1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/swapops.h>
23 #include <linux/interrupt.h>
24 #include <linux/pagemap.h>
25 #include <linux/jiffies.h>
26 #include <linux/memblock.h>
27 #include <linux/compiler.h>
28 #include <linux/kernel.h>
29 #include <linux/kasan.h>
30 #include <linux/kmsan.h>
31 #include <linux/module.h>
32 #include <linux/suspend.h>
33 #include <linux/pagevec.h>
34 #include <linux/blkdev.h>
35 #include <linux/slab.h>
36 #include <linux/ratelimit.h>
37 #include <linux/oom.h>
38 #include <linux/topology.h>
39 #include <linux/sysctl.h>
40 #include <linux/cpu.h>
41 #include <linux/cpuset.h>
42 #include <linux/memory_hotplug.h>
43 #include <linux/nodemask.h>
44 #include <linux/vmalloc.h>
45 #include <linux/vmstat.h>
46 #include <linux/mempolicy.h>
47 #include <linux/memremap.h>
48 #include <linux/stop_machine.h>
49 #include <linux/random.h>
50 #include <linux/sort.h>
51 #include <linux/pfn.h>
52 #include <linux/backing-dev.h>
53 #include <linux/fault-inject.h>
54 #include <linux/page-isolation.h>
55 #include <linux/debugobjects.h>
56 #include <linux/kmemleak.h>
57 #include <linux/compaction.h>
58 #include <trace/events/kmem.h>
59 #include <trace/events/oom.h>
60 #include <linux/prefetch.h>
61 #include <linux/mm_inline.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/migrate.h>
64 #include <linux/hugetlb.h>
65 #include <linux/sched/rt.h>
66 #include <linux/sched/mm.h>
67 #include <linux/page_owner.h>
68 #include <linux/page_table_check.h>
69 #include <linux/kthread.h>
70 #include <linux/memcontrol.h>
71 #include <linux/ftrace.h>
72 #include <linux/lockdep.h>
73 #include <linux/nmi.h>
74 #include <linux/psi.h>
75 #include <linux/padata.h>
76 #include <linux/khugepaged.h>
77 #include <linux/buffer_head.h>
78 #include <linux/delayacct.h>
79 #include <asm/sections.h>
80 #include <asm/tlbflush.h>
81 #include <asm/div64.h>
84 #include "page_reporting.h"
87 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
88 typedef int __bitwise fpi_t;
90 /* No special request */
91 #define FPI_NONE ((__force fpi_t)0)
94 * Skip free page reporting notification for the (possibly merged) page.
95 * This does not hinder free page reporting from grabbing the page,
96 * reporting it and marking it "reported" - it only skips notifying
97 * the free page reporting infrastructure about a newly freed page. For
98 * example, used when temporarily pulling a page from a freelist and
99 * putting it back unmodified.
101 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
104 * Place the (possibly merged) page to the tail of the freelist. Will ignore
105 * page shuffling (relevant code - e.g., memory onlining - is expected to
106 * shuffle the whole zone).
108 * Note: No code should rely on this flag for correctness - it's purely
109 * to allow for optimizations when handing back either fresh pages
110 * (memory onlining) or untouched pages (page isolation, free page
113 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
116 * Don't poison memory with KASAN (only for the tag-based modes).
117 * During boot, all non-reserved memblock memory is exposed to page_alloc.
118 * Poisoning all that memory lengthens boot time, especially on systems with
119 * large amount of RAM. This flag is used to skip that poisoning.
120 * This is only done for the tag-based KASAN modes, as those are able to
121 * detect memory corruptions with the memory tags assigned by default.
122 * All memory allocated normally after boot gets poisoned as usual.
124 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
126 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
127 static DEFINE_MUTEX(pcp_batch_high_lock);
128 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
130 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
132 * On SMP, spin_trylock is sufficient protection.
133 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
135 #define pcp_trylock_prepare(flags) do { } while (0)
136 #define pcp_trylock_finish(flag) do { } while (0)
139 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
140 #define pcp_trylock_prepare(flags) local_irq_save(flags)
141 #define pcp_trylock_finish(flags) local_irq_restore(flags)
145 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
146 * a migration causing the wrong PCP to be locked and remote memory being
147 * potentially allocated, pin the task to the CPU for the lookup+lock.
148 * preempt_disable is used on !RT because it is faster than migrate_disable.
149 * migrate_disable is used on RT because otherwise RT spinlock usage is
150 * interfered with and a high priority task cannot preempt the allocator.
152 #ifndef CONFIG_PREEMPT_RT
153 #define pcpu_task_pin() preempt_disable()
154 #define pcpu_task_unpin() preempt_enable()
156 #define pcpu_task_pin() migrate_disable()
157 #define pcpu_task_unpin() migrate_enable()
161 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
162 * Return value should be used with equivalent unlock helper.
164 #define pcpu_spin_lock(type, member, ptr) \
168 _ret = this_cpu_ptr(ptr); \
169 spin_lock(&_ret->member); \
173 #define pcpu_spin_lock_irqsave(type, member, ptr, flags) \
177 _ret = this_cpu_ptr(ptr); \
178 spin_lock_irqsave(&_ret->member, flags); \
182 #define pcpu_spin_trylock_irqsave(type, member, ptr, flags) \
186 _ret = this_cpu_ptr(ptr); \
187 if (!spin_trylock_irqsave(&_ret->member, flags)) { \
194 #define pcpu_spin_unlock(member, ptr) \
196 spin_unlock(&ptr->member); \
200 #define pcpu_spin_unlock_irqrestore(member, ptr, flags) \
202 spin_unlock_irqrestore(&ptr->member, flags); \
206 /* struct per_cpu_pages specific helpers. */
207 #define pcp_spin_lock(ptr) \
208 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
210 #define pcp_spin_lock_irqsave(ptr, flags) \
211 pcpu_spin_lock_irqsave(struct per_cpu_pages, lock, ptr, flags)
213 #define pcp_spin_trylock_irqsave(ptr, flags) \
214 pcpu_spin_trylock_irqsave(struct per_cpu_pages, lock, ptr, flags)
216 #define pcp_spin_unlock(ptr) \
217 pcpu_spin_unlock(lock, ptr)
219 #define pcp_spin_unlock_irqrestore(ptr, flags) \
220 pcpu_spin_unlock_irqrestore(lock, ptr, flags)
221 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
222 DEFINE_PER_CPU(int, numa_node);
223 EXPORT_PER_CPU_SYMBOL(numa_node);
226 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
228 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
230 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
231 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
232 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
233 * defined in <linux/topology.h>.
235 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
236 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
239 static DEFINE_MUTEX(pcpu_drain_mutex);
241 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
242 volatile unsigned long latent_entropy __latent_entropy;
243 EXPORT_SYMBOL(latent_entropy);
247 * Array of node states.
249 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
250 [N_POSSIBLE] = NODE_MASK_ALL,
251 [N_ONLINE] = { { [0] = 1UL } },
253 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
254 #ifdef CONFIG_HIGHMEM
255 [N_HIGH_MEMORY] = { { [0] = 1UL } },
257 [N_MEMORY] = { { [0] = 1UL } },
258 [N_CPU] = { { [0] = 1UL } },
261 EXPORT_SYMBOL(node_states);
263 atomic_long_t _totalram_pages __read_mostly;
264 EXPORT_SYMBOL(_totalram_pages);
265 unsigned long totalreserve_pages __read_mostly;
266 unsigned long totalcma_pages __read_mostly;
268 int percpu_pagelist_high_fraction;
269 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
270 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
271 EXPORT_SYMBOL(init_on_alloc);
273 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
274 EXPORT_SYMBOL(init_on_free);
276 static bool _init_on_alloc_enabled_early __read_mostly
277 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
278 static int __init early_init_on_alloc(char *buf)
281 return kstrtobool(buf, &_init_on_alloc_enabled_early);
283 early_param("init_on_alloc", early_init_on_alloc);
285 static bool _init_on_free_enabled_early __read_mostly
286 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
287 static int __init early_init_on_free(char *buf)
289 return kstrtobool(buf, &_init_on_free_enabled_early);
291 early_param("init_on_free", early_init_on_free);
294 * A cached value of the page's pageblock's migratetype, used when the page is
295 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
296 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
297 * Also the migratetype set in the page does not necessarily match the pcplist
298 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
299 * other index - this ensures that it will be put on the correct CMA freelist.
301 static inline int get_pcppage_migratetype(struct page *page)
306 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
308 page->index = migratetype;
311 #ifdef CONFIG_PM_SLEEP
313 * The following functions are used by the suspend/hibernate code to temporarily
314 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
315 * while devices are suspended. To avoid races with the suspend/hibernate code,
316 * they should always be called with system_transition_mutex held
317 * (gfp_allowed_mask also should only be modified with system_transition_mutex
318 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
319 * with that modification).
