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/swap.h>
22 #include <linux/swapops.h>
23 #include <linux/interrupt.h>
24 #include <linux/pagemap.h>
25 #include <linux/jiffies.h>
26 #include <linux/memblock.h>
27 #include <linux/compiler.h>
28 #include <linux/kernel.h>
29 #include <linux/kasan.h>
30 #include <linux/kmsan.h>
31 #include <linux/module.h>
32 #include <linux/suspend.h>
33 #include <linux/pagevec.h>
34 #include <linux/blkdev.h>
35 #include <linux/slab.h>
36 #include <linux/ratelimit.h>
37 #include <linux/oom.h>
38 #include <linux/topology.h>
39 #include <linux/sysctl.h>
40 #include <linux/cpu.h>
41 #include <linux/cpuset.h>
42 #include <linux/memory_hotplug.h>
43 #include <linux/nodemask.h>
44 #include <linux/vmalloc.h>
45 #include <linux/vmstat.h>
46 #include <linux/mempolicy.h>
47 #include <linux/memremap.h>
48 #include <linux/stop_machine.h>
49 #include <linux/random.h>
50 #include <linux/sort.h>
51 #include <linux/pfn.h>
52 #include <linux/backing-dev.h>
53 #include <linux/fault-inject.h>
54 #include <linux/page-isolation.h>
55 #include <linux/debugobjects.h>
56 #include <linux/kmemleak.h>
57 #include <linux/compaction.h>
58 #include <trace/events/kmem.h>
59 #include <trace/events/oom.h>
60 #include <linux/prefetch.h>
61 #include <linux/mm_inline.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/migrate.h>
64 #include <linux/hugetlb.h>
65 #include <linux/sched/rt.h>
66 #include <linux/sched/mm.h>
67 #include <linux/page_owner.h>
68 #include <linux/page_table_check.h>
69 #include <linux/kthread.h>
70 #include <linux/memcontrol.h>
71 #include <linux/ftrace.h>
72 #include <linux/lockdep.h>
73 #include <linux/nmi.h>
74 #include <linux/psi.h>
75 #include <linux/khugepaged.h>
76 #include <linux/delayacct.h>
77 #include <asm/sections.h>
78 #include <asm/tlbflush.h>
79 #include <asm/div64.h>
82 #include "page_reporting.h"
85 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
86 typedef int __bitwise fpi_t;
88 /* No special request */
89 #define FPI_NONE ((__force fpi_t)0)
92 * Skip free page reporting notification for the (possibly merged) page.
93 * This does not hinder free page reporting from grabbing the page,
94 * reporting it and marking it "reported" - it only skips notifying
95 * the free page reporting infrastructure about a newly freed page. For
96 * example, used when temporarily pulling a page from a freelist and
97 * putting it back unmodified.
99 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
102 * Place the (possibly merged) page to the tail of the freelist. Will ignore
103 * page shuffling (relevant code - e.g., memory onlining - is expected to
104 * shuffle the whole zone).
106 * Note: No code should rely on this flag for correctness - it's purely
107 * to allow for optimizations when handing back either fresh pages
108 * (memory onlining) or untouched pages (page isolation, free page
111 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
113 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
114 static DEFINE_MUTEX(pcp_batch_high_lock);
115 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
117 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
119 * On SMP, spin_trylock is sufficient protection.
120 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
122 #define pcp_trylock_prepare(flags) do { } while (0)
123 #define pcp_trylock_finish(flag) do { } while (0)
126 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
127 #define pcp_trylock_prepare(flags) local_irq_save(flags)
128 #define pcp_trylock_finish(flags) local_irq_restore(flags)
132 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
133 * a migration causing the wrong PCP to be locked and remote memory being
134 * potentially allocated, pin the task to the CPU for the lookup+lock.
135 * preempt_disable is used on !RT because it is faster than migrate_disable.
136 * migrate_disable is used on RT because otherwise RT spinlock usage is
137 * interfered with and a high priority task cannot preempt the allocator.
139 #ifndef CONFIG_PREEMPT_RT
140 #define pcpu_task_pin() preempt_disable()
141 #define pcpu_task_unpin() preempt_enable()
143 #define pcpu_task_pin() migrate_disable()
144 #define pcpu_task_unpin() migrate_enable()
148 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
149 * Return value should be used with equivalent unlock helper.
151 #define pcpu_spin_lock(type, member, ptr) \
155 _ret = this_cpu_ptr(ptr); \
156 spin_lock(&_ret->member); \
160 #define pcpu_spin_trylock(type, member, ptr) \
164 _ret = this_cpu_ptr(ptr); \
165 if (!spin_trylock(&_ret->member)) { \
172 #define pcpu_spin_unlock(member, ptr) \
174 spin_unlock(&ptr->member); \
178 /* struct per_cpu_pages specific helpers. */
179 #define pcp_spin_lock(ptr) \
180 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
182 #define pcp_spin_trylock(ptr) \
183 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
185 #define pcp_spin_unlock(ptr) \
186 pcpu_spin_unlock(lock, ptr)
188 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
189 DEFINE_PER_CPU(int, numa_node);
190 EXPORT_PER_CPU_SYMBOL(numa_node);
193 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
195 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
197 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
198 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
199 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
200 * defined in <linux/topology.h>.
202 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
203 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
206 static DEFINE_MUTEX(pcpu_drain_mutex);
208 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
209 volatile unsigned long latent_entropy __latent_entropy;
210 EXPORT_SYMBOL(latent_entropy);
214 * Array of node states.
216 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
217 [N_POSSIBLE] = NODE_MASK_ALL,
218 [N_ONLINE] = { { [0] = 1UL } },
220 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
221 #ifdef CONFIG_HIGHMEM
222 [N_HIGH_MEMORY] = { { [0] = 1UL } },
224 [N_MEMORY] = { { [0] = 1UL } },
225 [N_CPU] = { { [0] = 1UL } },
228 EXPORT_SYMBOL(node_states);
230 atomic_long_t _totalram_pages __read_mostly;
231 EXPORT_SYMBOL(_totalram_pages);
232 unsigned long totalreserve_pages __read_mostly;
233 unsigned long totalcma_pages __read_mostly;
235 int percpu_pagelist_high_fraction;
236 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
237 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
238 EXPORT_SYMBOL(init_on_alloc);
240 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
241 EXPORT_SYMBOL(init_on_free);
243 /* perform sanity checks on struct pages being allocated or freed */
244 static DEFINE_STATIC_KEY_MAYBE(CONFIG_DEBUG_VM, check_pages_enabled);
246 static inline bool is_check_pages_enabled(void)
248 return static_branch_unlikely(&check_pages_enabled);
251 static bool _init_on_alloc_enabled_early __read_mostly
252 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
253 static int __init early_init_on_alloc(char *buf)
256 return kstrtobool(buf, &_init_on_alloc_enabled_early);
258 early_param("init_on_alloc", early_init_on_alloc);
260 static bool _init_on_free_enabled_early __read_mostly
261 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
262 static int __init early_init_on_free(char *buf)
264 return kstrtobool(buf, &_init_on_free_enabled_early);
266 early_param("init_on_free", early_init_on_free);
269 * A cached value of the page's pageblock's migratetype, used when the page is
270 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
271 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
272 * Also the migratetype set in the page does not necessarily match the pcplist
273 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
274 * other index - this ensures that it will be put on the correct CMA freelist.
276 static inline int get_pcppage_migratetype(struct page *page)
281 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
283 page->index = migratetype;
286 #ifdef CONFIG_PM_SLEEP
288 * The following functions are used by the suspend/hibernate code to temporarily
289 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
290 * while devices are suspended. To avoid races with the suspend/hibernate code,
291 * they should always be called with system_transition_mutex held
292 * (gfp_allowed_mask also should only be modified with system_transition_mutex
293 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
294 * with that modification).
297 static gfp_t saved_gfp_mask;
299 void pm_restore_gfp_mask(void)
301 WARN_ON(!mutex_is_locked(&system_transition_mutex));
302 if (saved_gfp_mask) {
303 gfp_allowed_mask = saved_gfp_mask;
308 void pm_restrict_gfp_mask(void)
310 WARN_ON(!mutex_is_locked(&system_transition_mutex));
311 WARN_ON(saved_gfp_mask);
312 saved_gfp_mask = gfp_allowed_mask;
313 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
316 bool pm_suspended_storage(void)
318 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
322 #endif /* CONFIG_PM_SLEEP */
324 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
325 unsigned int pageblock_order __read_mostly;
328 static void __free_pages_ok(struct page *page, unsigned int order,
332 * results with 256, 32 in the lowmem_reserve sysctl:
333 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
334 * 1G machine -> (16M dma, 784M normal, 224M high)
335 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
336 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
337 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
339 * TBD: should special case ZONE_DMA32 machines here - in those we normally
340 * don't need any ZONE_NORMAL reservation
342 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
343 #ifdef CONFIG_ZONE_DMA
346 #ifdef CONFIG_ZONE_DMA32
350 #ifdef CONFIG_HIGHMEM
356 char * const zone_names[MAX_NR_ZONES] = {
357 #ifdef CONFIG_ZONE_DMA
360 #ifdef CONFIG_ZONE_DMA32
364 #ifdef CONFIG_HIGHMEM
368 #ifdef CONFIG_ZONE_DEVICE
373 const char * const migratetype_names[MIGRATE_TYPES] = {
381 #ifdef CONFIG_MEMORY_ISOLATION
386 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
387 [NULL_COMPOUND_DTOR] = NULL,
388 [COMPOUND_PAGE_DTOR] = free_compound_page,
389 #ifdef CONFIG_HUGETLB_PAGE
390 [HUGETLB_PAGE_DTOR] = free_huge_page,
392 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
393 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
397 int min_free_kbytes = 1024;
398 int user_min_free_kbytes = -1;
399 int watermark_boost_factor __read_mostly = 15000;
400 int watermark_scale_factor = 10;
402 bool mirrored_kernelcore __initdata_memblock;
404 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
406 EXPORT_SYMBOL(movable_zone);
409 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
410 unsigned int nr_online_nodes __read_mostly = 1;
411 EXPORT_SYMBOL(nr_node_ids);
412 EXPORT_SYMBOL(nr_online_nodes);
415 int page_group_by_mobility_disabled __read_mostly;
417 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
419 * During boot we initialize deferred pages on-demand, as needed, but once
420 * page_alloc_init_late() has finished, the deferred pages are all initialized,
421 * and we can permanently disable that path.
423 DEFINE_STATIC_KEY_TRUE(deferred_pages);
425 static inline bool deferred_pages_enabled(void)
427 return static_branch_unlikely(&deferred_pages);
431 * deferred_grow_zone() is __init, but it is called from
432 * get_page_from_freelist() during early boot until deferred_pages permanently
433 * disables this call. This is why we have refdata wrapper to avoid warning,
434 * and to ensure that the function body gets unloaded.
437 _deferred_grow_zone(struct zone *zone, unsigned int order)
439 return deferred_grow_zone(zone, order);
442 static inline bool deferred_pages_enabled(void)
446 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
448 /* Return a pointer to the bitmap storing bits affecting a block of pages */
449 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
452 #ifdef CONFIG_SPARSEMEM
453 return section_to_usemap(__pfn_to_section(pfn));
455 return page_zone(page)->pageblock_flags;
456 #endif /* CONFIG_SPARSEMEM */
459 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
461 #ifdef CONFIG_SPARSEMEM
462 pfn &= (PAGES_PER_SECTION-1);
464 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
465 #endif /* CONFIG_SPARSEMEM */
466 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
469 static __always_inline
470 unsigned long __get_pfnblock_flags_mask(const struct page *page,
474 unsigned long *bitmap;
475 unsigned long bitidx, word_bitidx;
478 bitmap = get_pageblock_bitmap(page, pfn);
479 bitidx = pfn_to_bitidx(page, pfn);
480 word_bitidx = bitidx / BITS_PER_LONG;
481 bitidx &= (BITS_PER_LONG-1);
483 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
484 * a consistent read of the memory array, so that results, even though
485 * racy, are not corrupted.
487 word = READ_ONCE(bitmap[word_bitidx]);
488 return (word >> bitidx) & mask;
492 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
493 * @page: The page within the block of interest
494 * @pfn: The target page frame number
495 * @mask: mask of bits that the caller is interested in
497 * Return: pageblock_bits flags
499 unsigned long get_pfnblock_flags_mask(const struct page *page,
500 unsigned long pfn, unsigned long mask)
502 return __get_pfnblock_flags_mask(page, pfn, mask);
505 static __always_inline int get_pfnblock_migratetype(const struct page *page,
508 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
512 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
513 * @page: The page within the block of interest
514 * @flags: The flags to set
515 * @pfn: The target page frame number
516 * @mask: mask of bits that the caller is interested in
518 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
522 unsigned long *bitmap;
523 unsigned long bitidx, word_bitidx;
526 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
527 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
529 bitmap = get_pageblock_bitmap(page, pfn);
530 bitidx = pfn_to_bitidx(page, pfn);
531 word_bitidx = bitidx / BITS_PER_LONG;
532 bitidx &= (BITS_PER_LONG-1);
534 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
539 word = READ_ONCE(bitmap[word_bitidx]);
541 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
544 void set_pageblock_migratetype(struct page *page, int migratetype)
546 if (unlikely(page_group_by_mobility_disabled &&
547 migratetype < MIGRATE_PCPTYPES))
548 migratetype = MIGRATE_UNMOVABLE;
550 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
551 page_to_pfn(page), MIGRATETYPE_MASK);
554 #ifdef CONFIG_DEBUG_VM
555 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
559 unsigned long pfn = page_to_pfn(page);
560 unsigned long sp, start_pfn;
563 seq = zone_span_seqbegin(zone);
564 start_pfn = zone->zone_start_pfn;
565 sp = zone->spanned_pages;
566 if (!zone_spans_pfn(zone, pfn))
568 } while (zone_span_seqretry(zone, seq));
571 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
572 pfn, zone_to_nid(zone), zone->name,
573 start_pfn, start_pfn + sp);
578 static int page_is_consistent(struct zone *zone, struct page *page)
580 if (zone != page_zone(page))
586 * Temporary debugging check for pages not lying within a given zone.
588 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
590 if (page_outside_zone_boundaries(zone, page))
592 if (!page_is_consistent(zone, page))
598 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
604 static void bad_page(struct page *page, const char *reason)
606 static unsigned long resume;
607 static unsigned long nr_shown;
608 static unsigned long nr_unshown;
611 * Allow a burst of 60 reports, then keep quiet for that minute;
612 * or allow a steady drip of one report per second.
614 if (nr_shown == 60) {
615 if (time_before(jiffies, resume)) {
621 "BUG: Bad page state: %lu messages suppressed\n",
628 resume = jiffies + 60 * HZ;
630 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
631 current->comm, page_to_pfn(page));
632 dump_page(page, reason);
637 /* Leave bad fields for debug, except PageBuddy could make trouble */
638 page_mapcount_reset(page); /* remove PageBuddy */
639 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
642 static inline unsigned int order_to_pindex(int migratetype, int order)
646 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
647 if (order > PAGE_ALLOC_COSTLY_ORDER) {
648 VM_BUG_ON(order != pageblock_order);
649 return NR_LOWORDER_PCP_LISTS;
652 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
655 return (MIGRATE_PCPTYPES * base) + migratetype;
658 static inline int pindex_to_order(unsigned int pindex)
660 int order = pindex / MIGRATE_PCPTYPES;
662 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
663 if (pindex == NR_LOWORDER_PCP_LISTS)
664 order = pageblock_order;
666 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
672 static inline bool pcp_allowed_order(unsigned int order)
674 if (order <= PAGE_ALLOC_COSTLY_ORDER)
676 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
677 if (order == pageblock_order)
683 static inline void free_the_page(struct page *page, unsigned int order)
685 if (pcp_allowed_order(order)) /* Via pcp? */
686 free_unref_page(page, order);
688 __free_pages_ok(page, order, FPI_NONE);
692 * Higher-order pages are called "compound pages". They are structured thusly:
694 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
696 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
697 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
699 * The first tail page's ->compound_dtor holds the offset in array of compound
700 * page destructors. See compound_page_dtors.
702 * The first tail page's ->compound_order holds the order of allocation.
703 * This usage means that zero-order pages may not be compound.
706 void free_compound_page(struct page *page)
708 mem_cgroup_uncharge(page_folio(page));
709 free_the_page(page, compound_order(page));
712 void prep_compound_page(struct page *page, unsigned int order)
715 int nr_pages = 1 << order;
718 for (i = 1; i < nr_pages; i++)
719 prep_compound_tail(page, i);
721 prep_compound_head(page, order);
724 void destroy_large_folio(struct folio *folio)
726 enum compound_dtor_id dtor = folio->_folio_dtor;
728 VM_BUG_ON_FOLIO(dtor >= NR_COMPOUND_DTORS, folio);
729 compound_page_dtors[dtor](&folio->page);
732 #ifdef CONFIG_DEBUG_PAGEALLOC
733 unsigned int _debug_guardpage_minorder;
735 bool _debug_pagealloc_enabled_early __read_mostly
736 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
737 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
738 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
739 EXPORT_SYMBOL(_debug_pagealloc_enabled);
741 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
743 static int __init early_debug_pagealloc(char *buf)
745 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
747 early_param("debug_pagealloc", early_debug_pagealloc);
749 static int __init debug_guardpage_minorder_setup(char *buf)
753 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
754 pr_err("Bad debug_guardpage_minorder value\n");
757 _debug_guardpage_minorder = res;
758 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
761 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
763 static inline bool set_page_guard(struct zone *zone, struct page *page,
764 unsigned int order, int migratetype)
766 if (!debug_guardpage_enabled())
769 if (order >= debug_guardpage_minorder())
772 __SetPageGuard(page);
773 INIT_LIST_HEAD(&page->buddy_list);
774 set_page_private(page, order);
775 /* Guard pages are not available for any usage */
776 if (!is_migrate_isolate(migratetype))
777 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
782 static inline void clear_page_guard(struct zone *zone, struct page *page,
783 unsigned int order, int migratetype)
785 if (!debug_guardpage_enabled())
788 __ClearPageGuard(page);
790 set_page_private(page, 0);
791 if (!is_migrate_isolate(migratetype))
792 __mod_zone_freepage_state(zone, (1 << order), migratetype);
795 static inline bool set_page_guard(struct zone *zone, struct page *page,
796 unsigned int order, int migratetype) { return false; }
797 static inline void clear_page_guard(struct zone *zone, struct page *page,
798 unsigned int order, int migratetype) {}
802 * Enable static keys related to various memory debugging and hardening options.
803 * Some override others, and depend on early params that are evaluated in the
804 * order of appearance. So we need to first gather the full picture of what was
805 * enabled, and then make decisions.
807 void __init init_mem_debugging_and_hardening(void)
809 bool page_poisoning_requested = false;
810 bool want_check_pages = false;
812 #ifdef CONFIG_PAGE_POISONING
814 * Page poisoning is debug page alloc for some arches. If
815 * either of those options are enabled, enable poisoning.
817 if (page_poisoning_enabled() ||
818 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
819 debug_pagealloc_enabled())) {
820 static_branch_enable(&_page_poisoning_enabled);
821 page_poisoning_requested = true;
822 want_check_pages = true;
826 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
827 page_poisoning_requested) {
828 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
829 "will take precedence over init_on_alloc and init_on_free\n");
830 _init_on_alloc_enabled_early = false;
831 _init_on_free_enabled_early = false;
834 if (_init_on_alloc_enabled_early) {
835 want_check_pages = true;
836 static_branch_enable(&init_on_alloc);
838 static_branch_disable(&init_on_alloc);
841 if (_init_on_free_enabled_early) {
842 want_check_pages = true;
843 static_branch_enable(&init_on_free);
845 static_branch_disable(&init_on_free);
848 if (IS_ENABLED(CONFIG_KMSAN) &&
849 (_init_on_alloc_enabled_early || _init_on_free_enabled_early))
850 pr_info("mem auto-init: please make sure init_on_alloc and init_on_free are disabled when running KMSAN\n");
852 #ifdef CONFIG_DEBUG_PAGEALLOC
853 if (debug_pagealloc_enabled()) {
854 want_check_pages = true;
855 static_branch_enable(&_debug_pagealloc_enabled);
857 if (debug_guardpage_minorder())
858 static_branch_enable(&_debug_guardpage_enabled);
863 * Any page debugging or hardening option also enables sanity checking
864 * of struct pages being allocated or freed. With CONFIG_DEBUG_VM it's
867 if (!IS_ENABLED(CONFIG_DEBUG_VM) && want_check_pages)
868 static_branch_enable(&check_pages_enabled);
871 static inline void set_buddy_order(struct page *page, unsigned int order)
873 set_page_private(page, order);
874 __SetPageBuddy(page);
877 #ifdef CONFIG_COMPACTION
878 static inline struct capture_control *task_capc(struct zone *zone)
880 struct capture_control *capc = current->capture_control;
882 return unlikely(capc) &&
883 !(current->flags & PF_KTHREAD) &&
885 capc->cc->zone == zone ? capc : NULL;
889 compaction_capture(struct capture_control *capc, struct page *page,
890 int order, int migratetype)
892 if (!capc || order != capc->cc->order)
895 /* Do not accidentally pollute CMA or isolated regions*/
896 if (is_migrate_cma(migratetype) ||
897 is_migrate_isolate(migratetype))
901 * Do not let lower order allocations pollute a movable pageblock.
902 * This might let an unmovable request use a reclaimable pageblock
903 * and vice-versa but no more than normal fallback logic which can
904 * have trouble finding a high-order free page.
