1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/interrupt.h>
22 #include <linux/jiffies.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.h>
25 #include <linux/kasan.h>
26 #include <linux/kmsan.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/ratelimit.h>
30 #include <linux/oom.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/pagevec.h>
36 #include <linux/memory_hotplug.h>
37 #include <linux/nodemask.h>
38 #include <linux/vmstat.h>
39 #include <linux/fault-inject.h>
40 #include <linux/compaction.h>
41 #include <trace/events/kmem.h>
42 #include <trace/events/oom.h>
43 #include <linux/prefetch.h>
44 #include <linux/mm_inline.h>
45 #include <linux/mmu_notifier.h>
46 #include <linux/migrate.h>
47 #include <linux/sched/mm.h>
48 #include <linux/page_owner.h>
49 #include <linux/page_table_check.h>
50 #include <linux/memcontrol.h>
51 #include <linux/ftrace.h>
52 #include <linux/lockdep.h>
53 #include <linux/psi.h>
54 #include <linux/khugepaged.h>
55 #include <linux/delayacct.h>
56 #include <linux/cacheinfo.h>
57 #include <linux/pgalloc_tag.h>
58 #include <asm/div64.h>
61 #include "page_reporting.h"
63 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
64 typedef int __bitwise fpi_t;
66 /* No special request */
67 #define FPI_NONE ((__force fpi_t)0)
70 * Skip free page reporting notification for the (possibly merged) page.
71 * This does not hinder free page reporting from grabbing the page,
72 * reporting it and marking it "reported" - it only skips notifying
73 * the free page reporting infrastructure about a newly freed page. For
74 * example, used when temporarily pulling a page from a freelist and
75 * putting it back unmodified.
77 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
80 * Place the (possibly merged) page to the tail of the freelist. Will ignore
81 * page shuffling (relevant code - e.g., memory onlining - is expected to
82 * shuffle the whole zone).
84 * Note: No code should rely on this flag for correctness - it's purely
85 * to allow for optimizations when handing back either fresh pages
86 * (memory onlining) or untouched pages (page isolation, free page
89 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
91 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
92 static DEFINE_MUTEX(pcp_batch_high_lock);
93 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
95 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
97 * On SMP, spin_trylock is sufficient protection.
98 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
100 #define pcp_trylock_prepare(flags) do { } while (0)
101 #define pcp_trylock_finish(flag) do { } while (0)
104 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
105 #define pcp_trylock_prepare(flags) local_irq_save(flags)
106 #define pcp_trylock_finish(flags) local_irq_restore(flags)
110 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
111 * a migration causing the wrong PCP to be locked and remote memory being
112 * potentially allocated, pin the task to the CPU for the lookup+lock.
113 * preempt_disable is used on !RT because it is faster than migrate_disable.
114 * migrate_disable is used on RT because otherwise RT spinlock usage is
115 * interfered with and a high priority task cannot preempt the allocator.
117 #ifndef CONFIG_PREEMPT_RT
118 #define pcpu_task_pin() preempt_disable()
119 #define pcpu_task_unpin() preempt_enable()
121 #define pcpu_task_pin() migrate_disable()
122 #define pcpu_task_unpin() migrate_enable()
126 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
127 * Return value should be used with equivalent unlock helper.
129 #define pcpu_spin_lock(type, member, ptr) \
133 _ret = this_cpu_ptr(ptr); \
134 spin_lock(&_ret->member); \
138 #define pcpu_spin_trylock(type, member, ptr) \
142 _ret = this_cpu_ptr(ptr); \
143 if (!spin_trylock(&_ret->member)) { \
150 #define pcpu_spin_unlock(member, ptr) \
152 spin_unlock(&ptr->member); \
156 /* struct per_cpu_pages specific helpers. */
157 #define pcp_spin_lock(ptr) \
158 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
160 #define pcp_spin_trylock(ptr) \
161 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
163 #define pcp_spin_unlock(ptr) \
164 pcpu_spin_unlock(lock, ptr)
166 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
167 DEFINE_PER_CPU(int, numa_node);
168 EXPORT_PER_CPU_SYMBOL(numa_node);
171 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
173 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
175 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
176 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
177 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
178 * defined in <linux/topology.h>.
180 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
181 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
184 static DEFINE_MUTEX(pcpu_drain_mutex);
186 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
187 volatile unsigned long latent_entropy __latent_entropy;
188 EXPORT_SYMBOL(latent_entropy);
192 * Array of node states.
194 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
195 [N_POSSIBLE] = NODE_MASK_ALL,
196 [N_ONLINE] = { { [0] = 1UL } },
198 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
199 #ifdef CONFIG_HIGHMEM
200 [N_HIGH_MEMORY] = { { [0] = 1UL } },
202 [N_MEMORY] = { { [0] = 1UL } },
203 [N_CPU] = { { [0] = 1UL } },
206 EXPORT_SYMBOL(node_states);
208 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
210 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
211 unsigned int pageblock_order __read_mostly;
214 static void __free_pages_ok(struct page *page, unsigned int order,
218 * results with 256, 32 in the lowmem_reserve sysctl:
219 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
220 * 1G machine -> (16M dma, 784M normal, 224M high)
221 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
222 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
223 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
225 * TBD: should special case ZONE_DMA32 machines here - in those we normally
226 * don't need any ZONE_NORMAL reservation
228 static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
229 #ifdef CONFIG_ZONE_DMA
232 #ifdef CONFIG_ZONE_DMA32
236 #ifdef CONFIG_HIGHMEM
242 char * const zone_names[MAX_NR_ZONES] = {
243 #ifdef CONFIG_ZONE_DMA
246 #ifdef CONFIG_ZONE_DMA32
250 #ifdef CONFIG_HIGHMEM
254 #ifdef CONFIG_ZONE_DEVICE
259 const char * const migratetype_names[MIGRATE_TYPES] = {
267 #ifdef CONFIG_MEMORY_ISOLATION
272 int min_free_kbytes = 1024;
273 int user_min_free_kbytes = -1;
274 static int watermark_boost_factor __read_mostly = 15000;
275 static int watermark_scale_factor = 10;
277 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
279 EXPORT_SYMBOL(movable_zone);
282 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
283 unsigned int nr_online_nodes __read_mostly = 1;
284 EXPORT_SYMBOL(nr_node_ids);
285 EXPORT_SYMBOL(nr_online_nodes);
288 static bool page_contains_unaccepted(struct page *page, unsigned int order);
289 static void accept_page(struct page *page, unsigned int order);
290 static bool cond_accept_memory(struct zone *zone, unsigned int order);
291 static inline bool has_unaccepted_memory(void);
292 static bool __free_unaccepted(struct page *page);
294 int page_group_by_mobility_disabled __read_mostly;
296 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
298 * During boot we initialize deferred pages on-demand, as needed, but once
299 * page_alloc_init_late() has finished, the deferred pages are all initialized,
300 * and we can permanently disable that path.
302 DEFINE_STATIC_KEY_TRUE(deferred_pages);
304 static inline bool deferred_pages_enabled(void)
306 return static_branch_unlikely(&deferred_pages);
310 * deferred_grow_zone() is __init, but it is called from
311 * get_page_from_freelist() during early boot until deferred_pages permanently
312 * disables this call. This is why we have refdata wrapper to avoid warning,
313 * and to ensure that the function body gets unloaded.
316 _deferred_grow_zone(struct zone *zone, unsigned int order)
318 return deferred_grow_zone(zone, order);
321 static inline bool deferred_pages_enabled(void)
325 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
327 /* Return a pointer to the bitmap storing bits affecting a block of pages */
328 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
331 #ifdef CONFIG_SPARSEMEM
332 return section_to_usemap(__pfn_to_section(pfn));
334 return page_zone(page)->pageblock_flags;
335 #endif /* CONFIG_SPARSEMEM */
338 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
340 #ifdef CONFIG_SPARSEMEM
341 pfn &= (PAGES_PER_SECTION-1);
343 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
344 #endif /* CONFIG_SPARSEMEM */
345 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
349 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
350 * @page: The page within the block of interest
351 * @pfn: The target page frame number
352 * @mask: mask of bits that the caller is interested in
354 * Return: pageblock_bits flags
356 unsigned long get_pfnblock_flags_mask(const struct page *page,
357 unsigned long pfn, unsigned long mask)
359 unsigned long *bitmap;
360 unsigned long bitidx, word_bitidx;
363 bitmap = get_pageblock_bitmap(page, pfn);
364 bitidx = pfn_to_bitidx(page, pfn);
365 word_bitidx = bitidx / BITS_PER_LONG;
366 bitidx &= (BITS_PER_LONG-1);
368 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
369 * a consistent read of the memory array, so that results, even though
370 * racy, are not corrupted.
372 word = READ_ONCE(bitmap[word_bitidx]);
373 return (word >> bitidx) & mask;
376 static __always_inline int get_pfnblock_migratetype(const struct page *page,
379 return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
383 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
384 * @page: The page within the block of interest
385 * @flags: The flags to set
386 * @pfn: The target page frame number
387 * @mask: mask of bits that the caller is interested in
389 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
393 unsigned long *bitmap;
394 unsigned long bitidx, word_bitidx;
397 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
398 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
400 bitmap = get_pageblock_bitmap(page, pfn);
401 bitidx = pfn_to_bitidx(page, pfn);
402 word_bitidx = bitidx / BITS_PER_LONG;
403 bitidx &= (BITS_PER_LONG-1);
405 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
410 word = READ_ONCE(bitmap[word_bitidx]);
412 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
415 void set_pageblock_migratetype(struct page *page, int migratetype)
417 if (unlikely(page_group_by_mobility_disabled &&
418 migratetype < MIGRATE_PCPTYPES))
419 migratetype = MIGRATE_UNMOVABLE;
421 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
422 page_to_pfn(page), MIGRATETYPE_MASK);
425 #ifdef CONFIG_DEBUG_VM
426 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
430 unsigned long pfn = page_to_pfn(page);
431 unsigned long sp, start_pfn;
434 seq = zone_span_seqbegin(zone);
435 start_pfn = zone->zone_start_pfn;
436 sp = zone->spanned_pages;
437 ret = !zone_spans_pfn(zone, pfn);
438 } while (zone_span_seqretry(zone, seq));
441 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
442 pfn, zone_to_nid(zone), zone->name,
443 start_pfn, start_pfn + sp);
449 * Temporary debugging check for pages not lying within a given zone.
451 static bool __maybe_unused bad_range(struct zone *zone, struct page *page)
453 if (page_outside_zone_boundaries(zone, page))
455 if (zone != page_zone(page))
461 static inline bool __maybe_unused bad_range(struct zone *zone, struct page *page)
467 static void bad_page(struct page *page, const char *reason)
469 static unsigned long resume;
470 static unsigned long nr_shown;
471 static unsigned long nr_unshown;
474 * Allow a burst of 60 reports, then keep quiet for that minute;
475 * or allow a steady drip of one report per second.
477 if (nr_shown == 60) {
478 if (time_before(jiffies, resume)) {
484 "BUG: Bad page state: %lu messages suppressed\n",
491 resume = jiffies + 60 * HZ;
493 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
494 current->comm, page_to_pfn(page));
495 dump_page(page, reason);
500 /* Leave bad fields for debug, except PageBuddy could make trouble */
502 __ClearPageBuddy(page);
503 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
506 static inline unsigned int order_to_pindex(int migratetype, int order)
508 bool __maybe_unused movable;
510 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
511 if (order > PAGE_ALLOC_COSTLY_ORDER) {
512 VM_BUG_ON(order != HPAGE_PMD_ORDER);
514 movable = migratetype == MIGRATE_MOVABLE;
516 return NR_LOWORDER_PCP_LISTS + movable;
519 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
522 return (MIGRATE_PCPTYPES * order) + migratetype;
525 static inline int pindex_to_order(unsigned int pindex)
527 int order = pindex / MIGRATE_PCPTYPES;
529 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
530 if (pindex >= NR_LOWORDER_PCP_LISTS)
531 order = HPAGE_PMD_ORDER;
533 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
539 static inline bool pcp_allowed_order(unsigned int order)
541 if (order <= PAGE_ALLOC_COSTLY_ORDER)
543 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
544 if (order == HPAGE_PMD_ORDER)
551 * Higher-order pages are called "compound pages". They are structured thusly:
553 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
555 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
556 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
558 * The first tail page's ->compound_order holds the order of allocation.
559 * This usage means that zero-order pages may not be compound.
562 void prep_compound_page(struct page *page, unsigned int order)
565 int nr_pages = 1 << order;
568 for (i = 1; i < nr_pages; i++)
569 prep_compound_tail(page, i);
571 prep_compound_head(page, order);
574 static inline void set_buddy_order(struct page *page, unsigned int order)
576 set_page_private(page, order);
577 __SetPageBuddy(page);
580 #ifdef CONFIG_COMPACTION
581 static inline struct capture_control *task_capc(struct zone *zone)
583 struct capture_control *capc = current->capture_control;
585 return unlikely(capc) &&
586 !(current->flags & PF_KTHREAD) &&
588 capc->cc->zone == zone ? capc : NULL;
592 compaction_capture(struct capture_control *capc, struct page *page,
593 int order, int migratetype)
595 if (!capc || order != capc->cc->order)
598 /* Do not accidentally pollute CMA or isolated regions*/
599 if (is_migrate_cma(migratetype) ||
600 is_migrate_isolate(migratetype))
604 * Do not let lower order allocations pollute a movable pageblock
605 * unless compaction is also requesting movable pages.
606 * This might let an unmovable request use a reclaimable pageblock
607 * and vice-versa but no more than normal fallback logic which can
608 * have trouble finding a high-order free page.
610 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE &&
611 capc->cc->migratetype != MIGRATE_MOVABLE)
619 static inline struct capture_control *task_capc(struct zone *zone)
625 compaction_capture(struct capture_control *capc, struct page *page,
626 int order, int migratetype)
630 #endif /* CONFIG_COMPACTION */
632 static inline void account_freepages(struct zone *zone, int nr_pages,
635 if (is_migrate_isolate(migratetype))
638 __mod_zone_page_state(zone, NR_FREE_PAGES, nr_pages);
640 if (is_migrate_cma(migratetype))
641 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, nr_pages);
644 /* Used for pages not on another list */
645 static inline void __add_to_free_list(struct page *page, struct zone *zone,
646 unsigned int order, int migratetype,
649 struct free_area *area = &zone->free_area[order];
651 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
652 "page type is %lu, passed migratetype is %d (nr=%d)\n",
653 get_pageblock_migratetype(page), migratetype, 1 << order);
656 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
658 list_add(&page->buddy_list, &area->free_list[migratetype]);
663 * Used for pages which are on another list. Move the pages to the tail
664 * of the list - so the moved pages won't immediately be considered for
665 * allocation again (e.g., optimization for memory onlining).
667 static inline void move_to_free_list(struct page *page, struct zone *zone,
668 unsigned int order, int old_mt, int new_mt)
670 struct free_area *area = &zone->free_area[order];
672 /* Free page moving can fail, so it happens before the type update */
673 VM_WARN_ONCE(get_pageblock_migratetype(page) != old_mt,
674 "page type is %lu, passed migratetype is %d (nr=%d)\n",
675 get_pageblock_migratetype(page), old_mt, 1 << order);
677 list_move_tail(&page->buddy_list, &area->free_list[new_mt]);
679 account_freepages(zone, -(1 << order), old_mt);
680 account_freepages(zone, 1 << order, new_mt);
683 static inline void __del_page_from_free_list(struct page *page, struct zone *zone,
684 unsigned int order, int migratetype)
686 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
687 "page type is %lu, passed migratetype is %d (nr=%d)\n",
688 get_pageblock_migratetype(page), migratetype, 1 << order);
690 /* clear reported state and update reported page count */
691 if (page_reported(page))
692 __ClearPageReported(page);
694 list_del(&page->buddy_list);
695 __ClearPageBuddy(page);
696 set_page_private(page, 0);
697 zone->free_area[order].nr_free--;
700 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
701 unsigned int order, int migratetype)
703 __del_page_from_free_list(page, zone, order, migratetype);
704 account_freepages(zone, -(1 << order), migratetype);
707 static inline struct page *get_page_from_free_area(struct free_area *area,
710 return list_first_entry_or_null(&area->free_list[migratetype],
711 struct page, buddy_list);
715 * If this is less than the 2nd largest possible page, check if the buddy
716 * of the next-higher order is free. If it is, it's possible
717 * that pages are being freed that will coalesce soon. In case,
718 * that is happening, add the free page to the tail of the list
719 * so it's less likely to be used soon and more likely to be merged
720 * as a 2-level higher order page
723 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
724 struct page *page, unsigned int order)
726 unsigned long higher_page_pfn;
727 struct page *higher_page;
729 if (order >= MAX_PAGE_ORDER - 1)
732 higher_page_pfn = buddy_pfn & pfn;
733 higher_page = page + (higher_page_pfn - pfn);
735 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
740 * Freeing function for a buddy system allocator.
742 * The concept of a buddy system is to maintain direct-mapped table
743 * (containing bit values) for memory blocks of various "orders".
744 * The bottom level table contains the map for the smallest allocatable
745 * units of memory (here, pages), and each level above it describes
746 * pairs of units from the levels below, hence, "buddies".
747 * At a high level, all that happens here is marking the table entry
748 * at the bottom level available, and propagating the changes upward
749 * as necessary, plus some accounting needed to play nicely with other
750 * parts of the VM system.
751 * At each level, we keep a list of pages, which are heads of continuous
752 * free pages of length of (1 << order) and marked with PageBuddy.
753 * Page's order is recorded in page_private(page) field.
754 * So when we are allocating or freeing one, we can derive the state of the
755 * other. That is, if we allocate a small block, and both were
756 * free, the remainder of the region must be split into blocks.
757 * If a block is freed, and its buddy is also free, then this
758 * triggers coalescing into a block of larger size.
763 static inline void __free_one_page(struct page *page,
765 struct zone *zone, unsigned int order,
766 int migratetype, fpi_t fpi_flags)
768 struct capture_control *capc = task_capc(zone);
769 unsigned long buddy_pfn = 0;
770 unsigned long combined_pfn;
774 VM_BUG_ON(!zone_is_initialized(zone));
775 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
777 VM_BUG_ON(migratetype == -1);
778 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
779 VM_BUG_ON_PAGE(bad_range(zone, page), page);
781 account_freepages(zone, 1 << order, migratetype);
783 while (order < MAX_PAGE_ORDER) {
784 int buddy_mt = migratetype;
786 if (compaction_capture(capc, page, order, migratetype)) {
787 account_freepages(zone, -(1 << order), migratetype);
791 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
795 if (unlikely(order >= pageblock_order)) {
797 * We want to prevent merge between freepages on pageblock
798 * without fallbacks and normal pageblock. Without this,
799 * pageblock isolation could cause incorrect freepage or CMA
800 * accounting or HIGHATOMIC accounting.
802 buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
804 if (migratetype != buddy_mt &&
805 (!migratetype_is_mergeable(migratetype) ||
806 !migratetype_is_mergeable(buddy_mt)))
811 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
812 * merge with it and move up one order.
814 if (page_is_guard(buddy))
815 clear_page_guard(zone, buddy, order);
817 __del_page_from_free_list(buddy, zone, order, buddy_mt);
819 if (unlikely(buddy_mt != migratetype)) {
821 * Match buddy type. This ensures that an
822 * expand() down the line puts the sub-blocks
823 * on the right freelists.
825 set_pageblock_migratetype(buddy, migratetype);
828 combined_pfn = buddy_pfn & pfn;
829 page = page + (combined_pfn - pfn);
835 set_buddy_order(page, order);
837 if (fpi_flags & FPI_TO_TAIL)
839 else if (is_shuffle_order(order))
840 to_tail = shuffle_pick_tail();
842 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
844 __add_to_free_list(page, zone, order, migratetype, to_tail);
846 /* Notify page reporting subsystem of freed page */
847 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
848 page_reporting_notify_free(order);
852 * A bad page could be due to a number of fields. Instead of multiple branches,
853 * try and check multiple fields with one check. The caller must do a detailed
854 * check if necessary.