322 static gfp_t saved_gfp_mask;
324 void pm_restore_gfp_mask(void)
326 WARN_ON(!mutex_is_locked(&system_transition_mutex));
327 if (saved_gfp_mask) {
328 gfp_allowed_mask = saved_gfp_mask;
333 void pm_restrict_gfp_mask(void)
335 WARN_ON(!mutex_is_locked(&system_transition_mutex));
336 WARN_ON(saved_gfp_mask);
337 saved_gfp_mask = gfp_allowed_mask;
338 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
341 bool pm_suspended_storage(void)
343 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
347 #endif /* CONFIG_PM_SLEEP */
349 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
350 unsigned int pageblock_order __read_mostly;
353 static void __free_pages_ok(struct page *page, unsigned int order,
357 * results with 256, 32 in the lowmem_reserve sysctl:
358 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
359 * 1G machine -> (16M dma, 784M normal, 224M high)
360 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
361 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
362 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
364 * TBD: should special case ZONE_DMA32 machines here - in those we normally
365 * don't need any ZONE_NORMAL reservation
367 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
368 #ifdef CONFIG_ZONE_DMA
371 #ifdef CONFIG_ZONE_DMA32
375 #ifdef CONFIG_HIGHMEM
381 static char * const zone_names[MAX_NR_ZONES] = {
382 #ifdef CONFIG_ZONE_DMA
385 #ifdef CONFIG_ZONE_DMA32
389 #ifdef CONFIG_HIGHMEM
393 #ifdef CONFIG_ZONE_DEVICE
398 const char * const migratetype_names[MIGRATE_TYPES] = {
406 #ifdef CONFIG_MEMORY_ISOLATION
411 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
412 [NULL_COMPOUND_DTOR] = NULL,
413 [COMPOUND_PAGE_DTOR] = free_compound_page,
414 #ifdef CONFIG_HUGETLB_PAGE
415 [HUGETLB_PAGE_DTOR] = free_huge_page,
417 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
418 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
422 int min_free_kbytes = 1024;
423 int user_min_free_kbytes = -1;
424 int watermark_boost_factor __read_mostly = 15000;
425 int watermark_scale_factor = 10;
427 static unsigned long nr_kernel_pages __initdata;
428 static unsigned long nr_all_pages __initdata;
429 static unsigned long dma_reserve __initdata;
431 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
432 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
433 static unsigned long required_kernelcore __initdata;
434 static unsigned long required_kernelcore_percent __initdata;
435 static unsigned long required_movablecore __initdata;
436 static unsigned long required_movablecore_percent __initdata;
437 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
438 bool mirrored_kernelcore __initdata_memblock;
440 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
442 EXPORT_SYMBOL(movable_zone);
445 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
446 unsigned int nr_online_nodes __read_mostly = 1;
447 EXPORT_SYMBOL(nr_node_ids);
448 EXPORT_SYMBOL(nr_online_nodes);
451 int page_group_by_mobility_disabled __read_mostly;
453 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
455 * During boot we initialize deferred pages on-demand, as needed, but once
456 * page_alloc_init_late() has finished, the deferred pages are all initialized,
457 * and we can permanently disable that path.
459 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
461 static inline bool deferred_pages_enabled(void)
463 return static_branch_unlikely(&deferred_pages);
466 /* Returns true if the struct page for the pfn is uninitialised */
467 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
469 int nid = early_pfn_to_nid(pfn);
471 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
478 * Returns true when the remaining initialisation should be deferred until
479 * later in the boot cycle when it can be parallelised.
481 static bool __meminit
482 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
484 static unsigned long prev_end_pfn, nr_initialised;
486 if (early_page_ext_enabled())
489 * prev_end_pfn static that contains the end of previous zone
490 * No need to protect because called very early in boot before smp_init.
492 if (prev_end_pfn != end_pfn) {
493 prev_end_pfn = end_pfn;
497 /* Always populate low zones for address-constrained allocations */
498 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
501 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
504 * We start only with one section of pages, more pages are added as
505 * needed until the rest of deferred pages are initialized.
508 if ((nr_initialised > PAGES_PER_SECTION) &&
509 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
510 NODE_DATA(nid)->first_deferred_pfn = pfn;
516 static inline bool deferred_pages_enabled(void)
521 static inline bool early_page_uninitialised(unsigned long pfn)
526 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
532 /* Return a pointer to the bitmap storing bits affecting a block of pages */
533 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
536 #ifdef CONFIG_SPARSEMEM
537 return section_to_usemap(__pfn_to_section(pfn));
539 return page_zone(page)->pageblock_flags;
540 #endif /* CONFIG_SPARSEMEM */
543 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
545 #ifdef CONFIG_SPARSEMEM
546 pfn &= (PAGES_PER_SECTION-1);
548 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
549 #endif /* CONFIG_SPARSEMEM */
550 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
553 static __always_inline
554 unsigned long __get_pfnblock_flags_mask(const struct page *page,
558 unsigned long *bitmap;
559 unsigned long bitidx, word_bitidx;
562 bitmap = get_pageblock_bitmap(page, pfn);
563 bitidx = pfn_to_bitidx(page, pfn);
564 word_bitidx = bitidx / BITS_PER_LONG;
565 bitidx &= (BITS_PER_LONG-1);
567 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
568 * a consistent read of the memory array, so that results, even though
569 * racy, are not corrupted.
571 word = READ_ONCE(bitmap[word_bitidx]);
572 return (word >> bitidx) & mask;
576 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
577 * @page: The page within the block of interest
578 * @pfn: The target page frame number
579 * @mask: mask of bits that the caller is interested in
581 * Return: pageblock_bits flags
583 unsigned long get_pfnblock_flags_mask(const struct page *page,
584 unsigned long pfn, unsigned long mask)
586 return __get_pfnblock_flags_mask(page, pfn, mask);
589 static __always_inline int get_pfnblock_migratetype(const struct page *page,
592 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
596 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
597 * @page: The page within the block of interest
598 * @flags: The flags to set
599 * @pfn: The target page frame number
600 * @mask: mask of bits that the caller is interested in
602 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
606 unsigned long *bitmap;
607 unsigned long bitidx, word_bitidx;
610 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
611 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
613 bitmap = get_pageblock_bitmap(page, pfn);
614 bitidx = pfn_to_bitidx(page, pfn);
615 word_bitidx = bitidx / BITS_PER_LONG;
616 bitidx &= (BITS_PER_LONG-1);
618 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
623 word = READ_ONCE(bitmap[word_bitidx]);
625 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
628 void set_pageblock_migratetype(struct page *page, int migratetype)
630 if (unlikely(page_group_by_mobility_disabled &&
631 migratetype < MIGRATE_PCPTYPES))
632 migratetype = MIGRATE_UNMOVABLE;
634 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
635 page_to_pfn(page), MIGRATETYPE_MASK);
638 #ifdef CONFIG_DEBUG_VM
639 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
643 unsigned long pfn = page_to_pfn(page);
644 unsigned long sp, start_pfn;
647 seq = zone_span_seqbegin(zone);
648 start_pfn = zone->zone_start_pfn;
649 sp = zone->spanned_pages;
650 if (!zone_spans_pfn(zone, pfn))
652 } while (zone_span_seqretry(zone, seq));
655 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
656 pfn, zone_to_nid(zone), zone->name,
657 start_pfn, start_pfn + sp);
662 static int page_is_consistent(struct zone *zone, struct page *page)
664 if (zone != page_zone(page))
670 * Temporary debugging check for pages not lying within a given zone.
672 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
674 if (page_outside_zone_boundaries(zone, page))
676 if (!page_is_consistent(zone, page))
682 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
688 static void bad_page(struct page *page, const char *reason)
690 static unsigned long resume;
691 static unsigned long nr_shown;
692 static unsigned long nr_unshown;
695 * Allow a burst of 60 reports, then keep quiet for that minute;
696 * or allow a steady drip of one report per second.
698 if (nr_shown == 60) {
699 if (time_before(jiffies, resume)) {
705 "BUG: Bad page state: %lu messages suppressed\n",
712 resume = jiffies + 60 * HZ;
714 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
715 current->comm, page_to_pfn(page));
716 dump_page(page, reason);
721 /* Leave bad fields for debug, except PageBuddy could make trouble */
722 page_mapcount_reset(page); /* remove PageBuddy */
723 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
726 static inline unsigned int order_to_pindex(int migratetype, int order)
730 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
731 if (order > PAGE_ALLOC_COSTLY_ORDER) {
732 VM_BUG_ON(order != pageblock_order);
733 return NR_LOWORDER_PCP_LISTS;
736 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
739 return (MIGRATE_PCPTYPES * base) + migratetype;
742 static inline int pindex_to_order(unsigned int pindex)
744 int order = pindex / MIGRATE_PCPTYPES;
746 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
747 if (pindex == NR_LOWORDER_PCP_LISTS)
748 order = pageblock_order;
750 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
756 static inline bool pcp_allowed_order(unsigned int order)
758 if (order <= PAGE_ALLOC_COSTLY_ORDER)
760 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
761 if (order == pageblock_order)
767 static inline void free_the_page(struct page *page, unsigned int order)
769 if (pcp_allowed_order(order)) /* Via pcp? */
770 free_unref_page(page, order);
772 __free_pages_ok(page, order, FPI_NONE);
776 * Higher-order pages are called "compound pages". They are structured thusly:
778 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
780 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
781 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
783 * The first tail page's ->compound_dtor holds the offset in array of compound
784 * page destructors. See compound_page_dtors.
786 * The first tail page's ->compound_order holds the order of allocation.
787 * This usage means that zero-order pages may not be compound.
790 void free_compound_page(struct page *page)
792 mem_cgroup_uncharge(page_folio(page));
793 free_the_page(page, compound_order(page));
796 static void prep_compound_head(struct page *page, unsigned int order)
798 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
799 set_compound_order(page, order);
800 atomic_set(compound_mapcount_ptr(page), -1);
801 atomic_set(compound_pincount_ptr(page), 0);
804 static void prep_compound_tail(struct page *head, int tail_idx)
806 struct page *p = head + tail_idx;
808 p->mapping = TAIL_MAPPING;
809 set_compound_head(p, head);
810 set_page_private(p, 0);
813 void prep_compound_page(struct page *page, unsigned int order)
816 int nr_pages = 1 << order;
819 for (i = 1; i < nr_pages; i++)
820 prep_compound_tail(page, i);
822 prep_compound_head(page, order);
825 void destroy_large_folio(struct folio *folio)
827 enum compound_dtor_id dtor = folio_page(folio, 1)->compound_dtor;
829 VM_BUG_ON_FOLIO(dtor >= NR_COMPOUND_DTORS, folio);
830 compound_page_dtors[dtor](&folio->page);
833 #ifdef CONFIG_DEBUG_PAGEALLOC
834 unsigned int _debug_guardpage_minorder;
836 bool _debug_pagealloc_enabled_early __read_mostly
837 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
838 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
839 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
840 EXPORT_SYMBOL(_debug_pagealloc_enabled);
842 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
844 static int __init early_debug_pagealloc(char *buf)
846 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
848 early_param("debug_pagealloc", early_debug_pagealloc);
850 static int __init debug_guardpage_minorder_setup(char *buf)
854 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
855 pr_err("Bad debug_guardpage_minorder value\n");
858 _debug_guardpage_minorder = res;
859 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
862 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
864 static inline bool set_page_guard(struct zone *zone, struct page *page,
865 unsigned int order, int migratetype)
867 if (!debug_guardpage_enabled())
870 if (order >= debug_guardpage_minorder())
873 __SetPageGuard(page);
874 INIT_LIST_HEAD(&page->buddy_list);
875 set_page_private(page, order);
876 /* Guard pages are not available for any usage */
877 if (!is_migrate_isolate(migratetype))
878 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
883 static inline void clear_page_guard(struct zone *zone, struct page *page,
884 unsigned int order, int migratetype)
886 if (!debug_guardpage_enabled())
889 __ClearPageGuard(page);
891 set_page_private(page, 0);
892 if (!is_migrate_isolate(migratetype))
893 __mod_zone_freepage_state(zone, (1 << order), migratetype);
896 static inline bool set_page_guard(struct zone *zone, struct page *page,
897 unsigned int order, int migratetype) { return false; }
898 static inline void clear_page_guard(struct zone *zone, struct page *page,
899 unsigned int order, int migratetype) {}
903 * Enable static keys related to various memory debugging and hardening options.
