 struct folio *folio = folios->folios[i];
2531 unsigned long pfn = folio_pfn(folio);
2532 unsigned int order = folio_order(folio);
2534 if (order > 0 && folio_test_large_rmappable(folio))
2535 folio_undo_large_rmappable(folio);
2536 if (!free_unref_page_prepare(&folio->page, pfn, order))
2540 * Free isolated folios and orders not handled on the PCP
2541 * directly to the allocator, see comment in free_unref_page.
2543 migratetype = get_pcppage_migratetype(&folio->page);
2544 if (!pcp_allowed_order(order) ||
2545 is_migrate_isolate(migratetype)) {
2546 free_one_page(folio_zone(folio), &folio->page, pfn,
2547 order, migratetype, FPI_NONE);
2550 folio->private = (void *)(unsigned long)order;
2552 folios->folios[j] = folio;
2557 for (i = 0; i < folios->nr; i++) {
2558 struct folio *folio = folios->folios[i];
2559 struct zone *zone = folio_zone(folio);
2560 unsigned int order = (unsigned long)folio->private;
2562 folio->private = NULL;
2563 migratetype = get_pcppage_migratetype(&folio->page);
2565 /* Different zone requires a different pcp lock */
2566 if (zone != locked_zone) {
2568 pcp_spin_unlock(pcp);
2569 pcp_trylock_finish(UP_flags);
2573 * trylock is necessary as folios may be getting freed
2574 * from IRQ or SoftIRQ context after an IO completion.
2576 pcp_trylock_prepare(UP_flags);
2577 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2578 if (unlikely(!pcp)) {
2579 pcp_trylock_finish(UP_flags);
2580 free_one_page(zone, &folio->page,
2581 folio_pfn(folio), order,
2582 migratetype, FPI_NONE);
2590 * Non-isolated types over MIGRATE_PCPTYPES get added
2591 * to the MIGRATE_MOVABLE pcp list.
2593 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2594 migratetype = MIGRATE_MOVABLE;
2596 trace_mm_page_free_batched(&folio->page);
2597 free_unref_page_commit(zone, pcp, &folio->page, migratetype,
2602 pcp_spin_unlock(pcp);
2603 pcp_trylock_finish(UP_flags);
2605 folio_batch_reinit(folios);
2609 * split_page takes a non-compound higher-order page, and splits it into
2610 * n (1<<order) sub-pages: page[0..n]
2611 * Each sub-page must be freed individually.
2613 * Note: this is probably too low level an operation for use in drivers.
2614 * Please consult with lkml before using this in your driver.
2616 void split_page(struct page *page, unsigned int order)
2620 VM_BUG_ON_PAGE(PageCompound(page), page);
2621 VM_BUG_ON_PAGE(!page_count(page), page);
2623 for (i = 1; i < (1 << order); i++)
2624 set_page_refcounted(page + i);
2625 split_page_owner(page, order, 0);
2626 split_page_memcg(page, order, 0);
2628 EXPORT_SYMBOL_GPL(split_page);
2630 int __isolate_free_page(struct page *page, unsigned int order)
2632 struct zone *zone = page_zone(page);
2633 int mt = get_pageblock_migratetype(page);
2635 if (!is_migrate_isolate(mt)) {
2636 unsigned long watermark;
2638 * Obey watermarks as if the page was being allocated. We can
2639 * emulate a high-order watermark check with a raised order-0
2640 * watermark, because we already know our high-order page
2643 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2644 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2647 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2650 del_page_from_free_list(page, zone, order);
2653 * Set the pageblock if the isolated page is at least half of a
2656 if (order >= pageblock_order - 1) {
2657 struct page *endpage = page + (1 << order) - 1;
2658 for (; page < endpage; page += pageblock_nr_pages) {
2659 int mt = get_pageblock_migratetype(page);
2661 * Only change normal pageblocks (i.e., they can merge
2664 if (migratetype_is_mergeable(mt))
2665 set_pageblock_migratetype(page,
2670 return 1UL << order;
2674 * __putback_isolated_page - Return a now-isolated page back where we got it
2675 * @page: Page that was isolated
2676 * @order: Order of the isolated page
2677 * @mt: The page's pageblock's migratetype
2679 * This function is meant to return a page pulled from the free lists via
2680 * __isolate_free_page back to the free lists they were pulled from.
2682 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2684 struct zone *zone = page_zone(page);
2686 /* zone lock should be held when this function is called */
2687 lockdep_assert_held(&zone->lock);
2689 /* Return isolated page to tail of freelist. */
2690 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2691 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2695 * Update NUMA hit/miss statistics
2697 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2701 enum numa_stat_item local_stat = NUMA_LOCAL;
2703 /* skip numa counters update if numa stats is disabled */
2704 if (!static_branch_likely(&vm_numa_stat_key))
2707 if (zone_to_nid(z) != numa_node_id())
2708 local_stat = NUMA_OTHER;
2710 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2711 __count_numa_events(z, NUMA_HIT, nr_account);
2713 __count_numa_events(z, NUMA_MISS, nr_account);
2714 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2716 __count_numa_events(z, local_stat, nr_account);
2720 static __always_inline
2721 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2722 unsigned int order, unsigned int alloc_flags,
2726 unsigned long flags;
2730 spin_lock_irqsave(&zone->lock, flags);
2731 if (alloc_flags & ALLOC_HIGHATOMIC)
2732 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2734 page = __rmqueue(zone, order, migratetype, alloc_flags);
2737 * If the allocation fails, allow OOM handling access
2738 * to HIGHATOMIC reserves as failing now is worse than
2739 * failing a high-order atomic allocation in the
2742 if (!page && (alloc_flags & ALLOC_OOM))
2743 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2746 spin_unlock_irqrestore(&zone->lock, flags);
2750 __mod_zone_freepage_state(zone, -(1 << order),
2751 get_pcppage_migratetype(page));
2752 spin_unlock_irqrestore(&zone->lock, flags);
2753 } while (check_new_pages(page, order));
2755 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2756 zone_statistics(preferred_zone, zone, 1);
2761 static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
2763 int high, base_batch, batch, max_nr_alloc;
2764 int high_max, high_min;
2766 base_batch = READ_ONCE(pcp->batch);
2767 high_min = READ_ONCE(pcp->high_min);
2768 high_max = READ_ONCE(pcp->high_max);
2769 high = pcp->high = clamp(pcp->high, high_min, high_max);
2771 /* Check for PCP disabled or boot pageset */
2772 if (unlikely(high < base_batch))
2778 batch = (base_batch << pcp->alloc_factor);
2781 * If we had larger pcp->high, we could avoid to allocate from
2784 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
2785 high = pcp->high = min(high + batch, high_max);
2788 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
2790 * Double the number of pages allocated each time there is
2791 * subsequent allocation of order-0 pages without any freeing.
2793 if (batch <= max_nr_alloc &&
2794 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
2795 pcp->alloc_factor++;
2796 batch = min(batch, max_nr_alloc);
2800 * Scale batch relative to order if batch implies free pages
2801 * can be stored on the PCP. Batch can be 1 for small zones or
2802 * for boot pagesets which should never store free pages as
2803 * the pages may belong to arbitrary zones.
2806 batch = max(batch >> order, 2);
2811 /* Remove page from the per-cpu list, caller must protect the list */
2813 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2815 unsigned int alloc_flags,
2816 struct per_cpu_pages *pcp,
2817 struct list_head *list)
2822 if (list_empty(list)) {
2823 int batch = nr_pcp_alloc(pcp, zone, order);
2826 alloced = rmqueue_bulk(zone, order,
2828 migratetype, alloc_flags);
2830 pcp->count += alloced << order;
2831 if (unlikely(list_empty(list)))
2835 page = list_first_entry(list, struct page, pcp_list);
2836 list_del(&page->pcp_list);
2837 pcp->count -= 1 << order;
2838 } while (check_new_pages(page, order));
2843 /* Lock and remove page from the per-cpu list */
2844 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2845 struct zone *zone, unsigned int order,
2846 int migratetype, unsigned int alloc_flags)
2848 struct per_cpu_pages *pcp;
2849 struct list_head *list;
2851 unsigned long __maybe_unused UP_flags;
2853 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2854 pcp_trylock_prepare(UP_flags);
2855 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2857 pcp_trylock_finish(UP_flags);
2862 * On allocation, reduce the number of pages that are batch freed.
2863 * See nr_pcp_free() where free_factor is increased for subsequent
2866 pcp->free_count >>= 1;
2867 list = &pcp->lists[order_to_pindex(migratetype, order)];
2868 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
2869 pcp_spin_unlock(pcp);
2870 pcp_trylock_finish(UP_flags);
2872 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2873 zone_statistics(preferred_zone, zone, 1);
2879 * Allocate a page from the given zone.
2880 * Use pcplists for THP or "cheap" high-order allocations.
2884 * Do not instrument rmqueue() with KMSAN. This function may call
2885 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
2886 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
2887 * may call rmqueue() again, which will result in a deadlock.
2889 __no_sanitize_memory
2891 struct page *rmqueue(struct zone *preferred_zone,
2892 struct zone *zone, unsigned int order,
2893 gfp_t gfp_flags, unsigned int alloc_flags,
2899 * We most definitely don't want callers attempting to
2900 * allocate greater than order-1 page units with __GFP_NOFAIL.
2902 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2904 if (likely(pcp_allowed_order(order))) {
2905 page = rmqueue_pcplist(preferred_zone, zone, order,
2906 migratetype, alloc_flags);
2911 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
2915 /* Separate test+clear to avoid unnecessary atomics */
2916 if ((alloc_flags & ALLOC_KSWAPD) &&
2917 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
2918 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2919 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
2922 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2926 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2928 return __should_fail_alloc_page(gfp_mask, order);
2930 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
2932 static inline long __zone_watermark_unusable_free(struct zone *z,
2933 unsigned int order, unsigned int alloc_flags)
2935 long unusable_free = (1 << order) - 1;
2938 * If the caller does not have rights to reserves below the min
2939 * watermark then subtract the high-atomic reserves. This will
2940 * over-estimate the size of the atomic reserve but it avoids a search.
2942 if (likely(!(alloc_flags & ALLOC_RESERVES)))
2943 unusable_free += z->nr_reserved_highatomic;
2946 /* If allocation can't use CMA areas don't use free CMA pages */
2947 if (!(alloc_flags & ALLOC_CMA))
2948 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
2950 #ifdef CONFIG_UNACCEPTED_MEMORY
2951 unusable_free += zone_page_state(z, NR_UNACCEPTED);
2954 return unusable_free;
2958 * Return true if free base pages are above 'mark'. For high-order checks it
2959 * will return true of the order-0 watermark is reached and there is at least
2960 * one free page of a suitable size. Checking now avoids taking the zone lock
2961 * to check in the allocation paths if no pages are free.
2963 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2964 int highest_zoneidx, unsigned int alloc_flags,
2970 /* free_pages may go negative - that's OK */
2971 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
2973 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
2975 * __GFP_HIGH allows access to 50% of the min reserve as well
2978 if (alloc_flags & ALLOC_MIN_RESERVE) {
2982 * Non-blocking allocations (e.g. GFP_ATOMIC) can
2983 * access more reserves than just __GFP_HIGH. Other
2984 * non-blocking allocations requests such as GFP_NOWAIT
2985 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
2986 * access to the min reserve.
2988 if (alloc_flags & ALLOC_NON_BLOCK)
2993 * OOM victims can try even harder than the normal reserve
2994 * users on the grounds that it's definitely going to be in
2995 * the exit path shortly and free memory. Any allocation it
2996 * makes during the free path will be small and short-lived.
2998 if (alloc_flags & ALLOC_OOM)
3003 * Check watermarks for an order-0 allocation request. If these
3004 * are not met, then a high-order request also cannot go ahead
3005 * even if a suitable page happened to be free.
3007 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3010 /* If this is an order-0 request then the watermark is fine */
3014 /* For a high-order request, check at least one suitable page is free */
3015 for (o = order; o < NR_PAGE_ORDERS; o++) {
3016 struct free_area *area = &z->free_area[o];
3022 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3023 if (!free_area_empty(area, mt))
3028 if ((alloc_flags & ALLOC_CMA) &&
3029 !free_area_empty(area, MIGRATE_CMA)) {
3033 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3034 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3041 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3042 int highest_zoneidx, unsigned int alloc_flags)
3044 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3045 zone_page_state(z, NR_FREE_PAGES));
3048 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3049 unsigned long mark, int highest_zoneidx,
3050 unsigned int alloc_flags, gfp_t gfp_mask)
3054 free_pages = zone_page_state(z, NR_FREE_PAGES);
3057 * Fast check for order-0 only. If this fails then the reserves
3058 * need to be calculated.
3064 usable_free = free_pages;
3065 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3067 /* reserved may over estimate high-atomic reserves. */
3068 usable_free -= min(usable_free, reserved);
3069 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3073 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3078 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3079 * when checking the min watermark. The min watermark is the
3080 * point where boosting is ignored so that kswapd is woken up
3081 * when below the low watermark.