856 static inline bool page_expected_state(struct page *page,
857 unsigned long check_flags)
859 if (unlikely(atomic_read(&page->_mapcount) != -1))
862 if (unlikely((unsigned long)page->mapping |
863 page_ref_count(page) |
867 #ifdef CONFIG_PAGE_POOL
868 ((page->pp_magic & ~0x3UL) == PP_SIGNATURE) |
870 (page->flags & check_flags)))
876 static const char *page_bad_reason(struct page *page, unsigned long flags)
878 const char *bad_reason = NULL;
880 if (unlikely(atomic_read(&page->_mapcount) != -1))
881 bad_reason = "nonzero mapcount";
882 if (unlikely(page->mapping != NULL))
883 bad_reason = "non-NULL mapping";
884 if (unlikely(page_ref_count(page) != 0))
885 bad_reason = "nonzero _refcount";
886 if (unlikely(page->flags & flags)) {
887 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
888 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
890 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
893 if (unlikely(page->memcg_data))
894 bad_reason = "page still charged to cgroup";
896 #ifdef CONFIG_PAGE_POOL
897 if (unlikely((page->pp_magic & ~0x3UL) == PP_SIGNATURE))
898 bad_reason = "page_pool leak";
903 static void free_page_is_bad_report(struct page *page)
906 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
909 static inline bool free_page_is_bad(struct page *page)
911 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
914 /* Something has gone sideways, find it */
915 free_page_is_bad_report(page);
919 static inline bool is_check_pages_enabled(void)
921 return static_branch_unlikely(&check_pages_enabled);
924 static int free_tail_page_prepare(struct page *head_page, struct page *page)
926 struct folio *folio = (struct folio *)head_page;
930 * We rely page->lru.next never has bit 0 set, unless the page
931 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
933 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
935 if (!is_check_pages_enabled()) {
939 switch (page - head_page) {
941 /* the first tail page: these may be in place of ->mapping */
942 if (unlikely(folio_entire_mapcount(folio))) {
943 bad_page(page, "nonzero entire_mapcount");
946 if (unlikely(folio_large_mapcount(folio))) {
947 bad_page(page, "nonzero large_mapcount");
950 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
951 bad_page(page, "nonzero nr_pages_mapped");
954 if (unlikely(atomic_read(&folio->_pincount))) {
955 bad_page(page, "nonzero pincount");
960 /* the second tail page: deferred_list overlaps ->mapping */
961 if (unlikely(!list_empty(&folio->_deferred_list))) {
962 bad_page(page, "on deferred list");
967 if (page->mapping != TAIL_MAPPING) {
968 bad_page(page, "corrupted mapping in tail page");
973 if (unlikely(!PageTail(page))) {
974 bad_page(page, "PageTail not set");
977 if (unlikely(compound_head(page) != head_page)) {
978 bad_page(page, "compound_head not consistent");
983 page->mapping = NULL;
984 clear_compound_head(page);
989 * Skip KASAN memory poisoning when either:
991 * 1. For generic KASAN: deferred memory initialization has not yet completed.
992 * Tag-based KASAN modes skip pages freed via deferred memory initialization
993 * using page tags instead (see below).
994 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
995 * that error detection is disabled for accesses via the page address.
997 * Pages will have match-all tags in the following circumstances:
999 * 1. Pages are being initialized for the first time, including during deferred
1000 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1001 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1002 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1003 * 3. The allocation was excluded from being checked due to sampling,
1004 * see the call to kasan_unpoison_pages.
1006 * Poisoning pages during deferred memory init will greatly lengthen the
1007 * process and cause problem in large memory systems as the deferred pages
1008 * initialization is done with interrupt disabled.
1010 * Assuming that there will be no reference to those newly initialized
1011 * pages before they are ever allocated, this should have no effect on
1012 * KASAN memory tracking as the poison will be properly inserted at page
1013 * allocation time. The only corner case is when pages are allocated by
1014 * on-demand allocation and then freed again before the deferred pages
1015 * initialization is done, but this is not likely to happen.
1017 static inline bool should_skip_kasan_poison(struct page *page)
1019 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1020 return deferred_pages_enabled();
1022 return page_kasan_tag(page) == KASAN_TAG_KERNEL;
1025 static void kernel_init_pages(struct page *page, int numpages)
1029 /* s390's use of memset() could override KASAN redzones. */
1030 kasan_disable_current();
1031 for (i = 0; i < numpages; i++)
1032 clear_highpage_kasan_tagged(page + i);
1033 kasan_enable_current();
1036 __always_inline bool free_pages_prepare(struct page *page,
1040 bool skip_kasan_poison = should_skip_kasan_poison(page);
1041 bool init = want_init_on_free();
1042 bool compound = PageCompound(page);
1044 VM_BUG_ON_PAGE(PageTail(page), page);
1046 trace_mm_page_free(page, order);
1047 kmsan_free_page(page, order);
1049 if (memcg_kmem_online() && PageMemcgKmem(page))
1050 __memcg_kmem_uncharge_page(page, order);
1052 if (unlikely(PageHWPoison(page)) && !order) {
1053 /* Do not let hwpoison pages hit pcplists/buddy */
1054 reset_page_owner(page, order);
1055 page_table_check_free(page, order);
1056 pgalloc_tag_sub(page, 1 << order);
1059 * The page is isolated and accounted for.
1060 * Mark the codetag as empty to avoid accounting error
1061 * when the page is freed by unpoison_memory().
1063 clear_page_tag_ref(page);
1067 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1070 * Check tail pages before head page information is cleared to
1071 * avoid checking PageCompound for order-0 pages.
1073 if (unlikely(order)) {
1077 page[1].flags &= ~PAGE_FLAGS_SECOND;
1078 for (i = 1; i < (1 << order); i++) {
1080 bad += free_tail_page_prepare(page, page + i);
1081 if (is_check_pages_enabled()) {
1082 if (free_page_is_bad(page + i)) {
1087 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1090 if (PageMappingFlags(page))
1091 page->mapping = NULL;
1092 if (is_check_pages_enabled()) {
1093 if (free_page_is_bad(page))
1099 page_cpupid_reset_last(page);
1100 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1101 reset_page_owner(page, order);
1102 page_table_check_free(page, order);
1103 pgalloc_tag_sub(page, 1 << order);
1105 if (!PageHighMem(page)) {
1106 debug_check_no_locks_freed(page_address(page),
1107 PAGE_SIZE << order);
1108 debug_check_no_obj_freed(page_address(page),
1109 PAGE_SIZE << order);
1112 kernel_poison_pages(page, 1 << order);
1115 * As memory initialization might be integrated into KASAN,
1116 * KASAN poisoning and memory initialization code must be
1117 * kept together to avoid discrepancies in behavior.
1119 * With hardware tag-based KASAN, memory tags must be set before the
1120 * page becomes unavailable via debug_pagealloc or arch_free_page.
1122 if (!skip_kasan_poison) {
1123 kasan_poison_pages(page, order, init);
1125 /* Memory is already initialized if KASAN did it internally. */
1126 if (kasan_has_integrated_init())
1130 kernel_init_pages(page, 1 << order);
1133 * arch_free_page() can make the page's contents inaccessible. s390
1134 * does this. So nothing which can access the page's contents should
1135 * happen after this.
1137 arch_free_page(page, order);
1139 debug_pagealloc_unmap_pages(page, 1 << order);
1145 * Frees a number of pages from the PCP lists
1146 * Assumes all pages on list are in same zone.
1147 * count is the number of pages to free.
1149 static void free_pcppages_bulk(struct zone *zone, int count,
1150 struct per_cpu_pages *pcp,
1153 unsigned long flags;
1158 * Ensure proper count is passed which otherwise would stuck in the
1159 * below while (list_empty(list)) loop.
1161 count = min(pcp->count, count);
1163 /* Ensure requested pindex is drained first. */
1164 pindex = pindex - 1;
1166 spin_lock_irqsave(&zone->lock, flags);
1169 struct list_head *list;
1172 /* Remove pages from lists in a round-robin fashion. */
1174 if (++pindex > NR_PCP_LISTS - 1)
1176 list = &pcp->lists[pindex];
1177 } while (list_empty(list));
1179 order = pindex_to_order(pindex);
1180 nr_pages = 1 << order;
1185 page = list_last_entry(list, struct page, pcp_list);
1186 pfn = page_to_pfn(page);
1187 mt = get_pfnblock_migratetype(page, pfn);
1189 /* must delete to avoid corrupting pcp list */
1190 list_del(&page->pcp_list);
1192 pcp->count -= nr_pages;
1194 __free_one_page(page, pfn, zone, order, mt, FPI_NONE);
1195 trace_mm_page_pcpu_drain(page, order, mt);
1196 } while (count > 0 && !list_empty(list));
1199 spin_unlock_irqrestore(&zone->lock, flags);
1202 static void free_one_page(struct zone *zone, struct page *page,
1203 unsigned long pfn, unsigned int order,
1206 unsigned long flags;
1209 spin_lock_irqsave(&zone->lock, flags);
1210 migratetype = get_pfnblock_migratetype(page, pfn);
1211 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1212 spin_unlock_irqrestore(&zone->lock, flags);
1215 static void __free_pages_ok(struct page *page, unsigned int order,
1218 unsigned long pfn = page_to_pfn(page);
1219 struct zone *zone = page_zone(page);
1221 if (!free_pages_prepare(page, order))
1224 free_one_page(zone, page, pfn, order, fpi_flags);
1226 __count_vm_events(PGFREE, 1 << order);
1229 void __meminit __free_pages_core(struct page *page, unsigned int order,
1230 enum meminit_context context)
1232 unsigned int nr_pages = 1 << order;
1233 struct page *p = page;
1237 * When initializing the memmap, __init_single_page() sets the refcount
1238 * of all pages to 1 ("allocated"/"not free"). We have to set the
1239 * refcount of all involved pages to 0.
1241 * Note that hotplugged memory pages are initialized to PageOffline().
1242 * Pages freed from memblock might be marked as reserved.
1244 if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG) &&
1245 unlikely(context == MEMINIT_HOTPLUG)) {
1246 for (loop = 0; loop < nr_pages; loop++, p++) {
1247 VM_WARN_ON_ONCE(PageReserved(p));
1248 __ClearPageOffline(p);
1249 set_page_count(p, 0);
1253 * Freeing the page with debug_pagealloc enabled will try to
1254 * unmap it; some archs don't like double-unmappings, so
1257 debug_pagealloc_map_pages(page, nr_pages);
1258 adjust_managed_page_count(page, nr_pages);
1260 for (loop = 0; loop < nr_pages; loop++, p++) {
1261 __ClearPageReserved(p);
1262 set_page_count(p, 0);
1265 /* memblock adjusts totalram_pages() manually. */
1266 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1269 if (page_contains_unaccepted(page, order)) {
1270 if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1273 accept_page(page, order);
1277 * Bypass PCP and place fresh pages right to the tail, primarily
1278 * relevant for memory onlining.
1280 __free_pages_ok(page, order, FPI_TO_TAIL);
1284 * Check that the whole (or subset of) a pageblock given by the interval of
1285 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1286 * with the migration of free compaction scanner.
1288 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1290 * It's possible on some configurations to have a setup like node0 node1 node0
1291 * i.e. it's possible that all pages within a zones range of pages do not
1292 * belong to a single zone. We assume that a border between node0 and node1
1293 * can occur within a single pageblock, but not a node0 node1 node0
1294 * interleaving within a single pageblock. It is therefore sufficient to check
1295 * the first and last page of a pageblock and avoid checking each individual
1296 * page in a pageblock.
1298 * Note: the function may return non-NULL struct page even for a page block
1299 * which contains a memory hole (i.e. there is no physical memory for a subset
1300 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1301 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1302 * even though the start pfn is online and valid. This should be safe most of
1303 * the time because struct pages are still initialized via init_unavailable_range()
1304 * and pfn walkers shouldn't touch any physical memory range for which they do
1305 * not recognize any specific metadata in struct pages.
1307 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1308 unsigned long end_pfn, struct zone *zone)
1310 struct page *start_page;
1311 struct page *end_page;
1313 /* end_pfn is one past the range we are checking */
1316 if (!pfn_valid(end_pfn))
1319 start_page = pfn_to_online_page(start_pfn);
1323 if (page_zone(start_page) != zone)
1326 end_page = pfn_to_page(end_pfn);
1328 /* This gives a shorter code than deriving page_zone(end_page) */
1329 if (page_zone_id(start_page) != page_zone_id(end_page))
1336 * The order of subdivision here is critical for the IO subsystem.
1337 * Please do not alter this order without good reasons and regression
1338 * testing. Specifically, as large blocks of memory are subdivided,
1339 * the order in which smaller blocks are delivered depends on the order
1340 * they're subdivided in this function. This is the primary factor
1341 * influencing the order in which pages are delivered to the IO
1342 * subsystem according to empirical testing, and this is also justified
1343 * by considering the behavior of a buddy system containing a single
1344 * large block of memory acted on by a series of small allocations.
1345 * This behavior is a critical factor in sglist merging's success.
1349 static inline void expand(struct zone *zone, struct page *page,
1350 int low, int high, int migratetype)
1352 unsigned long size = 1 << high;
1353 unsigned long nr_added = 0;
1355 while (high > low) {
1358 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1361 * Mark as guard pages (or page), that will allow to
1362 * merge back to allocator when buddy will be freed.
1363 * Corresponding page table entries will not be touched,
1364 * pages will stay not present in virtual address space
1366 if (set_page_guard(zone, &page[size], high))
1369 __add_to_free_list(&page[size], zone, high, migratetype, false);
1370 set_buddy_order(&page[size], high);
1373 account_freepages(zone, nr_added, migratetype);
1376 static void check_new_page_bad(struct page *page)
1378 if (unlikely(page->flags & __PG_HWPOISON)) {
1379 /* Don't complain about hwpoisoned pages */
1380 if (PageBuddy(page))
1381 __ClearPageBuddy(page);
1386 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1390 * This page is about to be returned from the page allocator
1392 static bool check_new_page(struct page *page)
1394 if (likely(page_expected_state(page,
1395 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1398 check_new_page_bad(page);
1402 static inline bool check_new_pages(struct page *page, unsigned int order)
1404 if (is_check_pages_enabled()) {
1405 for (int i = 0; i < (1 << order); i++) {
1406 struct page *p = page + i;
1408 if (check_new_page(p))
1416 static inline bool should_skip_kasan_unpoison(gfp_t flags)
1418 /* Don't skip if a software KASAN mode is enabled. */
1419 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1420 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1423 /* Skip, if hardware tag-based KASAN is not enabled. */
1424 if (!kasan_hw_tags_enabled())
1428 * With hardware tag-based KASAN enabled, skip if this has been
1429 * requested via __GFP_SKIP_KASAN.
1431 return flags & __GFP_SKIP_KASAN;
1434 static inline bool should_skip_init(gfp_t flags)
1436 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1437 if (!kasan_hw_tags_enabled())
1440 /* For hardware tag-based KASAN, skip if requested. */
1441 return (flags & __GFP_SKIP_ZERO);
1444 inline void post_alloc_hook(struct page *page, unsigned int order,
1447 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1448 !should_skip_init(gfp_flags);
1449 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1452 set_page_private(page, 0);
1453 set_page_refcounted(page);
1455 arch_alloc_page(page, order);
1456 debug_pagealloc_map_pages(page, 1 << order);
1459 * Page unpoisoning must happen before memory initialization.
1460 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1461 * allocations and the page unpoisoning code will complain.
1463 kernel_unpoison_pages(page, 1 << order);
1466 * As memory initialization might be integrated into KASAN,
1467 * KASAN unpoisoning and memory initializion code must be
1468 * kept together to avoid discrepancies in behavior.
1472 * If memory tags should be zeroed
1473 * (which happens only when memory should be initialized as well).
1476 /* Initialize both memory and memory tags. */
1477 for (i = 0; i != 1 << order; ++i)
1478 tag_clear_highpage(page + i);
1480 /* Take note that memory was initialized by the loop above. */
1483 if (!should_skip_kasan_unpoison(gfp_flags) &&
1484 kasan_unpoison_pages(page, order, init)) {
1485 /* Take note that memory was initialized by KASAN. */
1486 if (kasan_has_integrated_init())
1490 * If memory tags have not been set by KASAN, reset the page
1491 * tags to ensure page_address() dereferencing does not fault.
1493 for (i = 0; i != 1 << order; ++i)
1494 page_kasan_tag_reset(page + i);
1496 /* If memory is still not initialized, initialize it now. */
1498 kernel_init_pages(page, 1 << order);
1500 set_page_owner(page, order, gfp_flags);
1501 page_table_check_alloc(page, order);
1502 pgalloc_tag_add(page, current, 1 << order);
1505 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1506 unsigned int alloc_flags)
1508 post_alloc_hook(page, order, gfp_flags);
1510 if (order && (gfp_flags & __GFP_COMP))
1511 prep_compound_page(page, order);
1514 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1515 * allocate the page. The expectation is that the caller is taking
1516 * steps that will free more memory. The caller should avoid the page
1517 * being used for !PFMEMALLOC purposes.
1519 if (alloc_flags & ALLOC_NO_WATERMARKS)
1520 set_page_pfmemalloc(page);
1522 clear_page_pfmemalloc(page);
1526 * Go through the free lists for the given migratetype and remove
1527 * the smallest available page from the freelists
1529 static __always_inline
1530 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1533 unsigned int current_order;
1534 struct free_area *area;
1537 /* Find a page of the appropriate size in the preferred list */
1538 for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
1539 area = &(zone->free_area[current_order]);
1540 page = get_page_from_free_area(area, migratetype);
1543 del_page_from_free_list(page, zone, current_order, migratetype);
1544 expand(zone, page, order, current_order, migratetype);
1545 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1546 pcp_allowed_order(order) &&
1547 migratetype < MIGRATE_PCPTYPES);
1556 * This array describes the order lists are fallen back to when
1557 * the free lists for the desirable migrate type are depleted
1559 * The other migratetypes do not have fallbacks.
1561 static int fallbacks[MIGRATE_PCPTYPES][MIGRATE_PCPTYPES - 1] = {
1562 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1563 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1564 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1568 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1571 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1574 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1575 unsigned int order) { return NULL; }
1579 * Change the type of a block and move all its free pages to that
1582 static int __move_freepages_block(struct zone *zone, unsigned long start_pfn,
1583 int old_mt, int new_mt)
1586 unsigned long pfn, end_pfn;
1588 int pages_moved = 0;
1590 VM_WARN_ON(start_pfn & (pageblock_nr_pages - 1));
1591 end_pfn = pageblock_end_pfn(start_pfn);
1593 for (pfn = start_pfn; pfn < end_pfn;) {
1594 page = pfn_to_page(pfn);
1595 if (!PageBuddy(page)) {
1600 /* Make sure we are not inadvertently changing nodes */
1601 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1602 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1604 order = buddy_order(page);
1606 move_to_free_list(page, zone, order, old_mt, new_mt);
1609 pages_moved += 1 << order;
1612 set_pageblock_migratetype(pfn_to_page(start_pfn), new_mt);
1617 static bool prep_move_freepages_block(struct zone *zone, struct page *page,
1618 unsigned long *start_pfn,
1619 int *num_free, int *num_movable)
1621 unsigned long pfn, start, end;
1623 pfn = page_to_pfn(page);
1624 start = pageblock_start_pfn(pfn);
1625 end = pageblock_end_pfn(pfn);
1628 * The caller only has the lock for @zone, don't touch ranges
1629 * that straddle into other zones. While we could move part of
1630 * the range that's inside the zone, this call is usually
1631 * accompanied by other operations such as migratetype updates
1632 * which also should be locked.
1634 if (!zone_spans_pfn(zone, start))
1636 if (!zone_spans_pfn(zone, end - 1))
1644 for (pfn = start; pfn < end;) {
1645 page = pfn_to_page(pfn);
1646 if (PageBuddy(page)) {
1647 int nr = 1 << buddy_order(page);
1654 * We assume that pages that could be isolated for
1655 * migration are movable. But we don't actually try
1656 * isolating, as that would be expensive.
1658 if (PageLRU(page) || __PageMovable(page))
1667 static int move_freepages_block(struct zone *zone, struct page *page,
1668 int old_mt, int new_mt)
1670 unsigned long start_pfn;
1672 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
1675 return __move_freepages_block(zone, start_pfn, old_mt, new_mt);
1678 #ifdef CONFIG_MEMORY_ISOLATION
1679 /* Look for a buddy that straddles start_pfn */
1680 static unsigned long find_large_buddy(unsigned long start_pfn)
1684 unsigned long pfn = start_pfn;
1686 while (!PageBuddy(page = pfn_to_page(pfn))) {
1688 if (++order > MAX_PAGE_ORDER)
1690 pfn &= ~0UL << order;
1694 * Found a preceding buddy, but does it straddle?
1696 if (pfn + (1 << buddy_order(page)) > start_pfn)
1703 /* Split a multi-block free page into its individual pageblocks */
1704 static void split_large_buddy(struct zone *zone, struct page *page,
1705 unsigned long pfn, int order)
1707 unsigned long end_pfn = pfn + (1 << order);
1709 VM_WARN_ON_ONCE(order <= pageblock_order);
1710 VM_WARN_ON_ONCE(pfn & (pageblock_nr_pages - 1));
1712 /* Caller removed page from freelist, buddy info cleared! */
1713 VM_WARN_ON_ONCE(PageBuddy(page));
1715 while (pfn != end_pfn) {
1716 int mt = get_pfnblock_migratetype(page, pfn);
1718 __free_one_page(page, pfn, zone, pageblock_order, mt, FPI_NONE);
1719 pfn += pageblock_nr_pages;
1720 page = pfn_to_page(pfn);
1725 * move_freepages_block_isolate - move free pages in block for page isolation
1727 * @page: the pageblock page
1728 * @migratetype: migratetype to set on the pageblock
1730 * This is similar to move_freepages_block(), but handles the special
1731 * case encountered in page isolation, where the block of interest
1732 * might be part of a larger buddy spanning multiple pageblocks.
1734 * Unlike the regular page allocator path, which moves pages while
1735 * stealing buddies off the freelist, page isolation is interested in
1736 * arbitrary pfn ranges that may have overlapping buddies on both ends.