904 * Some override others, and depend on early params that are evaluated in the
905 * order of appearance. So we need to first gather the full picture of what was
906 * enabled, and then make decisions.
908 void __init init_mem_debugging_and_hardening(void)
910 bool page_poisoning_requested = false;
912 #ifdef CONFIG_PAGE_POISONING
914 * Page poisoning is debug page alloc for some arches. If
915 * either of those options are enabled, enable poisoning.
917 if (page_poisoning_enabled() ||
918 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
919 debug_pagealloc_enabled())) {
920 static_branch_enable(&_page_poisoning_enabled);
921 page_poisoning_requested = true;
925 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
926 page_poisoning_requested) {
927 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
928 "will take precedence over init_on_alloc and init_on_free\n");
929 _init_on_alloc_enabled_early = false;
930 _init_on_free_enabled_early = false;
933 if (_init_on_alloc_enabled_early)
934 static_branch_enable(&init_on_alloc);
936 static_branch_disable(&init_on_alloc);
938 if (_init_on_free_enabled_early)
939 static_branch_enable(&init_on_free);
941 static_branch_disable(&init_on_free);
943 if (IS_ENABLED(CONFIG_KMSAN) &&
944 (_init_on_alloc_enabled_early || _init_on_free_enabled_early))
945 pr_info("mem auto-init: please make sure init_on_alloc and init_on_free are disabled when running KMSAN\n");
947 #ifdef CONFIG_DEBUG_PAGEALLOC
948 if (!debug_pagealloc_enabled())
951 static_branch_enable(&_debug_pagealloc_enabled);
953 if (!debug_guardpage_minorder())
956 static_branch_enable(&_debug_guardpage_enabled);
960 static inline void set_buddy_order(struct page *page, unsigned int order)
962 set_page_private(page, order);
963 __SetPageBuddy(page);
966 #ifdef CONFIG_COMPACTION
967 static inline struct capture_control *task_capc(struct zone *zone)
969 struct capture_control *capc = current->capture_control;
971 return unlikely(capc) &&
972 !(current->flags & PF_KTHREAD) &&
974 capc->cc->zone == zone ? capc : NULL;
978 compaction_capture(struct capture_control *capc, struct page *page,
979 int order, int migratetype)
981 if (!capc || order != capc->cc->order)
984 /* Do not accidentally pollute CMA or isolated regions*/
985 if (is_migrate_cma(migratetype) ||
986 is_migrate_isolate(migratetype))
990 * Do not let lower order allocations pollute a movable pageblock.
991 * This might let an unmovable request use a reclaimable pageblock
992 * and vice-versa but no more than normal fallback logic which can
993 * have trouble finding a high-order free page.
995 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
1003 static inline struct capture_control *task_capc(struct zone *zone)
1009 compaction_capture(struct capture_control *capc, struct page *page,
1010 int order, int migratetype)
1014 #endif /* CONFIG_COMPACTION */
1016 /* Used for pages not on another list */
1017 static inline void add_to_free_list(struct page *page, struct zone *zone,
1018 unsigned int order, int migratetype)
1020 struct free_area *area = &zone->free_area[order];
1022 list_add(&page->buddy_list, &area->free_list[migratetype]);
1026 /* Used for pages not on another list */
1027 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
1028 unsigned int order, int migratetype)
1030 struct free_area *area = &zone->free_area[order];
1032 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
1037 * Used for pages which are on another list. Move the pages to the tail
1038 * of the list - so the moved pages won't immediately be considered for
1039 * allocation again (e.g., optimization for memory onlining).
1041 static inline void move_to_free_list(struct page *page, struct zone *zone,
1042 unsigned int order, int migratetype)
1044 struct free_area *area = &zone->free_area[order];
1046 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
1049 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
1052 /* clear reported state and update reported page count */
1053 if (page_reported(page))
1054 __ClearPageReported(page);
1056 list_del(&page->buddy_list);
1057 __ClearPageBuddy(page);
1058 set_page_private(page, 0);
1059 zone->free_area[order].nr_free--;
1063 * If this is not the largest possible page, check if the buddy
1064 * of the next-highest order is free. If it is, it's possible
1065 * that pages are being freed that will coalesce soon. In case,
1066 * that is happening, add the free page to the tail of the list
1067 * so it's less likely to be used soon and more likely to be merged
1068 * as a higher order page
1071 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1072 struct page *page, unsigned int order)
1074 unsigned long higher_page_pfn;
1075 struct page *higher_page;
1077 if (order >= MAX_ORDER - 2)
1080 higher_page_pfn = buddy_pfn & pfn;
1081 higher_page = page + (higher_page_pfn - pfn);
1083 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
1088 * Freeing function for a buddy system allocator.
1090 * The concept of a buddy system is to maintain direct-mapped table
1091 * (containing bit values) for memory blocks of various "orders".
1092 * The bottom level table contains the map for the smallest allocatable
1093 * units of memory (here, pages), and each level above it describes
1094 * pairs of units from the levels below, hence, "buddies".
1095 * At a high level, all that happens here is marking the table entry
1096 * at the bottom level available, and propagating the changes upward
1097 * as necessary, plus some accounting needed to play nicely with other
1098 * parts of the VM system.
1099 * At each level, we keep a list of pages, which are heads of continuous
1100 * free pages of length of (1 << order) and marked with PageBuddy.
1101 * Page's order is recorded in page_private(page) field.
1102 * So when we are allocating or freeing one, we can derive the state of the
1103 * other. That is, if we allocate a small block, and both were
1104 * free, the remainder of the region must be split into blocks.
1105 * If a block is freed, and its buddy is also free, then this
1106 * triggers coalescing into a block of larger size.
1111 static inline void __free_one_page(struct page *page,
1113 struct zone *zone, unsigned int order,
1114 int migratetype, fpi_t fpi_flags)
1116 struct capture_control *capc = task_capc(zone);
1117 unsigned long buddy_pfn = 0;
1118 unsigned long combined_pfn;
1122 VM_BUG_ON(!zone_is_initialized(zone));
1123 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1125 VM_BUG_ON(migratetype == -1);
1126 if (likely(!is_migrate_isolate(migratetype)))
1127 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1129 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1130 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1132 while (order < MAX_ORDER - 1) {
1133 if (compaction_capture(capc, page, order, migratetype)) {
1134 __mod_zone_freepage_state(zone, -(1 << order),
1139 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
1143 if (unlikely(order >= pageblock_order)) {
1145 * We want to prevent merge between freepages on pageblock
1146 * without fallbacks and normal pageblock. Without this,
1147 * pageblock isolation could cause incorrect freepage or CMA
1148 * accounting or HIGHATOMIC accounting.
1150 int buddy_mt = get_pageblock_migratetype(buddy);
1152 if (migratetype != buddy_mt
1153 && (!migratetype_is_mergeable(migratetype) ||
1154 !migratetype_is_mergeable(buddy_mt)))
1159 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1160 * merge with it and move up one order.
1162 if (page_is_guard(buddy))
1163 clear_page_guard(zone, buddy, order, migratetype);
1165 del_page_from_free_list(buddy, zone, order);
1166 combined_pfn = buddy_pfn & pfn;
1167 page = page + (combined_pfn - pfn);
1173 set_buddy_order(page, order);
1175 if (fpi_flags & FPI_TO_TAIL)
1177 else if (is_shuffle_order(order))
1178 to_tail = shuffle_pick_tail();
1180 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1183 add_to_free_list_tail(page, zone, order, migratetype);
1185 add_to_free_list(page, zone, order, migratetype);
1187 /* Notify page reporting subsystem of freed page */
1188 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1189 page_reporting_notify_free(order);
1193 * split_free_page() -- split a free page at split_pfn_offset
1194 * @free_page: the original free page
1195 * @order: the order of the page
1196 * @split_pfn_offset: split offset within the page
1198 * Return -ENOENT if the free page is changed, otherwise 0
1200 * It is used when the free page crosses two pageblocks with different migratetypes
1201 * at split_pfn_offset within the page. The split free page will be put into
1202 * separate migratetype lists afterwards. Otherwise, the function achieves
1205 int split_free_page(struct page *free_page,
1206 unsigned int order, unsigned long split_pfn_offset)
1208 struct zone *zone = page_zone(free_page);
1209 unsigned long free_page_pfn = page_to_pfn(free_page);
1211 unsigned long flags;
1212 int free_page_order;
1216 if (split_pfn_offset == 0)
1219 spin_lock_irqsave(&zone->lock, flags);
1221 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
1226 mt = get_pageblock_migratetype(free_page);
1227 if (likely(!is_migrate_isolate(mt)))
1228 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1230 del_page_from_free_list(free_page, zone, order);
1231 for (pfn = free_page_pfn;
1232 pfn < free_page_pfn + (1UL << order);) {
1233 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
1235 free_page_order = min_t(unsigned int,
1236 pfn ? __ffs(pfn) : order,
1237 __fls(split_pfn_offset));
1238 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
1240 pfn += 1UL << free_page_order;
1241 split_pfn_offset -= (1UL << free_page_order);
1242 /* we have done the first part, now switch to second part */
1243 if (split_pfn_offset == 0)
1244 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
1247 spin_unlock_irqrestore(&zone->lock, flags);
1251 * A bad page could be due to a number of fields. Instead of multiple branches,
1252 * try and check multiple fields with one check. The caller must do a detailed
1253 * check if necessary.