906 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
914 static inline struct capture_control *task_capc(struct zone *zone)
920 compaction_capture(struct capture_control *capc, struct page *page,
921 int order, int migratetype)
925 #endif /* CONFIG_COMPACTION */
927 /* Used for pages not on another list */
928 static inline void add_to_free_list(struct page *page, struct zone *zone,
929 unsigned int order, int migratetype)
931 struct free_area *area = &zone->free_area[order];
933 list_add(&page->buddy_list, &area->free_list[migratetype]);
937 /* Used for pages not on another list */
938 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
939 unsigned int order, int migratetype)
941 struct free_area *area = &zone->free_area[order];
943 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
948 * Used for pages which are on another list. Move the pages to the tail
949 * of the list - so the moved pages won't immediately be considered for
950 * allocation again (e.g., optimization for memory onlining).
952 static inline void move_to_free_list(struct page *page, struct zone *zone,
953 unsigned int order, int migratetype)
955 struct free_area *area = &zone->free_area[order];
957 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
960 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
963 /* clear reported state and update reported page count */
964 if (page_reported(page))
965 __ClearPageReported(page);
967 list_del(&page->buddy_list);
968 __ClearPageBuddy(page);
969 set_page_private(page, 0);
970 zone->free_area[order].nr_free--;
973 static inline struct page *get_page_from_free_area(struct free_area *area,
976 return list_first_entry_or_null(&area->free_list[migratetype],
981 * If this is not the largest possible page, check if the buddy
982 * of the next-highest order is free. If it is, it's possible
983 * that pages are being freed that will coalesce soon. In case,
984 * that is happening, add the free page to the tail of the list
985 * so it's less likely to be used soon and more likely to be merged
986 * as a higher order page
989 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
990 struct page *page, unsigned int order)
992 unsigned long higher_page_pfn;
993 struct page *higher_page;
995 if (order >= MAX_ORDER - 1)
998 higher_page_pfn = buddy_pfn & pfn;
999 higher_page = page + (higher_page_pfn - pfn);
1001 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
1006 * Freeing function for a buddy system allocator.
1008 * The concept of a buddy system is to maintain direct-mapped table
1009 * (containing bit values) for memory blocks of various "orders".
1010 * The bottom level table contains the map for the smallest allocatable
1011 * units of memory (here, pages), and each level above it describes
1012 * pairs of units from the levels below, hence, "buddies".
1013 * At a high level, all that happens here is marking the table entry
1014 * at the bottom level available, and propagating the changes upward
1015 * as necessary, plus some accounting needed to play nicely with other
1016 * parts of the VM system.
1017 * At each level, we keep a list of pages, which are heads of continuous
1018 * free pages of length of (1 << order) and marked with PageBuddy.
1019 * Page's order is recorded in page_private(page) field.
1020 * So when we are allocating or freeing one, we can derive the state of the
1021 * other. That is, if we allocate a small block, and both were
1022 * free, the remainder of the region must be split into blocks.
1023 * If a block is freed, and its buddy is also free, then this
1024 * triggers coalescing into a block of larger size.
1029 static inline void __free_one_page(struct page *page,
1031 struct zone *zone, unsigned int order,
1032 int migratetype, fpi_t fpi_flags)
1034 struct capture_control *capc = task_capc(zone);
1035 unsigned long buddy_pfn = 0;
1036 unsigned long combined_pfn;
1040 VM_BUG_ON(!zone_is_initialized(zone));
1041 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1043 VM_BUG_ON(migratetype == -1);
1044 if (likely(!is_migrate_isolate(migratetype)))
1045 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1047 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1048 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1050 while (order < MAX_ORDER) {
1051 if (compaction_capture(capc, page, order, migratetype)) {
1052 __mod_zone_freepage_state(zone, -(1 << order),
1057 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
1061 if (unlikely(order >= pageblock_order)) {
1063 * We want to prevent merge between freepages on pageblock
1064 * without fallbacks and normal pageblock. Without this,
1065 * pageblock isolation could cause incorrect freepage or CMA
1066 * accounting or HIGHATOMIC accounting.
1068 int buddy_mt = get_pageblock_migratetype(buddy);
1070 if (migratetype != buddy_mt
1071 && (!migratetype_is_mergeable(migratetype) ||
1072 !migratetype_is_mergeable(buddy_mt)))
1077 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1078 * merge with it and move up one order.
1080 if (page_is_guard(buddy))
1081 clear_page_guard(zone, buddy, order, migratetype);
1083 del_page_from_free_list(buddy, zone, order);
1084 combined_pfn = buddy_pfn & pfn;
1085 page = page + (combined_pfn - pfn);
1091 set_buddy_order(page, order);
1093 if (fpi_flags & FPI_TO_TAIL)
1095 else if (is_shuffle_order(order))
1096 to_tail = shuffle_pick_tail();
1098 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1101 add_to_free_list_tail(page, zone, order, migratetype);
1103 add_to_free_list(page, zone, order, migratetype);
1105 /* Notify page reporting subsystem of freed page */
1106 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1107 page_reporting_notify_free(order);
1111 * split_free_page() -- split a free page at split_pfn_offset
1112 * @free_page: the original free page
1113 * @order: the order of the page
1114 * @split_pfn_offset: split offset within the page
1116 * Return -ENOENT if the free page is changed, otherwise 0
1118 * It is used when the free page crosses two pageblocks with different migratetypes
1119 * at split_pfn_offset within the page. The split free page will be put into
1120 * separate migratetype lists afterwards. Otherwise, the function achieves
1123 int split_free_page(struct page *free_page,
1124 unsigned int order, unsigned long split_pfn_offset)
1126 struct zone *zone = page_zone(free_page);
1127 unsigned long free_page_pfn = page_to_pfn(free_page);
1129 unsigned long flags;
1130 int free_page_order;
1134 if (split_pfn_offset == 0)
1137 spin_lock_irqsave(&zone->lock, flags);
1139 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
1144 mt = get_pageblock_migratetype(free_page);
1145 if (likely(!is_migrate_isolate(mt)))
1146 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1148 del_page_from_free_list(free_page, zone, order);
1149 for (pfn = free_page_pfn;
1150 pfn < free_page_pfn + (1UL << order);) {
1151 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
1153 free_page_order = min_t(unsigned int,
1154 pfn ? __ffs(pfn) : order,
1155 __fls(split_pfn_offset));
1156 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
1158 pfn += 1UL << free_page_order;
1159 split_pfn_offset -= (1UL << free_page_order);
1160 /* we have done the first part, now switch to second part */
1161 if (split_pfn_offset == 0)
1162 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
1165 spin_unlock_irqrestore(&zone->lock, flags);
1169 * A bad page could be due to a number of fields. Instead of multiple branches,
1170 * try and check multiple fields with one check. The caller must do a detailed
1171 * check if necessary.
1173 static inline bool page_expected_state(struct page *page,
1174 unsigned long check_flags)
1176 if (unlikely(atomic_read(&page->_mapcount) != -1))
1179 if (unlikely((unsigned long)page->mapping |
1180 page_ref_count(page) |
1184 (page->flags & check_flags)))
1190 static const char *page_bad_reason(struct page *page, unsigned long flags)
1192 const char *bad_reason = NULL;
1194 if (unlikely(atomic_read(&page->_mapcount) != -1))
1195 bad_reason = "nonzero mapcount";
1196 if (unlikely(page->mapping != NULL))
1197 bad_reason = "non-NULL mapping";
1198 if (unlikely(page_ref_count(page) != 0))
1199 bad_reason = "nonzero _refcount";
1200 if (unlikely(page->flags & flags)) {
1201 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1202 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1204 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1207 if (unlikely(page->memcg_data))
1208 bad_reason = "page still charged to cgroup";
1213 static void free_page_is_bad_report(struct page *page)
1216 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1219 static inline bool free_page_is_bad(struct page *page)
1221 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1224 /* Something has gone sideways, find it */
1225 free_page_is_bad_report(page);
1229 static int free_tail_pages_check(struct page *head_page, struct page *page)
1231 struct folio *folio = (struct folio *)head_page;
1235 * We rely page->lru.next never has bit 0 set, unless the page
1236 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1238 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1240 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1244 switch (page - head_page) {
1246 /* the first tail page: these may be in place of ->mapping */
1247 if (unlikely(folio_entire_mapcount(folio))) {
1248 bad_page(page, "nonzero entire_mapcount");
1251 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
1252 bad_page(page, "nonzero nr_pages_mapped");
1255 if (unlikely(atomic_read(&folio->_pincount))) {
1256 bad_page(page, "nonzero pincount");
1262 * the second tail page: ->mapping is
1263 * deferred_list.next -- ignore value.
1267 if (page->mapping != TAIL_MAPPING) {
1268 bad_page(page, "corrupted mapping in tail page");
1273 if (unlikely(!PageTail(page))) {
1274 bad_page(page, "PageTail not set");
1277 if (unlikely(compound_head(page) != head_page)) {
1278 bad_page(page, "compound_head not consistent");
1283 page->mapping = NULL;
1284 clear_compound_head(page);
1289 * Skip KASAN memory poisoning when either:
1291 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1292 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1293 * using page tags instead (see below).
1294 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1295 * that error detection is disabled for accesses via the page address.
1297 * Pages will have match-all tags in the following circumstances:
1299 * 1. Pages are being initialized for the first time, including during deferred
1300 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1301 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1302 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1303 * 3. The allocation was excluded from being checked due to sampling,
1304 * see the call to kasan_unpoison_pages.
1306 * Poisoning pages during deferred memory init will greatly lengthen the
1307 * process and cause problem in large memory systems as the deferred pages
1308 * initialization is done with interrupt disabled.
1310 * Assuming that there will be no reference to those newly initialized
1311 * pages before they are ever allocated, this should have no effect on
1312 * KASAN memory tracking as the poison will be properly inserted at page
1313 * allocation time. The only corner case is when pages are allocated by
1314 * on-demand allocation and then freed again before the deferred pages
1315 * initialization is done, but this is not likely to happen.
1317 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1319 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1320 return deferred_pages_enabled();
1322 return page_kasan_tag(page) == 0xff;
1325 static void kernel_init_pages(struct page *page, int numpages)
1329 /* s390's use of memset() could override KASAN redzones. */
1330 kasan_disable_current();
1331 for (i = 0; i < numpages; i++)
1332 clear_highpage_kasan_tagged(page + i);
1333 kasan_enable_current();
1336 static __always_inline bool free_pages_prepare(struct page *page,
1337 unsigned int order, fpi_t fpi_flags)
1340 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1341 bool init = want_init_on_free();
1343 VM_BUG_ON_PAGE(PageTail(page), page);
1345 trace_mm_page_free(page, order);
1346 kmsan_free_page(page, order);
1348 if (unlikely(PageHWPoison(page)) && !order) {
1350 * Do not let hwpoison pages hit pcplists/buddy
1351 * Untie memcg state and reset page's owner
1353 if (memcg_kmem_online() && PageMemcgKmem(page))
1354 __memcg_kmem_uncharge_page(page, order);
1355 reset_page_owner(page, order);
1356 page_table_check_free(page, order);
1361 * Check tail pages before head page information is cleared to
1362 * avoid checking PageCompound for order-0 pages.
1364 if (unlikely(order)) {
1365 bool compound = PageCompound(page);
1368 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1371 ClearPageHasHWPoisoned(page);
1372 for (i = 1; i < (1 << order); i++) {
1374 bad += free_tail_pages_check(page, page + i);
1375 if (is_check_pages_enabled()) {
1376 if (unlikely(free_page_is_bad(page + i))) {
1381 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1384 if (PageMappingFlags(page))
1385 page->mapping = NULL;
1386 if (memcg_kmem_online() && PageMemcgKmem(page))
1387 __memcg_kmem_uncharge_page(page, order);
1388 if (is_check_pages_enabled()) {
1389 if (free_page_is_bad(page))
1395 page_cpupid_reset_last(page);
1396 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1397 reset_page_owner(page, order);
1398 page_table_check_free(page, order);
1400 if (!PageHighMem(page)) {
1401 debug_check_no_locks_freed(page_address(page),
1402 PAGE_SIZE << order);
1403 debug_check_no_obj_freed(page_address(page),
1404 PAGE_SIZE << order);
1407 kernel_poison_pages(page, 1 << order);
1410 * As memory initialization might be integrated into KASAN,
1411 * KASAN poisoning and memory initialization code must be
1412 * kept together to avoid discrepancies in behavior.
1414 * With hardware tag-based KASAN, memory tags must be set before the
1415 * page becomes unavailable via debug_pagealloc or arch_free_page.
1417 if (!skip_kasan_poison) {
1418 kasan_poison_pages(page, order, init);
1420 /* Memory is already initialized if KASAN did it internally. */
1421 if (kasan_has_integrated_init())
1425 kernel_init_pages(page, 1 << order);
1428 * arch_free_page() can make the page's contents inaccessible. s390
1429 * does this. So nothing which can access the page's contents should
1430 * happen after this.
1432 arch_free_page(page, order);
1434 debug_pagealloc_unmap_pages(page, 1 << order);
1440 * Frees a number of pages from the PCP lists
1441 * Assumes all pages on list are in same zone.
1442 * count is the number of pages to free.
1444 static void free_pcppages_bulk(struct zone *zone, int count,
1445 struct per_cpu_pages *pcp,
1448 unsigned long flags;
1450 int max_pindex = NR_PCP_LISTS - 1;
1452 bool isolated_pageblocks;
1456 * Ensure proper count is passed which otherwise would stuck in the
1457 * below while (list_empty(list)) loop.
1459 count = min(pcp->count, count);
1461 /* Ensure requested pindex is drained first. */
1462 pindex = pindex - 1;
1464 spin_lock_irqsave(&zone->lock, flags);
1465 isolated_pageblocks = has_isolate_pageblock(zone);
1468 struct list_head *list;
1471 /* Remove pages from lists in a round-robin fashion. */
1473 if (++pindex > max_pindex)
1474 pindex = min_pindex;
1475 list = &pcp->lists[pindex];
1476 if (!list_empty(list))
1479 if (pindex == max_pindex)
1481 if (pindex == min_pindex)
1485 order = pindex_to_order(pindex);
1486 nr_pages = 1 << order;
1490 page = list_last_entry(list, struct page, pcp_list);
1491 mt = get_pcppage_migratetype(page);
1493 /* must delete to avoid corrupting pcp list */
1494 list_del(&page->pcp_list);
1496 pcp->count -= nr_pages;
1498 /* MIGRATE_ISOLATE page should not go to pcplists */
1499 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1500 /* Pageblock could have been isolated meanwhile */
1501 if (unlikely(isolated_pageblocks))
1502 mt = get_pageblock_migratetype(page);
1504 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1505 trace_mm_page_pcpu_drain(page, order, mt);
1506 } while (count > 0 && !list_empty(list));
1509 spin_unlock_irqrestore(&zone->lock, flags);
1512 static void free_one_page(struct zone *zone,
1513 struct page *page, unsigned long pfn,
1515 int migratetype, fpi_t fpi_flags)
1517 unsigned long flags;
1519 spin_lock_irqsave(&zone->lock, flags);
1520 if (unlikely(has_isolate_pageblock(zone) ||
1521 is_migrate_isolate(migratetype))) {
1522 migratetype = get_pfnblock_migratetype(page, pfn);
1524 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1525 spin_unlock_irqrestore(&zone->lock, flags);
1528 static void __free_pages_ok(struct page *page, unsigned int order,
1531 unsigned long flags;
1533 unsigned long pfn = page_to_pfn(page);
1534 struct zone *zone = page_zone(page);
1536 if (!free_pages_prepare(page, order, fpi_flags))
1540 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1541 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1542 * This will reduce the lock holding time.
1544 migratetype = get_pfnblock_migratetype(page, pfn);
1546 spin_lock_irqsave(&zone->lock, flags);
1547 if (unlikely(has_isolate_pageblock(zone) ||
1548 is_migrate_isolate(migratetype))) {
1549 migratetype = get_pfnblock_migratetype(page, pfn);
1551 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1552 spin_unlock_irqrestore(&zone->lock, flags);
1554 __count_vm_events(PGFREE, 1 << order);
1557 void __free_pages_core(struct page *page, unsigned int order)
1559 unsigned int nr_pages = 1 << order;
1560 struct page *p = page;
1564 * When initializing the memmap, __init_single_page() sets the refcount
1565 * of all pages to 1 ("allocated"/"not free"). We have to set the
1566 * refcount of all involved pages to 0.
1569 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1571 __ClearPageReserved(p);
1572 set_page_count(p, 0);
1574 __ClearPageReserved(p);
1575 set_page_count(p, 0);
1577 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1580 * Bypass PCP and place fresh pages right to the tail, primarily
1581 * relevant for memory onlining.
1583 __free_pages_ok(page, order, FPI_TO_TAIL);
1587 * Check that the whole (or subset of) a pageblock given by the interval of
1588 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1589 * with the migration of free compaction scanner.
1591 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1593 * It's possible on some configurations to have a setup like node0 node1 node0
1594 * i.e. it's possible that all pages within a zones range of pages do not
1595 * belong to a single zone. We assume that a border between node0 and node1
1596 * can occur within a single pageblock, but not a node0 node1 node0
1597 * interleaving within a single pageblock. It is therefore sufficient to check
1598 * the first and last page of a pageblock and avoid checking each individual
1599 * page in a pageblock.
1601 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1602 unsigned long end_pfn, struct zone *zone)
1604 struct page *start_page;
1605 struct page *end_page;
1607 /* end_pfn is one past the range we are checking */
1610 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1613 start_page = pfn_to_online_page(start_pfn);
1617 if (page_zone(start_page) != zone)
1620 end_page = pfn_to_page(end_pfn);
1622 /* This gives a shorter code than deriving page_zone(end_page) */
1623 if (page_zone_id(start_page) != page_zone_id(end_page))
1629 void set_zone_contiguous(struct zone *zone)
1631 unsigned long block_start_pfn = zone->zone_start_pfn;
1632 unsigned long block_end_pfn;
1634 block_end_pfn = pageblock_end_pfn(block_start_pfn);
1635 for (; block_start_pfn < zone_end_pfn(zone);
1636 block_start_pfn = block_end_pfn,
1637 block_end_pfn += pageblock_nr_pages) {
1639 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1641 if (!__pageblock_pfn_to_page(block_start_pfn,
1642 block_end_pfn, zone))
1647 /* We confirm that there is no hole */
1648 zone->contiguous = true;
1651 void clear_zone_contiguous(struct zone *zone)
1653 zone->contiguous = false;
1657 * The order of subdivision here is critical for the IO subsystem.
1658 * Please do not alter this order without good reasons and regression
1659 * testing. Specifically, as large blocks of memory are subdivided,
1660 * the order in which smaller blocks are delivered depends on the order
1661 * they're subdivided in this function. This is the primary factor
1662 * influencing the order in which pages are delivered to the IO
1663 * subsystem according to empirical testing, and this is also justified
1664 * by considering the behavior of a buddy system containing a single
1665 * large block of memory acted on by a series of small allocations.
1666 * This behavior is a critical factor in sglist merging's success.
1670 static inline void expand(struct zone *zone, struct page *page,
1671 int low, int high, int migratetype)
1673 unsigned long size = 1 << high;
1675 while (high > low) {
1678 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1681 * Mark as guard pages (or page), that will allow to
1682 * merge back to allocator when buddy will be freed.
1683 * Corresponding page table entries will not be touched,
1684 * pages will stay not present in virtual address space
1686 if (set_page_guard(zone, &page[size], high, migratetype))
1689 add_to_free_list(&page[size], zone, high, migratetype);
1690 set_buddy_order(&page[size], high);
1694 static void check_new_page_bad(struct page *page)
1696 if (unlikely(page->flags & __PG_HWPOISON)) {
1697 /* Don't complain about hwpoisoned pages */
1698 page_mapcount_reset(page); /* remove PageBuddy */
1703 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1707 * This page is about to be returned from the page allocator
1709 static int check_new_page(struct page *page)
1711 if (likely(page_expected_state(page,
1712 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1715 check_new_page_bad(page);
1719 static inline bool check_new_pages(struct page *page, unsigned int order)
1721 if (is_check_pages_enabled()) {
1722 for (int i = 0; i < (1 << order); i++) {
1723 struct page *p = page + i;
1725 if (unlikely(check_new_page(p)))
1733 static inline bool should_skip_kasan_unpoison(gfp_t flags)
1735 /* Don't skip if a software KASAN mode is enabled. */
1736 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1737 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1740 /* Skip, if hardware tag-based KASAN is not enabled. */
1741 if (!kasan_hw_tags_enabled())
1745 * With hardware tag-based KASAN enabled, skip if this has been
1746 * requested via __GFP_SKIP_KASAN.
1748 return flags & __GFP_SKIP_KASAN;
1751 static inline bool should_skip_init(gfp_t flags)
1753 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1754 if (!kasan_hw_tags_enabled())
1757 /* For hardware tag-based KASAN, skip if requested. */
1758 return (flags & __GFP_SKIP_ZERO);
1761 inline void post_alloc_hook(struct page *page, unsigned int order,
1764 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1765 !should_skip_init(gfp_flags);
1766 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1769 set_page_private(page, 0);
1770 set_page_refcounted(page);
1772 arch_alloc_page(page, order);
1773 debug_pagealloc_map_pages(page, 1 << order);
1776 * Page unpoisoning must happen before memory initialization.
1777 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1778 * allocations and the page unpoisoning code will complain.
1780 kernel_unpoison_pages(page, 1 << order);
1783 * As memory initialization might be integrated into KASAN,
1784 * KASAN unpoisoning and memory initializion code must be
1785 * kept together to avoid discrepancies in behavior.