1738 * This function handles that. Straddling buddies are split into
1739 * individual pageblocks. Only the block of interest is moved.
1741 * Returns %true if pages could be moved, %false otherwise.
1743 bool move_freepages_block_isolate(struct zone *zone, struct page *page,
1746 unsigned long start_pfn, pfn;
1748 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
1751 /* No splits needed if buddies can't span multiple blocks */
1752 if (pageblock_order == MAX_PAGE_ORDER)
1755 /* We're a tail block in a larger buddy */
1756 pfn = find_large_buddy(start_pfn);
1757 if (pfn != start_pfn) {
1758 struct page *buddy = pfn_to_page(pfn);
1759 int order = buddy_order(buddy);
1761 del_page_from_free_list(buddy, zone, order,
1762 get_pfnblock_migratetype(buddy, pfn));
1763 set_pageblock_migratetype(page, migratetype);
1764 split_large_buddy(zone, buddy, pfn, order);
1768 /* We're the starting block of a larger buddy */
1769 if (PageBuddy(page) && buddy_order(page) > pageblock_order) {
1770 int order = buddy_order(page);
1772 del_page_from_free_list(page, zone, order,
1773 get_pfnblock_migratetype(page, pfn));
1774 set_pageblock_migratetype(page, migratetype);
1775 split_large_buddy(zone, page, pfn, order);
1779 __move_freepages_block(zone, start_pfn,
1780 get_pfnblock_migratetype(page, start_pfn),
1784 #endif /* CONFIG_MEMORY_ISOLATION */
1786 static void change_pageblock_range(struct page *pageblock_page,
1787 int start_order, int migratetype)
1789 int nr_pageblocks = 1 << (start_order - pageblock_order);
1791 while (nr_pageblocks--) {
1792 set_pageblock_migratetype(pageblock_page, migratetype);
1793 pageblock_page += pageblock_nr_pages;
1798 * When we are falling back to another migratetype during allocation, try to
1799 * steal extra free pages from the same pageblocks to satisfy further
1800 * allocations, instead of polluting multiple pageblocks.
1802 * If we are stealing a relatively large buddy page, it is likely there will
1803 * be more free pages in the pageblock, so try to steal them all. For
1804 * reclaimable and unmovable allocations, we steal regardless of page size,
1805 * as fragmentation caused by those allocations polluting movable pageblocks
1806 * is worse than movable allocations stealing from unmovable and reclaimable
1809 static bool can_steal_fallback(unsigned int order, int start_mt)
1812 * Leaving this order check is intended, although there is
1813 * relaxed order check in next check. The reason is that
1814 * we can actually steal whole pageblock if this condition met,
1815 * but, below check doesn't guarantee it and that is just heuristic
1816 * so could be changed anytime.
1818 if (order >= pageblock_order)
1821 if (order >= pageblock_order / 2 ||
1822 start_mt == MIGRATE_RECLAIMABLE ||
1823 start_mt == MIGRATE_UNMOVABLE ||
1824 page_group_by_mobility_disabled)
1830 static inline bool boost_watermark(struct zone *zone)
1832 unsigned long max_boost;
1834 if (!watermark_boost_factor)
1837 * Don't bother in zones that are unlikely to produce results.
1838 * On small machines, including kdump capture kernels running
1839 * in a small area, boosting the watermark can cause an out of
1840 * memory situation immediately.
1842 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1845 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1846 watermark_boost_factor, 10000);
1849 * high watermark may be uninitialised if fragmentation occurs
1850 * very early in boot so do not boost. We do not fall
1851 * through and boost by pageblock_nr_pages as failing
1852 * allocations that early means that reclaim is not going
1853 * to help and it may even be impossible to reclaim the
1854 * boosted watermark resulting in a hang.
1859 max_boost = max(pageblock_nr_pages, max_boost);
1861 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1868 * This function implements actual steal behaviour. If order is large enough, we
1869 * can claim the whole pageblock for the requested migratetype. If not, we check
1870 * the pageblock for constituent pages; if at least half of the pages are free
1871 * or compatible, we can still claim the whole block, so pages freed in the
1872 * future will be put on the correct free list. Otherwise, we isolate exactly
1873 * the order we need from the fallback block and leave its migratetype alone.
1875 static struct page *
1876 steal_suitable_fallback(struct zone *zone, struct page *page,
1877 int current_order, int order, int start_type,
1878 unsigned int alloc_flags, bool whole_block)
1880 int free_pages, movable_pages, alike_pages;
1881 unsigned long start_pfn;
1884 block_type = get_pageblock_migratetype(page);
1887 * This can happen due to races and we want to prevent broken
1888 * highatomic accounting.
1890 if (is_migrate_highatomic(block_type))
1893 /* Take ownership for orders >= pageblock_order */
1894 if (current_order >= pageblock_order) {
1895 del_page_from_free_list(page, zone, current_order, block_type);
1896 change_pageblock_range(page, current_order, start_type);
1897 expand(zone, page, order, current_order, start_type);
1902 * Boost watermarks to increase reclaim pressure to reduce the
1903 * likelihood of future fallbacks. Wake kswapd now as the node
1904 * may be balanced overall and kswapd will not wake naturally.
1906 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1907 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1909 /* We are not allowed to try stealing from the whole block */
1913 /* moving whole block can fail due to zone boundary conditions */
1914 if (!prep_move_freepages_block(zone, page, &start_pfn, &free_pages,
1919 * Determine how many pages are compatible with our allocation.
1920 * For movable allocation, it's the number of movable pages which
1921 * we just obtained. For other types it's a bit more tricky.
1923 if (start_type == MIGRATE_MOVABLE) {
1924 alike_pages = movable_pages;
1927 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1928 * to MOVABLE pageblock, consider all non-movable pages as
1929 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1930 * vice versa, be conservative since we can't distinguish the
1931 * exact migratetype of non-movable pages.
1933 if (block_type == MIGRATE_MOVABLE)
1934 alike_pages = pageblock_nr_pages
1935 - (free_pages + movable_pages);
1940 * If a sufficient number of pages in the block are either free or of
1941 * compatible migratability as our allocation, claim the whole block.
1943 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1944 page_group_by_mobility_disabled) {
1945 __move_freepages_block(zone, start_pfn, block_type, start_type);
1946 return __rmqueue_smallest(zone, order, start_type);
1950 del_page_from_free_list(page, zone, current_order, block_type);
1951 expand(zone, page, order, current_order, block_type);
1956 * Check whether there is a suitable fallback freepage with requested order.
1957 * If only_stealable is true, this function returns fallback_mt only if
1958 * we can steal other freepages all together. This would help to reduce
1959 * fragmentation due to mixed migratetype pages in one pageblock.
1961 int find_suitable_fallback(struct free_area *area, unsigned int order,
1962 int migratetype, bool only_stealable, bool *can_steal)
1967 if (area->nr_free == 0)
1971 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
1972 fallback_mt = fallbacks[migratetype][i];
1973 if (free_area_empty(area, fallback_mt))
1976 if (can_steal_fallback(order, migratetype))
1979 if (!only_stealable)
1990 * Reserve the pageblock(s) surrounding an allocation request for
1991 * exclusive use of high-order atomic allocations if there are no
1992 * empty page blocks that contain a page with a suitable order
1994 static void reserve_highatomic_pageblock(struct page *page, int order,
1998 unsigned long max_managed, flags;
2001 * The number reserved as: minimum is 1 pageblock, maximum is
2002 * roughly 1% of a zone. But if 1% of a zone falls below a
2003 * pageblock size, then don't reserve any pageblocks.
2004 * Check is race-prone but harmless.
2006 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
2008 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
2009 if (zone->nr_reserved_highatomic >= max_managed)
2012 spin_lock_irqsave(&zone->lock, flags);
2014 /* Recheck the nr_reserved_highatomic limit under the lock */
2015 if (zone->nr_reserved_highatomic >= max_managed)
2019 mt = get_pageblock_migratetype(page);
2020 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2021 if (!migratetype_is_mergeable(mt))
2024 if (order < pageblock_order) {
2025 if (move_freepages_block(zone, page, mt, MIGRATE_HIGHATOMIC) == -1)
2027 zone->nr_reserved_highatomic += pageblock_nr_pages;
2029 change_pageblock_range(page, order, MIGRATE_HIGHATOMIC);
2030 zone->nr_reserved_highatomic += 1 << order;
2034 spin_unlock_irqrestore(&zone->lock, flags);
2038 * Used when an allocation is about to fail under memory pressure. This
2039 * potentially hurts the reliability of high-order allocations when under
2040 * intense memory pressure but failed atomic allocations should be easier
2041 * to recover from than an OOM.
2043 * If @force is true, try to unreserve pageblocks even though highatomic
2044 * pageblock is exhausted.
2046 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2049 struct zonelist *zonelist = ac->zonelist;
2050 unsigned long flags;
2057 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2060 * Preserve at least one pageblock unless memory pressure
2063 if (!force && zone->nr_reserved_highatomic <=
2067 spin_lock_irqsave(&zone->lock, flags);
2068 for (order = 0; order < NR_PAGE_ORDERS; order++) {
2069 struct free_area *area = &(zone->free_area[order]);
2072 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2076 mt = get_pageblock_migratetype(page);
2078 * In page freeing path, migratetype change is racy so
2079 * we can counter several free pages in a pageblock
2080 * in this loop although we changed the pageblock type
2081 * from highatomic to ac->migratetype. So we should
2082 * adjust the count once.
2084 if (is_migrate_highatomic(mt)) {
2087 * It should never happen but changes to
2088 * locking could inadvertently allow a per-cpu
2089 * drain to add pages to MIGRATE_HIGHATOMIC
2090 * while unreserving so be safe and watch for
2093 size = max(pageblock_nr_pages, 1UL << order);
2094 size = min(size, zone->nr_reserved_highatomic);
2095 zone->nr_reserved_highatomic -= size;
2099 * Convert to ac->migratetype and avoid the normal
2100 * pageblock stealing heuristics. Minimally, the caller
2101 * is doing the work and needs the pages. More
2102 * importantly, if the block was always converted to
2103 * MIGRATE_UNMOVABLE or another type then the number
2104 * of pageblocks that cannot be completely freed
2107 if (order < pageblock_order)
2108 ret = move_freepages_block(zone, page, mt,
2111 move_to_free_list(page, zone, order, mt,
2113 change_pageblock_range(page, order,
2118 * Reserving the block(s) already succeeded,
2119 * so this should not fail on zone boundaries.
2121 WARN_ON_ONCE(ret == -1);
2123 spin_unlock_irqrestore(&zone->lock, flags);
2127 spin_unlock_irqrestore(&zone->lock, flags);
2134 * Try finding a free buddy page on the fallback list and put it on the free
2135 * list of requested migratetype, possibly along with other pages from the same
2136 * block, depending on fragmentation avoidance heuristics. Returns true if
2137 * fallback was found so that __rmqueue_smallest() can grab it.
2139 * The use of signed ints for order and current_order is a deliberate
2140 * deviation from the rest of this file, to make the for loop
2141 * condition simpler.
2143 static __always_inline struct page *
2144 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2145 unsigned int alloc_flags)
2147 struct free_area *area;
2149 int min_order = order;
2155 * Do not steal pages from freelists belonging to other pageblocks
2156 * i.e. orders < pageblock_order. If there are no local zones free,
2157 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2159 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2160 min_order = pageblock_order;
2163 * Find the largest available free page in the other list. This roughly
2164 * approximates finding the pageblock with the most free pages, which
2165 * would be too costly to do exactly.
2167 for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2169 area = &(zone->free_area[current_order]);
2170 fallback_mt = find_suitable_fallback(area, current_order,
2171 start_migratetype, false, &can_steal);
2172 if (fallback_mt == -1)
2176 * We cannot steal all free pages from the pageblock and the
2177 * requested migratetype is movable. In that case it's better to
2178 * steal and split the smallest available page instead of the
2179 * largest available page, because even if the next movable
2180 * allocation falls back into a different pageblock than this
2181 * one, it won't cause permanent fragmentation.
2183 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2184 && current_order > order)
2193 for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
2194 area = &(zone->free_area[current_order]);
2195 fallback_mt = find_suitable_fallback(area, current_order,
2196 start_migratetype, false, &can_steal);
2197 if (fallback_mt != -1)
2202 * This should not happen - we already found a suitable fallback
2203 * when looking for the largest page.
2205 VM_BUG_ON(current_order > MAX_PAGE_ORDER);
2208 page = get_page_from_free_area(area, fallback_mt);
2210 /* take off list, maybe claim block, expand remainder */
2211 page = steal_suitable_fallback(zone, page, current_order, order,
2212 start_migratetype, alloc_flags, can_steal);
2214 trace_mm_page_alloc_extfrag(page, order, current_order,
2215 start_migratetype, fallback_mt);
2221 * Do the hard work of removing an element from the buddy allocator.
2222 * Call me with the zone->lock already held.
2224 static __always_inline struct page *
2225 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2226 unsigned int alloc_flags)
2230 if (IS_ENABLED(CONFIG_CMA)) {
2232 * Balance movable allocations between regular and CMA areas by
2233 * allocating from CMA when over half of the zone's free memory
2234 * is in the CMA area.
2236 if (alloc_flags & ALLOC_CMA &&
2237 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2238 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2239 page = __rmqueue_cma_fallback(zone, order);
2245 page = __rmqueue_smallest(zone, order, migratetype);
2246 if (unlikely(!page)) {
2247 if (alloc_flags & ALLOC_CMA)
2248 page = __rmqueue_cma_fallback(zone, order);
2251 page = __rmqueue_fallback(zone, order, migratetype,
2258 * Obtain a specified number of elements from the buddy allocator, all under
2259 * a single hold of the lock, for efficiency. Add them to the supplied list.
2260 * Returns the number of new pages which were placed at *list.
2262 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2263 unsigned long count, struct list_head *list,
2264 int migratetype, unsigned int alloc_flags)
2266 unsigned long flags;
2269 spin_lock_irqsave(&zone->lock, flags);
2270 for (i = 0; i < count; ++i) {
2271 struct page *page = __rmqueue(zone, order, migratetype,
2273 if (unlikely(page == NULL))
2277 * Split buddy pages returned by expand() are received here in
2278 * physical page order. The page is added to the tail of
2279 * caller's list. From the callers perspective, the linked list
2280 * is ordered by page number under some conditions. This is
2281 * useful for IO devices that can forward direction from the
2282 * head, thus also in the physical page order. This is useful
2283 * for IO devices that can merge IO requests if the physical
2284 * pages are ordered properly.
2286 list_add_tail(&page->pcp_list, list);
2288 spin_unlock_irqrestore(&zone->lock, flags);
2294 * Called from the vmstat counter updater to decay the PCP high.
2295 * Return whether there are addition works to do.
2297 int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2299 int high_min, to_drain, batch;
2302 high_min = READ_ONCE(pcp->high_min);
2303 batch = READ_ONCE(pcp->batch);
2305 * Decrease pcp->high periodically to try to free possible
2306 * idle PCP pages. And, avoid to free too many pages to
2307 * control latency. This caps pcp->high decrement too.
2309 if (pcp->high > high_min) {
2310 pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2311 pcp->high - (pcp->high >> 3), high_min);
2312 if (pcp->high > high_min)
2316 to_drain = pcp->count - pcp->high;
2318 spin_lock(&pcp->lock);
2319 free_pcppages_bulk(zone, to_drain, pcp, 0);
2320 spin_unlock(&pcp->lock);
2329 * Called from the vmstat counter updater to drain pagesets of this
2330 * currently executing processor on remote nodes after they have
2333 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2335 int to_drain, batch;
2337 batch = READ_ONCE(pcp->batch);
2338 to_drain = min(pcp->count, batch);
2340 spin_lock(&pcp->lock);
2341 free_pcppages_bulk(zone, to_drain, pcp, 0);
2342 spin_unlock(&pcp->lock);
2348 * Drain pcplists of the indicated processor and zone.
2350 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2352 struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2356 spin_lock(&pcp->lock);
2359 int to_drain = min(count,
2360 pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
2362 free_pcppages_bulk(zone, to_drain, pcp, 0);
2365 spin_unlock(&pcp->lock);
2370 * Drain pcplists of all zones on the indicated processor.
2372 static void drain_pages(unsigned int cpu)
2376 for_each_populated_zone(zone) {
2377 drain_pages_zone(cpu, zone);
2382 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2384 void drain_local_pages(struct zone *zone)
2386 int cpu = smp_processor_id();
2389 drain_pages_zone(cpu, zone);
2395 * The implementation of drain_all_pages(), exposing an extra parameter to
2396 * drain on all cpus.
2398 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2399 * not empty. The check for non-emptiness can however race with a free to
2400 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2401 * that need the guarantee that every CPU has drained can disable the
2402 * optimizing racy check.
2404 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2409 * Allocate in the BSS so we won't require allocation in
2410 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2412 static cpumask_t cpus_with_pcps;
2415 * Do not drain if one is already in progress unless it's specific to
2416 * a zone. Such callers are primarily CMA and memory hotplug and need
2417 * the drain to be complete when the call returns.
2419 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2422 mutex_lock(&pcpu_drain_mutex);
2426 * We don't care about racing with CPU hotplug event
2427 * as offline notification will cause the notified
2428 * cpu to drain that CPU pcps and on_each_cpu_mask
2429 * disables preemption as part of its processing
2431 for_each_online_cpu(cpu) {
2432 struct per_cpu_pages *pcp;
2434 bool has_pcps = false;
2436 if (force_all_cpus) {
2438 * The pcp.count check is racy, some callers need a
2439 * guarantee that no cpu is missed.
2443 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2447 for_each_populated_zone(z) {
2448 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2457 cpumask_set_cpu(cpu, &cpus_with_pcps);
2459 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2462 for_each_cpu(cpu, &cpus_with_pcps) {
2464 drain_pages_zone(cpu, zone);
2469 mutex_unlock(&pcpu_drain_mutex);
2473 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2475 * When zone parameter is non-NULL, spill just the single zone's pages.
2477 void drain_all_pages(struct zone *zone)
2479 __drain_all_pages(zone, false);
2482 static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2484 int min_nr_free, max_nr_free;
2486 /* Free as much as possible if batch freeing high-order pages. */
2487 if (unlikely(free_high))
2488 return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2490 /* Check for PCP disabled or boot pageset */
2491 if (unlikely(high < batch))
2494 /* Leave at least pcp->batch pages on the list */
2495 min_nr_free = batch;
2496 max_nr_free = high - batch;
2499 * Increase the batch number to the number of the consecutive
2500 * freed pages to reduce zone lock contention.
2502 batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2507 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2508 int batch, bool free_high)
2510 int high, high_min, high_max;
2512 high_min = READ_ONCE(pcp->high_min);
2513 high_max = READ_ONCE(pcp->high_max);
2514 high = pcp->high = clamp(pcp->high, high_min, high_max);
2516 if (unlikely(!high))
2519 if (unlikely(free_high)) {
2520 pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2526 * If reclaim is active, limit the number of pages that can be
2527 * stored on pcp lists
2529 if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2530 int free_count = max_t(int, pcp->free_count, batch);
2532 pcp->high = max(high - free_count, high_min);
2533 return min(batch << 2, pcp->high);
2536 if (high_min == high_max)
2539 if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2540 int free_count = max_t(int, pcp->free_count, batch);
2542 pcp->high = max(high - free_count, high_min);
2543 high = max(pcp->count, high_min);
2544 } else if (pcp->count >= high) {
2545 int need_high = pcp->free_count + batch;
2547 /* pcp->high should be large enough to hold batch freed pages */
2548 if (pcp->high < need_high)
2549 pcp->high = clamp(need_high, high_min, high_max);
2555 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2556 struct page *page, int migratetype,
2561 bool free_high = false;
2564 * On freeing, reduce the number of pages that are batch allocated.
2565 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2568 pcp->alloc_factor >>= 1;
2569 __count_vm_events(PGFREE, 1 << order);
2570 pindex = order_to_pindex(migratetype, order);
2571 list_add(&page->pcp_list, &pcp->lists[pindex]);
2572 pcp->count += 1 << order;
2574 batch = READ_ONCE(pcp->batch);
2576 * As high-order pages other than THP's stored on PCP can contribute
2577 * to fragmentation, limit the number stored when PCP is heavily
2578 * freeing without allocation. The remainder after bulk freeing
2579 * stops will be drained from vmstat refresh context.
2581 if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2582 free_high = (pcp->free_count >= batch &&
2583 (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2584 (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2585 pcp->count >= READ_ONCE(batch)));
2586 pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2587 } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2588 pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2590 if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2591 pcp->free_count += (1 << order);
2592 high = nr_pcp_high(pcp, zone, batch, free_high);
2593 if (pcp->count >= high) {
2594 free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high),
2596 if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2597 zone_watermark_ok(zone, 0, high_wmark_pages(zone),
2599 clear_bit(ZONE_BELOW_HIGH, &zone->flags);
2606 void free_unref_page(struct page *page, unsigned int order)
2608 unsigned long __maybe_unused UP_flags;
2609 struct per_cpu_pages *pcp;
2611 unsigned long pfn = page_to_pfn(page);
2614 if (!pcp_allowed_order(order)) {
2615 __free_pages_ok(page, order, FPI_NONE);
2619 if (!free_pages_prepare(page, order))
2623 * We only track unmovable, reclaimable and movable on pcp lists.