1255 static inline bool page_expected_state(struct page *page,
1256 unsigned long check_flags)
1258 if (unlikely(atomic_read(&page->_mapcount) != -1))
1261 if (unlikely((unsigned long)page->mapping |
1262 page_ref_count(page) |
1266 (page->flags & check_flags)))
1272 static const char *page_bad_reason(struct page *page, unsigned long flags)
1274 const char *bad_reason = NULL;
1276 if (unlikely(atomic_read(&page->_mapcount) != -1))
1277 bad_reason = "nonzero mapcount";
1278 if (unlikely(page->mapping != NULL))
1279 bad_reason = "non-NULL mapping";
1280 if (unlikely(page_ref_count(page) != 0))
1281 bad_reason = "nonzero _refcount";
1282 if (unlikely(page->flags & flags)) {
1283 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1284 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1286 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1289 if (unlikely(page->memcg_data))
1290 bad_reason = "page still charged to cgroup";
1295 static void free_page_is_bad_report(struct page *page)
1298 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1301 static inline bool free_page_is_bad(struct page *page)
1303 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1306 /* Something has gone sideways, find it */
1307 free_page_is_bad_report(page);
1311 static int free_tail_pages_check(struct page *head_page, struct page *page)
1316 * We rely page->lru.next never has bit 0 set, unless the page
1317 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1319 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1321 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1325 switch (page - head_page) {
1327 /* the first tail page: ->mapping may be compound_mapcount() */
1328 if (unlikely(compound_mapcount(page))) {
1329 bad_page(page, "nonzero compound_mapcount");
1335 * the second tail page: ->mapping is
1336 * deferred_list.next -- ignore value.
1340 if (page->mapping != TAIL_MAPPING) {
1341 bad_page(page, "corrupted mapping in tail page");
1346 if (unlikely(!PageTail(page))) {
1347 bad_page(page, "PageTail not set");
1350 if (unlikely(compound_head(page) != head_page)) {
1351 bad_page(page, "compound_head not consistent");
1356 page->mapping = NULL;
1357 clear_compound_head(page);
1362 * Skip KASAN memory poisoning when either:
1364 * 1. Deferred memory initialization has not yet completed,
1365 * see the explanation below.
1366 * 2. Skipping poisoning is requested via FPI_SKIP_KASAN_POISON,
1367 * see the comment next to it.
1368 * 3. Skipping poisoning is requested via __GFP_SKIP_KASAN_POISON,
1369 * see the comment next to it.
1371 * Poisoning pages during deferred memory init will greatly lengthen the
1372 * process and cause problem in large memory systems as the deferred pages
1373 * initialization is done with interrupt disabled.
1375 * Assuming that there will be no reference to those newly initialized
1376 * pages before they are ever allocated, this should have no effect on
1377 * KASAN memory tracking as the poison will be properly inserted at page
1378 * allocation time. The only corner case is when pages are allocated by
1379 * on-demand allocation and then freed again before the deferred pages
1380 * initialization is done, but this is not likely to happen.
1382 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1384 return deferred_pages_enabled() ||
1385 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
1386 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
1387 PageSkipKASanPoison(page);
1390 static void kernel_init_pages(struct page *page, int numpages)
1394 /* s390's use of memset() could override KASAN redzones. */
1395 kasan_disable_current();
1396 for (i = 0; i < numpages; i++)
1397 clear_highpage_kasan_tagged(page + i);
1398 kasan_enable_current();
1401 static __always_inline bool free_pages_prepare(struct page *page,
1402 unsigned int order, bool check_free, fpi_t fpi_flags)
1405 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1406 bool init = want_init_on_free();
1408 VM_BUG_ON_PAGE(PageTail(page), page);
1410 trace_mm_page_free(page, order);
1411 kmsan_free_page(page, order);
1413 if (unlikely(PageHWPoison(page)) && !order) {
1415 * Do not let hwpoison pages hit pcplists/buddy
1416 * Untie memcg state and reset page's owner
1418 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1419 __memcg_kmem_uncharge_page(page, order);
1420 reset_page_owner(page, order);
1421 page_table_check_free(page, order);
1426 * Check tail pages before head page information is cleared to
1427 * avoid checking PageCompound for order-0 pages.
1429 if (unlikely(order)) {
1430 bool compound = PageCompound(page);
1433 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1436 ClearPageDoubleMap(page);
1437 ClearPageHasHWPoisoned(page);
1439 for (i = 1; i < (1 << order); i++) {
1441 bad += free_tail_pages_check(page, page + i);
1442 if (unlikely(free_page_is_bad(page + i))) {
1446 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1449 if (PageMappingFlags(page))
1450 page->mapping = NULL;
1451 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1452 __memcg_kmem_uncharge_page(page, order);
1453 if (check_free && free_page_is_bad(page))
1458 page_cpupid_reset_last(page);
1459 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1460 reset_page_owner(page, order);
1461 page_table_check_free(page, order);
1463 if (!PageHighMem(page)) {
1464 debug_check_no_locks_freed(page_address(page),
1465 PAGE_SIZE << order);
1466 debug_check_no_obj_freed(page_address(page),
1467 PAGE_SIZE << order);
1470 kernel_poison_pages(page, 1 << order);
1473 * As memory initialization might be integrated into KASAN,
1474 * KASAN poisoning and memory initialization code must be
1475 * kept together to avoid discrepancies in behavior.
1477 * With hardware tag-based KASAN, memory tags must be set before the
1478 * page becomes unavailable via debug_pagealloc or arch_free_page.
1480 if (!skip_kasan_poison) {
1481 kasan_poison_pages(page, order, init);
1483 /* Memory is already initialized if KASAN did it internally. */
1484 if (kasan_has_integrated_init())
1488 kernel_init_pages(page, 1 << order);
1491 * arch_free_page() can make the page's contents inaccessible. s390
1492 * does this. So nothing which can access the page's contents should
1493 * happen after this.
1495 arch_free_page(page, order);
1497 debug_pagealloc_unmap_pages(page, 1 << order);
1502 #ifdef CONFIG_DEBUG_VM
1504 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1505 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1506 * moved from pcp lists to free lists.
1508 static bool free_pcp_prepare(struct page *page, unsigned int order)
1510 return free_pages_prepare(page, order, true, FPI_NONE);
1513 /* return true if this page has an inappropriate state */
1514 static bool bulkfree_pcp_prepare(struct page *page)
1516 if (debug_pagealloc_enabled_static())
1517 return free_page_is_bad(page);
1523 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1524 * moving from pcp lists to free list in order to reduce overhead. With
1525 * debug_pagealloc enabled, they are checked also immediately when being freed
1528 static bool free_pcp_prepare(struct page *page, unsigned int order)
1530 if (debug_pagealloc_enabled_static())
1531 return free_pages_prepare(page, order, true, FPI_NONE);
1533 return free_pages_prepare(page, order, false, FPI_NONE);
1536 static bool bulkfree_pcp_prepare(struct page *page)
1538 return free_page_is_bad(page);
1540 #endif /* CONFIG_DEBUG_VM */
1543 * Frees a number of pages from the PCP lists
1544 * Assumes all pages on list are in same zone.
1545 * count is the number of pages to free.
1547 static void free_pcppages_bulk(struct zone *zone, int count,
1548 struct per_cpu_pages *pcp,
1552 int max_pindex = NR_PCP_LISTS - 1;
1554 bool isolated_pageblocks;
1558 * Ensure proper count is passed which otherwise would stuck in the
1559 * below while (list_empty(list)) loop.
1561 count = min(pcp->count, count);
1563 /* Ensure requested pindex is drained first. */
1564 pindex = pindex - 1;
1566 /* Caller must hold IRQ-safe pcp->lock so IRQs are disabled. */
1567 spin_lock(&zone->lock);
1568 isolated_pageblocks = has_isolate_pageblock(zone);
1571 struct list_head *list;
1574 /* Remove pages from lists in a round-robin fashion. */
1576 if (++pindex > max_pindex)
1577 pindex = min_pindex;
1578 list = &pcp->lists[pindex];
1579 if (!list_empty(list))
1582 if (pindex == max_pindex)
1584 if (pindex == min_pindex)
1588 order = pindex_to_order(pindex);
1589 nr_pages = 1 << order;
1593 page = list_last_entry(list, struct page, pcp_list);
1594 mt = get_pcppage_migratetype(page);
1596 /* must delete to avoid corrupting pcp list */
1597 list_del(&page->pcp_list);
1599 pcp->count -= nr_pages;
1601 if (bulkfree_pcp_prepare(page))
1604 /* MIGRATE_ISOLATE page should not go to pcplists */
1605 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1606 /* Pageblock could have been isolated meanwhile */
1607 if (unlikely(isolated_pageblocks))
1608 mt = get_pageblock_migratetype(page);
1610 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1611 trace_mm_page_pcpu_drain(page, order, mt);
1612 } while (count > 0 && !list_empty(list));
1615 spin_unlock(&zone->lock);
1618 static void free_one_page(struct zone *zone,
1619 struct page *page, unsigned long pfn,
1621 int migratetype, fpi_t fpi_flags)
1623 unsigned long flags;
1625 spin_lock_irqsave(&zone->lock, flags);
1626 if (unlikely(has_isolate_pageblock(zone) ||
1627 is_migrate_isolate(migratetype))) {
1628 migratetype = get_pfnblock_migratetype(page, pfn);
1630 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1631 spin_unlock_irqrestore(&zone->lock, flags);
1634 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1635 unsigned long zone, int nid)
1637 mm_zero_struct_page(page);
1638 set_page_links(page, zone, nid, pfn);
1639 init_page_count(page);
1640 page_mapcount_reset(page);
1641 page_cpupid_reset_last(page);
1642 page_kasan_tag_reset(page);
1644 INIT_LIST_HEAD(&page->lru);
1645 #ifdef WANT_PAGE_VIRTUAL
1646 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1647 if (!is_highmem_idx(zone))
1648 set_page_address(page, __va(pfn << PAGE_SHIFT));
1652 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1653 static void __meminit init_reserved_page(unsigned long pfn)
1658 if (!early_page_uninitialised(pfn))
1661 nid = early_pfn_to_nid(pfn);
1662 pgdat = NODE_DATA(nid);
1664 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1665 struct zone *zone = &pgdat->node_zones[zid];
1667 if (zone_spans_pfn(zone, pfn))
1670 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1673 static inline void init_reserved_page(unsigned long pfn)
1676 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1679 * Initialised pages do not have PageReserved set. This function is
1680 * called for each range allocated by the bootmem allocator and
1681 * marks the pages PageReserved. The remaining valid pages are later
1682 * sent to the buddy page allocator.
