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/module.h>
31 #include <linux/suspend.h>
32 #include <linux/pagevec.h>
33 #include <linux/blkdev.h>
34 #include <linux/slab.h>
35 #include <linux/ratelimit.h>
36 #include <linux/oom.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/random.h>
49 #include <linux/sort.h>
50 #include <linux/pfn.h>
51 #include <linux/backing-dev.h>
52 #include <linux/fault-inject.h>
53 #include <linux/page-isolation.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/migrate.h>
63 #include <linux/hugetlb.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/mm.h>
66 #include <linux/page_owner.h>
67 #include <linux/page_table_check.h>
68 #include <linux/kthread.h>
69 #include <linux/memcontrol.h>
70 #include <linux/ftrace.h>
71 #include <linux/lockdep.h>
72 #include <linux/nmi.h>
73 #include <linux/psi.h>
74 #include <linux/padata.h>
75 #include <linux/khugepaged.h>
76 #include <linux/buffer_head.h>
77 #include <linux/delayacct.h>
78 #include <asm/sections.h>
79 #include <asm/tlbflush.h>
80 #include <asm/div64.h>
83 #include "page_reporting.h"
86 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
87 typedef int __bitwise fpi_t;
89 /* No special request */
90 #define FPI_NONE ((__force fpi_t)0)
93 * Skip free page reporting notification for the (possibly merged) page.
94 * This does not hinder free page reporting from grabbing the page,
95 * reporting it and marking it "reported" - it only skips notifying
96 * the free page reporting infrastructure about a newly freed page. For
97 * example, used when temporarily pulling a page from a freelist and
98 * putting it back unmodified.
100 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
103 * Place the (possibly merged) page to the tail of the freelist. Will ignore
104 * page shuffling (relevant code - e.g., memory onlining - is expected to
105 * shuffle the whole zone).
107 * Note: No code should rely on this flag for correctness - it's purely
108 * to allow for optimizations when handing back either fresh pages
109 * (memory onlining) or untouched pages (page isolation, free page
112 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
115 * Don't poison memory with KASAN (only for the tag-based modes).
116 * During boot, all non-reserved memblock memory is exposed to page_alloc.
117 * Poisoning all that memory lengthens boot time, especially on systems with
118 * large amount of RAM. This flag is used to skip that poisoning.
119 * This is only done for the tag-based KASAN modes, as those are able to
120 * detect memory corruptions with the memory tags assigned by default.
121 * All memory allocated normally after boot gets poisoned as usual.
123 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
125 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
126 static DEFINE_MUTEX(pcp_batch_high_lock);
127 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
129 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
131 * On SMP, spin_trylock is sufficient protection.
132 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
134 #define pcp_trylock_prepare(flags) do { } while (0)
135 #define pcp_trylock_finish(flag) do { } while (0)
138 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
139 #define pcp_trylock_prepare(flags) local_irq_save(flags)
140 #define pcp_trylock_finish(flags) local_irq_restore(flags)
144 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
145 * a migration causing the wrong PCP to be locked and remote memory being
146 * potentially allocated, pin the task to the CPU for the lookup+lock.
147 * preempt_disable is used on !RT because it is faster than migrate_disable.
148 * migrate_disable is used on RT because otherwise RT spinlock usage is
149 * interfered with and a high priority task cannot preempt the allocator.
151 #ifndef CONFIG_PREEMPT_RT
152 #define pcpu_task_pin() preempt_disable()
153 #define pcpu_task_unpin() preempt_enable()
155 #define pcpu_task_pin() migrate_disable()
156 #define pcpu_task_unpin() migrate_enable()
160 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
161 * Return value should be used with equivalent unlock helper.
163 #define pcpu_spin_lock(type, member, ptr) \
167 _ret = this_cpu_ptr(ptr); \
168 spin_lock(&_ret->member); \
172 #define pcpu_spin_lock_irqsave(type, member, ptr, flags) \
176 _ret = this_cpu_ptr(ptr); \
177 spin_lock_irqsave(&_ret->member, flags); \
181 #define pcpu_spin_trylock_irqsave(type, member, ptr, flags) \
185 _ret = this_cpu_ptr(ptr); \
186 if (!spin_trylock_irqsave(&_ret->member, flags)) { \
193 #define pcpu_spin_unlock(member, ptr) \
195 spin_unlock(&ptr->member); \
199 #define pcpu_spin_unlock_irqrestore(member, ptr, flags) \
201 spin_unlock_irqrestore(&ptr->member, flags); \
205 /* struct per_cpu_pages specific helpers. */
206 #define pcp_spin_lock(ptr) \
207 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
209 #define pcp_spin_lock_irqsave(ptr, flags) \
210 pcpu_spin_lock_irqsave(struct per_cpu_pages, lock, ptr, flags)
212 #define pcp_spin_trylock_irqsave(ptr, flags) \
213 pcpu_spin_trylock_irqsave(struct per_cpu_pages, lock, ptr, flags)
215 #define pcp_spin_unlock(ptr) \
216 pcpu_spin_unlock(lock, ptr)
218 #define pcp_spin_unlock_irqrestore(ptr, flags) \
219 pcpu_spin_unlock_irqrestore(lock, ptr, flags)
220 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
221 DEFINE_PER_CPU(int, numa_node);
222 EXPORT_PER_CPU_SYMBOL(numa_node);
225 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
227 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
229 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
230 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
231 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
232 * defined in <linux/topology.h>.
234 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
235 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
238 static DEFINE_MUTEX(pcpu_drain_mutex);
240 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
241 volatile unsigned long latent_entropy __latent_entropy;
242 EXPORT_SYMBOL(latent_entropy);
246 * Array of node states.
248 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
249 [N_POSSIBLE] = NODE_MASK_ALL,
250 [N_ONLINE] = { { [0] = 1UL } },
252 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
253 #ifdef CONFIG_HIGHMEM
254 [N_HIGH_MEMORY] = { { [0] = 1UL } },
256 [N_MEMORY] = { { [0] = 1UL } },
257 [N_CPU] = { { [0] = 1UL } },
260 EXPORT_SYMBOL(node_states);
262 atomic_long_t _totalram_pages __read_mostly;
263 EXPORT_SYMBOL(_totalram_pages);
264 unsigned long totalreserve_pages __read_mostly;
265 unsigned long totalcma_pages __read_mostly;
267 int percpu_pagelist_high_fraction;
268 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
269 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
270 EXPORT_SYMBOL(init_on_alloc);
272 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
273 EXPORT_SYMBOL(init_on_free);
275 static bool _init_on_alloc_enabled_early __read_mostly
276 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
277 static int __init early_init_on_alloc(char *buf)
280 return kstrtobool(buf, &_init_on_alloc_enabled_early);
282 early_param("init_on_alloc", early_init_on_alloc);
284 static bool _init_on_free_enabled_early __read_mostly
285 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
286 static int __init early_init_on_free(char *buf)
288 return kstrtobool(buf, &_init_on_free_enabled_early);
290 early_param("init_on_free", early_init_on_free);
293 * A cached value of the page's pageblock's migratetype, used when the page is
294 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
295 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
296 * Also the migratetype set in the page does not necessarily match the pcplist
297 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
298 * other index - this ensures that it will be put on the correct CMA freelist.
300 static inline int get_pcppage_migratetype(struct page *page)
305 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
307 page->index = migratetype;
310 #ifdef CONFIG_PM_SLEEP
312 * The following functions are used by the suspend/hibernate code to temporarily
313 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
314 * while devices are suspended. To avoid races with the suspend/hibernate code,
315 * they should always be called with system_transition_mutex held
316 * (gfp_allowed_mask also should only be modified with system_transition_mutex
317 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
318 * with that modification).
321 static gfp_t saved_gfp_mask;
323 void pm_restore_gfp_mask(void)
325 WARN_ON(!mutex_is_locked(&system_transition_mutex));
326 if (saved_gfp_mask) {
327 gfp_allowed_mask = saved_gfp_mask;
332 void pm_restrict_gfp_mask(void)
334 WARN_ON(!mutex_is_locked(&system_transition_mutex));
335 WARN_ON(saved_gfp_mask);
336 saved_gfp_mask = gfp_allowed_mask;
337 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
340 bool pm_suspended_storage(void)
342 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
346 #endif /* CONFIG_PM_SLEEP */
348 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
349 unsigned int pageblock_order __read_mostly;
352 static void __free_pages_ok(struct page *page, unsigned int order,
356 * results with 256, 32 in the lowmem_reserve sysctl:
357 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
358 * 1G machine -> (16M dma, 784M normal, 224M high)
359 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
360 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
361 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
363 * TBD: should special case ZONE_DMA32 machines here - in those we normally
364 * don't need any ZONE_NORMAL reservation
366 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
367 #ifdef CONFIG_ZONE_DMA
370 #ifdef CONFIG_ZONE_DMA32
374 #ifdef CONFIG_HIGHMEM
380 static char * const zone_names[MAX_NR_ZONES] = {
381 #ifdef CONFIG_ZONE_DMA
384 #ifdef CONFIG_ZONE_DMA32
388 #ifdef CONFIG_HIGHMEM
392 #ifdef CONFIG_ZONE_DEVICE
397 const char * const migratetype_names[MIGRATE_TYPES] = {
405 #ifdef CONFIG_MEMORY_ISOLATION
410 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
411 [NULL_COMPOUND_DTOR] = NULL,
412 [COMPOUND_PAGE_DTOR] = free_compound_page,
413 #ifdef CONFIG_HUGETLB_PAGE
414 [HUGETLB_PAGE_DTOR] = free_huge_page,
416 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
417 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
421 int min_free_kbytes = 1024;
422 int user_min_free_kbytes = -1;
423 int watermark_boost_factor __read_mostly = 15000;
424 int watermark_scale_factor = 10;
426 static unsigned long nr_kernel_pages __initdata;
427 static unsigned long nr_all_pages __initdata;
428 static unsigned long dma_reserve __initdata;
430 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
431 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
432 static unsigned long required_kernelcore __initdata;
433 static unsigned long required_kernelcore_percent __initdata;
434 static unsigned long required_movablecore __initdata;
435 static unsigned long required_movablecore_percent __initdata;
436 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
437 static bool mirrored_kernelcore __meminitdata;
439 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
441 EXPORT_SYMBOL(movable_zone);
444 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
445 unsigned int nr_online_nodes __read_mostly = 1;
446 EXPORT_SYMBOL(nr_node_ids);
447 EXPORT_SYMBOL(nr_online_nodes);
450 int page_group_by_mobility_disabled __read_mostly;
452 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
454 * During boot we initialize deferred pages on-demand, as needed, but once
455 * page_alloc_init_late() has finished, the deferred pages are all initialized,
456 * and we can permanently disable that path.
458 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
460 static inline bool deferred_pages_enabled(void)
462 return static_branch_unlikely(&deferred_pages);
465 /* Returns true if the struct page for the pfn is uninitialised */
466 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
468 int nid = early_pfn_to_nid(pfn);
470 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
477 * Returns true when the remaining initialisation should be deferred until
478 * later in the boot cycle when it can be parallelised.
480 static bool __meminit
481 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
483 static unsigned long prev_end_pfn, nr_initialised;
486 * prev_end_pfn static that contains the end of previous zone
487 * No need to protect because called very early in boot before smp_init.
489 if (prev_end_pfn != end_pfn) {
490 prev_end_pfn = end_pfn;
494 /* Always populate low zones for address-constrained allocations */
495 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
498 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
501 * We start only with one section of pages, more pages are added as
502 * needed until the rest of deferred pages are initialized.
505 if ((nr_initialised > PAGES_PER_SECTION) &&
506 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
507 NODE_DATA(nid)->first_deferred_pfn = pfn;
513 static inline bool deferred_pages_enabled(void)
518 static inline bool early_page_uninitialised(unsigned long pfn)
523 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
529 /* Return a pointer to the bitmap storing bits affecting a block of pages */
530 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
533 #ifdef CONFIG_SPARSEMEM
534 return section_to_usemap(__pfn_to_section(pfn));
536 return page_zone(page)->pageblock_flags;
537 #endif /* CONFIG_SPARSEMEM */
540 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
542 #ifdef CONFIG_SPARSEMEM
543 pfn &= (PAGES_PER_SECTION-1);
545 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
546 #endif /* CONFIG_SPARSEMEM */
547 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
550 static __always_inline
551 unsigned long __get_pfnblock_flags_mask(const struct page *page,
555 unsigned long *bitmap;
556 unsigned long bitidx, word_bitidx;
559 bitmap = get_pageblock_bitmap(page, pfn);
560 bitidx = pfn_to_bitidx(page, pfn);
561 word_bitidx = bitidx / BITS_PER_LONG;
562 bitidx &= (BITS_PER_LONG-1);
564 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
565 * a consistent read of the memory array, so that results, even though
566 * racy, are not corrupted.
568 word = READ_ONCE(bitmap[word_bitidx]);
569 return (word >> bitidx) & mask;
573 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
574 * @page: The page within the block of interest
575 * @pfn: The target page frame number
576 * @mask: mask of bits that the caller is interested in
578 * Return: pageblock_bits flags
580 unsigned long get_pfnblock_flags_mask(const struct page *page,
581 unsigned long pfn, unsigned long mask)
583 return __get_pfnblock_flags_mask(page, pfn, mask);
586 static __always_inline int get_pfnblock_migratetype(const struct page *page,
589 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
593 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
594 * @page: The page within the block of interest
595 * @flags: The flags to set
596 * @pfn: The target page frame number
597 * @mask: mask of bits that the caller is interested in
599 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
603 unsigned long *bitmap;
604 unsigned long bitidx, word_bitidx;
607 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
608 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
610 bitmap = get_pageblock_bitmap(page, pfn);
611 bitidx = pfn_to_bitidx(page, pfn);
612 word_bitidx = bitidx / BITS_PER_LONG;
613 bitidx &= (BITS_PER_LONG-1);
615 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
620 word = READ_ONCE(bitmap[word_bitidx]);
622 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
625 void set_pageblock_migratetype(struct page *page, int migratetype)
627 if (unlikely(page_group_by_mobility_disabled &&
628 migratetype < MIGRATE_PCPTYPES))
629 migratetype = MIGRATE_UNMOVABLE;
631 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
632 page_to_pfn(page), MIGRATETYPE_MASK);
635 #ifdef CONFIG_DEBUG_VM
636 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
640 unsigned long pfn = page_to_pfn(page);
641 unsigned long sp, start_pfn;
644 seq = zone_span_seqbegin(zone);
645 start_pfn = zone->zone_start_pfn;
646 sp = zone->spanned_pages;
647 if (!zone_spans_pfn(zone, pfn))
649 } while (zone_span_seqretry(zone, seq));
652 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
653 pfn, zone_to_nid(zone), zone->name,
654 start_pfn, start_pfn + sp);
659 static int page_is_consistent(struct zone *zone, struct page *page)
661 if (zone != page_zone(page))
667 * Temporary debugging check for pages not lying within a given zone.
669 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
671 if (page_outside_zone_boundaries(zone, page))
673 if (!page_is_consistent(zone, page))
679 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
685 static void bad_page(struct page *page, const char *reason)
687 static unsigned long resume;
688 static unsigned long nr_shown;
689 static unsigned long nr_unshown;
692 * Allow a burst of 60 reports, then keep quiet for that minute;
693 * or allow a steady drip of one report per second.
695 if (nr_shown == 60) {
696 if (time_before(jiffies, resume)) {
702 "BUG: Bad page state: %lu messages suppressed\n",
709 resume = jiffies + 60 * HZ;
711 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
712 current->comm, page_to_pfn(page));
713 dump_page(page, reason);
718 /* Leave bad fields for debug, except PageBuddy could make trouble */
719 page_mapcount_reset(page); /* remove PageBuddy */
720 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
723 static inline unsigned int order_to_pindex(int migratetype, int order)
727 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
728 if (order > PAGE_ALLOC_COSTLY_ORDER) {
729 VM_BUG_ON(order != pageblock_order);
730 return NR_LOWORDER_PCP_LISTS;
733 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
736 return (MIGRATE_PCPTYPES * base) + migratetype;
739 static inline int pindex_to_order(unsigned int pindex)
741 int order = pindex / MIGRATE_PCPTYPES;
743 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
744 if (pindex == NR_LOWORDER_PCP_LISTS)
745 order = pageblock_order;
747 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
753 static inline bool pcp_allowed_order(unsigned int order)
755 if (order <= PAGE_ALLOC_COSTLY_ORDER)
757 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
758 if (order == pageblock_order)
764 static inline void free_the_page(struct page *page, unsigned int order)
766 if (pcp_allowed_order(order)) /* Via pcp? */
767 free_unref_page(page, order);
769 __free_pages_ok(page, order, FPI_NONE);
773 * Higher-order pages are called "compound pages". They are structured thusly:
775 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
777 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
778 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
780 * The first tail page's ->compound_dtor holds the offset in array of compound
781 * page destructors. See compound_page_dtors.
783 * The first tail page's ->compound_order holds the order of allocation.
784 * This usage means that zero-order pages may not be compound.
787 void free_compound_page(struct page *page)
789 mem_cgroup_uncharge(page_folio(page));
790 free_the_page(page, compound_order(page));
793 static void prep_compound_head(struct page *page, unsigned int order)
795 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
796 set_compound_order(page, order);
797 atomic_set(compound_mapcount_ptr(page), -1);
798 atomic_set(compound_pincount_ptr(page), 0);
801 static void prep_compound_tail(struct page *head, int tail_idx)
803 struct page *p = head + tail_idx;
805 p->mapping = TAIL_MAPPING;
806 set_compound_head(p, head);
809 void prep_compound_page(struct page *page, unsigned int order)
812 int nr_pages = 1 << order;
815 for (i = 1; i < nr_pages; i++)
816 prep_compound_tail(page, i);
818 prep_compound_head(page, order);
821 void destroy_large_folio(struct folio *folio)
823 enum compound_dtor_id dtor = folio_page(folio, 1)->compound_dtor;
825 VM_BUG_ON_FOLIO(dtor >= NR_COMPOUND_DTORS, folio);
826 compound_page_dtors[dtor](&folio->page);
829 #ifdef CONFIG_DEBUG_PAGEALLOC
830 unsigned int _debug_guardpage_minorder;
832 bool _debug_pagealloc_enabled_early __read_mostly
833 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
834 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
835 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
836 EXPORT_SYMBOL(_debug_pagealloc_enabled);
838 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
840 static int __init early_debug_pagealloc(char *buf)
842 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
844 early_param("debug_pagealloc", early_debug_pagealloc);
846 static int __init debug_guardpage_minorder_setup(char *buf)
850 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
851 pr_err("Bad debug_guardpage_minorder value\n");
854 _debug_guardpage_minorder = res;
855 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
858 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
860 static inline bool set_page_guard(struct zone *zone, struct page *page,
861 unsigned int order, int migratetype)
863 if (!debug_guardpage_enabled())
866 if (order >= debug_guardpage_minorder())
869 __SetPageGuard(page);
870 INIT_LIST_HEAD(&page->buddy_list);
871 set_page_private(page, order);
872 /* Guard pages are not available for any usage */
873 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
878 static inline void clear_page_guard(struct zone *zone, struct page *page,
879 unsigned int order, int migratetype)
881 if (!debug_guardpage_enabled())
884 __ClearPageGuard(page);
886 set_page_private(page, 0);
887 if (!is_migrate_isolate(migratetype))
888 __mod_zone_freepage_state(zone, (1 << order), migratetype);
891 static inline bool set_page_guard(struct zone *zone, struct page *page,
892 unsigned int order, int migratetype) { return false; }
893 static inline void clear_page_guard(struct zone *zone, struct page *page,
894 unsigned int order, int migratetype) {}
898 * Enable static keys related to various memory debugging and hardening options.
899 * Some override others, and depend on early params that are evaluated in the
900 * order of appearance. So we need to first gather the full picture of what was
901 * enabled, and then make decisions.
903 void init_mem_debugging_and_hardening(void)
905 bool page_poisoning_requested = false;
907 #ifdef CONFIG_PAGE_POISONING
909 * Page poisoning is debug page alloc for some arches. If
910 * either of those options are enabled, enable poisoning.
912 if (page_poisoning_enabled() ||
913 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
914 debug_pagealloc_enabled())) {
915 static_branch_enable(&_page_poisoning_enabled);
916 page_poisoning_requested = true;
920 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
921 page_poisoning_requested) {
922 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
923 "will take precedence over init_on_alloc and init_on_free\n");
924 _init_on_alloc_enabled_early = false;
925 _init_on_free_enabled_early = false;
928 if (_init_on_alloc_enabled_early)
929 static_branch_enable(&init_on_alloc);
931 static_branch_disable(&init_on_alloc);
933 if (_init_on_free_enabled_early)
934 static_branch_enable(&init_on_free);
936 static_branch_disable(&init_on_free);
938 #ifdef CONFIG_DEBUG_PAGEALLOC
939 if (!debug_pagealloc_enabled())
942 static_branch_enable(&_debug_pagealloc_enabled);
944 if (!debug_guardpage_minorder())
947 static_branch_enable(&_debug_guardpage_enabled);
951 static inline void set_buddy_order(struct page *page, unsigned int order)
953 set_page_private(page, order);
954 __SetPageBuddy(page);
957 #ifdef CONFIG_COMPACTION
958 static inline struct capture_control *task_capc(struct zone *zone)
960 struct capture_control *capc = current->capture_control;
962 return unlikely(capc) &&
963 !(current->flags & PF_KTHREAD) &&
965 capc->cc->zone == zone ? capc : NULL;
969 compaction_capture(struct capture_control *capc, struct page *page,
970 int order, int migratetype)
972 if (!capc || order != capc->cc->order)
975 /* Do not accidentally pollute CMA or isolated regions*/
976 if (is_migrate_cma(migratetype) ||
977 is_migrate_isolate(migratetype))
981 * Do not let lower order allocations pollute a movable pageblock.
982 * This might let an unmovable request use a reclaimable pageblock
983 * and vice-versa but no more than normal fallback logic which can
984 * have trouble finding a high-order free page.
986 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
994 static inline struct capture_control *task_capc(struct zone *zone)
1000 compaction_capture(struct capture_control *capc, struct page *page,
1001 int order, int migratetype)
1005 #endif /* CONFIG_COMPACTION */
1007 /* Used for pages not on another list */
1008 static inline void add_to_free_list(struct page *page, struct zone *zone,
1009 unsigned int order, int migratetype)
1011 struct free_area *area = &zone->free_area[order];
1013 list_add(&page->buddy_list, &area->free_list[migratetype]);
1017 /* Used for pages not on another list */
1018 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
1019 unsigned int order, int migratetype)
1021 struct free_area *area = &zone->free_area[order];
1023 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
1028 * Used for pages which are on another list. Move the pages to the tail
1029 * of the list - so the moved pages won't immediately be considered for
1030 * allocation again (e.g., optimization for memory onlining).
1032 static inline void move_to_free_list(struct page *page, struct zone *zone,
1033 unsigned int order, int migratetype)
1035 struct free_area *area = &zone->free_area[order];
1037 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
1040 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
1043 /* clear reported state and update reported page count */
1044 if (page_reported(page))
1045 __ClearPageReported(page);
1047 list_del(&page->buddy_list);
1048 __ClearPageBuddy(page);
1049 set_page_private(page, 0);
1050 zone->free_area[order].nr_free--;
1054 * If this is not the largest possible page, check if the buddy
1055 * of the next-highest order is free. If it is, it's possible
1056 * that pages are being freed that will coalesce soon. In case,
1057 * that is happening, add the free page to the tail of the list
1058 * so it's less likely to be used soon and more likely to be merged
1059 * as a higher order page
1062 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1063 struct page *page, unsigned int order)
1065 unsigned long higher_page_pfn;
1066 struct page *higher_page;
1068 if (order >= MAX_ORDER - 2)
1071 higher_page_pfn = buddy_pfn & pfn;
1072 higher_page = page + (higher_page_pfn - pfn);
1074 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
1079 * Freeing function for a buddy system allocator.
1081 * The concept of a buddy system is to maintain direct-mapped table
1082 * (containing bit values) for memory blocks of various "orders".
1083 * The bottom level table contains the map for the smallest allocatable
1084 * units of memory (here, pages), and each level above it describes
1085 * pairs of units from the levels below, hence, "buddies".
1086 * At a high level, all that happens here is marking the table entry
1087 * at the bottom level available, and propagating the changes upward
1088 * as necessary, plus some accounting needed to play nicely with other
1089 * parts of the VM system.
1090 * At each level, we keep a list of pages, which are heads of continuous
1091 * free pages of length of (1 << order) and marked with PageBuddy.
1092 * Page's order is recorded in page_private(page) field.
1093 * So when we are allocating or freeing one, we can derive the state of the
1094 * other. That is, if we allocate a small block, and both were
1095 * free, the remainder of the region must be split into blocks.
1096 * If a block is freed, and its buddy is also free, then this
1097 * triggers coalescing into a block of larger size.
1102 static inline void __free_one_page(struct page *page,
1104 struct zone *zone, unsigned int order,
1105 int migratetype, fpi_t fpi_flags)
1107 struct capture_control *capc = task_capc(zone);
1108 unsigned long buddy_pfn;
1109 unsigned long combined_pfn;
1113 VM_BUG_ON(!zone_is_initialized(zone));
1114 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1116 VM_BUG_ON(migratetype == -1);
1117 if (likely(!is_migrate_isolate(migratetype)))
1118 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1120 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1121 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1123 while (order < MAX_ORDER - 1) {
1124 if (compaction_capture(capc, page, order, migratetype)) {
1125 __mod_zone_freepage_state(zone, -(1 << order),
1130 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
1134 if (unlikely(order >= pageblock_order)) {
1136 * We want to prevent merge between freepages on pageblock
1137 * without fallbacks and normal pageblock. Without this,
1138 * pageblock isolation could cause incorrect freepage or CMA
1139 * accounting or HIGHATOMIC accounting.
1141 int buddy_mt = get_pageblock_migratetype(buddy);
1143 if (migratetype != buddy_mt
1144 && (!migratetype_is_mergeable(migratetype) ||
1145 !migratetype_is_mergeable(buddy_mt)))
1150 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1151 * merge with it and move up one order.
1153 if (page_is_guard(buddy))
1154 clear_page_guard(zone, buddy, order, migratetype);
1156 del_page_from_free_list(buddy, zone, order);
1157 combined_pfn = buddy_pfn & pfn;
1158 page = page + (combined_pfn - pfn);
1164 set_buddy_order(page, order);
1166 if (fpi_flags & FPI_TO_TAIL)
1168 else if (is_shuffle_order(order))
1169 to_tail = shuffle_pick_tail();
1171 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1174 add_to_free_list_tail(page, zone, order, migratetype);
1176 add_to_free_list(page, zone, order, migratetype);
1178 /* Notify page reporting subsystem of freed page */
1179 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1180 page_reporting_notify_free(order);
1184 * split_free_page() -- split a free page at split_pfn_offset
1185 * @free_page: the original free page
1186 * @order: the order of the page
1187 * @split_pfn_offset: split offset within the page
1189 * Return -ENOENT if the free page is changed, otherwise 0
1191 * It is used when the free page crosses two pageblocks with different migratetypes
1192 * at split_pfn_offset within the page. The split free page will be put into
1193 * separate migratetype lists afterwards. Otherwise, the function achieves
1196 int split_free_page(struct page *free_page,
1197 unsigned int order, unsigned long split_pfn_offset)
1199 struct zone *zone = page_zone(free_page);
1200 unsigned long free_page_pfn = page_to_pfn(free_page);
1202 unsigned long flags;
1203 int free_page_order;
1207 if (split_pfn_offset == 0)
1210 spin_lock_irqsave(&zone->lock, flags);
1212 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
1217 mt = get_pageblock_migratetype(free_page);
1218 if (likely(!is_migrate_isolate(mt)))
1219 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1221 del_page_from_free_list(free_page, zone, order);
1222 for (pfn = free_page_pfn;
1223 pfn < free_page_pfn + (1UL << order);) {
1224 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
1226 free_page_order = min_t(unsigned int,
1227 pfn ? __ffs(pfn) : order,
1228 __fls(split_pfn_offset));
1229 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
1231 pfn += 1UL << free_page_order;
1232 split_pfn_offset -= (1UL << free_page_order);
1233 /* we have done the first part, now switch to second part */
1234 if (split_pfn_offset == 0)
1235 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
1238 spin_unlock_irqrestore(&zone->lock, flags);
1242 * A bad page could be due to a number of fields. Instead of multiple branches,
1243 * try and check multiple fields with one check. The caller must do a detailed
1244 * check if necessary.