1789 * If memory tags should be zeroed
1790 * (which happens only when memory should be initialized as well).
1793 /* Initialize both memory and memory tags. */
1794 for (i = 0; i != 1 << order; ++i)
1795 tag_clear_highpage(page + i);
1797 /* Take note that memory was initialized by the loop above. */
1800 if (!should_skip_kasan_unpoison(gfp_flags) &&
1801 kasan_unpoison_pages(page, order, init)) {
1802 /* Take note that memory was initialized by KASAN. */
1803 if (kasan_has_integrated_init())
1807 * If memory tags have not been set by KASAN, reset the page
1808 * tags to ensure page_address() dereferencing does not fault.
1810 for (i = 0; i != 1 << order; ++i)
1811 page_kasan_tag_reset(page + i);
1813 /* If memory is still not initialized, initialize it now. */
1815 kernel_init_pages(page, 1 << order);
1817 set_page_owner(page, order, gfp_flags);
1818 page_table_check_alloc(page, order);
1821 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1822 unsigned int alloc_flags)
1824 post_alloc_hook(page, order, gfp_flags);
1826 if (order && (gfp_flags & __GFP_COMP))
1827 prep_compound_page(page, order);
1830 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1831 * allocate the page. The expectation is that the caller is taking
1832 * steps that will free more memory. The caller should avoid the page
1833 * being used for !PFMEMALLOC purposes.
1835 if (alloc_flags & ALLOC_NO_WATERMARKS)
1836 set_page_pfmemalloc(page);
1838 clear_page_pfmemalloc(page);
1842 * Go through the free lists for the given migratetype and remove
1843 * the smallest available page from the freelists
1845 static __always_inline
1846 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1849 unsigned int current_order;
1850 struct free_area *area;
1853 /* Find a page of the appropriate size in the preferred list */
1854 for (current_order = order; current_order <= MAX_ORDER; ++current_order) {
1855 area = &(zone->free_area[current_order]);
1856 page = get_page_from_free_area(area, migratetype);
1859 del_page_from_free_list(page, zone, current_order);
1860 expand(zone, page, order, current_order, migratetype);
1861 set_pcppage_migratetype(page, migratetype);
1862 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1863 pcp_allowed_order(order) &&
1864 migratetype < MIGRATE_PCPTYPES);
1873 * This array describes the order lists are fallen back to when
1874 * the free lists for the desirable migrate type are depleted
1876 * The other migratetypes do not have fallbacks.
1878 static int fallbacks[MIGRATE_TYPES][MIGRATE_PCPTYPES - 1] = {
1879 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1880 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1881 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1885 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1888 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1891 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1892 unsigned int order) { return NULL; }
1896 * Move the free pages in a range to the freelist tail of the requested type.
1897 * Note that start_page and end_pages are not aligned on a pageblock
1898 * boundary. If alignment is required, use move_freepages_block()
1900 static int move_freepages(struct zone *zone,
1901 unsigned long start_pfn, unsigned long end_pfn,
1902 int migratetype, int *num_movable)
1907 int pages_moved = 0;
1909 for (pfn = start_pfn; pfn <= end_pfn;) {
1910 page = pfn_to_page(pfn);
1911 if (!PageBuddy(page)) {
1913 * We assume that pages that could be isolated for
1914 * migration are movable. But we don't actually try
1915 * isolating, as that would be expensive.
1918 (PageLRU(page) || __PageMovable(page)))
1924 /* Make sure we are not inadvertently changing nodes */
1925 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1926 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1928 order = buddy_order(page);
1929 move_to_free_list(page, zone, order, migratetype);
1931 pages_moved += 1 << order;
1937 int move_freepages_block(struct zone *zone, struct page *page,
1938 int migratetype, int *num_movable)
1940 unsigned long start_pfn, end_pfn, pfn;
1945 pfn = page_to_pfn(page);
1946 start_pfn = pageblock_start_pfn(pfn);
1947 end_pfn = pageblock_end_pfn(pfn) - 1;
1949 /* Do not cross zone boundaries */
1950 if (!zone_spans_pfn(zone, start_pfn))
1952 if (!zone_spans_pfn(zone, end_pfn))
1955 return move_freepages(zone, start_pfn, end_pfn, migratetype,
1959 static void change_pageblock_range(struct page *pageblock_page,
1960 int start_order, int migratetype)
1962 int nr_pageblocks = 1 << (start_order - pageblock_order);
1964 while (nr_pageblocks--) {
1965 set_pageblock_migratetype(pageblock_page, migratetype);
1966 pageblock_page += pageblock_nr_pages;
1971 * When we are falling back to another migratetype during allocation, try to
1972 * steal extra free pages from the same pageblocks to satisfy further
1973 * allocations, instead of polluting multiple pageblocks.
1975 * If we are stealing a relatively large buddy page, it is likely there will
1976 * be more free pages in the pageblock, so try to steal them all. For
1977 * reclaimable and unmovable allocations, we steal regardless of page size,
1978 * as fragmentation caused by those allocations polluting movable pageblocks
1979 * is worse than movable allocations stealing from unmovable and reclaimable
1982 static bool can_steal_fallback(unsigned int order, int start_mt)
1985 * Leaving this order check is intended, although there is
1986 * relaxed order check in next check. The reason is that
1987 * we can actually steal whole pageblock if this condition met,
1988 * but, below check doesn't guarantee it and that is just heuristic
1989 * so could be changed anytime.
1991 if (order >= pageblock_order)
1994 if (order >= pageblock_order / 2 ||
1995 start_mt == MIGRATE_RECLAIMABLE ||
1996 start_mt == MIGRATE_UNMOVABLE ||
1997 page_group_by_mobility_disabled)
2003 static inline bool boost_watermark(struct zone *zone)
2005 unsigned long max_boost;
2007 if (!watermark_boost_factor)
2010 * Don't bother in zones that are unlikely to produce results.
2011 * On small machines, including kdump capture kernels running
2012 * in a small area, boosting the watermark can cause an out of
2013 * memory situation immediately.
2015 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2018 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2019 watermark_boost_factor, 10000);
2022 * high watermark may be uninitialised if fragmentation occurs
2023 * very early in boot so do not boost. We do not fall
2024 * through and boost by pageblock_nr_pages as failing
2025 * allocations that early means that reclaim is not going
2026 * to help and it may even be impossible to reclaim the
2027 * boosted watermark resulting in a hang.
2032 max_boost = max(pageblock_nr_pages, max_boost);
2034 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2041 * This function implements actual steal behaviour. If order is large enough,
2042 * we can steal whole pageblock. If not, we first move freepages in this
2043 * pageblock to our migratetype and determine how many already-allocated pages
2044 * are there in the pageblock with a compatible migratetype. If at least half
2045 * of pages are free or compatible, we can change migratetype of the pageblock
2046 * itself, so pages freed in the future will be put on the correct free list.
2048 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2049 unsigned int alloc_flags, int start_type, bool whole_block)
2051 unsigned int current_order = buddy_order(page);
2052 int free_pages, movable_pages, alike_pages;
2055 old_block_type = get_pageblock_migratetype(page);
2058 * This can happen due to races and we want to prevent broken
2059 * highatomic accounting.
2061 if (is_migrate_highatomic(old_block_type))
2064 /* Take ownership for orders >= pageblock_order */
2065 if (current_order >= pageblock_order) {
2066 change_pageblock_range(page, current_order, start_type);
2071 * Boost watermarks to increase reclaim pressure to reduce the
2072 * likelihood of future fallbacks. Wake kswapd now as the node
2073 * may be balanced overall and kswapd will not wake naturally.
2075 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2076 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2078 /* We are not allowed to try stealing from the whole block */
2082 free_pages = move_freepages_block(zone, page, start_type,
2085 * Determine how many pages are compatible with our allocation.
2086 * For movable allocation, it's the number of movable pages which
2087 * we just obtained. For other types it's a bit more tricky.
2089 if (start_type == MIGRATE_MOVABLE) {
2090 alike_pages = movable_pages;
2093 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2094 * to MOVABLE pageblock, consider all non-movable pages as
2095 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2096 * vice versa, be conservative since we can't distinguish the
2097 * exact migratetype of non-movable pages.
2099 if (old_block_type == MIGRATE_MOVABLE)
2100 alike_pages = pageblock_nr_pages
2101 - (free_pages + movable_pages);
2106 /* moving whole block can fail due to zone boundary conditions */
2111 * If a sufficient number of pages in the block are either free or of
2112 * comparable migratability as our allocation, claim the whole block.
2114 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2115 page_group_by_mobility_disabled)
2116 set_pageblock_migratetype(page, start_type);
2121 move_to_free_list(page, zone, current_order, start_type);
2125 * Check whether there is a suitable fallback freepage with requested order.
2126 * If only_stealable is true, this function returns fallback_mt only if
2127 * we can steal other freepages all together. This would help to reduce
2128 * fragmentation due to mixed migratetype pages in one pageblock.
2130 int find_suitable_fallback(struct free_area *area, unsigned int order,
2131 int migratetype, bool only_stealable, bool *can_steal)
2136 if (area->nr_free == 0)
2140 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
2141 fallback_mt = fallbacks[migratetype][i];
2142 if (free_area_empty(area, fallback_mt))
2145 if (can_steal_fallback(order, migratetype))
2148 if (!only_stealable)
2159 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2160 * there are no empty page blocks that contain a page with a suitable order
2162 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2163 unsigned int alloc_order)
2166 unsigned long max_managed, flags;
2169 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2170 * Check is race-prone but harmless.
2172 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2173 if (zone->nr_reserved_highatomic >= max_managed)
2176 spin_lock_irqsave(&zone->lock, flags);
2178 /* Recheck the nr_reserved_highatomic limit under the lock */
2179 if (zone->nr_reserved_highatomic >= max_managed)
2183 mt = get_pageblock_migratetype(page);
2184 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2185 if (migratetype_is_mergeable(mt)) {
2186 zone->nr_reserved_highatomic += pageblock_nr_pages;
2187 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2188 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2192 spin_unlock_irqrestore(&zone->lock, flags);
2196 * Used when an allocation is about to fail under memory pressure. This
2197 * potentially hurts the reliability of high-order allocations when under
2198 * intense memory pressure but failed atomic allocations should be easier
2199 * to recover from than an OOM.
2201 * If @force is true, try to unreserve a pageblock even though highatomic
2202 * pageblock is exhausted.
2204 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2207 struct zonelist *zonelist = ac->zonelist;
2208 unsigned long flags;
2215 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2218 * Preserve at least one pageblock unless memory pressure
2221 if (!force && zone->nr_reserved_highatomic <=
2225 spin_lock_irqsave(&zone->lock, flags);
2226 for (order = 0; order <= MAX_ORDER; order++) {
2227 struct free_area *area = &(zone->free_area[order]);
2229 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2234 * In page freeing path, migratetype change is racy so
2235 * we can counter several free pages in a pageblock
2236 * in this loop although we changed the pageblock type
2237 * from highatomic to ac->migratetype. So we should
2238 * adjust the count once.
2240 if (is_migrate_highatomic_page(page)) {
2242 * It should never happen but changes to
2243 * locking could inadvertently allow a per-cpu
2244 * drain to add pages to MIGRATE_HIGHATOMIC
2245 * while unreserving so be safe and watch for
2248 zone->nr_reserved_highatomic -= min(
2250 zone->nr_reserved_highatomic);
2254 * Convert to ac->migratetype and avoid the normal
2255 * pageblock stealing heuristics. Minimally, the caller
2256 * is doing the work and needs the pages. More
2257 * importantly, if the block was always converted to
2258 * MIGRATE_UNMOVABLE or another type then the number
2259 * of pageblocks that cannot be completely freed
2262 set_pageblock_migratetype(page, ac->migratetype);
2263 ret = move_freepages_block(zone, page, ac->migratetype,
2266 spin_unlock_irqrestore(&zone->lock, flags);
2270 spin_unlock_irqrestore(&zone->lock, flags);
2277 * Try finding a free buddy page on the fallback list and put it on the free
2278 * list of requested migratetype, possibly along with other pages from the same
2279 * block, depending on fragmentation avoidance heuristics. Returns true if
2280 * fallback was found so that __rmqueue_smallest() can grab it.
2282 * The use of signed ints for order and current_order is a deliberate
2283 * deviation from the rest of this file, to make the for loop
2284 * condition simpler.
2286 static __always_inline bool
2287 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2288 unsigned int alloc_flags)
2290 struct free_area *area;
2292 int min_order = order;
2298 * Do not steal pages from freelists belonging to other pageblocks
2299 * i.e. orders < pageblock_order. If there are no local zones free,
2300 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2302 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2303 min_order = pageblock_order;
2306 * Find the largest available free page in the other list. This roughly
2307 * approximates finding the pageblock with the most free pages, which
2308 * would be too costly to do exactly.
2310 for (current_order = MAX_ORDER; current_order >= min_order;
2312 area = &(zone->free_area[current_order]);
2313 fallback_mt = find_suitable_fallback(area, current_order,
2314 start_migratetype, false, &can_steal);
2315 if (fallback_mt == -1)
2319 * We cannot steal all free pages from the pageblock and the
2320 * requested migratetype is movable. In that case it's better to
2321 * steal and split the smallest available page instead of the
2322 * largest available page, because even if the next movable
2323 * allocation falls back into a different pageblock than this
2324 * one, it won't cause permanent fragmentation.
2326 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2327 && current_order > order)
2336 for (current_order = order; current_order <= MAX_ORDER;
2338 area = &(zone->free_area[current_order]);
2339 fallback_mt = find_suitable_fallback(area, current_order,
2340 start_migratetype, false, &can_steal);
2341 if (fallback_mt != -1)
2346 * This should not happen - we already found a suitable fallback
2347 * when looking for the largest page.
2349 VM_BUG_ON(current_order > MAX_ORDER);
2352 page = get_page_from_free_area(area, fallback_mt);
2354 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2357 trace_mm_page_alloc_extfrag(page, order, current_order,
2358 start_migratetype, fallback_mt);
2365 * Do the hard work of removing an element from the buddy allocator.
2366 * Call me with the zone->lock already held.
2368 static __always_inline struct page *
2369 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2370 unsigned int alloc_flags)
2374 if (IS_ENABLED(CONFIG_CMA)) {
2376 * Balance movable allocations between regular and CMA areas by
2377 * allocating from CMA when over half of the zone's free memory
2378 * is in the CMA area.
2380 if (alloc_flags & ALLOC_CMA &&
2381 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2382 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2383 page = __rmqueue_cma_fallback(zone, order);
2389 page = __rmqueue_smallest(zone, order, migratetype);
2390 if (unlikely(!page)) {
2391 if (alloc_flags & ALLOC_CMA)
2392 page = __rmqueue_cma_fallback(zone, order);
2394 if (!page && __rmqueue_fallback(zone, order, migratetype,
2402 * Obtain a specified number of elements from the buddy allocator, all under
2403 * a single hold of the lock, for efficiency. Add them to the supplied list.
2404 * Returns the number of new pages which were placed at *list.
2406 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2407 unsigned long count, struct list_head *list,
2408 int migratetype, unsigned int alloc_flags)
2410 unsigned long flags;
2413 spin_lock_irqsave(&zone->lock, flags);
2414 for (i = 0; i < count; ++i) {
2415 struct page *page = __rmqueue(zone, order, migratetype,
2417 if (unlikely(page == NULL))
2421 * Split buddy pages returned by expand() are received here in
2422 * physical page order. The page is added to the tail of
2423 * caller's list. From the callers perspective, the linked list
2424 * is ordered by page number under some conditions. This is
2425 * useful for IO devices that can forward direction from the
2426 * head, thus also in the physical page order. This is useful
2427 * for IO devices that can merge IO requests if the physical
2428 * pages are ordered properly.
2430 list_add_tail(&page->pcp_list, list);
2431 if (is_migrate_cma(get_pcppage_migratetype(page)))
2432 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2436 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2437 spin_unlock_irqrestore(&zone->lock, flags);
2444 * Called from the vmstat counter updater to drain pagesets of this
2445 * currently executing processor on remote nodes after they have
2448 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2450 int to_drain, batch;
2452 batch = READ_ONCE(pcp->batch);
2453 to_drain = min(pcp->count, batch);
2455 spin_lock(&pcp->lock);
2456 free_pcppages_bulk(zone, to_drain, pcp, 0);
2457 spin_unlock(&pcp->lock);
2463 * Drain pcplists of the indicated processor and zone.
2465 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2467 struct per_cpu_pages *pcp;
2469 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2471 spin_lock(&pcp->lock);
2472 free_pcppages_bulk(zone, pcp->count, pcp, 0);
2473 spin_unlock(&pcp->lock);
2478 * Drain pcplists of all zones on the indicated processor.
2480 static void drain_pages(unsigned int cpu)
2484 for_each_populated_zone(zone) {
2485 drain_pages_zone(cpu, zone);
2490 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2492 void drain_local_pages(struct zone *zone)
2494 int cpu = smp_processor_id();
2497 drain_pages_zone(cpu, zone);
2503 * The implementation of drain_all_pages(), exposing an extra parameter to
2504 * drain on all cpus.
2506 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2507 * not empty. The check for non-emptiness can however race with a free to
2508 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2509 * that need the guarantee that every CPU has drained can disable the
2510 * optimizing racy check.
2512 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2517 * Allocate in the BSS so we won't require allocation in
2518 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2520 static cpumask_t cpus_with_pcps;
2523 * Do not drain if one is already in progress unless it's specific to
2524 * a zone. Such callers are primarily CMA and memory hotplug and need
2525 * the drain to be complete when the call returns.
2527 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2530 mutex_lock(&pcpu_drain_mutex);
2534 * We don't care about racing with CPU hotplug event
2535 * as offline notification will cause the notified
2536 * cpu to drain that CPU pcps and on_each_cpu_mask
2537 * disables preemption as part of its processing
2539 for_each_online_cpu(cpu) {
2540 struct per_cpu_pages *pcp;
2542 bool has_pcps = false;
2544 if (force_all_cpus) {
2546 * The pcp.count check is racy, some callers need a
2547 * guarantee that no cpu is missed.
2551 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2555 for_each_populated_zone(z) {
2556 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2565 cpumask_set_cpu(cpu, &cpus_with_pcps);
2567 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2570 for_each_cpu(cpu, &cpus_with_pcps) {
2572 drain_pages_zone(cpu, zone);
2577 mutex_unlock(&pcpu_drain_mutex);
2581 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2583 * When zone parameter is non-NULL, spill just the single zone's pages.
2585 void drain_all_pages(struct zone *zone)
2587 __drain_all_pages(zone, false);
2590 #ifdef CONFIG_HIBERNATION
2593 * Touch the watchdog for every WD_PAGE_COUNT pages.
2595 #define WD_PAGE_COUNT (128*1024)
2597 void mark_free_pages(struct zone *zone)
2599 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2600 unsigned long flags;
2601 unsigned int order, t;
2604 if (zone_is_empty(zone))
2607 spin_lock_irqsave(&zone->lock, flags);
2609 max_zone_pfn = zone_end_pfn(zone);
2610 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2611 if (pfn_valid(pfn)) {
2612 page = pfn_to_page(pfn);
2614 if (!--page_count) {
2615 touch_nmi_watchdog();
2616 page_count = WD_PAGE_COUNT;
2619 if (page_zone(page) != zone)
2622 if (!swsusp_page_is_forbidden(page))
2623 swsusp_unset_page_free(page);
2626 for_each_migratetype_order(order, t) {
2627 list_for_each_entry(page,
2628 &zone->free_area[order].free_list[t], buddy_list) {
2631 pfn = page_to_pfn(page);
2632 for (i = 0; i < (1UL << order); i++) {
2633 if (!--page_count) {
2634 touch_nmi_watchdog();
2635 page_count = WD_PAGE_COUNT;
2637 swsusp_set_page_free(pfn_to_page(pfn + i));
2641 spin_unlock_irqrestore(&zone->lock, flags);
2643 #endif /* CONFIG_PM */
2645 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
2650 if (!free_pages_prepare(page, order, FPI_NONE))
2653 migratetype = get_pfnblock_migratetype(page, pfn);
2654 set_pcppage_migratetype(page, migratetype);
2658 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
2661 int min_nr_free, max_nr_free;
2663 /* Free everything if batch freeing high-order pages. */
2664 if (unlikely(free_high))
2667 /* Check for PCP disabled or boot pageset */
2668 if (unlikely(high < batch))
2671 /* Leave at least pcp->batch pages on the list */
2672 min_nr_free = batch;
2673 max_nr_free = high - batch;
2676 * Double the number of pages freed each time there is subsequent
2677 * freeing of pages without any allocation.
2679 batch <<= pcp->free_factor;
2680 if (batch < max_nr_free)
2682 batch = clamp(batch, min_nr_free, max_nr_free);
2687 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2690 int high = READ_ONCE(pcp->high);
2692 if (unlikely(!high || free_high))
2695 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
2699 * If reclaim is active, limit the number of pages that can be
2700 * stored on pcp lists
2702 return min(READ_ONCE(pcp->batch) << 2, high);
2705 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2706 struct page *page, int migratetype,
2713 __count_vm_events(PGFREE, 1 << order);
2714 pindex = order_to_pindex(migratetype, order);
2715 list_add(&page->pcp_list, &pcp->lists[pindex]);
2716 pcp->count += 1 << order;
2719 * As high-order pages other than THP's stored on PCP can contribute
2720 * to fragmentation, limit the number stored when PCP is heavily
2721 * freeing without allocation. The remainder after bulk freeing
2722 * stops will be drained from vmstat refresh context.