3083 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3084 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3085 mark = z->_watermark[WMARK_MIN];
3086 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3087 alloc_flags, free_pages);
3093 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3094 unsigned long mark, int highest_zoneidx)
3096 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3098 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3099 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3101 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3106 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3108 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3110 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3111 node_reclaim_distance;
3113 #else /* CONFIG_NUMA */
3114 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3118 #endif /* CONFIG_NUMA */
3121 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3122 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3123 * premature use of a lower zone may cause lowmem pressure problems that
3124 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3125 * probably too small. It only makes sense to spread allocations to avoid
3126 * fragmentation between the Normal and DMA32 zones.
3128 static inline unsigned int
3129 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3131 unsigned int alloc_flags;
3134 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3137 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3139 #ifdef CONFIG_ZONE_DMA32
3143 if (zone_idx(zone) != ZONE_NORMAL)
3147 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3148 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3149 * on UMA that if Normal is populated then so is DMA32.
3151 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3152 if (nr_online_nodes > 1 && !populated_zone(--zone))
3155 alloc_flags |= ALLOC_NOFRAGMENT;
3156 #endif /* CONFIG_ZONE_DMA32 */
3160 /* Must be called after current_gfp_context() which can change gfp_mask */
3161 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3162 unsigned int alloc_flags)
3165 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3166 alloc_flags |= ALLOC_CMA;
3172 * get_page_from_freelist goes through the zonelist trying to allocate
3175 static struct page *
3176 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3177 const struct alloc_context *ac)
3181 struct pglist_data *last_pgdat = NULL;
3182 bool last_pgdat_dirty_ok = false;
3187 * Scan zonelist, looking for a zone with enough free.
3188 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3190 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3191 z = ac->preferred_zoneref;
3192 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3197 if (cpusets_enabled() &&
3198 (alloc_flags & ALLOC_CPUSET) &&
3199 !__cpuset_zone_allowed(zone, gfp_mask))
3202 * When allocating a page cache page for writing, we
3203 * want to get it from a node that is within its dirty
3204 * limit, such that no single node holds more than its
3205 * proportional share of globally allowed dirty pages.
3206 * The dirty limits take into account the node's
3207 * lowmem reserves and high watermark so that kswapd
3208 * should be able to balance it without having to
3209 * write pages from its LRU list.
3211 * XXX: For now, allow allocations to potentially
3212 * exceed the per-node dirty limit in the slowpath
3213 * (spread_dirty_pages unset) before going into reclaim,
3214 * which is important when on a NUMA setup the allowed
3215 * nodes are together not big enough to reach the
3216 * global limit. The proper fix for these situations
3217 * will require awareness of nodes in the
3218 * dirty-throttling and the flusher threads.
3220 if (ac->spread_dirty_pages) {
3221 if (last_pgdat != zone->zone_pgdat) {
3222 last_pgdat = zone->zone_pgdat;
3223 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3226 if (!last_pgdat_dirty_ok)
3230 if (no_fallback && nr_online_nodes > 1 &&
3231 zone != ac->preferred_zoneref->zone) {
3235 * If moving to a remote node, retry but allow
3236 * fragmenting fallbacks. Locality is more important
3237 * than fragmentation avoidance.
3239 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3240 if (zone_to_nid(zone) != local_nid) {
3241 alloc_flags &= ~ALLOC_NOFRAGMENT;
3247 * Detect whether the number of free pages is below high
3248 * watermark. If so, we will decrease pcp->high and free
3249 * PCP pages in free path to reduce the possibility of
3250 * premature page reclaiming. Detection is done here to
3251 * avoid to do that in hotter free path.
3253 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3254 goto check_alloc_wmark;
3256 mark = high_wmark_pages(zone);
3257 if (zone_watermark_fast(zone, order, mark,
3258 ac->highest_zoneidx, alloc_flags,
3262 set_bit(ZONE_BELOW_HIGH, &zone->flags);
3265 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3266 if (!zone_watermark_fast(zone, order, mark,
3267 ac->highest_zoneidx, alloc_flags,
3271 if (has_unaccepted_memory()) {
3272 if (try_to_accept_memory(zone, order))
3276 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3278 * Watermark failed for this zone, but see if we can
3279 * grow this zone if it contains deferred pages.
3281 if (deferred_pages_enabled()) {
3282 if (_deferred_grow_zone(zone, order))
3286 /* Checked here to keep the fast path fast */
3287 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3288 if (alloc_flags & ALLOC_NO_WATERMARKS)
3291 if (!node_reclaim_enabled() ||
3292 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3295 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3297 case NODE_RECLAIM_NOSCAN:
3300 case NODE_RECLAIM_FULL:
3301 /* scanned but unreclaimable */
3304 /* did we reclaim enough */
3305 if (zone_watermark_ok(zone, order, mark,
3306 ac->highest_zoneidx, alloc_flags))
3314 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3315 gfp_mask, alloc_flags, ac->migratetype);
3317 prep_new_page(page, order, gfp_mask, alloc_flags);
3320 * If this is a high-order atomic allocation then check
3321 * if the pageblock should be reserved for the future
3323 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3324 reserve_highatomic_pageblock(page, zone);
3328 if (has_unaccepted_memory()) {
3329 if (try_to_accept_memory(zone, order))
3333 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3334 /* Try again if zone has deferred pages */
3335 if (deferred_pages_enabled()) {
3336 if (_deferred_grow_zone(zone, order))
3344 * It's possible on a UMA machine to get through all zones that are
3345 * fragmented. If avoiding fragmentation, reset and try again.
3348 alloc_flags &= ~ALLOC_NOFRAGMENT;
3355 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3357 unsigned int filter = SHOW_MEM_FILTER_NODES;
3360 * This documents exceptions given to allocations in certain
3361 * contexts that are allowed to allocate outside current's set
3364 if (!(gfp_mask & __GFP_NOMEMALLOC))
3365 if (tsk_is_oom_victim(current) ||
3366 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3367 filter &= ~SHOW_MEM_FILTER_NODES;
3368 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3369 filter &= ~SHOW_MEM_FILTER_NODES;
3371 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3374 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3376 struct va_format vaf;
3378 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3380 if ((gfp_mask & __GFP_NOWARN) ||
3381 !__ratelimit(&nopage_rs) ||
3382 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3385 va_start(args, fmt);
3388 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3389 current->comm, &vaf, gfp_mask, &gfp_mask,
3390 nodemask_pr_args(nodemask));
3393 cpuset_print_current_mems_allowed();
3396 warn_alloc_show_mem(gfp_mask, nodemask);
3399 static inline struct page *
3400 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3401 unsigned int alloc_flags,
3402 const struct alloc_context *ac)
3406 page = get_page_from_freelist(gfp_mask, order,
3407 alloc_flags|ALLOC_CPUSET, ac);
3409 * fallback to ignore cpuset restriction if our nodes
3413 page = get_page_from_freelist(gfp_mask, order,
3419 static inline struct page *
3420 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3421 const struct alloc_context *ac, unsigned long *did_some_progress)
3423 struct oom_control oc = {
3424 .zonelist = ac->zonelist,
3425 .nodemask = ac->nodemask,
3427 .gfp_mask = gfp_mask,
3432 *did_some_progress = 0;
3435 * Acquire the oom lock. If that fails, somebody else is
3436 * making progress for us.
3438 if (!mutex_trylock(&oom_lock)) {
3439 *did_some_progress = 1;
3440 schedule_timeout_uninterruptible(1);
3445 * Go through the zonelist yet one more time, keep very high watermark
3446 * here, this is only to catch a parallel oom killing, we must fail if
3447 * we're still under heavy pressure. But make sure that this reclaim
3448 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3449 * allocation which will never fail due to oom_lock already held.
3451 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3452 ~__GFP_DIRECT_RECLAIM, order,
3453 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3457 /* Coredumps can quickly deplete all memory reserves */
3458 if (current->flags & PF_DUMPCORE)
3460 /* The OOM killer will not help higher order allocs */
3461 if (order > PAGE_ALLOC_COSTLY_ORDER)
3464 * We have already exhausted all our reclaim opportunities without any
3465 * success so it is time to admit defeat. We will skip the OOM killer
3466 * because it is very likely that the caller has a more reasonable
3467 * fallback than shooting a random task.
3469 * The OOM killer may not free memory on a specific node.
3471 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3473 /* The OOM killer does not needlessly kill tasks for lowmem */
3474 if (ac->highest_zoneidx < ZONE_NORMAL)
3476 if (pm_suspended_storage())
3479 * XXX: GFP_NOFS allocations should rather fail than rely on
3480 * other request to make a forward progress.
3481 * We are in an unfortunate situation where out_of_memory cannot
3482 * do much for this context but let's try it to at least get
3483 * access to memory reserved if the current task is killed (see
3484 * out_of_memory). Once filesystems are ready to handle allocation
3485 * failures more gracefully we should just bail out here.
3488 /* Exhausted what can be done so it's blame time */
3489 if (out_of_memory(&oc) ||
3490 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3491 *did_some_progress = 1;
3494 * Help non-failing allocations by giving them access to memory
3497 if (gfp_mask & __GFP_NOFAIL)
3498 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3499 ALLOC_NO_WATERMARKS, ac);
3502 mutex_unlock(&oom_lock);
3507 * Maximum number of compaction retries with a progress before OOM
3508 * killer is consider as the only way to move forward.
3510 #define MAX_COMPACT_RETRIES 16
3512 #ifdef CONFIG_COMPACTION
3513 /* Try memory compaction for high-order allocations before reclaim */
3514 static struct page *
3515 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3516 unsigned int alloc_flags, const struct alloc_context *ac,
3517 enum compact_priority prio, enum compact_result *compact_result)
3519 struct page *page = NULL;
3520 unsigned long pflags;
3521 unsigned int noreclaim_flag;
3526 psi_memstall_enter(&pflags);
3527 delayacct_compact_start();
3528 noreclaim_flag = memalloc_noreclaim_save();
3530 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3533 memalloc_noreclaim_restore(noreclaim_flag);
3534 psi_memstall_leave(&pflags);
3535 delayacct_compact_end();
3537 if (*compact_result == COMPACT_SKIPPED)
3540 * At least in one zone compaction wasn't deferred or skipped, so let's
3541 * count a compaction stall
3543 count_vm_event(COMPACTSTALL);
3545 /* Prep a captured page if available */
3547 prep_new_page(page, order, gfp_mask, alloc_flags);
3549 /* Try get a page from the freelist if available */
3551 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3554 struct zone *zone = page_zone(page);
3556 zone->compact_blockskip_flush = false;
3557 compaction_defer_reset(zone, order, true);
3558 count_vm_event(COMPACTSUCCESS);
3563 * It's bad if compaction run occurs and fails. The most likely reason
3564 * is that pages exist, but not enough to satisfy watermarks.
3566 count_vm_event(COMPACTFAIL);
3574 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3575 enum compact_result compact_result,
3576 enum compact_priority *compact_priority,
3577 int *compaction_retries)
3579 int max_retries = MAX_COMPACT_RETRIES;
3582 int retries = *compaction_retries;
3583 enum compact_priority priority = *compact_priority;
3588 if (fatal_signal_pending(current))
3592 * Compaction was skipped due to a lack of free order-0
3593 * migration targets. Continue if reclaim can help.
3595 if (compact_result == COMPACT_SKIPPED) {
3596 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3601 * Compaction managed to coalesce some page blocks, but the
3602 * allocation failed presumably due to a race. Retry some.
3604 if (compact_result == COMPACT_SUCCESS) {
3606 * !costly requests are much more important than
3607 * __GFP_RETRY_MAYFAIL costly ones because they are de
3608 * facto nofail and invoke OOM killer to move on while
3609 * costly can fail and users are ready to cope with
3610 * that. 1/4 retries is rather arbitrary but we would
3611 * need much more detailed feedback from compaction to
3612 * make a better decision.
3614 if (order > PAGE_ALLOC_COSTLY_ORDER)
3617 if (++(*compaction_retries) <= max_retries) {
3624 * Compaction failed. Retry with increasing priority.
3626 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3627 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3629 if (*compact_priority > min_priority) {
3630 (*compact_priority)--;
3631 *compaction_retries = 0;
3635 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3639 static inline struct page *
3640 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3641 unsigned int alloc_flags, const struct alloc_context *ac,
3642 enum compact_priority prio, enum compact_result *compact_result)
3644 *compact_result = COMPACT_SKIPPED;
3649 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3650 enum compact_result compact_result,
3651 enum compact_priority *compact_priority,
3652 int *compaction_retries)
3657 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3661 * There are setups with compaction disabled which would prefer to loop
3662 * inside the allocator rather than hit the oom killer prematurely.
3663 * Let's give them a good hope and keep retrying while the order-0
3664 * watermarks are OK.