1246 static inline bool page_expected_state(struct page *page,
1247 unsigned long check_flags)
1249 if (unlikely(atomic_read(&page->_mapcount) != -1))
1252 if (unlikely((unsigned long)page->mapping |
1253 page_ref_count(page) |
1257 (page->flags & check_flags)))
1263 static const char *page_bad_reason(struct page *page, unsigned long flags)
1265 const char *bad_reason = NULL;
1267 if (unlikely(atomic_read(&page->_mapcount) != -1))
1268 bad_reason = "nonzero mapcount";
1269 if (unlikely(page->mapping != NULL))
1270 bad_reason = "non-NULL mapping";
1271 if (unlikely(page_ref_count(page) != 0))
1272 bad_reason = "nonzero _refcount";
1273 if (unlikely(page->flags & flags)) {
1274 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1275 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1277 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1280 if (unlikely(page->memcg_data))
1281 bad_reason = "page still charged to cgroup";
1286 static void check_free_page_bad(struct page *page)
1289 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1292 static inline int check_free_page(struct page *page)
1294 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1297 /* Something has gone sideways, find it */
1298 check_free_page_bad(page);
1302 static int free_tail_pages_check(struct page *head_page, struct page *page)
1307 * We rely page->lru.next never has bit 0 set, unless the page
1308 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1310 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1312 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1316 switch (page - head_page) {
1318 /* the first tail page: ->mapping may be compound_mapcount() */
1319 if (unlikely(compound_mapcount(page))) {
1320 bad_page(page, "nonzero compound_mapcount");
1326 * the second tail page: ->mapping is
1327 * deferred_list.next -- ignore value.
1331 if (page->mapping != TAIL_MAPPING) {
1332 bad_page(page, "corrupted mapping in tail page");
1337 if (unlikely(!PageTail(page))) {
1338 bad_page(page, "PageTail not set");
1341 if (unlikely(compound_head(page) != head_page)) {
1342 bad_page(page, "compound_head not consistent");
1347 page->mapping = NULL;
1348 clear_compound_head(page);
1353 * Skip KASAN memory poisoning when either:
1355 * 1. Deferred memory initialization has not yet completed,
1356 * see the explanation below.
1357 * 2. Skipping poisoning is requested via FPI_SKIP_KASAN_POISON,
1358 * see the comment next to it.
1359 * 3. Skipping poisoning is requested via __GFP_SKIP_KASAN_POISON,
1360 * see the comment next to it.
1362 * Poisoning pages during deferred memory init will greatly lengthen the
1363 * process and cause problem in large memory systems as the deferred pages
1364 * initialization is done with interrupt disabled.
1366 * Assuming that there will be no reference to those newly initialized
1367 * pages before they are ever allocated, this should have no effect on
1368 * KASAN memory tracking as the poison will be properly inserted at page
1369 * allocation time. The only corner case is when pages are allocated by
1370 * on-demand allocation and then freed again before the deferred pages
1371 * initialization is done, but this is not likely to happen.
1373 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1375 return deferred_pages_enabled() ||
1376 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
1377 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
1378 PageSkipKASanPoison(page);
1381 static void kernel_init_pages(struct page *page, int numpages)
1385 /* s390's use of memset() could override KASAN redzones. */
1386 kasan_disable_current();
1387 for (i = 0; i < numpages; i++)
1388 clear_highpage_kasan_tagged(page + i);
1389 kasan_enable_current();
1392 static __always_inline bool free_pages_prepare(struct page *page,
1393 unsigned int order, bool check_free, fpi_t fpi_flags)
1396 bool init = want_init_on_free();
1398 VM_BUG_ON_PAGE(PageTail(page), page);
1400 trace_mm_page_free(page, order);
1402 if (unlikely(PageHWPoison(page)) && !order) {
1404 * Do not let hwpoison pages hit pcplists/buddy
1405 * Untie memcg state and reset page's owner
1407 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1408 __memcg_kmem_uncharge_page(page, order);
1409 reset_page_owner(page, order);
1410 page_table_check_free(page, order);
1415 * Check tail pages before head page information is cleared to
1416 * avoid checking PageCompound for order-0 pages.
1418 if (unlikely(order)) {
1419 bool compound = PageCompound(page);
1422 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1425 ClearPageDoubleMap(page);
1426 ClearPageHasHWPoisoned(page);
1428 for (i = 1; i < (1 << order); i++) {
1430 bad += free_tail_pages_check(page, page + i);
1431 if (unlikely(check_free_page(page + i))) {
1435 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1438 if (PageMappingFlags(page))
1439 page->mapping = NULL;
1440 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1441 __memcg_kmem_uncharge_page(page, order);
1443 bad += check_free_page(page);
1447 page_cpupid_reset_last(page);
1448 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1449 reset_page_owner(page, order);
1450 page_table_check_free(page, order);
1452 if (!PageHighMem(page)) {
1453 debug_check_no_locks_freed(page_address(page),
1454 PAGE_SIZE << order);
1455 debug_check_no_obj_freed(page_address(page),
1456 PAGE_SIZE << order);
1459 kernel_poison_pages(page, 1 << order);
1462 * As memory initialization might be integrated into KASAN,
1463 * KASAN poisoning and memory initialization code must be
1464 * kept together to avoid discrepancies in behavior.
1466 * With hardware tag-based KASAN, memory tags must be set before the
1467 * page becomes unavailable via debug_pagealloc or arch_free_page.
1469 if (!should_skip_kasan_poison(page, fpi_flags)) {
1470 kasan_poison_pages(page, order, init);
1472 /* Memory is already initialized if KASAN did it internally. */
1473 if (kasan_has_integrated_init())
1477 kernel_init_pages(page, 1 << order);
1480 * arch_free_page() can make the page's contents inaccessible. s390
1481 * does this. So nothing which can access the page's contents should
1482 * happen after this.
1484 arch_free_page(page, order);
1486 debug_pagealloc_unmap_pages(page, 1 << order);
1491 #ifdef CONFIG_DEBUG_VM
1493 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1494 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1495 * moved from pcp lists to free lists.
1497 static bool free_pcp_prepare(struct page *page, unsigned int order)
1499 return free_pages_prepare(page, order, true, FPI_NONE);
1502 static bool bulkfree_pcp_prepare(struct page *page)
1504 if (debug_pagealloc_enabled_static())
1505 return check_free_page(page);
1511 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1512 * moving from pcp lists to free list in order to reduce overhead. With
1513 * debug_pagealloc enabled, they are checked also immediately when being freed
1516 static bool free_pcp_prepare(struct page *page, unsigned int order)
1518 if (debug_pagealloc_enabled_static())
1519 return free_pages_prepare(page, order, true, FPI_NONE);
1521 return free_pages_prepare(page, order, false, FPI_NONE);
1524 static bool bulkfree_pcp_prepare(struct page *page)
1526 return check_free_page(page);
1528 #endif /* CONFIG_DEBUG_VM */
1531 * Frees a number of pages from the PCP lists
1532 * Assumes all pages on list are in same zone.
1533 * count is the number of pages to free.
1535 static void free_pcppages_bulk(struct zone *zone, int count,
1536 struct per_cpu_pages *pcp,
1540 int max_pindex = NR_PCP_LISTS - 1;
1542 bool isolated_pageblocks;
1546 * Ensure proper count is passed which otherwise would stuck in the
1547 * below while (list_empty(list)) loop.
1549 count = min(pcp->count, count);
1551 /* Ensure requested pindex is drained first. */
1552 pindex = pindex - 1;
1554 /* Caller must hold IRQ-safe pcp->lock so IRQs are disabled. */
1555 spin_lock(&zone->lock);
1556 isolated_pageblocks = has_isolate_pageblock(zone);
1559 struct list_head *list;
1562 /* Remove pages from lists in a round-robin fashion. */
1564 if (++pindex > max_pindex)
1565 pindex = min_pindex;
1566 list = &pcp->lists[pindex];
1567 if (!list_empty(list))
1570 if (pindex == max_pindex)
1572 if (pindex == min_pindex)
1576 order = pindex_to_order(pindex);
1577 nr_pages = 1 << order;
1578 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1582 page = list_last_entry(list, struct page, pcp_list);
1583 mt = get_pcppage_migratetype(page);
1585 /* must delete to avoid corrupting pcp list */
1586 list_del(&page->pcp_list);
1588 pcp->count -= nr_pages;
1590 if (bulkfree_pcp_prepare(page))
1593 /* MIGRATE_ISOLATE page should not go to pcplists */
1594 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1595 /* Pageblock could have been isolated meanwhile */
1596 if (unlikely(isolated_pageblocks))
1597 mt = get_pageblock_migratetype(page);
1599 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1600 trace_mm_page_pcpu_drain(page, order, mt);
1601 } while (count > 0 && !list_empty(list));
1604 spin_unlock(&zone->lock);
1607 static void free_one_page(struct zone *zone,
1608 struct page *page, unsigned long pfn,
1610 int migratetype, fpi_t fpi_flags)
1612 unsigned long flags;
1614 spin_lock_irqsave(&zone->lock, flags);
1615 if (unlikely(has_isolate_pageblock(zone) ||
1616 is_migrate_isolate(migratetype))) {
1617 migratetype = get_pfnblock_migratetype(page, pfn);
1619 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1620 spin_unlock_irqrestore(&zone->lock, flags);
1623 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1624 unsigned long zone, int nid)
1626 mm_zero_struct_page(page);
1627 set_page_links(page, zone, nid, pfn);
1628 init_page_count(page);
1629 page_mapcount_reset(page);
1630 page_cpupid_reset_last(page);
1631 page_kasan_tag_reset(page);
1633 INIT_LIST_HEAD(&page->lru);
1634 #ifdef WANT_PAGE_VIRTUAL
1635 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1636 if (!is_highmem_idx(zone))
1637 set_page_address(page, __va(pfn << PAGE_SHIFT));
1641 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1642 static void __meminit init_reserved_page(unsigned long pfn)
1647 if (!early_page_uninitialised(pfn))
1650 nid = early_pfn_to_nid(pfn);
1651 pgdat = NODE_DATA(nid);
1653 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1654 struct zone *zone = &pgdat->node_zones[zid];
1656 if (zone_spans_pfn(zone, pfn))
1659 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1662 static inline void init_reserved_page(unsigned long pfn)
1665 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1668 * Initialised pages do not have PageReserved set. This function is
1669 * called for each range allocated by the bootmem allocator and
1670 * marks the pages PageReserved. The remaining valid pages are later
1671 * sent to the buddy page allocator.
1673 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1675 unsigned long start_pfn = PFN_DOWN(start);
1676 unsigned long end_pfn = PFN_UP(end);
1678 for (; start_pfn < end_pfn; start_pfn++) {
1679 if (pfn_valid(start_pfn)) {
1680 struct page *page = pfn_to_page(start_pfn);
1682 init_reserved_page(start_pfn);
1684 /* Avoid false-positive PageTail() */
1685 INIT_LIST_HEAD(&page->lru);
1688 * no need for atomic set_bit because the struct
1689 * page is not visible yet so nobody should
1692 __SetPageReserved(page);
1697 static void __free_pages_ok(struct page *page, unsigned int order,
1700 unsigned long flags;
1702 unsigned long pfn = page_to_pfn(page);
1703 struct zone *zone = page_zone(page);
1705 if (!free_pages_prepare(page, order, true, fpi_flags))
1708 migratetype = get_pfnblock_migratetype(page, pfn);
1710 spin_lock_irqsave(&zone->lock, flags);
1711 if (unlikely(has_isolate_pageblock(zone) ||
1712 is_migrate_isolate(migratetype))) {
1713 migratetype = get_pfnblock_migratetype(page, pfn);
1715 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1716 spin_unlock_irqrestore(&zone->lock, flags);
1718 __count_vm_events(PGFREE, 1 << order);
1721 void __free_pages_core(struct page *page, unsigned int order)
1723 unsigned int nr_pages = 1 << order;
1724 struct page *p = page;
1728 * When initializing the memmap, __init_single_page() sets the refcount
1729 * of all pages to 1 ("allocated"/"not free"). We have to set the
1730 * refcount of all involved pages to 0.
1733 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1735 __ClearPageReserved(p);
1736 set_page_count(p, 0);
1738 __ClearPageReserved(p);
1739 set_page_count(p, 0);
1741 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1744 * Bypass PCP and place fresh pages right to the tail, primarily
1745 * relevant for memory onlining.
1747 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1753 * During memory init memblocks map pfns to nids. The search is expensive and
1754 * this caches recent lookups. The implementation of __early_pfn_to_nid
1755 * treats start/end as pfns.
1757 struct mminit_pfnnid_cache {
1758 unsigned long last_start;
1759 unsigned long last_end;
1763 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1766 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1768 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1769 struct mminit_pfnnid_cache *state)
1771 unsigned long start_pfn, end_pfn;
1774 if (state->last_start <= pfn && pfn < state->last_end)
1775 return state->last_nid;
1777 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1778 if (nid != NUMA_NO_NODE) {
1779 state->last_start = start_pfn;
1780 state->last_end = end_pfn;
1781 state->last_nid = nid;
1787 int __meminit early_pfn_to_nid(unsigned long pfn)
1789 static DEFINE_SPINLOCK(early_pfn_lock);
1792 spin_lock(&early_pfn_lock);
1793 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1795 nid = first_online_node;
1796 spin_unlock(&early_pfn_lock);
1800 #endif /* CONFIG_NUMA */
1802 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1805 if (early_page_uninitialised(pfn))
1807 __free_pages_core(page, order);
1811 * Check that the whole (or subset of) a pageblock given by the interval of
1812 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1813 * with the migration of free compaction scanner.
1815 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1817 * It's possible on some configurations to have a setup like node0 node1 node0
1818 * i.e. it's possible that all pages within a zones range of pages do not
1819 * belong to a single zone. We assume that a border between node0 and node1
1820 * can occur within a single pageblock, but not a node0 node1 node0
1821 * interleaving within a single pageblock. It is therefore sufficient to check
1822 * the first and last page of a pageblock and avoid checking each individual
1823 * page in a pageblock.
1825 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1826 unsigned long end_pfn, struct zone *zone)
1828 struct page *start_page;
1829 struct page *end_page;
1831 /* end_pfn is one past the range we are checking */
1834 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1837 start_page = pfn_to_online_page(start_pfn);
1841 if (page_zone(start_page) != zone)
1844 end_page = pfn_to_page(end_pfn);
1846 /* This gives a shorter code than deriving page_zone(end_page) */
1847 if (page_zone_id(start_page) != page_zone_id(end_page))
1853 void set_zone_contiguous(struct zone *zone)
1855 unsigned long block_start_pfn = zone->zone_start_pfn;
1856 unsigned long block_end_pfn;
1858 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1859 for (; block_start_pfn < zone_end_pfn(zone);
1860 block_start_pfn = block_end_pfn,
1861 block_end_pfn += pageblock_nr_pages) {
1863 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1865 if (!__pageblock_pfn_to_page(block_start_pfn,
1866 block_end_pfn, zone))
1871 /* We confirm that there is no hole */
1872 zone->contiguous = true;
1875 void clear_zone_contiguous(struct zone *zone)
1877 zone->contiguous = false;
1880 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1881 static void __init deferred_free_range(unsigned long pfn,
1882 unsigned long nr_pages)
1890 page = pfn_to_page(pfn);
1892 /* Free a large naturally-aligned chunk if possible */
1893 if (nr_pages == pageblock_nr_pages &&
1894 (pfn & (pageblock_nr_pages - 1)) == 0) {
1895 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1896 __free_pages_core(page, pageblock_order);
1900 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1901 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1902 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1903 __free_pages_core(page, 0);
1907 /* Completion tracking for deferred_init_memmap() threads */
1908 static atomic_t pgdat_init_n_undone __initdata;
1909 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1911 static inline void __init pgdat_init_report_one_done(void)
1913 if (atomic_dec_and_test(&pgdat_init_n_undone))
1914 complete(&pgdat_init_all_done_comp);
1918 * Returns true if page needs to be initialized or freed to buddy allocator.
1920 * First we check if pfn is valid on architectures where it is possible to have
1921 * holes within pageblock_nr_pages. On systems where it is not possible, this
1922 * function is optimized out.
1924 * Then, we check if a current large page is valid by only checking the validity
1927 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1929 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1935 * Free pages to buddy allocator. Try to free aligned pages in
1936 * pageblock_nr_pages sizes.
1938 static void __init deferred_free_pages(unsigned long pfn,
1939 unsigned long end_pfn)
1941 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1942 unsigned long nr_free = 0;
1944 for (; pfn < end_pfn; pfn++) {
1945 if (!deferred_pfn_valid(pfn)) {
1946 deferred_free_range(pfn - nr_free, nr_free);
1948 } else if (!(pfn & nr_pgmask)) {
1949 deferred_free_range(pfn - nr_free, nr_free);
1955 /* Free the last block of pages to allocator */
1956 deferred_free_range(pfn - nr_free, nr_free);
1960 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1961 * by performing it only once every pageblock_nr_pages.
1962 * Return number of pages initialized.
1964 static unsigned long __init deferred_init_pages(struct zone *zone,
1966 unsigned long end_pfn)
1968 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1969 int nid = zone_to_nid(zone);
1970 unsigned long nr_pages = 0;
1971 int zid = zone_idx(zone);
1972 struct page *page = NULL;
1974 for (; pfn < end_pfn; pfn++) {
1975 if (!deferred_pfn_valid(pfn)) {
1978 } else if (!page || !(pfn & nr_pgmask)) {
1979 page = pfn_to_page(pfn);
1983 __init_single_page(page, pfn, zid, nid);
1990 * This function is meant to pre-load the iterator for the zone init.
1991 * Specifically it walks through the ranges until we are caught up to the
1992 * first_init_pfn value and exits there. If we never encounter the value we
1993 * return false indicating there are no valid ranges left.
1996 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1997 unsigned long *spfn, unsigned long *epfn,
1998 unsigned long first_init_pfn)
2003 * Start out by walking through the ranges in this zone that have
2004 * already been initialized. We don't need to do anything with them
2005 * so we just need to flush them out of the system.
2007 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
2008 if (*epfn <= first_init_pfn)
2010 if (*spfn < first_init_pfn)
2011 *spfn = first_init_pfn;
2020 * Initialize and free pages. We do it in two loops: first we initialize
2021 * struct page, then free to buddy allocator, because while we are
2022 * freeing pages we can access pages that are ahead (computing buddy
2023 * page in __free_one_page()).
2025 * In order to try and keep some memory in the cache we have the loop
2026 * broken along max page order boundaries. This way we will not cause
2027 * any issues with the buddy page computation.
2029 static unsigned long __init
2030 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
2031 unsigned long *end_pfn)
2033 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
2034 unsigned long spfn = *start_pfn, epfn = *end_pfn;
2035 unsigned long nr_pages = 0;
2038 /* First we loop through and initialize the page values */
2039 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
2042 if (mo_pfn <= *start_pfn)
2045 t = min(mo_pfn, *end_pfn);
2046 nr_pages += deferred_init_pages(zone, *start_pfn, t);
2048 if (mo_pfn < *end_pfn) {
2049 *start_pfn = mo_pfn;
2054 /* Reset values and now loop through freeing pages as needed */
2057 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2063 t = min(mo_pfn, epfn);
2064 deferred_free_pages(spfn, t);
2074 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2077 unsigned long spfn, epfn;
2078 struct zone *zone = arg;
2081 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2084 * Initialize and free pages in MAX_ORDER sized increments so that we
2085 * can avoid introducing any issues with the buddy allocator.
2087 while (spfn < end_pfn) {
2088 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2093 /* An arch may override for more concurrency. */
2095 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2100 /* Initialise remaining memory on a node */
2101 static int __init deferred_init_memmap(void *data)
2103 pg_data_t *pgdat = data;
2104 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2105 unsigned long spfn = 0, epfn = 0;
2106 unsigned long first_init_pfn, flags;
2107 unsigned long start = jiffies;
2109 int zid, max_threads;
2112 /* Bind memory initialisation thread to a local node if possible */
2113 if (!cpumask_empty(cpumask))
2114 set_cpus_allowed_ptr(current, cpumask);
2116 pgdat_resize_lock(pgdat, &flags);
2117 first_init_pfn = pgdat->first_deferred_pfn;
2118 if (first_init_pfn == ULONG_MAX) {
2119 pgdat_resize_unlock(pgdat, &flags);
2120 pgdat_init_report_one_done();
2124 /* Sanity check boundaries */
2125 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2126 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2127 pgdat->first_deferred_pfn = ULONG_MAX;
2130 * Once we unlock here, the zone cannot be grown anymore, thus if an
2131 * interrupt thread must allocate this early in boot, zone must be
2132 * pre-grown prior to start of deferred page initialization.
2134 pgdat_resize_unlock(pgdat, &flags);
2136 /* Only the highest zone is deferred so find it */
2137 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2138 zone = pgdat->node_zones + zid;
2139 if (first_init_pfn < zone_end_pfn(zone))
2143 /* If the zone is empty somebody else may have cleared out the zone */
2144 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2148 max_threads = deferred_page_init_max_threads(cpumask);
2150 while (spfn < epfn) {
2151 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2152 struct padata_mt_job job = {
2153 .thread_fn = deferred_init_memmap_chunk,
2156 .size = epfn_align - spfn,
2157 .align = PAGES_PER_SECTION,
2158 .min_chunk = PAGES_PER_SECTION,
2159 .max_threads = max_threads,
2162 padata_do_multithreaded(&job);
2163 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2167 /* Sanity check that the next zone really is unpopulated */
2168 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2170 pr_info("node %d deferred pages initialised in %ums\n",
2171 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2173 pgdat_init_report_one_done();
2178 * If this zone has deferred pages, try to grow it by initializing enough
2179 * deferred pages to satisfy the allocation specified by order, rounded up to
2180 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2181 * of SECTION_SIZE bytes by initializing struct pages in increments of
2182 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2184 * Return true when zone was grown, otherwise return false. We return true even
2185 * when we grow less than requested, to let the caller decide if there are
2186 * enough pages to satisfy the allocation.
2188 * Note: We use noinline because this function is needed only during boot, and
2189 * it is called from a __ref function _deferred_grow_zone. This way we are
2190 * making sure that it is not inlined into permanent text section.
2192 static noinline bool __init
2193 deferred_grow_zone(struct zone *zone, unsigned int order)
2195 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2196 pg_data_t *pgdat = zone->zone_pgdat;
2197 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2198 unsigned long spfn, epfn, flags;
2199 unsigned long nr_pages = 0;
2202 /* Only the last zone may have deferred pages */
2203 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2206 pgdat_resize_lock(pgdat, &flags);
2209 * If someone grew this zone while we were waiting for spinlock, return
2210 * true, as there might be enough pages already.
2212 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2213 pgdat_resize_unlock(pgdat, &flags);
2217 /* If the zone is empty somebody else may have cleared out the zone */
2218 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2219 first_deferred_pfn)) {
2220 pgdat->first_deferred_pfn = ULONG_MAX;
2221 pgdat_resize_unlock(pgdat, &flags);
2222 /* Retry only once. */
2223 return first_deferred_pfn != ULONG_MAX;
2227 * Initialize and free pages in MAX_ORDER sized increments so
2228 * that we can avoid introducing any issues with the buddy
2231 while (spfn < epfn) {
2232 /* update our first deferred PFN for this section */
2233 first_deferred_pfn = spfn;
2235 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2236 touch_nmi_watchdog();
2238 /* We should only stop along section boundaries */
2239 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2242 /* If our quota has been met we can stop here */
2243 if (nr_pages >= nr_pages_needed)
2247 pgdat->first_deferred_pfn = spfn;
2248 pgdat_resize_unlock(pgdat, &flags);
2250 return nr_pages > 0;
2254 * deferred_grow_zone() is __init, but it is called from
2255 * get_page_from_freelist() during early boot until deferred_pages permanently
2256 * disables this call. This is why we have refdata wrapper to avoid warning,
2257 * and to ensure that the function body gets unloaded.
2260 _deferred_grow_zone(struct zone *zone, unsigned int order)
2262 return deferred_grow_zone(zone, order);
2265 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2267 void __init page_alloc_init_late(void)
2272 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2274 /* There will be num_node_state(N_MEMORY) threads */
2275 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2276 for_each_node_state(nid, N_MEMORY) {
2277 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2280 /* Block until all are initialised */
2281 wait_for_completion(&pgdat_init_all_done_comp);
2284 * We initialized the rest of the deferred pages. Permanently disable
2285 * on-demand struct page initialization.
2287 static_branch_disable(&deferred_pages);
2289 /* Reinit limits that are based on free pages after the kernel is up */
2290 files_maxfiles_init();
2295 /* Discard memblock private memory */
2298 for_each_node_state(nid, N_MEMORY)
2299 shuffle_free_memory(NODE_DATA(nid));
2301 for_each_populated_zone(zone)
2302 set_zone_contiguous(zone);
2306 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2307 void __init init_cma_reserved_pageblock(struct page *page)
2309 unsigned i = pageblock_nr_pages;
2310 struct page *p = page;
2313 __ClearPageReserved(p);
2314 set_page_count(p, 0);
2317 set_pageblock_migratetype(page, MIGRATE_CMA);
2318 set_page_refcounted(page);
2319 __free_pages(page, pageblock_order);
2321 adjust_managed_page_count(page, pageblock_nr_pages);
2322 page_zone(page)->cma_pages += pageblock_nr_pages;
2327 * The order of subdivision here is critical for the IO subsystem.
2328 * Please do not alter this order without good reasons and regression
2329 * testing. Specifically, as large blocks of memory are subdivided,
2330 * the order in which smaller blocks are delivered depends on the order
2331 * they're subdivided in this function. This is the primary factor
2332 * influencing the order in which pages are delivered to the IO
2333 * subsystem according to empirical testing, and this is also justified
2334 * by considering the behavior of a buddy system containing a single
2335 * large block of memory acted on by a series of small allocations.
2336 * This behavior is a critical factor in sglist merging's success.
2340 static inline void expand(struct zone *zone, struct page *page,
2341 int low, int high, int migratetype)
2343 unsigned long size = 1 << high;
2345 while (high > low) {
2348 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2351 * Mark as guard pages (or page), that will allow to
2352 * merge back to allocator when buddy will be freed.
2353 * Corresponding page table entries will not be touched,
2354 * pages will stay not present in virtual address space
2356 if (set_page_guard(zone, &page[size], high, migratetype))
2359 add_to_free_list(&page[size], zone, high, migratetype);
2360 set_buddy_order(&page[size], high);
2364 static void check_new_page_bad(struct page *page)
2366 if (unlikely(page->flags & __PG_HWPOISON)) {
2367 /* Don't complain about hwpoisoned pages */
2368 page_mapcount_reset(page); /* remove PageBuddy */
2373 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2377 * This page is about to be returned from the page allocator
2379 static inline int check_new_page(struct page *page)
2381 if (likely(page_expected_state(page,
2382 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2385 check_new_page_bad(page);
2389 static bool check_new_pages(struct page *page, unsigned int order)
2392 for (i = 0; i < (1 << order); i++) {
2393 struct page *p = page + i;
2395 if (unlikely(check_new_page(p)))
2402 #ifdef CONFIG_DEBUG_VM
2404 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2405 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2406 * also checked when pcp lists are refilled from the free lists.
2408 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2410 if (debug_pagealloc_enabled_static())
2411 return check_new_pages(page, order);
2416 static inline bool check_new_pcp(struct page *page, unsigned int order)
2418 return check_new_pages(page, order);
2422 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2423 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2424 * enabled, they are also checked when being allocated from the pcp lists.
2426 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2428 return check_new_pages(page, order);
2430 static inline bool check_new_pcp(struct page *page, unsigned int order)
2432 if (debug_pagealloc_enabled_static())
2433 return check_new_pages(page, order);
2437 #endif /* CONFIG_DEBUG_VM */
2439 static inline bool should_skip_kasan_unpoison(gfp_t flags, bool init_tags)
2441 /* Don't skip if a software KASAN mode is enabled. */
2442 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
2443 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
2446 /* Skip, if hardware tag-based KASAN is not enabled. */
2447 if (!kasan_hw_tags_enabled())
2451 * With hardware tag-based KASAN enabled, skip if either:
2453 * 1. Memory tags have already been cleared via tag_clear_highpage().
2454 * 2. Skipping has been requested via __GFP_SKIP_KASAN_UNPOISON.