2724 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
2726 high = nr_pcp_high(pcp, zone, free_high);
2727 if (pcp->count >= high) {
2728 int batch = READ_ONCE(pcp->batch);
2730 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
2737 void free_unref_page(struct page *page, unsigned int order)
2739 unsigned long __maybe_unused UP_flags;
2740 struct per_cpu_pages *pcp;
2742 unsigned long pfn = page_to_pfn(page);
2745 if (!free_unref_page_prepare(page, pfn, order))
2749 * We only track unmovable, reclaimable and movable on pcp lists.
2750 * Place ISOLATE pages on the isolated list because they are being
2751 * offlined but treat HIGHATOMIC as movable pages so we can get those
2752 * areas back if necessary. Otherwise, we may have to free
2753 * excessively into the page allocator
2755 migratetype = get_pcppage_migratetype(page);
2756 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2757 if (unlikely(is_migrate_isolate(migratetype))) {
2758 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
2761 migratetype = MIGRATE_MOVABLE;
2764 zone = page_zone(page);
2765 pcp_trylock_prepare(UP_flags);
2766 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2768 free_unref_page_commit(zone, pcp, page, migratetype, order);
2769 pcp_spin_unlock(pcp);
2771 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
2773 pcp_trylock_finish(UP_flags);
2777 * Free a list of 0-order pages
2779 void free_unref_page_list(struct list_head *list)
2781 unsigned long __maybe_unused UP_flags;
2782 struct page *page, *next;
2783 struct per_cpu_pages *pcp = NULL;
2784 struct zone *locked_zone = NULL;
2785 int batch_count = 0;
2788 /* Prepare pages for freeing */
2789 list_for_each_entry_safe(page, next, list, lru) {
2790 unsigned long pfn = page_to_pfn(page);
2791 if (!free_unref_page_prepare(page, pfn, 0)) {
2792 list_del(&page->lru);
2797 * Free isolated pages directly to the allocator, see
2798 * comment in free_unref_page.
2800 migratetype = get_pcppage_migratetype(page);
2801 if (unlikely(is_migrate_isolate(migratetype))) {
2802 list_del(&page->lru);
2803 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
2808 list_for_each_entry_safe(page, next, list, lru) {
2809 struct zone *zone = page_zone(page);
2811 list_del(&page->lru);
2812 migratetype = get_pcppage_migratetype(page);
2815 * Either different zone requiring a different pcp lock or
2816 * excessive lock hold times when freeing a large list of
2819 if (zone != locked_zone || batch_count == SWAP_CLUSTER_MAX) {
2821 pcp_spin_unlock(pcp);
2822 pcp_trylock_finish(UP_flags);
2828 * trylock is necessary as pages may be getting freed
2829 * from IRQ or SoftIRQ context after an IO completion.
2831 pcp_trylock_prepare(UP_flags);
2832 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2833 if (unlikely(!pcp)) {
2834 pcp_trylock_finish(UP_flags);
2835 free_one_page(zone, page, page_to_pfn(page),
2836 0, migratetype, FPI_NONE);
2844 * Non-isolated types over MIGRATE_PCPTYPES get added
2845 * to the MIGRATE_MOVABLE pcp list.
2847 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2848 migratetype = MIGRATE_MOVABLE;
2850 trace_mm_page_free_batched(page);
2851 free_unref_page_commit(zone, pcp, page, migratetype, 0);
2856 pcp_spin_unlock(pcp);
2857 pcp_trylock_finish(UP_flags);
2862 * split_page takes a non-compound higher-order page, and splits it into
2863 * n (1<<order) sub-pages: page[0..n]
2864 * Each sub-page must be freed individually.
2866 * Note: this is probably too low level an operation for use in drivers.
2867 * Please consult with lkml before using this in your driver.
2869 void split_page(struct page *page, unsigned int order)
2873 VM_BUG_ON_PAGE(PageCompound(page), page);
2874 VM_BUG_ON_PAGE(!page_count(page), page);
2876 for (i = 1; i < (1 << order); i++)
2877 set_page_refcounted(page + i);
2878 split_page_owner(page, 1 << order);
2879 split_page_memcg(page, 1 << order);
2881 EXPORT_SYMBOL_GPL(split_page);
2883 int __isolate_free_page(struct page *page, unsigned int order)
2885 struct zone *zone = page_zone(page);
2886 int mt = get_pageblock_migratetype(page);
2888 if (!is_migrate_isolate(mt)) {
2889 unsigned long watermark;
2891 * Obey watermarks as if the page was being allocated. We can
2892 * emulate a high-order watermark check with a raised order-0
2893 * watermark, because we already know our high-order page
2896 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2897 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2900 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2903 del_page_from_free_list(page, zone, order);
2906 * Set the pageblock if the isolated page is at least half of a
2909 if (order >= pageblock_order - 1) {
2910 struct page *endpage = page + (1 << order) - 1;
2911 for (; page < endpage; page += pageblock_nr_pages) {
2912 int mt = get_pageblock_migratetype(page);
2914 * Only change normal pageblocks (i.e., they can merge
2917 if (migratetype_is_mergeable(mt))
2918 set_pageblock_migratetype(page,
2923 return 1UL << order;
2927 * __putback_isolated_page - Return a now-isolated page back where we got it
2928 * @page: Page that was isolated
2929 * @order: Order of the isolated page
2930 * @mt: The page's pageblock's migratetype
2932 * This function is meant to return a page pulled from the free lists via
2933 * __isolate_free_page back to the free lists they were pulled from.
2935 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2937 struct zone *zone = page_zone(page);
2939 /* zone lock should be held when this function is called */
2940 lockdep_assert_held(&zone->lock);
2942 /* Return isolated page to tail of freelist. */
2943 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2944 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2948 * Update NUMA hit/miss statistics
2950 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2954 enum numa_stat_item local_stat = NUMA_LOCAL;
2956 /* skip numa counters update if numa stats is disabled */
2957 if (!static_branch_likely(&vm_numa_stat_key))
2960 if (zone_to_nid(z) != numa_node_id())
2961 local_stat = NUMA_OTHER;
2963 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2964 __count_numa_events(z, NUMA_HIT, nr_account);
2966 __count_numa_events(z, NUMA_MISS, nr_account);
2967 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2969 __count_numa_events(z, local_stat, nr_account);
2973 static __always_inline
2974 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2975 unsigned int order, unsigned int alloc_flags,
2979 unsigned long flags;
2983 spin_lock_irqsave(&zone->lock, flags);
2985 * order-0 request can reach here when the pcplist is skipped
2986 * due to non-CMA allocation context. HIGHATOMIC area is
2987 * reserved for high-order atomic allocation, so order-0
2988 * request should skip it.
2990 if (alloc_flags & ALLOC_HIGHATOMIC)
2991 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2993 page = __rmqueue(zone, order, migratetype, alloc_flags);
2996 * If the allocation fails, allow OOM handling access
2997 * to HIGHATOMIC reserves as failing now is worse than
2998 * failing a high-order atomic allocation in the
3001 if (!page && (alloc_flags & ALLOC_OOM))
3002 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3005 spin_unlock_irqrestore(&zone->lock, flags);
3009 __mod_zone_freepage_state(zone, -(1 << order),
3010 get_pcppage_migratetype(page));
3011 spin_unlock_irqrestore(&zone->lock, flags);
3012 } while (check_new_pages(page, order));
3014 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3015 zone_statistics(preferred_zone, zone, 1);
3020 /* Remove page from the per-cpu list, caller must protect the list */
3022 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3024 unsigned int alloc_flags,
3025 struct per_cpu_pages *pcp,
3026 struct list_head *list)
3031 if (list_empty(list)) {
3032 int batch = READ_ONCE(pcp->batch);
3036 * Scale batch relative to order if batch implies
3037 * free pages can be stored on the PCP. Batch can
3038 * be 1 for small zones or for boot pagesets which
3039 * should never store free pages as the pages may
3040 * belong to arbitrary zones.
3043 batch = max(batch >> order, 2);
3044 alloced = rmqueue_bulk(zone, order,
3046 migratetype, alloc_flags);
3048 pcp->count += alloced << order;
3049 if (unlikely(list_empty(list)))
3053 page = list_first_entry(list, struct page, pcp_list);
3054 list_del(&page->pcp_list);
3055 pcp->count -= 1 << order;
3056 } while (check_new_pages(page, order));
3061 /* Lock and remove page from the per-cpu list */
3062 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3063 struct zone *zone, unsigned int order,
3064 int migratetype, unsigned int alloc_flags)
3066 struct per_cpu_pages *pcp;
3067 struct list_head *list;
3069 unsigned long __maybe_unused UP_flags;
3071 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
3072 pcp_trylock_prepare(UP_flags);
3073 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3075 pcp_trylock_finish(UP_flags);
3080 * On allocation, reduce the number of pages that are batch freed.
3081 * See nr_pcp_free() where free_factor is increased for subsequent
3084 pcp->free_factor >>= 1;
3085 list = &pcp->lists[order_to_pindex(migratetype, order)];
3086 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3087 pcp_spin_unlock(pcp);
3088 pcp_trylock_finish(UP_flags);
3090 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3091 zone_statistics(preferred_zone, zone, 1);
3097 * Allocate a page from the given zone.
3098 * Use pcplists for THP or "cheap" high-order allocations.
3102 * Do not instrument rmqueue() with KMSAN. This function may call
3103 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
3104 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3105 * may call rmqueue() again, which will result in a deadlock.
3107 __no_sanitize_memory
3109 struct page *rmqueue(struct zone *preferred_zone,
3110 struct zone *zone, unsigned int order,
3111 gfp_t gfp_flags, unsigned int alloc_flags,
3117 * We most definitely don't want callers attempting to
3118 * allocate greater than order-1 page units with __GFP_NOFAIL.
3120 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3122 if (likely(pcp_allowed_order(order))) {
3124 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3125 * we need to skip it when CMA area isn't allowed.
3127 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3128 migratetype != MIGRATE_MOVABLE) {
3129 page = rmqueue_pcplist(preferred_zone, zone, order,
3130 migratetype, alloc_flags);
3136 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3140 /* Separate test+clear to avoid unnecessary atomics */
3141 if (unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3142 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3143 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3146 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3150 #ifdef CONFIG_FAIL_PAGE_ALLOC
3153 struct fault_attr attr;
3155 bool ignore_gfp_highmem;
3156 bool ignore_gfp_reclaim;
3158 } fail_page_alloc = {
3159 .attr = FAULT_ATTR_INITIALIZER,
3160 .ignore_gfp_reclaim = true,
3161 .ignore_gfp_highmem = true,
3165 static int __init setup_fail_page_alloc(char *str)
3167 return setup_fault_attr(&fail_page_alloc.attr, str);
3169 __setup("fail_page_alloc=", setup_fail_page_alloc);
3171 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3175 if (order < fail_page_alloc.min_order)
3177 if (gfp_mask & __GFP_NOFAIL)
3179 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3181 if (fail_page_alloc.ignore_gfp_reclaim &&
3182 (gfp_mask & __GFP_DIRECT_RECLAIM))
3185 /* See comment in __should_failslab() */
3186 if (gfp_mask & __GFP_NOWARN)
3187 flags |= FAULT_NOWARN;
3189 return should_fail_ex(&fail_page_alloc.attr, 1 << order, flags);
3192 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3194 static int __init fail_page_alloc_debugfs(void)
3196 umode_t mode = S_IFREG | 0600;
3199 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3200 &fail_page_alloc.attr);
3202 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3203 &fail_page_alloc.ignore_gfp_reclaim);
3204 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3205 &fail_page_alloc.ignore_gfp_highmem);
3206 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3211 late_initcall(fail_page_alloc_debugfs);
3213 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3215 #else /* CONFIG_FAIL_PAGE_ALLOC */
3217 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3222 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3224 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3226 return __should_fail_alloc_page(gfp_mask, order);
3228 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3230 static inline long __zone_watermark_unusable_free(struct zone *z,
3231 unsigned int order, unsigned int alloc_flags)
3233 long unusable_free = (1 << order) - 1;
3236 * If the caller does not have rights to reserves below the min
3237 * watermark then subtract the high-atomic reserves. This will
3238 * over-estimate the size of the atomic reserve but it avoids a search.
3240 if (likely(!(alloc_flags & ALLOC_RESERVES)))
3241 unusable_free += z->nr_reserved_highatomic;
3244 /* If allocation can't use CMA areas don't use free CMA pages */
3245 if (!(alloc_flags & ALLOC_CMA))
3246 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3249 return unusable_free;
3253 * Return true if free base pages are above 'mark'. For high-order checks it
3254 * will return true of the order-0 watermark is reached and there is at least
3255 * one free page of a suitable size. Checking now avoids taking the zone lock
3256 * to check in the allocation paths if no pages are free.
3258 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3259 int highest_zoneidx, unsigned int alloc_flags,
3265 /* free_pages may go negative - that's OK */
3266 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3268 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
3270 * __GFP_HIGH allows access to 50% of the min reserve as well
3273 if (alloc_flags & ALLOC_MIN_RESERVE) {
3277 * Non-blocking allocations (e.g. GFP_ATOMIC) can
3278 * access more reserves than just __GFP_HIGH. Other
3279 * non-blocking allocations requests such as GFP_NOWAIT
3280 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
3281 * access to the min reserve.
3283 if (alloc_flags & ALLOC_NON_BLOCK)
3288 * OOM victims can try even harder than the normal reserve
3289 * users on the grounds that it's definitely going to be in
3290 * the exit path shortly and free memory. Any allocation it
3291 * makes during the free path will be small and short-lived.
3293 if (alloc_flags & ALLOC_OOM)
3298 * Check watermarks for an order-0 allocation request. If these
3299 * are not met, then a high-order request also cannot go ahead
3300 * even if a suitable page happened to be free.
3302 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3305 /* If this is an order-0 request then the watermark is fine */
3309 /* For a high-order request, check at least one suitable page is free */
3310 for (o = order; o <= MAX_ORDER; o++) {
3311 struct free_area *area = &z->free_area[o];
3317 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3318 if (!free_area_empty(area, mt))
3323 if ((alloc_flags & ALLOC_CMA) &&
3324 !free_area_empty(area, MIGRATE_CMA)) {
3328 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3329 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3336 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3337 int highest_zoneidx, unsigned int alloc_flags)
3339 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3340 zone_page_state(z, NR_FREE_PAGES));
3343 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3344 unsigned long mark, int highest_zoneidx,
3345 unsigned int alloc_flags, gfp_t gfp_mask)
3349 free_pages = zone_page_state(z, NR_FREE_PAGES);
3352 * Fast check for order-0 only. If this fails then the reserves
3353 * need to be calculated.
3359 usable_free = free_pages;
3360 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3362 /* reserved may over estimate high-atomic reserves. */
3363 usable_free -= min(usable_free, reserved);
3364 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3368 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3373 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3374 * when checking the min watermark. The min watermark is the
3375 * point where boosting is ignored so that kswapd is woken up
3376 * when below the low watermark.
3378 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3379 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3380 mark = z->_watermark[WMARK_MIN];
3381 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3382 alloc_flags, free_pages);
3388 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3389 unsigned long mark, int highest_zoneidx)
3391 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3393 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3394 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3396 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3401 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3403 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3405 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3406 node_reclaim_distance;
3408 #else /* CONFIG_NUMA */
3409 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3413 #endif /* CONFIG_NUMA */
3416 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3417 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3418 * premature use of a lower zone may cause lowmem pressure problems that
3419 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3420 * probably too small. It only makes sense to spread allocations to avoid
3421 * fragmentation between the Normal and DMA32 zones.
3423 static inline unsigned int
3424 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3426 unsigned int alloc_flags;
3429 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3432 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3434 #ifdef CONFIG_ZONE_DMA32
3438 if (zone_idx(zone) != ZONE_NORMAL)
3442 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3443 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3444 * on UMA that if Normal is populated then so is DMA32.
3446 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3447 if (nr_online_nodes > 1 && !populated_zone(--zone))
3450 alloc_flags |= ALLOC_NOFRAGMENT;
3451 #endif /* CONFIG_ZONE_DMA32 */
3455 /* Must be called after current_gfp_context() which can change gfp_mask */
3456 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3457 unsigned int alloc_flags)
3460 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3461 alloc_flags |= ALLOC_CMA;
3467 * get_page_from_freelist goes through the zonelist trying to allocate
3470 static struct page *
3471 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3472 const struct alloc_context *ac)
3476 struct pglist_data *last_pgdat = NULL;
3477 bool last_pgdat_dirty_ok = false;
3482 * Scan zonelist, looking for a zone with enough free.
3483 * See also __cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3485 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3486 z = ac->preferred_zoneref;
3487 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3492 if (cpusets_enabled() &&
3493 (alloc_flags & ALLOC_CPUSET) &&
3494 !__cpuset_zone_allowed(zone, gfp_mask))
3497 * When allocating a page cache page for writing, we
3498 * want to get it from a node that is within its dirty
3499 * limit, such that no single node holds more than its
3500 * proportional share of globally allowed dirty pages.
3501 * The dirty limits take into account the node's
3502 * lowmem reserves and high watermark so that kswapd
3503 * should be able to balance it without having to
3504 * write pages from its LRU list.
3506 * XXX: For now, allow allocations to potentially
3507 * exceed the per-node dirty limit in the slowpath
3508 * (spread_dirty_pages unset) before going into reclaim,
3509 * which is important when on a NUMA setup the allowed
3510 * nodes are together not big enough to reach the
3511 * global limit. The proper fix for these situations
3512 * will require awareness of nodes in the
3513 * dirty-throttling and the flusher threads.
3515 if (ac->spread_dirty_pages) {
3516 if (last_pgdat != zone->zone_pgdat) {
3517 last_pgdat = zone->zone_pgdat;
3518 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3521 if (!last_pgdat_dirty_ok)
3525 if (no_fallback && nr_online_nodes > 1 &&
3526 zone != ac->preferred_zoneref->zone) {
3530 * If moving to a remote node, retry but allow
3531 * fragmenting fallbacks. Locality is more important
3532 * than fragmentation avoidance.
3534 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3535 if (zone_to_nid(zone) != local_nid) {
3536 alloc_flags &= ~ALLOC_NOFRAGMENT;
3541 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3542 if (!zone_watermark_fast(zone, order, mark,
3543 ac->highest_zoneidx, alloc_flags,
3547 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3549 * Watermark failed for this zone, but see if we can
3550 * grow this zone if it contains deferred pages.
3552 if (deferred_pages_enabled()) {
3553 if (_deferred_grow_zone(zone, order))
3557 /* Checked here to keep the fast path fast */
3558 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3559 if (alloc_flags & ALLOC_NO_WATERMARKS)
3562 if (!node_reclaim_enabled() ||
3563 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3566 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3568 case NODE_RECLAIM_NOSCAN:
3571 case NODE_RECLAIM_FULL:
3572 /* scanned but unreclaimable */
3575 /* did we reclaim enough */
3576 if (zone_watermark_ok(zone, order, mark,
3577 ac->highest_zoneidx, alloc_flags))
3585 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3586 gfp_mask, alloc_flags, ac->migratetype);
3588 prep_new_page(page, order, gfp_mask, alloc_flags);
3591 * If this is a high-order atomic allocation then check
3592 * if the pageblock should be reserved for the future
3594 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3595 reserve_highatomic_pageblock(page, zone, order);
3599 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3600 /* Try again if zone has deferred pages */
3601 if (deferred_pages_enabled()) {
3602 if (_deferred_grow_zone(zone, order))
3610 * It's possible on a UMA machine to get through all zones that are
3611 * fragmented. If avoiding fragmentation, reset and try again.
3614 alloc_flags &= ~ALLOC_NOFRAGMENT;
3621 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3623 unsigned int filter = SHOW_MEM_FILTER_NODES;
3626 * This documents exceptions given to allocations in certain
3627 * contexts that are allowed to allocate outside current's set
3630 if (!(gfp_mask & __GFP_NOMEMALLOC))
3631 if (tsk_is_oom_victim(current) ||
3632 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3633 filter &= ~SHOW_MEM_FILTER_NODES;
3634 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3635 filter &= ~SHOW_MEM_FILTER_NODES;
3637 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3640 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3642 struct va_format vaf;
3644 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3646 if ((gfp_mask & __GFP_NOWARN) ||
3647 !__ratelimit(&nopage_rs) ||
3648 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3651 va_start(args, fmt);
3654 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3655 current->comm, &vaf, gfp_mask, &gfp_mask,
3656 nodemask_pr_args(nodemask));
3659 cpuset_print_current_mems_allowed();
3662 warn_alloc_show_mem(gfp_mask, nodemask);
3665 static inline struct page *
3666 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3667 unsigned int alloc_flags,
3668 const struct alloc_context *ac)
3672 page = get_page_from_freelist(gfp_mask, order,
3673 alloc_flags|ALLOC_CPUSET, ac);
3675 * fallback to ignore cpuset restriction if our nodes
3679 page = get_page_from_freelist(gfp_mask, order,
3685 static inline struct page *
3686 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3687 const struct alloc_context *ac, unsigned long *did_some_progress)
3689 struct oom_control oc = {
3690 .zonelist = ac->zonelist,
3691 .nodemask = ac->nodemask,
3693 .gfp_mask = gfp_mask,
3698 *did_some_progress = 0;
3701 * Acquire the oom lock. If that fails, somebody else is
3702 * making progress for us.
3704 if (!mutex_trylock(&oom_lock)) {
3705 *did_some_progress = 1;
3706 schedule_timeout_uninterruptible(1);
3711 * Go through the zonelist yet one more time, keep very high watermark
3712 * here, this is only to catch a parallel oom killing, we must fail if
3713 * we're still under heavy pressure. But make sure that this reclaim
3714 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3715 * allocation which will never fail due to oom_lock already held.