2624 * Place ISOLATE pages on the isolated list because they are being
2625 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2626 * get those areas back if necessary. Otherwise, we may have to free
2627 * excessively into the page allocator
2629 migratetype = get_pfnblock_migratetype(page, pfn);
2630 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2631 if (unlikely(is_migrate_isolate(migratetype))) {
2632 free_one_page(page_zone(page), page, pfn, order, FPI_NONE);
2635 migratetype = MIGRATE_MOVABLE;
2638 zone = page_zone(page);
2639 pcp_trylock_prepare(UP_flags);
2640 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2642 free_unref_page_commit(zone, pcp, page, migratetype, order);
2643 pcp_spin_unlock(pcp);
2645 free_one_page(zone, page, pfn, order, FPI_NONE);
2647 pcp_trylock_finish(UP_flags);
2651 * Free a batch of folios
2653 void free_unref_folios(struct folio_batch *folios)
2655 unsigned long __maybe_unused UP_flags;
2656 struct per_cpu_pages *pcp = NULL;
2657 struct zone *locked_zone = NULL;
2660 /* Prepare folios for freeing */
2661 for (i = 0, j = 0; i < folios->nr; i++) {
2662 struct folio *folio = folios->folios[i];
2663 unsigned long pfn = folio_pfn(folio);
2664 unsigned int order = folio_order(folio);
2666 folio_undo_large_rmappable(folio);
2667 if (!free_pages_prepare(&folio->page, order))
2670 * Free orders not handled on the PCP directly to the
2673 if (!pcp_allowed_order(order)) {
2674 free_one_page(folio_zone(folio), &folio->page,
2675 pfn, order, FPI_NONE);
2678 folio->private = (void *)(unsigned long)order;
2680 folios->folios[j] = folio;
2685 for (i = 0; i < folios->nr; i++) {
2686 struct folio *folio = folios->folios[i];
2687 struct zone *zone = folio_zone(folio);
2688 unsigned long pfn = folio_pfn(folio);
2689 unsigned int order = (unsigned long)folio->private;
2692 folio->private = NULL;
2693 migratetype = get_pfnblock_migratetype(&folio->page, pfn);
2695 /* Different zone requires a different pcp lock */
2696 if (zone != locked_zone ||
2697 is_migrate_isolate(migratetype)) {
2699 pcp_spin_unlock(pcp);
2700 pcp_trylock_finish(UP_flags);
2706 * Free isolated pages directly to the
2707 * allocator, see comment in free_unref_page.
2709 if (is_migrate_isolate(migratetype)) {
2710 free_one_page(zone, &folio->page, pfn,
2716 * trylock is necessary as folios may be getting freed
2717 * from IRQ or SoftIRQ context after an IO completion.
2719 pcp_trylock_prepare(UP_flags);
2720 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2721 if (unlikely(!pcp)) {
2722 pcp_trylock_finish(UP_flags);
2723 free_one_page(zone, &folio->page, pfn,
2731 * Non-isolated types over MIGRATE_PCPTYPES get added
2732 * to the MIGRATE_MOVABLE pcp list.
2734 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2735 migratetype = MIGRATE_MOVABLE;
2737 trace_mm_page_free_batched(&folio->page);
2738 free_unref_page_commit(zone, pcp, &folio->page, migratetype,
2743 pcp_spin_unlock(pcp);
2744 pcp_trylock_finish(UP_flags);
2746 folio_batch_reinit(folios);
2750 * split_page takes a non-compound higher-order page, and splits it into
2751 * n (1<<order) sub-pages: page[0..n]
2752 * Each sub-page must be freed individually.
2754 * Note: this is probably too low level an operation for use in drivers.
2755 * Please consult with lkml before using this in your driver.
2757 void split_page(struct page *page, unsigned int order)
2761 VM_BUG_ON_PAGE(PageCompound(page), page);
2762 VM_BUG_ON_PAGE(!page_count(page), page);
2764 for (i = 1; i < (1 << order); i++)
2765 set_page_refcounted(page + i);
2766 split_page_owner(page, order, 0);
2767 pgalloc_tag_split(page, 1 << order);
2768 split_page_memcg(page, order, 0);
2770 EXPORT_SYMBOL_GPL(split_page);
2772 int __isolate_free_page(struct page *page, unsigned int order)
2774 struct zone *zone = page_zone(page);
2775 int mt = get_pageblock_migratetype(page);
2777 if (!is_migrate_isolate(mt)) {
2778 unsigned long watermark;
2780 * Obey watermarks as if the page was being allocated. We can
2781 * emulate a high-order watermark check with a raised order-0
2782 * watermark, because we already know our high-order page
2785 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2786 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2790 del_page_from_free_list(page, zone, order, mt);
2793 * Set the pageblock if the isolated page is at least half of a
2796 if (order >= pageblock_order - 1) {
2797 struct page *endpage = page + (1 << order) - 1;
2798 for (; page < endpage; page += pageblock_nr_pages) {
2799 int mt = get_pageblock_migratetype(page);
2801 * Only change normal pageblocks (i.e., they can merge
2804 if (migratetype_is_mergeable(mt))
2805 move_freepages_block(zone, page, mt,
2810 return 1UL << order;
2814 * __putback_isolated_page - Return a now-isolated page back where we got it
2815 * @page: Page that was isolated
2816 * @order: Order of the isolated page
2817 * @mt: The page's pageblock's migratetype
2819 * This function is meant to return a page pulled from the free lists via
2820 * __isolate_free_page back to the free lists they were pulled from.
2822 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2824 struct zone *zone = page_zone(page);
2826 /* zone lock should be held when this function is called */
2827 lockdep_assert_held(&zone->lock);
2829 /* Return isolated page to tail of freelist. */
2830 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2831 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2835 * Update NUMA hit/miss statistics
2837 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2841 enum numa_stat_item local_stat = NUMA_LOCAL;
2843 /* skip numa counters update if numa stats is disabled */
2844 if (!static_branch_likely(&vm_numa_stat_key))
2847 if (zone_to_nid(z) != numa_node_id())
2848 local_stat = NUMA_OTHER;
2850 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2851 __count_numa_events(z, NUMA_HIT, nr_account);
2853 __count_numa_events(z, NUMA_MISS, nr_account);
2854 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2856 __count_numa_events(z, local_stat, nr_account);
2860 static __always_inline
2861 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2862 unsigned int order, unsigned int alloc_flags,
2866 unsigned long flags;
2870 spin_lock_irqsave(&zone->lock, flags);
2871 if (alloc_flags & ALLOC_HIGHATOMIC)
2872 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2874 page = __rmqueue(zone, order, migratetype, alloc_flags);
2877 * If the allocation fails, allow OOM handling access
2878 * to HIGHATOMIC reserves as failing now is worse than
2879 * failing a high-order atomic allocation in the
2882 if (!page && (alloc_flags & ALLOC_OOM))
2883 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2886 spin_unlock_irqrestore(&zone->lock, flags);
2890 spin_unlock_irqrestore(&zone->lock, flags);
2891 } while (check_new_pages(page, order));
2893 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2894 zone_statistics(preferred_zone, zone, 1);
2899 static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
2901 int high, base_batch, batch, max_nr_alloc;
2902 int high_max, high_min;
2904 base_batch = READ_ONCE(pcp->batch);
2905 high_min = READ_ONCE(pcp->high_min);
2906 high_max = READ_ONCE(pcp->high_max);
2907 high = pcp->high = clamp(pcp->high, high_min, high_max);
2909 /* Check for PCP disabled or boot pageset */
2910 if (unlikely(high < base_batch))
2916 batch = (base_batch << pcp->alloc_factor);
2919 * If we had larger pcp->high, we could avoid to allocate from
2922 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
2923 high = pcp->high = min(high + batch, high_max);
2926 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
2928 * Double the number of pages allocated each time there is
2929 * subsequent allocation of order-0 pages without any freeing.
2931 if (batch <= max_nr_alloc &&
2932 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
2933 pcp->alloc_factor++;
2934 batch = min(batch, max_nr_alloc);
2938 * Scale batch relative to order if batch implies free pages
2939 * can be stored on the PCP. Batch can be 1 for small zones or
2940 * for boot pagesets which should never store free pages as
2941 * the pages may belong to arbitrary zones.
2944 batch = max(batch >> order, 2);
2949 /* Remove page from the per-cpu list, caller must protect the list */
2951 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2953 unsigned int alloc_flags,
2954 struct per_cpu_pages *pcp,
2955 struct list_head *list)
2960 if (list_empty(list)) {
2961 int batch = nr_pcp_alloc(pcp, zone, order);
2964 alloced = rmqueue_bulk(zone, order,
2966 migratetype, alloc_flags);
2968 pcp->count += alloced << order;
2969 if (unlikely(list_empty(list)))
2973 page = list_first_entry(list, struct page, pcp_list);
2974 list_del(&page->pcp_list);
2975 pcp->count -= 1 << order;
2976 } while (check_new_pages(page, order));
2981 /* Lock and remove page from the per-cpu list */
2982 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2983 struct zone *zone, unsigned int order,
2984 int migratetype, unsigned int alloc_flags)
2986 struct per_cpu_pages *pcp;
2987 struct list_head *list;
2989 unsigned long __maybe_unused UP_flags;
2991 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2992 pcp_trylock_prepare(UP_flags);
2993 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2995 pcp_trylock_finish(UP_flags);
3000 * On allocation, reduce the number of pages that are batch freed.
3001 * See nr_pcp_free() where free_factor is increased for subsequent
3004 pcp->free_count >>= 1;
3005 list = &pcp->lists[order_to_pindex(migratetype, order)];
3006 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3007 pcp_spin_unlock(pcp);
3008 pcp_trylock_finish(UP_flags);
3010 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3011 zone_statistics(preferred_zone, zone, 1);
3017 * Allocate a page from the given zone.
3018 * Use pcplists for THP or "cheap" high-order allocations.
3022 * Do not instrument rmqueue() with KMSAN. This function may call
3023 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
3024 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3025 * may call rmqueue() again, which will result in a deadlock.
3027 __no_sanitize_memory
3029 struct page *rmqueue(struct zone *preferred_zone,
3030 struct zone *zone, unsigned int order,
3031 gfp_t gfp_flags, unsigned int alloc_flags,
3037 * We most definitely don't want callers attempting to
3038 * allocate greater than order-1 page units with __GFP_NOFAIL.
3040 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3042 if (likely(pcp_allowed_order(order))) {
3043 page = rmqueue_pcplist(preferred_zone, zone, order,
3044 migratetype, alloc_flags);
3049 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3053 /* Separate test+clear to avoid unnecessary atomics */
3054 if ((alloc_flags & ALLOC_KSWAPD) &&
3055 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3056 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3057 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3060 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3064 static inline long __zone_watermark_unusable_free(struct zone *z,
3065 unsigned int order, unsigned int alloc_flags)
3067 long unusable_free = (1 << order) - 1;
3070 * If the caller does not have rights to reserves below the min
3071 * watermark then subtract the high-atomic reserves. This will
3072 * over-estimate the size of the atomic reserve but it avoids a search.
3074 if (likely(!(alloc_flags & ALLOC_RESERVES)))
3075 unusable_free += z->nr_reserved_highatomic;
3078 /* If allocation can't use CMA areas don't use free CMA pages */
3079 if (!(alloc_flags & ALLOC_CMA))
3080 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3083 return unusable_free;
3087 * Return true if free base pages are above 'mark'. For high-order checks it
3088 * will return true of the order-0 watermark is reached and there is at least
3089 * one free page of a suitable size. Checking now avoids taking the zone lock
3090 * to check in the allocation paths if no pages are free.
3092 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3093 int highest_zoneidx, unsigned int alloc_flags,
3099 /* free_pages may go negative - that's OK */
3100 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3102 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
3104 * __GFP_HIGH allows access to 50% of the min reserve as well
3107 if (alloc_flags & ALLOC_MIN_RESERVE) {
3111 * Non-blocking allocations (e.g. GFP_ATOMIC) can
3112 * access more reserves than just __GFP_HIGH. Other
3113 * non-blocking allocations requests such as GFP_NOWAIT
3114 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
3115 * access to the min reserve.
3117 if (alloc_flags & ALLOC_NON_BLOCK)
3122 * OOM victims can try even harder than the normal reserve
3123 * users on the grounds that it's definitely going to be in
3124 * the exit path shortly and free memory. Any allocation it
3125 * makes during the free path will be small and short-lived.
3127 if (alloc_flags & ALLOC_OOM)
3132 * Check watermarks for an order-0 allocation request. If these
3133 * are not met, then a high-order request also cannot go ahead
3134 * even if a suitable page happened to be free.
3136 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3139 /* If this is an order-0 request then the watermark is fine */
3143 /* For a high-order request, check at least one suitable page is free */
3144 for (o = order; o < NR_PAGE_ORDERS; o++) {
3145 struct free_area *area = &z->free_area[o];
3151 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3152 if (!free_area_empty(area, mt))
3157 if ((alloc_flags & ALLOC_CMA) &&
3158 !free_area_empty(area, MIGRATE_CMA)) {
3162 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3163 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3170 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3171 int highest_zoneidx, unsigned int alloc_flags)
3173 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3174 zone_page_state(z, NR_FREE_PAGES));
3177 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3178 unsigned long mark, int highest_zoneidx,
3179 unsigned int alloc_flags, gfp_t gfp_mask)
3183 free_pages = zone_page_state(z, NR_FREE_PAGES);
3186 * Fast check for order-0 only. If this fails then the reserves
3187 * need to be calculated.
3193 usable_free = free_pages;
3194 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3196 /* reserved may over estimate high-atomic reserves. */
3197 usable_free -= min(usable_free, reserved);
3198 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3202 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3207 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3208 * when checking the min watermark. The min watermark is the
3209 * point where boosting is ignored so that kswapd is woken up
3210 * when below the low watermark.
3212 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3213 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3214 mark = z->_watermark[WMARK_MIN];
3215 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3216 alloc_flags, free_pages);
3222 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3223 unsigned long mark, int highest_zoneidx)
3225 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3227 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3228 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3230 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3235 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3237 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3239 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3240 node_reclaim_distance;
3242 #else /* CONFIG_NUMA */
3243 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3247 #endif /* CONFIG_NUMA */
3250 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3251 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3252 * premature use of a lower zone may cause lowmem pressure problems that
3253 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3254 * probably too small. It only makes sense to spread allocations to avoid
3255 * fragmentation between the Normal and DMA32 zones.
3257 static inline unsigned int
3258 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3260 unsigned int alloc_flags;
3263 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3266 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3268 #ifdef CONFIG_ZONE_DMA32
3272 if (zone_idx(zone) != ZONE_NORMAL)
3276 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3277 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3278 * on UMA that if Normal is populated then so is DMA32.
3280 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3281 if (nr_online_nodes > 1 && !populated_zone(--zone))
3284 alloc_flags |= ALLOC_NOFRAGMENT;
3285 #endif /* CONFIG_ZONE_DMA32 */
3289 /* Must be called after current_gfp_context() which can change gfp_mask */
3290 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3291 unsigned int alloc_flags)
3294 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3295 alloc_flags |= ALLOC_CMA;
3301 * get_page_from_freelist goes through the zonelist trying to allocate
3304 static struct page *
3305 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3306 const struct alloc_context *ac)
3310 struct pglist_data *last_pgdat = NULL;
3311 bool last_pgdat_dirty_ok = false;
3316 * Scan zonelist, looking for a zone with enough free.
3317 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3319 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3320 z = ac->preferred_zoneref;
3321 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3326 if (cpusets_enabled() &&
3327 (alloc_flags & ALLOC_CPUSET) &&
3328 !__cpuset_zone_allowed(zone, gfp_mask))
3331 * When allocating a page cache page for writing, we
3332 * want to get it from a node that is within its dirty
3333 * limit, such that no single node holds more than its
3334 * proportional share of globally allowed dirty pages.
3335 * The dirty limits take into account the node's
3336 * lowmem reserves and high watermark so that kswapd
3337 * should be able to balance it without having to
3338 * write pages from its LRU list.
3340 * XXX: For now, allow allocations to potentially
3341 * exceed the per-node dirty limit in the slowpath
3342 * (spread_dirty_pages unset) before going into reclaim,
3343 * which is important when on a NUMA setup the allowed
3344 * nodes are together not big enough to reach the
3345 * global limit. The proper fix for these situations
3346 * will require awareness of nodes in the
3347 * dirty-throttling and the flusher threads.
3349 if (ac->spread_dirty_pages) {
3350 if (last_pgdat != zone->zone_pgdat) {
3351 last_pgdat = zone->zone_pgdat;
3352 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3355 if (!last_pgdat_dirty_ok)
3359 if (no_fallback && nr_online_nodes > 1 &&
3360 zone != ac->preferred_zoneref->zone) {
3364 * If moving to a remote node, retry but allow
3365 * fragmenting fallbacks. Locality is more important
3366 * than fragmentation avoidance.
3368 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3369 if (zone_to_nid(zone) != local_nid) {
3370 alloc_flags &= ~ALLOC_NOFRAGMENT;
3375 cond_accept_memory(zone, order);
3378 * Detect whether the number of free pages is below high
3379 * watermark. If so, we will decrease pcp->high and free
3380 * PCP pages in free path to reduce the possibility of
3381 * premature page reclaiming. Detection is done here to
3382 * avoid to do that in hotter free path.
3384 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3385 goto check_alloc_wmark;
3387 mark = high_wmark_pages(zone);
3388 if (zone_watermark_fast(zone, order, mark,
3389 ac->highest_zoneidx, alloc_flags,
3393 set_bit(ZONE_BELOW_HIGH, &zone->flags);
3396 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3397 if (!zone_watermark_fast(zone, order, mark,
3398 ac->highest_zoneidx, alloc_flags,
3402 if (cond_accept_memory(zone, order))
3405 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3407 * Watermark failed for this zone, but see if we can
3408 * grow this zone if it contains deferred pages.
3410 if (deferred_pages_enabled()) {
3411 if (_deferred_grow_zone(zone, order))
3415 /* Checked here to keep the fast path fast */
3416 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3417 if (alloc_flags & ALLOC_NO_WATERMARKS)
3420 if (!node_reclaim_enabled() ||
3421 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3424 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3426 case NODE_RECLAIM_NOSCAN:
3429 case NODE_RECLAIM_FULL:
3430 /* scanned but unreclaimable */
3433 /* did we reclaim enough */
3434 if (zone_watermark_ok(zone, order, mark,
3435 ac->highest_zoneidx, alloc_flags))
3443 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3444 gfp_mask, alloc_flags, ac->migratetype);
3446 prep_new_page(page, order, gfp_mask, alloc_flags);
3449 * If this is a high-order atomic allocation then check
3450 * if the pageblock should be reserved for the future
3452 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3453 reserve_highatomic_pageblock(page, order, zone);
3457 if (cond_accept_memory(zone, order))
3460 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3461 /* Try again if zone has deferred pages */
3462 if (deferred_pages_enabled()) {
3463 if (_deferred_grow_zone(zone, order))
3471 * It's possible on a UMA machine to get through all zones that are
3472 * fragmented. If avoiding fragmentation, reset and try again.
3475 alloc_flags &= ~ALLOC_NOFRAGMENT;
3482 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3484 unsigned int filter = SHOW_MEM_FILTER_NODES;
3487 * This documents exceptions given to allocations in certain
3488 * contexts that are allowed to allocate outside current's set
3491 if (!(gfp_mask & __GFP_NOMEMALLOC))
3492 if (tsk_is_oom_victim(current) ||
3493 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3494 filter &= ~SHOW_MEM_FILTER_NODES;
3495 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3496 filter &= ~SHOW_MEM_FILTER_NODES;
3498 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3501 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3503 struct va_format vaf;
3505 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3507 if ((gfp_mask & __GFP_NOWARN) ||
3508 !__ratelimit(&nopage_rs) ||
3509 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3512 va_start(args, fmt);
3515 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3516 current->comm, &vaf, gfp_mask, &gfp_mask,
3517 nodemask_pr_args(nodemask));
3520 cpuset_print_current_mems_allowed();
3523 warn_alloc_show_mem(gfp_mask, nodemask);
3526 static inline struct page *
3527 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3528 unsigned int alloc_flags,
3529 const struct alloc_context *ac)
3533 page = get_page_from_freelist(gfp_mask, order,
3534 alloc_flags|ALLOC_CPUSET, ac);
3536 * fallback to ignore cpuset restriction if our nodes
3540 page = get_page_from_freelist(gfp_mask, order,
3546 static inline struct page *
3547 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3548 const struct alloc_context *ac, unsigned long *did_some_progress)
3550 struct oom_control oc = {
3551 .zonelist = ac->zonelist,
3552 .nodemask = ac->nodemask,
3554 .gfp_mask = gfp_mask,
3559 *did_some_progress = 0;
3562 * Acquire the oom lock. If that fails, somebody else is
3563 * making progress for us.