2456 return init_tags || (flags & __GFP_SKIP_KASAN_UNPOISON);
2459 static inline bool should_skip_init(gfp_t flags)
2461 /* Don't skip, if hardware tag-based KASAN is not enabled. */
2462 if (!kasan_hw_tags_enabled())
2465 /* For hardware tag-based KASAN, skip if requested. */
2466 return (flags & __GFP_SKIP_ZERO);
2469 inline void post_alloc_hook(struct page *page, unsigned int order,
2472 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
2473 !should_skip_init(gfp_flags);
2474 bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
2476 set_page_private(page, 0);
2477 set_page_refcounted(page);
2479 arch_alloc_page(page, order);
2480 debug_pagealloc_map_pages(page, 1 << order);
2483 * Page unpoisoning must happen before memory initialization.
2484 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2485 * allocations and the page unpoisoning code will complain.
2487 kernel_unpoison_pages(page, 1 << order);
2490 * As memory initialization might be integrated into KASAN,
2491 * KASAN unpoisoning and memory initializion code must be
2492 * kept together to avoid discrepancies in behavior.
2496 * If memory tags should be zeroed (which happens only when memory
2497 * should be initialized as well).
2502 /* Initialize both memory and tags. */
2503 for (i = 0; i != 1 << order; ++i)
2504 tag_clear_highpage(page + i);
2506 /* Note that memory is already initialized by the loop above. */
2509 if (!should_skip_kasan_unpoison(gfp_flags, init_tags)) {
2510 /* Unpoison shadow memory or set memory tags. */
2511 kasan_unpoison_pages(page, order, init);
2513 /* Note that memory is already initialized by KASAN. */
2514 if (kasan_has_integrated_init())
2517 /* If memory is still not initialized, do it now. */
2519 kernel_init_pages(page, 1 << order);
2520 /* Propagate __GFP_SKIP_KASAN_POISON to page flags. */
2521 if (kasan_hw_tags_enabled() && (gfp_flags & __GFP_SKIP_KASAN_POISON))
2522 SetPageSkipKASanPoison(page);
2524 set_page_owner(page, order, gfp_flags);
2525 page_table_check_alloc(page, order);
2528 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2529 unsigned int alloc_flags)
2531 post_alloc_hook(page, order, gfp_flags);
2533 if (order && (gfp_flags & __GFP_COMP))
2534 prep_compound_page(page, order);
2537 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2538 * allocate the page. The expectation is that the caller is taking
2539 * steps that will free more memory. The caller should avoid the page
2540 * being used for !PFMEMALLOC purposes.
2542 if (alloc_flags & ALLOC_NO_WATERMARKS)
2543 set_page_pfmemalloc(page);
2545 clear_page_pfmemalloc(page);
2549 * Go through the free lists for the given migratetype and remove
2550 * the smallest available page from the freelists
2552 static __always_inline
2553 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2556 unsigned int current_order;
2557 struct free_area *area;
2560 /* Find a page of the appropriate size in the preferred list */
2561 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2562 area = &(zone->free_area[current_order]);
2563 page = get_page_from_free_area(area, migratetype);
2566 del_page_from_free_list(page, zone, current_order);
2567 expand(zone, page, order, current_order, migratetype);
2568 set_pcppage_migratetype(page, migratetype);
2569 trace_mm_page_alloc_zone_locked(page, order, migratetype,
2570 pcp_allowed_order(order) &&
2571 migratetype < MIGRATE_PCPTYPES);
2580 * This array describes the order lists are fallen back to when
2581 * the free lists for the desirable migrate type are depleted
2583 * The other migratetypes do not have fallbacks.
2585 static int fallbacks[MIGRATE_TYPES][3] = {
2586 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2587 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2588 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2592 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2595 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2598 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2599 unsigned int order) { return NULL; }
2603 * Move the free pages in a range to the freelist tail of the requested type.
2604 * Note that start_page and end_pages are not aligned on a pageblock
2605 * boundary. If alignment is required, use move_freepages_block()
2607 static int move_freepages(struct zone *zone,
2608 unsigned long start_pfn, unsigned long end_pfn,
2609 int migratetype, int *num_movable)
2614 int pages_moved = 0;
2616 for (pfn = start_pfn; pfn <= end_pfn;) {
2617 page = pfn_to_page(pfn);
2618 if (!PageBuddy(page)) {
2620 * We assume that pages that could be isolated for
2621 * migration are movable. But we don't actually try
2622 * isolating, as that would be expensive.
2625 (PageLRU(page) || __PageMovable(page)))
2631 /* Make sure we are not inadvertently changing nodes */
2632 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2633 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2635 order = buddy_order(page);
2636 move_to_free_list(page, zone, order, migratetype);
2638 pages_moved += 1 << order;
2644 int move_freepages_block(struct zone *zone, struct page *page,
2645 int migratetype, int *num_movable)
2647 unsigned long start_pfn, end_pfn, pfn;
2652 pfn = page_to_pfn(page);
2653 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2654 end_pfn = start_pfn + pageblock_nr_pages - 1;
2656 /* Do not cross zone boundaries */
2657 if (!zone_spans_pfn(zone, start_pfn))
2659 if (!zone_spans_pfn(zone, end_pfn))
2662 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2666 static void change_pageblock_range(struct page *pageblock_page,
2667 int start_order, int migratetype)
2669 int nr_pageblocks = 1 << (start_order - pageblock_order);
2671 while (nr_pageblocks--) {
2672 set_pageblock_migratetype(pageblock_page, migratetype);
2673 pageblock_page += pageblock_nr_pages;
2678 * When we are falling back to another migratetype during allocation, try to
2679 * steal extra free pages from the same pageblocks to satisfy further
2680 * allocations, instead of polluting multiple pageblocks.
2682 * If we are stealing a relatively large buddy page, it is likely there will
2683 * be more free pages in the pageblock, so try to steal them all. For
2684 * reclaimable and unmovable allocations, we steal regardless of page size,
2685 * as fragmentation caused by those allocations polluting movable pageblocks
2686 * is worse than movable allocations stealing from unmovable and reclaimable
2689 static bool can_steal_fallback(unsigned int order, int start_mt)
2692 * Leaving this order check is intended, although there is
2693 * relaxed order check in next check. The reason is that
2694 * we can actually steal whole pageblock if this condition met,
2695 * but, below check doesn't guarantee it and that is just heuristic
2696 * so could be changed anytime.
2698 if (order >= pageblock_order)
2701 if (order >= pageblock_order / 2 ||
2702 start_mt == MIGRATE_RECLAIMABLE ||
2703 start_mt == MIGRATE_UNMOVABLE ||
2704 page_group_by_mobility_disabled)
2710 static inline bool boost_watermark(struct zone *zone)
2712 unsigned long max_boost;
2714 if (!watermark_boost_factor)
2717 * Don't bother in zones that are unlikely to produce results.
2718 * On small machines, including kdump capture kernels running
2719 * in a small area, boosting the watermark can cause an out of
2720 * memory situation immediately.
2722 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2725 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2726 watermark_boost_factor, 10000);
2729 * high watermark may be uninitialised if fragmentation occurs
2730 * very early in boot so do not boost. We do not fall
2731 * through and boost by pageblock_nr_pages as failing
2732 * allocations that early means that reclaim is not going
2733 * to help and it may even be impossible to reclaim the
2734 * boosted watermark resulting in a hang.
2739 max_boost = max(pageblock_nr_pages, max_boost);
2741 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2748 * This function implements actual steal behaviour. If order is large enough,
2749 * we can steal whole pageblock. If not, we first move freepages in this
2750 * pageblock to our migratetype and determine how many already-allocated pages
2751 * are there in the pageblock with a compatible migratetype. If at least half
2752 * of pages are free or compatible, we can change migratetype of the pageblock
2753 * itself, so pages freed in the future will be put on the correct free list.
2755 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2756 unsigned int alloc_flags, int start_type, bool whole_block)
2758 unsigned int current_order = buddy_order(page);
2759 int free_pages, movable_pages, alike_pages;
2762 old_block_type = get_pageblock_migratetype(page);
2765 * This can happen due to races and we want to prevent broken
2766 * highatomic accounting.
2768 if (is_migrate_highatomic(old_block_type))
2771 /* Take ownership for orders >= pageblock_order */
2772 if (current_order >= pageblock_order) {
2773 change_pageblock_range(page, current_order, start_type);
2778 * Boost watermarks to increase reclaim pressure to reduce the
2779 * likelihood of future fallbacks. Wake kswapd now as the node
2780 * may be balanced overall and kswapd will not wake naturally.
2782 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2783 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2785 /* We are not allowed to try stealing from the whole block */
2789 free_pages = move_freepages_block(zone, page, start_type,
2792 * Determine how many pages are compatible with our allocation.
2793 * For movable allocation, it's the number of movable pages which
2794 * we just obtained. For other types it's a bit more tricky.
2796 if (start_type == MIGRATE_MOVABLE) {
2797 alike_pages = movable_pages;
2800 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2801 * to MOVABLE pageblock, consider all non-movable pages as
2802 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2803 * vice versa, be conservative since we can't distinguish the
2804 * exact migratetype of non-movable pages.
2806 if (old_block_type == MIGRATE_MOVABLE)
2807 alike_pages = pageblock_nr_pages
2808 - (free_pages + movable_pages);
2813 /* moving whole block can fail due to zone boundary conditions */
2818 * If a sufficient number of pages in the block are either free or of
2819 * comparable migratability as our allocation, claim the whole block.
2821 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2822 page_group_by_mobility_disabled)
2823 set_pageblock_migratetype(page, start_type);
2828 move_to_free_list(page, zone, current_order, start_type);
2832 * Check whether there is a suitable fallback freepage with requested order.
2833 * If only_stealable is true, this function returns fallback_mt only if
2834 * we can steal other freepages all together. This would help to reduce
2835 * fragmentation due to mixed migratetype pages in one pageblock.
2837 int find_suitable_fallback(struct free_area *area, unsigned int order,
2838 int migratetype, bool only_stealable, bool *can_steal)
2843 if (area->nr_free == 0)
2848 fallback_mt = fallbacks[migratetype][i];
2849 if (fallback_mt == MIGRATE_TYPES)
2852 if (free_area_empty(area, fallback_mt))
2855 if (can_steal_fallback(order, migratetype))
2858 if (!only_stealable)
2869 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2870 * there are no empty page blocks that contain a page with a suitable order
2872 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2873 unsigned int alloc_order)
2876 unsigned long max_managed, flags;
2879 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2880 * Check is race-prone but harmless.
2882 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2883 if (zone->nr_reserved_highatomic >= max_managed)
2886 spin_lock_irqsave(&zone->lock, flags);
2888 /* Recheck the nr_reserved_highatomic limit under the lock */
2889 if (zone->nr_reserved_highatomic >= max_managed)
2893 mt = get_pageblock_migratetype(page);
2894 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2895 if (migratetype_is_mergeable(mt)) {
2896 zone->nr_reserved_highatomic += pageblock_nr_pages;
2897 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2898 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2902 spin_unlock_irqrestore(&zone->lock, flags);
2906 * Used when an allocation is about to fail under memory pressure. This
2907 * potentially hurts the reliability of high-order allocations when under
2908 * intense memory pressure but failed atomic allocations should be easier
2909 * to recover from than an OOM.
2911 * If @force is true, try to unreserve a pageblock even though highatomic
2912 * pageblock is exhausted.
2914 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2917 struct zonelist *zonelist = ac->zonelist;
2918 unsigned long flags;
2925 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2928 * Preserve at least one pageblock unless memory pressure
2931 if (!force && zone->nr_reserved_highatomic <=
2935 spin_lock_irqsave(&zone->lock, flags);
2936 for (order = 0; order < MAX_ORDER; order++) {
2937 struct free_area *area = &(zone->free_area[order]);
2939 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2944 * In page freeing path, migratetype change is racy so
2945 * we can counter several free pages in a pageblock
2946 * in this loop although we changed the pageblock type
2947 * from highatomic to ac->migratetype. So we should
2948 * adjust the count once.
2950 if (is_migrate_highatomic_page(page)) {
2952 * It should never happen but changes to
2953 * locking could inadvertently allow a per-cpu
2954 * drain to add pages to MIGRATE_HIGHATOMIC
2955 * while unreserving so be safe and watch for
2958 zone->nr_reserved_highatomic -= min(
2960 zone->nr_reserved_highatomic);
2964 * Convert to ac->migratetype and avoid the normal
2965 * pageblock stealing heuristics. Minimally, the caller
2966 * is doing the work and needs the pages. More
2967 * importantly, if the block was always converted to
2968 * MIGRATE_UNMOVABLE or another type then the number
2969 * of pageblocks that cannot be completely freed
2972 set_pageblock_migratetype(page, ac->migratetype);
2973 ret = move_freepages_block(zone, page, ac->migratetype,
2976 spin_unlock_irqrestore(&zone->lock, flags);
2980 spin_unlock_irqrestore(&zone->lock, flags);
2987 * Try finding a free buddy page on the fallback list and put it on the free
2988 * list of requested migratetype, possibly along with other pages from the same
2989 * block, depending on fragmentation avoidance heuristics. Returns true if
2990 * fallback was found so that __rmqueue_smallest() can grab it.
2992 * The use of signed ints for order and current_order is a deliberate
2993 * deviation from the rest of this file, to make the for loop
2994 * condition simpler.
2996 static __always_inline bool
2997 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2998 unsigned int alloc_flags)
3000 struct free_area *area;
3002 int min_order = order;
3008 * Do not steal pages from freelists belonging to other pageblocks
3009 * i.e. orders < pageblock_order. If there are no local zones free,
3010 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
3012 if (alloc_flags & ALLOC_NOFRAGMENT)
3013 min_order = pageblock_order;
3016 * Find the largest available free page in the other list. This roughly
3017 * approximates finding the pageblock with the most free pages, which
3018 * would be too costly to do exactly.
3020 for (current_order = MAX_ORDER - 1; current_order >= min_order;
3022 area = &(zone->free_area[current_order]);
3023 fallback_mt = find_suitable_fallback(area, current_order,
3024 start_migratetype, false, &can_steal);
3025 if (fallback_mt == -1)
3029 * We cannot steal all free pages from the pageblock and the
3030 * requested migratetype is movable. In that case it's better to
3031 * steal and split the smallest available page instead of the
3032 * largest available page, because even if the next movable
3033 * allocation falls back into a different pageblock than this
3034 * one, it won't cause permanent fragmentation.
3036 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
3037 && current_order > order)
3046 for (current_order = order; current_order < MAX_ORDER;
3048 area = &(zone->free_area[current_order]);
3049 fallback_mt = find_suitable_fallback(area, current_order,
3050 start_migratetype, false, &can_steal);
3051 if (fallback_mt != -1)
3056 * This should not happen - we already found a suitable fallback
3057 * when looking for the largest page.
3059 VM_BUG_ON(current_order == MAX_ORDER);
3062 page = get_page_from_free_area(area, fallback_mt);
3064 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
3067 trace_mm_page_alloc_extfrag(page, order, current_order,
3068 start_migratetype, fallback_mt);
3075 * Do the hard work of removing an element from the buddy allocator.
3076 * Call me with the zone->lock already held.
3078 static __always_inline struct page *
3079 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
3080 unsigned int alloc_flags)
3084 if (IS_ENABLED(CONFIG_CMA)) {
3086 * Balance movable allocations between regular and CMA areas by
3087 * allocating from CMA when over half of the zone's free memory
3088 * is in the CMA area.
3090 if (alloc_flags & ALLOC_CMA &&
3091 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3092 zone_page_state(zone, NR_FREE_PAGES) / 2) {
3093 page = __rmqueue_cma_fallback(zone, order);
3099 page = __rmqueue_smallest(zone, order, migratetype);
3100 if (unlikely(!page)) {
3101 if (alloc_flags & ALLOC_CMA)
3102 page = __rmqueue_cma_fallback(zone, order);
3104 if (!page && __rmqueue_fallback(zone, order, migratetype,
3112 * Obtain a specified number of elements from the buddy allocator, all under
3113 * a single hold of the lock, for efficiency. Add them to the supplied list.
3114 * Returns the number of new pages which were placed at *list.
3116 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3117 unsigned long count, struct list_head *list,
3118 int migratetype, unsigned int alloc_flags)
3120 int i, allocated = 0;
3122 /* Caller must hold IRQ-safe pcp->lock so IRQs are disabled. */
3123 spin_lock(&zone->lock);
3124 for (i = 0; i < count; ++i) {
3125 struct page *page = __rmqueue(zone, order, migratetype,
3127 if (unlikely(page == NULL))
3130 if (unlikely(check_pcp_refill(page, order)))
3134 * Split buddy pages returned by expand() are received here in
3135 * physical page order. The page is added to the tail of
3136 * caller's list. From the callers perspective, the linked list
3137 * is ordered by page number under some conditions. This is
3138 * useful for IO devices that can forward direction from the
3139 * head, thus also in the physical page order. This is useful
3140 * for IO devices that can merge IO requests if the physical
3141 * pages are ordered properly.
3143 list_add_tail(&page->pcp_list, list);
3145 if (is_migrate_cma(get_pcppage_migratetype(page)))
3146 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3151 * i pages were removed from the buddy list even if some leak due
3152 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3153 * on i. Do not confuse with 'allocated' which is the number of
3154 * pages added to the pcp list.
3156 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3157 spin_unlock(&zone->lock);
3163 * Called from the vmstat counter updater to drain pagesets of this
3164 * currently executing processor on remote nodes after they have
3167 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3169 int to_drain, batch;
3171 batch = READ_ONCE(pcp->batch);
3172 to_drain = min(pcp->count, batch);
3174 unsigned long flags;
3177 * free_pcppages_bulk expects IRQs disabled for zone->lock
3178 * so even though pcp->lock is not intended to be IRQ-safe,
3179 * it's needed in this context.
3181 spin_lock_irqsave(&pcp->lock, flags);
3182 free_pcppages_bulk(zone, to_drain, pcp, 0);
3183 spin_unlock_irqrestore(&pcp->lock, flags);
3189 * Drain pcplists of the indicated processor and zone.
3191 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3193 struct per_cpu_pages *pcp;
3195 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3197 unsigned long flags;
3199 /* See drain_zone_pages on why this is disabling IRQs */
3200 spin_lock_irqsave(&pcp->lock, flags);
3201 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3202 spin_unlock_irqrestore(&pcp->lock, flags);
3207 * Drain pcplists of all zones on the indicated processor.
3209 static void drain_pages(unsigned int cpu)
3213 for_each_populated_zone(zone) {
3214 drain_pages_zone(cpu, zone);
3219 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3221 void drain_local_pages(struct zone *zone)
3223 int cpu = smp_processor_id();
3226 drain_pages_zone(cpu, zone);
3232 * The implementation of drain_all_pages(), exposing an extra parameter to
3233 * drain on all cpus.
3235 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3236 * not empty. The check for non-emptiness can however race with a free to
3237 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3238 * that need the guarantee that every CPU has drained can disable the
3239 * optimizing racy check.
3241 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3246 * Allocate in the BSS so we won't require allocation in
3247 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3249 static cpumask_t cpus_with_pcps;
3252 * Do not drain if one is already in progress unless it's specific to
3253 * a zone. Such callers are primarily CMA and memory hotplug and need
3254 * the drain to be complete when the call returns.
3256 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3259 mutex_lock(&pcpu_drain_mutex);
3263 * We don't care about racing with CPU hotplug event
3264 * as offline notification will cause the notified
3265 * cpu to drain that CPU pcps and on_each_cpu_mask
3266 * disables preemption as part of its processing
3268 for_each_online_cpu(cpu) {
3269 struct per_cpu_pages *pcp;
3271 bool has_pcps = false;
3273 if (force_all_cpus) {
3275 * The pcp.count check is racy, some callers need a
3276 * guarantee that no cpu is missed.
3280 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3284 for_each_populated_zone(z) {
3285 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3294 cpumask_set_cpu(cpu, &cpus_with_pcps);
3296 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3299 for_each_cpu(cpu, &cpus_with_pcps) {
3301 drain_pages_zone(cpu, zone);
3306 mutex_unlock(&pcpu_drain_mutex);
3310 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3312 * When zone parameter is non-NULL, spill just the single zone's pages.
3314 void drain_all_pages(struct zone *zone)
3316 __drain_all_pages(zone, false);
3319 #ifdef CONFIG_HIBERNATION
3322 * Touch the watchdog for every WD_PAGE_COUNT pages.
3324 #define WD_PAGE_COUNT (128*1024)
3326 void mark_free_pages(struct zone *zone)
3328 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3329 unsigned long flags;
3330 unsigned int order, t;
3333 if (zone_is_empty(zone))
3336 spin_lock_irqsave(&zone->lock, flags);
3338 max_zone_pfn = zone_end_pfn(zone);
3339 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3340 if (pfn_valid(pfn)) {
3341 page = pfn_to_page(pfn);
3343 if (!--page_count) {
3344 touch_nmi_watchdog();
3345 page_count = WD_PAGE_COUNT;
3348 if (page_zone(page) != zone)
3351 if (!swsusp_page_is_forbidden(page))
3352 swsusp_unset_page_free(page);
3355 for_each_migratetype_order(order, t) {
3356 list_for_each_entry(page,
3357 &zone->free_area[order].free_list[t], buddy_list) {
3360 pfn = page_to_pfn(page);
3361 for (i = 0; i < (1UL << order); i++) {
3362 if (!--page_count) {
3363 touch_nmi_watchdog();
3364 page_count = WD_PAGE_COUNT;
3366 swsusp_set_page_free(pfn_to_page(pfn + i));
3370 spin_unlock_irqrestore(&zone->lock, flags);
3372 #endif /* CONFIG_PM */
3374 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3379 if (!free_pcp_prepare(page, order))
3382 migratetype = get_pfnblock_migratetype(page, pfn);
3383 set_pcppage_migratetype(page, migratetype);
3387 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3390 int min_nr_free, max_nr_free;
3392 /* Free everything if batch freeing high-order pages. */
3393 if (unlikely(free_high))
3396 /* Check for PCP disabled or boot pageset */
3397 if (unlikely(high < batch))
3400 /* Leave at least pcp->batch pages on the list */
3401 min_nr_free = batch;
3402 max_nr_free = high - batch;
3405 * Double the number of pages freed each time there is subsequent
3406 * freeing of pages without any allocation.
3408 batch <<= pcp->free_factor;
3409 if (batch < max_nr_free)
3411 batch = clamp(batch, min_nr_free, max_nr_free);
3416 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3419 int high = READ_ONCE(pcp->high);
3421 if (unlikely(!high || free_high))
3424 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3428 * If reclaim is active, limit the number of pages that can be
3429 * stored on pcp lists
3431 return min(READ_ONCE(pcp->batch) << 2, high);
3434 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
3435 struct page *page, int migratetype,
3442 __count_vm_event(PGFREE);
3443 pindex = order_to_pindex(migratetype, order);
3444 list_add(&page->pcp_list, &pcp->lists[pindex]);
3445 pcp->count += 1 << order;
3448 * As high-order pages other than THP's stored on PCP can contribute
3449 * to fragmentation, limit the number stored when PCP is heavily
3450 * freeing without allocation. The remainder after bulk freeing
3451 * stops will be drained from vmstat refresh context.
3453 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3455 high = nr_pcp_high(pcp, zone, free_high);
3456 if (pcp->count >= high) {
3457 int batch = READ_ONCE(pcp->batch);
3459 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3466 void free_unref_page(struct page *page, unsigned int order)
3468 unsigned long flags;
3469 unsigned long __maybe_unused UP_flags;
3470 struct per_cpu_pages *pcp;
3472 unsigned long pfn = page_to_pfn(page);
3475 if (!free_unref_page_prepare(page, pfn, order))
3479 * We only track unmovable, reclaimable and movable on pcp lists.
3480 * Place ISOLATE pages on the isolated list because they are being
3481 * offlined but treat HIGHATOMIC as movable pages so we can get those
3482 * areas back if necessary. Otherwise, we may have to free
3483 * excessively into the page allocator
3485 migratetype = get_pcppage_migratetype(page);
3486 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3487 if (unlikely(is_migrate_isolate(migratetype))) {
3488 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3491 migratetype = MIGRATE_MOVABLE;
3494 zone = page_zone(page);
3495 pcp_trylock_prepare(UP_flags);
3496 pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
3498 free_unref_page_commit(zone, pcp, page, migratetype, order);
3499 pcp_spin_unlock_irqrestore(pcp, flags);
3501 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
3503 pcp_trylock_finish(UP_flags);
3507 * Free a list of 0-order pages
3509 void free_unref_page_list(struct list_head *list)
3511 struct page *page, *next;
3512 struct per_cpu_pages *pcp = NULL;
3513 struct zone *locked_zone = NULL;
3514 unsigned long flags;
3515 int batch_count = 0;
3518 /* Prepare pages for freeing */
3519 list_for_each_entry_safe(page, next, list, lru) {
3520 unsigned long pfn = page_to_pfn(page);
3521 if (!free_unref_page_prepare(page, pfn, 0)) {
3522 list_del(&page->lru);
3527 * Free isolated pages directly to the allocator, see
3528 * comment in free_unref_page.
3530 migratetype = get_pcppage_migratetype(page);
3531 if (unlikely(is_migrate_isolate(migratetype))) {
3532 list_del(&page->lru);
3533 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3538 list_for_each_entry_safe(page, next, list, lru) {
3539 struct zone *zone = page_zone(page);
3541 /* Different zone, different pcp lock. */
3542 if (zone != locked_zone) {
3544 pcp_spin_unlock_irqrestore(pcp, flags);
3547 pcp = pcp_spin_lock_irqsave(locked_zone->per_cpu_pageset, flags);
3551 * Non-isolated types over MIGRATE_PCPTYPES get added
3552 * to the MIGRATE_MOVABLE pcp list.
3554 migratetype = get_pcppage_migratetype(page);
3555 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3556 migratetype = MIGRATE_MOVABLE;
3558 trace_mm_page_free_batched(page);
3559 free_unref_page_commit(zone, pcp, page, migratetype, 0);
3562 * Guard against excessive IRQ disabled times when we get
3563 * a large list of pages to free.
3565 if (++batch_count == SWAP_CLUSTER_MAX) {
3566 pcp_spin_unlock_irqrestore(pcp, flags);
3568 pcp = pcp_spin_lock_irqsave(locked_zone->per_cpu_pageset, flags);
3573 pcp_spin_unlock_irqrestore(pcp, flags);
3577 * split_page takes a non-compound higher-order page, and splits it into
3578 * n (1<<order) sub-pages: page[0..n]
3579 * Each sub-page must be freed individually.
3581 * Note: this is probably too low level an operation for use in drivers.
3582 * Please consult with lkml before using this in your driver.
3584 void split_page(struct page *page, unsigned int order)
3588 VM_BUG_ON_PAGE(PageCompound(page), page);
3589 VM_BUG_ON_PAGE(!page_count(page), page);
3591 for (i = 1; i < (1 << order); i++)
3592 set_page_refcounted(page + i);
3593 split_page_owner(page, 1 << order);
3594 split_page_memcg(page, 1 << order);
3596 EXPORT_SYMBOL_GPL(split_page);
3598 int __isolate_free_page(struct page *page, unsigned int order)
3600 unsigned long watermark;
3604 BUG_ON(!PageBuddy(page));
3606 zone = page_zone(page);
3607 mt = get_pageblock_migratetype(page);
3609 if (!is_migrate_isolate(mt)) {
3611 * Obey watermarks as if the page was being allocated. We can
3612 * emulate a high-order watermark check with a raised order-0
3613 * watermark, because we already know our high-order page
3616 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3617 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3620 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3623 /* Remove page from free list */
3625 del_page_from_free_list(page, zone, order);
3628 * Set the pageblock if the isolated page is at least half of a
3631 if (order >= pageblock_order - 1) {
3632 struct page *endpage = page + (1 << order) - 1;
3633 for (; page < endpage; page += pageblock_nr_pages) {
3634 int mt = get_pageblock_migratetype(page);
3636 * Only change normal pageblocks (i.e., they can merge
3639 if (migratetype_is_mergeable(mt))
3640 set_pageblock_migratetype(page,
3646 return 1UL << order;
3650 * __putback_isolated_page - Return a now-isolated page back where we got it
3651 * @page: Page that was isolated
3652 * @order: Order of the isolated page
3653 * @mt: The page's pageblock's migratetype
3655 * This function is meant to return a page pulled from the free lists via
3656 * __isolate_free_page back to the free lists they were pulled from.