3717 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3718 ~__GFP_DIRECT_RECLAIM, order,
3719 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3723 /* Coredumps can quickly deplete all memory reserves */
3724 if (current->flags & PF_DUMPCORE)
3726 /* The OOM killer will not help higher order allocs */
3727 if (order > PAGE_ALLOC_COSTLY_ORDER)
3730 * We have already exhausted all our reclaim opportunities without any
3731 * success so it is time to admit defeat. We will skip the OOM killer
3732 * because it is very likely that the caller has a more reasonable
3733 * fallback than shooting a random task.
3735 * The OOM killer may not free memory on a specific node.
3737 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3739 /* The OOM killer does not needlessly kill tasks for lowmem */
3740 if (ac->highest_zoneidx < ZONE_NORMAL)
3742 if (pm_suspended_storage())
3745 * XXX: GFP_NOFS allocations should rather fail than rely on
3746 * other request to make a forward progress.
3747 * We are in an unfortunate situation where out_of_memory cannot
3748 * do much for this context but let's try it to at least get
3749 * access to memory reserved if the current task is killed (see
3750 * out_of_memory). Once filesystems are ready to handle allocation
3751 * failures more gracefully we should just bail out here.
3754 /* Exhausted what can be done so it's blame time */
3755 if (out_of_memory(&oc) ||
3756 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3757 *did_some_progress = 1;
3760 * Help non-failing allocations by giving them access to memory
3763 if (gfp_mask & __GFP_NOFAIL)
3764 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3765 ALLOC_NO_WATERMARKS, ac);
3768 mutex_unlock(&oom_lock);
3773 * Maximum number of compaction retries with a progress before OOM
3774 * killer is consider as the only way to move forward.
3776 #define MAX_COMPACT_RETRIES 16
3778 #ifdef CONFIG_COMPACTION
3779 /* Try memory compaction for high-order allocations before reclaim */
3780 static struct page *
3781 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3782 unsigned int alloc_flags, const struct alloc_context *ac,
3783 enum compact_priority prio, enum compact_result *compact_result)
3785 struct page *page = NULL;
3786 unsigned long pflags;
3787 unsigned int noreclaim_flag;
3792 psi_memstall_enter(&pflags);
3793 delayacct_compact_start();
3794 noreclaim_flag = memalloc_noreclaim_save();
3796 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3799 memalloc_noreclaim_restore(noreclaim_flag);
3800 psi_memstall_leave(&pflags);
3801 delayacct_compact_end();
3803 if (*compact_result == COMPACT_SKIPPED)
3806 * At least in one zone compaction wasn't deferred or skipped, so let's
3807 * count a compaction stall
3809 count_vm_event(COMPACTSTALL);
3811 /* Prep a captured page if available */
3813 prep_new_page(page, order, gfp_mask, alloc_flags);
3815 /* Try get a page from the freelist if available */
3817 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3820 struct zone *zone = page_zone(page);
3822 zone->compact_blockskip_flush = false;
3823 compaction_defer_reset(zone, order, true);
3824 count_vm_event(COMPACTSUCCESS);
3829 * It's bad if compaction run occurs and fails. The most likely reason
3830 * is that pages exist, but not enough to satisfy watermarks.
3832 count_vm_event(COMPACTFAIL);
3840 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3841 enum compact_result compact_result,
3842 enum compact_priority *compact_priority,
3843 int *compaction_retries)
3845 int max_retries = MAX_COMPACT_RETRIES;
3848 int retries = *compaction_retries;
3849 enum compact_priority priority = *compact_priority;
3854 if (fatal_signal_pending(current))
3857 if (compaction_made_progress(compact_result))
3858 (*compaction_retries)++;
3861 * compaction considers all the zone as desperately out of memory
3862 * so it doesn't really make much sense to retry except when the
3863 * failure could be caused by insufficient priority
3865 if (compaction_failed(compact_result))
3866 goto check_priority;
3869 * compaction was skipped because there are not enough order-0 pages
3870 * to work with, so we retry only if it looks like reclaim can help.
3872 if (compaction_needs_reclaim(compact_result)) {
3873 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3878 * make sure the compaction wasn't deferred or didn't bail out early
3879 * due to locks contention before we declare that we should give up.
3880 * But the next retry should use a higher priority if allowed, so
3881 * we don't just keep bailing out endlessly.
3883 if (compaction_withdrawn(compact_result)) {
3884 goto check_priority;
3888 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3889 * costly ones because they are de facto nofail and invoke OOM
3890 * killer to move on while costly can fail and users are ready
3891 * to cope with that. 1/4 retries is rather arbitrary but we
3892 * would need much more detailed feedback from compaction to
3893 * make a better decision.
3895 if (order > PAGE_ALLOC_COSTLY_ORDER)
3897 if (*compaction_retries <= max_retries) {
3903 * Make sure there are attempts at the highest priority if we exhausted
3904 * all retries or failed at the lower priorities.
3907 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3908 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3910 if (*compact_priority > min_priority) {
3911 (*compact_priority)--;
3912 *compaction_retries = 0;
3916 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3920 static inline struct page *
3921 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3922 unsigned int alloc_flags, const struct alloc_context *ac,
3923 enum compact_priority prio, enum compact_result *compact_result)
3925 *compact_result = COMPACT_SKIPPED;
3930 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3931 enum compact_result compact_result,
3932 enum compact_priority *compact_priority,
3933 int *compaction_retries)
3938 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3942 * There are setups with compaction disabled which would prefer to loop
3943 * inside the allocator rather than hit the oom killer prematurely.
3944 * Let's give them a good hope and keep retrying while the order-0
3945 * watermarks are OK.
3947 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3948 ac->highest_zoneidx, ac->nodemask) {
3949 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3950 ac->highest_zoneidx, alloc_flags))
3955 #endif /* CONFIG_COMPACTION */
3957 #ifdef CONFIG_LOCKDEP
3958 static struct lockdep_map __fs_reclaim_map =
3959 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3961 static bool __need_reclaim(gfp_t gfp_mask)
3963 /* no reclaim without waiting on it */
3964 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3967 /* this guy won't enter reclaim */
3968 if (current->flags & PF_MEMALLOC)
3971 if (gfp_mask & __GFP_NOLOCKDEP)
3977 void __fs_reclaim_acquire(unsigned long ip)
3979 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3982 void __fs_reclaim_release(unsigned long ip)
3984 lock_release(&__fs_reclaim_map, ip);
3987 void fs_reclaim_acquire(gfp_t gfp_mask)
3989 gfp_mask = current_gfp_context(gfp_mask);
3991 if (__need_reclaim(gfp_mask)) {
3992 if (gfp_mask & __GFP_FS)
3993 __fs_reclaim_acquire(_RET_IP_);
3995 #ifdef CONFIG_MMU_NOTIFIER
3996 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3997 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4002 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4004 void fs_reclaim_release(gfp_t gfp_mask)
4006 gfp_mask = current_gfp_context(gfp_mask);
4008 if (__need_reclaim(gfp_mask)) {
4009 if (gfp_mask & __GFP_FS)
4010 __fs_reclaim_release(_RET_IP_);
4013 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4017 * Zonelists may change due to hotplug during allocation. Detect when zonelists
4018 * have been rebuilt so allocation retries. Reader side does not lock and
4019 * retries the allocation if zonelist changes. Writer side is protected by the
4020 * embedded spin_lock.
4022 static DEFINE_SEQLOCK(zonelist_update_seq);
4024 static unsigned int zonelist_iter_begin(void)
4026 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4027 return read_seqbegin(&zonelist_update_seq);
4032 static unsigned int check_retry_zonelist(unsigned int seq)
4034 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4035 return read_seqretry(&zonelist_update_seq, seq);
4040 /* Perform direct synchronous page reclaim */
4041 static unsigned long
4042 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4043 const struct alloc_context *ac)
4045 unsigned int noreclaim_flag;
4046 unsigned long progress;
4050 /* We now go into synchronous reclaim */
4051 cpuset_memory_pressure_bump();
4052 fs_reclaim_acquire(gfp_mask);
4053 noreclaim_flag = memalloc_noreclaim_save();
4055 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4058 memalloc_noreclaim_restore(noreclaim_flag);
4059 fs_reclaim_release(gfp_mask);
4066 /* The really slow allocator path where we enter direct reclaim */
4067 static inline struct page *
4068 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4069 unsigned int alloc_flags, const struct alloc_context *ac,
4070 unsigned long *did_some_progress)
4072 struct page *page = NULL;
4073 unsigned long pflags;
4074 bool drained = false;
4076 psi_memstall_enter(&pflags);
4077 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4078 if (unlikely(!(*did_some_progress)))
4082 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4085 * If an allocation failed after direct reclaim, it could be because
4086 * pages are pinned on the per-cpu lists or in high alloc reserves.
4087 * Shrink them and try again
4089 if (!page && !drained) {
4090 unreserve_highatomic_pageblock(ac, false);
4091 drain_all_pages(NULL);
4096 psi_memstall_leave(&pflags);
4101 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4102 const struct alloc_context *ac)
4106 pg_data_t *last_pgdat = NULL;
4107 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4109 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4111 if (!managed_zone(zone))
4113 if (last_pgdat != zone->zone_pgdat) {
4114 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4115 last_pgdat = zone->zone_pgdat;
4120 static inline unsigned int
4121 gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
4123 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4126 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
4127 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4128 * to save two branches.
4130 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
4131 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4134 * The caller may dip into page reserves a bit more if the caller
4135 * cannot run direct reclaim, or if the caller has realtime scheduling
4136 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4137 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
4139 alloc_flags |= (__force int)
4140 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4142 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
4144 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4145 * if it can't schedule.
4147 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
4148 alloc_flags |= ALLOC_NON_BLOCK;
4151 alloc_flags |= ALLOC_HIGHATOMIC;
4155 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
4156 * GFP_ATOMIC) rather than fail, see the comment for
4157 * __cpuset_node_allowed().
4159 if (alloc_flags & ALLOC_MIN_RESERVE)
4160 alloc_flags &= ~ALLOC_CPUSET;
4161 } else if (unlikely(rt_task(current)) && in_task())
4162 alloc_flags |= ALLOC_MIN_RESERVE;
4164 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4169 static bool oom_reserves_allowed(struct task_struct *tsk)
4171 if (!tsk_is_oom_victim(tsk))
4175 * !MMU doesn't have oom reaper so give access to memory reserves
4176 * only to the thread with TIF_MEMDIE set
4178 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4185 * Distinguish requests which really need access to full memory
4186 * reserves from oom victims which can live with a portion of it
4188 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4190 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4192 if (gfp_mask & __GFP_MEMALLOC)
4193 return ALLOC_NO_WATERMARKS;
4194 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4195 return ALLOC_NO_WATERMARKS;
4196 if (!in_interrupt()) {
4197 if (current->flags & PF_MEMALLOC)
4198 return ALLOC_NO_WATERMARKS;
4199 else if (oom_reserves_allowed(current))
4206 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4208 return !!__gfp_pfmemalloc_flags(gfp_mask);
4212 * Checks whether it makes sense to retry the reclaim to make a forward progress
4213 * for the given allocation request.
4215 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4216 * without success, or when we couldn't even meet the watermark if we
4217 * reclaimed all remaining pages on the LRU lists.
4219 * Returns true if a retry is viable or false to enter the oom path.
4222 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4223 struct alloc_context *ac, int alloc_flags,
4224 bool did_some_progress, int *no_progress_loops)
4231 * Costly allocations might have made a progress but this doesn't mean
4232 * their order will become available due to high fragmentation so
4233 * always increment the no progress counter for them
4235 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4236 *no_progress_loops = 0;
4238 (*no_progress_loops)++;
4241 * Make sure we converge to OOM if we cannot make any progress
4242 * several times in the row.
4244 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4245 /* Before OOM, exhaust highatomic_reserve */
4246 return unreserve_highatomic_pageblock(ac, true);
4250 * Keep reclaiming pages while there is a chance this will lead
4251 * somewhere. If none of the target zones can satisfy our allocation
4252 * request even if all reclaimable pages are considered then we are
4253 * screwed and have to go OOM.
4255 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4256 ac->highest_zoneidx, ac->nodemask) {
4257 unsigned long available;
4258 unsigned long reclaimable;
4259 unsigned long min_wmark = min_wmark_pages(zone);
4262 available = reclaimable = zone_reclaimable_pages(zone);
4263 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4266 * Would the allocation succeed if we reclaimed all
4267 * reclaimable pages?
4269 wmark = __zone_watermark_ok(zone, order, min_wmark,
4270 ac->highest_zoneidx, alloc_flags, available);
4271 trace_reclaim_retry_zone(z, order, reclaimable,
4272 available, min_wmark, *no_progress_loops, wmark);
4280 * Memory allocation/reclaim might be called from a WQ context and the
4281 * current implementation of the WQ concurrency control doesn't
4282 * recognize that a particular WQ is congested if the worker thread is
4283 * looping without ever sleeping. Therefore we have to do a short sleep
4284 * here rather than calling cond_resched().
4286 if (current->flags & PF_WQ_WORKER)
4287 schedule_timeout_uninterruptible(1);
4294 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4297 * It's possible that cpuset's mems_allowed and the nodemask from
4298 * mempolicy don't intersect. This should be normally dealt with by
4299 * policy_nodemask(), but it's possible to race with cpuset update in
4300 * such a way the check therein was true, and then it became false
4301 * before we got our cpuset_mems_cookie here.
4302 * This assumes that for all allocations, ac->nodemask can come only
4303 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4304 * when it does not intersect with the cpuset restrictions) or the
4305 * caller can deal with a violated nodemask.
4307 if (cpusets_enabled() && ac->nodemask &&
4308 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4309 ac->nodemask = NULL;
4314 * When updating a task's mems_allowed or mempolicy nodemask, it is
4315 * possible to race with parallel threads in such a way that our
4316 * allocation can fail while the mask is being updated. If we are about
4317 * to fail, check if the cpuset changed during allocation and if so,
4320 if (read_mems_allowed_retry(cpuset_mems_cookie))
4326 static inline struct page *
4327 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4328 struct alloc_context *ac)
4330 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4331 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4332 struct page *page = NULL;
4333 unsigned int alloc_flags;
4334 unsigned long did_some_progress;
4335 enum compact_priority compact_priority;
4336 enum compact_result compact_result;
4337 int compaction_retries;
4338 int no_progress_loops;
4339 unsigned int cpuset_mems_cookie;
4340 unsigned int zonelist_iter_cookie;
4344 compaction_retries = 0;
4345 no_progress_loops = 0;
4346 compact_priority = DEF_COMPACT_PRIORITY;
4347 cpuset_mems_cookie = read_mems_allowed_begin();
4348 zonelist_iter_cookie = zonelist_iter_begin();
4351 * The fast path uses conservative alloc_flags to succeed only until
4352 * kswapd needs to be woken up, and to avoid the cost of setting up
4353 * alloc_flags precisely. So we do that now.
4355 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4358 * We need to recalculate the starting point for the zonelist iterator
4359 * because we might have used different nodemask in the fast path, or
4360 * there was a cpuset modification and we are retrying - otherwise we
4361 * could end up iterating over non-eligible zones endlessly.
4363 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4364 ac->highest_zoneidx, ac->nodemask);
4365 if (!ac->preferred_zoneref->zone)
4369 * Check for insane configurations where the cpuset doesn't contain
4370 * any suitable zone to satisfy the request - e.g. non-movable
4371 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4373 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4374 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4375 ac->highest_zoneidx,
4376 &cpuset_current_mems_allowed);
4381 if (alloc_flags & ALLOC_KSWAPD)
4382 wake_all_kswapds(order, gfp_mask, ac);
4385 * The adjusted alloc_flags might result in immediate success, so try
4388 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4393 * For costly allocations, try direct compaction first, as it's likely
4394 * that we have enough base pages and don't need to reclaim. For non-
4395 * movable high-order allocations, do that as well, as compaction will
4396 * try prevent permanent fragmentation by migrating from blocks of the
4398 * Don't try this for allocations that are allowed to ignore
4399 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4401 if (can_direct_reclaim &&
4403 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4404 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4405 page = __alloc_pages_direct_compact(gfp_mask, order,
4407 INIT_COMPACT_PRIORITY,
4413 * Checks for costly allocations with __GFP_NORETRY, which
4414 * includes some THP page fault allocations
4416 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4418 * If allocating entire pageblock(s) and compaction
4419 * failed because all zones are below low watermarks
4420 * or is prohibited because it recently failed at this
4421 * order, fail immediately unless the allocator has
4422 * requested compaction and reclaim retry.
4425 * - potentially very expensive because zones are far
4426 * below their low watermarks or this is part of very
4427 * bursty high order allocations,
4428 * - not guaranteed to help because isolate_freepages()
4429 * may not iterate over freed pages as part of its
4431 * - unlikely to make entire pageblocks free on its
4434 if (compact_result == COMPACT_SKIPPED ||
4435 compact_result == COMPACT_DEFERRED)
4439 * Looks like reclaim/compaction is worth trying, but
4440 * sync compaction could be very expensive, so keep
4441 * using async compaction.
4443 compact_priority = INIT_COMPACT_PRIORITY;
4448 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4449 if (alloc_flags & ALLOC_KSWAPD)
4450 wake_all_kswapds(order, gfp_mask, ac);
4452 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4454 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4455 (alloc_flags & ALLOC_KSWAPD);
4458 * Reset the nodemask and zonelist iterators if memory policies can be
4459 * ignored. These allocations are high priority and system rather than
4462 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4463 ac->nodemask = NULL;
4464 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4465 ac->highest_zoneidx, ac->nodemask);
4468 /* Attempt with potentially adjusted zonelist and alloc_flags */
4469 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4473 /* Caller is not willing to reclaim, we can't balance anything */
4474 if (!can_direct_reclaim)
4477 /* Avoid recursion of direct reclaim */
4478 if (current->flags & PF_MEMALLOC)
4481 /* Try direct reclaim and then allocating */
4482 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4483 &did_some_progress);
4487 /* Try direct compaction and then allocating */
4488 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4489 compact_priority, &compact_result);
4493 /* Do not loop if specifically requested */
4494 if (gfp_mask & __GFP_NORETRY)
4498 * Do not retry costly high order allocations unless they are
4499 * __GFP_RETRY_MAYFAIL
4501 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4504 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4505 did_some_progress > 0, &no_progress_loops))
4509 * It doesn't make any sense to retry for the compaction if the order-0
4510 * reclaim is not able to make any progress because the current
4511 * implementation of the compaction depends on the sufficient amount
4512 * of free memory (see __compaction_suitable)
4514 if (did_some_progress > 0 &&
4515 should_compact_retry(ac, order, alloc_flags,
4516 compact_result, &compact_priority,
4517 &compaction_retries))
4522 * Deal with possible cpuset update races or zonelist updates to avoid
4523 * a unnecessary OOM kill.
4525 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4526 check_retry_zonelist(zonelist_iter_cookie))
4529 /* Reclaim has failed us, start killing things */
4530 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4534 /* Avoid allocations with no watermarks from looping endlessly */
4535 if (tsk_is_oom_victim(current) &&
4536 (alloc_flags & ALLOC_OOM ||
4537 (gfp_mask & __GFP_NOMEMALLOC)))
4540 /* Retry as long as the OOM killer is making progress */
4541 if (did_some_progress) {
4542 no_progress_loops = 0;
4548 * Deal with possible cpuset update races or zonelist updates to avoid
4549 * a unnecessary OOM kill.
4551 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4552 check_retry_zonelist(zonelist_iter_cookie))
4556 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4559 if (gfp_mask & __GFP_NOFAIL) {
4561 * All existing users of the __GFP_NOFAIL are blockable, so warn
4562 * of any new users that actually require GFP_NOWAIT
4564 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4568 * PF_MEMALLOC request from this context is rather bizarre
4569 * because we cannot reclaim anything and only can loop waiting
4570 * for somebody to do a work for us
4572 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4575 * non failing costly orders are a hard requirement which we
4576 * are not prepared for much so let's warn about these users
4577 * so that we can identify them and convert them to something
4580 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4583 * Help non-failing allocations by giving some access to memory
4584 * reserves normally used for high priority non-blocking
4585 * allocations but do not use ALLOC_NO_WATERMARKS because this
4586 * could deplete whole memory reserves which would just make
4587 * the situation worse.
4589 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4597 warn_alloc(gfp_mask, ac->nodemask,
4598 "page allocation failure: order:%u", order);
4603 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4604 int preferred_nid, nodemask_t *nodemask,
4605 struct alloc_context *ac, gfp_t *alloc_gfp,
4606 unsigned int *alloc_flags)
4608 ac->highest_zoneidx = gfp_zone(gfp_mask);
4609 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4610 ac->nodemask = nodemask;
4611 ac->migratetype = gfp_migratetype(gfp_mask);
4613 if (cpusets_enabled()) {
4614 *alloc_gfp |= __GFP_HARDWALL;
4616 * When we are in the interrupt context, it is irrelevant
4617 * to the current task context. It means that any node ok.
4619 if (in_task() && !ac->nodemask)
4620 ac->nodemask = &cpuset_current_mems_allowed;
4622 *alloc_flags |= ALLOC_CPUSET;
4625 might_alloc(gfp_mask);
4627 if (should_fail_alloc_page(gfp_mask, order))
4630 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4632 /* Dirty zone balancing only done in the fast path */
4633 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4636 * The preferred zone is used for statistics but crucially it is
4637 * also used as the starting point for the zonelist iterator. It
4638 * may get reset for allocations that ignore memory policies.