1684 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1686 unsigned long start_pfn = PFN_DOWN(start);
1687 unsigned long end_pfn = PFN_UP(end);
1689 for (; start_pfn < end_pfn; start_pfn++) {
1690 if (pfn_valid(start_pfn)) {
1691 struct page *page = pfn_to_page(start_pfn);
1693 init_reserved_page(start_pfn);
1695 /* Avoid false-positive PageTail() */
1696 INIT_LIST_HEAD(&page->lru);
1699 * no need for atomic set_bit because the struct
1700 * page is not visible yet so nobody should
1703 __SetPageReserved(page);
1708 static void __free_pages_ok(struct page *page, unsigned int order,
1711 unsigned long flags;
1713 unsigned long pfn = page_to_pfn(page);
1714 struct zone *zone = page_zone(page);
1716 if (!free_pages_prepare(page, order, true, fpi_flags))
1719 migratetype = get_pfnblock_migratetype(page, pfn);
1721 spin_lock_irqsave(&zone->lock, flags);
1722 if (unlikely(has_isolate_pageblock(zone) ||
1723 is_migrate_isolate(migratetype))) {
1724 migratetype = get_pfnblock_migratetype(page, pfn);
1726 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1727 spin_unlock_irqrestore(&zone->lock, flags);
1729 __count_vm_events(PGFREE, 1 << order);
1732 void __free_pages_core(struct page *page, unsigned int order)
1734 unsigned int nr_pages = 1 << order;
1735 struct page *p = page;
1739 * When initializing the memmap, __init_single_page() sets the refcount
1740 * of all pages to 1 ("allocated"/"not free"). We have to set the
1741 * refcount of all involved pages to 0.
1744 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1746 __ClearPageReserved(p);
1747 set_page_count(p, 0);
1749 __ClearPageReserved(p);
1750 set_page_count(p, 0);
1752 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1755 * Bypass PCP and place fresh pages right to the tail, primarily
1756 * relevant for memory onlining.
1758 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1764 * During memory init memblocks map pfns to nids. The search is expensive and
1765 * this caches recent lookups. The implementation of __early_pfn_to_nid
1766 * treats start/end as pfns.
1768 struct mminit_pfnnid_cache {
1769 unsigned long last_start;
1770 unsigned long last_end;
1774 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1777 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1779 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1780 struct mminit_pfnnid_cache *state)
1782 unsigned long start_pfn, end_pfn;
1785 if (state->last_start <= pfn && pfn < state->last_end)
1786 return state->last_nid;
1788 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1789 if (nid != NUMA_NO_NODE) {
1790 state->last_start = start_pfn;
1791 state->last_end = end_pfn;
1792 state->last_nid = nid;
1798 int __meminit early_pfn_to_nid(unsigned long pfn)
1800 static DEFINE_SPINLOCK(early_pfn_lock);
1803 spin_lock(&early_pfn_lock);
1804 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1806 nid = first_online_node;
1807 spin_unlock(&early_pfn_lock);
1811 #endif /* CONFIG_NUMA */
1813 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1816 if (early_page_uninitialised(pfn))
1818 if (!kmsan_memblock_free_pages(page, order)) {
1819 /* KMSAN will take care of these pages. */
1822 __free_pages_core(page, order);
1826 * Check that the whole (or subset of) a pageblock given by the interval of
1827 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1828 * with the migration of free compaction scanner.
1830 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1832 * It's possible on some configurations to have a setup like node0 node1 node0
1833 * i.e. it's possible that all pages within a zones range of pages do not
1834 * belong to a single zone. We assume that a border between node0 and node1
1835 * can occur within a single pageblock, but not a node0 node1 node0
1836 * interleaving within a single pageblock. It is therefore sufficient to check
1837 * the first and last page of a pageblock and avoid checking each individual
1838 * page in a pageblock.
1840 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1841 unsigned long end_pfn, struct zone *zone)
1843 struct page *start_page;
1844 struct page *end_page;
1846 /* end_pfn is one past the range we are checking */
1849 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1852 start_page = pfn_to_online_page(start_pfn);
1856 if (page_zone(start_page) != zone)
1859 end_page = pfn_to_page(end_pfn);
1861 /* This gives a shorter code than deriving page_zone(end_page) */
1862 if (page_zone_id(start_page) != page_zone_id(end_page))
1868 void set_zone_contiguous(struct zone *zone)
1870 unsigned long block_start_pfn = zone->zone_start_pfn;
1871 unsigned long block_end_pfn;
1873 block_end_pfn = pageblock_end_pfn(block_start_pfn);
1874 for (; block_start_pfn < zone_end_pfn(zone);
1875 block_start_pfn = block_end_pfn,
1876 block_end_pfn += pageblock_nr_pages) {
1878 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1880 if (!__pageblock_pfn_to_page(block_start_pfn,
1881 block_end_pfn, zone))
1886 /* We confirm that there is no hole */
1887 zone->contiguous = true;
1890 void clear_zone_contiguous(struct zone *zone)
1892 zone->contiguous = false;
1895 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1896 static void __init deferred_free_range(unsigned long pfn,
1897 unsigned long nr_pages)
1905 page = pfn_to_page(pfn);
1907 /* Free a large naturally-aligned chunk if possible */
1908 if (nr_pages == pageblock_nr_pages && pageblock_aligned(pfn)) {
1909 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1910 __free_pages_core(page, pageblock_order);
1914 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1915 if (pageblock_aligned(pfn))
1916 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1917 __free_pages_core(page, 0);
1921 /* Completion tracking for deferred_init_memmap() threads */
1922 static atomic_t pgdat_init_n_undone __initdata;
1923 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1925 static inline void __init pgdat_init_report_one_done(void)
1927 if (atomic_dec_and_test(&pgdat_init_n_undone))
1928 complete(&pgdat_init_all_done_comp);
1932 * Returns true if page needs to be initialized or freed to buddy allocator.
1934 * We check if a current large page is valid by only checking the validity
1937 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1939 if (pageblock_aligned(pfn) && !pfn_valid(pfn))
1945 * Free pages to buddy allocator. Try to free aligned pages in
1946 * pageblock_nr_pages sizes.
1948 static void __init deferred_free_pages(unsigned long pfn,
1949 unsigned long end_pfn)
1951 unsigned long nr_free = 0;
1953 for (; pfn < end_pfn; pfn++) {
1954 if (!deferred_pfn_valid(pfn)) {
1955 deferred_free_range(pfn - nr_free, nr_free);
1957 } else if (pageblock_aligned(pfn)) {
1958 deferred_free_range(pfn - nr_free, nr_free);
1964 /* Free the last block of pages to allocator */
1965 deferred_free_range(pfn - nr_free, nr_free);
1969 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1970 * by performing it only once every pageblock_nr_pages.
1971 * Return number of pages initialized.
1973 static unsigned long __init deferred_init_pages(struct zone *zone,
1975 unsigned long end_pfn)
1977 int nid = zone_to_nid(zone);
1978 unsigned long nr_pages = 0;
1979 int zid = zone_idx(zone);
1980 struct page *page = NULL;
1982 for (; pfn < end_pfn; pfn++) {
1983 if (!deferred_pfn_valid(pfn)) {
1986 } else if (!page || pageblock_aligned(pfn)) {
1987 page = pfn_to_page(pfn);
1991 __init_single_page(page, pfn, zid, nid);
1998 * This function is meant to pre-load the iterator for the zone init.
1999 * Specifically it walks through the ranges until we are caught up to the
2000 * first_init_pfn value and exits there. If we never encounter the value we
2001 * return false indicating there are no valid ranges left.
2004 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
2005 unsigned long *spfn, unsigned long *epfn,
2006 unsigned long first_init_pfn)
2011 * Start out by walking through the ranges in this zone that have
2012 * already been initialized. We don't need to do anything with them
2013 * so we just need to flush them out of the system.
2015 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
2016 if (*epfn <= first_init_pfn)
2018 if (*spfn < first_init_pfn)
2019 *spfn = first_init_pfn;
2028 * Initialize and free pages. We do it in two loops: first we initialize
2029 * struct page, then free to buddy allocator, because while we are
2030 * freeing pages we can access pages that are ahead (computing buddy
2031 * page in __free_one_page()).
2033 * In order to try and keep some memory in the cache we have the loop
2034 * broken along max page order boundaries. This way we will not cause
2035 * any issues with the buddy page computation.
2037 static unsigned long __init
2038 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
2039 unsigned long *end_pfn)
2041 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
2042 unsigned long spfn = *start_pfn, epfn = *end_pfn;
2043 unsigned long nr_pages = 0;
2046 /* First we loop through and initialize the page values */
2047 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
2050 if (mo_pfn <= *start_pfn)
2053 t = min(mo_pfn, *end_pfn);
2054 nr_pages += deferred_init_pages(zone, *start_pfn, t);
2056 if (mo_pfn < *end_pfn) {
2057 *start_pfn = mo_pfn;
2062 /* Reset values and now loop through freeing pages as needed */
2065 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2071 t = min(mo_pfn, epfn);
2072 deferred_free_pages(spfn, t);
2082 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2085 unsigned long spfn, epfn;
2086 struct zone *zone = arg;
2089 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2092 * Initialize and free pages in MAX_ORDER sized increments so that we
2093 * can avoid introducing any issues with the buddy allocator.