3565 if (!mutex_trylock(&oom_lock)) {
3566 *did_some_progress = 1;
3567 schedule_timeout_uninterruptible(1);
3572 * Go through the zonelist yet one more time, keep very high watermark
3573 * here, this is only to catch a parallel oom killing, we must fail if
3574 * we're still under heavy pressure. But make sure that this reclaim
3575 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3576 * allocation which will never fail due to oom_lock already held.
3578 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3579 ~__GFP_DIRECT_RECLAIM, order,
3580 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3584 /* Coredumps can quickly deplete all memory reserves */
3585 if (current->flags & PF_DUMPCORE)
3587 /* The OOM killer will not help higher order allocs */
3588 if (order > PAGE_ALLOC_COSTLY_ORDER)
3591 * We have already exhausted all our reclaim opportunities without any
3592 * success so it is time to admit defeat. We will skip the OOM killer
3593 * because it is very likely that the caller has a more reasonable
3594 * fallback than shooting a random task.
3596 * The OOM killer may not free memory on a specific node.
3598 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3600 /* The OOM killer does not needlessly kill tasks for lowmem */
3601 if (ac->highest_zoneidx < ZONE_NORMAL)
3603 if (pm_suspended_storage())
3606 * XXX: GFP_NOFS allocations should rather fail than rely on
3607 * other request to make a forward progress.
3608 * We are in an unfortunate situation where out_of_memory cannot
3609 * do much for this context but let's try it to at least get
3610 * access to memory reserved if the current task is killed (see
3611 * out_of_memory). Once filesystems are ready to handle allocation
3612 * failures more gracefully we should just bail out here.
3615 /* Exhausted what can be done so it's blame time */
3616 if (out_of_memory(&oc) ||
3617 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3618 *did_some_progress = 1;
3621 * Help non-failing allocations by giving them access to memory
3624 if (gfp_mask & __GFP_NOFAIL)
3625 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3626 ALLOC_NO_WATERMARKS, ac);
3629 mutex_unlock(&oom_lock);
3634 * Maximum number of compaction retries with a progress before OOM
3635 * killer is consider as the only way to move forward.
3637 #define MAX_COMPACT_RETRIES 16
3639 #ifdef CONFIG_COMPACTION
3640 /* Try memory compaction for high-order allocations before reclaim */
3641 static struct page *
3642 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3643 unsigned int alloc_flags, const struct alloc_context *ac,
3644 enum compact_priority prio, enum compact_result *compact_result)
3646 struct page *page = NULL;
3647 unsigned long pflags;
3648 unsigned int noreclaim_flag;
3653 psi_memstall_enter(&pflags);
3654 delayacct_compact_start();
3655 noreclaim_flag = memalloc_noreclaim_save();
3657 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3660 memalloc_noreclaim_restore(noreclaim_flag);
3661 psi_memstall_leave(&pflags);
3662 delayacct_compact_end();
3664 if (*compact_result == COMPACT_SKIPPED)
3667 * At least in one zone compaction wasn't deferred or skipped, so let's
3668 * count a compaction stall
3670 count_vm_event(COMPACTSTALL);
3672 /* Prep a captured page if available */
3674 prep_new_page(page, order, gfp_mask, alloc_flags);
3676 /* Try get a page from the freelist if available */
3678 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3681 struct zone *zone = page_zone(page);
3683 zone->compact_blockskip_flush = false;
3684 compaction_defer_reset(zone, order, true);
3685 count_vm_event(COMPACTSUCCESS);
3690 * It's bad if compaction run occurs and fails. The most likely reason
3691 * is that pages exist, but not enough to satisfy watermarks.
3693 count_vm_event(COMPACTFAIL);
3701 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3702 enum compact_result compact_result,
3703 enum compact_priority *compact_priority,
3704 int *compaction_retries)
3706 int max_retries = MAX_COMPACT_RETRIES;
3709 int retries = *compaction_retries;
3710 enum compact_priority priority = *compact_priority;
3715 if (fatal_signal_pending(current))
3719 * Compaction was skipped due to a lack of free order-0
3720 * migration targets. Continue if reclaim can help.
3722 if (compact_result == COMPACT_SKIPPED) {
3723 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3728 * Compaction managed to coalesce some page blocks, but the
3729 * allocation failed presumably due to a race. Retry some.
3731 if (compact_result == COMPACT_SUCCESS) {
3733 * !costly requests are much more important than
3734 * __GFP_RETRY_MAYFAIL costly ones because they are de
3735 * facto nofail and invoke OOM killer to move on while
3736 * costly can fail and users are ready to cope with
3737 * that. 1/4 retries is rather arbitrary but we would
3738 * need much more detailed feedback from compaction to
3739 * make a better decision.
3741 if (order > PAGE_ALLOC_COSTLY_ORDER)
3744 if (++(*compaction_retries) <= max_retries) {
3751 * Compaction failed. Retry with increasing priority.
3753 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3754 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3756 if (*compact_priority > min_priority) {
3757 (*compact_priority)--;
3758 *compaction_retries = 0;
3762 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3766 static inline struct page *
3767 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3768 unsigned int alloc_flags, const struct alloc_context *ac,
3769 enum compact_priority prio, enum compact_result *compact_result)
3771 *compact_result = COMPACT_SKIPPED;
3776 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3777 enum compact_result compact_result,
3778 enum compact_priority *compact_priority,
3779 int *compaction_retries)
3784 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3788 * There are setups with compaction disabled which would prefer to loop
3789 * inside the allocator rather than hit the oom killer prematurely.
3790 * Let's give them a good hope and keep retrying while the order-0
3791 * watermarks are OK.
3793 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3794 ac->highest_zoneidx, ac->nodemask) {
3795 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3796 ac->highest_zoneidx, alloc_flags))
3801 #endif /* CONFIG_COMPACTION */
3803 #ifdef CONFIG_LOCKDEP
3804 static struct lockdep_map __fs_reclaim_map =
3805 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3807 static bool __need_reclaim(gfp_t gfp_mask)
3809 /* no reclaim without waiting on it */
3810 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3813 /* this guy won't enter reclaim */
3814 if (current->flags & PF_MEMALLOC)
3817 if (gfp_mask & __GFP_NOLOCKDEP)
3823 void __fs_reclaim_acquire(unsigned long ip)
3825 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3828 void __fs_reclaim_release(unsigned long ip)
3830 lock_release(&__fs_reclaim_map, ip);
3833 void fs_reclaim_acquire(gfp_t gfp_mask)
3835 gfp_mask = current_gfp_context(gfp_mask);
3837 if (__need_reclaim(gfp_mask)) {
3838 if (gfp_mask & __GFP_FS)
3839 __fs_reclaim_acquire(_RET_IP_);
3841 #ifdef CONFIG_MMU_NOTIFIER
3842 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3843 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3848 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3850 void fs_reclaim_release(gfp_t gfp_mask)
3852 gfp_mask = current_gfp_context(gfp_mask);
3854 if (__need_reclaim(gfp_mask)) {
3855 if (gfp_mask & __GFP_FS)
3856 __fs_reclaim_release(_RET_IP_);
3859 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3863 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3864 * have been rebuilt so allocation retries. Reader side does not lock and
3865 * retries the allocation if zonelist changes. Writer side is protected by the
3866 * embedded spin_lock.
3868 static DEFINE_SEQLOCK(zonelist_update_seq);
3870 static unsigned int zonelist_iter_begin(void)
3872 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3873 return read_seqbegin(&zonelist_update_seq);
3878 static unsigned int check_retry_zonelist(unsigned int seq)
3880 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3881 return read_seqretry(&zonelist_update_seq, seq);
3886 /* Perform direct synchronous page reclaim */
3887 static unsigned long
3888 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3889 const struct alloc_context *ac)
3891 unsigned int noreclaim_flag;
3892 unsigned long progress;
3896 /* We now go into synchronous reclaim */
3897 cpuset_memory_pressure_bump();
3898 fs_reclaim_acquire(gfp_mask);
3899 noreclaim_flag = memalloc_noreclaim_save();
3901 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3904 memalloc_noreclaim_restore(noreclaim_flag);
3905 fs_reclaim_release(gfp_mask);
3912 /* The really slow allocator path where we enter direct reclaim */
3913 static inline struct page *
3914 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3915 unsigned int alloc_flags, const struct alloc_context *ac,
3916 unsigned long *did_some_progress)
3918 struct page *page = NULL;
3919 unsigned long pflags;
3920 bool drained = false;
3922 psi_memstall_enter(&pflags);
3923 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3924 if (unlikely(!(*did_some_progress)))
3928 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3931 * If an allocation failed after direct reclaim, it could be because
3932 * pages are pinned on the per-cpu lists or in high alloc reserves.
3933 * Shrink them and try again
3935 if (!page && !drained) {
3936 unreserve_highatomic_pageblock(ac, false);
3937 drain_all_pages(NULL);
3942 psi_memstall_leave(&pflags);
3947 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3948 const struct alloc_context *ac)
3952 pg_data_t *last_pgdat = NULL;
3953 enum zone_type highest_zoneidx = ac->highest_zoneidx;
3955 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3957 if (!managed_zone(zone))
3959 if (last_pgdat != zone->zone_pgdat) {
3960 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3961 last_pgdat = zone->zone_pgdat;
3966 static inline unsigned int
3967 gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3969 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3972 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3973 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3974 * to save two branches.
3976 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
3977 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
3980 * The caller may dip into page reserves a bit more if the caller
3981 * cannot run direct reclaim, or if the caller has realtime scheduling
3982 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3983 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3985 alloc_flags |= (__force int)
3986 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
3988 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
3990 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3991 * if it can't schedule.
3993 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
3994 alloc_flags |= ALLOC_NON_BLOCK;
3997 alloc_flags |= ALLOC_HIGHATOMIC;
4001 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
4002 * GFP_ATOMIC) rather than fail, see the comment for
4003 * cpuset_node_allowed().
4005 if (alloc_flags & ALLOC_MIN_RESERVE)
4006 alloc_flags &= ~ALLOC_CPUSET;
4007 } else if (unlikely(rt_or_dl_task(current)) && in_task())
4008 alloc_flags |= ALLOC_MIN_RESERVE;
4010 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4015 static bool oom_reserves_allowed(struct task_struct *tsk)
4017 if (!tsk_is_oom_victim(tsk))
4021 * !MMU doesn't have oom reaper so give access to memory reserves
4022 * only to the thread with TIF_MEMDIE set
4024 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4031 * Distinguish requests which really need access to full memory
4032 * reserves from oom victims which can live with a portion of it
4034 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4036 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4038 if (gfp_mask & __GFP_MEMALLOC)
4039 return ALLOC_NO_WATERMARKS;
4040 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4041 return ALLOC_NO_WATERMARKS;
4042 if (!in_interrupt()) {
4043 if (current->flags & PF_MEMALLOC)
4044 return ALLOC_NO_WATERMARKS;
4045 else if (oom_reserves_allowed(current))
4052 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4054 return !!__gfp_pfmemalloc_flags(gfp_mask);
4058 * Checks whether it makes sense to retry the reclaim to make a forward progress
4059 * for the given allocation request.
4061 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4062 * without success, or when we couldn't even meet the watermark if we
4063 * reclaimed all remaining pages on the LRU lists.
4065 * Returns true if a retry is viable or false to enter the oom path.
4068 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4069 struct alloc_context *ac, int alloc_flags,
4070 bool did_some_progress, int *no_progress_loops)
4077 * Costly allocations might have made a progress but this doesn't mean
4078 * their order will become available due to high fragmentation so
4079 * always increment the no progress counter for them
4081 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4082 *no_progress_loops = 0;
4084 (*no_progress_loops)++;
4086 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
4091 * Keep reclaiming pages while there is a chance this will lead
4092 * somewhere. If none of the target zones can satisfy our allocation
4093 * request even if all reclaimable pages are considered then we are
4094 * screwed and have to go OOM.
4096 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4097 ac->highest_zoneidx, ac->nodemask) {
4098 unsigned long available;
4099 unsigned long reclaimable;
4100 unsigned long min_wmark = min_wmark_pages(zone);
4103 available = reclaimable = zone_reclaimable_pages(zone);
4104 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4107 * Would the allocation succeed if we reclaimed all
4108 * reclaimable pages?
4110 wmark = __zone_watermark_ok(zone, order, min_wmark,
4111 ac->highest_zoneidx, alloc_flags, available);
4112 trace_reclaim_retry_zone(z, order, reclaimable,
4113 available, min_wmark, *no_progress_loops, wmark);
4121 * Memory allocation/reclaim might be called from a WQ context and the
4122 * current implementation of the WQ concurrency control doesn't
4123 * recognize that a particular WQ is congested if the worker thread is
4124 * looping without ever sleeping. Therefore we have to do a short sleep
4125 * here rather than calling cond_resched().
4127 if (current->flags & PF_WQ_WORKER)
4128 schedule_timeout_uninterruptible(1);
4132 /* Before OOM, exhaust highatomic_reserve */
4134 return unreserve_highatomic_pageblock(ac, true);
4140 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4143 * It's possible that cpuset's mems_allowed and the nodemask from
4144 * mempolicy don't intersect. This should be normally dealt with by
4145 * policy_nodemask(), but it's possible to race with cpuset update in
4146 * such a way the check therein was true, and then it became false
4147 * before we got our cpuset_mems_cookie here.
4148 * This assumes that for all allocations, ac->nodemask can come only
4149 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4150 * when it does not intersect with the cpuset restrictions) or the
4151 * caller can deal with a violated nodemask.
4153 if (cpusets_enabled() && ac->nodemask &&
4154 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4155 ac->nodemask = NULL;
4160 * When updating a task's mems_allowed or mempolicy nodemask, it is
4161 * possible to race with parallel threads in such a way that our
4162 * allocation can fail while the mask is being updated. If we are about
4163 * to fail, check if the cpuset changed during allocation and if so,
4166 if (read_mems_allowed_retry(cpuset_mems_cookie))
4172 static inline struct page *
4173 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4174 struct alloc_context *ac)
4176 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4177 bool can_compact = gfp_compaction_allowed(gfp_mask);
4178 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4179 struct page *page = NULL;
4180 unsigned int alloc_flags;
4181 unsigned long did_some_progress;
4182 enum compact_priority compact_priority;
4183 enum compact_result compact_result;
4184 int compaction_retries;
4185 int no_progress_loops;
4186 unsigned int cpuset_mems_cookie;
4187 unsigned int zonelist_iter_cookie;
4191 compaction_retries = 0;
4192 no_progress_loops = 0;
4193 compact_priority = DEF_COMPACT_PRIORITY;
4194 cpuset_mems_cookie = read_mems_allowed_begin();
4195 zonelist_iter_cookie = zonelist_iter_begin();
4198 * The fast path uses conservative alloc_flags to succeed only until
4199 * kswapd needs to be woken up, and to avoid the cost of setting up
4200 * alloc_flags precisely. So we do that now.
4202 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4205 * We need to recalculate the starting point for the zonelist iterator
4206 * because we might have used different nodemask in the fast path, or
4207 * there was a cpuset modification and we are retrying - otherwise we
4208 * could end up iterating over non-eligible zones endlessly.
4210 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4211 ac->highest_zoneidx, ac->nodemask);
4212 if (!ac->preferred_zoneref->zone)
4216 * Check for insane configurations where the cpuset doesn't contain
4217 * any suitable zone to satisfy the request - e.g. non-movable
4218 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4220 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4221 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4222 ac->highest_zoneidx,
4223 &cpuset_current_mems_allowed);
4228 if (alloc_flags & ALLOC_KSWAPD)
4229 wake_all_kswapds(order, gfp_mask, ac);
4232 * The adjusted alloc_flags might result in immediate success, so try
4235 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4240 * For costly allocations, try direct compaction first, as it's likely
4241 * that we have enough base pages and don't need to reclaim. For non-
4242 * movable high-order allocations, do that as well, as compaction will
4243 * try prevent permanent fragmentation by migrating from blocks of the
4245 * Don't try this for allocations that are allowed to ignore
4246 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4248 if (can_direct_reclaim && can_compact &&
4250 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4251 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4252 page = __alloc_pages_direct_compact(gfp_mask, order,
4254 INIT_COMPACT_PRIORITY,
4260 * Checks for costly allocations with __GFP_NORETRY, which
4261 * includes some THP page fault allocations
4263 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4265 * If allocating entire pageblock(s) and compaction
4266 * failed because all zones are below low watermarks
4267 * or is prohibited because it recently failed at this
4268 * order, fail immediately unless the allocator has
4269 * requested compaction and reclaim retry.
4272 * - potentially very expensive because zones are far
4273 * below their low watermarks or this is part of very
4274 * bursty high order allocations,
4275 * - not guaranteed to help because isolate_freepages()
4276 * may not iterate over freed pages as part of its
4278 * - unlikely to make entire pageblocks free on its
4281 if (compact_result == COMPACT_SKIPPED ||
4282 compact_result == COMPACT_DEFERRED)
4286 * Looks like reclaim/compaction is worth trying, but
4287 * sync compaction could be very expensive, so keep
4288 * using async compaction.
4290 compact_priority = INIT_COMPACT_PRIORITY;
4295 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4296 if (alloc_flags & ALLOC_KSWAPD)
4297 wake_all_kswapds(order, gfp_mask, ac);
4299 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4301 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4302 (alloc_flags & ALLOC_KSWAPD);
4305 * Reset the nodemask and zonelist iterators if memory policies can be
4306 * ignored. These allocations are high priority and system rather than
4309 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4310 ac->nodemask = NULL;
4311 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4312 ac->highest_zoneidx, ac->nodemask);
4315 /* Attempt with potentially adjusted zonelist and alloc_flags */
4316 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4320 /* Caller is not willing to reclaim, we can't balance anything */
4321 if (!can_direct_reclaim)
4324 /* Avoid recursion of direct reclaim */
4325 if (current->flags & PF_MEMALLOC)
4328 /* Try direct reclaim and then allocating */
4329 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4330 &did_some_progress);
4334 /* Try direct compaction and then allocating */
4335 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4336 compact_priority, &compact_result);
4340 /* Do not loop if specifically requested */
4341 if (gfp_mask & __GFP_NORETRY)
4345 * Do not retry costly high order allocations unless they are
4346 * __GFP_RETRY_MAYFAIL and we can compact
4348 if (costly_order && (!can_compact ||
4349 !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4352 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4353 did_some_progress > 0, &no_progress_loops))
4357 * It doesn't make any sense to retry for the compaction if the order-0
4358 * reclaim is not able to make any progress because the current
4359 * implementation of the compaction depends on the sufficient amount
4360 * of free memory (see __compaction_suitable)
4362 if (did_some_progress > 0 && can_compact &&
4363 should_compact_retry(ac, order, alloc_flags,
4364 compact_result, &compact_priority,
4365 &compaction_retries))
4370 * Deal with possible cpuset update races or zonelist updates to avoid
4371 * a unnecessary OOM kill.
4373 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4374 check_retry_zonelist(zonelist_iter_cookie))
4377 /* Reclaim has failed us, start killing things */
4378 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4382 /* Avoid allocations with no watermarks from looping endlessly */
4383 if (tsk_is_oom_victim(current) &&
4384 (alloc_flags & ALLOC_OOM ||
4385 (gfp_mask & __GFP_NOMEMALLOC)))
4388 /* Retry as long as the OOM killer is making progress */
4389 if (did_some_progress) {
4390 no_progress_loops = 0;
4396 * Deal with possible cpuset update races or zonelist updates to avoid
4397 * a unnecessary OOM kill.
4399 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4400 check_retry_zonelist(zonelist_iter_cookie))
4404 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4407 if (gfp_mask & __GFP_NOFAIL) {
4409 * All existing users of the __GFP_NOFAIL are blockable, so warn
4410 * of any new users that actually require GFP_NOWAIT
4412 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4416 * PF_MEMALLOC request from this context is rather bizarre
4417 * because we cannot reclaim anything and only can loop waiting
4418 * for somebody to do a work for us
4420 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4423 * non failing costly orders are a hard requirement which we
4424 * are not prepared for much so let's warn about these users
4425 * so that we can identify them and convert them to something
4428 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4431 * Help non-failing allocations by giving some access to memory
4432 * reserves normally used for high priority non-blocking
4433 * allocations but do not use ALLOC_NO_WATERMARKS because this
4434 * could deplete whole memory reserves which would just make
4435 * the situation worse.
4437 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4445 warn_alloc(gfp_mask, ac->nodemask,
4446 "page allocation failure: order:%u", order);
4451 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4452 int preferred_nid, nodemask_t *nodemask,
4453 struct alloc_context *ac, gfp_t *alloc_gfp,
4454 unsigned int *alloc_flags)
4456 ac->highest_zoneidx = gfp_zone(gfp_mask);
4457 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4458 ac->nodemask = nodemask;
4459 ac->migratetype = gfp_migratetype(gfp_mask);
4461 if (cpusets_enabled()) {
4462 *alloc_gfp |= __GFP_HARDWALL;
4464 * When we are in the interrupt context, it is irrelevant
4465 * to the current task context. It means that any node ok.