4640 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4641 ac->highest_zoneidx, ac->nodemask);
4647 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4648 * @gfp: GFP flags for the allocation
4649 * @preferred_nid: The preferred NUMA node ID to allocate from
4650 * @nodemask: Set of nodes to allocate from, may be NULL
4651 * @nr_pages: The number of pages desired on the list or array
4652 * @page_list: Optional list to store the allocated pages
4653 * @page_array: Optional array to store the pages
4655 * This is a batched version of the page allocator that attempts to
4656 * allocate nr_pages quickly. Pages are added to page_list if page_list
4657 * is not NULL, otherwise it is assumed that the page_array is valid.
4659 * For lists, nr_pages is the number of pages that should be allocated.
4661 * For arrays, only NULL elements are populated with pages and nr_pages
4662 * is the maximum number of pages that will be stored in the array.
4664 * Returns the number of pages on the list or array.
4666 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
4667 nodemask_t *nodemask, int nr_pages,
4668 struct list_head *page_list,
4669 struct page **page_array)
4672 unsigned long __maybe_unused UP_flags;
4675 struct per_cpu_pages *pcp;
4676 struct list_head *pcp_list;
4677 struct alloc_context ac;
4679 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4680 int nr_populated = 0, nr_account = 0;
4683 * Skip populated array elements to determine if any pages need
4684 * to be allocated before disabling IRQs.
4686 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4689 /* No pages requested? */
4690 if (unlikely(nr_pages <= 0))
4693 /* Already populated array? */
4694 if (unlikely(page_array && nr_pages - nr_populated == 0))
4697 /* Bulk allocator does not support memcg accounting. */
4698 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4701 /* Use the single page allocator for one page. */
4702 if (nr_pages - nr_populated == 1)
4705 #ifdef CONFIG_PAGE_OWNER
4707 * PAGE_OWNER may recurse into the allocator to allocate space to
4708 * save the stack with pagesets.lock held. Releasing/reacquiring
4709 * removes much of the performance benefit of bulk allocation so
4710 * force the caller to allocate one page at a time as it'll have
4711 * similar performance to added complexity to the bulk allocator.
4713 if (static_branch_unlikely(&page_owner_inited))
4717 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4718 gfp &= gfp_allowed_mask;
4720 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4724 /* Find an allowed local zone that meets the low watermark. */
4725 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4728 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4729 !__cpuset_zone_allowed(zone, gfp)) {
4733 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4734 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4738 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4739 if (zone_watermark_fast(zone, 0, mark,
4740 zonelist_zone_idx(ac.preferred_zoneref),
4741 alloc_flags, gfp)) {
4747 * If there are no allowed local zones that meets the watermarks then
4748 * try to allocate a single page and reclaim if necessary.
4750 if (unlikely(!zone))
4753 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4754 pcp_trylock_prepare(UP_flags);
4755 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4759 /* Attempt the batch allocation */
4760 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4761 while (nr_populated < nr_pages) {
4763 /* Skip existing pages */
4764 if (page_array && page_array[nr_populated]) {
4769 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4771 if (unlikely(!page)) {
4772 /* Try and allocate at least one page */
4774 pcp_spin_unlock(pcp);
4781 prep_new_page(page, 0, gfp, 0);
4783 list_add(&page->lru, page_list);
4785 page_array[nr_populated] = page;
4789 pcp_spin_unlock(pcp);
4790 pcp_trylock_finish(UP_flags);
4792 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4793 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4796 return nr_populated;
4799 pcp_trylock_finish(UP_flags);
4802 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
4805 list_add(&page->lru, page_list);
4807 page_array[nr_populated] = page;
4813 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
4816 * This is the 'heart' of the zoned buddy allocator.
4818 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
4819 nodemask_t *nodemask)
4822 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4823 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4824 struct alloc_context ac = { };
4827 * There are several places where we assume that the order value is sane
4828 * so bail out early if the request is out of bound.
4830 if (WARN_ON_ONCE_GFP(order > MAX_ORDER, gfp))
4833 gfp &= gfp_allowed_mask;
4835 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4836 * resp. GFP_NOIO which has to be inherited for all allocation requests
4837 * from a particular context which has been marked by
4838 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4839 * movable zones are not used during allocation.
4841 gfp = current_gfp_context(gfp);
4843 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4844 &alloc_gfp, &alloc_flags))
4848 * Forbid the first pass from falling back to types that fragment
4849 * memory until all local zones are considered.
4851 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4853 /* First allocation attempt */
4854 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4859 ac.spread_dirty_pages = false;
4862 * Restore the original nodemask if it was potentially replaced with
4863 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4865 ac.nodemask = nodemask;
4867 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4870 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4871 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4872 __free_pages(page, order);
4876 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4877 kmsan_alloc_page(page, order, alloc_gfp);
4881 EXPORT_SYMBOL(__alloc_pages);
4883 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
4884 nodemask_t *nodemask)
4886 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
4887 preferred_nid, nodemask);
4889 if (page && order > 1)
4890 prep_transhuge_page(page);
4891 return (struct folio *)page;
4893 EXPORT_SYMBOL(__folio_alloc);
4896 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4897 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4898 * you need to access high mem.
4900 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4904 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4907 return (unsigned long) page_address(page);
4909 EXPORT_SYMBOL(__get_free_pages);
4911 unsigned long get_zeroed_page(gfp_t gfp_mask)
4913 return __get_free_page(gfp_mask | __GFP_ZERO);
4915 EXPORT_SYMBOL(get_zeroed_page);
4918 * __free_pages - Free pages allocated with alloc_pages().
4919 * @page: The page pointer returned from alloc_pages().
4920 * @order: The order of the allocation.
4922 * This function can free multi-page allocations that are not compound
4923 * pages. It does not check that the @order passed in matches that of
4924 * the allocation, so it is easy to leak memory. Freeing more memory
4925 * than was allocated will probably emit a warning.
4927 * If the last reference to this page is speculative, it will be released
4928 * by put_page() which only frees the first page of a non-compound
4929 * allocation. To prevent the remaining pages from being leaked, we free
4930 * the subsequent pages here. If you want to use the page's reference
4931 * count to decide when to free the allocation, you should allocate a
4932 * compound page, and use put_page() instead of __free_pages().
4934 * Context: May be called in interrupt context or while holding a normal
4935 * spinlock, but not in NMI context or while holding a raw spinlock.
4937 void __free_pages(struct page *page, unsigned int order)
4939 /* get PageHead before we drop reference */
4940 int head = PageHead(page);
4942 if (put_page_testzero(page))
4943 free_the_page(page, order);
4946 free_the_page(page + (1 << order), order);
4948 EXPORT_SYMBOL(__free_pages);
4950 void free_pages(unsigned long addr, unsigned int order)
4953 VM_BUG_ON(!virt_addr_valid((void *)addr));
4954 __free_pages(virt_to_page((void *)addr), order);
4958 EXPORT_SYMBOL(free_pages);
4962 * An arbitrary-length arbitrary-offset area of memory which resides
4963 * within a 0 or higher order page. Multiple fragments within that page
4964 * are individually refcounted, in the page's reference counter.
4966 * The page_frag functions below provide a simple allocation framework for
4967 * page fragments. This is used by the network stack and network device
4968 * drivers to provide a backing region of memory for use as either an
4969 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4971 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4974 struct page *page = NULL;
4975 gfp_t gfp = gfp_mask;
4977 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4978 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4980 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4981 PAGE_FRAG_CACHE_MAX_ORDER);
4982 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4984 if (unlikely(!page))
4985 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4987 nc->va = page ? page_address(page) : NULL;
4992 void __page_frag_cache_drain(struct page *page, unsigned int count)
4994 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4996 if (page_ref_sub_and_test(page, count))
4997 free_the_page(page, compound_order(page));
4999 EXPORT_SYMBOL(__page_frag_cache_drain);
5001 void *page_frag_alloc_align(struct page_frag_cache *nc,
5002 unsigned int fragsz, gfp_t gfp_mask,
5003 unsigned int align_mask)
5005 unsigned int size = PAGE_SIZE;
5009 if (unlikely(!nc->va)) {
5011 page = __page_frag_cache_refill(nc, gfp_mask);
5015 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5016 /* if size can vary use size else just use PAGE_SIZE */
5019 /* Even if we own the page, we do not use atomic_set().
5020 * This would break get_page_unless_zero() users.
5022 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5024 /* reset page count bias and offset to start of new frag */
5025 nc->pfmemalloc = page_is_pfmemalloc(page);
5026 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5030 offset = nc->offset - fragsz;
5031 if (unlikely(offset < 0)) {
5032 page = virt_to_page(nc->va);
5034 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5037 if (unlikely(nc->pfmemalloc)) {
5038 free_the_page(page, compound_order(page));
5042 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5043 /* if size can vary use size else just use PAGE_SIZE */
5046 /* OK, page count is 0, we can safely set it */
5047 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5049 /* reset page count bias and offset to start of new frag */
5050 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5051 offset = size - fragsz;
5052 if (unlikely(offset < 0)) {
5054 * The caller is trying to allocate a fragment
5055 * with fragsz > PAGE_SIZE but the cache isn't big
5056 * enough to satisfy the request, this may
5057 * happen in low memory conditions.
5058 * We don't release the cache page because
5059 * it could make memory pressure worse
5060 * so we simply return NULL here.
5067 offset &= align_mask;
5068 nc->offset = offset;
5070 return nc->va + offset;
5072 EXPORT_SYMBOL(page_frag_alloc_align);
5075 * Frees a page fragment allocated out of either a compound or order 0 page.
5077 void page_frag_free(void *addr)
5079 struct page *page = virt_to_head_page(addr);
5081 if (unlikely(put_page_testzero(page)))
5082 free_the_page(page, compound_order(page));
5084 EXPORT_SYMBOL(page_frag_free);
5086 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5090 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
5091 struct page *page = virt_to_page((void *)addr);
5092 struct page *last = page + nr;
5094 split_page_owner(page, 1 << order);
5095 split_page_memcg(page, 1 << order);
5096 while (page < --last)
5097 set_page_refcounted(last);
5099 last = page + (1UL << order);
5100 for (page += nr; page < last; page++)
5101 __free_pages_ok(page, 0, FPI_TO_TAIL);
5103 return (void *)addr;
5107 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5108 * @size: the number of bytes to allocate
5109 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5111 * This function is similar to alloc_pages(), except that it allocates the
5112 * minimum number of pages to satisfy the request. alloc_pages() can only
5113 * allocate memory in power-of-two pages.
5115 * This function is also limited by MAX_ORDER.
5117 * Memory allocated by this function must be released by free_pages_exact().
5119 * Return: pointer to the allocated area or %NULL in case of error.
5121 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5123 unsigned int order = get_order(size);
5126 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5127 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5129 addr = __get_free_pages(gfp_mask, order);
5130 return make_alloc_exact(addr, order, size);
5132 EXPORT_SYMBOL(alloc_pages_exact);
5135 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5137 * @nid: the preferred node ID where memory should be allocated
5138 * @size: the number of bytes to allocate
5139 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5141 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5144 * Return: pointer to the allocated area or %NULL in case of error.
5146 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5148 unsigned int order = get_order(size);
5151 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5152 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5154 p = alloc_pages_node(nid, gfp_mask, order);
5157 return make_alloc_exact((unsigned long)page_address(p), order, size);
5161 * free_pages_exact - release memory allocated via alloc_pages_exact()
5162 * @virt: the value returned by alloc_pages_exact.
5163 * @size: size of allocation, same value as passed to alloc_pages_exact().
5165 * Release the memory allocated by a previous call to alloc_pages_exact.
5167 void free_pages_exact(void *virt, size_t size)
5169 unsigned long addr = (unsigned long)virt;
5170 unsigned long end = addr + PAGE_ALIGN(size);
5172 while (addr < end) {
5177 EXPORT_SYMBOL(free_pages_exact);
5180 * nr_free_zone_pages - count number of pages beyond high watermark
5181 * @offset: The zone index of the highest zone
5183 * nr_free_zone_pages() counts the number of pages which are beyond the
5184 * high watermark within all zones at or below a given zone index. For each
5185 * zone, the number of pages is calculated as:
5187 * nr_free_zone_pages = managed_pages - high_pages
5189 * Return: number of pages beyond high watermark.
5191 static unsigned long nr_free_zone_pages(int offset)
5196 /* Just pick one node, since fallback list is circular */
5197 unsigned long sum = 0;
5199 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5201 for_each_zone_zonelist(zone, z, zonelist, offset) {
5202 unsigned long size = zone_managed_pages(zone);
5203 unsigned long high = high_wmark_pages(zone);
5212 * nr_free_buffer_pages - count number of pages beyond high watermark
5214 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5215 * watermark within ZONE_DMA and ZONE_NORMAL.
5217 * Return: number of pages beyond high watermark within ZONE_DMA and
5220 unsigned long nr_free_buffer_pages(void)
5222 return nr_free_zone_pages(gfp_zone(GFP_USER));
5224 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5226 static inline void show_node(struct zone *zone)
5228 if (IS_ENABLED(CONFIG_NUMA))
5229 printk("Node %d ", zone_to_nid(zone));
5232 long si_mem_available(void)
5235 unsigned long pagecache;
5236 unsigned long wmark_low = 0;
5237 unsigned long pages[NR_LRU_LISTS];
5238 unsigned long reclaimable;
5242 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5243 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5246 wmark_low += low_wmark_pages(zone);
5249 * Estimate the amount of memory available for userspace allocations,
5250 * without causing swapping or OOM.
5252 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5255 * Not all the page cache can be freed, otherwise the system will
5256 * start swapping or thrashing. Assume at least half of the page
5257 * cache, or the low watermark worth of cache, needs to stay.
5259 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5260 pagecache -= min(pagecache / 2, wmark_low);
5261 available += pagecache;
5264 * Part of the reclaimable slab and other kernel memory consists of
5265 * items that are in use, and cannot be freed. Cap this estimate at the
5268 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5269 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5270 available += reclaimable - min(reclaimable / 2, wmark_low);
5276 EXPORT_SYMBOL_GPL(si_mem_available);
5278 void si_meminfo(struct sysinfo *val)
5280 val->totalram = totalram_pages();
5281 val->sharedram = global_node_page_state(NR_SHMEM);
5282 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5283 val->bufferram = nr_blockdev_pages();
5284 val->totalhigh = totalhigh_pages();
5285 val->freehigh = nr_free_highpages();
5286 val->mem_unit = PAGE_SIZE;
5289 EXPORT_SYMBOL(si_meminfo);
5292 void si_meminfo_node(struct sysinfo *val, int nid)
5294 int zone_type; /* needs to be signed */
5295 unsigned long managed_pages = 0;
5296 unsigned long managed_highpages = 0;
5297 unsigned long free_highpages = 0;
5298 pg_data_t *pgdat = NODE_DATA(nid);
5300 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5301 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5302 val->totalram = managed_pages;
5303 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5304 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5305 #ifdef CONFIG_HIGHMEM
5306 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5307 struct zone *zone = &pgdat->node_zones[zone_type];
5309 if (is_highmem(zone)) {
5310 managed_highpages += zone_managed_pages(zone);
5311 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5314 val->totalhigh = managed_highpages;
5315 val->freehigh = free_highpages;
5317 val->totalhigh = managed_highpages;
5318 val->freehigh = free_highpages;
5320 val->mem_unit = PAGE_SIZE;
5325 * Determine whether the node should be displayed or not, depending on whether
5326 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5328 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5330 if (!(flags & SHOW_MEM_FILTER_NODES))
5334 * no node mask - aka implicit memory numa policy. Do not bother with
5335 * the synchronization - read_mems_allowed_begin - because we do not
5336 * have to be precise here.
5339 nodemask = &cpuset_current_mems_allowed;
5341 return !node_isset(nid, *nodemask);
5344 #define K(x) ((x) << (PAGE_SHIFT-10))
5346 static void show_migration_types(unsigned char type)
5348 static const char types[MIGRATE_TYPES] = {
5349 [MIGRATE_UNMOVABLE] = 'U',
5350 [MIGRATE_MOVABLE] = 'M',
5351 [MIGRATE_RECLAIMABLE] = 'E',
5352 [MIGRATE_HIGHATOMIC] = 'H',
5354 [MIGRATE_CMA] = 'C',
5356 #ifdef CONFIG_MEMORY_ISOLATION
5357 [MIGRATE_ISOLATE] = 'I',
5360 char tmp[MIGRATE_TYPES + 1];
5364 for (i = 0; i < MIGRATE_TYPES; i++) {
5365 if (type & (1 << i))
5370 printk(KERN_CONT "(%s) ", tmp);
5373 static bool node_has_managed_zones(pg_data_t *pgdat, int max_zone_idx)
5376 for (zone_idx = 0; zone_idx <= max_zone_idx; zone_idx++)
5377 if (zone_managed_pages(pgdat->node_zones + zone_idx))
5383 * Show free area list (used inside shift_scroll-lock stuff)
5384 * We also calculate the percentage fragmentation. We do this by counting the
5385 * memory on each free list with the exception of the first item on the list.
5388 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5391 void __show_free_areas(unsigned int filter, nodemask_t *nodemask, int max_zone_idx)
5393 unsigned long free_pcp = 0;
5398 for_each_populated_zone(zone) {
5399 if (zone_idx(zone) > max_zone_idx)
5401 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5404 for_each_online_cpu(cpu)
5405 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5408 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5409 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5410 " unevictable:%lu dirty:%lu writeback:%lu\n"
5411 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5412 " mapped:%lu shmem:%lu pagetables:%lu\n"
5413 " sec_pagetables:%lu bounce:%lu\n"
5414 " kernel_misc_reclaimable:%lu\n"
5415 " free:%lu free_pcp:%lu free_cma:%lu\n",
5416 global_node_page_state(NR_ACTIVE_ANON),
5417 global_node_page_state(NR_INACTIVE_ANON),
5418 global_node_page_state(NR_ISOLATED_ANON),
5419 global_node_page_state(NR_ACTIVE_FILE),
5420 global_node_page_state(NR_INACTIVE_FILE),
5421 global_node_page_state(NR_ISOLATED_FILE),
5422 global_node_page_state(NR_UNEVICTABLE),
5423 global_node_page_state(NR_FILE_DIRTY),
5424 global_node_page_state(NR_WRITEBACK),
5425 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5426 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5427 global_node_page_state(NR_FILE_MAPPED),
5428 global_node_page_state(NR_SHMEM),
5429 global_node_page_state(NR_PAGETABLE),
5430 global_node_page_state(NR_SECONDARY_PAGETABLE),
5431 global_zone_page_state(NR_BOUNCE),
5432 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5433 global_zone_page_state(NR_FREE_PAGES),
5435 global_zone_page_state(NR_FREE_CMA_PAGES));
5437 for_each_online_pgdat(pgdat) {
5438 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5440 if (!node_has_managed_zones(pgdat, max_zone_idx))
5444 " active_anon:%lukB"
5445 " inactive_anon:%lukB"
5446 " active_file:%lukB"
5447 " inactive_file:%lukB"
5448 " unevictable:%lukB"
5449 " isolated(anon):%lukB"
5450 " isolated(file):%lukB"
5455 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5457 " shmem_pmdmapped: %lukB"
5460 " writeback_tmp:%lukB"
5461 " kernel_stack:%lukB"
5462 #ifdef CONFIG_SHADOW_CALL_STACK
5463 " shadow_call_stack:%lukB"
5466 " sec_pagetables:%lukB"
5467 " all_unreclaimable? %s"
5470 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5471 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5472 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5473 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5474 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5475 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5476 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5477 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5478 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5479 K(node_page_state(pgdat, NR_WRITEBACK)),
5480 K(node_page_state(pgdat, NR_SHMEM)),
5481 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5482 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5483 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5484 K(node_page_state(pgdat, NR_ANON_THPS)),
5486 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5487 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5488 #ifdef CONFIG_SHADOW_CALL_STACK
5489 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5491 K(node_page_state(pgdat, NR_PAGETABLE)),
5492 K(node_page_state(pgdat, NR_SECONDARY_PAGETABLE)),
5493 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5497 for_each_populated_zone(zone) {
5500 if (zone_idx(zone) > max_zone_idx)
5502 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5506 for_each_online_cpu(cpu)
5507 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5517 " reserved_highatomic:%luKB"
5518 " active_anon:%lukB"
5519 " inactive_anon:%lukB"
5520 " active_file:%lukB"
5521 " inactive_file:%lukB"
5522 " unevictable:%lukB"
5523 " writepending:%lukB"
5533 K(zone_page_state(zone, NR_FREE_PAGES)),
5534 K(zone->watermark_boost),
5535 K(min_wmark_pages(zone)),
5536 K(low_wmark_pages(zone)),
5537 K(high_wmark_pages(zone)),
5538 K(zone->nr_reserved_highatomic),
5539 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5540 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5541 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5542 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5543 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5544 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5545 K(zone->present_pages),
5546 K(zone_managed_pages(zone)),
5547 K(zone_page_state(zone, NR_MLOCK)),
5548 K(zone_page_state(zone, NR_BOUNCE)),
5550 K(this_cpu_read(zone->per_cpu_pageset->count)),
5551 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5552 printk("lowmem_reserve[]:");
5553 for (i = 0; i < MAX_NR_ZONES; i++)
5554 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5555 printk(KERN_CONT "\n");
5558 for_each_populated_zone(zone) {
5560 unsigned long nr[MAX_ORDER + 1], flags, total = 0;
5561 unsigned char types[MAX_ORDER + 1];
5563 if (zone_idx(zone) > max_zone_idx)
5565 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5568 printk(KERN_CONT "%s: ", zone->name);
5570 spin_lock_irqsave(&zone->lock, flags);
5571 for (order = 0; order <= MAX_ORDER; order++) {
5572 struct free_area *area = &zone->free_area[order];
5575 nr[order] = area->nr_free;
5576 total += nr[order] << order;
5579 for (type = 0; type < MIGRATE_TYPES; type++) {
5580 if (!free_area_empty(area, type))
5581 types[order] |= 1 << type;
5584 spin_unlock_irqrestore(&zone->lock, flags);
5585 for (order = 0; order <= MAX_ORDER; order++) {
5586 printk(KERN_CONT "%lu*%lukB ",
5587 nr[order], K(1UL) << order);
5589 show_migration_types(types[order]);
5591 printk(KERN_CONT "= %lukB\n", K(total));
5594 for_each_online_node(nid) {
5595 if (show_mem_node_skip(filter, nid, nodemask))
5597 hugetlb_show_meminfo_node(nid);
5600 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5602 show_swap_cache_info();
5605 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5607 zoneref->zone = zone;
5608 zoneref->zone_idx = zone_idx(zone);
5612 * Builds allocation fallback zone lists.