3658 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3660 struct zone *zone = page_zone(page);
3662 /* zone lock should be held when this function is called */
3663 lockdep_assert_held(&zone->lock);
3665 /* Return isolated page to tail of freelist. */
3666 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3667 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3671 * Update NUMA hit/miss statistics
3673 * Must be called with interrupts disabled.
3675 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3679 enum numa_stat_item local_stat = NUMA_LOCAL;
3681 /* skip numa counters update if numa stats is disabled */
3682 if (!static_branch_likely(&vm_numa_stat_key))
3685 if (zone_to_nid(z) != numa_node_id())
3686 local_stat = NUMA_OTHER;
3688 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3689 __count_numa_events(z, NUMA_HIT, nr_account);
3691 __count_numa_events(z, NUMA_MISS, nr_account);
3692 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3694 __count_numa_events(z, local_stat, nr_account);
3698 static __always_inline
3699 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
3700 unsigned int order, unsigned int alloc_flags,
3704 unsigned long flags;
3708 spin_lock_irqsave(&zone->lock, flags);
3710 * order-0 request can reach here when the pcplist is skipped
3711 * due to non-CMA allocation context. HIGHATOMIC area is
3712 * reserved for high-order atomic allocation, so order-0
3713 * request should skip it.
3715 if (order > 0 && alloc_flags & ALLOC_HARDER)
3716 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3718 page = __rmqueue(zone, order, migratetype, alloc_flags);
3720 spin_unlock_irqrestore(&zone->lock, flags);
3724 __mod_zone_freepage_state(zone, -(1 << order),
3725 get_pcppage_migratetype(page));
3726 spin_unlock_irqrestore(&zone->lock, flags);
3727 } while (check_new_pages(page, order));
3729 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3730 zone_statistics(preferred_zone, zone, 1);
3735 /* Remove page from the per-cpu list, caller must protect the list */
3737 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3739 unsigned int alloc_flags,
3740 struct per_cpu_pages *pcp,
3741 struct list_head *list)
3746 if (list_empty(list)) {
3747 int batch = READ_ONCE(pcp->batch);
3751 * Scale batch relative to order if batch implies
3752 * free pages can be stored on the PCP. Batch can
3753 * be 1 for small zones or for boot pagesets which
3754 * should never store free pages as the pages may
3755 * belong to arbitrary zones.
3758 batch = max(batch >> order, 2);
3759 alloced = rmqueue_bulk(zone, order,
3761 migratetype, alloc_flags);
3763 pcp->count += alloced << order;
3764 if (unlikely(list_empty(list)))
3768 page = list_first_entry(list, struct page, pcp_list);
3769 list_del(&page->pcp_list);
3770 pcp->count -= 1 << order;
3771 } while (check_new_pcp(page, order));
3776 /* Lock and remove page from the per-cpu list */
3777 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3778 struct zone *zone, unsigned int order,
3779 gfp_t gfp_flags, int migratetype,
3780 unsigned int alloc_flags)
3782 struct per_cpu_pages *pcp;
3783 struct list_head *list;
3785 unsigned long flags;
3786 unsigned long __maybe_unused UP_flags;
3789 * spin_trylock may fail due to a parallel drain. In the future, the
3790 * trylock will also protect against IRQ reentrancy.
3792 pcp_trylock_prepare(UP_flags);
3793 pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
3795 pcp_trylock_finish(UP_flags);
3800 * On allocation, reduce the number of pages that are batch freed.
3801 * See nr_pcp_free() where free_factor is increased for subsequent
3804 pcp->free_factor >>= 1;
3805 list = &pcp->lists[order_to_pindex(migratetype, order)];
3806 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3807 pcp_spin_unlock_irqrestore(pcp, flags);
3808 pcp_trylock_finish(UP_flags);
3810 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3811 zone_statistics(preferred_zone, zone, 1);
3817 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3820 struct page *rmqueue(struct zone *preferred_zone,
3821 struct zone *zone, unsigned int order,
3822 gfp_t gfp_flags, unsigned int alloc_flags,
3828 * We most definitely don't want callers attempting to
3829 * allocate greater than order-1 page units with __GFP_NOFAIL.
3831 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3833 if (likely(pcp_allowed_order(order))) {
3835 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3836 * we need to skip it when CMA area isn't allowed.
3838 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3839 migratetype != MIGRATE_MOVABLE) {
3840 page = rmqueue_pcplist(preferred_zone, zone, order,
3841 gfp_flags, migratetype, alloc_flags);
3847 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3851 /* Separate test+clear to avoid unnecessary atomics */
3852 if (unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3853 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3854 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3857 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3861 #ifdef CONFIG_FAIL_PAGE_ALLOC
3864 struct fault_attr attr;
3866 bool ignore_gfp_highmem;
3867 bool ignore_gfp_reclaim;
3869 } fail_page_alloc = {
3870 .attr = FAULT_ATTR_INITIALIZER,
3871 .ignore_gfp_reclaim = true,
3872 .ignore_gfp_highmem = true,
3876 static int __init setup_fail_page_alloc(char *str)
3878 return setup_fault_attr(&fail_page_alloc.attr, str);
3880 __setup("fail_page_alloc=", setup_fail_page_alloc);
3882 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3884 if (order < fail_page_alloc.min_order)
3886 if (gfp_mask & __GFP_NOFAIL)
3888 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3890 if (fail_page_alloc.ignore_gfp_reclaim &&
3891 (gfp_mask & __GFP_DIRECT_RECLAIM))
3894 if (gfp_mask & __GFP_NOWARN)
3895 fail_page_alloc.attr.no_warn = true;
3897 return should_fail(&fail_page_alloc.attr, 1 << order);
3900 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3902 static int __init fail_page_alloc_debugfs(void)
3904 umode_t mode = S_IFREG | 0600;
3907 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3908 &fail_page_alloc.attr);
3910 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3911 &fail_page_alloc.ignore_gfp_reclaim);
3912 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3913 &fail_page_alloc.ignore_gfp_highmem);
3914 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3919 late_initcall(fail_page_alloc_debugfs);
3921 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3923 #else /* CONFIG_FAIL_PAGE_ALLOC */
3925 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3930 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3932 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3934 return __should_fail_alloc_page(gfp_mask, order);
3936 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3938 static inline long __zone_watermark_unusable_free(struct zone *z,
3939 unsigned int order, unsigned int alloc_flags)
3941 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3942 long unusable_free = (1 << order) - 1;
3945 * If the caller does not have rights to ALLOC_HARDER then subtract
3946 * the high-atomic reserves. This will over-estimate the size of the
3947 * atomic reserve but it avoids a search.
3949 if (likely(!alloc_harder))
3950 unusable_free += z->nr_reserved_highatomic;
3953 /* If allocation can't use CMA areas don't use free CMA pages */
3954 if (!(alloc_flags & ALLOC_CMA))
3955 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3958 return unusable_free;
3962 * Return true if free base pages are above 'mark'. For high-order checks it
3963 * will return true of the order-0 watermark is reached and there is at least
3964 * one free page of a suitable size. Checking now avoids taking the zone lock
3965 * to check in the allocation paths if no pages are free.
3967 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3968 int highest_zoneidx, unsigned int alloc_flags,
3973 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3975 /* free_pages may go negative - that's OK */
3976 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3978 if (alloc_flags & ALLOC_HIGH)
3981 if (unlikely(alloc_harder)) {
3983 * OOM victims can try even harder than normal ALLOC_HARDER
3984 * users on the grounds that it's definitely going to be in
3985 * the exit path shortly and free memory. Any allocation it
3986 * makes during the free path will be small and short-lived.
3988 if (alloc_flags & ALLOC_OOM)
3995 * Check watermarks for an order-0 allocation request. If these
3996 * are not met, then a high-order request also cannot go ahead
3997 * even if a suitable page happened to be free.
3999 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
4002 /* If this is an order-0 request then the watermark is fine */
4006 /* For a high-order request, check at least one suitable page is free */
4007 for (o = order; o < MAX_ORDER; o++) {
4008 struct free_area *area = &z->free_area[o];
4014 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
4015 if (!free_area_empty(area, mt))
4020 if ((alloc_flags & ALLOC_CMA) &&
4021 !free_area_empty(area, MIGRATE_CMA)) {
4025 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
4031 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
4032 int highest_zoneidx, unsigned int alloc_flags)
4034 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4035 zone_page_state(z, NR_FREE_PAGES));
4038 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
4039 unsigned long mark, int highest_zoneidx,
4040 unsigned int alloc_flags, gfp_t gfp_mask)
4044 free_pages = zone_page_state(z, NR_FREE_PAGES);
4047 * Fast check for order-0 only. If this fails then the reserves
4048 * need to be calculated.
4053 fast_free = free_pages;
4054 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
4055 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
4059 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4063 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
4064 * when checking the min watermark. The min watermark is the
4065 * point where boosting is ignored so that kswapd is woken up
4066 * when below the low watermark.
4068 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
4069 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
4070 mark = z->_watermark[WMARK_MIN];
4071 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
4072 alloc_flags, free_pages);
4078 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
4079 unsigned long mark, int highest_zoneidx)
4081 long free_pages = zone_page_state(z, NR_FREE_PAGES);
4083 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
4084 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
4086 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
4091 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
4093 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4095 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
4096 node_reclaim_distance;
4098 #else /* CONFIG_NUMA */
4099 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4103 #endif /* CONFIG_NUMA */
4106 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4107 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4108 * premature use of a lower zone may cause lowmem pressure problems that
4109 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4110 * probably too small. It only makes sense to spread allocations to avoid
4111 * fragmentation between the Normal and DMA32 zones.
4113 static inline unsigned int
4114 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4116 unsigned int alloc_flags;
4119 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4122 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4124 #ifdef CONFIG_ZONE_DMA32
4128 if (zone_idx(zone) != ZONE_NORMAL)
4132 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4133 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4134 * on UMA that if Normal is populated then so is DMA32.
4136 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4137 if (nr_online_nodes > 1 && !populated_zone(--zone))
4140 alloc_flags |= ALLOC_NOFRAGMENT;
4141 #endif /* CONFIG_ZONE_DMA32 */
4145 /* Must be called after current_gfp_context() which can change gfp_mask */
4146 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4147 unsigned int alloc_flags)
4150 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4151 alloc_flags |= ALLOC_CMA;
4157 * get_page_from_freelist goes through the zonelist trying to allocate
4160 static struct page *
4161 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4162 const struct alloc_context *ac)
4166 struct pglist_data *last_pgdat = NULL;
4167 bool last_pgdat_dirty_ok = false;
4172 * Scan zonelist, looking for a zone with enough free.
4173 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4175 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4176 z = ac->preferred_zoneref;
4177 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4182 if (cpusets_enabled() &&
4183 (alloc_flags & ALLOC_CPUSET) &&
4184 !__cpuset_zone_allowed(zone, gfp_mask))
4187 * When allocating a page cache page for writing, we
4188 * want to get it from a node that is within its dirty
4189 * limit, such that no single node holds more than its
4190 * proportional share of globally allowed dirty pages.
4191 * The dirty limits take into account the node's
4192 * lowmem reserves and high watermark so that kswapd
4193 * should be able to balance it without having to
4194 * write pages from its LRU list.
4196 * XXX: For now, allow allocations to potentially
4197 * exceed the per-node dirty limit in the slowpath
4198 * (spread_dirty_pages unset) before going into reclaim,
4199 * which is important when on a NUMA setup the allowed
4200 * nodes are together not big enough to reach the
4201 * global limit. The proper fix for these situations
4202 * will require awareness of nodes in the
4203 * dirty-throttling and the flusher threads.
4205 if (ac->spread_dirty_pages) {
4206 if (last_pgdat != zone->zone_pgdat) {
4207 last_pgdat = zone->zone_pgdat;
4208 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
4211 if (!last_pgdat_dirty_ok)
4215 if (no_fallback && nr_online_nodes > 1 &&
4216 zone != ac->preferred_zoneref->zone) {
4220 * If moving to a remote node, retry but allow
4221 * fragmenting fallbacks. Locality is more important
4222 * than fragmentation avoidance.
4224 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4225 if (zone_to_nid(zone) != local_nid) {
4226 alloc_flags &= ~ALLOC_NOFRAGMENT;
4231 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4232 if (!zone_watermark_fast(zone, order, mark,
4233 ac->highest_zoneidx, alloc_flags,
4237 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4239 * Watermark failed for this zone, but see if we can
4240 * grow this zone if it contains deferred pages.
4242 if (static_branch_unlikely(&deferred_pages)) {
4243 if (_deferred_grow_zone(zone, order))
4247 /* Checked here to keep the fast path fast */
4248 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4249 if (alloc_flags & ALLOC_NO_WATERMARKS)
4252 if (!node_reclaim_enabled() ||
4253 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4256 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4258 case NODE_RECLAIM_NOSCAN:
4261 case NODE_RECLAIM_FULL:
4262 /* scanned but unreclaimable */
4265 /* did we reclaim enough */
4266 if (zone_watermark_ok(zone, order, mark,
4267 ac->highest_zoneidx, alloc_flags))
4275 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4276 gfp_mask, alloc_flags, ac->migratetype);
4278 prep_new_page(page, order, gfp_mask, alloc_flags);
4281 * If this is a high-order atomic allocation then check
4282 * if the pageblock should be reserved for the future
4284 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4285 reserve_highatomic_pageblock(page, zone, order);
4289 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4290 /* Try again if zone has deferred pages */
4291 if (static_branch_unlikely(&deferred_pages)) {
4292 if (_deferred_grow_zone(zone, order))
4300 * It's possible on a UMA machine to get through all zones that are
4301 * fragmented. If avoiding fragmentation, reset and try again.
4304 alloc_flags &= ~ALLOC_NOFRAGMENT;
4311 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4313 unsigned int filter = SHOW_MEM_FILTER_NODES;
4316 * This documents exceptions given to allocations in certain
4317 * contexts that are allowed to allocate outside current's set
4320 if (!(gfp_mask & __GFP_NOMEMALLOC))
4321 if (tsk_is_oom_victim(current) ||
4322 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4323 filter &= ~SHOW_MEM_FILTER_NODES;
4324 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4325 filter &= ~SHOW_MEM_FILTER_NODES;
4327 show_mem(filter, nodemask);
4330 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4332 struct va_format vaf;
4334 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4336 if ((gfp_mask & __GFP_NOWARN) ||
4337 !__ratelimit(&nopage_rs) ||
4338 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4341 va_start(args, fmt);
4344 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4345 current->comm, &vaf, gfp_mask, &gfp_mask,
4346 nodemask_pr_args(nodemask));
4349 cpuset_print_current_mems_allowed();
4352 warn_alloc_show_mem(gfp_mask, nodemask);
4355 static inline struct page *
4356 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4357 unsigned int alloc_flags,
4358 const struct alloc_context *ac)
4362 page = get_page_from_freelist(gfp_mask, order,
4363 alloc_flags|ALLOC_CPUSET, ac);
4365 * fallback to ignore cpuset restriction if our nodes
4369 page = get_page_from_freelist(gfp_mask, order,
4375 static inline struct page *
4376 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4377 const struct alloc_context *ac, unsigned long *did_some_progress)
4379 struct oom_control oc = {
4380 .zonelist = ac->zonelist,
4381 .nodemask = ac->nodemask,
4383 .gfp_mask = gfp_mask,
4388 *did_some_progress = 0;
4391 * Acquire the oom lock. If that fails, somebody else is
4392 * making progress for us.
4394 if (!mutex_trylock(&oom_lock)) {
4395 *did_some_progress = 1;
4396 schedule_timeout_uninterruptible(1);
4401 * Go through the zonelist yet one more time, keep very high watermark
4402 * here, this is only to catch a parallel oom killing, we must fail if
4403 * we're still under heavy pressure. But make sure that this reclaim
4404 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4405 * allocation which will never fail due to oom_lock already held.
4407 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4408 ~__GFP_DIRECT_RECLAIM, order,
4409 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4413 /* Coredumps can quickly deplete all memory reserves */
4414 if (current->flags & PF_DUMPCORE)
4416 /* The OOM killer will not help higher order allocs */
4417 if (order > PAGE_ALLOC_COSTLY_ORDER)
4420 * We have already exhausted all our reclaim opportunities without any
4421 * success so it is time to admit defeat. We will skip the OOM killer
4422 * because it is very likely that the caller has a more reasonable
4423 * fallback than shooting a random task.
4425 * The OOM killer may not free memory on a specific node.
4427 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4429 /* The OOM killer does not needlessly kill tasks for lowmem */
4430 if (ac->highest_zoneidx < ZONE_NORMAL)
4432 if (pm_suspended_storage())
4435 * XXX: GFP_NOFS allocations should rather fail than rely on
4436 * other request to make a forward progress.
4437 * We are in an unfortunate situation where out_of_memory cannot
4438 * do much for this context but let's try it to at least get
4439 * access to memory reserved if the current task is killed (see
4440 * out_of_memory). Once filesystems are ready to handle allocation
4441 * failures more gracefully we should just bail out here.
4444 /* Exhausted what can be done so it's blame time */
4445 if (out_of_memory(&oc) ||
4446 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4447 *did_some_progress = 1;
4450 * Help non-failing allocations by giving them access to memory
4453 if (gfp_mask & __GFP_NOFAIL)
4454 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4455 ALLOC_NO_WATERMARKS, ac);
4458 mutex_unlock(&oom_lock);
4463 * Maximum number of compaction retries with a progress before OOM
4464 * killer is consider as the only way to move forward.
4466 #define MAX_COMPACT_RETRIES 16
4468 #ifdef CONFIG_COMPACTION
4469 /* Try memory compaction for high-order allocations before reclaim */
4470 static struct page *
4471 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4472 unsigned int alloc_flags, const struct alloc_context *ac,
4473 enum compact_priority prio, enum compact_result *compact_result)
4475 struct page *page = NULL;
4476 unsigned long pflags;
4477 unsigned int noreclaim_flag;
4482 psi_memstall_enter(&pflags);
4483 delayacct_compact_start();
4484 noreclaim_flag = memalloc_noreclaim_save();
4486 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4489 memalloc_noreclaim_restore(noreclaim_flag);
4490 psi_memstall_leave(&pflags);
4491 delayacct_compact_end();
4493 if (*compact_result == COMPACT_SKIPPED)
4496 * At least in one zone compaction wasn't deferred or skipped, so let's
4497 * count a compaction stall
4499 count_vm_event(COMPACTSTALL);
4501 /* Prep a captured page if available */
4503 prep_new_page(page, order, gfp_mask, alloc_flags);
4505 /* Try get a page from the freelist if available */
4507 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4510 struct zone *zone = page_zone(page);
4512 zone->compact_blockskip_flush = false;
4513 compaction_defer_reset(zone, order, true);
4514 count_vm_event(COMPACTSUCCESS);
4519 * It's bad if compaction run occurs and fails. The most likely reason
4520 * is that pages exist, but not enough to satisfy watermarks.
4522 count_vm_event(COMPACTFAIL);
4530 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4531 enum compact_result compact_result,
4532 enum compact_priority *compact_priority,
4533 int *compaction_retries)
4535 int max_retries = MAX_COMPACT_RETRIES;
4538 int retries = *compaction_retries;
4539 enum compact_priority priority = *compact_priority;
4544 if (fatal_signal_pending(current))
4547 if (compaction_made_progress(compact_result))
4548 (*compaction_retries)++;
4551 * compaction considers all the zone as desperately out of memory
4552 * so it doesn't really make much sense to retry except when the
4553 * failure could be caused by insufficient priority
4555 if (compaction_failed(compact_result))
4556 goto check_priority;
4559 * compaction was skipped because there are not enough order-0 pages
4560 * to work with, so we retry only if it looks like reclaim can help.
4562 if (compaction_needs_reclaim(compact_result)) {
4563 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4568 * make sure the compaction wasn't deferred or didn't bail out early
4569 * due to locks contention before we declare that we should give up.
4570 * But the next retry should use a higher priority if allowed, so
4571 * we don't just keep bailing out endlessly.
4573 if (compaction_withdrawn(compact_result)) {
4574 goto check_priority;
4578 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4579 * costly ones because they are de facto nofail and invoke OOM
4580 * killer to move on while costly can fail and users are ready
4581 * to cope with that. 1/4 retries is rather arbitrary but we
4582 * would need much more detailed feedback from compaction to
4583 * make a better decision.
4585 if (order > PAGE_ALLOC_COSTLY_ORDER)
4587 if (*compaction_retries <= max_retries) {
4593 * Make sure there are attempts at the highest priority if we exhausted
4594 * all retries or failed at the lower priorities.
4597 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4598 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4600 if (*compact_priority > min_priority) {
4601 (*compact_priority)--;
4602 *compaction_retries = 0;
4606 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4610 static inline struct page *
4611 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4612 unsigned int alloc_flags, const struct alloc_context *ac,
4613 enum compact_priority prio, enum compact_result *compact_result)
4615 *compact_result = COMPACT_SKIPPED;
4620 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4621 enum compact_result compact_result,
4622 enum compact_priority *compact_priority,
4623 int *compaction_retries)
4628 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4632 * There are setups with compaction disabled which would prefer to loop
4633 * inside the allocator rather than hit the oom killer prematurely.
4634 * Let's give them a good hope and keep retrying while the order-0
4635 * watermarks are OK.
4637 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4638 ac->highest_zoneidx, ac->nodemask) {
4639 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4640 ac->highest_zoneidx, alloc_flags))
4645 #endif /* CONFIG_COMPACTION */
4647 #ifdef CONFIG_LOCKDEP
4648 static struct lockdep_map __fs_reclaim_map =
4649 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4651 static bool __need_reclaim(gfp_t gfp_mask)
4653 /* no reclaim without waiting on it */
4654 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4657 /* this guy won't enter reclaim */
4658 if (current->flags & PF_MEMALLOC)
4661 if (gfp_mask & __GFP_NOLOCKDEP)
4667 void __fs_reclaim_acquire(unsigned long ip)
4669 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4672 void __fs_reclaim_release(unsigned long ip)
4674 lock_release(&__fs_reclaim_map, ip);
4677 void fs_reclaim_acquire(gfp_t gfp_mask)
4679 gfp_mask = current_gfp_context(gfp_mask);
4681 if (__need_reclaim(gfp_mask)) {
4682 if (gfp_mask & __GFP_FS)
4683 __fs_reclaim_acquire(_RET_IP_);
4685 #ifdef CONFIG_MMU_NOTIFIER
4686 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4687 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4692 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4694 void fs_reclaim_release(gfp_t gfp_mask)
4696 gfp_mask = current_gfp_context(gfp_mask);
4698 if (__need_reclaim(gfp_mask)) {
4699 if (gfp_mask & __GFP_FS)
4700 __fs_reclaim_release(_RET_IP_);
4703 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4706 /* Perform direct synchronous page reclaim */
4707 static unsigned long
4708 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4709 const struct alloc_context *ac)
4711 unsigned int noreclaim_flag;
4712 unsigned long progress;
4716 /* We now go into synchronous reclaim */
4717 cpuset_memory_pressure_bump();
4718 fs_reclaim_acquire(gfp_mask);
4719 noreclaim_flag = memalloc_noreclaim_save();
4721 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4724 memalloc_noreclaim_restore(noreclaim_flag);
4725 fs_reclaim_release(gfp_mask);
4732 /* The really slow allocator path where we enter direct reclaim */
4733 static inline struct page *
4734 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4735 unsigned int alloc_flags, const struct alloc_context *ac,
4736 unsigned long *did_some_progress)
4738 struct page *page = NULL;
4739 unsigned long pflags;
4740 bool drained = false;
4742 psi_memstall_enter(&pflags);
4743 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4744 if (unlikely(!(*did_some_progress)))
4748 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4751 * If an allocation failed after direct reclaim, it could be because
4752 * pages are pinned on the per-cpu lists or in high alloc reserves.
4753 * Shrink them and try again
4755 if (!page && !drained) {
4756 unreserve_highatomic_pageblock(ac, false);
4757 drain_all_pages(NULL);
4762 psi_memstall_leave(&pflags);
4767 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4768 const struct alloc_context *ac)
4772 pg_data_t *last_pgdat = NULL;
4773 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4775 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4777 if (!managed_zone(zone))
4779 if (last_pgdat != zone->zone_pgdat) {
4780 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4781 last_pgdat = zone->zone_pgdat;
4786 static inline unsigned int
4787 gfp_to_alloc_flags(gfp_t gfp_mask)
4789 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4792 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4793 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4794 * to save two branches.
4796 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4797 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4800 * The caller may dip into page reserves a bit more if the caller
4801 * cannot run direct reclaim, or if the caller has realtime scheduling
4802 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4803 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4805 alloc_flags |= (__force int)
4806 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4808 if (gfp_mask & __GFP_ATOMIC) {
4810 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4811 * if it can't schedule.
4813 if (!(gfp_mask & __GFP_NOMEMALLOC))
4814 alloc_flags |= ALLOC_HARDER;
4816 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4817 * comment for __cpuset_node_allowed().
4819 alloc_flags &= ~ALLOC_CPUSET;
4820 } else if (unlikely(rt_task(current)) && in_task())
4821 alloc_flags |= ALLOC_HARDER;
4823 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4828 static bool oom_reserves_allowed(struct task_struct *tsk)
4830 if (!tsk_is_oom_victim(tsk))
4834 * !MMU doesn't have oom reaper so give access to memory reserves
4835 * only to the thread with TIF_MEMDIE set
4837 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4844 * Distinguish requests which really need access to full memory
4845 * reserves from oom victims which can live with a portion of it
4847 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4849 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4851 if (gfp_mask & __GFP_MEMALLOC)
4852 return ALLOC_NO_WATERMARKS;
4853 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4854 return ALLOC_NO_WATERMARKS;
4855 if (!in_interrupt()) {
4856 if (current->flags & PF_MEMALLOC)
4857 return ALLOC_NO_WATERMARKS;
4858 else if (oom_reserves_allowed(current))
4865 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4867 return !!__gfp_pfmemalloc_flags(gfp_mask);
4871 * Checks whether it makes sense to retry the reclaim to make a forward progress
4872 * for the given allocation request.
4874 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4875 * without success, or when we couldn't even meet the watermark if we
4876 * reclaimed all remaining pages on the LRU lists.
4878 * Returns true if a retry is viable or false to enter the oom path.
4881 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4882 struct alloc_context *ac, int alloc_flags,
4883 bool did_some_progress, int *no_progress_loops)
4890 * Costly allocations might have made a progress but this doesn't mean
4891 * their order will become available due to high fragmentation so
4892 * always increment the no progress counter for them
4894 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4895 *no_progress_loops = 0;
4897 (*no_progress_loops)++;
4900 * Make sure we converge to OOM if we cannot make any progress
4901 * several times in the row.
4903 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4904 /* Before OOM, exhaust highatomic_reserve */
4905 return unreserve_highatomic_pageblock(ac, true);
4909 * Keep reclaiming pages while there is a chance this will lead
4910 * somewhere. If none of the target zones can satisfy our allocation
4911 * request even if all reclaimable pages are considered then we are
4912 * screwed and have to go OOM.
4914 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4915 ac->highest_zoneidx, ac->nodemask) {
4916 unsigned long available;
4917 unsigned long reclaimable;
4918 unsigned long min_wmark = min_wmark_pages(zone);
4921 available = reclaimable = zone_reclaimable_pages(zone);
4922 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4925 * Would the allocation succeed if we reclaimed all
4926 * reclaimable pages?
4928 wmark = __zone_watermark_ok(zone, order, min_wmark,
4929 ac->highest_zoneidx, alloc_flags, available);
4930 trace_reclaim_retry_zone(z, order, reclaimable,
4931 available, min_wmark, *no_progress_loops, wmark);
4939 * Memory allocation/reclaim might be called from a WQ context and the
4940 * current implementation of the WQ concurrency control doesn't
4941 * recognize that a particular WQ is congested if the worker thread is
4942 * looping without ever sleeping. Therefore we have to do a short sleep
4943 * here rather than calling cond_resched().
4945 if (current->flags & PF_WQ_WORKER)
4946 schedule_timeout_uninterruptible(1);
4953 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4956 * It's possible that cpuset's mems_allowed and the nodemask from
4957 * mempolicy don't intersect. This should be normally dealt with by
4958 * policy_nodemask(), but it's possible to race with cpuset update in
4959 * such a way the check therein was true, and then it became false
4960 * before we got our cpuset_mems_cookie here.
4961 * This assumes that for all allocations, ac->nodemask can come only
4962 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4963 * when it does not intersect with the cpuset restrictions) or the
4964 * caller can deal with a violated nodemask.