2095 while (spfn < end_pfn) {
2096 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2101 /* An arch may override for more concurrency. */
2103 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2108 /* Initialise remaining memory on a node */
2109 static int __init deferred_init_memmap(void *data)
2111 pg_data_t *pgdat = data;
2112 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2113 unsigned long spfn = 0, epfn = 0;
2114 unsigned long first_init_pfn, flags;
2115 unsigned long start = jiffies;
2117 int zid, max_threads;
2120 /* Bind memory initialisation thread to a local node if possible */
2121 if (!cpumask_empty(cpumask))
2122 set_cpus_allowed_ptr(current, cpumask);
2124 pgdat_resize_lock(pgdat, &flags);
2125 first_init_pfn = pgdat->first_deferred_pfn;
2126 if (first_init_pfn == ULONG_MAX) {
2127 pgdat_resize_unlock(pgdat, &flags);
2128 pgdat_init_report_one_done();
2132 /* Sanity check boundaries */
2133 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2134 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2135 pgdat->first_deferred_pfn = ULONG_MAX;
2138 * Once we unlock here, the zone cannot be grown anymore, thus if an
2139 * interrupt thread must allocate this early in boot, zone must be
2140 * pre-grown prior to start of deferred page initialization.
2142 pgdat_resize_unlock(pgdat, &flags);
2144 /* Only the highest zone is deferred so find it */
2145 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2146 zone = pgdat->node_zones + zid;
2147 if (first_init_pfn < zone_end_pfn(zone))
2151 /* If the zone is empty somebody else may have cleared out the zone */
2152 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2156 max_threads = deferred_page_init_max_threads(cpumask);
2158 while (spfn < epfn) {
2159 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2160 struct padata_mt_job job = {
2161 .thread_fn = deferred_init_memmap_chunk,
2164 .size = epfn_align - spfn,
2165 .align = PAGES_PER_SECTION,
2166 .min_chunk = PAGES_PER_SECTION,
2167 .max_threads = max_threads,
2170 padata_do_multithreaded(&job);
2171 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2175 /* Sanity check that the next zone really is unpopulated */
2176 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2178 pr_info("node %d deferred pages initialised in %ums\n",
2179 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2181 pgdat_init_report_one_done();
2186 * If this zone has deferred pages, try to grow it by initializing enough
2187 * deferred pages to satisfy the allocation specified by order, rounded up to
2188 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2189 * of SECTION_SIZE bytes by initializing struct pages in increments of
2190 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2192 * Return true when zone was grown, otherwise return false. We return true even
2193 * when we grow less than requested, to let the caller decide if there are
2194 * enough pages to satisfy the allocation.
2196 * Note: We use noinline because this function is needed only during boot, and
2197 * it is called from a __ref function _deferred_grow_zone. This way we are
2198 * making sure that it is not inlined into permanent text section.
2200 static noinline bool __init
2201 deferred_grow_zone(struct zone *zone, unsigned int order)
2203 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2204 pg_data_t *pgdat = zone->zone_pgdat;
2205 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2206 unsigned long spfn, epfn, flags;
2207 unsigned long nr_pages = 0;
2210 /* Only the last zone may have deferred pages */
2211 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2214 pgdat_resize_lock(pgdat, &flags);
2217 * If someone grew this zone while we were waiting for spinlock, return
2218 * true, as there might be enough pages already.
2220 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2221 pgdat_resize_unlock(pgdat, &flags);
2225 /* If the zone is empty somebody else may have cleared out the zone */
2226 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2227 first_deferred_pfn)) {
2228 pgdat->first_deferred_pfn = ULONG_MAX;
2229 pgdat_resize_unlock(pgdat, &flags);
2230 /* Retry only once. */
2231 return first_deferred_pfn != ULONG_MAX;
2235 * Initialize and free pages in MAX_ORDER sized increments so
2236 * that we can avoid introducing any issues with the buddy
2239 while (spfn < epfn) {
2240 /* update our first deferred PFN for this section */
2241 first_deferred_pfn = spfn;
2243 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2244 touch_nmi_watchdog();
2246 /* We should only stop along section boundaries */
2247 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2250 /* If our quota has been met we can stop here */
2251 if (nr_pages >= nr_pages_needed)
2255 pgdat->first_deferred_pfn = spfn;
2256 pgdat_resize_unlock(pgdat, &flags);
2258 return nr_pages > 0;
2262 * deferred_grow_zone() is __init, but it is called from
2263 * get_page_from_freelist() during early boot until deferred_pages permanently
2264 * disables this call. This is why we have refdata wrapper to avoid warning,
2265 * and to ensure that the function body gets unloaded.
2268 _deferred_grow_zone(struct zone *zone, unsigned int order)
2270 return deferred_grow_zone(zone, order);
2273 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2275 void __init page_alloc_init_late(void)
2280 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2282 /* There will be num_node_state(N_MEMORY) threads */
2283 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2284 for_each_node_state(nid, N_MEMORY) {
2285 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2288 /* Block until all are initialised */
2289 wait_for_completion(&pgdat_init_all_done_comp);
2292 * We initialized the rest of the deferred pages. Permanently disable
2293 * on-demand struct page initialization.
2295 static_branch_disable(&deferred_pages);
2297 /* Reinit limits that are based on free pages after the kernel is up */
2298 files_maxfiles_init();
2303 /* Discard memblock private memory */
2306 for_each_node_state(nid, N_MEMORY)
2307 shuffle_free_memory(NODE_DATA(nid));
2309 for_each_populated_zone(zone)
2310 set_zone_contiguous(zone);
2314 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2315 void __init init_cma_reserved_pageblock(struct page *page)
2317 unsigned i = pageblock_nr_pages;
2318 struct page *p = page;
2321 __ClearPageReserved(p);
2322 set_page_count(p, 0);
2325 set_pageblock_migratetype(page, MIGRATE_CMA);
2326 set_page_refcounted(page);
2327 __free_pages(page, pageblock_order);
2329 adjust_managed_page_count(page, pageblock_nr_pages);
2330 page_zone(page)->cma_pages += pageblock_nr_pages;
2335 * The order of subdivision here is critical for the IO subsystem.
2336 * Please do not alter this order without good reasons and regression
2337 * testing. Specifically, as large blocks of memory are subdivided,
2338 * the order in which smaller blocks are delivered depends on the order
2339 * they're subdivided in this function. This is the primary factor
2340 * influencing the order in which pages are delivered to the IO
2341 * subsystem according to empirical testing, and this is also justified
2342 * by considering the behavior of a buddy system containing a single
2343 * large block of memory acted on by a series of small allocations.
2344 * This behavior is a critical factor in sglist merging's success.
2348 static inline void expand(struct zone *zone, struct page *page,
2349 int low, int high, int migratetype)
2351 unsigned long size = 1 << high;
2353 while (high > low) {
2356 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2359 * Mark as guard pages (or page), that will allow to
2360 * merge back to allocator when buddy will be freed.
2361 * Corresponding page table entries will not be touched,
2362 * pages will stay not present in virtual address space
2364 if (set_page_guard(zone, &page[size], high, migratetype))
2367 add_to_free_list(&page[size], zone, high, migratetype);
2368 set_buddy_order(&page[size], high);
2372 static void check_new_page_bad(struct page *page)
2374 if (unlikely(page->flags & __PG_HWPOISON)) {
2375 /* Don't complain about hwpoisoned pages */
2376 page_mapcount_reset(page); /* remove PageBuddy */
2381 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2385 * This page is about to be returned from the page allocator
2387 static inline int check_new_page(struct page *page)
2389 if (likely(page_expected_state(page,
2390 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2393 check_new_page_bad(page);
2397 static bool check_new_pages(struct page *page, unsigned int order)
2400 for (i = 0; i < (1 << order); i++) {
2401 struct page *p = page + i;
2403 if (unlikely(check_new_page(p)))
2410 #ifdef CONFIG_DEBUG_VM
2412 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2413 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2414 * also checked when pcp lists are refilled from the free lists.
2416 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2418 if (debug_pagealloc_enabled_static())
2419 return check_new_pages(page, order);
2424 static inline bool check_new_pcp(struct page *page, unsigned int order)
2426 return check_new_pages(page, order);
2430 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2431 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2432 * enabled, they are also checked when being allocated from the pcp lists.
2434 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2436 return check_new_pages(page, order);
2438 static inline bool check_new_pcp(struct page *page, unsigned int order)
2440 if (debug_pagealloc_enabled_static())
2441 return check_new_pages(page, order);
2445 #endif /* CONFIG_DEBUG_VM */
2447 static inline bool should_skip_kasan_unpoison(gfp_t flags)
2449 /* Don't skip if a software KASAN mode is enabled. */
2450 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
2451 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
2454 /* Skip, if hardware tag-based KASAN is not enabled. */
2455 if (!kasan_hw_tags_enabled())
2459 * With hardware tag-based KASAN enabled, skip if this has been
2460 * requested via __GFP_SKIP_KASAN_UNPOISON.
2462 return flags & __GFP_SKIP_KASAN_UNPOISON;
2465 static inline bool should_skip_init(gfp_t flags)
2467 /* Don't skip, if hardware tag-based KASAN is not enabled. */
2468 if (!kasan_hw_tags_enabled())
2471 /* For hardware tag-based KASAN, skip if requested. */
2472 return (flags & __GFP_SKIP_ZERO);
2475 inline void post_alloc_hook(struct page *page, unsigned int order,
2478 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
2479 !should_skip_init(gfp_flags);
2480 bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
2483 set_page_private(page, 0);
2484 set_page_refcounted(page);
2486 arch_alloc_page(page, order);
2487 debug_pagealloc_map_pages(page, 1 << order);
2490 * Page unpoisoning must happen before memory initialization.
2491 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2492 * allocations and the page unpoisoning code will complain.
2494 kernel_unpoison_pages(page, 1 << order);
2497 * As memory initialization might be integrated into KASAN,
2498 * KASAN unpoisoning and memory initializion code must be
2499 * kept together to avoid discrepancies in behavior.
2503 * If memory tags should be zeroed (which happens only when memory
2504 * should be initialized as well).