3666 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3667 ac->highest_zoneidx, ac->nodemask) {
3668 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3669 ac->highest_zoneidx, alloc_flags))
3674 #endif /* CONFIG_COMPACTION */
3676 #ifdef CONFIG_LOCKDEP
3677 static struct lockdep_map __fs_reclaim_map =
3678 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3680 static bool __need_reclaim(gfp_t gfp_mask)
3682 /* no reclaim without waiting on it */
3683 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3686 /* this guy won't enter reclaim */
3687 if (current->flags & PF_MEMALLOC)
3690 if (gfp_mask & __GFP_NOLOCKDEP)
3696 void __fs_reclaim_acquire(unsigned long ip)
3698 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3701 void __fs_reclaim_release(unsigned long ip)
3703 lock_release(&__fs_reclaim_map, ip);
3706 void fs_reclaim_acquire(gfp_t gfp_mask)
3708 gfp_mask = current_gfp_context(gfp_mask);
3710 if (__need_reclaim(gfp_mask)) {
3711 if (gfp_mask & __GFP_FS)
3712 __fs_reclaim_acquire(_RET_IP_);
3714 #ifdef CONFIG_MMU_NOTIFIER
3715 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3716 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3721 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3723 void fs_reclaim_release(gfp_t gfp_mask)
3725 gfp_mask = current_gfp_context(gfp_mask);
3727 if (__need_reclaim(gfp_mask)) {
3728 if (gfp_mask & __GFP_FS)
3729 __fs_reclaim_release(_RET_IP_);
3732 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3736 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3737 * have been rebuilt so allocation retries. Reader side does not lock and
3738 * retries the allocation if zonelist changes. Writer side is protected by the
3739 * embedded spin_lock.
3741 static DEFINE_SEQLOCK(zonelist_update_seq);
3743 static unsigned int zonelist_iter_begin(void)
3745 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3746 return read_seqbegin(&zonelist_update_seq);
3751 static unsigned int check_retry_zonelist(unsigned int seq)
3753 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3754 return read_seqretry(&zonelist_update_seq, seq);
3759 /* Perform direct synchronous page reclaim */
3760 static unsigned long
3761 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3762 const struct alloc_context *ac)
3764 unsigned int noreclaim_flag;
3765 unsigned long progress;
3769 /* We now go into synchronous reclaim */
3770 cpuset_memory_pressure_bump();
3771 fs_reclaim_acquire(gfp_mask);
3772 noreclaim_flag = memalloc_noreclaim_save();
3774 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3777 memalloc_noreclaim_restore(noreclaim_flag);
3778 fs_reclaim_release(gfp_mask);
3785 /* The really slow allocator path where we enter direct reclaim */
3786 static inline struct page *
3787 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3788 unsigned int alloc_flags, const struct alloc_context *ac,
3789 unsigned long *did_some_progress)
3791 struct page *page = NULL;
3792 unsigned long pflags;
3793 bool drained = false;
3795 psi_memstall_enter(&pflags);
3796 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3797 if (unlikely(!(*did_some_progress)))
3801 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3804 * If an allocation failed after direct reclaim, it could be because
3805 * pages are pinned on the per-cpu lists or in high alloc reserves.
3806 * Shrink them and try again
3808 if (!page && !drained) {
3809 unreserve_highatomic_pageblock(ac, false);
3810 drain_all_pages(NULL);
3815 psi_memstall_leave(&pflags);
3820 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3821 const struct alloc_context *ac)
3825 pg_data_t *last_pgdat = NULL;
3826 enum zone_type highest_zoneidx = ac->highest_zoneidx;
3828 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3830 if (!managed_zone(zone))
3832 if (last_pgdat != zone->zone_pgdat) {
3833 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3834 last_pgdat = zone->zone_pgdat;
3839 static inline unsigned int
3840 gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3842 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3845 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3846 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3847 * to save two branches.
3849 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
3850 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
3853 * The caller may dip into page reserves a bit more if the caller
3854 * cannot run direct reclaim, or if the caller has realtime scheduling
3855 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3856 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3858 alloc_flags |= (__force int)
3859 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
3861 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
3863 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3864 * if it can't schedule.
3866 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
3867 alloc_flags |= ALLOC_NON_BLOCK;
3870 alloc_flags |= ALLOC_HIGHATOMIC;
3874 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
3875 * GFP_ATOMIC) rather than fail, see the comment for
3876 * cpuset_node_allowed().
3878 if (alloc_flags & ALLOC_MIN_RESERVE)
3879 alloc_flags &= ~ALLOC_CPUSET;
3880 } else if (unlikely(rt_task(current)) && in_task())
3881 alloc_flags |= ALLOC_MIN_RESERVE;
3883 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
3888 static bool oom_reserves_allowed(struct task_struct *tsk)
3890 if (!tsk_is_oom_victim(tsk))
3894 * !MMU doesn't have oom reaper so give access to memory reserves
3895 * only to the thread with TIF_MEMDIE set
3897 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3904 * Distinguish requests which really need access to full memory
3905 * reserves from oom victims which can live with a portion of it
3907 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3909 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3911 if (gfp_mask & __GFP_MEMALLOC)
3912 return ALLOC_NO_WATERMARKS;
3913 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3914 return ALLOC_NO_WATERMARKS;
3915 if (!in_interrupt()) {
3916 if (current->flags & PF_MEMALLOC)
3917 return ALLOC_NO_WATERMARKS;
3918 else if (oom_reserves_allowed(current))
3925 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3927 return !!__gfp_pfmemalloc_flags(gfp_mask);
3931 * Checks whether it makes sense to retry the reclaim to make a forward progress
3932 * for the given allocation request.
3934 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3935 * without success, or when we couldn't even meet the watermark if we
3936 * reclaimed all remaining pages on the LRU lists.
3938 * Returns true if a retry is viable or false to enter the oom path.
3941 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3942 struct alloc_context *ac, int alloc_flags,
3943 bool did_some_progress, int *no_progress_loops)
3950 * Costly allocations might have made a progress but this doesn't mean
3951 * their order will become available due to high fragmentation so
3952 * always increment the no progress counter for them
3954 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3955 *no_progress_loops = 0;
3957 (*no_progress_loops)++;
3959 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
3964 * Keep reclaiming pages while there is a chance this will lead
3965 * somewhere. If none of the target zones can satisfy our allocation
3966 * request even if all reclaimable pages are considered then we are
3967 * screwed and have to go OOM.
3969 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3970 ac->highest_zoneidx, ac->nodemask) {
3971 unsigned long available;
3972 unsigned long reclaimable;
3973 unsigned long min_wmark = min_wmark_pages(zone);
3976 available = reclaimable = zone_reclaimable_pages(zone);
3977 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3980 * Would the allocation succeed if we reclaimed all
3981 * reclaimable pages?
3983 wmark = __zone_watermark_ok(zone, order, min_wmark,
3984 ac->highest_zoneidx, alloc_flags, available);
3985 trace_reclaim_retry_zone(z, order, reclaimable,
3986 available, min_wmark, *no_progress_loops, wmark);
3994 * Memory allocation/reclaim might be called from a WQ context and the
3995 * current implementation of the WQ concurrency control doesn't
3996 * recognize that a particular WQ is congested if the worker thread is
3997 * looping without ever sleeping. Therefore we have to do a short sleep
3998 * here rather than calling cond_resched().
4000 if (current->flags & PF_WQ_WORKER)
4001 schedule_timeout_uninterruptible(1);
4005 /* Before OOM, exhaust highatomic_reserve */
4007 return unreserve_highatomic_pageblock(ac, true);
4013 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4016 * It's possible that cpuset's mems_allowed and the nodemask from
4017 * mempolicy don't intersect. This should be normally dealt with by
4018 * policy_nodemask(), but it's possible to race with cpuset update in
4019 * such a way the check therein was true, and then it became false
4020 * before we got our cpuset_mems_cookie here.
4021 * This assumes that for all allocations, ac->nodemask can come only
4022 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4023 * when it does not intersect with the cpuset restrictions) or the
4024 * caller can deal with a violated nodemask.
4026 if (cpusets_enabled() && ac->nodemask &&
4027 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4028 ac->nodemask = NULL;
4033 * When updating a task's mems_allowed or mempolicy nodemask, it is
4034 * possible to race with parallel threads in such a way that our
4035 * allocation can fail while the mask is being updated. If we are about
4036 * to fail, check if the cpuset changed during allocation and if so,
4039 if (read_mems_allowed_retry(cpuset_mems_cookie))
4045 static inline struct page *
4046 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4047 struct alloc_context *ac)
4049 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4050 bool can_compact = gfp_compaction_allowed(gfp_mask);
4051 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4052 struct page *page = NULL;
4053 unsigned int alloc_flags;
4054 unsigned long did_some_progress;
4055 enum compact_priority compact_priority;
4056 enum compact_result compact_result;
4057 int compaction_retries;
4058 int no_progress_loops;
4059 unsigned int cpuset_mems_cookie;
4060 unsigned int zonelist_iter_cookie;
4064 compaction_retries = 0;
4065 no_progress_loops = 0;
4066 compact_priority = DEF_COMPACT_PRIORITY;
4067 cpuset_mems_cookie = read_mems_allowed_begin();
4068 zonelist_iter_cookie = zonelist_iter_begin();
4071 * The fast path uses conservative alloc_flags to succeed only until
4072 * kswapd needs to be woken up, and to avoid the cost of setting up
4073 * alloc_flags precisely. So we do that now.
4075 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4078 * We need to recalculate the starting point for the zonelist iterator
4079 * because we might have used different nodemask in the fast path, or
4080 * there was a cpuset modification and we are retrying - otherwise we
4081 * could end up iterating over non-eligible zones endlessly.
4083 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4084 ac->highest_zoneidx, ac->nodemask);
4085 if (!ac->preferred_zoneref->zone)
4089 * Check for insane configurations where the cpuset doesn't contain
4090 * any suitable zone to satisfy the request - e.g. non-movable
4091 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4093 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4094 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4095 ac->highest_zoneidx,
4096 &cpuset_current_mems_allowed);
4101 if (alloc_flags & ALLOC_KSWAPD)
4102 wake_all_kswapds(order, gfp_mask, ac);
4105 * The adjusted alloc_flags might result in immediate success, so try
4108 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4113 * For costly allocations, try direct compaction first, as it's likely
4114 * that we have enough base pages and don't need to reclaim. For non-
4115 * movable high-order allocations, do that as well, as compaction will
4116 * try prevent permanent fragmentation by migrating from blocks of the
4118 * Don't try this for allocations that are allowed to ignore
4119 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4121 if (can_direct_reclaim && can_compact &&
4123 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4124 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4125 page = __alloc_pages_direct_compact(gfp_mask, order,
4127 INIT_COMPACT_PRIORITY,
4133 * Checks for costly allocations with __GFP_NORETRY, which
4134 * includes some THP page fault allocations
4136 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4138 * If allocating entire pageblock(s) and compaction
4139 * failed because all zones are below low watermarks
4140 * or is prohibited because it recently failed at this
4141 * order, fail immediately unless the allocator has
4142 * requested compaction and reclaim retry.
4145 * - potentially very expensive because zones are far
4146 * below their low watermarks or this is part of very
4147 * bursty high order allocations,
4148 * - not guaranteed to help because isolate_freepages()
4149 * may not iterate over freed pages as part of its
4151 * - unlikely to make entire pageblocks free on its
4154 if (compact_result == COMPACT_SKIPPED ||
4155 compact_result == COMPACT_DEFERRED)
4159 * Looks like reclaim/compaction is worth trying, but
4160 * sync compaction could be very expensive, so keep
4161 * using async compaction.
4163 compact_priority = INIT_COMPACT_PRIORITY;
4168 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4169 if (alloc_flags & ALLOC_KSWAPD)
4170 wake_all_kswapds(order, gfp_mask, ac);
4172 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4174 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4175 (alloc_flags & ALLOC_KSWAPD);
4178 * Reset the nodemask and zonelist iterators if memory policies can be
4179 * ignored. These allocations are high priority and system rather than
4182 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4183 ac->nodemask = NULL;
4184 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4185 ac->highest_zoneidx, ac->nodemask);
4188 /* Attempt with potentially adjusted zonelist and alloc_flags */
4189 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4193 /* Caller is not willing to reclaim, we can't balance anything */
4194 if (!can_direct_reclaim)
4197 /* Avoid recursion of direct reclaim */
4198 if (current->flags & PF_MEMALLOC)
4201 /* Try direct reclaim and then allocating */
4202 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4203 &did_some_progress);
4207 /* Try direct compaction and then allocating */
4208 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4209 compact_priority, &compact_result);
4213 /* Do not loop if specifically requested */
4214 if (gfp_mask & __GFP_NORETRY)
4218 * Do not retry costly high order allocations unless they are
4219 * __GFP_RETRY_MAYFAIL and we can compact
4221 if (costly_order && (!can_compact ||
4222 !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4225 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4226 did_some_progress > 0, &no_progress_loops))
4230 * It doesn't make any sense to retry for the compaction if the order-0
4231 * reclaim is not able to make any progress because the current
4232 * implementation of the compaction depends on the sufficient amount
4233 * of free memory (see __compaction_suitable)
4235 if (did_some_progress > 0 && can_compact &&
4236 should_compact_retry(ac, order, alloc_flags,
4237 compact_result, &compact_priority,
4238 &compaction_retries))
4243 * Deal with possible cpuset update races or zonelist updates to avoid
4244 * a unnecessary OOM kill.