4467 if (in_task() && !ac->nodemask)
4468 ac->nodemask = &cpuset_current_mems_allowed;
4470 *alloc_flags |= ALLOC_CPUSET;
4473 might_alloc(gfp_mask);
4475 if (should_fail_alloc_page(gfp_mask, order))
4478 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4480 /* Dirty zone balancing only done in the fast path */
4481 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4484 * The preferred zone is used for statistics but crucially it is
4485 * also used as the starting point for the zonelist iterator. It
4486 * may get reset for allocations that ignore memory policies.
4488 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4489 ac->highest_zoneidx, ac->nodemask);
4495 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4496 * @gfp: GFP flags for the allocation
4497 * @preferred_nid: The preferred NUMA node ID to allocate from
4498 * @nodemask: Set of nodes to allocate from, may be NULL
4499 * @nr_pages: The number of pages desired on the list or array
4500 * @page_list: Optional list to store the allocated pages
4501 * @page_array: Optional array to store the pages
4503 * This is a batched version of the page allocator that attempts to
4504 * allocate nr_pages quickly. Pages are added to page_list if page_list
4505 * is not NULL, otherwise it is assumed that the page_array is valid.
4507 * For lists, nr_pages is the number of pages that should be allocated.
4509 * For arrays, only NULL elements are populated with pages and nr_pages
4510 * is the maximum number of pages that will be stored in the array.
4512 * Returns the number of pages on the list or array.
4514 unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
4515 nodemask_t *nodemask, int nr_pages,
4516 struct list_head *page_list,
4517 struct page **page_array)
4520 unsigned long __maybe_unused UP_flags;
4523 struct per_cpu_pages *pcp;
4524 struct list_head *pcp_list;
4525 struct alloc_context ac;
4527 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4528 int nr_populated = 0, nr_account = 0;
4531 * Skip populated array elements to determine if any pages need
4532 * to be allocated before disabling IRQs.
4534 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4537 /* No pages requested? */
4538 if (unlikely(nr_pages <= 0))
4541 /* Already populated array? */
4542 if (unlikely(page_array && nr_pages - nr_populated == 0))
4545 /* Bulk allocator does not support memcg accounting. */
4546 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4549 /* Use the single page allocator for one page. */
4550 if (nr_pages - nr_populated == 1)
4553 #ifdef CONFIG_PAGE_OWNER
4555 * PAGE_OWNER may recurse into the allocator to allocate space to
4556 * save the stack with pagesets.lock held. Releasing/reacquiring
4557 * removes much of the performance benefit of bulk allocation so
4558 * force the caller to allocate one page at a time as it'll have
4559 * similar performance to added complexity to the bulk allocator.
4561 if (static_branch_unlikely(&page_owner_inited))
4565 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4566 gfp &= gfp_allowed_mask;
4568 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4572 /* Find an allowed local zone that meets the low watermark. */
4573 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4576 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4577 !__cpuset_zone_allowed(zone, gfp)) {
4581 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4582 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4586 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4587 if (zone_watermark_fast(zone, 0, mark,
4588 zonelist_zone_idx(ac.preferred_zoneref),
4589 alloc_flags, gfp)) {
4595 * If there are no allowed local zones that meets the watermarks then
4596 * try to allocate a single page and reclaim if necessary.
4598 if (unlikely(!zone))
4601 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4602 pcp_trylock_prepare(UP_flags);
4603 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4607 /* Attempt the batch allocation */
4608 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4609 while (nr_populated < nr_pages) {
4611 /* Skip existing pages */
4612 if (page_array && page_array[nr_populated]) {
4617 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4619 if (unlikely(!page)) {
4620 /* Try and allocate at least one page */
4622 pcp_spin_unlock(pcp);
4629 prep_new_page(page, 0, gfp, 0);
4631 list_add(&page->lru, page_list);
4633 page_array[nr_populated] = page;
4637 pcp_spin_unlock(pcp);
4638 pcp_trylock_finish(UP_flags);
4640 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4641 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4644 return nr_populated;
4647 pcp_trylock_finish(UP_flags);
4650 page = __alloc_pages_noprof(gfp, 0, preferred_nid, nodemask);
4653 list_add(&page->lru, page_list);
4655 page_array[nr_populated] = page;
4661 EXPORT_SYMBOL_GPL(alloc_pages_bulk_noprof);
4664 * This is the 'heart' of the zoned buddy allocator.
4666 struct page *__alloc_pages_noprof(gfp_t gfp, unsigned int order,
4667 int preferred_nid, nodemask_t *nodemask)
4670 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4671 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4672 struct alloc_context ac = { };
4675 * There are several places where we assume that the order value is sane
4676 * so bail out early if the request is out of bound.
4678 if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
4681 gfp &= gfp_allowed_mask;
4683 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4684 * resp. GFP_NOIO which has to be inherited for all allocation requests
4685 * from a particular context which has been marked by
4686 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4687 * movable zones are not used during allocation.
4689 gfp = current_gfp_context(gfp);
4691 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4692 &alloc_gfp, &alloc_flags))
4696 * Forbid the first pass from falling back to types that fragment
4697 * memory until all local zones are considered.
4699 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4701 /* First allocation attempt */
4702 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4707 ac.spread_dirty_pages = false;
4710 * Restore the original nodemask if it was potentially replaced with
4711 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4713 ac.nodemask = nodemask;
4715 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4718 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4719 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4720 __free_pages(page, order);
4724 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4725 kmsan_alloc_page(page, order, alloc_gfp);
4729 EXPORT_SYMBOL(__alloc_pages_noprof);
4731 struct folio *__folio_alloc_noprof(gfp_t gfp, unsigned int order, int preferred_nid,
4732 nodemask_t *nodemask)
4734 struct page *page = __alloc_pages_noprof(gfp | __GFP_COMP, order,
4735 preferred_nid, nodemask);
4736 return page_rmappable_folio(page);
4738 EXPORT_SYMBOL(__folio_alloc_noprof);
4741 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4742 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4743 * you need to access high mem.
4745 unsigned long get_free_pages_noprof(gfp_t gfp_mask, unsigned int order)
4749 page = alloc_pages_noprof(gfp_mask & ~__GFP_HIGHMEM, order);
4752 return (unsigned long) page_address(page);
4754 EXPORT_SYMBOL(get_free_pages_noprof);
4756 unsigned long get_zeroed_page_noprof(gfp_t gfp_mask)
4758 return get_free_pages_noprof(gfp_mask | __GFP_ZERO, 0);
4760 EXPORT_SYMBOL(get_zeroed_page_noprof);
4763 * __free_pages - Free pages allocated with alloc_pages().
4764 * @page: The page pointer returned from alloc_pages().
4765 * @order: The order of the allocation.
4767 * This function can free multi-page allocations that are not compound
4768 * pages. It does not check that the @order passed in matches that of
4769 * the allocation, so it is easy to leak memory. Freeing more memory
4770 * than was allocated will probably emit a warning.
4772 * If the last reference to this page is speculative, it will be released
4773 * by put_page() which only frees the first page of a non-compound
4774 * allocation. To prevent the remaining pages from being leaked, we free
4775 * the subsequent pages here. If you want to use the page's reference
4776 * count to decide when to free the allocation, you should allocate a
4777 * compound page, and use put_page() instead of __free_pages().
4779 * Context: May be called in interrupt context or while holding a normal
4780 * spinlock, but not in NMI context or while holding a raw spinlock.
4782 void __free_pages(struct page *page, unsigned int order)
4784 /* get PageHead before we drop reference */
4785 int head = PageHead(page);
4786 struct alloc_tag *tag = pgalloc_tag_get(page);
4788 if (put_page_testzero(page))
4789 free_unref_page(page, order);
4791 pgalloc_tag_sub_pages(tag, (1 << order) - 1);
4793 free_unref_page(page + (1 << order), order);
4796 EXPORT_SYMBOL(__free_pages);
4798 void free_pages(unsigned long addr, unsigned int order)
4801 VM_BUG_ON(!virt_addr_valid((void *)addr));
4802 __free_pages(virt_to_page((void *)addr), order);
4806 EXPORT_SYMBOL(free_pages);
4810 * An arbitrary-length arbitrary-offset area of memory which resides
4811 * within a 0 or higher order page. Multiple fragments within that page
4812 * are individually refcounted, in the page's reference counter.
4814 * The page_frag functions below provide a simple allocation framework for
4815 * page fragments. This is used by the network stack and network device
4816 * drivers to provide a backing region of memory for use as either an
4817 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4819 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4822 struct page *page = NULL;
4823 gfp_t gfp = gfp_mask;
4825 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4826 gfp_mask = (gfp_mask & ~__GFP_DIRECT_RECLAIM) | __GFP_COMP |
4827 __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC;
4828 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4829 PAGE_FRAG_CACHE_MAX_ORDER);
4830 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4832 if (unlikely(!page))
4833 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4835 nc->va = page ? page_address(page) : NULL;
4840 void page_frag_cache_drain(struct page_frag_cache *nc)
4845 __page_frag_cache_drain(virt_to_head_page(nc->va), nc->pagecnt_bias);
4848 EXPORT_SYMBOL(page_frag_cache_drain);
4850 void __page_frag_cache_drain(struct page *page, unsigned int count)
4852 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4854 if (page_ref_sub_and_test(page, count))
4855 free_unref_page(page, compound_order(page));
4857 EXPORT_SYMBOL(__page_frag_cache_drain);
4859 void *__page_frag_alloc_align(struct page_frag_cache *nc,
4860 unsigned int fragsz, gfp_t gfp_mask,
4861 unsigned int align_mask)
4863 unsigned int size = PAGE_SIZE;
4867 if (unlikely(!nc->va)) {
4869 page = __page_frag_cache_refill(nc, gfp_mask);
4873 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4874 /* if size can vary use size else just use PAGE_SIZE */
4877 /* Even if we own the page, we do not use atomic_set().
4878 * This would break get_page_unless_zero() users.
4880 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4882 /* reset page count bias and offset to start of new frag */
4883 nc->pfmemalloc = page_is_pfmemalloc(page);
4884 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4888 offset = nc->offset - fragsz;
4889 if (unlikely(offset < 0)) {
4890 page = virt_to_page(nc->va);
4892 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4895 if (unlikely(nc->pfmemalloc)) {
4896 free_unref_page(page, compound_order(page));
4900 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4901 /* if size can vary use size else just use PAGE_SIZE */
4904 /* OK, page count is 0, we can safely set it */
4905 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4907 /* reset page count bias and offset to start of new frag */
4908 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4909 offset = size - fragsz;
4910 if (unlikely(offset < 0)) {
4912 * The caller is trying to allocate a fragment
4913 * with fragsz > PAGE_SIZE but the cache isn't big
4914 * enough to satisfy the request, this may
4915 * happen in low memory conditions.
4916 * We don't release the cache page because
4917 * it could make memory pressure worse
4918 * so we simply return NULL here.
4925 offset &= align_mask;
4926 nc->offset = offset;
4928 return nc->va + offset;
4930 EXPORT_SYMBOL(__page_frag_alloc_align);
4933 * Frees a page fragment allocated out of either a compound or order 0 page.
4935 void page_frag_free(void *addr)
4937 struct page *page = virt_to_head_page(addr);
4939 if (unlikely(put_page_testzero(page)))
4940 free_unref_page(page, compound_order(page));
4942 EXPORT_SYMBOL(page_frag_free);
4944 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4948 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4949 struct page *page = virt_to_page((void *)addr);
4950 struct page *last = page + nr;
4952 split_page_owner(page, order, 0);
4953 pgalloc_tag_split(page, 1 << order);
4954 split_page_memcg(page, order, 0);
4955 while (page < --last)
4956 set_page_refcounted(last);
4958 last = page + (1UL << order);
4959 for (page += nr; page < last; page++)
4960 __free_pages_ok(page, 0, FPI_TO_TAIL);
4962 return (void *)addr;
4966 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4967 * @size: the number of bytes to allocate
4968 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4970 * This function is similar to alloc_pages(), except that it allocates the
4971 * minimum number of pages to satisfy the request. alloc_pages() can only
4972 * allocate memory in power-of-two pages.
4974 * This function is also limited by MAX_PAGE_ORDER.
4976 * Memory allocated by this function must be released by free_pages_exact().
4978 * Return: pointer to the allocated area or %NULL in case of error.
4980 void *alloc_pages_exact_noprof(size_t size, gfp_t gfp_mask)
4982 unsigned int order = get_order(size);
4985 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4986 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4988 addr = get_free_pages_noprof(gfp_mask, order);
4989 return make_alloc_exact(addr, order, size);
4991 EXPORT_SYMBOL(alloc_pages_exact_noprof);
4994 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4996 * @nid: the preferred node ID where memory should be allocated
4997 * @size: the number of bytes to allocate
4998 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5000 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5003 * Return: pointer to the allocated area or %NULL in case of error.
5005 void * __meminit alloc_pages_exact_nid_noprof(int nid, size_t size, gfp_t gfp_mask)
5007 unsigned int order = get_order(size);
5010 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5011 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5013 p = alloc_pages_node_noprof(nid, gfp_mask, order);
5016 return make_alloc_exact((unsigned long)page_address(p), order, size);
5020 * free_pages_exact - release memory allocated via alloc_pages_exact()
5021 * @virt: the value returned by alloc_pages_exact.
5022 * @size: size of allocation, same value as passed to alloc_pages_exact().
5024 * Release the memory allocated by a previous call to alloc_pages_exact.
5026 void free_pages_exact(void *virt, size_t size)
5028 unsigned long addr = (unsigned long)virt;
5029 unsigned long end = addr + PAGE_ALIGN(size);
5031 while (addr < end) {
5036 EXPORT_SYMBOL(free_pages_exact);
5039 * nr_free_zone_pages - count number of pages beyond high watermark
5040 * @offset: The zone index of the highest zone
5042 * nr_free_zone_pages() counts the number of pages which are beyond the
5043 * high watermark within all zones at or below a given zone index. For each
5044 * zone, the number of pages is calculated as:
5046 * nr_free_zone_pages = managed_pages - high_pages
5048 * Return: number of pages beyond high watermark.
5050 static unsigned long nr_free_zone_pages(int offset)
5055 /* Just pick one node, since fallback list is circular */
5056 unsigned long sum = 0;
5058 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5060 for_each_zone_zonelist(zone, z, zonelist, offset) {
5061 unsigned long size = zone_managed_pages(zone);
5062 unsigned long high = high_wmark_pages(zone);
5071 * nr_free_buffer_pages - count number of pages beyond high watermark
5073 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5074 * watermark within ZONE_DMA and ZONE_NORMAL.
5076 * Return: number of pages beyond high watermark within ZONE_DMA and
5079 unsigned long nr_free_buffer_pages(void)
5081 return nr_free_zone_pages(gfp_zone(GFP_USER));
5083 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5085 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5087 zoneref->zone = zone;
5088 zoneref->zone_idx = zone_idx(zone);
5092 * Builds allocation fallback zone lists.
5094 * Add all populated zones of a node to the zonelist.
5096 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5099 enum zone_type zone_type = MAX_NR_ZONES;
5104 zone = pgdat->node_zones + zone_type;
5105 if (populated_zone(zone)) {
5106 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5107 check_highest_zone(zone_type);
5109 } while (zone_type);
5116 static int __parse_numa_zonelist_order(char *s)
5119 * We used to support different zonelists modes but they turned
5120 * out to be just not useful. Let's keep the warning in place
5121 * if somebody still use the cmd line parameter so that we do
5122 * not fail it silently
5124 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5125 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5131 static char numa_zonelist_order[] = "Node";
5132 #define NUMA_ZONELIST_ORDER_LEN 16
5134 * sysctl handler for numa_zonelist_order
5136 static int numa_zonelist_order_handler(const struct ctl_table *table, int write,
5137 void *buffer, size_t *length, loff_t *ppos)
5140 return __parse_numa_zonelist_order(buffer);
5141 return proc_dostring(table, write, buffer, length, ppos);
5144 static int node_load[MAX_NUMNODES];
5147 * find_next_best_node - find the next node that should appear in a given node's fallback list
5148 * @node: node whose fallback list we're appending
5149 * @used_node_mask: nodemask_t of already used nodes
5151 * We use a number of factors to determine which is the next node that should
5152 * appear on a given node's fallback list. The node should not have appeared
5153 * already in @node's fallback list, and it should be the next closest node
5154 * according to the distance array (which contains arbitrary distance values
5155 * from each node to each node in the system), and should also prefer nodes
5156 * with no CPUs, since presumably they'll have very little allocation pressure
5157 * on them otherwise.
5159 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5161 int find_next_best_node(int node, nodemask_t *used_node_mask)
5164 int min_val = INT_MAX;
5165 int best_node = NUMA_NO_NODE;
5168 * Use the local node if we haven't already, but for memoryless local
5169 * node, we should skip it and fall back to other nodes.
5171 if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5172 node_set(node, *used_node_mask);
5176 for_each_node_state(n, N_MEMORY) {
5178 /* Don't want a node to appear more than once */
5179 if (node_isset(n, *used_node_mask))
5182 /* Use the distance array to find the distance */
5183 val = node_distance(node, n);
5185 /* Penalize nodes under us ("prefer the next node") */
5188 /* Give preference to headless and unused nodes */
5189 if (!cpumask_empty(cpumask_of_node(n)))
5190 val += PENALTY_FOR_NODE_WITH_CPUS;
5192 /* Slight preference for less loaded node */
5193 val *= MAX_NUMNODES;
5194 val += node_load[n];
5196 if (val < min_val) {
5203 node_set(best_node, *used_node_mask);
5210 * Build zonelists ordered by node and zones within node.
5211 * This results in maximum locality--normal zone overflows into local
5212 * DMA zone, if any--but risks exhausting DMA zone.
5214 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5217 struct zoneref *zonerefs;
5220 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5222 for (i = 0; i < nr_nodes; i++) {
5225 pg_data_t *node = NODE_DATA(node_order[i]);
5227 nr_zones = build_zonerefs_node(node, zonerefs);
5228 zonerefs += nr_zones;
5230 zonerefs->zone = NULL;
5231 zonerefs->zone_idx = 0;
5235 * Build __GFP_THISNODE zonelists
5237 static void build_thisnode_zonelists(pg_data_t *pgdat)
5239 struct zoneref *zonerefs;
5242 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5243 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5244 zonerefs += nr_zones;
5245 zonerefs->zone = NULL;
5246 zonerefs->zone_idx = 0;
5250 * Build zonelists ordered by zone and nodes within zones.
5251 * This results in conserving DMA zone[s] until all Normal memory is
5252 * exhausted, but results in overflowing to remote node while memory
5253 * may still exist in local DMA zone.
5256 static void build_zonelists(pg_data_t *pgdat)
5258 static int node_order[MAX_NUMNODES];
5259 int node, nr_nodes = 0;
5260 nodemask_t used_mask = NODE_MASK_NONE;
5261 int local_node, prev_node;
5263 /* NUMA-aware ordering of nodes */
5264 local_node = pgdat->node_id;
5265 prev_node = local_node;
5267 memset(node_order, 0, sizeof(node_order));
5268 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5270 * We don't want to pressure a particular node.
5271 * So adding penalty to the first node in same
5272 * distance group to make it round-robin.
5274 if (node_distance(local_node, node) !=
5275 node_distance(local_node, prev_node))
5276 node_load[node] += 1;
5278 node_order[nr_nodes++] = node;
5282 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5283 build_thisnode_zonelists(pgdat);
5284 pr_info("Fallback order for Node %d: ", local_node);
5285 for (node = 0; node < nr_nodes; node++)
5286 pr_cont("%d ", node_order[node]);
5290 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5292 * Return node id of node used for "local" allocations.
5293 * I.e., first node id of first zone in arg node's generic zonelist.
5294 * Used for initializing percpu 'numa_mem', which is used primarily
5295 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5297 int local_memory_node(int node)
5301 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5302 gfp_zone(GFP_KERNEL),
5304 return zone_to_nid(z->zone);
5308 static void setup_min_unmapped_ratio(void);
5309 static void setup_min_slab_ratio(void);
5310 #else /* CONFIG_NUMA */
5312 static void build_zonelists(pg_data_t *pgdat)
5314 struct zoneref *zonerefs;
5317 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5318 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5319 zonerefs += nr_zones;
5321 zonerefs->zone = NULL;
5322 zonerefs->zone_idx = 0;
5325 #endif /* CONFIG_NUMA */
5328 * Boot pageset table. One per cpu which is going to be used for all
5329 * zones and all nodes. The parameters will be set in such a way
5330 * that an item put on a list will immediately be handed over to
5331 * the buddy list. This is safe since pageset manipulation is done
5332 * with interrupts disabled.
5334 * The boot_pagesets must be kept even after bootup is complete for
5335 * unused processors and/or zones. They do play a role for bootstrapping
5336 * hotplugged processors.
5338 * zoneinfo_show() and maybe other functions do
5339 * not check if the processor is online before following the pageset pointer.
5340 * Other parts of the kernel may not check if the zone is available.
5342 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5343 /* These effectively disable the pcplists in the boot pageset completely */
5344 #define BOOT_PAGESET_HIGH 0
5345 #define BOOT_PAGESET_BATCH 1
5346 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5347 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5349 static void __build_all_zonelists(void *data)
5352 int __maybe_unused cpu;
5353 pg_data_t *self = data;
5354 unsigned long flags;
5357 * The zonelist_update_seq must be acquired with irqsave because the
5358 * reader can be invoked from IRQ with GFP_ATOMIC.