2507 /* Initialize both memory and tags. */
2508 for (i = 0; i != 1 << order; ++i)
2509 tag_clear_highpage(page + i);
2511 /* Note that memory is already initialized by the loop above. */
2514 if (!should_skip_kasan_unpoison(gfp_flags)) {
2515 /* Unpoison shadow memory or set memory tags. */
2516 kasan_unpoison_pages(page, order, init);
2518 /* Note that memory is already initialized by KASAN. */
2519 if (kasan_has_integrated_init())
2522 /* Ensure page_address() dereferencing does not fault. */
2523 for (i = 0; i != 1 << order; ++i)
2524 page_kasan_tag_reset(page + i);
2526 /* If memory is still not initialized, do it now. */
2528 kernel_init_pages(page, 1 << order);
2529 /* Propagate __GFP_SKIP_KASAN_POISON to page flags. */
2530 if (kasan_hw_tags_enabled() && (gfp_flags & __GFP_SKIP_KASAN_POISON))
2531 SetPageSkipKASanPoison(page);
2533 set_page_owner(page, order, gfp_flags);
2534 page_table_check_alloc(page, order);
2537 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2538 unsigned int alloc_flags)
2540 post_alloc_hook(page, order, gfp_flags);
2542 if (order && (gfp_flags & __GFP_COMP))
2543 prep_compound_page(page, order);
2546 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2547 * allocate the page. The expectation is that the caller is taking
2548 * steps that will free more memory. The caller should avoid the page
2549 * being used for !PFMEMALLOC purposes.
2551 if (alloc_flags & ALLOC_NO_WATERMARKS)
2552 set_page_pfmemalloc(page);
2554 clear_page_pfmemalloc(page);
2558 * Go through the free lists for the given migratetype and remove
2559 * the smallest available page from the freelists
2561 static __always_inline
2562 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2565 unsigned int current_order;
2566 struct free_area *area;
2569 /* Find a page of the appropriate size in the preferred list */
2570 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2571 area = &(zone->free_area[current_order]);
2572 page = get_page_from_free_area(area, migratetype);
2575 del_page_from_free_list(page, zone, current_order);
2576 expand(zone, page, order, current_order, migratetype);
2577 set_pcppage_migratetype(page, migratetype);
2578 trace_mm_page_alloc_zone_locked(page, order, migratetype,
2579 pcp_allowed_order(order) &&
2580 migratetype < MIGRATE_PCPTYPES);
2589 * This array describes the order lists are fallen back to when
2590 * the free lists for the desirable migrate type are depleted
2592 * The other migratetypes do not have fallbacks.
2594 static int fallbacks[MIGRATE_TYPES][3] = {
2595 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2596 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2597 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2601 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2604 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2607 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2608 unsigned int order) { return NULL; }
2612 * Move the free pages in a range to the freelist tail of the requested type.
2613 * Note that start_page and end_pages are not aligned on a pageblock
2614 * boundary. If alignment is required, use move_freepages_block()
2616 static int move_freepages(struct zone *zone,
2617 unsigned long start_pfn, unsigned long end_pfn,
2618 int migratetype, int *num_movable)
2623 int pages_moved = 0;
2625 for (pfn = start_pfn; pfn <= end_pfn;) {
2626 page = pfn_to_page(pfn);
2627 if (!PageBuddy(page)) {
2629 * We assume that pages that could be isolated for
2630 * migration are movable. But we don't actually try
2631 * isolating, as that would be expensive.
2634 (PageLRU(page) || __PageMovable(page)))
2640 /* Make sure we are not inadvertently changing nodes */
2641 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2642 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2644 order = buddy_order(page);
2645 move_to_free_list(page, zone, order, migratetype);
2647 pages_moved += 1 << order;
2653 int move_freepages_block(struct zone *zone, struct page *page,
2654 int migratetype, int *num_movable)
2656 unsigned long start_pfn, end_pfn, pfn;
2661 pfn = page_to_pfn(page);
2662 start_pfn = pageblock_start_pfn(pfn);
2663 end_pfn = pageblock_end_pfn(pfn) - 1;
2665 /* Do not cross zone boundaries */
2666 if (!zone_spans_pfn(zone, start_pfn))
2668 if (!zone_spans_pfn(zone, end_pfn))
2671 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2675 static void change_pageblock_range(struct page *pageblock_page,
2676 int start_order, int migratetype)
2678 int nr_pageblocks = 1 << (start_order - pageblock_order);
2680 while (nr_pageblocks--) {
2681 set_pageblock_migratetype(pageblock_page, migratetype);
2682 pageblock_page += pageblock_nr_pages;
2687 * When we are falling back to another migratetype during allocation, try to
2688 * steal extra free pages from the same pageblocks to satisfy further
2689 * allocations, instead of polluting multiple pageblocks.
2691 * If we are stealing a relatively large buddy page, it is likely there will
2692 * be more free pages in the pageblock, so try to steal them all. For
2693 * reclaimable and unmovable allocations, we steal regardless of page size,
2694 * as fragmentation caused by those allocations polluting movable pageblocks
2695 * is worse than movable allocations stealing from unmovable and reclaimable
2698 static bool can_steal_fallback(unsigned int order, int start_mt)
2701 * Leaving this order check is intended, although there is
2702 * relaxed order check in next check. The reason is that
2703 * we can actually steal whole pageblock if this condition met,
2704 * but, below check doesn't guarantee it and that is just heuristic
2705 * so could be changed anytime.
2707 if (order >= pageblock_order)
2710 if (order >= pageblock_order / 2 ||
2711 start_mt == MIGRATE_RECLAIMABLE ||
2712 start_mt == MIGRATE_UNMOVABLE ||
2713 page_group_by_mobility_disabled)
2719 static inline bool boost_watermark(struct zone *zone)
2721 unsigned long max_boost;
2723 if (!watermark_boost_factor)
2726 * Don't bother in zones that are unlikely to produce results.
2727 * On small machines, including kdump capture kernels running
2728 * in a small area, boosting the watermark can cause an out of
2729 * memory situation immediately.
2731 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2734 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2735 watermark_boost_factor, 10000);
2738 * high watermark may be uninitialised if fragmentation occurs
2739 * very early in boot so do not boost. We do not fall
2740 * through and boost by pageblock_nr_pages as failing
2741 * allocations that early means that reclaim is not going
2742 * to help and it may even be impossible to reclaim the
2743 * boosted watermark resulting in a hang.
2748 max_boost = max(pageblock_nr_pages, max_boost);
2750 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2757 * This function implements actual steal behaviour. If order is large enough,
2758 * we can steal whole pageblock. If not, we first move freepages in this
2759 * pageblock to our migratetype and determine how many already-allocated pages
2760 * are there in the pageblock with a compatible migratetype. If at least half
2761 * of pages are free or compatible, we can change migratetype of the pageblock
2762 * itself, so pages freed in the future will be put on the correct free list.
2764 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2765 unsigned int alloc_flags, int start_type, bool whole_block)
2767 unsigned int current_order = buddy_order(page);
2768 int free_pages, movable_pages, alike_pages;
2771 old_block_type = get_pageblock_migratetype(page);
2774 * This can happen due to races and we want to prevent broken
2775 * highatomic accounting.
2777 if (is_migrate_highatomic(old_block_type))
2780 /* Take ownership for orders >= pageblock_order */
2781 if (current_order >= pageblock_order) {
2782 change_pageblock_range(page, current_order, start_type);
2787 * Boost watermarks to increase reclaim pressure to reduce the
2788 * likelihood of future fallbacks. Wake kswapd now as the node
2789 * may be balanced overall and kswapd will not wake naturally.
2791 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2792 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2794 /* We are not allowed to try stealing from the whole block */
2798 free_pages = move_freepages_block(zone, page, start_type,
2801 * Determine how many pages are compatible with our allocation.
2802 * For movable allocation, it's the number of movable pages which
2803 * we just obtained. For other types it's a bit more tricky.
2805 if (start_type == MIGRATE_MOVABLE) {
2806 alike_pages = movable_pages;
2809 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2810 * to MOVABLE pageblock, consider all non-movable pages as
2811 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2812 * vice versa, be conservative since we can't distinguish the
2813 * exact migratetype of non-movable pages.
2815 if (old_block_type == MIGRATE_MOVABLE)
2816 alike_pages = pageblock_nr_pages
2817 - (free_pages + movable_pages);
2822 /* moving whole block can fail due to zone boundary conditions */
2827 * If a sufficient number of pages in the block are either free or of
2828 * comparable migratability as our allocation, claim the whole block.
2830 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2831 page_group_by_mobility_disabled)
2832 set_pageblock_migratetype(page, start_type);
2837 move_to_free_list(page, zone, current_order, start_type);
2841 * Check whether there is a suitable fallback freepage with requested order.
2842 * If only_stealable is true, this function returns fallback_mt only if
2843 * we can steal other freepages all together. This would help to reduce
2844 * fragmentation due to mixed migratetype pages in one pageblock.
2846 int find_suitable_fallback(struct free_area *area, unsigned int order,
2847 int migratetype, bool only_stealable, bool *can_steal)
2852 if (area->nr_free == 0)
2857 fallback_mt = fallbacks[migratetype][i];
2858 if (fallback_mt == MIGRATE_TYPES)
2861 if (free_area_empty(area, fallback_mt))
2864 if (can_steal_fallback(order, migratetype))
2867 if (!only_stealable)
2878 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2879 * there are no empty page blocks that contain a page with a suitable order
2881 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2882 unsigned int alloc_order)
2885 unsigned long max_managed, flags;
2888 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2889 * Check is race-prone but harmless.
2891 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2892 if (zone->nr_reserved_highatomic >= max_managed)
2895 spin_lock_irqsave(&zone->lock, flags);
2897 /* Recheck the nr_reserved_highatomic limit under the lock */
2898 if (zone->nr_reserved_highatomic >= max_managed)
2902 mt = get_pageblock_migratetype(page);
2903 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2904 if (migratetype_is_mergeable(mt)) {
2905 zone->nr_reserved_highatomic += pageblock_nr_pages;
2906 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2907 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2911 spin_unlock_irqrestore(&zone->lock, flags);
2915 * Used when an allocation is about to fail under memory pressure. This
2916 * potentially hurts the reliability of high-order allocations when under
2917 * intense memory pressure but failed atomic allocations should be easier
2918 * to recover from than an OOM.