4246 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4247 check_retry_zonelist(zonelist_iter_cookie))
4250 /* Reclaim has failed us, start killing things */
4251 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4255 /* Avoid allocations with no watermarks from looping endlessly */
4256 if (tsk_is_oom_victim(current) &&
4257 (alloc_flags & ALLOC_OOM ||
4258 (gfp_mask & __GFP_NOMEMALLOC)))
4261 /* Retry as long as the OOM killer is making progress */
4262 if (did_some_progress) {
4263 no_progress_loops = 0;
4269 * Deal with possible cpuset update races or zonelist updates to avoid
4270 * a unnecessary OOM kill.
4272 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4273 check_retry_zonelist(zonelist_iter_cookie))
4277 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4280 if (gfp_mask & __GFP_NOFAIL) {
4282 * All existing users of the __GFP_NOFAIL are blockable, so warn
4283 * of any new users that actually require GFP_NOWAIT
4285 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4289 * PF_MEMALLOC request from this context is rather bizarre
4290 * because we cannot reclaim anything and only can loop waiting
4291 * for somebody to do a work for us
4293 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4296 * non failing costly orders are a hard requirement which we
4297 * are not prepared for much so let's warn about these users
4298 * so that we can identify them and convert them to something
4301 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4304 * Help non-failing allocations by giving some access to memory
4305 * reserves normally used for high priority non-blocking
4306 * allocations but do not use ALLOC_NO_WATERMARKS because this
4307 * could deplete whole memory reserves which would just make
4308 * the situation worse.
4310 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4318 warn_alloc(gfp_mask, ac->nodemask,
4319 "page allocation failure: order:%u", order);
4324 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4325 int preferred_nid, nodemask_t *nodemask,
4326 struct alloc_context *ac, gfp_t *alloc_gfp,
4327 unsigned int *alloc_flags)
4329 ac->highest_zoneidx = gfp_zone(gfp_mask);
4330 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4331 ac->nodemask = nodemask;
4332 ac->migratetype = gfp_migratetype(gfp_mask);
4334 if (cpusets_enabled()) {
4335 *alloc_gfp |= __GFP_HARDWALL;
4337 * When we are in the interrupt context, it is irrelevant
4338 * to the current task context. It means that any node ok.
4340 if (in_task() && !ac->nodemask)
4341 ac->nodemask = &cpuset_current_mems_allowed;
4343 *alloc_flags |= ALLOC_CPUSET;
4346 might_alloc(gfp_mask);
4348 if (should_fail_alloc_page(gfp_mask, order))
4351 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4353 /* Dirty zone balancing only done in the fast path */
4354 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4357 * The preferred zone is used for statistics but crucially it is
4358 * also used as the starting point for the zonelist iterator. It
4359 * may get reset for allocations that ignore memory policies.
4361 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4362 ac->highest_zoneidx, ac->nodemask);
4368 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4369 * @gfp: GFP flags for the allocation
4370 * @preferred_nid: The preferred NUMA node ID to allocate from
4371 * @nodemask: Set of nodes to allocate from, may be NULL
4372 * @nr_pages: The number of pages desired on the list or array
4373 * @page_list: Optional list to store the allocated pages
4374 * @page_array: Optional array to store the pages
4376 * This is a batched version of the page allocator that attempts to
4377 * allocate nr_pages quickly. Pages are added to page_list if page_list
4378 * is not NULL, otherwise it is assumed that the page_array is valid.
4380 * For lists, nr_pages is the number of pages that should be allocated.
4382 * For arrays, only NULL elements are populated with pages and nr_pages
4383 * is the maximum number of pages that will be stored in the array.
4385 * Returns the number of pages on the list or array.
4387 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
4388 nodemask_t *nodemask, int nr_pages,
4389 struct list_head *page_list,
4390 struct page **page_array)
4393 unsigned long __maybe_unused UP_flags;
4396 struct per_cpu_pages *pcp;
4397 struct list_head *pcp_list;
4398 struct alloc_context ac;
4400 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4401 int nr_populated = 0, nr_account = 0;
4404 * Skip populated array elements to determine if any pages need
4405 * to be allocated before disabling IRQs.
4407 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4410 /* No pages requested? */
4411 if (unlikely(nr_pages <= 0))
4414 /* Already populated array? */
4415 if (unlikely(page_array && nr_pages - nr_populated == 0))
4418 /* Bulk allocator does not support memcg accounting. */
4419 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4422 /* Use the single page allocator for one page. */
4423 if (nr_pages - nr_populated == 1)
4426 #ifdef CONFIG_PAGE_OWNER
4428 * PAGE_OWNER may recurse into the allocator to allocate space to
4429 * save the stack with pagesets.lock held. Releasing/reacquiring
4430 * removes much of the performance benefit of bulk allocation so
4431 * force the caller to allocate one page at a time as it'll have
4432 * similar performance to added complexity to the bulk allocator.
4434 if (static_branch_unlikely(&page_owner_inited))
4438 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4439 gfp &= gfp_allowed_mask;
4441 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4445 /* Find an allowed local zone that meets the low watermark. */
4446 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4449 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4450 !__cpuset_zone_allowed(zone, gfp)) {
4454 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4455 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4459 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4460 if (zone_watermark_fast(zone, 0, mark,
4461 zonelist_zone_idx(ac.preferred_zoneref),
4462 alloc_flags, gfp)) {
4468 * If there are no allowed local zones that meets the watermarks then
4469 * try to allocate a single page and reclaim if necessary.
4471 if (unlikely(!zone))
4474 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4475 pcp_trylock_prepare(UP_flags);
4476 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4480 /* Attempt the batch allocation */
4481 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4482 while (nr_populated < nr_pages) {
4484 /* Skip existing pages */
4485 if (page_array && page_array[nr_populated]) {
4490 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4492 if (unlikely(!page)) {
4493 /* Try and allocate at least one page */
4495 pcp_spin_unlock(pcp);
4502 prep_new_page(page, 0, gfp, 0);
4504 list_add(&page->lru, page_list);
4506 page_array[nr_populated] = page;
4510 pcp_spin_unlock(pcp);
4511 pcp_trylock_finish(UP_flags);
4513 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4514 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4517 return nr_populated;
4520 pcp_trylock_finish(UP_flags);
4523 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
4526 list_add(&page->lru, page_list);
4528 page_array[nr_populated] = page;
4534 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
4537 * This is the 'heart' of the zoned buddy allocator.
4539 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
4540 nodemask_t *nodemask)
4543 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4544 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4545 struct alloc_context ac = { };
4548 * There are several places where we assume that the order value is sane
4549 * so bail out early if the request is out of bound.
4551 if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
4554 gfp &= gfp_allowed_mask;
4556 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4557 * resp. GFP_NOIO which has to be inherited for all allocation requests
4558 * from a particular context which has been marked by
4559 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4560 * movable zones are not used during allocation.
4562 gfp = current_gfp_context(gfp);
4564 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4565 &alloc_gfp, &alloc_flags))
4569 * Forbid the first pass from falling back to types that fragment
4570 * memory until all local zones are considered.
4572 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4574 /* First allocation attempt */
4575 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4580 ac.spread_dirty_pages = false;
4583 * Restore the original nodemask if it was potentially replaced with
4584 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4586 ac.nodemask = nodemask;
4588 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4591 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4592 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4593 __free_pages(page, order);
4597 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4598 kmsan_alloc_page(page, order, alloc_gfp);
4602 EXPORT_SYMBOL(__alloc_pages);
4604 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
4605 nodemask_t *nodemask)
4607 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
4608 preferred_nid, nodemask);
4609 return page_rmappable_folio(page);
4611 EXPORT_SYMBOL(__folio_alloc);
4614 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4615 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4616 * you need to access high mem.
4618 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4622 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4625 return (unsigned long) page_address(page);
4627 EXPORT_SYMBOL(__get_free_pages);
4629 unsigned long get_zeroed_page(gfp_t gfp_mask)
4631 return __get_free_page(gfp_mask | __GFP_ZERO);
4633 EXPORT_SYMBOL(get_zeroed_page);
4636 * __free_pages - Free pages allocated with alloc_pages().
4637 * @page: The page pointer returned from alloc_pages().
4638 * @order: The order of the allocation.
4640 * This function can free multi-page allocations that are not compound
4641 * pages. It does not check that the @order passed in matches that of
4642 * the allocation, so it is easy to leak memory. Freeing more memory
4643 * than was allocated will probably emit a warning.
4645 * If the last reference to this page is speculative, it will be released
4646 * by put_page() which only frees the first page of a non-compound
4647 * allocation. To prevent the remaining pages from being leaked, we free
4648 * the subsequent pages here. If you want to use the page's reference
4649 * count to decide when to free the allocation, you should allocate a
4650 * compound page, and use put_page() instead of __free_pages().
4652 * Context: May be called in interrupt context or while holding a normal
4653 * spinlock, but not in NMI context or while holding a raw spinlock.
4655 void __free_pages(struct page *page, unsigned int order)
4657 /* get PageHead before we drop reference */
4658 int head = PageHead(page);
4660 if (put_page_testzero(page))
4661 free_the_page(page, order);
4664 free_the_page(page + (1 << order), order);
4666 EXPORT_SYMBOL(__free_pages);
4668 void free_pages(unsigned long addr, unsigned int order)
4671 VM_BUG_ON(!virt_addr_valid((void *)addr));
4672 __free_pages(virt_to_page((void *)addr), order);
4676 EXPORT_SYMBOL(free_pages);
4680 * An arbitrary-length arbitrary-offset area of memory which resides
4681 * within a 0 or higher order page. Multiple fragments within that page
4682 * are individually refcounted, in the page's reference counter.
4684 * The page_frag functions below provide a simple allocation framework for
4685 * page fragments. This is used by the network stack and network device
4686 * drivers to provide a backing region of memory for use as either an
4687 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4689 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4692 struct page *page = NULL;
4693 gfp_t gfp = gfp_mask;
4695 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4696 gfp_mask = (gfp_mask & ~__GFP_DIRECT_RECLAIM) | __GFP_COMP |
4697 __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC;
4698 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4699 PAGE_FRAG_CACHE_MAX_ORDER);
4700 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4702 if (unlikely(!page))
4703 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4705 nc->va = page ? page_address(page) : NULL;
4710 void page_frag_cache_drain(struct page_frag_cache *nc)
4715 __page_frag_cache_drain(virt_to_head_page(nc->va), nc->pagecnt_bias);
4718 EXPORT_SYMBOL(page_frag_cache_drain);
4720 void __page_frag_cache_drain(struct page *page, unsigned int count)
4722 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4724 if (page_ref_sub_and_test(page, count))
4725 free_the_page(page, compound_order(page));
4727 EXPORT_SYMBOL(__page_frag_cache_drain);
4729 void *__page_frag_alloc_align(struct page_frag_cache *nc,
4730 unsigned int fragsz, gfp_t gfp_mask,
4731 unsigned int align_mask)
4733 unsigned int size = PAGE_SIZE;
4737 if (unlikely(!nc->va)) {
4739 page = __page_frag_cache_refill(nc, gfp_mask);
4743 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4744 /* if size can vary use size else just use PAGE_SIZE */
4747 /* Even if we own the page, we do not use atomic_set().
4748 * This would break get_page_unless_zero() users.
4750 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4752 /* reset page count bias and offset to start of new frag */
4753 nc->pfmemalloc = page_is_pfmemalloc(page);
4754 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4758 offset = nc->offset - fragsz;
4759 if (unlikely(offset < 0)) {
4760 page = virt_to_page(nc->va);
4762 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4765 if (unlikely(nc->pfmemalloc)) {
4766 free_the_page(page, compound_order(page));
4770 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4771 /* if size can vary use size else just use PAGE_SIZE */
4774 /* OK, page count is 0, we can safely set it */
4775 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4777 /* reset page count bias and offset to start of new frag */
4778 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4779 offset = size - fragsz;
4780 if (unlikely(offset < 0)) {
4782 * The caller is trying to allocate a fragment
4783 * with fragsz > PAGE_SIZE but the cache isn't big
4784 * enough to satisfy the request, this may
4785 * happen in low memory conditions.
4786 * We don't release the cache page because
4787 * it could make memory pressure worse
4788 * so we simply return NULL here.
4795 offset &= align_mask;
4796 nc->offset = offset;
4798 return nc->va + offset;
4800 EXPORT_SYMBOL(__page_frag_alloc_align);
4803 * Frees a page fragment allocated out of either a compound or order 0 page.