5360 write_seqlock_irqsave(&zonelist_update_seq, flags);
5362 * Also disable synchronous printk() to prevent any printk() from
5363 * trying to hold port->lock, for
5364 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5365 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5367 printk_deferred_enter();
5370 memset(node_load, 0, sizeof(node_load));
5374 * This node is hotadded and no memory is yet present. So just
5375 * building zonelists is fine - no need to touch other nodes.
5377 if (self && !node_online(self->node_id)) {
5378 build_zonelists(self);
5381 * All possible nodes have pgdat preallocated
5384 for_each_node(nid) {
5385 pg_data_t *pgdat = NODE_DATA(nid);
5387 build_zonelists(pgdat);
5390 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5392 * We now know the "local memory node" for each node--
5393 * i.e., the node of the first zone in the generic zonelist.
5394 * Set up numa_mem percpu variable for on-line cpus. During
5395 * boot, only the boot cpu should be on-line; we'll init the
5396 * secondary cpus' numa_mem as they come on-line. During
5397 * node/memory hotplug, we'll fixup all on-line cpus.
5399 for_each_online_cpu(cpu)
5400 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5404 printk_deferred_exit();
5405 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5408 static noinline void __init
5409 build_all_zonelists_init(void)
5413 __build_all_zonelists(NULL);
5416 * Initialize the boot_pagesets that are going to be used
5417 * for bootstrapping processors. The real pagesets for
5418 * each zone will be allocated later when the per cpu
5419 * allocator is available.
5421 * boot_pagesets are used also for bootstrapping offline
5422 * cpus if the system is already booted because the pagesets
5423 * are needed to initialize allocators on a specific cpu too.
5424 * F.e. the percpu allocator needs the page allocator which
5425 * needs the percpu allocator in order to allocate its pagesets
5426 * (a chicken-egg dilemma).
5428 for_each_possible_cpu(cpu)
5429 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5431 mminit_verify_zonelist();
5432 cpuset_init_current_mems_allowed();
5436 * unless system_state == SYSTEM_BOOTING.
5438 * __ref due to call of __init annotated helper build_all_zonelists_init
5439 * [protected by SYSTEM_BOOTING].
5441 void __ref build_all_zonelists(pg_data_t *pgdat)
5443 unsigned long vm_total_pages;
5445 if (system_state == SYSTEM_BOOTING) {
5446 build_all_zonelists_init();
5448 __build_all_zonelists(pgdat);
5449 /* cpuset refresh routine should be here */
5451 /* Get the number of free pages beyond high watermark in all zones. */
5452 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5454 * Disable grouping by mobility if the number of pages in the
5455 * system is too low to allow the mechanism to work. It would be
5456 * more accurate, but expensive to check per-zone. This check is
5457 * made on memory-hotadd so a system can start with mobility
5458 * disabled and enable it later
5460 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5461 page_group_by_mobility_disabled = 1;
5463 page_group_by_mobility_disabled = 0;
5465 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5467 page_group_by_mobility_disabled ? "off" : "on",
5470 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5474 static int zone_batchsize(struct zone *zone)
5480 * The number of pages to batch allocate is either ~0.1%
5481 * of the zone or 1MB, whichever is smaller. The batch
5482 * size is striking a balance between allocation latency
5483 * and zone lock contention.
5485 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5486 batch /= 4; /* We effectively *= 4 below */
5491 * Clamp the batch to a 2^n - 1 value. Having a power
5492 * of 2 value was found to be more likely to have
5493 * suboptimal cache aliasing properties in some cases.
5495 * For example if 2 tasks are alternately allocating
5496 * batches of pages, one task can end up with a lot
5497 * of pages of one half of the possible page colors
5498 * and the other with pages of the other colors.
5500 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5505 /* The deferral and batching of frees should be suppressed under NOMMU
5508 * The problem is that NOMMU needs to be able to allocate large chunks
5509 * of contiguous memory as there's no hardware page translation to
5510 * assemble apparent contiguous memory from discontiguous pages.
5512 * Queueing large contiguous runs of pages for batching, however,
5513 * causes the pages to actually be freed in smaller chunks. As there
5514 * can be a significant delay between the individual batches being
5515 * recycled, this leads to the once large chunks of space being
5516 * fragmented and becoming unavailable for high-order allocations.
5522 static int percpu_pagelist_high_fraction;
5523 static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5529 unsigned long total_pages;
5531 if (!high_fraction) {
5533 * By default, the high value of the pcp is based on the zone
5534 * low watermark so that if they are full then background
5535 * reclaim will not be started prematurely.
5537 total_pages = low_wmark_pages(zone);
5540 * If percpu_pagelist_high_fraction is configured, the high
5541 * value is based on a fraction of the managed pages in the
5544 total_pages = zone_managed_pages(zone) / high_fraction;
5548 * Split the high value across all online CPUs local to the zone. Note
5549 * that early in boot that CPUs may not be online yet and that during
5550 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5551 * onlined. For memory nodes that have no CPUs, split the high value
5552 * across all online CPUs to mitigate the risk that reclaim is triggered
5553 * prematurely due to pages stored on pcp lists.
5555 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5557 nr_split_cpus = num_online_cpus();
5558 high = total_pages / nr_split_cpus;
5561 * Ensure high is at least batch*4. The multiple is based on the
5562 * historical relationship between high and batch.
5564 high = max(high, batch << 2);
5573 * pcp->high and pcp->batch values are related and generally batch is lower
5574 * than high. They are also related to pcp->count such that count is lower
5575 * than high, and as soon as it reaches high, the pcplist is flushed.
5577 * However, guaranteeing these relations at all times would require e.g. write
5578 * barriers here but also careful usage of read barriers at the read side, and
5579 * thus be prone to error and bad for performance. Thus the update only prevents
5580 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5581 * should ensure they can cope with those fields changing asynchronously, and
5582 * fully trust only the pcp->count field on the local CPU with interrupts
5585 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5586 * outside of boot time (or some other assurance that no concurrent updaters
5589 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5590 unsigned long high_max, unsigned long batch)
5592 WRITE_ONCE(pcp->batch, batch);
5593 WRITE_ONCE(pcp->high_min, high_min);
5594 WRITE_ONCE(pcp->high_max, high_max);
5597 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5601 memset(pcp, 0, sizeof(*pcp));
5602 memset(pzstats, 0, sizeof(*pzstats));
5604 spin_lock_init(&pcp->lock);
5605 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5606 INIT_LIST_HEAD(&pcp->lists[pindex]);
5609 * Set batch and high values safe for a boot pageset. A true percpu
5610 * pageset's initialization will update them subsequently. Here we don't
5611 * need to be as careful as pageset_update() as nobody can access the
5614 pcp->high_min = BOOT_PAGESET_HIGH;
5615 pcp->high_max = BOOT_PAGESET_HIGH;
5616 pcp->batch = BOOT_PAGESET_BATCH;
5617 pcp->free_count = 0;
5620 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
5621 unsigned long high_max, unsigned long batch)
5623 struct per_cpu_pages *pcp;
5626 for_each_possible_cpu(cpu) {
5627 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5628 pageset_update(pcp, high_min, high_max, batch);
5633 * Calculate and set new high and batch values for all per-cpu pagesets of a
5634 * zone based on the zone's size.
5636 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5638 int new_high_min, new_high_max, new_batch;
5640 new_batch = max(1, zone_batchsize(zone));
5641 if (percpu_pagelist_high_fraction) {
5642 new_high_min = zone_highsize(zone, new_batch, cpu_online,
5643 percpu_pagelist_high_fraction);
5645 * PCP high is tuned manually, disable auto-tuning via
5646 * setting high_min and high_max to the manual value.
5648 new_high_max = new_high_min;
5650 new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
5651 new_high_max = zone_highsize(zone, new_batch, cpu_online,
5652 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
5655 if (zone->pageset_high_min == new_high_min &&
5656 zone->pageset_high_max == new_high_max &&
5657 zone->pageset_batch == new_batch)
5660 zone->pageset_high_min = new_high_min;
5661 zone->pageset_high_max = new_high_max;
5662 zone->pageset_batch = new_batch;
5664 __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
5668 void __meminit setup_zone_pageset(struct zone *zone)
5672 /* Size may be 0 on !SMP && !NUMA */
5673 if (sizeof(struct per_cpu_zonestat) > 0)
5674 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5676 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5677 for_each_possible_cpu(cpu) {
5678 struct per_cpu_pages *pcp;
5679 struct per_cpu_zonestat *pzstats;
5681 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5682 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5683 per_cpu_pages_init(pcp, pzstats);
5686 zone_set_pageset_high_and_batch(zone, 0);
5690 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5691 * page high values need to be recalculated.
5693 static void zone_pcp_update(struct zone *zone, int cpu_online)
5695 mutex_lock(&pcp_batch_high_lock);
5696 zone_set_pageset_high_and_batch(zone, cpu_online);
5697 mutex_unlock(&pcp_batch_high_lock);
5700 static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu)
5702 struct per_cpu_pages *pcp;
5703 struct cpu_cacheinfo *cci;
5705 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5706 cci = get_cpu_cacheinfo(cpu);
5708 * If data cache slice of CPU is large enough, "pcp->batch"
5709 * pages can be preserved in PCP before draining PCP for
5710 * consecutive high-order pages freeing without allocation.
5711 * This can reduce zone lock contention without hurting
5712 * cache-hot pages sharing.
5714 spin_lock(&pcp->lock);
5715 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
5716 pcp->flags |= PCPF_FREE_HIGH_BATCH;
5718 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
5719 spin_unlock(&pcp->lock);
5722 void setup_pcp_cacheinfo(unsigned int cpu)
5726 for_each_populated_zone(zone)
5727 zone_pcp_update_cacheinfo(zone, cpu);
5731 * Allocate per cpu pagesets and initialize them.
5732 * Before this call only boot pagesets were available.
5734 void __init setup_per_cpu_pageset(void)
5736 struct pglist_data *pgdat;
5738 int __maybe_unused cpu;
5740 for_each_populated_zone(zone)
5741 setup_zone_pageset(zone);
5745 * Unpopulated zones continue using the boot pagesets.
5746 * The numa stats for these pagesets need to be reset.
5747 * Otherwise, they will end up skewing the stats of
5748 * the nodes these zones are associated with.
5750 for_each_possible_cpu(cpu) {
5751 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5752 memset(pzstats->vm_numa_event, 0,
5753 sizeof(pzstats->vm_numa_event));
5757 for_each_online_pgdat(pgdat)
5758 pgdat->per_cpu_nodestats =
5759 alloc_percpu(struct per_cpu_nodestat);
5762 __meminit void zone_pcp_init(struct zone *zone)
5765 * per cpu subsystem is not up at this point. The following code
5766 * relies on the ability of the linker to provide the
5767 * offset of a (static) per cpu variable into the per cpu area.
5769 zone->per_cpu_pageset = &boot_pageset;
5770 zone->per_cpu_zonestats = &boot_zonestats;
5771 zone->pageset_high_min = BOOT_PAGESET_HIGH;
5772 zone->pageset_high_max = BOOT_PAGESET_HIGH;
5773 zone->pageset_batch = BOOT_PAGESET_BATCH;
5775 if (populated_zone(zone))
5776 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5777 zone->present_pages, zone_batchsize(zone));
5780 void adjust_managed_page_count(struct page *page, long count)
5782 atomic_long_add(count, &page_zone(page)->managed_pages);
5783 totalram_pages_add(count);
5785 EXPORT_SYMBOL(adjust_managed_page_count);
5787 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5790 unsigned long pages = 0;
5792 start = (void *)PAGE_ALIGN((unsigned long)start);
5793 end = (void *)((unsigned long)end & PAGE_MASK);
5794 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5795 struct page *page = virt_to_page(pos);
5796 void *direct_map_addr;
5799 * 'direct_map_addr' might be different from 'pos'
5800 * because some architectures' virt_to_page()
5801 * work with aliases. Getting the direct map
5802 * address ensures that we get a _writeable_
5803 * alias for the memset().
5805 direct_map_addr = page_address(page);
5807 * Perform a kasan-unchecked memset() since this memory
5808 * has not been initialized.
5810 direct_map_addr = kasan_reset_tag(direct_map_addr);
5811 if ((unsigned int)poison <= 0xFF)
5812 memset(direct_map_addr, poison, PAGE_SIZE);
5814 free_reserved_page(page);
5818 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5823 void free_reserved_page(struct page *page)
5825 clear_page_tag_ref(page);
5826 ClearPageReserved(page);
5827 init_page_count(page);
5829 adjust_managed_page_count(page, 1);
5831 EXPORT_SYMBOL(free_reserved_page);
5833 static int page_alloc_cpu_dead(unsigned int cpu)
5837 lru_add_drain_cpu(cpu);
5838 mlock_drain_remote(cpu);
5842 * Spill the event counters of the dead processor
5843 * into the current processors event counters.
5844 * This artificially elevates the count of the current
5847 vm_events_fold_cpu(cpu);
5850 * Zero the differential counters of the dead processor
5851 * so that the vm statistics are consistent.
5853 * This is only okay since the processor is dead and cannot
5854 * race with what we are doing.
5856 cpu_vm_stats_fold(cpu);
5858 for_each_populated_zone(zone)
5859 zone_pcp_update(zone, 0);
5864 static int page_alloc_cpu_online(unsigned int cpu)
5868 for_each_populated_zone(zone)
5869 zone_pcp_update(zone, 1);
5873 void __init page_alloc_init_cpuhp(void)
5877 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5878 "mm/page_alloc:pcp",
5879 page_alloc_cpu_online,
5880 page_alloc_cpu_dead);
5885 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5886 * or min_free_kbytes changes.
5888 static void calculate_totalreserve_pages(void)
5890 struct pglist_data *pgdat;
5891 unsigned long reserve_pages = 0;
5892 enum zone_type i, j;
5894 for_each_online_pgdat(pgdat) {
5896 pgdat->totalreserve_pages = 0;
5898 for (i = 0; i < MAX_NR_ZONES; i++) {
5899 struct zone *zone = pgdat->node_zones + i;
5901 unsigned long managed_pages = zone_managed_pages(zone);
5903 /* Find valid and maximum lowmem_reserve in the zone */
5904 for (j = i; j < MAX_NR_ZONES; j++) {
5905 if (zone->lowmem_reserve[j] > max)
5906 max = zone->lowmem_reserve[j];
5909 /* we treat the high watermark as reserved pages. */
5910 max += high_wmark_pages(zone);
5912 if (max > managed_pages)
5913 max = managed_pages;
5915 pgdat->totalreserve_pages += max;
5917 reserve_pages += max;
5920 totalreserve_pages = reserve_pages;
5924 * setup_per_zone_lowmem_reserve - called whenever
5925 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5926 * has a correct pages reserved value, so an adequate number of
5927 * pages are left in the zone after a successful __alloc_pages().
5929 static void setup_per_zone_lowmem_reserve(void)
5931 struct pglist_data *pgdat;
5932 enum zone_type i, j;
5934 for_each_online_pgdat(pgdat) {
5935 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5936 struct zone *zone = &pgdat->node_zones[i];
5937 int ratio = sysctl_lowmem_reserve_ratio[i];
5938 bool clear = !ratio || !zone_managed_pages(zone);
5939 unsigned long managed_pages = 0;
5941 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5942 struct zone *upper_zone = &pgdat->node_zones[j];
5943 bool empty = !zone_managed_pages(upper_zone);
5945 managed_pages += zone_managed_pages(upper_zone);
5948 zone->lowmem_reserve[j] = 0;
5950 zone->lowmem_reserve[j] = managed_pages / ratio;
5955 /* update totalreserve_pages */
5956 calculate_totalreserve_pages();
5959 static void __setup_per_zone_wmarks(void)
5961 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5962 unsigned long lowmem_pages = 0;
5964 unsigned long flags;
5966 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5967 for_each_zone(zone) {
5968 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5969 lowmem_pages += zone_managed_pages(zone);
5972 for_each_zone(zone) {
5975 spin_lock_irqsave(&zone->lock, flags);
5976 tmp = (u64)pages_min * zone_managed_pages(zone);
5977 tmp = div64_ul(tmp, lowmem_pages);
5978 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5980 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5981 * need highmem and movable zones pages, so cap pages_min
5982 * to a small value here.
5984 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5985 * deltas control async page reclaim, and so should
5986 * not be capped for highmem and movable zones.
5988 unsigned long min_pages;
5990 min_pages = zone_managed_pages(zone) / 1024;
5991 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5992 zone->_watermark[WMARK_MIN] = min_pages;
5995 * If it's a lowmem zone, reserve a number of pages
5996 * proportionate to the zone's size.
5998 zone->_watermark[WMARK_MIN] = tmp;
6002 * Set the kswapd watermarks distance according to the
6003 * scale factor in proportion to available memory, but
6004 * ensure a minimum size on small systems.
6006 tmp = max_t(u64, tmp >> 2,
6007 mult_frac(zone_managed_pages(zone),
6008 watermark_scale_factor, 10000));
6010 zone->watermark_boost = 0;
6011 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6012 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
6013 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
6015 spin_unlock_irqrestore(&zone->lock, flags);
6018 /* update totalreserve_pages */
6019 calculate_totalreserve_pages();
6023 * setup_per_zone_wmarks - called when min_free_kbytes changes
6024 * or when memory is hot-{added|removed}
6026 * Ensures that the watermark[min,low,high] values for each zone are set
6027 * correctly with respect to min_free_kbytes.
6029 void setup_per_zone_wmarks(void)
6032 static DEFINE_SPINLOCK(lock);
6035 __setup_per_zone_wmarks();
6039 * The watermark size have changed so update the pcpu batch
6040 * and high limits or the limits may be inappropriate.
6043 zone_pcp_update(zone, 0);
6047 * Initialise min_free_kbytes.
6049 * For small machines we want it small (128k min). For large machines
6050 * we want it large (256MB max). But it is not linear, because network
6051 * bandwidth does not increase linearly with machine size. We use
6053 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6054 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6070 void calculate_min_free_kbytes(void)
6072 unsigned long lowmem_kbytes;
6073 int new_min_free_kbytes;
6075 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6076 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6078 if (new_min_free_kbytes > user_min_free_kbytes)
6079 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
6081 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6082 new_min_free_kbytes, user_min_free_kbytes);
6086 int __meminit init_per_zone_wmark_min(void)
6088 calculate_min_free_kbytes();
6089 setup_per_zone_wmarks();
6090 refresh_zone_stat_thresholds();
6091 setup_per_zone_lowmem_reserve();
6094 setup_min_unmapped_ratio();
6095 setup_min_slab_ratio();
6098 khugepaged_min_free_kbytes_update();
6102 postcore_initcall(init_per_zone_wmark_min)
6105 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6106 * that we can call two helper functions whenever min_free_kbytes
6109 static int min_free_kbytes_sysctl_handler(const struct ctl_table *table, int write,
6110 void *buffer, size_t *length, loff_t *ppos)
6114 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6119 user_min_free_kbytes = min_free_kbytes;
6120 setup_per_zone_wmarks();
6125 static int watermark_scale_factor_sysctl_handler(const struct ctl_table *table, int write,
6126 void *buffer, size_t *length, loff_t *ppos)
6130 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6135 setup_per_zone_wmarks();
6141 static void setup_min_unmapped_ratio(void)
6146 for_each_online_pgdat(pgdat)
6147 pgdat->min_unmapped_pages = 0;
6150 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6151 sysctl_min_unmapped_ratio) / 100;
6155 static int sysctl_min_unmapped_ratio_sysctl_handler(const struct ctl_table *table, int write,
6156 void *buffer, size_t *length, loff_t *ppos)
6160 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6164 setup_min_unmapped_ratio();
6169 static void setup_min_slab_ratio(void)
6174 for_each_online_pgdat(pgdat)
6175 pgdat->min_slab_pages = 0;
6178 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6179 sysctl_min_slab_ratio) / 100;
6182 static int sysctl_min_slab_ratio_sysctl_handler(const struct ctl_table *table, int write,
6183 void *buffer, size_t *length, loff_t *ppos)
6187 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6191 setup_min_slab_ratio();
6198 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6199 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6200 * whenever sysctl_lowmem_reserve_ratio changes.
6202 * The reserve ratio obviously has absolutely no relation with the
6203 * minimum watermarks. The lowmem reserve ratio can only make sense
6204 * if in function of the boot time zone sizes.
6206 static int lowmem_reserve_ratio_sysctl_handler(const struct ctl_table *table,
6207 int write, void *buffer, size_t *length, loff_t *ppos)
6211 proc_dointvec_minmax(table, write, buffer, length, ppos);
6213 for (i = 0; i < MAX_NR_ZONES; i++) {
6214 if (sysctl_lowmem_reserve_ratio[i] < 1)
6215 sysctl_lowmem_reserve_ratio[i] = 0;
6218 setup_per_zone_lowmem_reserve();
6223 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6224 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6225 * pagelist can have before it gets flushed back to buddy allocator.