5614 * Add all populated zones of a node to the zonelist.
5616 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5619 enum zone_type zone_type = MAX_NR_ZONES;
5624 zone = pgdat->node_zones + zone_type;
5625 if (populated_zone(zone)) {
5626 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5627 check_highest_zone(zone_type);
5629 } while (zone_type);
5636 static int __parse_numa_zonelist_order(char *s)
5639 * We used to support different zonelists modes but they turned
5640 * out to be just not useful. Let's keep the warning in place
5641 * if somebody still use the cmd line parameter so that we do
5642 * not fail it silently
5644 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5645 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5651 char numa_zonelist_order[] = "Node";
5654 * sysctl handler for numa_zonelist_order
5656 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5657 void *buffer, size_t *length, loff_t *ppos)
5660 return __parse_numa_zonelist_order(buffer);
5661 return proc_dostring(table, write, buffer, length, ppos);
5665 static int node_load[MAX_NUMNODES];
5668 * find_next_best_node - find the next node that should appear in a given node's fallback list
5669 * @node: node whose fallback list we're appending
5670 * @used_node_mask: nodemask_t of already used nodes
5672 * We use a number of factors to determine which is the next node that should
5673 * appear on a given node's fallback list. The node should not have appeared
5674 * already in @node's fallback list, and it should be the next closest node
5675 * according to the distance array (which contains arbitrary distance values
5676 * from each node to each node in the system), and should also prefer nodes
5677 * with no CPUs, since presumably they'll have very little allocation pressure
5678 * on them otherwise.
5680 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5682 int find_next_best_node(int node, nodemask_t *used_node_mask)
5685 int min_val = INT_MAX;
5686 int best_node = NUMA_NO_NODE;
5688 /* Use the local node if we haven't already */
5689 if (!node_isset(node, *used_node_mask)) {
5690 node_set(node, *used_node_mask);
5694 for_each_node_state(n, N_MEMORY) {
5696 /* Don't want a node to appear more than once */
5697 if (node_isset(n, *used_node_mask))
5700 /* Use the distance array to find the distance */
5701 val = node_distance(node, n);
5703 /* Penalize nodes under us ("prefer the next node") */
5706 /* Give preference to headless and unused nodes */
5707 if (!cpumask_empty(cpumask_of_node(n)))
5708 val += PENALTY_FOR_NODE_WITH_CPUS;
5710 /* Slight preference for less loaded node */
5711 val *= MAX_NUMNODES;
5712 val += node_load[n];
5714 if (val < min_val) {
5721 node_set(best_node, *used_node_mask);
5728 * Build zonelists ordered by node and zones within node.
5729 * This results in maximum locality--normal zone overflows into local
5730 * DMA zone, if any--but risks exhausting DMA zone.
5732 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5735 struct zoneref *zonerefs;
5738 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5740 for (i = 0; i < nr_nodes; i++) {
5743 pg_data_t *node = NODE_DATA(node_order[i]);
5745 nr_zones = build_zonerefs_node(node, zonerefs);
5746 zonerefs += nr_zones;
5748 zonerefs->zone = NULL;
5749 zonerefs->zone_idx = 0;
5753 * Build gfp_thisnode zonelists
5755 static void build_thisnode_zonelists(pg_data_t *pgdat)
5757 struct zoneref *zonerefs;
5760 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5761 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5762 zonerefs += nr_zones;
5763 zonerefs->zone = NULL;
5764 zonerefs->zone_idx = 0;
5768 * Build zonelists ordered by zone and nodes within zones.
5769 * This results in conserving DMA zone[s] until all Normal memory is
5770 * exhausted, but results in overflowing to remote node while memory
5771 * may still exist in local DMA zone.
5774 static void build_zonelists(pg_data_t *pgdat)
5776 static int node_order[MAX_NUMNODES];
5777 int node, nr_nodes = 0;
5778 nodemask_t used_mask = NODE_MASK_NONE;
5779 int local_node, prev_node;
5781 /* NUMA-aware ordering of nodes */
5782 local_node = pgdat->node_id;
5783 prev_node = local_node;
5785 memset(node_order, 0, sizeof(node_order));
5786 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5788 * We don't want to pressure a particular node.
5789 * So adding penalty to the first node in same
5790 * distance group to make it round-robin.
5792 if (node_distance(local_node, node) !=
5793 node_distance(local_node, prev_node))
5794 node_load[node] += 1;
5796 node_order[nr_nodes++] = node;
5800 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5801 build_thisnode_zonelists(pgdat);
5802 pr_info("Fallback order for Node %d: ", local_node);
5803 for (node = 0; node < nr_nodes; node++)
5804 pr_cont("%d ", node_order[node]);
5808 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5810 * Return node id of node used for "local" allocations.
5811 * I.e., first node id of first zone in arg node's generic zonelist.
5812 * Used for initializing percpu 'numa_mem', which is used primarily
5813 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5815 int local_memory_node(int node)
5819 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5820 gfp_zone(GFP_KERNEL),
5822 return zone_to_nid(z->zone);
5826 static void setup_min_unmapped_ratio(void);
5827 static void setup_min_slab_ratio(void);
5828 #else /* CONFIG_NUMA */
5830 static void build_zonelists(pg_data_t *pgdat)
5832 int node, local_node;
5833 struct zoneref *zonerefs;
5836 local_node = pgdat->node_id;
5838 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5839 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5840 zonerefs += nr_zones;
5843 * Now we build the zonelist so that it contains the zones
5844 * of all the other nodes.
5845 * We don't want to pressure a particular node, so when
5846 * building the zones for node N, we make sure that the
5847 * zones coming right after the local ones are those from
5848 * node N+1 (modulo N)
5850 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5851 if (!node_online(node))
5853 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5854 zonerefs += nr_zones;
5856 for (node = 0; node < local_node; node++) {
5857 if (!node_online(node))
5859 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5860 zonerefs += nr_zones;
5863 zonerefs->zone = NULL;
5864 zonerefs->zone_idx = 0;
5867 #endif /* CONFIG_NUMA */
5870 * Boot pageset table. One per cpu which is going to be used for all
5871 * zones and all nodes. The parameters will be set in such a way
5872 * that an item put on a list will immediately be handed over to
5873 * the buddy list. This is safe since pageset manipulation is done
5874 * with interrupts disabled.
5876 * The boot_pagesets must be kept even after bootup is complete for
5877 * unused processors and/or zones. They do play a role for bootstrapping
5878 * hotplugged processors.
5880 * zoneinfo_show() and maybe other functions do
5881 * not check if the processor is online before following the pageset pointer.
5882 * Other parts of the kernel may not check if the zone is available.
5884 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5885 /* These effectively disable the pcplists in the boot pageset completely */
5886 #define BOOT_PAGESET_HIGH 0
5887 #define BOOT_PAGESET_BATCH 1
5888 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5889 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5891 static void __build_all_zonelists(void *data)
5894 int __maybe_unused cpu;
5895 pg_data_t *self = data;
5897 write_seqlock(&zonelist_update_seq);
5900 memset(node_load, 0, sizeof(node_load));
5904 * This node is hotadded and no memory is yet present. So just
5905 * building zonelists is fine - no need to touch other nodes.
5907 if (self && !node_online(self->node_id)) {
5908 build_zonelists(self);
5911 * All possible nodes have pgdat preallocated
5914 for_each_node(nid) {
5915 pg_data_t *pgdat = NODE_DATA(nid);
5917 build_zonelists(pgdat);
5920 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5922 * We now know the "local memory node" for each node--
5923 * i.e., the node of the first zone in the generic zonelist.
5924 * Set up numa_mem percpu variable for on-line cpus. During
5925 * boot, only the boot cpu should be on-line; we'll init the
5926 * secondary cpus' numa_mem as they come on-line. During
5927 * node/memory hotplug, we'll fixup all on-line cpus.
5929 for_each_online_cpu(cpu)
5930 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5934 write_sequnlock(&zonelist_update_seq);
5937 static noinline void __init
5938 build_all_zonelists_init(void)
5942 __build_all_zonelists(NULL);
5945 * Initialize the boot_pagesets that are going to be used
5946 * for bootstrapping processors. The real pagesets for
5947 * each zone will be allocated later when the per cpu
5948 * allocator is available.
5950 * boot_pagesets are used also for bootstrapping offline
5951 * cpus if the system is already booted because the pagesets
5952 * are needed to initialize allocators on a specific cpu too.
5953 * F.e. the percpu allocator needs the page allocator which
5954 * needs the percpu allocator in order to allocate its pagesets
5955 * (a chicken-egg dilemma).
5957 for_each_possible_cpu(cpu)
5958 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5960 mminit_verify_zonelist();
5961 cpuset_init_current_mems_allowed();
5965 * unless system_state == SYSTEM_BOOTING.
5967 * __ref due to call of __init annotated helper build_all_zonelists_init
5968 * [protected by SYSTEM_BOOTING].
5970 void __ref build_all_zonelists(pg_data_t *pgdat)
5972 unsigned long vm_total_pages;
5974 if (system_state == SYSTEM_BOOTING) {
5975 build_all_zonelists_init();
5977 __build_all_zonelists(pgdat);
5978 /* cpuset refresh routine should be here */
5980 /* Get the number of free pages beyond high watermark in all zones. */
5981 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5983 * Disable grouping by mobility if the number of pages in the
5984 * system is too low to allow the mechanism to work. It would be
5985 * more accurate, but expensive to check per-zone. This check is
5986 * made on memory-hotadd so a system can start with mobility
5987 * disabled and enable it later
5989 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5990 page_group_by_mobility_disabled = 1;
5992 page_group_by_mobility_disabled = 0;
5994 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5996 page_group_by_mobility_disabled ? "off" : "on",
5999 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6003 static int zone_batchsize(struct zone *zone)
6009 * The number of pages to batch allocate is either ~0.1%
6010 * of the zone or 1MB, whichever is smaller. The batch
6011 * size is striking a balance between allocation latency
6012 * and zone lock contention.
6014 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
6015 batch /= 4; /* We effectively *= 4 below */
6020 * Clamp the batch to a 2^n - 1 value. Having a power
6021 * of 2 value was found to be more likely to have
6022 * suboptimal cache aliasing properties in some cases.
6024 * For example if 2 tasks are alternately allocating
6025 * batches of pages, one task can end up with a lot
6026 * of pages of one half of the possible page colors
6027 * and the other with pages of the other colors.
6029 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6034 /* The deferral and batching of frees should be suppressed under NOMMU
6037 * The problem is that NOMMU needs to be able to allocate large chunks
6038 * of contiguous memory as there's no hardware page translation to
6039 * assemble apparent contiguous memory from discontiguous pages.
6041 * Queueing large contiguous runs of pages for batching, however,
6042 * causes the pages to actually be freed in smaller chunks. As there
6043 * can be a significant delay between the individual batches being
6044 * recycled, this leads to the once large chunks of space being
6045 * fragmented and becoming unavailable for high-order allocations.
6051 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6056 unsigned long total_pages;
6058 if (!percpu_pagelist_high_fraction) {
6060 * By default, the high value of the pcp is based on the zone
6061 * low watermark so that if they are full then background
6062 * reclaim will not be started prematurely.
6064 total_pages = low_wmark_pages(zone);
6067 * If percpu_pagelist_high_fraction is configured, the high
6068 * value is based on a fraction of the managed pages in the
6071 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6075 * Split the high value across all online CPUs local to the zone. Note
6076 * that early in boot that CPUs may not be online yet and that during
6077 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6078 * onlined. For memory nodes that have no CPUs, split pcp->high across
6079 * all online CPUs to mitigate the risk that reclaim is triggered
6080 * prematurely due to pages stored on pcp lists.
6082 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6084 nr_split_cpus = num_online_cpus();
6085 high = total_pages / nr_split_cpus;
6088 * Ensure high is at least batch*4. The multiple is based on the
6089 * historical relationship between high and batch.
6091 high = max(high, batch << 2);
6100 * pcp->high and pcp->batch values are related and generally batch is lower
6101 * than high. They are also related to pcp->count such that count is lower
6102 * than high, and as soon as it reaches high, the pcplist is flushed.
6104 * However, guaranteeing these relations at all times would require e.g. write
6105 * barriers here but also careful usage of read barriers at the read side, and
6106 * thus be prone to error and bad for performance. Thus the update only prevents
6107 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6108 * can cope with those fields changing asynchronously, and fully trust only the
6109 * pcp->count field on the local CPU with interrupts disabled.
6111 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6112 * outside of boot time (or some other assurance that no concurrent updaters
6115 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6116 unsigned long batch)
6118 WRITE_ONCE(pcp->batch, batch);
6119 WRITE_ONCE(pcp->high, high);
6122 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6126 memset(pcp, 0, sizeof(*pcp));
6127 memset(pzstats, 0, sizeof(*pzstats));
6129 spin_lock_init(&pcp->lock);
6130 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6131 INIT_LIST_HEAD(&pcp->lists[pindex]);
6134 * Set batch and high values safe for a boot pageset. A true percpu
6135 * pageset's initialization will update them subsequently. Here we don't
6136 * need to be as careful as pageset_update() as nobody can access the
6139 pcp->high = BOOT_PAGESET_HIGH;
6140 pcp->batch = BOOT_PAGESET_BATCH;
6141 pcp->free_factor = 0;
6144 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6145 unsigned long batch)
6147 struct per_cpu_pages *pcp;
6150 for_each_possible_cpu(cpu) {
6151 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6152 pageset_update(pcp, high, batch);
6157 * Calculate and set new high and batch values for all per-cpu pagesets of a
6158 * zone based on the zone's size.
6160 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6162 int new_high, new_batch;
6164 new_batch = max(1, zone_batchsize(zone));
6165 new_high = zone_highsize(zone, new_batch, cpu_online);
6167 if (zone->pageset_high == new_high &&
6168 zone->pageset_batch == new_batch)
6171 zone->pageset_high = new_high;
6172 zone->pageset_batch = new_batch;
6174 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6177 void __meminit setup_zone_pageset(struct zone *zone)
6181 /* Size may be 0 on !SMP && !NUMA */
6182 if (sizeof(struct per_cpu_zonestat) > 0)
6183 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6185 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6186 for_each_possible_cpu(cpu) {
6187 struct per_cpu_pages *pcp;
6188 struct per_cpu_zonestat *pzstats;
6190 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6191 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6192 per_cpu_pages_init(pcp, pzstats);
6195 zone_set_pageset_high_and_batch(zone, 0);
6199 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6200 * page high values need to be recalculated.
6202 static void zone_pcp_update(struct zone *zone, int cpu_online)
6204 mutex_lock(&pcp_batch_high_lock);
6205 zone_set_pageset_high_and_batch(zone, cpu_online);
6206 mutex_unlock(&pcp_batch_high_lock);
6210 * Allocate per cpu pagesets and initialize them.
6211 * Before this call only boot pagesets were available.
6213 void __init setup_per_cpu_pageset(void)
6215 struct pglist_data *pgdat;
6217 int __maybe_unused cpu;
6219 for_each_populated_zone(zone)
6220 setup_zone_pageset(zone);
6224 * Unpopulated zones continue using the boot pagesets.
6225 * The numa stats for these pagesets need to be reset.
6226 * Otherwise, they will end up skewing the stats of
6227 * the nodes these zones are associated with.
6229 for_each_possible_cpu(cpu) {
6230 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
6231 memset(pzstats->vm_numa_event, 0,
6232 sizeof(pzstats->vm_numa_event));
6236 for_each_online_pgdat(pgdat)
6237 pgdat->per_cpu_nodestats =
6238 alloc_percpu(struct per_cpu_nodestat);
6241 __meminit void zone_pcp_init(struct zone *zone)
6244 * per cpu subsystem is not up at this point. The following code
6245 * relies on the ability of the linker to provide the
6246 * offset of a (static) per cpu variable into the per cpu area.
6248 zone->per_cpu_pageset = &boot_pageset;
6249 zone->per_cpu_zonestats = &boot_zonestats;
6250 zone->pageset_high = BOOT_PAGESET_HIGH;
6251 zone->pageset_batch = BOOT_PAGESET_BATCH;
6253 if (populated_zone(zone))
6254 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
6255 zone->present_pages, zone_batchsize(zone));
6258 void adjust_managed_page_count(struct page *page, long count)
6260 atomic_long_add(count, &page_zone(page)->managed_pages);
6261 totalram_pages_add(count);
6262 #ifdef CONFIG_HIGHMEM
6263 if (PageHighMem(page))
6264 totalhigh_pages_add(count);
6267 EXPORT_SYMBOL(adjust_managed_page_count);
6269 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
6272 unsigned long pages = 0;
6274 start = (void *)PAGE_ALIGN((unsigned long)start);
6275 end = (void *)((unsigned long)end & PAGE_MASK);
6276 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6277 struct page *page = virt_to_page(pos);
6278 void *direct_map_addr;
6281 * 'direct_map_addr' might be different from 'pos'
6282 * because some architectures' virt_to_page()
6283 * work with aliases. Getting the direct map
6284 * address ensures that we get a _writeable_
6285 * alias for the memset().
6287 direct_map_addr = page_address(page);
6289 * Perform a kasan-unchecked memset() since this memory
6290 * has not been initialized.
6292 direct_map_addr = kasan_reset_tag(direct_map_addr);
6293 if ((unsigned int)poison <= 0xFF)
6294 memset(direct_map_addr, poison, PAGE_SIZE);
6296 free_reserved_page(page);
6300 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
6305 void __init mem_init_print_info(void)
6307 unsigned long physpages, codesize, datasize, rosize, bss_size;
6308 unsigned long init_code_size, init_data_size;
6310 physpages = get_num_physpages();
6311 codesize = _etext - _stext;
6312 datasize = _edata - _sdata;
6313 rosize = __end_rodata - __start_rodata;
6314 bss_size = __bss_stop - __bss_start;
6315 init_data_size = __init_end - __init_begin;
6316 init_code_size = _einittext - _sinittext;
6319 * Detect special cases and adjust section sizes accordingly:
6320 * 1) .init.* may be embedded into .data sections
6321 * 2) .init.text.* may be out of [__init_begin, __init_end],
6322 * please refer to arch/tile/kernel/vmlinux.lds.S.
6323 * 3) .rodata.* may be embedded into .text or .data sections.
6325 #define adj_init_size(start, end, size, pos, adj) \
6327 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
6331 adj_init_size(__init_begin, __init_end, init_data_size,
6332 _sinittext, init_code_size);
6333 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6334 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6335 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6336 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6338 #undef adj_init_size
6340 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6341 #ifdef CONFIG_HIGHMEM
6345 K(nr_free_pages()), K(physpages),
6346 codesize / SZ_1K, datasize / SZ_1K, rosize / SZ_1K,
6347 (init_data_size + init_code_size) / SZ_1K, bss_size / SZ_1K,
6348 K(physpages - totalram_pages() - totalcma_pages),
6350 #ifdef CONFIG_HIGHMEM
6351 , K(totalhigh_pages())
6356 static int page_alloc_cpu_dead(unsigned int cpu)
6360 lru_add_drain_cpu(cpu);
6361 mlock_drain_remote(cpu);
6365 * Spill the event counters of the dead processor
6366 * into the current processors event counters.
6367 * This artificially elevates the count of the current
6370 vm_events_fold_cpu(cpu);
6373 * Zero the differential counters of the dead processor
6374 * so that the vm statistics are consistent.
6376 * This is only okay since the processor is dead and cannot
6377 * race with what we are doing.
6379 cpu_vm_stats_fold(cpu);
6381 for_each_populated_zone(zone)
6382 zone_pcp_update(zone, 0);
6387 static int page_alloc_cpu_online(unsigned int cpu)
6391 for_each_populated_zone(zone)
6392 zone_pcp_update(zone, 1);
6397 int hashdist = HASHDIST_DEFAULT;
6399 static int __init set_hashdist(char *str)
6403 hashdist = simple_strtoul(str, &str, 0);
6406 __setup("hashdist=", set_hashdist);
6409 void __init page_alloc_init(void)
6414 if (num_node_state(N_MEMORY) == 1)
6418 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
6419 "mm/page_alloc:pcp",
6420 page_alloc_cpu_online,
6421 page_alloc_cpu_dead);
6426 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6427 * or min_free_kbytes changes.