4805 void page_frag_free(void *addr)
4807 struct page *page = virt_to_head_page(addr);
4809 if (unlikely(put_page_testzero(page)))
4810 free_the_page(page, compound_order(page));
4812 EXPORT_SYMBOL(page_frag_free);
4814 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4818 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4819 struct page *page = virt_to_page((void *)addr);
4820 struct page *last = page + nr;
4822 split_page_owner(page, order, 0);
4823 split_page_memcg(page, order, 0);
4824 while (page < --last)
4825 set_page_refcounted(last);
4827 last = page + (1UL << order);
4828 for (page += nr; page < last; page++)
4829 __free_pages_ok(page, 0, FPI_TO_TAIL);
4831 return (void *)addr;
4835 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4836 * @size: the number of bytes to allocate
4837 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4839 * This function is similar to alloc_pages(), except that it allocates the
4840 * minimum number of pages to satisfy the request. alloc_pages() can only
4841 * allocate memory in power-of-two pages.
4843 * This function is also limited by MAX_PAGE_ORDER.
4845 * Memory allocated by this function must be released by free_pages_exact().
4847 * Return: pointer to the allocated area or %NULL in case of error.
4849 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4851 unsigned int order = get_order(size);
4854 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4855 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4857 addr = __get_free_pages(gfp_mask, order);
4858 return make_alloc_exact(addr, order, size);
4860 EXPORT_SYMBOL(alloc_pages_exact);
4863 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4865 * @nid: the preferred node ID where memory should be allocated
4866 * @size: the number of bytes to allocate
4867 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4869 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4872 * Return: pointer to the allocated area or %NULL in case of error.
4874 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4876 unsigned int order = get_order(size);
4879 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4880 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4882 p = alloc_pages_node(nid, gfp_mask, order);
4885 return make_alloc_exact((unsigned long)page_address(p), order, size);
4889 * free_pages_exact - release memory allocated via alloc_pages_exact()
4890 * @virt: the value returned by alloc_pages_exact.
4891 * @size: size of allocation, same value as passed to alloc_pages_exact().
4893 * Release the memory allocated by a previous call to alloc_pages_exact.
4895 void free_pages_exact(void *virt, size_t size)
4897 unsigned long addr = (unsigned long)virt;
4898 unsigned long end = addr + PAGE_ALIGN(size);
4900 while (addr < end) {
4905 EXPORT_SYMBOL(free_pages_exact);
4908 * nr_free_zone_pages - count number of pages beyond high watermark
4909 * @offset: The zone index of the highest zone
4911 * nr_free_zone_pages() counts the number of pages which are beyond the
4912 * high watermark within all zones at or below a given zone index. For each
4913 * zone, the number of pages is calculated as:
4915 * nr_free_zone_pages = managed_pages - high_pages
4917 * Return: number of pages beyond high watermark.
4919 static unsigned long nr_free_zone_pages(int offset)
4924 /* Just pick one node, since fallback list is circular */
4925 unsigned long sum = 0;
4927 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4929 for_each_zone_zonelist(zone, z, zonelist, offset) {
4930 unsigned long size = zone_managed_pages(zone);
4931 unsigned long high = high_wmark_pages(zone);
4940 * nr_free_buffer_pages - count number of pages beyond high watermark
4942 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4943 * watermark within ZONE_DMA and ZONE_NORMAL.
4945 * Return: number of pages beyond high watermark within ZONE_DMA and
4948 unsigned long nr_free_buffer_pages(void)
4950 return nr_free_zone_pages(gfp_zone(GFP_USER));
4952 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4954 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4956 zoneref->zone = zone;
4957 zoneref->zone_idx = zone_idx(zone);
4961 * Builds allocation fallback zone lists.
4963 * Add all populated zones of a node to the zonelist.
4965 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4968 enum zone_type zone_type = MAX_NR_ZONES;
4973 zone = pgdat->node_zones + zone_type;
4974 if (populated_zone(zone)) {
4975 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4976 check_highest_zone(zone_type);
4978 } while (zone_type);
4985 static int __parse_numa_zonelist_order(char *s)
4988 * We used to support different zonelists modes but they turned
4989 * out to be just not useful. Let's keep the warning in place
4990 * if somebody still use the cmd line parameter so that we do
4991 * not fail it silently
4993 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4994 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5000 static char numa_zonelist_order[] = "Node";
5001 #define NUMA_ZONELIST_ORDER_LEN 16
5003 * sysctl handler for numa_zonelist_order
5005 static int numa_zonelist_order_handler(struct ctl_table *table, int write,
5006 void *buffer, size_t *length, loff_t *ppos)
5009 return __parse_numa_zonelist_order(buffer);
5010 return proc_dostring(table, write, buffer, length, ppos);
5013 static int node_load[MAX_NUMNODES];
5016 * find_next_best_node - find the next node that should appear in a given node's fallback list
5017 * @node: node whose fallback list we're appending
5018 * @used_node_mask: nodemask_t of already used nodes
5020 * We use a number of factors to determine which is the next node that should
5021 * appear on a given node's fallback list. The node should not have appeared
5022 * already in @node's fallback list, and it should be the next closest node
5023 * according to the distance array (which contains arbitrary distance values
5024 * from each node to each node in the system), and should also prefer nodes
5025 * with no CPUs, since presumably they'll have very little allocation pressure
5026 * on them otherwise.
5028 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5030 int find_next_best_node(int node, nodemask_t *used_node_mask)
5033 int min_val = INT_MAX;
5034 int best_node = NUMA_NO_NODE;
5037 * Use the local node if we haven't already, but for memoryless local
5038 * node, we should skip it and fall back to other nodes.
5040 if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5041 node_set(node, *used_node_mask);
5045 for_each_node_state(n, N_MEMORY) {
5047 /* Don't want a node to appear more than once */
5048 if (node_isset(n, *used_node_mask))
5051 /* Use the distance array to find the distance */
5052 val = node_distance(node, n);
5054 /* Penalize nodes under us ("prefer the next node") */
5057 /* Give preference to headless and unused nodes */
5058 if (!cpumask_empty(cpumask_of_node(n)))
5059 val += PENALTY_FOR_NODE_WITH_CPUS;
5061 /* Slight preference for less loaded node */
5062 val *= MAX_NUMNODES;
5063 val += node_load[n];
5065 if (val < min_val) {
5072 node_set(best_node, *used_node_mask);
5079 * Build zonelists ordered by node and zones within node.
5080 * This results in maximum locality--normal zone overflows into local
5081 * DMA zone, if any--but risks exhausting DMA zone.
5083 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5086 struct zoneref *zonerefs;
5089 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5091 for (i = 0; i < nr_nodes; i++) {
5094 pg_data_t *node = NODE_DATA(node_order[i]);
5096 nr_zones = build_zonerefs_node(node, zonerefs);
5097 zonerefs += nr_zones;
5099 zonerefs->zone = NULL;
5100 zonerefs->zone_idx = 0;
5104 * Build gfp_thisnode zonelists
5106 static void build_thisnode_zonelists(pg_data_t *pgdat)
5108 struct zoneref *zonerefs;
5111 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5112 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5113 zonerefs += nr_zones;
5114 zonerefs->zone = NULL;
5115 zonerefs->zone_idx = 0;
5119 * Build zonelists ordered by zone and nodes within zones.
5120 * This results in conserving DMA zone[s] until all Normal memory is
5121 * exhausted, but results in overflowing to remote node while memory
5122 * may still exist in local DMA zone.
5125 static void build_zonelists(pg_data_t *pgdat)
5127 static int node_order[MAX_NUMNODES];
5128 int node, nr_nodes = 0;
5129 nodemask_t used_mask = NODE_MASK_NONE;
5130 int local_node, prev_node;
5132 /* NUMA-aware ordering of nodes */
5133 local_node = pgdat->node_id;
5134 prev_node = local_node;
5136 memset(node_order, 0, sizeof(node_order));
5137 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5139 * We don't want to pressure a particular node.
5140 * So adding penalty to the first node in same
5141 * distance group to make it round-robin.
5143 if (node_distance(local_node, node) !=
5144 node_distance(local_node, prev_node))
5145 node_load[node] += 1;
5147 node_order[nr_nodes++] = node;
5151 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5152 build_thisnode_zonelists(pgdat);
5153 pr_info("Fallback order for Node %d: ", local_node);
5154 for (node = 0; node < nr_nodes; node++)
5155 pr_cont("%d ", node_order[node]);
5159 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5161 * Return node id of node used for "local" allocations.
5162 * I.e., first node id of first zone in arg node's generic zonelist.
5163 * Used for initializing percpu 'numa_mem', which is used primarily
5164 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5166 int local_memory_node(int node)
5170 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5171 gfp_zone(GFP_KERNEL),
5173 return zone_to_nid(z->zone);
5177 static void setup_min_unmapped_ratio(void);
5178 static void setup_min_slab_ratio(void);
5179 #else /* CONFIG_NUMA */
5181 static void build_zonelists(pg_data_t *pgdat)
5183 int node, local_node;
5184 struct zoneref *zonerefs;
5187 local_node = pgdat->node_id;
5189 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5190 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5191 zonerefs += nr_zones;
5194 * Now we build the zonelist so that it contains the zones
5195 * of all the other nodes.
5196 * We don't want to pressure a particular node, so when
5197 * building the zones for node N, we make sure that the
5198 * zones coming right after the local ones are those from
5199 * node N+1 (modulo N)
5201 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5202 if (!node_online(node))
5204 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5205 zonerefs += nr_zones;
5207 for (node = 0; node < local_node; node++) {
5208 if (!node_online(node))
5210 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5211 zonerefs += nr_zones;
5214 zonerefs->zone = NULL;
5215 zonerefs->zone_idx = 0;
5218 #endif /* CONFIG_NUMA */
5221 * Boot pageset table. One per cpu which is going to be used for all
5222 * zones and all nodes. The parameters will be set in such a way
5223 * that an item put on a list will immediately be handed over to
5224 * the buddy list. This is safe since pageset manipulation is done
5225 * with interrupts disabled.
5227 * The boot_pagesets must be kept even after bootup is complete for
5228 * unused processors and/or zones. They do play a role for bootstrapping
5229 * hotplugged processors.
5231 * zoneinfo_show() and maybe other functions do
5232 * not check if the processor is online before following the pageset pointer.
5233 * Other parts of the kernel may not check if the zone is available.
5235 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5236 /* These effectively disable the pcplists in the boot pageset completely */
5237 #define BOOT_PAGESET_HIGH 0
5238 #define BOOT_PAGESET_BATCH 1
5239 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5240 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5242 static void __build_all_zonelists(void *data)
5245 int __maybe_unused cpu;
5246 pg_data_t *self = data;
5247 unsigned long flags;
5250 * The zonelist_update_seq must be acquired with irqsave because the
5251 * reader can be invoked from IRQ with GFP_ATOMIC.
5253 write_seqlock_irqsave(&zonelist_update_seq, flags);
5255 * Also disable synchronous printk() to prevent any printk() from
5256 * trying to hold port->lock, for
5257 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5258 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5260 printk_deferred_enter();
5263 memset(node_load, 0, sizeof(node_load));
5267 * This node is hotadded and no memory is yet present. So just
5268 * building zonelists is fine - no need to touch other nodes.
5270 if (self && !node_online(self->node_id)) {
5271 build_zonelists(self);
5274 * All possible nodes have pgdat preallocated
5277 for_each_node(nid) {
5278 pg_data_t *pgdat = NODE_DATA(nid);
5280 build_zonelists(pgdat);
5283 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5285 * We now know the "local memory node" for each node--
5286 * i.e., the node of the first zone in the generic zonelist.
5287 * Set up numa_mem percpu variable for on-line cpus. During
5288 * boot, only the boot cpu should be on-line; we'll init the
5289 * secondary cpus' numa_mem as they come on-line. During
5290 * node/memory hotplug, we'll fixup all on-line cpus.
5292 for_each_online_cpu(cpu)
5293 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5297 printk_deferred_exit();
5298 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5301 static noinline void __init
5302 build_all_zonelists_init(void)
5306 __build_all_zonelists(NULL);
5309 * Initialize the boot_pagesets that are going to be used
5310 * for bootstrapping processors. The real pagesets for
5311 * each zone will be allocated later when the per cpu
5312 * allocator is available.
5314 * boot_pagesets are used also for bootstrapping offline
5315 * cpus if the system is already booted because the pagesets
5316 * are needed to initialize allocators on a specific cpu too.
5317 * F.e. the percpu allocator needs the page allocator which
5318 * needs the percpu allocator in order to allocate its pagesets
5319 * (a chicken-egg dilemma).
5321 for_each_possible_cpu(cpu)
5322 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5324 mminit_verify_zonelist();
5325 cpuset_init_current_mems_allowed();
5329 * unless system_state == SYSTEM_BOOTING.
5331 * __ref due to call of __init annotated helper build_all_zonelists_init
5332 * [protected by SYSTEM_BOOTING].