4966 if (cpusets_enabled() && ac->nodemask &&
4967 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4968 ac->nodemask = NULL;
4973 * When updating a task's mems_allowed or mempolicy nodemask, it is
4974 * possible to race with parallel threads in such a way that our
4975 * allocation can fail while the mask is being updated. If we are about
4976 * to fail, check if the cpuset changed during allocation and if so,
4979 if (read_mems_allowed_retry(cpuset_mems_cookie))
4985 static inline struct page *
4986 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4987 struct alloc_context *ac)
4989 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4990 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4991 struct page *page = NULL;
4992 unsigned int alloc_flags;
4993 unsigned long did_some_progress;
4994 enum compact_priority compact_priority;
4995 enum compact_result compact_result;
4996 int compaction_retries;
4997 int no_progress_loops;
4998 unsigned int cpuset_mems_cookie;
5002 * We also sanity check to catch abuse of atomic reserves being used by
5003 * callers that are not in atomic context.
5005 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
5006 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
5007 gfp_mask &= ~__GFP_ATOMIC;
5010 compaction_retries = 0;
5011 no_progress_loops = 0;
5012 compact_priority = DEF_COMPACT_PRIORITY;
5013 cpuset_mems_cookie = read_mems_allowed_begin();
5016 * The fast path uses conservative alloc_flags to succeed only until
5017 * kswapd needs to be woken up, and to avoid the cost of setting up
5018 * alloc_flags precisely. So we do that now.
5020 alloc_flags = gfp_to_alloc_flags(gfp_mask);
5023 * We need to recalculate the starting point for the zonelist iterator
5024 * because we might have used different nodemask in the fast path, or
5025 * there was a cpuset modification and we are retrying - otherwise we
5026 * could end up iterating over non-eligible zones endlessly.
5028 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5029 ac->highest_zoneidx, ac->nodemask);
5030 if (!ac->preferred_zoneref->zone)
5034 * Check for insane configurations where the cpuset doesn't contain
5035 * any suitable zone to satisfy the request - e.g. non-movable
5036 * GFP_HIGHUSER allocations from MOVABLE nodes only.
5038 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
5039 struct zoneref *z = first_zones_zonelist(ac->zonelist,
5040 ac->highest_zoneidx,
5041 &cpuset_current_mems_allowed);
5046 if (alloc_flags & ALLOC_KSWAPD)
5047 wake_all_kswapds(order, gfp_mask, ac);
5050 * The adjusted alloc_flags might result in immediate success, so try
5053 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5058 * For costly allocations, try direct compaction first, as it's likely
5059 * that we have enough base pages and don't need to reclaim. For non-
5060 * movable high-order allocations, do that as well, as compaction will
5061 * try prevent permanent fragmentation by migrating from blocks of the
5063 * Don't try this for allocations that are allowed to ignore
5064 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
5066 if (can_direct_reclaim &&
5068 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
5069 && !gfp_pfmemalloc_allowed(gfp_mask)) {
5070 page = __alloc_pages_direct_compact(gfp_mask, order,
5072 INIT_COMPACT_PRIORITY,
5078 * Checks for costly allocations with __GFP_NORETRY, which
5079 * includes some THP page fault allocations
5081 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
5083 * If allocating entire pageblock(s) and compaction
5084 * failed because all zones are below low watermarks
5085 * or is prohibited because it recently failed at this
5086 * order, fail immediately unless the allocator has
5087 * requested compaction and reclaim retry.
5090 * - potentially very expensive because zones are far
5091 * below their low watermarks or this is part of very
5092 * bursty high order allocations,
5093 * - not guaranteed to help because isolate_freepages()
5094 * may not iterate over freed pages as part of its
5096 * - unlikely to make entire pageblocks free on its
5099 if (compact_result == COMPACT_SKIPPED ||
5100 compact_result == COMPACT_DEFERRED)
5104 * Looks like reclaim/compaction is worth trying, but
5105 * sync compaction could be very expensive, so keep
5106 * using async compaction.
5108 compact_priority = INIT_COMPACT_PRIORITY;
5113 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5114 if (alloc_flags & ALLOC_KSWAPD)
5115 wake_all_kswapds(order, gfp_mask, ac);
5117 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5119 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
5122 * Reset the nodemask and zonelist iterators if memory policies can be
5123 * ignored. These allocations are high priority and system rather than
5126 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5127 ac->nodemask = NULL;
5128 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5129 ac->highest_zoneidx, ac->nodemask);
5132 /* Attempt with potentially adjusted zonelist and alloc_flags */
5133 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5137 /* Caller is not willing to reclaim, we can't balance anything */
5138 if (!can_direct_reclaim)
5141 /* Avoid recursion of direct reclaim */
5142 if (current->flags & PF_MEMALLOC)
5145 /* Try direct reclaim and then allocating */
5146 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5147 &did_some_progress);
5151 /* Try direct compaction and then allocating */
5152 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5153 compact_priority, &compact_result);
5157 /* Do not loop if specifically requested */
5158 if (gfp_mask & __GFP_NORETRY)
5162 * Do not retry costly high order allocations unless they are
5163 * __GFP_RETRY_MAYFAIL
5165 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5168 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5169 did_some_progress > 0, &no_progress_loops))
5173 * It doesn't make any sense to retry for the compaction if the order-0
5174 * reclaim is not able to make any progress because the current
5175 * implementation of the compaction depends on the sufficient amount
5176 * of free memory (see __compaction_suitable)
5178 if (did_some_progress > 0 &&
5179 should_compact_retry(ac, order, alloc_flags,
5180 compact_result, &compact_priority,
5181 &compaction_retries))
5185 /* Deal with possible cpuset update races before we start OOM killing */
5186 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5189 /* Reclaim has failed us, start killing things */
5190 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5194 /* Avoid allocations with no watermarks from looping endlessly */
5195 if (tsk_is_oom_victim(current) &&
5196 (alloc_flags & ALLOC_OOM ||
5197 (gfp_mask & __GFP_NOMEMALLOC)))
5200 /* Retry as long as the OOM killer is making progress */
5201 if (did_some_progress) {
5202 no_progress_loops = 0;
5207 /* Deal with possible cpuset update races before we fail */
5208 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5212 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5215 if (gfp_mask & __GFP_NOFAIL) {
5217 * All existing users of the __GFP_NOFAIL are blockable, so warn
5218 * of any new users that actually require GFP_NOWAIT
5220 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
5224 * PF_MEMALLOC request from this context is rather bizarre
5225 * because we cannot reclaim anything and only can loop waiting
5226 * for somebody to do a work for us
5228 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
5231 * non failing costly orders are a hard requirement which we
5232 * are not prepared for much so let's warn about these users
5233 * so that we can identify them and convert them to something
5236 WARN_ON_ONCE_GFP(order > PAGE_ALLOC_COSTLY_ORDER, gfp_mask);
5239 * Help non-failing allocations by giving them access to memory
5240 * reserves but do not use ALLOC_NO_WATERMARKS because this
5241 * could deplete whole memory reserves which would just make
5242 * the situation worse
5244 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5252 warn_alloc(gfp_mask, ac->nodemask,
5253 "page allocation failure: order:%u", order);
5258 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5259 int preferred_nid, nodemask_t *nodemask,
5260 struct alloc_context *ac, gfp_t *alloc_gfp,
5261 unsigned int *alloc_flags)
5263 ac->highest_zoneidx = gfp_zone(gfp_mask);
5264 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5265 ac->nodemask = nodemask;
5266 ac->migratetype = gfp_migratetype(gfp_mask);
5268 if (cpusets_enabled()) {
5269 *alloc_gfp |= __GFP_HARDWALL;
5271 * When we are in the interrupt context, it is irrelevant
5272 * to the current task context. It means that any node ok.
5274 if (in_task() && !ac->nodemask)
5275 ac->nodemask = &cpuset_current_mems_allowed;
5277 *alloc_flags |= ALLOC_CPUSET;
5280 might_alloc(gfp_mask);
5282 if (should_fail_alloc_page(gfp_mask, order))
5285 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5287 /* Dirty zone balancing only done in the fast path */
5288 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5291 * The preferred zone is used for statistics but crucially it is
5292 * also used as the starting point for the zonelist iterator. It
5293 * may get reset for allocations that ignore memory policies.
5295 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5296 ac->highest_zoneidx, ac->nodemask);
5302 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5303 * @gfp: GFP flags for the allocation
5304 * @preferred_nid: The preferred NUMA node ID to allocate from
5305 * @nodemask: Set of nodes to allocate from, may be NULL
5306 * @nr_pages: The number of pages desired on the list or array
5307 * @page_list: Optional list to store the allocated pages
5308 * @page_array: Optional array to store the pages
5310 * This is a batched version of the page allocator that attempts to
5311 * allocate nr_pages quickly. Pages are added to page_list if page_list
5312 * is not NULL, otherwise it is assumed that the page_array is valid.
5314 * For lists, nr_pages is the number of pages that should be allocated.
5316 * For arrays, only NULL elements are populated with pages and nr_pages
5317 * is the maximum number of pages that will be stored in the array.
5319 * Returns the number of pages on the list or array.
5321 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5322 nodemask_t *nodemask, int nr_pages,
5323 struct list_head *page_list,
5324 struct page **page_array)
5327 unsigned long flags;
5328 unsigned long __maybe_unused UP_flags;
5331 struct per_cpu_pages *pcp;
5332 struct list_head *pcp_list;
5333 struct alloc_context ac;
5335 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5336 int nr_populated = 0, nr_account = 0;
5339 * Skip populated array elements to determine if any pages need
5340 * to be allocated before disabling IRQs.
5342 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5345 /* No pages requested? */
5346 if (unlikely(nr_pages <= 0))
5349 /* Already populated array? */
5350 if (unlikely(page_array && nr_pages - nr_populated == 0))
5353 /* Bulk allocator does not support memcg accounting. */
5354 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5357 /* Use the single page allocator for one page. */
5358 if (nr_pages - nr_populated == 1)
5361 #ifdef CONFIG_PAGE_OWNER
5363 * PAGE_OWNER may recurse into the allocator to allocate space to
5364 * save the stack with pagesets.lock held. Releasing/reacquiring
5365 * removes much of the performance benefit of bulk allocation so
5366 * force the caller to allocate one page at a time as it'll have
5367 * similar performance to added complexity to the bulk allocator.
5369 if (static_branch_unlikely(&page_owner_inited))
5373 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5374 gfp &= gfp_allowed_mask;
5376 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5380 /* Find an allowed local zone that meets the low watermark. */
5381 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5384 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5385 !__cpuset_zone_allowed(zone, gfp)) {
5389 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5390 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5394 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5395 if (zone_watermark_fast(zone, 0, mark,
5396 zonelist_zone_idx(ac.preferred_zoneref),
5397 alloc_flags, gfp)) {
5403 * If there are no allowed local zones that meets the watermarks then
5404 * try to allocate a single page and reclaim if necessary.
5406 if (unlikely(!zone))
5409 /* Is a parallel drain in progress? */
5410 pcp_trylock_prepare(UP_flags);
5411 pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
5415 /* Attempt the batch allocation */
5416 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5417 while (nr_populated < nr_pages) {
5419 /* Skip existing pages */
5420 if (page_array && page_array[nr_populated]) {
5425 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5427 if (unlikely(!page)) {
5428 /* Try and allocate at least one page */
5430 pcp_spin_unlock_irqrestore(pcp, flags);
5437 prep_new_page(page, 0, gfp, 0);
5439 list_add(&page->lru, page_list);
5441 page_array[nr_populated] = page;
5445 pcp_spin_unlock_irqrestore(pcp, flags);
5446 pcp_trylock_finish(UP_flags);
5448 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5449 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5452 return nr_populated;
5455 pcp_trylock_finish(UP_flags);
5458 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5461 list_add(&page->lru, page_list);
5463 page_array[nr_populated] = page;
5469 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5472 * This is the 'heart' of the zoned buddy allocator.
5474 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5475 nodemask_t *nodemask)
5478 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5479 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5480 struct alloc_context ac = { };
5483 * There are several places where we assume that the order value is sane
5484 * so bail out early if the request is out of bound.
5486 if (WARN_ON_ONCE_GFP(order >= MAX_ORDER, gfp))
5489 gfp &= gfp_allowed_mask;
5491 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5492 * resp. GFP_NOIO which has to be inherited for all allocation requests
5493 * from a particular context which has been marked by
5494 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5495 * movable zones are not used during allocation.
5497 gfp = current_gfp_context(gfp);
5499 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5500 &alloc_gfp, &alloc_flags))
5504 * Forbid the first pass from falling back to types that fragment
5505 * memory until all local zones are considered.
5507 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5509 /* First allocation attempt */
5510 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5515 ac.spread_dirty_pages = false;
5518 * Restore the original nodemask if it was potentially replaced with
5519 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5521 ac.nodemask = nodemask;
5523 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5526 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5527 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5528 __free_pages(page, order);
5532 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5536 EXPORT_SYMBOL(__alloc_pages);
5538 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5539 nodemask_t *nodemask)
5541 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5542 preferred_nid, nodemask);
5544 if (page && order > 1)
5545 prep_transhuge_page(page);
5546 return (struct folio *)page;
5548 EXPORT_SYMBOL(__folio_alloc);
5551 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5552 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5553 * you need to access high mem.
5555 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5559 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5562 return (unsigned long) page_address(page);
5564 EXPORT_SYMBOL(__get_free_pages);
5566 unsigned long get_zeroed_page(gfp_t gfp_mask)
5568 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5570 EXPORT_SYMBOL(get_zeroed_page);
5573 * __free_pages - Free pages allocated with alloc_pages().
5574 * @page: The page pointer returned from alloc_pages().
5575 * @order: The order of the allocation.
5577 * This function can free multi-page allocations that are not compound
5578 * pages. It does not check that the @order passed in matches that of
5579 * the allocation, so it is easy to leak memory. Freeing more memory
5580 * than was allocated will probably emit a warning.
5582 * If the last reference to this page is speculative, it will be released
5583 * by put_page() which only frees the first page of a non-compound
5584 * allocation. To prevent the remaining pages from being leaked, we free
5585 * the subsequent pages here. If you want to use the page's reference
5586 * count to decide when to free the allocation, you should allocate a
5587 * compound page, and use put_page() instead of __free_pages().
5589 * Context: May be called in interrupt context or while holding a normal
5590 * spinlock, but not in NMI context or while holding a raw spinlock.
5592 void __free_pages(struct page *page, unsigned int order)
5594 if (put_page_testzero(page))
5595 free_the_page(page, order);
5596 else if (!PageHead(page))
5598 free_the_page(page + (1 << order), order);
5600 EXPORT_SYMBOL(__free_pages);
5602 void free_pages(unsigned long addr, unsigned int order)
5605 VM_BUG_ON(!virt_addr_valid((void *)addr));
5606 __free_pages(virt_to_page((void *)addr), order);
5610 EXPORT_SYMBOL(free_pages);
5614 * An arbitrary-length arbitrary-offset area of memory which resides
5615 * within a 0 or higher order page. Multiple fragments within that page
5616 * are individually refcounted, in the page's reference counter.
5618 * The page_frag functions below provide a simple allocation framework for
5619 * page fragments. This is used by the network stack and network device
5620 * drivers to provide a backing region of memory for use as either an
5621 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5623 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5626 struct page *page = NULL;
5627 gfp_t gfp = gfp_mask;
5629 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5630 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5632 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5633 PAGE_FRAG_CACHE_MAX_ORDER);
5634 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5636 if (unlikely(!page))
5637 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5639 nc->va = page ? page_address(page) : NULL;
5644 void __page_frag_cache_drain(struct page *page, unsigned int count)
5646 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5648 if (page_ref_sub_and_test(page, count))
5649 free_the_page(page, compound_order(page));
5651 EXPORT_SYMBOL(__page_frag_cache_drain);
5653 void *page_frag_alloc_align(struct page_frag_cache *nc,
5654 unsigned int fragsz, gfp_t gfp_mask,
5655 unsigned int align_mask)
5657 unsigned int size = PAGE_SIZE;
5661 if (unlikely(!nc->va)) {
5663 page = __page_frag_cache_refill(nc, gfp_mask);
5667 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5668 /* if size can vary use size else just use PAGE_SIZE */
5671 /* Even if we own the page, we do not use atomic_set().
5672 * This would break get_page_unless_zero() users.
5674 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5676 /* reset page count bias and offset to start of new frag */
5677 nc->pfmemalloc = page_is_pfmemalloc(page);
5678 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5682 offset = nc->offset - fragsz;
5683 if (unlikely(offset < 0)) {
5684 page = virt_to_page(nc->va);
5686 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5689 if (unlikely(nc->pfmemalloc)) {
5690 free_the_page(page, compound_order(page));
5694 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5695 /* if size can vary use size else just use PAGE_SIZE */
5698 /* OK, page count is 0, we can safely set it */
5699 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5701 /* reset page count bias and offset to start of new frag */
5702 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5703 offset = size - fragsz;
5707 offset &= align_mask;
5708 nc->offset = offset;
5710 return nc->va + offset;
5712 EXPORT_SYMBOL(page_frag_alloc_align);
5715 * Frees a page fragment allocated out of either a compound or order 0 page.
5717 void page_frag_free(void *addr)
5719 struct page *page = virt_to_head_page(addr);
5721 if (unlikely(put_page_testzero(page)))
5722 free_the_page(page, compound_order(page));
5724 EXPORT_SYMBOL(page_frag_free);
5726 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5730 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5731 unsigned long used = addr + PAGE_ALIGN(size);
5733 split_page(virt_to_page((void *)addr), order);
5734 while (used < alloc_end) {
5739 return (void *)addr;
5743 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5744 * @size: the number of bytes to allocate
5745 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5747 * This function is similar to alloc_pages(), except that it allocates the
5748 * minimum number of pages to satisfy the request. alloc_pages() can only
5749 * allocate memory in power-of-two pages.
5751 * This function is also limited by MAX_ORDER.
5753 * Memory allocated by this function must be released by free_pages_exact().
5755 * Return: pointer to the allocated area or %NULL in case of error.
5757 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5759 unsigned int order = get_order(size);
5762 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5763 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5765 addr = __get_free_pages(gfp_mask, order);
5766 return make_alloc_exact(addr, order, size);
5768 EXPORT_SYMBOL(alloc_pages_exact);
5771 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5773 * @nid: the preferred node ID where memory should be allocated
5774 * @size: the number of bytes to allocate
5775 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5777 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5780 * Return: pointer to the allocated area or %NULL in case of error.
5782 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5784 unsigned int order = get_order(size);
5787 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5788 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5790 p = alloc_pages_node(nid, gfp_mask, order);
5793 return make_alloc_exact((unsigned long)page_address(p), order, size);
5797 * free_pages_exact - release memory allocated via alloc_pages_exact()
5798 * @virt: the value returned by alloc_pages_exact.
5799 * @size: size of allocation, same value as passed to alloc_pages_exact().
5801 * Release the memory allocated by a previous call to alloc_pages_exact.
5803 void free_pages_exact(void *virt, size_t size)
5805 unsigned long addr = (unsigned long)virt;
5806 unsigned long end = addr + PAGE_ALIGN(size);
5808 while (addr < end) {
5813 EXPORT_SYMBOL(free_pages_exact);
5816 * nr_free_zone_pages - count number of pages beyond high watermark
5817 * @offset: The zone index of the highest zone
5819 * nr_free_zone_pages() counts the number of pages which are beyond the
5820 * high watermark within all zones at or below a given zone index. For each
5821 * zone, the number of pages is calculated as:
5823 * nr_free_zone_pages = managed_pages - high_pages
5825 * Return: number of pages beyond high watermark.
5827 static unsigned long nr_free_zone_pages(int offset)
5832 /* Just pick one node, since fallback list is circular */
5833 unsigned long sum = 0;
5835 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5837 for_each_zone_zonelist(zone, z, zonelist, offset) {
5838 unsigned long size = zone_managed_pages(zone);
5839 unsigned long high = high_wmark_pages(zone);
5848 * nr_free_buffer_pages - count number of pages beyond high watermark
5850 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5851 * watermark within ZONE_DMA and ZONE_NORMAL.
5853 * Return: number of pages beyond high watermark within ZONE_DMA and
5856 unsigned long nr_free_buffer_pages(void)
5858 return nr_free_zone_pages(gfp_zone(GFP_USER));
5860 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5862 static inline void show_node(struct zone *zone)
5864 if (IS_ENABLED(CONFIG_NUMA))
5865 printk("Node %d ", zone_to_nid(zone));
5868 long si_mem_available(void)
5871 unsigned long pagecache;
5872 unsigned long wmark_low = 0;
5873 unsigned long pages[NR_LRU_LISTS];
5874 unsigned long reclaimable;
5878 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5879 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5882 wmark_low += low_wmark_pages(zone);
5885 * Estimate the amount of memory available for userspace allocations,
5886 * without causing swapping or OOM.
5888 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5891 * Not all the page cache can be freed, otherwise the system will
5892 * start swapping or thrashing. Assume at least half of the page
5893 * cache, or the low watermark worth of cache, needs to stay.
5895 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5896 pagecache -= min(pagecache / 2, wmark_low);
5897 available += pagecache;
5900 * Part of the reclaimable slab and other kernel memory consists of
5901 * items that are in use, and cannot be freed. Cap this estimate at the
5904 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5905 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5906 available += reclaimable - min(reclaimable / 2, wmark_low);
5912 EXPORT_SYMBOL_GPL(si_mem_available);
5914 void si_meminfo(struct sysinfo *val)
5916 val->totalram = totalram_pages();
5917 val->sharedram = global_node_page_state(NR_SHMEM);
5918 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5919 val->bufferram = nr_blockdev_pages();
5920 val->totalhigh = totalhigh_pages();
5921 val->freehigh = nr_free_highpages();
5922 val->mem_unit = PAGE_SIZE;
5925 EXPORT_SYMBOL(si_meminfo);
5928 void si_meminfo_node(struct sysinfo *val, int nid)
5930 int zone_type; /* needs to be signed */
5931 unsigned long managed_pages = 0;
5932 unsigned long managed_highpages = 0;
5933 unsigned long free_highpages = 0;
5934 pg_data_t *pgdat = NODE_DATA(nid);
5936 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5937 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5938 val->totalram = managed_pages;
5939 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5940 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5941 #ifdef CONFIG_HIGHMEM
5942 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5943 struct zone *zone = &pgdat->node_zones[zone_type];
5945 if (is_highmem(zone)) {
5946 managed_highpages += zone_managed_pages(zone);
5947 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5950 val->totalhigh = managed_highpages;
5951 val->freehigh = free_highpages;
5953 val->totalhigh = managed_highpages;
5954 val->freehigh = free_highpages;
5956 val->mem_unit = PAGE_SIZE;
5961 * Determine whether the node should be displayed or not, depending on whether
5962 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5964 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5966 if (!(flags & SHOW_MEM_FILTER_NODES))
5970 * no node mask - aka implicit memory numa policy. Do not bother with
5971 * the synchronization - read_mems_allowed_begin - because we do not
5972 * have to be precise here.
5975 nodemask = &cpuset_current_mems_allowed;
5977 return !node_isset(nid, *nodemask);
5980 #define K(x) ((x) << (PAGE_SHIFT-10))
5982 static void show_migration_types(unsigned char type)
5984 static const char types[MIGRATE_TYPES] = {
5985 [MIGRATE_UNMOVABLE] = 'U',
5986 [MIGRATE_MOVABLE] = 'M',
5987 [MIGRATE_RECLAIMABLE] = 'E',
5988 [MIGRATE_HIGHATOMIC] = 'H',
5990 [MIGRATE_CMA] = 'C',
5992 #ifdef CONFIG_MEMORY_ISOLATION
5993 [MIGRATE_ISOLATE] = 'I',
5996 char tmp[MIGRATE_TYPES + 1];
6000 for (i = 0; i < MIGRATE_TYPES; i++) {
6001 if (type & (1 << i))
6006 printk(KERN_CONT "(%s) ", tmp);
6010 * Show free area list (used inside shift_scroll-lock stuff)
6011 * We also calculate the percentage fragmentation. We do this by counting the
6012 * memory on each free list with the exception of the first item on the list.
6015 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
6018 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
6020 unsigned long free_pcp = 0;
6025 for_each_populated_zone(zone) {
6026 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6029 for_each_online_cpu(cpu)
6030 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6033 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
6034 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
6035 " unevictable:%lu dirty:%lu writeback:%lu\n"
6036 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
6037 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
6038 " kernel_misc_reclaimable:%lu\n"
6039 " free:%lu free_pcp:%lu free_cma:%lu\n",
6040 global_node_page_state(NR_ACTIVE_ANON),
6041 global_node_page_state(NR_INACTIVE_ANON),
6042 global_node_page_state(NR_ISOLATED_ANON),
6043 global_node_page_state(NR_ACTIVE_FILE),
6044 global_node_page_state(NR_INACTIVE_FILE),
6045 global_node_page_state(NR_ISOLATED_FILE),
6046 global_node_page_state(NR_UNEVICTABLE),
6047 global_node_page_state(NR_FILE_DIRTY),
6048 global_node_page_state(NR_WRITEBACK),
6049 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
6050 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
6051 global_node_page_state(NR_FILE_MAPPED),
6052 global_node_page_state(NR_SHMEM),
6053 global_node_page_state(NR_PAGETABLE),
6054 global_zone_page_state(NR_BOUNCE),
6055 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
6056 global_zone_page_state(NR_FREE_PAGES),
6058 global_zone_page_state(NR_FREE_CMA_PAGES));
6060 for_each_online_pgdat(pgdat) {
6061 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
6065 " active_anon:%lukB"
6066 " inactive_anon:%lukB"
6067 " active_file:%lukB"
6068 " inactive_file:%lukB"
6069 " unevictable:%lukB"
6070 " isolated(anon):%lukB"
6071 " isolated(file):%lukB"
6076 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6078 " shmem_pmdmapped: %lukB"
6081 " writeback_tmp:%lukB"
6082 " kernel_stack:%lukB"
6083 #ifdef CONFIG_SHADOW_CALL_STACK
6084 " shadow_call_stack:%lukB"
6087 " all_unreclaimable? %s"
6090 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
6091 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
6092 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
6093 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
6094 K(node_page_state(pgdat, NR_UNEVICTABLE)),
6095 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
6096 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
6097 K(node_page_state(pgdat, NR_FILE_MAPPED)),
6098 K(node_page_state(pgdat, NR_FILE_DIRTY)),
6099 K(node_page_state(pgdat, NR_WRITEBACK)),
6100 K(node_page_state(pgdat, NR_SHMEM)),
6101 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6102 K(node_page_state(pgdat, NR_SHMEM_THPS)),
6103 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6104 K(node_page_state(pgdat, NR_ANON_THPS)),
6106 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6107 node_page_state(pgdat, NR_KERNEL_STACK_KB),
6108 #ifdef CONFIG_SHADOW_CALL_STACK
6109 node_page_state(pgdat, NR_KERNEL_SCS_KB),
6111 K(node_page_state(pgdat, NR_PAGETABLE)),
6112 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6116 for_each_populated_zone(zone) {
6119 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6123 for_each_online_cpu(cpu)
6124 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6134 " reserved_highatomic:%luKB"
6135 " active_anon:%lukB"
6136 " inactive_anon:%lukB"
6137 " active_file:%lukB"
6138 " inactive_file:%lukB"
6139 " unevictable:%lukB"
6140 " writepending:%lukB"
6150 K(zone_page_state(zone, NR_FREE_PAGES)),
6151 K(zone->watermark_boost),
6152 K(min_wmark_pages(zone)),
6153 K(low_wmark_pages(zone)),
6154 K(high_wmark_pages(zone)),
6155 K(zone->nr_reserved_highatomic),
6156 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6157 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6158 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6159 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6160 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6161 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6162 K(zone->present_pages),
6163 K(zone_managed_pages(zone)),
6164 K(zone_page_state(zone, NR_MLOCK)),
6165 K(zone_page_state(zone, NR_BOUNCE)),
6167 K(this_cpu_read(zone->per_cpu_pageset->count)),
6168 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6169 printk("lowmem_reserve[]:");
6170 for (i = 0; i < MAX_NR_ZONES; i++)
6171 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6172 printk(KERN_CONT "\n");
6175 for_each_populated_zone(zone) {
6177 unsigned long nr[MAX_ORDER], flags, total = 0;
6178 unsigned char types[MAX_ORDER];
6180 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6183 printk(KERN_CONT "%s: ", zone->name);
6185 spin_lock_irqsave(&zone->lock, flags);
6186 for (order = 0; order < MAX_ORDER; order++) {
6187 struct free_area *area = &zone->free_area[order];
6190 nr[order] = area->nr_free;
6191 total += nr[order] << order;
6194 for (type = 0; type < MIGRATE_TYPES; type++) {
6195 if (!free_area_empty(area, type))
6196 types[order] |= 1 << type;
6199 spin_unlock_irqrestore(&zone->lock, flags);
6200 for (order = 0; order < MAX_ORDER; order++) {
6201 printk(KERN_CONT "%lu*%lukB ",
6202 nr[order], K(1UL) << order);
6204 show_migration_types(types[order]);
6206 printk(KERN_CONT "= %lukB\n", K(total));
6209 for_each_online_node(nid) {
6210 if (show_mem_node_skip(filter, nid, nodemask))
6212 hugetlb_show_meminfo_node(nid);
6215 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6217 show_swap_cache_info();
6220 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6222 zoneref->zone = zone;
6223 zoneref->zone_idx = zone_idx(zone);
6227 * Builds allocation fallback zone lists.