6227 static int percpu_pagelist_high_fraction_sysctl_handler(const struct ctl_table *table,
6228 int write, void *buffer, size_t *length, loff_t *ppos)
6231 int old_percpu_pagelist_high_fraction;
6234 mutex_lock(&pcp_batch_high_lock);
6235 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6237 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6238 if (!write || ret < 0)
6241 /* Sanity checking to avoid pcp imbalance */
6242 if (percpu_pagelist_high_fraction &&
6243 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6244 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6250 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6253 for_each_populated_zone(zone)
6254 zone_set_pageset_high_and_batch(zone, 0);
6256 mutex_unlock(&pcp_batch_high_lock);
6260 static struct ctl_table page_alloc_sysctl_table[] = {
6262 .procname = "min_free_kbytes",
6263 .data = &min_free_kbytes,
6264 .maxlen = sizeof(min_free_kbytes),
6266 .proc_handler = min_free_kbytes_sysctl_handler,
6267 .extra1 = SYSCTL_ZERO,
6270 .procname = "watermark_boost_factor",
6271 .data = &watermark_boost_factor,
6272 .maxlen = sizeof(watermark_boost_factor),
6274 .proc_handler = proc_dointvec_minmax,
6275 .extra1 = SYSCTL_ZERO,
6278 .procname = "watermark_scale_factor",
6279 .data = &watermark_scale_factor,
6280 .maxlen = sizeof(watermark_scale_factor),
6282 .proc_handler = watermark_scale_factor_sysctl_handler,
6283 .extra1 = SYSCTL_ONE,
6284 .extra2 = SYSCTL_THREE_THOUSAND,
6287 .procname = "percpu_pagelist_high_fraction",
6288 .data = &percpu_pagelist_high_fraction,
6289 .maxlen = sizeof(percpu_pagelist_high_fraction),
6291 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6292 .extra1 = SYSCTL_ZERO,
6295 .procname = "lowmem_reserve_ratio",
6296 .data = &sysctl_lowmem_reserve_ratio,
6297 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6299 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6303 .procname = "numa_zonelist_order",
6304 .data = &numa_zonelist_order,
6305 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6307 .proc_handler = numa_zonelist_order_handler,
6310 .procname = "min_unmapped_ratio",
6311 .data = &sysctl_min_unmapped_ratio,
6312 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6314 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6315 .extra1 = SYSCTL_ZERO,
6316 .extra2 = SYSCTL_ONE_HUNDRED,
6319 .procname = "min_slab_ratio",
6320 .data = &sysctl_min_slab_ratio,
6321 .maxlen = sizeof(sysctl_min_slab_ratio),
6323 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6324 .extra1 = SYSCTL_ZERO,
6325 .extra2 = SYSCTL_ONE_HUNDRED,
6330 void __init page_alloc_sysctl_init(void)
6332 register_sysctl_init("vm", page_alloc_sysctl_table);
6335 #ifdef CONFIG_CONTIG_ALLOC
6336 /* Usage: See admin-guide/dynamic-debug-howto.rst */
6337 static void alloc_contig_dump_pages(struct list_head *page_list)
6339 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6341 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6345 list_for_each_entry(page, page_list, lru)
6346 dump_page(page, "migration failure");
6351 * [start, end) must belong to a single zone.
6352 * @migratetype: using migratetype to filter the type of migration in
6353 * trace_mm_alloc_contig_migrate_range_info.
6355 int __alloc_contig_migrate_range(struct compact_control *cc,
6356 unsigned long start, unsigned long end,
6359 /* This function is based on compact_zone() from compaction.c. */
6360 unsigned int nr_reclaimed;
6361 unsigned long pfn = start;
6362 unsigned int tries = 0;
6364 struct migration_target_control mtc = {
6365 .nid = zone_to_nid(cc->zone),
6366 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6367 .reason = MR_CONTIG_RANGE,
6370 unsigned long total_mapped = 0;
6371 unsigned long total_migrated = 0;
6372 unsigned long total_reclaimed = 0;
6374 lru_cache_disable();
6376 while (pfn < end || !list_empty(&cc->migratepages)) {
6377 if (fatal_signal_pending(current)) {
6382 if (list_empty(&cc->migratepages)) {
6383 cc->nr_migratepages = 0;
6384 ret = isolate_migratepages_range(cc, pfn, end);
6385 if (ret && ret != -EAGAIN)
6387 pfn = cc->migrate_pfn;
6389 } else if (++tries == 5) {
6394 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6396 cc->nr_migratepages -= nr_reclaimed;
6398 if (trace_mm_alloc_contig_migrate_range_info_enabled()) {
6399 total_reclaimed += nr_reclaimed;
6400 list_for_each_entry(page, &cc->migratepages, lru) {
6401 struct folio *folio = page_folio(page);
6403 total_mapped += folio_mapped(folio) *
6404 folio_nr_pages(folio);
6408 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6409 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6411 if (trace_mm_alloc_contig_migrate_range_info_enabled() && !ret)
6412 total_migrated += cc->nr_migratepages;
6415 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6416 * to retry again over this error, so do the same here.
6424 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6425 alloc_contig_dump_pages(&cc->migratepages);
6426 putback_movable_pages(&cc->migratepages);
6429 trace_mm_alloc_contig_migrate_range_info(start, end, migratetype,
6433 return (ret < 0) ? ret : 0;
6437 * alloc_contig_range() -- tries to allocate given range of pages
6438 * @start: start PFN to allocate
6439 * @end: one-past-the-last PFN to allocate
6440 * @migratetype: migratetype of the underlying pageblocks (either
6441 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6442 * in range must have the same migratetype and it must
6443 * be either of the two.
6444 * @gfp_mask: GFP mask to use during compaction
6446 * The PFN range does not have to be pageblock aligned. The PFN range must
6447 * belong to a single zone.
6449 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6450 * pageblocks in the range. Once isolated, the pageblocks should not
6451 * be modified by others.
6453 * Return: zero on success or negative error code. On success all
6454 * pages which PFN is in [start, end) are allocated for the caller and
6455 * need to be freed with free_contig_range().
6457 int alloc_contig_range_noprof(unsigned long start, unsigned long end,
6458 unsigned migratetype, gfp_t gfp_mask)
6460 unsigned long outer_start, outer_end;
6463 struct compact_control cc = {
6464 .nr_migratepages = 0,
6466 .zone = page_zone(pfn_to_page(start)),
6467 .mode = MIGRATE_SYNC,
6468 .ignore_skip_hint = true,
6469 .no_set_skip_hint = true,
6470 .gfp_mask = current_gfp_context(gfp_mask),
6471 .alloc_contig = true,
6473 INIT_LIST_HEAD(&cc.migratepages);
6476 * What we do here is we mark all pageblocks in range as
6477 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6478 * have different sizes, and due to the way page allocator
6479 * work, start_isolate_page_range() has special handlings for this.
6481 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6482 * migrate the pages from an unaligned range (ie. pages that
6483 * we are interested in). This will put all the pages in
6484 * range back to page allocator as MIGRATE_ISOLATE.
6486 * When this is done, we take the pages in range from page
6487 * allocator removing them from the buddy system. This way
6488 * page allocator will never consider using them.
6490 * This lets us mark the pageblocks back as
6491 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6492 * aligned range but not in the unaligned, original range are
6493 * put back to page allocator so that buddy can use them.
6496 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6500 drain_all_pages(cc.zone);
6503 * In case of -EBUSY, we'd like to know which page causes problem.
6504 * So, just fall through. test_pages_isolated() has a tracepoint
6505 * which will report the busy page.
6507 * It is possible that busy pages could become available before
6508 * the call to test_pages_isolated, and the range will actually be
6509 * allocated. So, if we fall through be sure to clear ret so that
6510 * -EBUSY is not accidentally used or returned to caller.
6512 ret = __alloc_contig_migrate_range(&cc, start, end, migratetype);
6513 if (ret && ret != -EBUSY)
6518 * Pages from [start, end) are within a pageblock_nr_pages
6519 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6520 * more, all pages in [start, end) are free in page allocator.
6521 * What we are going to do is to allocate all pages from
6522 * [start, end) (that is remove them from page allocator).
6524 * The only problem is that pages at the beginning and at the
6525 * end of interesting range may be not aligned with pages that
6526 * page allocator holds, ie. they can be part of higher order
6527 * pages. Because of this, we reserve the bigger range and
6528 * once this is done free the pages we are not interested in.
6530 * We don't have to hold zone->lock here because the pages are
6531 * isolated thus they won't get removed from buddy.
6533 outer_start = find_large_buddy(start);
6535 /* Make sure the range is really isolated. */
6536 if (test_pages_isolated(outer_start, end, 0)) {
6541 /* Grab isolated pages from freelists. */
6542 outer_end = isolate_freepages_range(&cc, outer_start, end);
6548 /* Free head and tail (if any) */
6549 if (start != outer_start)
6550 free_contig_range(outer_start, start - outer_start);
6551 if (end != outer_end)
6552 free_contig_range(end, outer_end - end);
6555 undo_isolate_page_range(start, end, migratetype);
6558 EXPORT_SYMBOL(alloc_contig_range_noprof);
6560 static int __alloc_contig_pages(unsigned long start_pfn,
6561 unsigned long nr_pages, gfp_t gfp_mask)
6563 unsigned long end_pfn = start_pfn + nr_pages;
6565 return alloc_contig_range_noprof(start_pfn, end_pfn, MIGRATE_MOVABLE,
6569 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6570 unsigned long nr_pages)
6572 unsigned long i, end_pfn = start_pfn + nr_pages;
6575 for (i = start_pfn; i < end_pfn; i++) {
6576 page = pfn_to_online_page(i);
6580 if (page_zone(page) != z)
6583 if (PageReserved(page))
6592 static bool zone_spans_last_pfn(const struct zone *zone,
6593 unsigned long start_pfn, unsigned long nr_pages)
6595 unsigned long last_pfn = start_pfn + nr_pages - 1;
6597 return zone_spans_pfn(zone, last_pfn);
6601 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6602 * @nr_pages: Number of contiguous pages to allocate
6603 * @gfp_mask: GFP mask to limit search and used during compaction
6605 * @nodemask: Mask for other possible nodes
6607 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6608 * on an applicable zonelist to find a contiguous pfn range which can then be
6609 * tried for allocation with alloc_contig_range(). This routine is intended
6610 * for allocation requests which can not be fulfilled with the buddy allocator.
6612 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6613 * power of two, then allocated range is also guaranteed to be aligned to same
6614 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6616 * Allocated pages can be freed with free_contig_range() or by manually calling
6617 * __free_page() on each allocated page.
6619 * Return: pointer to contiguous pages on success, or NULL if not successful.
6621 struct page *alloc_contig_pages_noprof(unsigned long nr_pages, gfp_t gfp_mask,
6622 int nid, nodemask_t *nodemask)
6624 unsigned long ret, pfn, flags;
6625 struct zonelist *zonelist;
6629 zonelist = node_zonelist(nid, gfp_mask);
6630 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6631 gfp_zone(gfp_mask), nodemask) {
6632 spin_lock_irqsave(&zone->lock, flags);
6634 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6635 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6636 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6638 * We release the zone lock here because
6639 * alloc_contig_range() will also lock the zone
6640 * at some point. If there's an allocation
6641 * spinning on this lock, it may win the race
6642 * and cause alloc_contig_range() to fail...
6644 spin_unlock_irqrestore(&zone->lock, flags);
6645 ret = __alloc_contig_pages(pfn, nr_pages,
6648 return pfn_to_page(pfn);
6649 spin_lock_irqsave(&zone->lock, flags);
6653 spin_unlock_irqrestore(&zone->lock, flags);
6657 #endif /* CONFIG_CONTIG_ALLOC */
6659 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6661 unsigned long count = 0;
6663 for (; nr_pages--; pfn++) {
6664 struct page *page = pfn_to_page(pfn);
6666 count += page_count(page) != 1;
6669 WARN(count != 0, "%lu pages are still in use!\n", count);
6671 EXPORT_SYMBOL(free_contig_range);
6674 * Effectively disable pcplists for the zone by setting the high limit to 0
6675 * and draining all cpus. A concurrent page freeing on another CPU that's about
6676 * to put the page on pcplist will either finish before the drain and the page
6677 * will be drained, or observe the new high limit and skip the pcplist.
6679 * Must be paired with a call to zone_pcp_enable().
6681 void zone_pcp_disable(struct zone *zone)
6683 mutex_lock(&pcp_batch_high_lock);
6684 __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
6685 __drain_all_pages(zone, true);
6688 void zone_pcp_enable(struct zone *zone)
6690 __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
6691 zone->pageset_high_max, zone->pageset_batch);
6692 mutex_unlock(&pcp_batch_high_lock);
6695 void zone_pcp_reset(struct zone *zone)
6698 struct per_cpu_zonestat *pzstats;
6700 if (zone->per_cpu_pageset != &boot_pageset) {
6701 for_each_online_cpu(cpu) {
6702 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6703 drain_zonestat(zone, pzstats);
6705 free_percpu(zone->per_cpu_pageset);
6706 zone->per_cpu_pageset = &boot_pageset;
6707 if (zone->per_cpu_zonestats != &boot_zonestats) {
6708 free_percpu(zone->per_cpu_zonestats);
6709 zone->per_cpu_zonestats = &boot_zonestats;
6714 #ifdef CONFIG_MEMORY_HOTREMOVE
6716 * All pages in the range must be in a single zone, must not contain holes,
6717 * must span full sections, and must be isolated before calling this function.
6719 * Returns the number of managed (non-PageOffline()) pages in the range: the
6720 * number of pages for which memory offlining code must adjust managed page
6721 * counters using adjust_managed_page_count().
6723 unsigned long __offline_isolated_pages(unsigned long start_pfn,
6724 unsigned long end_pfn)
6726 unsigned long already_offline = 0, flags;
6727 unsigned long pfn = start_pfn;
6732 offline_mem_sections(pfn, end_pfn);
6733 zone = page_zone(pfn_to_page(pfn));
6734 spin_lock_irqsave(&zone->lock, flags);
6735 while (pfn < end_pfn) {
6736 page = pfn_to_page(pfn);
6738 * The HWPoisoned page may be not in buddy system, and
6739 * page_count() is not 0.
6741 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6746 * At this point all remaining PageOffline() pages have a
6747 * reference count of 0 and can simply be skipped.
6749 if (PageOffline(page)) {
6750 BUG_ON(page_count(page));
6751 BUG_ON(PageBuddy(page));
6757 BUG_ON(page_count(page));
6758 BUG_ON(!PageBuddy(page));
6759 VM_WARN_ON(get_pageblock_migratetype(page) != MIGRATE_ISOLATE);
6760 order = buddy_order(page);
6761 del_page_from_free_list(page, zone, order, MIGRATE_ISOLATE);
6762 pfn += (1 << order);
6764 spin_unlock_irqrestore(&zone->lock, flags);
6766 return end_pfn - start_pfn - already_offline;
6771 * This function returns a stable result only if called under zone lock.
6773 bool is_free_buddy_page(const struct page *page)
6775 unsigned long pfn = page_to_pfn(page);
6778 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6779 const struct page *head = page - (pfn & ((1 << order) - 1));
6781 if (PageBuddy(head) &&
6782 buddy_order_unsafe(head) >= order)
6786 return order <= MAX_PAGE_ORDER;
6788 EXPORT_SYMBOL(is_free_buddy_page);
6790 #ifdef CONFIG_MEMORY_FAILURE
6791 static inline void add_to_free_list(struct page *page, struct zone *zone,
6792 unsigned int order, int migratetype,
6795 __add_to_free_list(page, zone, order, migratetype, tail);
6796 account_freepages(zone, 1 << order, migratetype);
6800 * Break down a higher-order page in sub-pages, and keep our target out of
6803 static void break_down_buddy_pages(struct zone *zone, struct page *page,
6804 struct page *target, int low, int high,
6807 unsigned long size = 1 << high;
6808 struct page *current_buddy;
6810 while (high > low) {
6814 if (target >= &page[size]) {
6815 current_buddy = page;
6818 current_buddy = page + size;
6821 if (set_page_guard(zone, current_buddy, high))
6824 add_to_free_list(current_buddy, zone, high, migratetype, false);
6825 set_buddy_order(current_buddy, high);
6830 * Take a page that will be marked as poisoned off the buddy allocator.
6832 bool take_page_off_buddy(struct page *page)
6834 struct zone *zone = page_zone(page);
6835 unsigned long pfn = page_to_pfn(page);
6836 unsigned long flags;
6840 spin_lock_irqsave(&zone->lock, flags);
6841 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6842 struct page *page_head = page - (pfn & ((1 << order) - 1));
6843 int page_order = buddy_order(page_head);
6845 if (PageBuddy(page_head) && page_order >= order) {
6846 unsigned long pfn_head = page_to_pfn(page_head);
6847 int migratetype = get_pfnblock_migratetype(page_head,
6850 del_page_from_free_list(page_head, zone, page_order,
6852 break_down_buddy_pages(zone, page_head, page, 0,
6853 page_order, migratetype);
6854 SetPageHWPoisonTakenOff(page);
6858 if (page_count(page_head) > 0)
6861 spin_unlock_irqrestore(&zone->lock, flags);
6866 * Cancel takeoff done by take_page_off_buddy().
6868 bool put_page_back_buddy(struct page *page)
6870 struct zone *zone = page_zone(page);
6871 unsigned long flags;
6874 spin_lock_irqsave(&zone->lock, flags);
6875 if (put_page_testzero(page)) {
6876 unsigned long pfn = page_to_pfn(page);
6877 int migratetype = get_pfnblock_migratetype(page, pfn);
6879 ClearPageHWPoisonTakenOff(page);
6880 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6881 if (TestClearPageHWPoison(page)) {
6885 spin_unlock_irqrestore(&zone->lock, flags);
6891 #ifdef CONFIG_ZONE_DMA
6892 bool has_managed_dma(void)
6894 struct pglist_data *pgdat;
6896 for_each_online_pgdat(pgdat) {
6897 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6899 if (managed_zone(zone))
6904 #endif /* CONFIG_ZONE_DMA */
6906 #ifdef CONFIG_UNACCEPTED_MEMORY
6908 /* Counts number of zones with unaccepted pages. */
6909 static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6911 static bool lazy_accept = true;
6913 static int __init accept_memory_parse(char *p)
6915 if (!strcmp(p, "lazy")) {
6918 } else if (!strcmp(p, "eager")) {
6919 lazy_accept = false;
6925 early_param("accept_memory", accept_memory_parse);
6927 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6929 phys_addr_t start = page_to_phys(page);
6930 phys_addr_t end = start + (PAGE_SIZE << order);
6932 return range_contains_unaccepted_memory(start, end);
6935 static void accept_page(struct page *page, unsigned int order)
6937 phys_addr_t start = page_to_phys(page);
6939 accept_memory(start, start + (PAGE_SIZE << order));
6942 static bool try_to_accept_memory_one(struct zone *zone)
6944 unsigned long flags;
6948 spin_lock_irqsave(&zone->lock, flags);
6949 page = list_first_entry_or_null(&zone->unaccepted_pages,
6952 spin_unlock_irqrestore(&zone->lock, flags);
6956 list_del(&page->lru);
6957 last = list_empty(&zone->unaccepted_pages);
6959 account_freepages(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6960 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6961 spin_unlock_irqrestore(&zone->lock, flags);
6963 accept_page(page, MAX_PAGE_ORDER);
6965 __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
6968 static_branch_dec(&zones_with_unaccepted_pages);
6973 static bool cond_accept_memory(struct zone *zone, unsigned int order)
6978 if (!has_unaccepted_memory())
6981 if (list_empty(&zone->unaccepted_pages))
6984 /* How much to accept to get to high watermark? */
6985 to_accept = high_wmark_pages(zone) -
6986 (zone_page_state(zone, NR_FREE_PAGES) -
6987 __zone_watermark_unusable_free(zone, order, 0) -
6988 zone_page_state(zone, NR_UNACCEPTED));
6990 while (to_accept > 0) {
6991 if (!try_to_accept_memory_one(zone))
6994 to_accept -= MAX_ORDER_NR_PAGES;
7000 static inline bool has_unaccepted_memory(void)
7002 return static_branch_unlikely(&zones_with_unaccepted_pages);
7005 static bool __free_unaccepted(struct page *page)
7007 struct zone *zone = page_zone(page);
7008 unsigned long flags;
7014 spin_lock_irqsave(&zone->lock, flags);
7015 first = list_empty(&zone->unaccepted_pages);
7016 list_add_tail(&page->lru, &zone->unaccepted_pages);
7017 account_freepages(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
7018 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
7019 spin_unlock_irqrestore(&zone->lock, flags);
7022 static_branch_inc(&zones_with_unaccepted_pages);
7029 static bool page_contains_unaccepted(struct page *page, unsigned int order)
7034 static void accept_page(struct page *page, unsigned int order)
7038 static bool cond_accept_memory(struct zone *zone, unsigned int order)
7043 static inline bool has_unaccepted_memory(void)
7048 static bool __free_unaccepted(struct page *page)
7054 #endif /* CONFIG_UNACCEPTED_MEMORY */