5334 void __ref build_all_zonelists(pg_data_t *pgdat)
5336 unsigned long vm_total_pages;
5338 if (system_state == SYSTEM_BOOTING) {
5339 build_all_zonelists_init();
5341 __build_all_zonelists(pgdat);
5342 /* cpuset refresh routine should be here */
5344 /* Get the number of free pages beyond high watermark in all zones. */
5345 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5347 * Disable grouping by mobility if the number of pages in the
5348 * system is too low to allow the mechanism to work. It would be
5349 * more accurate, but expensive to check per-zone. This check is
5350 * made on memory-hotadd so a system can start with mobility
5351 * disabled and enable it later
5353 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5354 page_group_by_mobility_disabled = 1;
5356 page_group_by_mobility_disabled = 0;
5358 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5360 page_group_by_mobility_disabled ? "off" : "on",
5363 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5367 static int zone_batchsize(struct zone *zone)
5373 * The number of pages to batch allocate is either ~0.1%
5374 * of the zone or 1MB, whichever is smaller. The batch
5375 * size is striking a balance between allocation latency
5376 * and zone lock contention.
5378 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5379 batch /= 4; /* We effectively *= 4 below */
5384 * Clamp the batch to a 2^n - 1 value. Having a power
5385 * of 2 value was found to be more likely to have
5386 * suboptimal cache aliasing properties in some cases.
5388 * For example if 2 tasks are alternately allocating
5389 * batches of pages, one task can end up with a lot
5390 * of pages of one half of the possible page colors
5391 * and the other with pages of the other colors.
5393 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5398 /* The deferral and batching of frees should be suppressed under NOMMU
5401 * The problem is that NOMMU needs to be able to allocate large chunks
5402 * of contiguous memory as there's no hardware page translation to
5403 * assemble apparent contiguous memory from discontiguous pages.
5405 * Queueing large contiguous runs of pages for batching, however,
5406 * causes the pages to actually be freed in smaller chunks. As there
5407 * can be a significant delay between the individual batches being
5408 * recycled, this leads to the once large chunks of space being
5409 * fragmented and becoming unavailable for high-order allocations.
5415 static int percpu_pagelist_high_fraction;
5416 static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5422 unsigned long total_pages;
5424 if (!high_fraction) {
5426 * By default, the high value of the pcp is based on the zone
5427 * low watermark so that if they are full then background
5428 * reclaim will not be started prematurely.
5430 total_pages = low_wmark_pages(zone);
5433 * If percpu_pagelist_high_fraction is configured, the high
5434 * value is based on a fraction of the managed pages in the
5437 total_pages = zone_managed_pages(zone) / high_fraction;
5441 * Split the high value across all online CPUs local to the zone. Note
5442 * that early in boot that CPUs may not be online yet and that during
5443 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5444 * onlined. For memory nodes that have no CPUs, split the high value
5445 * across all online CPUs to mitigate the risk that reclaim is triggered
5446 * prematurely due to pages stored on pcp lists.
5448 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5450 nr_split_cpus = num_online_cpus();
5451 high = total_pages / nr_split_cpus;
5454 * Ensure high is at least batch*4. The multiple is based on the
5455 * historical relationship between high and batch.
5457 high = max(high, batch << 2);
5466 * pcp->high and pcp->batch values are related and generally batch is lower
5467 * than high. They are also related to pcp->count such that count is lower
5468 * than high, and as soon as it reaches high, the pcplist is flushed.
5470 * However, guaranteeing these relations at all times would require e.g. write
5471 * barriers here but also careful usage of read barriers at the read side, and
5472 * thus be prone to error and bad for performance. Thus the update only prevents
5473 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5474 * should ensure they can cope with those fields changing asynchronously, and
5475 * fully trust only the pcp->count field on the local CPU with interrupts
5478 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5479 * outside of boot time (or some other assurance that no concurrent updaters
5482 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5483 unsigned long high_max, unsigned long batch)
5485 WRITE_ONCE(pcp->batch, batch);
5486 WRITE_ONCE(pcp->high_min, high_min);
5487 WRITE_ONCE(pcp->high_max, high_max);
5490 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5494 memset(pcp, 0, sizeof(*pcp));
5495 memset(pzstats, 0, sizeof(*pzstats));
5497 spin_lock_init(&pcp->lock);
5498 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5499 INIT_LIST_HEAD(&pcp->lists[pindex]);
5502 * Set batch and high values safe for a boot pageset. A true percpu
5503 * pageset's initialization will update them subsequently. Here we don't
5504 * need to be as careful as pageset_update() as nobody can access the
5507 pcp->high_min = BOOT_PAGESET_HIGH;
5508 pcp->high_max = BOOT_PAGESET_HIGH;
5509 pcp->batch = BOOT_PAGESET_BATCH;
5510 pcp->free_count = 0;
5513 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
5514 unsigned long high_max, unsigned long batch)
5516 struct per_cpu_pages *pcp;
5519 for_each_possible_cpu(cpu) {
5520 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5521 pageset_update(pcp, high_min, high_max, batch);
5526 * Calculate and set new high and batch values for all per-cpu pagesets of a
5527 * zone based on the zone's size.
5529 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5531 int new_high_min, new_high_max, new_batch;
5533 new_batch = max(1, zone_batchsize(zone));
5534 if (percpu_pagelist_high_fraction) {
5535 new_high_min = zone_highsize(zone, new_batch, cpu_online,
5536 percpu_pagelist_high_fraction);
5538 * PCP high is tuned manually, disable auto-tuning via
5539 * setting high_min and high_max to the manual value.
5541 new_high_max = new_high_min;
5543 new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
5544 new_high_max = zone_highsize(zone, new_batch, cpu_online,
5545 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
5548 if (zone->pageset_high_min == new_high_min &&
5549 zone->pageset_high_max == new_high_max &&
5550 zone->pageset_batch == new_batch)
5553 zone->pageset_high_min = new_high_min;
5554 zone->pageset_high_max = new_high_max;
5555 zone->pageset_batch = new_batch;
5557 __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
5561 void __meminit setup_zone_pageset(struct zone *zone)
5565 /* Size may be 0 on !SMP && !NUMA */
5566 if (sizeof(struct per_cpu_zonestat) > 0)
5567 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5569 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5570 for_each_possible_cpu(cpu) {
5571 struct per_cpu_pages *pcp;
5572 struct per_cpu_zonestat *pzstats;
5574 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5575 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5576 per_cpu_pages_init(pcp, pzstats);
5579 zone_set_pageset_high_and_batch(zone, 0);
5583 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5584 * page high values need to be recalculated.
5586 static void zone_pcp_update(struct zone *zone, int cpu_online)
5588 mutex_lock(&pcp_batch_high_lock);
5589 zone_set_pageset_high_and_batch(zone, cpu_online);
5590 mutex_unlock(&pcp_batch_high_lock);
5593 static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu)
5595 struct per_cpu_pages *pcp;
5596 struct cpu_cacheinfo *cci;
5598 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5599 cci = get_cpu_cacheinfo(cpu);
5601 * If data cache slice of CPU is large enough, "pcp->batch"
5602 * pages can be preserved in PCP before draining PCP for
5603 * consecutive high-order pages freeing without allocation.
5604 * This can reduce zone lock contention without hurting
5605 * cache-hot pages sharing.
5607 spin_lock(&pcp->lock);
5608 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
5609 pcp->flags |= PCPF_FREE_HIGH_BATCH;
5611 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
5612 spin_unlock(&pcp->lock);
5615 void setup_pcp_cacheinfo(unsigned int cpu)
5619 for_each_populated_zone(zone)
5620 zone_pcp_update_cacheinfo(zone, cpu);
5624 * Allocate per cpu pagesets and initialize them.
5625 * Before this call only boot pagesets were available.
5627 void __init setup_per_cpu_pageset(void)
5629 struct pglist_data *pgdat;
5631 int __maybe_unused cpu;
5633 for_each_populated_zone(zone)
5634 setup_zone_pageset(zone);
5638 * Unpopulated zones continue using the boot pagesets.
5639 * The numa stats for these pagesets need to be reset.
5640 * Otherwise, they will end up skewing the stats of
5641 * the nodes these zones are associated with.
5643 for_each_possible_cpu(cpu) {
5644 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5645 memset(pzstats->vm_numa_event, 0,
5646 sizeof(pzstats->vm_numa_event));
5650 for_each_online_pgdat(pgdat)
5651 pgdat->per_cpu_nodestats =
5652 alloc_percpu(struct per_cpu_nodestat);
5655 __meminit void zone_pcp_init(struct zone *zone)
5658 * per cpu subsystem is not up at this point. The following code
5659 * relies on the ability of the linker to provide the
5660 * offset of a (static) per cpu variable into the per cpu area.
5662 zone->per_cpu_pageset = &boot_pageset;
5663 zone->per_cpu_zonestats = &boot_zonestats;
5664 zone->pageset_high_min = BOOT_PAGESET_HIGH;
5665 zone->pageset_high_max = BOOT_PAGESET_HIGH;
5666 zone->pageset_batch = BOOT_PAGESET_BATCH;
5668 if (populated_zone(zone))
5669 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5670 zone->present_pages, zone_batchsize(zone));
5673 void adjust_managed_page_count(struct page *page, long count)
5675 atomic_long_add(count, &page_zone(page)->managed_pages);
5676 totalram_pages_add(count);
5677 #ifdef CONFIG_HIGHMEM
5678 if (PageHighMem(page))
5679 totalhigh_pages_add(count);
5682 EXPORT_SYMBOL(adjust_managed_page_count);
5684 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5687 unsigned long pages = 0;
5689 start = (void *)PAGE_ALIGN((unsigned long)start);
5690 end = (void *)((unsigned long)end & PAGE_MASK);
5691 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5692 struct page *page = virt_to_page(pos);
5693 void *direct_map_addr;
5696 * 'direct_map_addr' might be different from 'pos'
5697 * because some architectures' virt_to_page()
5698 * work with aliases. Getting the direct map
5699 * address ensures that we get a _writeable_
5700 * alias for the memset().
5702 direct_map_addr = page_address(page);
5704 * Perform a kasan-unchecked memset() since this memory
5705 * has not been initialized.
5707 direct_map_addr = kasan_reset_tag(direct_map_addr);
5708 if ((unsigned int)poison <= 0xFF)
5709 memset(direct_map_addr, poison, PAGE_SIZE);
5711 free_reserved_page(page);
5715 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5720 static int page_alloc_cpu_dead(unsigned int cpu)
5724 lru_add_drain_cpu(cpu);
5725 mlock_drain_remote(cpu);
5729 * Spill the event counters of the dead processor
5730 * into the current processors event counters.
5731 * This artificially elevates the count of the current
5734 vm_events_fold_cpu(cpu);
5737 * Zero the differential counters of the dead processor
5738 * so that the vm statistics are consistent.
5740 * This is only okay since the processor is dead and cannot
5741 * race with what we are doing.
5743 cpu_vm_stats_fold(cpu);
5745 for_each_populated_zone(zone)
5746 zone_pcp_update(zone, 0);
5751 static int page_alloc_cpu_online(unsigned int cpu)
5755 for_each_populated_zone(zone)
5756 zone_pcp_update(zone, 1);
5760 void __init page_alloc_init_cpuhp(void)
5764 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5765 "mm/page_alloc:pcp",
5766 page_alloc_cpu_online,
5767 page_alloc_cpu_dead);
5772 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5773 * or min_free_kbytes changes.
5775 static void calculate_totalreserve_pages(void)
5777 struct pglist_data *pgdat;
5778 unsigned long reserve_pages = 0;
5779 enum zone_type i, j;
5781 for_each_online_pgdat(pgdat) {
5783 pgdat->totalreserve_pages = 0;
5785 for (i = 0; i < MAX_NR_ZONES; i++) {
5786 struct zone *zone = pgdat->node_zones + i;
5788 unsigned long managed_pages = zone_managed_pages(zone);
5790 /* Find valid and maximum lowmem_reserve in the zone */
5791 for (j = i; j < MAX_NR_ZONES; j++) {
5792 if (zone->lowmem_reserve[j] > max)
5793 max = zone->lowmem_reserve[j];
5796 /* we treat the high watermark as reserved pages. */
5797 max += high_wmark_pages(zone);
5799 if (max > managed_pages)
5800 max = managed_pages;
5802 pgdat->totalreserve_pages += max;
5804 reserve_pages += max;
5807 totalreserve_pages = reserve_pages;
5811 * setup_per_zone_lowmem_reserve - called whenever
5812 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5813 * has a correct pages reserved value, so an adequate number of
5814 * pages are left in the zone after a successful __alloc_pages().