6229 * Add all populated zones of a node to the zonelist.
6231 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6234 enum zone_type zone_type = MAX_NR_ZONES;
6239 zone = pgdat->node_zones + zone_type;
6240 if (populated_zone(zone)) {
6241 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6242 check_highest_zone(zone_type);
6244 } while (zone_type);
6251 static int __parse_numa_zonelist_order(char *s)
6254 * We used to support different zonelists modes but they turned
6255 * out to be just not useful. Let's keep the warning in place
6256 * if somebody still use the cmd line parameter so that we do
6257 * not fail it silently
6259 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6260 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6266 char numa_zonelist_order[] = "Node";
6269 * sysctl handler for numa_zonelist_order
6271 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6272 void *buffer, size_t *length, loff_t *ppos)
6275 return __parse_numa_zonelist_order(buffer);
6276 return proc_dostring(table, write, buffer, length, ppos);
6280 static int node_load[MAX_NUMNODES];
6283 * find_next_best_node - find the next node that should appear in a given node's fallback list
6284 * @node: node whose fallback list we're appending
6285 * @used_node_mask: nodemask_t of already used nodes
6287 * We use a number of factors to determine which is the next node that should
6288 * appear on a given node's fallback list. The node should not have appeared
6289 * already in @node's fallback list, and it should be the next closest node
6290 * according to the distance array (which contains arbitrary distance values
6291 * from each node to each node in the system), and should also prefer nodes
6292 * with no CPUs, since presumably they'll have very little allocation pressure
6293 * on them otherwise.
6295 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6297 int find_next_best_node(int node, nodemask_t *used_node_mask)
6300 int min_val = INT_MAX;
6301 int best_node = NUMA_NO_NODE;
6303 /* Use the local node if we haven't already */
6304 if (!node_isset(node, *used_node_mask)) {
6305 node_set(node, *used_node_mask);
6309 for_each_node_state(n, N_MEMORY) {
6311 /* Don't want a node to appear more than once */
6312 if (node_isset(n, *used_node_mask))
6315 /* Use the distance array to find the distance */
6316 val = node_distance(node, n);
6318 /* Penalize nodes under us ("prefer the next node") */
6321 /* Give preference to headless and unused nodes */
6322 if (!cpumask_empty(cpumask_of_node(n)))
6323 val += PENALTY_FOR_NODE_WITH_CPUS;
6325 /* Slight preference for less loaded node */
6326 val *= MAX_NUMNODES;
6327 val += node_load[n];
6329 if (val < min_val) {
6336 node_set(best_node, *used_node_mask);
6343 * Build zonelists ordered by node and zones within node.
6344 * This results in maximum locality--normal zone overflows into local
6345 * DMA zone, if any--but risks exhausting DMA zone.
6347 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6350 struct zoneref *zonerefs;
6353 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6355 for (i = 0; i < nr_nodes; i++) {
6358 pg_data_t *node = NODE_DATA(node_order[i]);
6360 nr_zones = build_zonerefs_node(node, zonerefs);
6361 zonerefs += nr_zones;
6363 zonerefs->zone = NULL;
6364 zonerefs->zone_idx = 0;
6368 * Build gfp_thisnode zonelists
6370 static void build_thisnode_zonelists(pg_data_t *pgdat)
6372 struct zoneref *zonerefs;
6375 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6376 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6377 zonerefs += nr_zones;
6378 zonerefs->zone = NULL;
6379 zonerefs->zone_idx = 0;
6383 * Build zonelists ordered by zone and nodes within zones.
6384 * This results in conserving DMA zone[s] until all Normal memory is
6385 * exhausted, but results in overflowing to remote node while memory
6386 * may still exist in local DMA zone.
6389 static void build_zonelists(pg_data_t *pgdat)
6391 static int node_order[MAX_NUMNODES];
6392 int node, nr_nodes = 0;
6393 nodemask_t used_mask = NODE_MASK_NONE;
6394 int local_node, prev_node;
6396 /* NUMA-aware ordering of nodes */
6397 local_node = pgdat->node_id;
6398 prev_node = local_node;
6400 memset(node_order, 0, sizeof(node_order));
6401 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6403 * We don't want to pressure a particular node.
6404 * So adding penalty to the first node in same
6405 * distance group to make it round-robin.
6407 if (node_distance(local_node, node) !=
6408 node_distance(local_node, prev_node))
6409 node_load[node] += 1;
6411 node_order[nr_nodes++] = node;
6415 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6416 build_thisnode_zonelists(pgdat);
6417 pr_info("Fallback order for Node %d: ", local_node);
6418 for (node = 0; node < nr_nodes; node++)
6419 pr_cont("%d ", node_order[node]);
6423 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6425 * Return node id of node used for "local" allocations.
6426 * I.e., first node id of first zone in arg node's generic zonelist.
6427 * Used for initializing percpu 'numa_mem', which is used primarily
6428 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6430 int local_memory_node(int node)
6434 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6435 gfp_zone(GFP_KERNEL),
6437 return zone_to_nid(z->zone);
6441 static void setup_min_unmapped_ratio(void);
6442 static void setup_min_slab_ratio(void);
6443 #else /* CONFIG_NUMA */
6445 static void build_zonelists(pg_data_t *pgdat)
6447 int node, local_node;
6448 struct zoneref *zonerefs;
6451 local_node = pgdat->node_id;
6453 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6454 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6455 zonerefs += nr_zones;
6458 * Now we build the zonelist so that it contains the zones
6459 * of all the other nodes.
6460 * We don't want to pressure a particular node, so when
6461 * building the zones for node N, we make sure that the
6462 * zones coming right after the local ones are those from
6463 * node N+1 (modulo N)
6465 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6466 if (!node_online(node))
6468 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6469 zonerefs += nr_zones;
6471 for (node = 0; node < local_node; node++) {
6472 if (!node_online(node))
6474 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6475 zonerefs += nr_zones;
6478 zonerefs->zone = NULL;
6479 zonerefs->zone_idx = 0;
6482 #endif /* CONFIG_NUMA */
6485 * Boot pageset table. One per cpu which is going to be used for all
6486 * zones and all nodes. The parameters will be set in such a way
6487 * that an item put on a list will immediately be handed over to
6488 * the buddy list. This is safe since pageset manipulation is done
6489 * with interrupts disabled.
6491 * The boot_pagesets must be kept even after bootup is complete for
6492 * unused processors and/or zones. They do play a role for bootstrapping
6493 * hotplugged processors.
6495 * zoneinfo_show() and maybe other functions do
6496 * not check if the processor is online before following the pageset pointer.
6497 * Other parts of the kernel may not check if the zone is available.
6499 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6500 /* These effectively disable the pcplists in the boot pageset completely */
6501 #define BOOT_PAGESET_HIGH 0
6502 #define BOOT_PAGESET_BATCH 1
6503 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6504 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6505 DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6507 static void __build_all_zonelists(void *data)
6510 int __maybe_unused cpu;
6511 pg_data_t *self = data;
6512 static DEFINE_SPINLOCK(lock);
6517 memset(node_load, 0, sizeof(node_load));
6521 * This node is hotadded and no memory is yet present. So just
6522 * building zonelists is fine - no need to touch other nodes.
6524 if (self && !node_online(self->node_id)) {
6525 build_zonelists(self);
6528 * All possible nodes have pgdat preallocated
6531 for_each_node(nid) {
6532 pg_data_t *pgdat = NODE_DATA(nid);
6534 build_zonelists(pgdat);
6537 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6539 * We now know the "local memory node" for each node--
6540 * i.e., the node of the first zone in the generic zonelist.
6541 * Set up numa_mem percpu variable for on-line cpus. During
6542 * boot, only the boot cpu should be on-line; we'll init the
6543 * secondary cpus' numa_mem as they come on-line. During
6544 * node/memory hotplug, we'll fixup all on-line cpus.
6546 for_each_online_cpu(cpu)
6547 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6554 static noinline void __init
6555 build_all_zonelists_init(void)
6559 __build_all_zonelists(NULL);
6562 * Initialize the boot_pagesets that are going to be used
6563 * for bootstrapping processors. The real pagesets for
6564 * each zone will be allocated later when the per cpu
6565 * allocator is available.
6567 * boot_pagesets are used also for bootstrapping offline
6568 * cpus if the system is already booted because the pagesets
6569 * are needed to initialize allocators on a specific cpu too.
6570 * F.e. the percpu allocator needs the page allocator which
6571 * needs the percpu allocator in order to allocate its pagesets
6572 * (a chicken-egg dilemma).
6574 for_each_possible_cpu(cpu)
6575 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6577 mminit_verify_zonelist();
6578 cpuset_init_current_mems_allowed();
6582 * unless system_state == SYSTEM_BOOTING.
6584 * __ref due to call of __init annotated helper build_all_zonelists_init
6585 * [protected by SYSTEM_BOOTING].
6587 void __ref build_all_zonelists(pg_data_t *pgdat)
6589 unsigned long vm_total_pages;
6591 if (system_state == SYSTEM_BOOTING) {
6592 build_all_zonelists_init();
6594 __build_all_zonelists(pgdat);
6595 /* cpuset refresh routine should be here */
6597 /* Get the number of free pages beyond high watermark in all zones. */
6598 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6600 * Disable grouping by mobility if the number of pages in the
6601 * system is too low to allow the mechanism to work. It would be
6602 * more accurate, but expensive to check per-zone. This check is
6603 * made on memory-hotadd so a system can start with mobility
6604 * disabled and enable it later
6606 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6607 page_group_by_mobility_disabled = 1;
6609 page_group_by_mobility_disabled = 0;
6611 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6613 page_group_by_mobility_disabled ? "off" : "on",
6616 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6620 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6621 static bool __meminit
6622 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6624 static struct memblock_region *r;
6626 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6627 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6628 for_each_mem_region(r) {
6629 if (*pfn < memblock_region_memory_end_pfn(r))
6633 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6634 memblock_is_mirror(r)) {
6635 *pfn = memblock_region_memory_end_pfn(r);
6643 * Initially all pages are reserved - free ones are freed
6644 * up by memblock_free_all() once the early boot process is
6645 * done. Non-atomic initialization, single-pass.
6647 * All aligned pageblocks are initialized to the specified migratetype
6648 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6649 * zone stats (e.g., nr_isolate_pageblock) are touched.
6651 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6652 unsigned long start_pfn, unsigned long zone_end_pfn,
6653 enum meminit_context context,
6654 struct vmem_altmap *altmap, int migratetype)
6656 unsigned long pfn, end_pfn = start_pfn + size;
6659 if (highest_memmap_pfn < end_pfn - 1)
6660 highest_memmap_pfn = end_pfn - 1;
6662 #ifdef CONFIG_ZONE_DEVICE
6664 * Honor reservation requested by the driver for this ZONE_DEVICE
6665 * memory. We limit the total number of pages to initialize to just
6666 * those that might contain the memory mapping. We will defer the
6667 * ZONE_DEVICE page initialization until after we have released
6670 if (zone == ZONE_DEVICE) {
6674 if (start_pfn == altmap->base_pfn)
6675 start_pfn += altmap->reserve;
6676 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6680 for (pfn = start_pfn; pfn < end_pfn; ) {
6682 * There can be holes in boot-time mem_map[]s handed to this
6683 * function. They do not exist on hotplugged memory.
6685 if (context == MEMINIT_EARLY) {
6686 if (overlap_memmap_init(zone, &pfn))
6688 if (defer_init(nid, pfn, zone_end_pfn))
6692 page = pfn_to_page(pfn);
6693 __init_single_page(page, pfn, zone, nid);
6694 if (context == MEMINIT_HOTPLUG)
6695 __SetPageReserved(page);
6698 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6699 * such that unmovable allocations won't be scattered all
6700 * over the place during system boot.
6702 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6703 set_pageblock_migratetype(page, migratetype);
6710 #ifdef CONFIG_ZONE_DEVICE
6711 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6712 unsigned long zone_idx, int nid,
6713 struct dev_pagemap *pgmap)
6716 __init_single_page(page, pfn, zone_idx, nid);
6719 * Mark page reserved as it will need to wait for onlining
6720 * phase for it to be fully associated with a zone.
6722 * We can use the non-atomic __set_bit operation for setting
6723 * the flag as we are still initializing the pages.
6725 __SetPageReserved(page);
6728 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6729 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6730 * ever freed or placed on a driver-private list.
6732 page->pgmap = pgmap;
6733 page->zone_device_data = NULL;
6736 * Mark the block movable so that blocks are reserved for
6737 * movable at startup. This will force kernel allocations
6738 * to reserve their blocks rather than leaking throughout
6739 * the address space during boot when many long-lived
6740 * kernel allocations are made.
6742 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6743 * because this is done early in section_activate()
6745 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6746 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6752 * With compound page geometry and when struct pages are stored in ram most
6753 * tail pages are reused. Consequently, the amount of unique struct pages to
6754 * initialize is a lot smaller that the total amount of struct pages being
6755 * mapped. This is a paired / mild layering violation with explicit knowledge
6756 * of how the sparse_vmemmap internals handle compound pages in the lack
6757 * of an altmap. See vmemmap_populate_compound_pages().
6759 static inline unsigned long compound_nr_pages(struct vmem_altmap *altmap,
6760 unsigned long nr_pages)
6762 return is_power_of_2(sizeof(struct page)) &&
6763 !altmap ? 2 * (PAGE_SIZE / sizeof(struct page)) : nr_pages;
6766 static void __ref memmap_init_compound(struct page *head,
6767 unsigned long head_pfn,
6768 unsigned long zone_idx, int nid,
6769 struct dev_pagemap *pgmap,
6770 unsigned long nr_pages)
6772 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6773 unsigned int order = pgmap->vmemmap_shift;
6775 __SetPageHead(head);
6776 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6777 struct page *page = pfn_to_page(pfn);
6779 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6780 prep_compound_tail(head, pfn - head_pfn);
6781 set_page_count(page, 0);
6784 * The first tail page stores compound_mapcount_ptr() and
6785 * compound_order() and the second tail page stores
6786 * compound_pincount_ptr(). Call prep_compound_head() after
6787 * the first and second tail pages have been initialized to
6788 * not have the data overwritten.
6790 if (pfn == head_pfn + 2)
6791 prep_compound_head(head, order);
6795 void __ref memmap_init_zone_device(struct zone *zone,
6796 unsigned long start_pfn,
6797 unsigned long nr_pages,
6798 struct dev_pagemap *pgmap)
6800 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6801 struct pglist_data *pgdat = zone->zone_pgdat;
6802 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6803 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6804 unsigned long zone_idx = zone_idx(zone);
6805 unsigned long start = jiffies;
6806 int nid = pgdat->node_id;
6808 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6812 * The call to memmap_init should have already taken care
6813 * of the pages reserved for the memmap, so we can just jump to
6814 * the end of that region and start processing the device pages.
6817 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6818 nr_pages = end_pfn - start_pfn;
6821 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6822 struct page *page = pfn_to_page(pfn);
6824 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6826 if (pfns_per_compound == 1)
6829 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6830 compound_nr_pages(altmap, pfns_per_compound));
6833 pr_info("%s initialised %lu pages in %ums\n", __func__,
6834 nr_pages, jiffies_to_msecs(jiffies - start));
6838 static void __meminit zone_init_free_lists(struct zone *zone)
6840 unsigned int order, t;
6841 for_each_migratetype_order(order, t) {
6842 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6843 zone->free_area[order].nr_free = 0;
6848 * Only struct pages that correspond to ranges defined by memblock.memory
6849 * are zeroed and initialized by going through __init_single_page() during
6850 * memmap_init_zone_range().
6852 * But, there could be struct pages that correspond to holes in
6853 * memblock.memory. This can happen because of the following reasons:
6854 * - physical memory bank size is not necessarily the exact multiple of the
6855 * arbitrary section size
6856 * - early reserved memory may not be listed in memblock.memory
6857 * - memory layouts defined with memmap= kernel parameter may not align
6858 * nicely with memmap sections
6860 * Explicitly initialize those struct pages so that:
6861 * - PG_Reserved is set
6862 * - zone and node links point to zone and node that span the page if the
6863 * hole is in the middle of a zone
6864 * - zone and node links point to adjacent zone/node if the hole falls on
6865 * the zone boundary; the pages in such holes will be prepended to the
6866 * zone/node above the hole except for the trailing pages in the last
6867 * section that will be appended to the zone/node below.
6869 static void __init init_unavailable_range(unsigned long spfn,
6876 for (pfn = spfn; pfn < epfn; pfn++) {
6877 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6878 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6879 + pageblock_nr_pages - 1;
6882 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6883 __SetPageReserved(pfn_to_page(pfn));
6888 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6889 node, zone_names[zone], pgcnt);
6892 static void __init memmap_init_zone_range(struct zone *zone,
6893 unsigned long start_pfn,
6894 unsigned long end_pfn,
6895 unsigned long *hole_pfn)
6897 unsigned long zone_start_pfn = zone->zone_start_pfn;
6898 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6899 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6901 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6902 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6904 if (start_pfn >= end_pfn)
6907 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6908 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6910 if (*hole_pfn < start_pfn)
6911 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6913 *hole_pfn = end_pfn;
6916 static void __init memmap_init(void)
6918 unsigned long start_pfn, end_pfn;
6919 unsigned long hole_pfn = 0;
6920 int i, j, zone_id = 0, nid;
6922 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6923 struct pglist_data *node = NODE_DATA(nid);
6925 for (j = 0; j < MAX_NR_ZONES; j++) {
6926 struct zone *zone = node->node_zones + j;
6928 if (!populated_zone(zone))
6931 memmap_init_zone_range(zone, start_pfn, end_pfn,
6937 #ifdef CONFIG_SPARSEMEM
6939 * Initialize the memory map for hole in the range [memory_end,
6941 * Append the pages in this hole to the highest zone in the last
6943 * The call to init_unavailable_range() is outside the ifdef to
6944 * silence the compiler warining about zone_id set but not used;
6945 * for FLATMEM it is a nop anyway
6947 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6948 if (hole_pfn < end_pfn)
6950 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6953 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6954 phys_addr_t min_addr, int nid, bool exact_nid)
6959 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6960 MEMBLOCK_ALLOC_ACCESSIBLE,
6963 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6964 MEMBLOCK_ALLOC_ACCESSIBLE,
6967 if (ptr && size > 0)
6968 page_init_poison(ptr, size);
6973 static int zone_batchsize(struct zone *zone)
6979 * The number of pages to batch allocate is either ~0.1%
6980 * of the zone or 1MB, whichever is smaller. The batch
6981 * size is striking a balance between allocation latency
6982 * and zone lock contention.
6984 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6985 batch /= 4; /* We effectively *= 4 below */
6990 * Clamp the batch to a 2^n - 1 value. Having a power
6991 * of 2 value was found to be more likely to have
6992 * suboptimal cache aliasing properties in some cases.
6994 * For example if 2 tasks are alternately allocating
6995 * batches of pages, one task can end up with a lot
6996 * of pages of one half of the possible page colors
6997 * and the other with pages of the other colors.
6999 batch = rounddown_pow_of_two(batch + batch/2) - 1;
7004 /* The deferral and batching of frees should be suppressed under NOMMU
7007 * The problem is that NOMMU needs to be able to allocate large chunks
7008 * of contiguous memory as there's no hardware page translation to
7009 * assemble apparent contiguous memory from discontiguous pages.
7011 * Queueing large contiguous runs of pages for batching, however,
7012 * causes the pages to actually be freed in smaller chunks. As there
7013 * can be a significant delay between the individual batches being
7014 * recycled, this leads to the once large chunks of space being
7015 * fragmented and becoming unavailable for high-order allocations.
7021 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
7026 unsigned long total_pages;
7028 if (!percpu_pagelist_high_fraction) {
7030 * By default, the high value of the pcp is based on the zone
7031 * low watermark so that if they are full then background
7032 * reclaim will not be started prematurely.
7034 total_pages = low_wmark_pages(zone);
7037 * If percpu_pagelist_high_fraction is configured, the high
7038 * value is based on a fraction of the managed pages in the
7041 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
7045 * Split the high value across all online CPUs local to the zone. Note
7046 * that early in boot that CPUs may not be online yet and that during
7047 * CPU hotplug that the cpumask is not yet updated when a CPU is being
7048 * onlined. For memory nodes that have no CPUs, split pcp->high across
7049 * all online CPUs to mitigate the risk that reclaim is triggered
7050 * prematurely due to pages stored on pcp lists.
7052 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
7054 nr_split_cpus = num_online_cpus();
7055 high = total_pages / nr_split_cpus;
7058 * Ensure high is at least batch*4. The multiple is based on the
7059 * historical relationship between high and batch.
7061 high = max(high, batch << 2);
7070 * pcp->high and pcp->batch values are related and generally batch is lower
7071 * than high. They are also related to pcp->count such that count is lower
7072 * than high, and as soon as it reaches high, the pcplist is flushed.
7074 * However, guaranteeing these relations at all times would require e.g. write
7075 * barriers here but also careful usage of read barriers at the read side, and
7076 * thus be prone to error and bad for performance. Thus the update only prevents
7077 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
7078 * can cope with those fields changing asynchronously, and fully trust only the
7079 * pcp->count field on the local CPU with interrupts disabled.
7081 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
7082 * outside of boot time (or some other assurance that no concurrent updaters
7085 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
7086 unsigned long batch)
7088 WRITE_ONCE(pcp->batch, batch);
7089 WRITE_ONCE(pcp->high, high);
7092 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
7096 memset(pcp, 0, sizeof(*pcp));
7097 memset(pzstats, 0, sizeof(*pzstats));
7099 spin_lock_init(&pcp->lock);
7100 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
7101 INIT_LIST_HEAD(&pcp->lists[pindex]);
7104 * Set batch and high values safe for a boot pageset. A true percpu
7105 * pageset's initialization will update them subsequently. Here we don't
7106 * need to be as careful as pageset_update() as nobody can access the
7109 pcp->high = BOOT_PAGESET_HIGH;
7110 pcp->batch = BOOT_PAGESET_BATCH;
7111 pcp->free_factor = 0;
7114 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
7115 unsigned long batch)
7117 struct per_cpu_pages *pcp;
7120 for_each_possible_cpu(cpu) {
7121 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7122 pageset_update(pcp, high, batch);
7127 * Calculate and set new high and batch values for all per-cpu pagesets of a
7128 * zone based on the zone's size.
7130 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
7132 int new_high, new_batch;
7134 new_batch = max(1, zone_batchsize(zone));
7135 new_high = zone_highsize(zone, new_batch, cpu_online);
7137 if (zone->pageset_high == new_high &&
7138 zone->pageset_batch == new_batch)
7141 zone->pageset_high = new_high;
7142 zone->pageset_batch = new_batch;
7144 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7147 void __meminit setup_zone_pageset(struct zone *zone)
7151 /* Size may be 0 on !SMP && !NUMA */
7152 if (sizeof(struct per_cpu_zonestat) > 0)
7153 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7155 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7156 for_each_possible_cpu(cpu) {
7157 struct per_cpu_pages *pcp;
7158 struct per_cpu_zonestat *pzstats;
7160 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7161 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7162 per_cpu_pages_init(pcp, pzstats);
7165 zone_set_pageset_high_and_batch(zone, 0);
7169 * Allocate per cpu pagesets and initialize them.
7170 * Before this call only boot pagesets were available.
7172 void __init setup_per_cpu_pageset(void)
7174 struct pglist_data *pgdat;
7176 int __maybe_unused cpu;
7178 for_each_populated_zone(zone)
7179 setup_zone_pageset(zone);
7183 * Unpopulated zones continue using the boot pagesets.
7184 * The numa stats for these pagesets need to be reset.
7185 * Otherwise, they will end up skewing the stats of
7186 * the nodes these zones are associated with.
7188 for_each_possible_cpu(cpu) {
7189 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7190 memset(pzstats->vm_numa_event, 0,
7191 sizeof(pzstats->vm_numa_event));
7195 for_each_online_pgdat(pgdat)
7196 pgdat->per_cpu_nodestats =
7197 alloc_percpu(struct per_cpu_nodestat);
7200 static __meminit void zone_pcp_init(struct zone *zone)
7203 * per cpu subsystem is not up at this point. The following code
7204 * relies on the ability of the linker to provide the
7205 * offset of a (static) per cpu variable into the per cpu area.
7207 zone->per_cpu_pageset = &boot_pageset;
7208 zone->per_cpu_zonestats = &boot_zonestats;
7209 zone->pageset_high = BOOT_PAGESET_HIGH;
7210 zone->pageset_batch = BOOT_PAGESET_BATCH;
7212 if (populated_zone(zone))
7213 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7214 zone->present_pages, zone_batchsize(zone));
7217 void __meminit init_currently_empty_zone(struct zone *zone,
7218 unsigned long zone_start_pfn,
7221 struct pglist_data *pgdat = zone->zone_pgdat;
7222 int zone_idx = zone_idx(zone) + 1;
7224 if (zone_idx > pgdat->nr_zones)
7225 pgdat->nr_zones = zone_idx;
7227 zone->zone_start_pfn = zone_start_pfn;
7229 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7230 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7232 (unsigned long)zone_idx(zone),
7233 zone_start_pfn, (zone_start_pfn + size));
7235 zone_init_free_lists(zone);
7236 zone->initialized = 1;
7240 * get_pfn_range_for_nid - Return the start and end page frames for a node
7241 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7242 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7243 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7245 * It returns the start and end page frame of a node based on information
7246 * provided by memblock_set_node(). If called for a node
7247 * with no available memory, a warning is printed and the start and end
7250 void __init get_pfn_range_for_nid(unsigned int nid,
7251 unsigned long *start_pfn, unsigned long *end_pfn)
7253 unsigned long this_start_pfn, this_end_pfn;
7259 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7260 *start_pfn = min(*start_pfn, this_start_pfn);
7261 *end_pfn = max(*end_pfn, this_end_pfn);
7264 if (*start_pfn == -1UL)
7269 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7270 * assumption is made that zones within a node are ordered in monotonic
7271 * increasing memory addresses so that the "highest" populated zone is used
7273 static void __init find_usable_zone_for_movable(void)
7276 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7277 if (zone_index == ZONE_MOVABLE)
7280 if (arch_zone_highest_possible_pfn[zone_index] >
7281 arch_zone_lowest_possible_pfn[zone_index])
7285 VM_BUG_ON(zone_index == -1);
7286 movable_zone = zone_index;
7290 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7291 * because it is sized independent of architecture. Unlike the other zones,
7292 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7293 * in each node depending on the size of each node and how evenly kernelcore
7294 * is distributed. This helper function adjusts the zone ranges
7295 * provided by the architecture for a given node by using the end of the
7296 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7297 * zones within a node are in order of monotonic increases memory addresses
7299 static void __init adjust_zone_range_for_zone_movable(int nid,
7300 unsigned long zone_type,
7301 unsigned long node_start_pfn,
7302 unsigned long node_end_pfn,
7303 unsigned long *zone_start_pfn,
7304 unsigned long *zone_end_pfn)
7306 /* Only adjust if ZONE_MOVABLE is on this node */
7307 if (zone_movable_pfn[nid]) {
7308 /* Size ZONE_MOVABLE */
7309 if (zone_type == ZONE_MOVABLE) {
7310 *zone_start_pfn = zone_movable_pfn[nid];
7311 *zone_end_pfn = min(node_end_pfn,
7312 arch_zone_highest_possible_pfn[movable_zone]);
7314 /* Adjust for ZONE_MOVABLE starting within this range */
7315 } else if (!mirrored_kernelcore &&
7316 *zone_start_pfn < zone_movable_pfn[nid] &&
7317 *zone_end_pfn > zone_movable_pfn[nid]) {
7318 *zone_end_pfn = zone_movable_pfn[nid];
7320 /* Check if this whole range is within ZONE_MOVABLE */
7321 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7322 *zone_start_pfn = *zone_end_pfn;
7327 * Return the number of pages a zone spans in a node, including holes
7328 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7330 static unsigned long __init zone_spanned_pages_in_node(int nid,
7331 unsigned long zone_type,
7332 unsigned long node_start_pfn,
7333 unsigned long node_end_pfn,
7334 unsigned long *zone_start_pfn,
7335 unsigned long *zone_end_pfn)
7337 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7338 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7339 /* When hotadd a new node from cpu_up(), the node should be empty */
7340 if (!node_start_pfn && !node_end_pfn)
7343 /* Get the start and end of the zone */
7344 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7345 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7346 adjust_zone_range_for_zone_movable(nid, zone_type,
7347 node_start_pfn, node_end_pfn,
7348 zone_start_pfn, zone_end_pfn);
7350 /* Check that this node has pages within the zone's required range */
7351 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7354 /* Move the zone boundaries inside the node if necessary */
7355 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7356 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7358 /* Return the spanned pages */
7359 return *zone_end_pfn - *zone_start_pfn;
7363 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7364 * then all holes in the requested range will be accounted for.