6429 static void calculate_totalreserve_pages(void)
6431 struct pglist_data *pgdat;
6432 unsigned long reserve_pages = 0;
6433 enum zone_type i, j;
6435 for_each_online_pgdat(pgdat) {
6437 pgdat->totalreserve_pages = 0;
6439 for (i = 0; i < MAX_NR_ZONES; i++) {
6440 struct zone *zone = pgdat->node_zones + i;
6442 unsigned long managed_pages = zone_managed_pages(zone);
6444 /* Find valid and maximum lowmem_reserve in the zone */
6445 for (j = i; j < MAX_NR_ZONES; j++) {
6446 if (zone->lowmem_reserve[j] > max)
6447 max = zone->lowmem_reserve[j];
6450 /* we treat the high watermark as reserved pages. */
6451 max += high_wmark_pages(zone);
6453 if (max > managed_pages)
6454 max = managed_pages;
6456 pgdat->totalreserve_pages += max;
6458 reserve_pages += max;
6461 totalreserve_pages = reserve_pages;
6465 * setup_per_zone_lowmem_reserve - called whenever
6466 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6467 * has a correct pages reserved value, so an adequate number of
6468 * pages are left in the zone after a successful __alloc_pages().
6470 static void setup_per_zone_lowmem_reserve(void)
6472 struct pglist_data *pgdat;
6473 enum zone_type i, j;
6475 for_each_online_pgdat(pgdat) {
6476 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
6477 struct zone *zone = &pgdat->node_zones[i];
6478 int ratio = sysctl_lowmem_reserve_ratio[i];
6479 bool clear = !ratio || !zone_managed_pages(zone);
6480 unsigned long managed_pages = 0;
6482 for (j = i + 1; j < MAX_NR_ZONES; j++) {
6483 struct zone *upper_zone = &pgdat->node_zones[j];
6485 managed_pages += zone_managed_pages(upper_zone);
6488 zone->lowmem_reserve[j] = 0;
6490 zone->lowmem_reserve[j] = managed_pages / ratio;
6495 /* update totalreserve_pages */
6496 calculate_totalreserve_pages();
6499 static void __setup_per_zone_wmarks(void)
6501 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6502 unsigned long lowmem_pages = 0;
6504 unsigned long flags;
6506 /* Calculate total number of !ZONE_HIGHMEM pages */
6507 for_each_zone(zone) {
6508 if (!is_highmem(zone))
6509 lowmem_pages += zone_managed_pages(zone);
6512 for_each_zone(zone) {
6515 spin_lock_irqsave(&zone->lock, flags);
6516 tmp = (u64)pages_min * zone_managed_pages(zone);
6517 do_div(tmp, lowmem_pages);
6518 if (is_highmem(zone)) {
6520 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6521 * need highmem pages, so cap pages_min to a small
6524 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6525 * deltas control async page reclaim, and so should
6526 * not be capped for highmem.
6528 unsigned long min_pages;
6530 min_pages = zone_managed_pages(zone) / 1024;
6531 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6532 zone->_watermark[WMARK_MIN] = min_pages;
6535 * If it's a lowmem zone, reserve a number of pages
6536 * proportionate to the zone's size.
6538 zone->_watermark[WMARK_MIN] = tmp;
6542 * Set the kswapd watermarks distance according to the
6543 * scale factor in proportion to available memory, but
6544 * ensure a minimum size on small systems.
6546 tmp = max_t(u64, tmp >> 2,
6547 mult_frac(zone_managed_pages(zone),
6548 watermark_scale_factor, 10000));
6550 zone->watermark_boost = 0;
6551 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6552 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
6553 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
6555 spin_unlock_irqrestore(&zone->lock, flags);
6558 /* update totalreserve_pages */
6559 calculate_totalreserve_pages();
6563 * setup_per_zone_wmarks - called when min_free_kbytes changes
6564 * or when memory is hot-{added|removed}
6566 * Ensures that the watermark[min,low,high] values for each zone are set
6567 * correctly with respect to min_free_kbytes.
6569 void setup_per_zone_wmarks(void)
6572 static DEFINE_SPINLOCK(lock);
6575 __setup_per_zone_wmarks();
6579 * The watermark size have changed so update the pcpu batch
6580 * and high limits or the limits may be inappropriate.
6583 zone_pcp_update(zone, 0);
6587 * Initialise min_free_kbytes.
6589 * For small machines we want it small (128k min). For large machines
6590 * we want it large (256MB max). But it is not linear, because network
6591 * bandwidth does not increase linearly with machine size. We use
6593 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6594 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6610 void calculate_min_free_kbytes(void)
6612 unsigned long lowmem_kbytes;
6613 int new_min_free_kbytes;
6615 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6616 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6618 if (new_min_free_kbytes > user_min_free_kbytes)
6619 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
6621 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6622 new_min_free_kbytes, user_min_free_kbytes);
6626 int __meminit init_per_zone_wmark_min(void)
6628 calculate_min_free_kbytes();
6629 setup_per_zone_wmarks();
6630 refresh_zone_stat_thresholds();
6631 setup_per_zone_lowmem_reserve();
6634 setup_min_unmapped_ratio();
6635 setup_min_slab_ratio();
6638 khugepaged_min_free_kbytes_update();
6642 postcore_initcall(init_per_zone_wmark_min)
6645 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6646 * that we can call two helper functions whenever min_free_kbytes
6649 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6650 void *buffer, size_t *length, loff_t *ppos)
6654 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6659 user_min_free_kbytes = min_free_kbytes;
6660 setup_per_zone_wmarks();
6665 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6666 void *buffer, size_t *length, loff_t *ppos)
6670 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6675 setup_per_zone_wmarks();
6681 static void setup_min_unmapped_ratio(void)
6686 for_each_online_pgdat(pgdat)
6687 pgdat->min_unmapped_pages = 0;
6690 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6691 sysctl_min_unmapped_ratio) / 100;
6695 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6696 void *buffer, size_t *length, loff_t *ppos)
6700 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6704 setup_min_unmapped_ratio();
6709 static void setup_min_slab_ratio(void)
6714 for_each_online_pgdat(pgdat)
6715 pgdat->min_slab_pages = 0;
6718 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6719 sysctl_min_slab_ratio) / 100;
6722 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6723 void *buffer, size_t *length, loff_t *ppos)
6727 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6731 setup_min_slab_ratio();
6738 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6739 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6740 * whenever sysctl_lowmem_reserve_ratio changes.
6742 * The reserve ratio obviously has absolutely no relation with the
6743 * minimum watermarks. The lowmem reserve ratio can only make sense
6744 * if in function of the boot time zone sizes.
6746 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6747 void *buffer, size_t *length, loff_t *ppos)
6751 proc_dointvec_minmax(table, write, buffer, length, ppos);
6753 for (i = 0; i < MAX_NR_ZONES; i++) {
6754 if (sysctl_lowmem_reserve_ratio[i] < 1)
6755 sysctl_lowmem_reserve_ratio[i] = 0;
6758 setup_per_zone_lowmem_reserve();
6763 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6764 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6765 * pagelist can have before it gets flushed back to buddy allocator.
6767 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
6768 int write, void *buffer, size_t *length, loff_t *ppos)
6771 int old_percpu_pagelist_high_fraction;
6774 mutex_lock(&pcp_batch_high_lock);
6775 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6777 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6778 if (!write || ret < 0)
6781 /* Sanity checking to avoid pcp imbalance */
6782 if (percpu_pagelist_high_fraction &&
6783 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6784 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6790 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6793 for_each_populated_zone(zone)
6794 zone_set_pageset_high_and_batch(zone, 0);
6796 mutex_unlock(&pcp_batch_high_lock);
6800 #ifdef CONFIG_CONTIG_ALLOC
6801 #if defined(CONFIG_DYNAMIC_DEBUG) || \
6802 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
6803 /* Usage: See admin-guide/dynamic-debug-howto.rst */
6804 static void alloc_contig_dump_pages(struct list_head *page_list)
6806 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6808 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6812 list_for_each_entry(page, page_list, lru)
6813 dump_page(page, "migration failure");
6817 static inline void alloc_contig_dump_pages(struct list_head *page_list)
6822 /* [start, end) must belong to a single zone. */
6823 int __alloc_contig_migrate_range(struct compact_control *cc,
6824 unsigned long start, unsigned long end)
6826 /* This function is based on compact_zone() from compaction.c. */
6827 unsigned int nr_reclaimed;
6828 unsigned long pfn = start;
6829 unsigned int tries = 0;
6831 struct migration_target_control mtc = {
6832 .nid = zone_to_nid(cc->zone),
6833 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6836 lru_cache_disable();
6838 while (pfn < end || !list_empty(&cc->migratepages)) {
6839 if (fatal_signal_pending(current)) {
6844 if (list_empty(&cc->migratepages)) {
6845 cc->nr_migratepages = 0;
6846 ret = isolate_migratepages_range(cc, pfn, end);
6847 if (ret && ret != -EAGAIN)
6849 pfn = cc->migrate_pfn;
6851 } else if (++tries == 5) {
6856 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6858 cc->nr_migratepages -= nr_reclaimed;
6860 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6861 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6864 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6865 * to retry again over this error, so do the same here.
6873 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6874 alloc_contig_dump_pages(&cc->migratepages);
6875 putback_movable_pages(&cc->migratepages);
6882 * alloc_contig_range() -- tries to allocate given range of pages
6883 * @start: start PFN to allocate
6884 * @end: one-past-the-last PFN to allocate
6885 * @migratetype: migratetype of the underlying pageblocks (either
6886 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6887 * in range must have the same migratetype and it must
6888 * be either of the two.
6889 * @gfp_mask: GFP mask to use during compaction
6891 * The PFN range does not have to be pageblock aligned. The PFN range must
6892 * belong to a single zone.
6894 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6895 * pageblocks in the range. Once isolated, the pageblocks should not
6896 * be modified by others.
6898 * Return: zero on success or negative error code. On success all
6899 * pages which PFN is in [start, end) are allocated for the caller and
6900 * need to be freed with free_contig_range().
6902 int alloc_contig_range(unsigned long start, unsigned long end,
6903 unsigned migratetype, gfp_t gfp_mask)
6905 unsigned long outer_start, outer_end;
6909 struct compact_control cc = {
6910 .nr_migratepages = 0,
6912 .zone = page_zone(pfn_to_page(start)),
6913 .mode = MIGRATE_SYNC,
6914 .ignore_skip_hint = true,
6915 .no_set_skip_hint = true,
6916 .gfp_mask = current_gfp_context(gfp_mask),
6917 .alloc_contig = true,
6919 INIT_LIST_HEAD(&cc.migratepages);
6922 * What we do here is we mark all pageblocks in range as
6923 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6924 * have different sizes, and due to the way page allocator
6925 * work, start_isolate_page_range() has special handlings for this.
6927 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6928 * migrate the pages from an unaligned range (ie. pages that
6929 * we are interested in). This will put all the pages in
6930 * range back to page allocator as MIGRATE_ISOLATE.
6932 * When this is done, we take the pages in range from page
6933 * allocator removing them from the buddy system. This way
6934 * page allocator will never consider using them.
6936 * This lets us mark the pageblocks back as
6937 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6938 * aligned range but not in the unaligned, original range are
6939 * put back to page allocator so that buddy can use them.
6942 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6946 drain_all_pages(cc.zone);
6949 * In case of -EBUSY, we'd like to know which page causes problem.
6950 * So, just fall through. test_pages_isolated() has a tracepoint
6951 * which will report the busy page.
6953 * It is possible that busy pages could become available before
6954 * the call to test_pages_isolated, and the range will actually be
6955 * allocated. So, if we fall through be sure to clear ret so that
6956 * -EBUSY is not accidentally used or returned to caller.
6958 ret = __alloc_contig_migrate_range(&cc, start, end);
6959 if (ret && ret != -EBUSY)
6964 * Pages from [start, end) are within a pageblock_nr_pages
6965 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6966 * more, all pages in [start, end) are free in page allocator.
6967 * What we are going to do is to allocate all pages from
6968 * [start, end) (that is remove them from page allocator).
6970 * The only problem is that pages at the beginning and at the
6971 * end of interesting range may be not aligned with pages that
6972 * page allocator holds, ie. they can be part of higher order
6973 * pages. Because of this, we reserve the bigger range and
6974 * once this is done free the pages we are not interested in.
6976 * We don't have to hold zone->lock here because the pages are
6977 * isolated thus they won't get removed from buddy.
6981 outer_start = start;
6982 while (!PageBuddy(pfn_to_page(outer_start))) {
6983 if (++order > MAX_ORDER) {
6984 outer_start = start;
6987 outer_start &= ~0UL << order;
6990 if (outer_start != start) {
6991 order = buddy_order(pfn_to_page(outer_start));
6994 * outer_start page could be small order buddy page and
6995 * it doesn't include start page. Adjust outer_start
6996 * in this case to report failed page properly
6997 * on tracepoint in test_pages_isolated()
6999 if (outer_start + (1UL << order) <= start)
7000 outer_start = start;
7003 /* Make sure the range is really isolated. */
7004 if (test_pages_isolated(outer_start, end, 0)) {
7009 /* Grab isolated pages from freelists. */
7010 outer_end = isolate_freepages_range(&cc, outer_start, end);
7016 /* Free head and tail (if any) */
7017 if (start != outer_start)
7018 free_contig_range(outer_start, start - outer_start);
7019 if (end != outer_end)
7020 free_contig_range(end, outer_end - end);
7023 undo_isolate_page_range(start, end, migratetype);
7026 EXPORT_SYMBOL(alloc_contig_range);
7028 static int __alloc_contig_pages(unsigned long start_pfn,
7029 unsigned long nr_pages, gfp_t gfp_mask)
7031 unsigned long end_pfn = start_pfn + nr_pages;
7033 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
7037 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
7038 unsigned long nr_pages)
7040 unsigned long i, end_pfn = start_pfn + nr_pages;
7043 for (i = start_pfn; i < end_pfn; i++) {
7044 page = pfn_to_online_page(i);
7048 if (page_zone(page) != z)
7051 if (PageReserved(page))
7057 static bool zone_spans_last_pfn(const struct zone *zone,
7058 unsigned long start_pfn, unsigned long nr_pages)
7060 unsigned long last_pfn = start_pfn + nr_pages - 1;
7062 return zone_spans_pfn(zone, last_pfn);
7066 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
7067 * @nr_pages: Number of contiguous pages to allocate
7068 * @gfp_mask: GFP mask to limit search and used during compaction
7070 * @nodemask: Mask for other possible nodes
7072 * This routine is a wrapper around alloc_contig_range(). It scans over zones
7073 * on an applicable zonelist to find a contiguous pfn range which can then be
7074 * tried for allocation with alloc_contig_range(). This routine is intended
7075 * for allocation requests which can not be fulfilled with the buddy allocator.
7077 * The allocated memory is always aligned to a page boundary. If nr_pages is a
7078 * power of two, then allocated range is also guaranteed to be aligned to same
7079 * nr_pages (e.g. 1GB request would be aligned to 1GB).
7081 * Allocated pages can be freed with free_contig_range() or by manually calling
7082 * __free_page() on each allocated page.
7084 * Return: pointer to contiguous pages on success, or NULL if not successful.
7086 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
7087 int nid, nodemask_t *nodemask)
7089 unsigned long ret, pfn, flags;
7090 struct zonelist *zonelist;
7094 zonelist = node_zonelist(nid, gfp_mask);
7095 for_each_zone_zonelist_nodemask(zone, z, zonelist,
7096 gfp_zone(gfp_mask), nodemask) {
7097 spin_lock_irqsave(&zone->lock, flags);
7099 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
7100 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
7101 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
7103 * We release the zone lock here because
7104 * alloc_contig_range() will also lock the zone
7105 * at some point. If there's an allocation
7106 * spinning on this lock, it may win the race
7107 * and cause alloc_contig_range() to fail...
7109 spin_unlock_irqrestore(&zone->lock, flags);
7110 ret = __alloc_contig_pages(pfn, nr_pages,
7113 return pfn_to_page(pfn);
7114 spin_lock_irqsave(&zone->lock, flags);
7118 spin_unlock_irqrestore(&zone->lock, flags);
7122 #endif /* CONFIG_CONTIG_ALLOC */
7124 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
7126 unsigned long count = 0;
7128 for (; nr_pages--; pfn++) {
7129 struct page *page = pfn_to_page(pfn);
7131 count += page_count(page) != 1;
7134 WARN(count != 0, "%lu pages are still in use!\n", count);
7136 EXPORT_SYMBOL(free_contig_range);
7139 * Effectively disable pcplists for the zone by setting the high limit to 0
7140 * and draining all cpus. A concurrent page freeing on another CPU that's about
7141 * to put the page on pcplist will either finish before the drain and the page
7142 * will be drained, or observe the new high limit and skip the pcplist.
7144 * Must be paired with a call to zone_pcp_enable().
7146 void zone_pcp_disable(struct zone *zone)
7148 mutex_lock(&pcp_batch_high_lock);
7149 __zone_set_pageset_high_and_batch(zone, 0, 1);
7150 __drain_all_pages(zone, true);
7153 void zone_pcp_enable(struct zone *zone)
7155 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
7156 mutex_unlock(&pcp_batch_high_lock);
7159 void zone_pcp_reset(struct zone *zone)
7162 struct per_cpu_zonestat *pzstats;
7164 if (zone->per_cpu_pageset != &boot_pageset) {
7165 for_each_online_cpu(cpu) {
7166 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7167 drain_zonestat(zone, pzstats);
7169 free_percpu(zone->per_cpu_pageset);
7170 zone->per_cpu_pageset = &boot_pageset;
7171 if (zone->per_cpu_zonestats != &boot_zonestats) {
7172 free_percpu(zone->per_cpu_zonestats);
7173 zone->per_cpu_zonestats = &boot_zonestats;
7178 #ifdef CONFIG_MEMORY_HOTREMOVE
7180 * All pages in the range must be in a single zone, must not contain holes,
7181 * must span full sections, and must be isolated before calling this function.
7183 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7185 unsigned long pfn = start_pfn;
7189 unsigned long flags;
7191 offline_mem_sections(pfn, end_pfn);
7192 zone = page_zone(pfn_to_page(pfn));
7193 spin_lock_irqsave(&zone->lock, flags);
7194 while (pfn < end_pfn) {
7195 page = pfn_to_page(pfn);
7197 * The HWPoisoned page may be not in buddy system, and
7198 * page_count() is not 0.
7200 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7205 * At this point all remaining PageOffline() pages have a
7206 * reference count of 0 and can simply be skipped.
7208 if (PageOffline(page)) {
7209 BUG_ON(page_count(page));
7210 BUG_ON(PageBuddy(page));
7215 BUG_ON(page_count(page));
7216 BUG_ON(!PageBuddy(page));
7217 order = buddy_order(page);
7218 del_page_from_free_list(page, zone, order);
7219 pfn += (1 << order);
7221 spin_unlock_irqrestore(&zone->lock, flags);
7226 * This function returns a stable result only if called under zone lock.
7228 bool is_free_buddy_page(struct page *page)
7230 unsigned long pfn = page_to_pfn(page);
7233 for (order = 0; order <= MAX_ORDER; order++) {
7234 struct page *page_head = page - (pfn & ((1 << order) - 1));
7236 if (PageBuddy(page_head) &&
7237 buddy_order_unsafe(page_head) >= order)
7241 return order <= MAX_ORDER;
7243 EXPORT_SYMBOL(is_free_buddy_page);
7245 #ifdef CONFIG_MEMORY_FAILURE
7247 * Break down a higher-order page in sub-pages, and keep our target out of
7250 static void break_down_buddy_pages(struct zone *zone, struct page *page,
7251 struct page *target, int low, int high,
7254 unsigned long size = 1 << high;
7255 struct page *current_buddy, *next_page;
7257 while (high > low) {
7261 if (target >= &page[size]) {
7262 next_page = page + size;
7263 current_buddy = page;
7266 current_buddy = page + size;
7269 if (set_page_guard(zone, current_buddy, high, migratetype))
7272 if (current_buddy != target) {
7273 add_to_free_list(current_buddy, zone, high, migratetype);
7274 set_buddy_order(current_buddy, high);
7281 * Take a page that will be marked as poisoned off the buddy allocator.
7283 bool take_page_off_buddy(struct page *page)
7285 struct zone *zone = page_zone(page);
7286 unsigned long pfn = page_to_pfn(page);
7287 unsigned long flags;
7291 spin_lock_irqsave(&zone->lock, flags);
7292 for (order = 0; order <= MAX_ORDER; order++) {
7293 struct page *page_head = page - (pfn & ((1 << order) - 1));
7294 int page_order = buddy_order(page_head);
7296 if (PageBuddy(page_head) && page_order >= order) {
7297 unsigned long pfn_head = page_to_pfn(page_head);
7298 int migratetype = get_pfnblock_migratetype(page_head,
7301 del_page_from_free_list(page_head, zone, page_order);
7302 break_down_buddy_pages(zone, page_head, page, 0,
7303 page_order, migratetype);
7304 SetPageHWPoisonTakenOff(page);
7305 if (!is_migrate_isolate(migratetype))
7306 __mod_zone_freepage_state(zone, -1, migratetype);
7310 if (page_count(page_head) > 0)
7313 spin_unlock_irqrestore(&zone->lock, flags);
7318 * Cancel takeoff done by take_page_off_buddy().
7320 bool put_page_back_buddy(struct page *page)
7322 struct zone *zone = page_zone(page);
7323 unsigned long pfn = page_to_pfn(page);
7324 unsigned long flags;
7325 int migratetype = get_pfnblock_migratetype(page, pfn);
7328 spin_lock_irqsave(&zone->lock, flags);
7329 if (put_page_testzero(page)) {
7330 ClearPageHWPoisonTakenOff(page);
7331 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
7332 if (TestClearPageHWPoison(page)) {
7336 spin_unlock_irqrestore(&zone->lock, flags);
7342 #ifdef CONFIG_ZONE_DMA
7343 bool has_managed_dma(void)
7345 struct pglist_data *pgdat;
7347 for_each_online_pgdat(pgdat) {
7348 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
7350 if (managed_zone(zone))
7355 #endif /* CONFIG_ZONE_DMA */