5816 static void setup_per_zone_lowmem_reserve(void)
5818 struct pglist_data *pgdat;
5819 enum zone_type i, j;
5821 for_each_online_pgdat(pgdat) {
5822 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5823 struct zone *zone = &pgdat->node_zones[i];
5824 int ratio = sysctl_lowmem_reserve_ratio[i];
5825 bool clear = !ratio || !zone_managed_pages(zone);
5826 unsigned long managed_pages = 0;
5828 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5829 struct zone *upper_zone = &pgdat->node_zones[j];
5831 managed_pages += zone_managed_pages(upper_zone);
5834 zone->lowmem_reserve[j] = 0;
5836 zone->lowmem_reserve[j] = managed_pages / ratio;
5841 /* update totalreserve_pages */
5842 calculate_totalreserve_pages();
5845 static void __setup_per_zone_wmarks(void)
5847 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5848 unsigned long lowmem_pages = 0;
5850 unsigned long flags;
5852 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5853 for_each_zone(zone) {
5854 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5855 lowmem_pages += zone_managed_pages(zone);
5858 for_each_zone(zone) {
5861 spin_lock_irqsave(&zone->lock, flags);
5862 tmp = (u64)pages_min * zone_managed_pages(zone);
5863 tmp = div64_ul(tmp, lowmem_pages);
5864 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5866 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5867 * need highmem and movable zones pages, so cap pages_min
5868 * to a small value here.
5870 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5871 * deltas control async page reclaim, and so should
5872 * not be capped for highmem and movable zones.
5874 unsigned long min_pages;
5876 min_pages = zone_managed_pages(zone) / 1024;
5877 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5878 zone->_watermark[WMARK_MIN] = min_pages;
5881 * If it's a lowmem zone, reserve a number of pages
5882 * proportionate to the zone's size.
5884 zone->_watermark[WMARK_MIN] = tmp;
5888 * Set the kswapd watermarks distance according to the
5889 * scale factor in proportion to available memory, but
5890 * ensure a minimum size on small systems.
5892 tmp = max_t(u64, tmp >> 2,
5893 mult_frac(zone_managed_pages(zone),
5894 watermark_scale_factor, 10000));
5896 zone->watermark_boost = 0;
5897 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
5898 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5899 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5901 spin_unlock_irqrestore(&zone->lock, flags);
5904 /* update totalreserve_pages */
5905 calculate_totalreserve_pages();
5909 * setup_per_zone_wmarks - called when min_free_kbytes changes
5910 * or when memory is hot-{added|removed}
5912 * Ensures that the watermark[min,low,high] values for each zone are set
5913 * correctly with respect to min_free_kbytes.
5915 void setup_per_zone_wmarks(void)
5918 static DEFINE_SPINLOCK(lock);
5921 __setup_per_zone_wmarks();
5925 * The watermark size have changed so update the pcpu batch
5926 * and high limits or the limits may be inappropriate.
5929 zone_pcp_update(zone, 0);
5933 * Initialise min_free_kbytes.
5935 * For small machines we want it small (128k min). For large machines
5936 * we want it large (256MB max). But it is not linear, because network
5937 * bandwidth does not increase linearly with machine size. We use
5939 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5940 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5956 void calculate_min_free_kbytes(void)
5958 unsigned long lowmem_kbytes;
5959 int new_min_free_kbytes;
5961 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5962 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5964 if (new_min_free_kbytes > user_min_free_kbytes)
5965 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
5967 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5968 new_min_free_kbytes, user_min_free_kbytes);
5972 int __meminit init_per_zone_wmark_min(void)
5974 calculate_min_free_kbytes();
5975 setup_per_zone_wmarks();
5976 refresh_zone_stat_thresholds();
5977 setup_per_zone_lowmem_reserve();
5980 setup_min_unmapped_ratio();
5981 setup_min_slab_ratio();
5984 khugepaged_min_free_kbytes_update();
5988 postcore_initcall(init_per_zone_wmark_min)
5991 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5992 * that we can call two helper functions whenever min_free_kbytes
5995 static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5996 void *buffer, size_t *length, loff_t *ppos)
6000 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6005 user_min_free_kbytes = min_free_kbytes;
6006 setup_per_zone_wmarks();
6011 static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6012 void *buffer, size_t *length, loff_t *ppos)
6016 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6021 setup_per_zone_wmarks();
6027 static void setup_min_unmapped_ratio(void)
6032 for_each_online_pgdat(pgdat)
6033 pgdat->min_unmapped_pages = 0;
6036 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6037 sysctl_min_unmapped_ratio) / 100;
6041 static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6042 void *buffer, size_t *length, loff_t *ppos)
6046 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6050 setup_min_unmapped_ratio();
6055 static void setup_min_slab_ratio(void)
6060 for_each_online_pgdat(pgdat)
6061 pgdat->min_slab_pages = 0;
6064 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6065 sysctl_min_slab_ratio) / 100;
6068 static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6069 void *buffer, size_t *length, loff_t *ppos)
6073 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6077 setup_min_slab_ratio();
6084 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6085 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6086 * whenever sysctl_lowmem_reserve_ratio changes.
6088 * The reserve ratio obviously has absolutely no relation with the
6089 * minimum watermarks. The lowmem reserve ratio can only make sense
6090 * if in function of the boot time zone sizes.
6092 static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table,
6093 int write, void *buffer, size_t *length, loff_t *ppos)
6097 proc_dointvec_minmax(table, write, buffer, length, ppos);
6099 for (i = 0; i < MAX_NR_ZONES; i++) {
6100 if (sysctl_lowmem_reserve_ratio[i] < 1)
6101 sysctl_lowmem_reserve_ratio[i] = 0;
6104 setup_per_zone_lowmem_reserve();
6109 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6110 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6111 * pagelist can have before it gets flushed back to buddy allocator.
6113 static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
6114 int write, void *buffer, size_t *length, loff_t *ppos)
6117 int old_percpu_pagelist_high_fraction;
6120 mutex_lock(&pcp_batch_high_lock);
6121 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6123 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6124 if (!write || ret < 0)
6127 /* Sanity checking to avoid pcp imbalance */
6128 if (percpu_pagelist_high_fraction &&
6129 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6130 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6136 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6139 for_each_populated_zone(zone)
6140 zone_set_pageset_high_and_batch(zone, 0);
6142 mutex_unlock(&pcp_batch_high_lock);
6146 static struct ctl_table page_alloc_sysctl_table[] = {
6148 .procname = "min_free_kbytes",
6149 .data = &min_free_kbytes,
6150 .maxlen = sizeof(min_free_kbytes),
6152 .proc_handler = min_free_kbytes_sysctl_handler,
6153 .extra1 = SYSCTL_ZERO,
6156 .procname = "watermark_boost_factor",
6157 .data = &watermark_boost_factor,
6158 .maxlen = sizeof(watermark_boost_factor),
6160 .proc_handler = proc_dointvec_minmax,
6161 .extra1 = SYSCTL_ZERO,
6164 .procname = "watermark_scale_factor",
6165 .data = &watermark_scale_factor,
6166 .maxlen = sizeof(watermark_scale_factor),
6168 .proc_handler = watermark_scale_factor_sysctl_handler,
6169 .extra1 = SYSCTL_ONE,
6170 .extra2 = SYSCTL_THREE_THOUSAND,
6173 .procname = "percpu_pagelist_high_fraction",
6174 .data = &percpu_pagelist_high_fraction,
6175 .maxlen = sizeof(percpu_pagelist_high_fraction),
6177 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6178 .extra1 = SYSCTL_ZERO,
6181 .procname = "lowmem_reserve_ratio",
6182 .data = &sysctl_lowmem_reserve_ratio,
6183 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6185 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6189 .procname = "numa_zonelist_order",
6190 .data = &numa_zonelist_order,
6191 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6193 .proc_handler = numa_zonelist_order_handler,
6196 .procname = "min_unmapped_ratio",
6197 .data = &sysctl_min_unmapped_ratio,
6198 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6200 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6201 .extra1 = SYSCTL_ZERO,
6202 .extra2 = SYSCTL_ONE_HUNDRED,
6205 .procname = "min_slab_ratio",
6206 .data = &sysctl_min_slab_ratio,
6207 .maxlen = sizeof(sysctl_min_slab_ratio),
6209 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6210 .extra1 = SYSCTL_ZERO,
6211 .extra2 = SYSCTL_ONE_HUNDRED,
6217 void __init page_alloc_sysctl_init(void)
6219 register_sysctl_init("vm", page_alloc_sysctl_table);
6222 #ifdef CONFIG_CONTIG_ALLOC
6223 /* Usage: See admin-guide/dynamic-debug-howto.rst */
6224 static void alloc_contig_dump_pages(struct list_head *page_list)
6226 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6228 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6232 list_for_each_entry(page, page_list, lru)
6233 dump_page(page, "migration failure");
6238 * [start, end) must belong to a single zone.
6239 * @migratetype: using migratetype to filter the type of migration in
6240 * trace_mm_alloc_contig_migrate_range_info.
6242 int __alloc_contig_migrate_range(struct compact_control *cc,
6243 unsigned long start, unsigned long end,
6246 /* This function is based on compact_zone() from compaction.c. */
6247 unsigned int nr_reclaimed;
6248 unsigned long pfn = start;
6249 unsigned int tries = 0;
6251 struct migration_target_control mtc = {
6252 .nid = zone_to_nid(cc->zone),
6253 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6256 unsigned long total_mapped = 0;
6257 unsigned long total_migrated = 0;
6258 unsigned long total_reclaimed = 0;
6260 lru_cache_disable();
6262 while (pfn < end || !list_empty(&cc->migratepages)) {
6263 if (fatal_signal_pending(current)) {
6268 if (list_empty(&cc->migratepages)) {
6269 cc->nr_migratepages = 0;
6270 ret = isolate_migratepages_range(cc, pfn, end);
6271 if (ret && ret != -EAGAIN)
6273 pfn = cc->migrate_pfn;
6275 } else if (++tries == 5) {
6280 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6282 cc->nr_migratepages -= nr_reclaimed;
6284 if (trace_mm_alloc_contig_migrate_range_info_enabled()) {
6285 total_reclaimed += nr_reclaimed;
6286 list_for_each_entry(page, &cc->migratepages, lru)
6287 total_mapped += page_mapcount(page);
6290 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6291 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6293 if (trace_mm_alloc_contig_migrate_range_info_enabled() && !ret)
6294 total_migrated += cc->nr_migratepages;
6297 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6298 * to retry again over this error, so do the same here.
6306 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6307 alloc_contig_dump_pages(&cc->migratepages);
6308 putback_movable_pages(&cc->migratepages);
6311 trace_mm_alloc_contig_migrate_range_info(start, end, migratetype,
6315 return (ret < 0) ? ret : 0;
6319 * alloc_contig_range() -- tries to allocate given range of pages
6320 * @start: start PFN to allocate
6321 * @end: one-past-the-last PFN to allocate
6322 * @migratetype: migratetype of the underlying pageblocks (either
6323 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6324 * in range must have the same migratetype and it must
6325 * be either of the two.
6326 * @gfp_mask: GFP mask to use during compaction
6328 * The PFN range does not have to be pageblock aligned. The PFN range must
6329 * belong to a single zone.
6331 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6332 * pageblocks in the range. Once isolated, the pageblocks should not
6333 * be modified by others.
6335 * Return: zero on success or negative error code. On success all
6336 * pages which PFN is in [start, end) are allocated for the caller and
6337 * need to be freed with free_contig_range().
6339 int alloc_contig_range(unsigned long start, unsigned long end,
6340 unsigned migratetype, gfp_t gfp_mask)
6342 unsigned long outer_start, outer_end;
6346 struct compact_control cc = {
6347 .nr_migratepages = 0,
6349 .zone = page_zone(pfn_to_page(start)),
6350 .mode = MIGRATE_SYNC,
6351 .ignore_skip_hint = true,
6352 .no_set_skip_hint = true,
6353 .gfp_mask = current_gfp_context(gfp_mask),
6354 .alloc_contig = true,
6356 INIT_LIST_HEAD(&cc.migratepages);
6359 * What we do here is we mark all pageblocks in range as
6360 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6361 * have different sizes, and due to the way page allocator
6362 * work, start_isolate_page_range() has special handlings for this.
6364 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6365 * migrate the pages from an unaligned range (ie. pages that
6366 * we are interested in). This will put all the pages in
6367 * range back to page allocator as MIGRATE_ISOLATE.
6369 * When this is done, we take the pages in range from page
6370 * allocator removing them from the buddy system. This way
6371 * page allocator will never consider using them.
6373 * This lets us mark the pageblocks back as
6374 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6375 * aligned range but not in the unaligned, original range are
6376 * put back to page allocator so that buddy can use them.
6379 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6383 drain_all_pages(cc.zone);
6386 * In case of -EBUSY, we'd like to know which page causes problem.
6387 * So, just fall through. test_pages_isolated() has a tracepoint
6388 * which will report the busy page.
6390 * It is possible that busy pages could become available before
6391 * the call to test_pages_isolated, and the range will actually be
6392 * allocated. So, if we fall through be sure to clear ret so that
6393 * -EBUSY is not accidentally used or returned to caller.