7366 unsigned long __init __absent_pages_in_range(int nid,
7367 unsigned long range_start_pfn,
7368 unsigned long range_end_pfn)
7370 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7371 unsigned long start_pfn, end_pfn;
7374 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7375 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7376 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7377 nr_absent -= end_pfn - start_pfn;
7383 * absent_pages_in_range - Return number of page frames in holes within a range
7384 * @start_pfn: The start PFN to start searching for holes
7385 * @end_pfn: The end PFN to stop searching for holes
7387 * Return: the number of pages frames in memory holes within a range.
7389 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7390 unsigned long end_pfn)
7392 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7395 /* Return the number of page frames in holes in a zone on a node */
7396 static unsigned long __init zone_absent_pages_in_node(int nid,
7397 unsigned long zone_type,
7398 unsigned long node_start_pfn,
7399 unsigned long node_end_pfn)
7401 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7402 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7403 unsigned long zone_start_pfn, zone_end_pfn;
7404 unsigned long nr_absent;
7406 /* When hotadd a new node from cpu_up(), the node should be empty */
7407 if (!node_start_pfn && !node_end_pfn)
7410 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7411 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7413 adjust_zone_range_for_zone_movable(nid, zone_type,
7414 node_start_pfn, node_end_pfn,
7415 &zone_start_pfn, &zone_end_pfn);
7416 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7419 * ZONE_MOVABLE handling.
7420 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7423 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7424 unsigned long start_pfn, end_pfn;
7425 struct memblock_region *r;
7427 for_each_mem_region(r) {
7428 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7429 zone_start_pfn, zone_end_pfn);
7430 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7431 zone_start_pfn, zone_end_pfn);
7433 if (zone_type == ZONE_MOVABLE &&
7434 memblock_is_mirror(r))
7435 nr_absent += end_pfn - start_pfn;
7437 if (zone_type == ZONE_NORMAL &&
7438 !memblock_is_mirror(r))
7439 nr_absent += end_pfn - start_pfn;
7446 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7447 unsigned long node_start_pfn,
7448 unsigned long node_end_pfn)
7450 unsigned long realtotalpages = 0, totalpages = 0;
7453 for (i = 0; i < MAX_NR_ZONES; i++) {
7454 struct zone *zone = pgdat->node_zones + i;
7455 unsigned long zone_start_pfn, zone_end_pfn;
7456 unsigned long spanned, absent;
7457 unsigned long size, real_size;
7459 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7464 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7469 real_size = size - absent;
7472 zone->zone_start_pfn = zone_start_pfn;
7474 zone->zone_start_pfn = 0;
7475 zone->spanned_pages = size;
7476 zone->present_pages = real_size;
7477 #if defined(CONFIG_MEMORY_HOTPLUG)
7478 zone->present_early_pages = real_size;
7482 realtotalpages += real_size;
7485 pgdat->node_spanned_pages = totalpages;
7486 pgdat->node_present_pages = realtotalpages;
7487 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7490 #ifndef CONFIG_SPARSEMEM
7492 * Calculate the size of the zone->blockflags rounded to an unsigned long
7493 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7494 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7495 * round what is now in bits to nearest long in bits, then return it in
7498 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7500 unsigned long usemapsize;
7502 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7503 usemapsize = roundup(zonesize, pageblock_nr_pages);
7504 usemapsize = usemapsize >> pageblock_order;
7505 usemapsize *= NR_PAGEBLOCK_BITS;
7506 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7508 return usemapsize / 8;
7511 static void __ref setup_usemap(struct zone *zone)
7513 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7514 zone->spanned_pages);
7515 zone->pageblock_flags = NULL;
7517 zone->pageblock_flags =
7518 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7520 if (!zone->pageblock_flags)
7521 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7522 usemapsize, zone->name, zone_to_nid(zone));
7526 static inline void setup_usemap(struct zone *zone) {}
7527 #endif /* CONFIG_SPARSEMEM */
7529 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7531 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7532 void __init set_pageblock_order(void)
7534 unsigned int order = MAX_ORDER - 1;
7536 /* Check that pageblock_nr_pages has not already been setup */
7537 if (pageblock_order)
7540 /* Don't let pageblocks exceed the maximum allocation granularity. */
7541 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7542 order = HUGETLB_PAGE_ORDER;
7545 * Assume the largest contiguous order of interest is a huge page.
7546 * This value may be variable depending on boot parameters on IA64 and
7549 pageblock_order = order;
7551 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7554 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7555 * is unused as pageblock_order is set at compile-time. See
7556 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7559 void __init set_pageblock_order(void)
7563 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7565 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7566 unsigned long present_pages)
7568 unsigned long pages = spanned_pages;
7571 * Provide a more accurate estimation if there are holes within
7572 * the zone and SPARSEMEM is in use. If there are holes within the
7573 * zone, each populated memory region may cost us one or two extra
7574 * memmap pages due to alignment because memmap pages for each
7575 * populated regions may not be naturally aligned on page boundary.
7576 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7578 if (spanned_pages > present_pages + (present_pages >> 4) &&
7579 IS_ENABLED(CONFIG_SPARSEMEM))
7580 pages = present_pages;
7582 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7585 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7586 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7588 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7590 spin_lock_init(&ds_queue->split_queue_lock);
7591 INIT_LIST_HEAD(&ds_queue->split_queue);
7592 ds_queue->split_queue_len = 0;
7595 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7598 #ifdef CONFIG_COMPACTION
7599 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7601 init_waitqueue_head(&pgdat->kcompactd_wait);
7604 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7607 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7611 pgdat_resize_init(pgdat);
7613 pgdat_init_split_queue(pgdat);
7614 pgdat_init_kcompactd(pgdat);
7616 init_waitqueue_head(&pgdat->kswapd_wait);
7617 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7619 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7620 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7622 pgdat_page_ext_init(pgdat);
7623 lruvec_init(&pgdat->__lruvec);
7626 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7627 unsigned long remaining_pages)
7629 atomic_long_set(&zone->managed_pages, remaining_pages);
7630 zone_set_nid(zone, nid);
7631 zone->name = zone_names[idx];
7632 zone->zone_pgdat = NODE_DATA(nid);
7633 spin_lock_init(&zone->lock);
7634 zone_seqlock_init(zone);
7635 zone_pcp_init(zone);
7639 * Set up the zone data structures
7640 * - init pgdat internals
7641 * - init all zones belonging to this node
7643 * NOTE: this function is only called during memory hotplug
7645 #ifdef CONFIG_MEMORY_HOTPLUG
7646 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7648 int nid = pgdat->node_id;
7652 pgdat_init_internals(pgdat);
7654 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7655 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7658 * Reset the nr_zones, order and highest_zoneidx before reuse.
7659 * Note that kswapd will init kswapd_highest_zoneidx properly
7660 * when it starts in the near future.
7662 pgdat->nr_zones = 0;
7663 pgdat->kswapd_order = 0;
7664 pgdat->kswapd_highest_zoneidx = 0;
7665 pgdat->node_start_pfn = 0;
7666 for_each_online_cpu(cpu) {
7667 struct per_cpu_nodestat *p;
7669 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7670 memset(p, 0, sizeof(*p));
7673 for (z = 0; z < MAX_NR_ZONES; z++)
7674 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7679 * Set up the zone data structures:
7680 * - mark all pages reserved
7681 * - mark all memory queues empty
7682 * - clear the memory bitmaps
7684 * NOTE: pgdat should get zeroed by caller.
7685 * NOTE: this function is only called during early init.
7687 static void __init free_area_init_core(struct pglist_data *pgdat)
7690 int nid = pgdat->node_id;
7692 pgdat_init_internals(pgdat);
7693 pgdat->per_cpu_nodestats = &boot_nodestats;
7695 for (j = 0; j < MAX_NR_ZONES; j++) {
7696 struct zone *zone = pgdat->node_zones + j;
7697 unsigned long size, freesize, memmap_pages;
7699 size = zone->spanned_pages;
7700 freesize = zone->present_pages;
7703 * Adjust freesize so that it accounts for how much memory
7704 * is used by this zone for memmap. This affects the watermark
7705 * and per-cpu initialisations
7707 memmap_pages = calc_memmap_size(size, freesize);
7708 if (!is_highmem_idx(j)) {
7709 if (freesize >= memmap_pages) {
7710 freesize -= memmap_pages;
7712 pr_debug(" %s zone: %lu pages used for memmap\n",
7713 zone_names[j], memmap_pages);
7715 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7716 zone_names[j], memmap_pages, freesize);
7719 /* Account for reserved pages */
7720 if (j == 0 && freesize > dma_reserve) {
7721 freesize -= dma_reserve;
7722 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7725 if (!is_highmem_idx(j))
7726 nr_kernel_pages += freesize;
7727 /* Charge for highmem memmap if there are enough kernel pages */
7728 else if (nr_kernel_pages > memmap_pages * 2)
7729 nr_kernel_pages -= memmap_pages;
7730 nr_all_pages += freesize;
7733 * Set an approximate value for lowmem here, it will be adjusted
7734 * when the bootmem allocator frees pages into the buddy system.
7735 * And all highmem pages will be managed by the buddy system.
7737 zone_init_internals(zone, j, nid, freesize);
7742 set_pageblock_order();
7744 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7748 #ifdef CONFIG_FLATMEM
7749 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7751 unsigned long __maybe_unused start = 0;
7752 unsigned long __maybe_unused offset = 0;
7754 /* Skip empty nodes */
7755 if (!pgdat->node_spanned_pages)
7758 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7759 offset = pgdat->node_start_pfn - start;
7760 /* ia64 gets its own node_mem_map, before this, without bootmem */
7761 if (!pgdat->node_mem_map) {
7762 unsigned long size, end;
7766 * The zone's endpoints aren't required to be MAX_ORDER
7767 * aligned but the node_mem_map endpoints must be in order
7768 * for the buddy allocator to function correctly.
7770 end = pgdat_end_pfn(pgdat);
7771 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7772 size = (end - start) * sizeof(struct page);
7773 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7774 pgdat->node_id, false);
7776 panic("Failed to allocate %ld bytes for node %d memory map\n",
7777 size, pgdat->node_id);
7778 pgdat->node_mem_map = map + offset;
7780 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7781 __func__, pgdat->node_id, (unsigned long)pgdat,
7782 (unsigned long)pgdat->node_mem_map);
7785 * With no DISCONTIG, the global mem_map is just set as node 0's
7787 if (pgdat == NODE_DATA(0)) {
7788 mem_map = NODE_DATA(0)->node_mem_map;
7789 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7795 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7796 #endif /* CONFIG_FLATMEM */
7798 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7799 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7801 pgdat->first_deferred_pfn = ULONG_MAX;
7804 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7807 static void __init free_area_init_node(int nid)
7809 pg_data_t *pgdat = NODE_DATA(nid);
7810 unsigned long start_pfn = 0;
7811 unsigned long end_pfn = 0;
7813 /* pg_data_t should be reset to zero when it's allocated */
7814 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7816 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7818 pgdat->node_id = nid;
7819 pgdat->node_start_pfn = start_pfn;
7820 pgdat->per_cpu_nodestats = NULL;
7822 if (start_pfn != end_pfn) {
7823 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7824 (u64)start_pfn << PAGE_SHIFT,
7825 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7827 pr_info("Initmem setup node %d as memoryless\n", nid);
7830 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7832 alloc_node_mem_map(pgdat);
7833 pgdat_set_deferred_range(pgdat);
7835 free_area_init_core(pgdat);
7838 static void __init free_area_init_memoryless_node(int nid)
7840 free_area_init_node(nid);
7843 #if MAX_NUMNODES > 1
7845 * Figure out the number of possible node ids.
7847 void __init setup_nr_node_ids(void)
7849 unsigned int highest;
7851 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7852 nr_node_ids = highest + 1;
7857 * node_map_pfn_alignment - determine the maximum internode alignment
7859 * This function should be called after node map is populated and sorted.
7860 * It calculates the maximum power of two alignment which can distinguish
7863 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7864 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7865 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7866 * shifted, 1GiB is enough and this function will indicate so.
7868 * This is used to test whether pfn -> nid mapping of the chosen memory
7869 * model has fine enough granularity to avoid incorrect mapping for the
7870 * populated node map.
7872 * Return: the determined alignment in pfn's. 0 if there is no alignment
7873 * requirement (single node).
7875 unsigned long __init node_map_pfn_alignment(void)
7877 unsigned long accl_mask = 0, last_end = 0;
7878 unsigned long start, end, mask;
7879 int last_nid = NUMA_NO_NODE;
7882 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7883 if (!start || last_nid < 0 || last_nid == nid) {
7890 * Start with a mask granular enough to pin-point to the
7891 * start pfn and tick off bits one-by-one until it becomes
7892 * too coarse to separate the current node from the last.
7894 mask = ~((1 << __ffs(start)) - 1);
7895 while (mask && last_end <= (start & (mask << 1)))
7898 /* accumulate all internode masks */
7902 /* convert mask to number of pages */
7903 return ~accl_mask + 1;
7907 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7909 * Return: the minimum PFN based on information provided via
7910 * memblock_set_node().
7912 unsigned long __init find_min_pfn_with_active_regions(void)
7914 return PHYS_PFN(memblock_start_of_DRAM());
7918 * early_calculate_totalpages()
7919 * Sum pages in active regions for movable zone.
7920 * Populate N_MEMORY for calculating usable_nodes.
7922 static unsigned long __init early_calculate_totalpages(void)
7924 unsigned long totalpages = 0;
7925 unsigned long start_pfn, end_pfn;
7928 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7929 unsigned long pages = end_pfn - start_pfn;
7931 totalpages += pages;
7933 node_set_state(nid, N_MEMORY);
7939 * Find the PFN the Movable zone begins in each node. Kernel memory
7940 * is spread evenly between nodes as long as the nodes have enough
7941 * memory. When they don't, some nodes will have more kernelcore than
7944 static void __init find_zone_movable_pfns_for_nodes(void)
7947 unsigned long usable_startpfn;
7948 unsigned long kernelcore_node, kernelcore_remaining;
7949 /* save the state before borrow the nodemask */
7950 nodemask_t saved_node_state = node_states[N_MEMORY];
7951 unsigned long totalpages = early_calculate_totalpages();
7952 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7953 struct memblock_region *r;
7955 /* Need to find movable_zone earlier when movable_node is specified. */
7956 find_usable_zone_for_movable();
7959 * If movable_node is specified, ignore kernelcore and movablecore
7962 if (movable_node_is_enabled()) {
7963 for_each_mem_region(r) {
7964 if (!memblock_is_hotpluggable(r))
7967 nid = memblock_get_region_node(r);
7969 usable_startpfn = PFN_DOWN(r->base);
7970 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7971 min(usable_startpfn, zone_movable_pfn[nid]) :
7979 * If kernelcore=mirror is specified, ignore movablecore option
7981 if (mirrored_kernelcore) {
7982 bool mem_below_4gb_not_mirrored = false;
7984 for_each_mem_region(r) {
7985 if (memblock_is_mirror(r))
7988 nid = memblock_get_region_node(r);
7990 usable_startpfn = memblock_region_memory_base_pfn(r);
7992 if (usable_startpfn < PHYS_PFN(SZ_4G)) {
7993 mem_below_4gb_not_mirrored = true;
7997 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7998 min(usable_startpfn, zone_movable_pfn[nid]) :
8002 if (mem_below_4gb_not_mirrored)
8003 pr_warn("This configuration results in unmirrored kernel memory.\n");
8009 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
8010 * amount of necessary memory.
8012 if (required_kernelcore_percent)
8013 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
8015 if (required_movablecore_percent)
8016 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
8020 * If movablecore= was specified, calculate what size of
8021 * kernelcore that corresponds so that memory usable for
8022 * any allocation type is evenly spread. If both kernelcore
8023 * and movablecore are specified, then the value of kernelcore
8024 * will be used for required_kernelcore if it's greater than
8025 * what movablecore would have allowed.
8027 if (required_movablecore) {
8028 unsigned long corepages;
8031 * Round-up so that ZONE_MOVABLE is at least as large as what
8032 * was requested by the user
8034 required_movablecore =
8035 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
8036 required_movablecore = min(totalpages, required_movablecore);
8037 corepages = totalpages - required_movablecore;
8039 required_kernelcore = max(required_kernelcore, corepages);
8043 * If kernelcore was not specified or kernelcore size is larger
8044 * than totalpages, there is no ZONE_MOVABLE.
8046 if (!required_kernelcore || required_kernelcore >= totalpages)
8049 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
8050 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
8053 /* Spread kernelcore memory as evenly as possible throughout nodes */
8054 kernelcore_node = required_kernelcore / usable_nodes;
8055 for_each_node_state(nid, N_MEMORY) {
8056 unsigned long start_pfn, end_pfn;
8059 * Recalculate kernelcore_node if the division per node
8060 * now exceeds what is necessary to satisfy the requested
8061 * amount of memory for the kernel
8063 if (required_kernelcore < kernelcore_node)
8064 kernelcore_node = required_kernelcore / usable_nodes;
8067 * As the map is walked, we track how much memory is usable
8068 * by the kernel using kernelcore_remaining. When it is
8069 * 0, the rest of the node is usable by ZONE_MOVABLE
8071 kernelcore_remaining = kernelcore_node;
8073 /* Go through each range of PFNs within this node */
8074 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
8075 unsigned long size_pages;
8077 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
8078 if (start_pfn >= end_pfn)
8081 /* Account for what is only usable for kernelcore */
8082 if (start_pfn < usable_startpfn) {
8083 unsigned long kernel_pages;
8084 kernel_pages = min(end_pfn, usable_startpfn)
8087 kernelcore_remaining -= min(kernel_pages,
8088 kernelcore_remaining);
8089 required_kernelcore -= min(kernel_pages,
8090 required_kernelcore);
8092 /* Continue if range is now fully accounted */
8093 if (end_pfn <= usable_startpfn) {
8096 * Push zone_movable_pfn to the end so
8097 * that if we have to rebalance
8098 * kernelcore across nodes, we will
8099 * not double account here
8101 zone_movable_pfn[nid] = end_pfn;
8104 start_pfn = usable_startpfn;
8108 * The usable PFN range for ZONE_MOVABLE is from
8109 * start_pfn->end_pfn. Calculate size_pages as the
8110 * number of pages used as kernelcore
8112 size_pages = end_pfn - start_pfn;
8113 if (size_pages > kernelcore_remaining)
8114 size_pages = kernelcore_remaining;
8115 zone_movable_pfn[nid] = start_pfn + size_pages;
8118 * Some kernelcore has been met, update counts and
8119 * break if the kernelcore for this node has been
8122 required_kernelcore -= min(required_kernelcore,
8124 kernelcore_remaining -= size_pages;
8125 if (!kernelcore_remaining)
8131 * If there is still required_kernelcore, we do another pass with one
8132 * less node in the count. This will push zone_movable_pfn[nid] further
8133 * along on the nodes that still have memory until kernelcore is
8137 if (usable_nodes && required_kernelcore > usable_nodes)
8141 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
8142 for (nid = 0; nid < MAX_NUMNODES; nid++) {
8143 unsigned long start_pfn, end_pfn;
8145 zone_movable_pfn[nid] =
8146 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
8148 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
8149 if (zone_movable_pfn[nid] >= end_pfn)
8150 zone_movable_pfn[nid] = 0;
8154 /* restore the node_state */
8155 node_states[N_MEMORY] = saved_node_state;
8158 /* Any regular or high memory on that node ? */
8159 static void check_for_memory(pg_data_t *pgdat, int nid)
8161 enum zone_type zone_type;
8163 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8164 struct zone *zone = &pgdat->node_zones[zone_type];
8165 if (populated_zone(zone)) {
8166 if (IS_ENABLED(CONFIG_HIGHMEM))
8167 node_set_state(nid, N_HIGH_MEMORY);
8168 if (zone_type <= ZONE_NORMAL)
8169 node_set_state(nid, N_NORMAL_MEMORY);
8176 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8177 * such cases we allow max_zone_pfn sorted in the descending order
8179 bool __weak arch_has_descending_max_zone_pfns(void)
8185 * free_area_init - Initialise all pg_data_t and zone data
8186 * @max_zone_pfn: an array of max PFNs for each zone
8188 * This will call free_area_init_node() for each active node in the system.
8189 * Using the page ranges provided by memblock_set_node(), the size of each
8190 * zone in each node and their holes is calculated. If the maximum PFN
8191 * between two adjacent zones match, it is assumed that the zone is empty.
8192 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8193 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8194 * starts where the previous one ended. For example, ZONE_DMA32 starts
8195 * at arch_max_dma_pfn.
8197 void __init free_area_init(unsigned long *max_zone_pfn)
8199 unsigned long start_pfn, end_pfn;
8203 /* Record where the zone boundaries are */
8204 memset(arch_zone_lowest_possible_pfn, 0,
8205 sizeof(arch_zone_lowest_possible_pfn));
8206 memset(arch_zone_highest_possible_pfn, 0,
8207 sizeof(arch_zone_highest_possible_pfn));
8209 start_pfn = find_min_pfn_with_active_regions();
8210 descending = arch_has_descending_max_zone_pfns();
8212 for (i = 0; i < MAX_NR_ZONES; i++) {
8214 zone = MAX_NR_ZONES - i - 1;
8218 if (zone == ZONE_MOVABLE)
8221 end_pfn = max(max_zone_pfn[zone], start_pfn);
8222 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8223 arch_zone_highest_possible_pfn[zone] = end_pfn;
8225 start_pfn = end_pfn;
8228 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8229 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8230 find_zone_movable_pfns_for_nodes();
8232 /* Print out the zone ranges */
8233 pr_info("Zone ranges:\n");
8234 for (i = 0; i < MAX_NR_ZONES; i++) {
8235 if (i == ZONE_MOVABLE)
8237 pr_info(" %-8s ", zone_names[i]);
8238 if (arch_zone_lowest_possible_pfn[i] ==
8239 arch_zone_highest_possible_pfn[i])
8242 pr_cont("[mem %#018Lx-%#018Lx]\n",
8243 (u64)arch_zone_lowest_possible_pfn[i]
8245 ((u64)arch_zone_highest_possible_pfn[i]
8246 << PAGE_SHIFT) - 1);
8249 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8250 pr_info("Movable zone start for each node\n");
8251 for (i = 0; i < MAX_NUMNODES; i++) {
8252 if (zone_movable_pfn[i])
8253 pr_info(" Node %d: %#018Lx\n", i,
8254 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8258 * Print out the early node map, and initialize the
8259 * subsection-map relative to active online memory ranges to
8260 * enable future "sub-section" extensions of the memory map.
8262 pr_info("Early memory node ranges\n");
8263 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8264 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8265 (u64)start_pfn << PAGE_SHIFT,
8266 ((u64)end_pfn << PAGE_SHIFT) - 1);
8267 subsection_map_init(start_pfn, end_pfn - start_pfn);
8270 /* Initialise every node */
8271 mminit_verify_pageflags_layout();
8272 setup_nr_node_ids();
8273 for_each_node(nid) {
8276 if (!node_online(nid)) {
8277 pr_info("Initializing node %d as memoryless\n", nid);
8279 /* Allocator not initialized yet */
8280 pgdat = arch_alloc_nodedata(nid);
8282 pr_err("Cannot allocate %zuB for node %d.\n",
8283 sizeof(*pgdat), nid);
8286 arch_refresh_nodedata(nid, pgdat);
8287 free_area_init_memoryless_node(nid);
8290 * We do not want to confuse userspace by sysfs
8291 * files/directories for node without any memory
8292 * attached to it, so this node is not marked as
8293 * N_MEMORY and not marked online so that no sysfs
8294 * hierarchy will be created via register_one_node for
8295 * it. The pgdat will get fully initialized by
8296 * hotadd_init_pgdat() when memory is hotplugged into
8302 pgdat = NODE_DATA(nid);
8303 free_area_init_node(nid);
8305 /* Any memory on that node */
8306 if (pgdat->node_present_pages)
8307 node_set_state(nid, N_MEMORY);
8308 check_for_memory(pgdat, nid);
8314 static int __init cmdline_parse_core(char *p, unsigned long *core,
8315 unsigned long *percent)
8317 unsigned long long coremem;
8323 /* Value may be a percentage of total memory, otherwise bytes */
8324 coremem = simple_strtoull(p, &endptr, 0);
8325 if (*endptr == '%') {
8326 /* Paranoid check for percent values greater than 100 */
8327 WARN_ON(coremem > 100);
8331 coremem = memparse(p, &p);
8332 /* Paranoid check that UL is enough for the coremem value */
8333 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8335 *core = coremem >> PAGE_SHIFT;
8342 * kernelcore=size sets the amount of memory for use for allocations that
8343 * cannot be reclaimed or migrated.
8345 static int __init cmdline_parse_kernelcore(char *p)
8347 /* parse kernelcore=mirror */
8348 if (parse_option_str(p, "mirror")) {
8349 mirrored_kernelcore = true;
8353 return cmdline_parse_core(p, &required_kernelcore,
8354 &required_kernelcore_percent);
8358 * movablecore=size sets the amount of memory for use for allocations that
8359 * can be reclaimed or migrated.
8361 static int __init cmdline_parse_movablecore(char *p)
8363 return cmdline_parse_core(p, &required_movablecore,
8364 &required_movablecore_percent);
8367 early_param("kernelcore", cmdline_parse_kernelcore);
8368 early_param("movablecore", cmdline_parse_movablecore);
8370 void adjust_managed_page_count(struct page *page, long count)
8372 atomic_long_add(count, &page_zone(page)->managed_pages);
8373 totalram_pages_add(count);
8374 #ifdef CONFIG_HIGHMEM
8375 if (PageHighMem(page))
8376 totalhigh_pages_add(count);
8379 EXPORT_SYMBOL(adjust_managed_page_count);
8381 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8384 unsigned long pages = 0;
8386 start = (void *)PAGE_ALIGN((unsigned long)start);
8387 end = (void *)((unsigned long)end & PAGE_MASK);
8388 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8389 struct page *page = virt_to_page(pos);
8390 void *direct_map_addr;
8393 * 'direct_map_addr' might be different from 'pos'
8394 * because some architectures' virt_to_page()
8395 * work with aliases. Getting the direct map
8396 * address ensures that we get a _writeable_
8397 * alias for the memset().
8399 direct_map_addr = page_address(page);
8401 * Perform a kasan-unchecked memset() since this memory
8402 * has not been initialized.
8404 direct_map_addr = kasan_reset_tag(direct_map_addr);
8405 if ((unsigned int)poison <= 0xFF)
8406 memset(direct_map_addr, poison, PAGE_SIZE);
8408 free_reserved_page(page);
8412 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8417 void __init mem_init_print_info(void)
8419 unsigned long physpages, codesize, datasize, rosize, bss_size;
8420 unsigned long init_code_size, init_data_size;
8422 physpages = get_num_physpages();
8423 codesize = _etext - _stext;
8424 datasize = _edata - _sdata;
8425 rosize = __end_rodata - __start_rodata;
8426 bss_size = __bss_stop - __bss_start;
8427 init_data_size = __init_end - __init_begin;
8428 init_code_size = _einittext - _sinittext;
8431 * Detect special cases and adjust section sizes accordingly:
8432 * 1) .init.* may be embedded into .data sections
8433 * 2) .init.text.* may be out of [__init_begin, __init_end],
8434 * please refer to arch/tile/kernel/vmlinux.lds.S.
8435 * 3) .rodata.* may be embedded into .text or .data sections.
8437 #define adj_init_size(start, end, size, pos, adj) \
8439 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8443 adj_init_size(__init_begin, __init_end, init_data_size,
8444 _sinittext, init_code_size);
8445 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8446 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8447 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8448 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8450 #undef adj_init_size
8452 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8453 #ifdef CONFIG_HIGHMEM
8457 K(nr_free_pages()), K(physpages),
8458 codesize >> 10, datasize >> 10, rosize >> 10,
8459 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8460 K(physpages - totalram_pages() - totalcma_pages),
8462 #ifdef CONFIG_HIGHMEM
8463 , K(totalhigh_pages())
8469 * set_dma_reserve - set the specified number of pages reserved in the first zone
8470 * @new_dma_reserve: The number of pages to mark reserved
8472 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8473 * In the DMA zone, a significant percentage may be consumed by kernel image
8474 * and other unfreeable allocations which can skew the watermarks badly. This
8475 * function may optionally be used to account for unfreeable pages in the
8476 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8477 * smaller per-cpu batchsize.
8479 void __init set_dma_reserve(unsigned long new_dma_reserve)
8481 dma_reserve = new_dma_reserve;
8484 static int page_alloc_cpu_dead(unsigned int cpu)
8488 lru_add_drain_cpu(cpu);
8489 mlock_page_drain_remote(cpu);
8493 * Spill the event counters of the dead processor
8494 * into the current processors event counters.
8495 * This artificially elevates the count of the current
8498 vm_events_fold_cpu(cpu);
8501 * Zero the differential counters of the dead processor
8502 * so that the vm statistics are consistent.
8504 * This is only okay since the processor is dead and cannot
8505 * race with what we are doing.
8507 cpu_vm_stats_fold(cpu);
8509 for_each_populated_zone(zone)
8510 zone_pcp_update(zone, 0);
8515 static int page_alloc_cpu_online(unsigned int cpu)
8519 for_each_populated_zone(zone)
8520 zone_pcp_update(zone, 1);
8525 int hashdist = HASHDIST_DEFAULT;
8527 static int __init set_hashdist(char *str)
8531 hashdist = simple_strtoul(str, &str, 0);
8534 __setup("hashdist=", set_hashdist);
8537 void __init page_alloc_init(void)
8542 if (num_node_state(N_MEMORY) == 1)
8546 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8547 "mm/page_alloc:pcp",
8548 page_alloc_cpu_online,
8549 page_alloc_cpu_dead);
8554 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8555 * or min_free_kbytes changes.
8557 static void calculate_totalreserve_pages(void)
8559 struct pglist_data *pgdat;
8560 unsigned long reserve_pages = 0;
8561 enum zone_type i, j;
8563 for_each_online_pgdat(pgdat) {
8565 pgdat->totalreserve_pages = 0;
8567 for (i = 0; i < MAX_NR_ZONES; i++) {
8568 struct zone *zone = pgdat->node_zones + i;
8570 unsigned long managed_pages = zone_managed_pages(zone);
8572 /* Find valid and maximum lowmem_reserve in the zone */
8573 for (j = i; j < MAX_NR_ZONES; j++) {
8574 if (zone->lowmem_reserve[j] > max)
8575 max = zone->lowmem_reserve[j];
8578 /* we treat the high watermark as reserved pages. */
8579 max += high_wmark_pages(zone);
8581 if (max > managed_pages)
8582 max = managed_pages;
8584 pgdat->totalreserve_pages += max;
8586 reserve_pages += max;
8589 totalreserve_pages = reserve_pages;
8593 * setup_per_zone_lowmem_reserve - called whenever
8594 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8595 * has a correct pages reserved value, so an adequate number of
8596 * pages are left in the zone after a successful __alloc_pages().
8598 static void setup_per_zone_lowmem_reserve(void)
8600 struct pglist_data *pgdat;
8601 enum zone_type i, j;
8603 for_each_online_pgdat(pgdat) {
8604 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8605 struct zone *zone = &pgdat->node_zones[i];
8606 int ratio = sysctl_lowmem_reserve_ratio[i];
8607 bool clear = !ratio || !zone_managed_pages(zone);
8608 unsigned long managed_pages = 0;
8610 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8611 struct zone *upper_zone = &pgdat->node_zones[j];
8613 managed_pages += zone_managed_pages(upper_zone);
8616 zone->lowmem_reserve[j] = 0;
8618 zone->lowmem_reserve[j] = managed_pages / ratio;
8623 /* update totalreserve_pages */
8624 calculate_totalreserve_pages();
8627 static void __setup_per_zone_wmarks(void)
8629 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8630 unsigned long lowmem_pages = 0;
8632 unsigned long flags;
8634 /* Calculate total number of !ZONE_HIGHMEM pages */
8635 for_each_zone(zone) {
8636 if (!is_highmem(zone))
8637 lowmem_pages += zone_managed_pages(zone);
8640 for_each_zone(zone) {
8643 spin_lock_irqsave(&zone->lock, flags);
8644 tmp = (u64)pages_min * zone_managed_pages(zone);
8645 do_div(tmp, lowmem_pages);
8646 if (is_highmem(zone)) {
8648 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8649 * need highmem pages, so cap pages_min to a small
8652 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8653 * deltas control async page reclaim, and so should
8654 * not be capped for highmem.
8656 unsigned long min_pages;
8658 min_pages = zone_managed_pages(zone) / 1024;
8659 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8660 zone->_watermark[WMARK_MIN] = min_pages;
8663 * If it's a lowmem zone, reserve a number of pages
8664 * proportionate to the zone's size.
8666 zone->_watermark[WMARK_MIN] = tmp;
8670 * Set the kswapd watermarks distance according to the
8671 * scale factor in proportion to available memory, but
8672 * ensure a minimum size on small systems.
8674 tmp = max_t(u64, tmp >> 2,
8675 mult_frac(zone_managed_pages(zone),
8676 watermark_scale_factor, 10000));
8678 zone->watermark_boost = 0;
8679 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8680 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8681 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8683 spin_unlock_irqrestore(&zone->lock, flags);
8686 /* update totalreserve_pages */
8687 calculate_totalreserve_pages();
8691 * setup_per_zone_wmarks - called when min_free_kbytes changes
8692 * or when memory is hot-{added|removed}
8694 * Ensures that the watermark[min,low,high] values for each zone are set
8695 * correctly with respect to min_free_kbytes.
8697 void setup_per_zone_wmarks(void)
8700 static DEFINE_SPINLOCK(lock);
8703 __setup_per_zone_wmarks();
8707 * The watermark size have changed so update the pcpu batch
8708 * and high limits or the limits may be inappropriate.
8711 zone_pcp_update(zone, 0);
8715 * Initialise min_free_kbytes.
8717 * For small machines we want it small (128k min). For large machines
8718 * we want it large (256MB max). But it is not linear, because network
8719 * bandwidth does not increase linearly with machine size. We use
8721 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8722 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8738 void calculate_min_free_kbytes(void)
8740 unsigned long lowmem_kbytes;
8741 int new_min_free_kbytes;
8743 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8744 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8746 if (new_min_free_kbytes > user_min_free_kbytes)
8747 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8749 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8750 new_min_free_kbytes, user_min_free_kbytes);
8754 int __meminit init_per_zone_wmark_min(void)
8756 calculate_min_free_kbytes();
8757 setup_per_zone_wmarks();
8758 refresh_zone_stat_thresholds();
8759 setup_per_zone_lowmem_reserve();
8762 setup_min_unmapped_ratio();
8763 setup_min_slab_ratio();
8766 khugepaged_min_free_kbytes_update();
8770 postcore_initcall(init_per_zone_wmark_min)
8773 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8774 * that we can call two helper functions whenever min_free_kbytes
8777 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8778 void *buffer, size_t *length, loff_t *ppos)
8782 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8787 user_min_free_kbytes = min_free_kbytes;
8788 setup_per_zone_wmarks();
8793 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8794 void *buffer, size_t *length, loff_t *ppos)
8798 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8803 setup_per_zone_wmarks();
8809 static void setup_min_unmapped_ratio(void)
8814 for_each_online_pgdat(pgdat)
8815 pgdat->min_unmapped_pages = 0;
8818 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8819 sysctl_min_unmapped_ratio) / 100;
8823 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8824 void *buffer, size_t *length, loff_t *ppos)
8828 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8832 setup_min_unmapped_ratio();
8837 static void setup_min_slab_ratio(void)
8842 for_each_online_pgdat(pgdat)
8843 pgdat->min_slab_pages = 0;
8846 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8847 sysctl_min_slab_ratio) / 100;
8850 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8851 void *buffer, size_t *length, loff_t *ppos)
8855 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8859 setup_min_slab_ratio();
8866 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8867 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8868 * whenever sysctl_lowmem_reserve_ratio changes.
8870 * The reserve ratio obviously has absolutely no relation with the
8871 * minimum watermarks. The lowmem reserve ratio can only make sense
8872 * if in function of the boot time zone sizes.
8874 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8875 void *buffer, size_t *length, loff_t *ppos)
8879 proc_dointvec_minmax(table, write, buffer, length, ppos);
8881 for (i = 0; i < MAX_NR_ZONES; i++) {
8882 if (sysctl_lowmem_reserve_ratio[i] < 1)
8883 sysctl_lowmem_reserve_ratio[i] = 0;
8886 setup_per_zone_lowmem_reserve();
8891 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8892 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8893 * pagelist can have before it gets flushed back to buddy allocator.
8895 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8896 int write, void *buffer, size_t *length, loff_t *ppos)
8899 int old_percpu_pagelist_high_fraction;
8902 mutex_lock(&pcp_batch_high_lock);
8903 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8905 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8906 if (!write || ret < 0)
8909 /* Sanity checking to avoid pcp imbalance */
8910 if (percpu_pagelist_high_fraction &&
8911 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8912 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8918 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8921 for_each_populated_zone(zone)
8922 zone_set_pageset_high_and_batch(zone, 0);
8924 mutex_unlock(&pcp_batch_high_lock);
8928 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8930 * Returns the number of pages that arch has reserved but
8931 * is not known to alloc_large_system_hash().
8933 static unsigned long __init arch_reserved_kernel_pages(void)
8940 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8941 * machines. As memory size is increased the scale is also increased but at
8942 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8943 * quadruples the scale is increased by one, which means the size of hash table
8944 * only doubles, instead of quadrupling as well.
8945 * Because 32-bit systems cannot have large physical memory, where this scaling
8946 * makes sense, it is disabled on such platforms.
8948 #if __BITS_PER_LONG > 32
8949 #define ADAPT_SCALE_BASE (64ul << 30)
8950 #define ADAPT_SCALE_SHIFT 2
8951 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8955 * allocate a large system hash table from bootmem
8956 * - it is assumed that the hash table must contain an exact power-of-2
8957 * quantity of entries
8958 * - limit is the number of hash buckets, not the total allocation size
8960 void *__init alloc_large_system_hash(const char *tablename,
8961 unsigned long bucketsize,
8962 unsigned long numentries,
8965 unsigned int *_hash_shift,
8966 unsigned int *_hash_mask,
8967 unsigned long low_limit,
8968 unsigned long high_limit)
8970 unsigned long long max = high_limit;
8971 unsigned long log2qty, size;
8977 /* allow the kernel cmdline to have a say */
8979 /* round applicable memory size up to nearest megabyte */
8980 numentries = nr_kernel_pages;
8981 numentries -= arch_reserved_kernel_pages();
8983 /* It isn't necessary when PAGE_SIZE >= 1MB */
8984 if (PAGE_SHIFT < 20)
8985 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8987 #if __BITS_PER_LONG > 32
8989 unsigned long adapt;
8991 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8992 adapt <<= ADAPT_SCALE_SHIFT)
8997 /* limit to 1 bucket per 2^scale bytes of low memory */
8998 if (scale > PAGE_SHIFT)
8999 numentries >>= (scale - PAGE_SHIFT);
9001 numentries <<= (PAGE_SHIFT - scale);
9003 /* Make sure we've got at least a 0-order allocation.. */
9004 if (unlikely(flags & HASH_SMALL)) {
9005 /* Makes no sense without HASH_EARLY */
9006 WARN_ON(!(flags & HASH_EARLY));
9007 if (!(numentries >> *_hash_shift)) {
9008 numentries = 1UL << *_hash_shift;
9009 BUG_ON(!numentries);
9011 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
9012 numentries = PAGE_SIZE / bucketsize;
9014 numentries = roundup_pow_of_two(numentries);
9016 /* limit allocation size to 1/16 total memory by default */
9018 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
9019 do_div(max, bucketsize);
9021 max = min(max, 0x80000000ULL);
9023 if (numentries < low_limit)
9024 numentries = low_limit;
9025 if (numentries > max)
9028 log2qty = ilog2(numentries);
9030 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
9033 size = bucketsize << log2qty;
9034 if (flags & HASH_EARLY) {
9035 if (flags & HASH_ZERO)
9036 table = memblock_alloc(size, SMP_CACHE_BYTES);
9038 table = memblock_alloc_raw(size,
9040 } else if (get_order(size) >= MAX_ORDER || hashdist) {
9041 table = vmalloc_huge(size, gfp_flags);
9044 huge = is_vm_area_hugepages(table);
9047 * If bucketsize is not a power-of-two, we may free
9048 * some pages at the end of hash table which
9049 * alloc_pages_exact() automatically does
9051 table = alloc_pages_exact(size, gfp_flags);
9052 kmemleak_alloc(table, size, 1, gfp_flags);
9054 } while (!table && size > PAGE_SIZE && --log2qty);
9057 panic("Failed to allocate %s hash table\n", tablename);
9059 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
9060 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
9061 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
9064 *_hash_shift = log2qty;
9066 *_hash_mask = (1 << log2qty) - 1;
9071 #ifdef CONFIG_CONTIG_ALLOC
9072 #if defined(CONFIG_DYNAMIC_DEBUG) || \
9073 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
9074 /* Usage: See admin-guide/dynamic-debug-howto.rst */
9075 static void alloc_contig_dump_pages(struct list_head *page_list)
9077 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9079 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9083 list_for_each_entry(page, page_list, lru)
9084 dump_page(page, "migration failure");
9088 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9093 /* [start, end) must belong to a single zone. */
9094 int __alloc_contig_migrate_range(struct compact_control *cc,
9095 unsigned long start, unsigned long end)
9097 /* This function is based on compact_zone() from compaction.c. */
9098 unsigned int nr_reclaimed;
9099 unsigned long pfn = start;
9100 unsigned int tries = 0;
9102 struct migration_target_control mtc = {
9103 .nid = zone_to_nid(cc->zone),
9104 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9107 lru_cache_disable();
9109 while (pfn < end || !list_empty(&cc->migratepages)) {
9110 if (fatal_signal_pending(current)) {
9115 if (list_empty(&cc->migratepages)) {
9116 cc->nr_migratepages = 0;
9117 ret = isolate_migratepages_range(cc, pfn, end);
9118 if (ret && ret != -EAGAIN)
9120 pfn = cc->migrate_pfn;
9122 } else if (++tries == 5) {
9127 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9129 cc->nr_migratepages -= nr_reclaimed;
9131 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9132 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9135 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9136 * to retry again over this error, so do the same here.
9144 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
9145 alloc_contig_dump_pages(&cc->migratepages);
9146 putback_movable_pages(&cc->migratepages);
9153 * alloc_contig_range() -- tries to allocate given range of pages
9154 * @start: start PFN to allocate
9155 * @end: one-past-the-last PFN to allocate
9156 * @migratetype: migratetype of the underlying pageblocks (either
9157 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9158 * in range must have the same migratetype and it must
9159 * be either of the two.
9160 * @gfp_mask: GFP mask to use during compaction
9162 * The PFN range does not have to be pageblock aligned. The PFN range must
9163 * belong to a single zone.
9165 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9166 * pageblocks in the range. Once isolated, the pageblocks should not
9167 * be modified by others.
9169 * Return: zero on success or negative error code. On success all
9170 * pages which PFN is in [start, end) are allocated for the caller and
9171 * need to be freed with free_contig_range().
9173 int alloc_contig_range(unsigned long start, unsigned long end,
9174 unsigned migratetype, gfp_t gfp_mask)
9176 unsigned long outer_start, outer_end;
9180 struct compact_control cc = {
9181 .nr_migratepages = 0,
9183 .zone = page_zone(pfn_to_page(start)),
9184 .mode = MIGRATE_SYNC,
9185 .ignore_skip_hint = true,
9186 .no_set_skip_hint = true,
9187 .gfp_mask = current_gfp_context(gfp_mask),
9188 .alloc_contig = true,
9190 INIT_LIST_HEAD(&cc.migratepages);
9193 * What we do here is we mark all pageblocks in range as
9194 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9195 * have different sizes, and due to the way page allocator
9196 * work, start_isolate_page_range() has special handlings for this.
9198 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9199 * migrate the pages from an unaligned range (ie. pages that
9200 * we are interested in). This will put all the pages in
9201 * range back to page allocator as MIGRATE_ISOLATE.
9203 * When this is done, we take the pages in range from page
9204 * allocator removing them from the buddy system. This way
9205 * page allocator will never consider using them.
9207 * This lets us mark the pageblocks back as
9208 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9209 * aligned range but not in the unaligned, original range are
9210 * put back to page allocator so that buddy can use them.
9213 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
9217 drain_all_pages(cc.zone);
9220 * In case of -EBUSY, we'd like to know which page causes problem.
9221 * So, just fall through. test_pages_isolated() has a tracepoint
9222 * which will report the busy page.
9224 * It is possible that busy pages could become available before
9225 * the call to test_pages_isolated, and the range will actually be
9226 * allocated. So, if we fall through be sure to clear ret so that
9227 * -EBUSY is not accidentally used or returned to caller.
9229 ret = __alloc_contig_migrate_range(&cc, start, end);
9230 if (ret && ret != -EBUSY)
9235 * Pages from [start, end) are within a pageblock_nr_pages
9236 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9237 * more, all pages in [start, end) are free in page allocator.
9238 * What we are going to do is to allocate all pages from
9239 * [start, end) (that is remove them from page allocator).
9241 * The only problem is that pages at the beginning and at the
9242 * end of interesting range may be not aligned with pages that
9243 * page allocator holds, ie. they can be part of higher order
9244 * pages. Because of this, we reserve the bigger range and
9245 * once this is done free the pages we are not interested in.
9247 * We don't have to hold zone->lock here because the pages are
9248 * isolated thus they won't get removed from buddy.
9252 outer_start = start;
9253 while (!PageBuddy(pfn_to_page(outer_start))) {
9254 if (++order >= MAX_ORDER) {
9255 outer_start = start;
9258 outer_start &= ~0UL << order;
9261 if (outer_start != start) {
9262 order = buddy_order(pfn_to_page(outer_start));
9265 * outer_start page could be small order buddy page and
9266 * it doesn't include start page. Adjust outer_start
9267 * in this case to report failed page properly
9268 * on tracepoint in test_pages_isolated()
9270 if (outer_start + (1UL << order) <= start)
9271 outer_start = start;
9274 /* Make sure the range is really isolated. */
9275 if (test_pages_isolated(outer_start, end, 0)) {
9280 /* Grab isolated pages from freelists. */
9281 outer_end = isolate_freepages_range(&cc, outer_start, end);
9287 /* Free head and tail (if any) */
9288 if (start != outer_start)
9289 free_contig_range(outer_start, start - outer_start);
9290 if (end != outer_end)
9291 free_contig_range(end, outer_end - end);
9294 undo_isolate_page_range(start, end, migratetype);
9297 EXPORT_SYMBOL(alloc_contig_range);
9299 static int __alloc_contig_pages(unsigned long start_pfn,
9300 unsigned long nr_pages, gfp_t gfp_mask)
9302 unsigned long end_pfn = start_pfn + nr_pages;
9304 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9308 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9309 unsigned long nr_pages)
9311 unsigned long i, end_pfn = start_pfn + nr_pages;
9314 for (i = start_pfn; i < end_pfn; i++) {
9315 page = pfn_to_online_page(i);
9319 if (page_zone(page) != z)
9322 if (PageReserved(page))
9328 static bool zone_spans_last_pfn(const struct zone *zone,
9329 unsigned long start_pfn, unsigned long nr_pages)
9331 unsigned long last_pfn = start_pfn + nr_pages - 1;
9333 return zone_spans_pfn(zone, last_pfn);
9337 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9338 * @nr_pages: Number of contiguous pages to allocate
9339 * @gfp_mask: GFP mask to limit search and used during compaction
9341 * @nodemask: Mask for other possible nodes
9343 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9344 * on an applicable zonelist to find a contiguous pfn range which can then be
9345 * tried for allocation with alloc_contig_range(). This routine is intended
9346 * for allocation requests which can not be fulfilled with the buddy allocator.
9348 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9349 * power of two, then allocated range is also guaranteed to be aligned to same
9350 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9352 * Allocated pages can be freed with free_contig_range() or by manually calling
9353 * __free_page() on each allocated page.
9355 * Return: pointer to contiguous pages on success, or NULL if not successful.
9357 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9358 int nid, nodemask_t *nodemask)
9360 unsigned long ret, pfn, flags;
9361 struct zonelist *zonelist;
9365 zonelist = node_zonelist(nid, gfp_mask);
9366 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9367 gfp_zone(gfp_mask), nodemask) {
9368 spin_lock_irqsave(&zone->lock, flags);
9370 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9371 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9372 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9374 * We release the zone lock here because
9375 * alloc_contig_range() will also lock the zone
9376 * at some point. If there's an allocation
9377 * spinning on this lock, it may win the race
9378 * and cause alloc_contig_range() to fail...
9380 spin_unlock_irqrestore(&zone->lock, flags);
9381 ret = __alloc_contig_pages(pfn, nr_pages,
9384 return pfn_to_page(pfn);
9385 spin_lock_irqsave(&zone->lock, flags);
9389 spin_unlock_irqrestore(&zone->lock, flags);
9393 #endif /* CONFIG_CONTIG_ALLOC */
9395 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9397 unsigned long count = 0;
9399 for (; nr_pages--; pfn++) {
9400 struct page *page = pfn_to_page(pfn);
9402 count += page_count(page) != 1;
9405 WARN(count != 0, "%lu pages are still in use!\n", count);
9407 EXPORT_SYMBOL(free_contig_range);
9410 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9411 * page high values need to be recalculated.
9413 void zone_pcp_update(struct zone *zone, int cpu_online)
9415 mutex_lock(&pcp_batch_high_lock);
9416 zone_set_pageset_high_and_batch(zone, cpu_online);
9417 mutex_unlock(&pcp_batch_high_lock);
9421 * Effectively disable pcplists for the zone by setting the high limit to 0
9422 * and draining all cpus. A concurrent page freeing on another CPU that's about
9423 * to put the page on pcplist will either finish before the drain and the page
9424 * will be drained, or observe the new high limit and skip the pcplist.
9426 * Must be paired with a call to zone_pcp_enable().
9428 void zone_pcp_disable(struct zone *zone)
9430 mutex_lock(&pcp_batch_high_lock);
9431 __zone_set_pageset_high_and_batch(zone, 0, 1);
9432 __drain_all_pages(zone, true);
9435 void zone_pcp_enable(struct zone *zone)
9437 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9438 mutex_unlock(&pcp_batch_high_lock);
9441 void zone_pcp_reset(struct zone *zone)
9444 struct per_cpu_zonestat *pzstats;
9446 if (zone->per_cpu_pageset != &boot_pageset) {
9447 for_each_online_cpu(cpu) {
9448 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9449 drain_zonestat(zone, pzstats);
9451 free_percpu(zone->per_cpu_pageset);
9452 free_percpu(zone->per_cpu_zonestats);
9453 zone->per_cpu_pageset = &boot_pageset;
9454 zone->per_cpu_zonestats = &boot_zonestats;
9458 #ifdef CONFIG_MEMORY_HOTREMOVE
9460 * All pages in the range must be in a single zone, must not contain holes,
9461 * must span full sections, and must be isolated before calling this function.
9463 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9465 unsigned long pfn = start_pfn;
9469 unsigned long flags;
9471 offline_mem_sections(pfn, end_pfn);
9472 zone = page_zone(pfn_to_page(pfn));
9473 spin_lock_irqsave(&zone->lock, flags);
9474 while (pfn < end_pfn) {
9475 page = pfn_to_page(pfn);
9477 * The HWPoisoned page may be not in buddy system, and
9478 * page_count() is not 0.
9480 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9485 * At this point all remaining PageOffline() pages have a
9486 * reference count of 0 and can simply be skipped.
9488 if (PageOffline(page)) {
9489 BUG_ON(page_count(page));
9490 BUG_ON(PageBuddy(page));
9495 BUG_ON(page_count(page));
9496 BUG_ON(!PageBuddy(page));
9497 order = buddy_order(page);
9498 del_page_from_free_list(page, zone, order);
9499 pfn += (1 << order);
9501 spin_unlock_irqrestore(&zone->lock, flags);
9506 * This function returns a stable result only if called under zone lock.
9508 bool is_free_buddy_page(struct page *page)
9510 unsigned long pfn = page_to_pfn(page);
9513 for (order = 0; order < MAX_ORDER; order++) {
9514 struct page *page_head = page - (pfn & ((1 << order) - 1));
9516 if (PageBuddy(page_head) &&
9517 buddy_order_unsafe(page_head) >= order)
9521 return order < MAX_ORDER;
9523 EXPORT_SYMBOL(is_free_buddy_page);
9525 #ifdef CONFIG_MEMORY_FAILURE
9527 * Break down a higher-order page in sub-pages, and keep our target out of
9530 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9531 struct page *target, int low, int high,
9534 unsigned long size = 1 << high;
9535 struct page *current_buddy, *next_page;
9537 while (high > low) {
9541 if (target >= &page[size]) {
9542 next_page = page + size;
9543 current_buddy = page;
9546 current_buddy = page + size;
9549 if (set_page_guard(zone, current_buddy, high, migratetype))
9552 if (current_buddy != target) {
9553 add_to_free_list(current_buddy, zone, high, migratetype);
9554 set_buddy_order(current_buddy, high);
9561 * Take a page that will be marked as poisoned off the buddy allocator.
9563 bool take_page_off_buddy(struct page *page)
9565 struct zone *zone = page_zone(page);
9566 unsigned long pfn = page_to_pfn(page);
9567 unsigned long flags;
9571 spin_lock_irqsave(&zone->lock, flags);
9572 for (order = 0; order < MAX_ORDER; order++) {
9573 struct page *page_head = page - (pfn & ((1 << order) - 1));
9574 int page_order = buddy_order(page_head);
9576 if (PageBuddy(page_head) && page_order >= order) {
9577 unsigned long pfn_head = page_to_pfn(page_head);
9578 int migratetype = get_pfnblock_migratetype(page_head,
9581 del_page_from_free_list(page_head, zone, page_order);
9582 break_down_buddy_pages(zone, page_head, page, 0,
9583 page_order, migratetype);
9584 SetPageHWPoisonTakenOff(page);
9585 if (!is_migrate_isolate(migratetype))
9586 __mod_zone_freepage_state(zone, -1, migratetype);
9590 if (page_count(page_head) > 0)
9593 spin_unlock_irqrestore(&zone->lock, flags);
9598 * Cancel takeoff done by take_page_off_buddy().
9600 bool put_page_back_buddy(struct page *page)
9602 struct zone *zone = page_zone(page);
9603 unsigned long pfn = page_to_pfn(page);
9604 unsigned long flags;
9605 int migratetype = get_pfnblock_migratetype(page, pfn);
9608 spin_lock_irqsave(&zone->lock, flags);
9609 if (put_page_testzero(page)) {
9610 ClearPageHWPoisonTakenOff(page);
9611 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9612 if (TestClearPageHWPoison(page)) {
9616 spin_unlock_irqrestore(&zone->lock, flags);
9622 #ifdef CONFIG_ZONE_DMA
9623 bool has_managed_dma(void)
9625 struct pglist_data *pgdat;
9627 for_each_online_pgdat(pgdat) {
9628 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9630 if (managed_zone(zone))
9635 #endif /* CONFIG_ZONE_DMA */