6395 ret = __alloc_contig_migrate_range(&cc, start, end, migratetype);
6396 if (ret && ret != -EBUSY)
6401 * Pages from [start, end) are within a pageblock_nr_pages
6402 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6403 * more, all pages in [start, end) are free in page allocator.
6404 * What we are going to do is to allocate all pages from
6405 * [start, end) (that is remove them from page allocator).
6407 * The only problem is that pages at the beginning and at the
6408 * end of interesting range may be not aligned with pages that
6409 * page allocator holds, ie. they can be part of higher order
6410 * pages. Because of this, we reserve the bigger range and
6411 * once this is done free the pages we are not interested in.
6413 * We don't have to hold zone->lock here because the pages are
6414 * isolated thus they won't get removed from buddy.
6418 outer_start = start;
6419 while (!PageBuddy(pfn_to_page(outer_start))) {
6420 if (++order > MAX_PAGE_ORDER) {
6421 outer_start = start;
6424 outer_start &= ~0UL << order;
6427 if (outer_start != start) {
6428 order = buddy_order(pfn_to_page(outer_start));
6431 * outer_start page could be small order buddy page and
6432 * it doesn't include start page. Adjust outer_start
6433 * in this case to report failed page properly
6434 * on tracepoint in test_pages_isolated()
6436 if (outer_start + (1UL << order) <= start)
6437 outer_start = start;
6440 /* Make sure the range is really isolated. */
6441 if (test_pages_isolated(outer_start, end, 0)) {
6446 /* Grab isolated pages from freelists. */
6447 outer_end = isolate_freepages_range(&cc, outer_start, end);
6453 /* Free head and tail (if any) */
6454 if (start != outer_start)
6455 free_contig_range(outer_start, start - outer_start);
6456 if (end != outer_end)
6457 free_contig_range(end, outer_end - end);
6460 undo_isolate_page_range(start, end, migratetype);
6463 EXPORT_SYMBOL(alloc_contig_range);
6465 static int __alloc_contig_pages(unsigned long start_pfn,
6466 unsigned long nr_pages, gfp_t gfp_mask)
6468 unsigned long end_pfn = start_pfn + nr_pages;
6470 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
6474 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6475 unsigned long nr_pages)
6477 unsigned long i, end_pfn = start_pfn + nr_pages;
6480 for (i = start_pfn; i < end_pfn; i++) {
6481 page = pfn_to_online_page(i);
6485 if (page_zone(page) != z)
6488 if (PageReserved(page))
6497 static bool zone_spans_last_pfn(const struct zone *zone,
6498 unsigned long start_pfn, unsigned long nr_pages)
6500 unsigned long last_pfn = start_pfn + nr_pages - 1;
6502 return zone_spans_pfn(zone, last_pfn);
6506 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6507 * @nr_pages: Number of contiguous pages to allocate
6508 * @gfp_mask: GFP mask to limit search and used during compaction
6510 * @nodemask: Mask for other possible nodes
6512 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6513 * on an applicable zonelist to find a contiguous pfn range which can then be
6514 * tried for allocation with alloc_contig_range(). This routine is intended
6515 * for allocation requests which can not be fulfilled with the buddy allocator.
6517 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6518 * power of two, then allocated range is also guaranteed to be aligned to same
6519 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6521 * Allocated pages can be freed with free_contig_range() or by manually calling
6522 * __free_page() on each allocated page.
6524 * Return: pointer to contiguous pages on success, or NULL if not successful.
6526 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
6527 int nid, nodemask_t *nodemask)
6529 unsigned long ret, pfn, flags;
6530 struct zonelist *zonelist;
6534 zonelist = node_zonelist(nid, gfp_mask);
6535 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6536 gfp_zone(gfp_mask), nodemask) {
6537 spin_lock_irqsave(&zone->lock, flags);
6539 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6540 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6541 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6543 * We release the zone lock here because
6544 * alloc_contig_range() will also lock the zone
6545 * at some point. If there's an allocation
6546 * spinning on this lock, it may win the race
6547 * and cause alloc_contig_range() to fail...
6549 spin_unlock_irqrestore(&zone->lock, flags);
6550 ret = __alloc_contig_pages(pfn, nr_pages,
6553 return pfn_to_page(pfn);
6554 spin_lock_irqsave(&zone->lock, flags);
6558 spin_unlock_irqrestore(&zone->lock, flags);
6562 #endif /* CONFIG_CONTIG_ALLOC */
6564 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6566 unsigned long count = 0;
6568 for (; nr_pages--; pfn++) {
6569 struct page *page = pfn_to_page(pfn);
6571 count += page_count(page) != 1;
6574 WARN(count != 0, "%lu pages are still in use!\n", count);
6576 EXPORT_SYMBOL(free_contig_range);
6579 * Effectively disable pcplists for the zone by setting the high limit to 0
6580 * and draining all cpus. A concurrent page freeing on another CPU that's about
6581 * to put the page on pcplist will either finish before the drain and the page
6582 * will be drained, or observe the new high limit and skip the pcplist.
6584 * Must be paired with a call to zone_pcp_enable().
6586 void zone_pcp_disable(struct zone *zone)
6588 mutex_lock(&pcp_batch_high_lock);
6589 __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
6590 __drain_all_pages(zone, true);
6593 void zone_pcp_enable(struct zone *zone)
6595 __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
6596 zone->pageset_high_max, zone->pageset_batch);
6597 mutex_unlock(&pcp_batch_high_lock);
6600 void zone_pcp_reset(struct zone *zone)
6603 struct per_cpu_zonestat *pzstats;
6605 if (zone->per_cpu_pageset != &boot_pageset) {
6606 for_each_online_cpu(cpu) {
6607 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6608 drain_zonestat(zone, pzstats);
6610 free_percpu(zone->per_cpu_pageset);
6611 zone->per_cpu_pageset = &boot_pageset;
6612 if (zone->per_cpu_zonestats != &boot_zonestats) {
6613 free_percpu(zone->per_cpu_zonestats);
6614 zone->per_cpu_zonestats = &boot_zonestats;
6619 #ifdef CONFIG_MEMORY_HOTREMOVE
6621 * All pages in the range must be in a single zone, must not contain holes,
6622 * must span full sections, and must be isolated before calling this function.
6624 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6626 unsigned long pfn = start_pfn;
6630 unsigned long flags;
6632 offline_mem_sections(pfn, end_pfn);
6633 zone = page_zone(pfn_to_page(pfn));
6634 spin_lock_irqsave(&zone->lock, flags);
6635 while (pfn < end_pfn) {
6636 page = pfn_to_page(pfn);
6638 * The HWPoisoned page may be not in buddy system, and
6639 * page_count() is not 0.
6641 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6646 * At this point all remaining PageOffline() pages have a
6647 * reference count of 0 and can simply be skipped.
6649 if (PageOffline(page)) {
6650 BUG_ON(page_count(page));
6651 BUG_ON(PageBuddy(page));
6656 BUG_ON(page_count(page));
6657 BUG_ON(!PageBuddy(page));
6658 order = buddy_order(page);
6659 del_page_from_free_list(page, zone, order);
6660 pfn += (1 << order);
6662 spin_unlock_irqrestore(&zone->lock, flags);
6667 * This function returns a stable result only if called under zone lock.
6669 bool is_free_buddy_page(struct page *page)
6671 unsigned long pfn = page_to_pfn(page);
6674 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6675 struct page *page_head = page - (pfn & ((1 << order) - 1));
6677 if (PageBuddy(page_head) &&
6678 buddy_order_unsafe(page_head) >= order)
6682 return order <= MAX_PAGE_ORDER;
6684 EXPORT_SYMBOL(is_free_buddy_page);
6686 #ifdef CONFIG_MEMORY_FAILURE
6688 * Break down a higher-order page in sub-pages, and keep our target out of
6691 static void break_down_buddy_pages(struct zone *zone, struct page *page,
6692 struct page *target, int low, int high,
6695 unsigned long size = 1 << high;
6696 struct page *current_buddy;
6698 while (high > low) {
6702 if (target >= &page[size]) {
6703 current_buddy = page;
6706 current_buddy = page + size;
6709 if (set_page_guard(zone, current_buddy, high, migratetype))
6712 add_to_free_list(current_buddy, zone, high, migratetype);
6713 set_buddy_order(current_buddy, high);
6718 * Take a page that will be marked as poisoned off the buddy allocator.
6720 bool take_page_off_buddy(struct page *page)
6722 struct zone *zone = page_zone(page);
6723 unsigned long pfn = page_to_pfn(page);
6724 unsigned long flags;
6728 spin_lock_irqsave(&zone->lock, flags);
6729 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6730 struct page *page_head = page - (pfn & ((1 << order) - 1));
6731 int page_order = buddy_order(page_head);
6733 if (PageBuddy(page_head) && page_order >= order) {
6734 unsigned long pfn_head = page_to_pfn(page_head);
6735 int migratetype = get_pfnblock_migratetype(page_head,
6738 del_page_from_free_list(page_head, zone, page_order);
6739 break_down_buddy_pages(zone, page_head, page, 0,
6740 page_order, migratetype);
6741 SetPageHWPoisonTakenOff(page);
6742 if (!is_migrate_isolate(migratetype))
6743 __mod_zone_freepage_state(zone, -1, migratetype);
6747 if (page_count(page_head) > 0)
6750 spin_unlock_irqrestore(&zone->lock, flags);
6755 * Cancel takeoff done by take_page_off_buddy().
6757 bool put_page_back_buddy(struct page *page)
6759 struct zone *zone = page_zone(page);
6760 unsigned long pfn = page_to_pfn(page);
6761 unsigned long flags;
6762 int migratetype = get_pfnblock_migratetype(page, pfn);
6765 spin_lock_irqsave(&zone->lock, flags);
6766 if (put_page_testzero(page)) {
6767 ClearPageHWPoisonTakenOff(page);
6768 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6769 if (TestClearPageHWPoison(page)) {
6773 spin_unlock_irqrestore(&zone->lock, flags);
6779 #ifdef CONFIG_ZONE_DMA
6780 bool has_managed_dma(void)
6782 struct pglist_data *pgdat;
6784 for_each_online_pgdat(pgdat) {
6785 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6787 if (managed_zone(zone))
6792 #endif /* CONFIG_ZONE_DMA */
6794 #ifdef CONFIG_UNACCEPTED_MEMORY
6796 /* Counts number of zones with unaccepted pages. */
6797 static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6799 static bool lazy_accept = true;
6801 static int __init accept_memory_parse(char *p)
6803 if (!strcmp(p, "lazy")) {
6806 } else if (!strcmp(p, "eager")) {
6807 lazy_accept = false;
6813 early_param("accept_memory", accept_memory_parse);
6815 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6817 phys_addr_t start = page_to_phys(page);
6818 phys_addr_t end = start + (PAGE_SIZE << order);
6820 return range_contains_unaccepted_memory(start, end);
6823 static void accept_page(struct page *page, unsigned int order)
6825 phys_addr_t start = page_to_phys(page);
6827 accept_memory(start, start + (PAGE_SIZE << order));
6830 static bool try_to_accept_memory_one(struct zone *zone)
6832 unsigned long flags;
6836 if (list_empty(&zone->unaccepted_pages))
6839 spin_lock_irqsave(&zone->lock, flags);
6840 page = list_first_entry_or_null(&zone->unaccepted_pages,
6843 spin_unlock_irqrestore(&zone->lock, flags);
6847 list_del(&page->lru);
6848 last = list_empty(&zone->unaccepted_pages);
6850 __mod_zone_freepage_state(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6851 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6852 spin_unlock_irqrestore(&zone->lock, flags);
6854 accept_page(page, MAX_PAGE_ORDER);
6856 __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
6859 static_branch_dec(&zones_with_unaccepted_pages);
6864 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6869 /* How much to accept to get to high watermark? */
6870 to_accept = high_wmark_pages(zone) -
6871 (zone_page_state(zone, NR_FREE_PAGES) -
6872 __zone_watermark_unusable_free(zone, order, 0));
6874 /* Accept at least one page */
6876 if (!try_to_accept_memory_one(zone))
6879 to_accept -= MAX_ORDER_NR_PAGES;
6880 } while (to_accept > 0);
6885 static inline bool has_unaccepted_memory(void)
6887 return static_branch_unlikely(&zones_with_unaccepted_pages);
6890 static bool __free_unaccepted(struct page *page)
6892 struct zone *zone = page_zone(page);
6893 unsigned long flags;
6899 spin_lock_irqsave(&zone->lock, flags);
6900 first = list_empty(&zone->unaccepted_pages);
6901 list_add_tail(&page->lru, &zone->unaccepted_pages);
6902 __mod_zone_freepage_state(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6903 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
6904 spin_unlock_irqrestore(&zone->lock, flags);
6907 static_branch_inc(&zones_with_unaccepted_pages);
6914 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6919 static void accept_page(struct page *page, unsigned int order)
6923 static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6928 static inline bool has_unaccepted_memory(void)
6933 static bool __free_unaccepted(struct page *page)
6939 #endif /* CONFIG_UNACCEPTED_MEMORY */