2920 * If @force is true, try to unreserve a pageblock even though highatomic
2921 * pageblock is exhausted.
2923 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2926 struct zonelist *zonelist = ac->zonelist;
2927 unsigned long flags;
2934 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2937 * Preserve at least one pageblock unless memory pressure
2940 if (!force && zone->nr_reserved_highatomic <=
2944 spin_lock_irqsave(&zone->lock, flags);
2945 for (order = 0; order < MAX_ORDER; order++) {
2946 struct free_area *area = &(zone->free_area[order]);
2948 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2953 * In page freeing path, migratetype change is racy so
2954 * we can counter several free pages in a pageblock
2955 * in this loop although we changed the pageblock type
2956 * from highatomic to ac->migratetype. So we should
2957 * adjust the count once.
2959 if (is_migrate_highatomic_page(page)) {
2961 * It should never happen but changes to
2962 * locking could inadvertently allow a per-cpu
2963 * drain to add pages to MIGRATE_HIGHATOMIC
2964 * while unreserving so be safe and watch for
2967 zone->nr_reserved_highatomic -= min(
2969 zone->nr_reserved_highatomic);
2973 * Convert to ac->migratetype and avoid the normal
2974 * pageblock stealing heuristics. Minimally, the caller
2975 * is doing the work and needs the pages. More
2976 * importantly, if the block was always converted to
2977 * MIGRATE_UNMOVABLE or another type then the number
2978 * of pageblocks that cannot be completely freed
2981 set_pageblock_migratetype(page, ac->migratetype);
2982 ret = move_freepages_block(zone, page, ac->migratetype,
2985 spin_unlock_irqrestore(&zone->lock, flags);
2989 spin_unlock_irqrestore(&zone->lock, flags);
2996 * Try finding a free buddy page on the fallback list and put it on the free
2997 * list of requested migratetype, possibly along with other pages from the same
2998 * block, depending on fragmentation avoidance heuristics. Returns true if
2999 * fallback was found so that __rmqueue_smallest() can grab it.
3001 * The use of signed ints for order and current_order is a deliberate
3002 * deviation from the rest of this file, to make the for loop
3003 * condition simpler.
3005 static __always_inline bool
3006 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
3007 unsigned int alloc_flags)
3009 struct free_area *area;
3011 int min_order = order;
3017 * Do not steal pages from freelists belonging to other pageblocks
3018 * i.e. orders < pageblock_order. If there are no local zones free,
3019 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
3021 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
3022 min_order = pageblock_order;
3025 * Find the largest available free page in the other list. This roughly
3026 * approximates finding the pageblock with the most free pages, which
3027 * would be too costly to do exactly.
3029 for (current_order = MAX_ORDER - 1; current_order >= min_order;
3031 area = &(zone->free_area[current_order]);
3032 fallback_mt = find_suitable_fallback(area, current_order,
3033 start_migratetype, false, &can_steal);
3034 if (fallback_mt == -1)
3038 * We cannot steal all free pages from the pageblock and the
3039 * requested migratetype is movable. In that case it's better to
3040 * steal and split the smallest available page instead of the
3041 * largest available page, because even if the next movable
3042 * allocation falls back into a different pageblock than this
3043 * one, it won't cause permanent fragmentation.
3045 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
3046 && current_order > order)
3055 for (current_order = order; current_order < MAX_ORDER;
3057 area = &(zone->free_area[current_order]);
3058 fallback_mt = find_suitable_fallback(area, current_order,
3059 start_migratetype, false, &can_steal);
3060 if (fallback_mt != -1)
3065 * This should not happen - we already found a suitable fallback
3066 * when looking for the largest page.
3068 VM_BUG_ON(current_order == MAX_ORDER);
3071 page = get_page_from_free_area(area, fallback_mt);
3073 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
3076 trace_mm_page_alloc_extfrag(page, order, current_order,
3077 start_migratetype, fallback_mt);
3084 * Do the hard work of removing an element from the buddy allocator.
3085 * Call me with the zone->lock already held.
3087 static __always_inline struct page *
3088 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
3089 unsigned int alloc_flags)
3093 if (IS_ENABLED(CONFIG_CMA)) {
3095 * Balance movable allocations between regular and CMA areas by
3096 * allocating from CMA when over half of the zone's free memory
3097 * is in the CMA area.
3099 if (alloc_flags & ALLOC_CMA &&
3100 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3101 zone_page_state(zone, NR_FREE_PAGES) / 2) {
3102 page = __rmqueue_cma_fallback(zone, order);
3108 page = __rmqueue_smallest(zone, order, migratetype);
3109 if (unlikely(!page)) {
3110 if (alloc_flags & ALLOC_CMA)
3111 page = __rmqueue_cma_fallback(zone, order);
3113 if (!page && __rmqueue_fallback(zone, order, migratetype,
3121 * Obtain a specified number of elements from the buddy allocator, all under
3122 * a single hold of the lock, for efficiency. Add them to the supplied list.
3123 * Returns the number of new pages which were placed at *list.
3125 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3126 unsigned long count, struct list_head *list,
3127 int migratetype, unsigned int alloc_flags)
3129 int i, allocated = 0;
3131 /* Caller must hold IRQ-safe pcp->lock so IRQs are disabled. */
3132 spin_lock(&zone->lock);
3133 for (i = 0; i < count; ++i) {
3134 struct page *page = __rmqueue(zone, order, migratetype,
3136 if (unlikely(page == NULL))
3139 if (unlikely(check_pcp_refill(page, order)))
3143 * Split buddy pages returned by expand() are received here in
3144 * physical page order. The page is added to the tail of
3145 * caller's list. From the callers perspective, the linked list
3146 * is ordered by page number under some conditions. This is
3147 * useful for IO devices that can forward direction from the
3148 * head, thus also in the physical page order. This is useful
3149 * for IO devices that can merge IO requests if the physical
3150 * pages are ordered properly.
3152 list_add_tail(&page->pcp_list, list);
3154 if (is_migrate_cma(get_pcppage_migratetype(page)))
3155 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3160 * i pages were removed from the buddy list even if some leak due
3161 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3162 * on i. Do not confuse with 'allocated' which is the number of
3163 * pages added to the pcp list.
3165 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3166 spin_unlock(&zone->lock);
3172 * Called from the vmstat counter updater to drain pagesets of this
3173 * currently executing processor on remote nodes after they have
3176 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3178 int to_drain, batch;
3180 batch = READ_ONCE(pcp->batch);
3181 to_drain = min(pcp->count, batch);
3183 unsigned long flags;
3186 * free_pcppages_bulk expects IRQs disabled for zone->lock
3187 * so even though pcp->lock is not intended to be IRQ-safe,
3188 * it's needed in this context.
3190 spin_lock_irqsave(&pcp->lock, flags);
3191 free_pcppages_bulk(zone, to_drain, pcp, 0);
3192 spin_unlock_irqrestore(&pcp->lock, flags);
3198 * Drain pcplists of the indicated processor and zone.
3200 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3202 struct per_cpu_pages *pcp;
3204 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3206 unsigned long flags;
3208 /* See drain_zone_pages on why this is disabling IRQs */
3209 spin_lock_irqsave(&pcp->lock, flags);
3210 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3211 spin_unlock_irqrestore(&pcp->lock, flags);
3216 * Drain pcplists of all zones on the indicated processor.
3218 static void drain_pages(unsigned int cpu)
3222 for_each_populated_zone(zone) {
3223 drain_pages_zone(cpu, zone);
3228 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3230 void drain_local_pages(struct zone *zone)
3232 int cpu = smp_processor_id();
3235 drain_pages_zone(cpu, zone);
3241 * The implementation of drain_all_pages(), exposing an extra parameter to
3242 * drain on all cpus.
3244 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3245 * not empty. The check for non-emptiness can however race with a free to
3246 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3247 * that need the guarantee that every CPU has drained can disable the
3248 * optimizing racy check.
3250 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3255 * Allocate in the BSS so we won't require allocation in
3256 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3258 static cpumask_t cpus_with_pcps;
3261 * Do not drain if one is already in progress unless it's specific to
3262 * a zone. Such callers are primarily CMA and memory hotplug and need
3263 * the drain to be complete when the call returns.
3265 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3268 mutex_lock(&pcpu_drain_mutex);
3272 * We don't care about racing with CPU hotplug event
3273 * as offline notification will cause the notified
3274 * cpu to drain that CPU pcps and on_each_cpu_mask
3275 * disables preemption as part of its processing
3277 for_each_online_cpu(cpu) {
3278 struct per_cpu_pages *pcp;
3280 bool has_pcps = false;
3282 if (force_all_cpus) {
3284 * The pcp.count check is racy, some callers need a
3285 * guarantee that no cpu is missed.
3289 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3293 for_each_populated_zone(z) {
3294 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3303 cpumask_set_cpu(cpu, &cpus_with_pcps);
3305 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3308 for_each_cpu(cpu, &cpus_with_pcps) {
3310 drain_pages_zone(cpu, zone);
3315 mutex_unlock(&pcpu_drain_mutex);
3319 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3321 * When zone parameter is non-NULL, spill just the single zone's pages.
3323 void drain_all_pages(struct zone *zone)
3325 __drain_all_pages(zone, false);
3328 #ifdef CONFIG_HIBERNATION
3331 * Touch the watchdog for every WD_PAGE_COUNT pages.
3333 #define WD_PAGE_COUNT (128*1024)
3335 void mark_free_pages(struct zone *zone)
3337 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3338 unsigned long flags;
3339 unsigned int order, t;
3342 if (zone_is_empty(zone))
3345 spin_lock_irqsave(&zone->lock, flags);
3347 max_zone_pfn = zone_end_pfn(zone);
3348 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3349 if (pfn_valid(pfn)) {
3350 page = pfn_to_page(pfn);
3352 if (!--page_count) {
3353 touch_nmi_watchdog();
3354 page_count = WD_PAGE_COUNT;
3357 if (page_zone(page) != zone)
3360 if (!swsusp_page_is_forbidden(page))
3361 swsusp_unset_page_free(page);
3364 for_each_migratetype_order(order, t) {
3365 list_for_each_entry(page,
3366 &zone->free_area[order].free_list[t], buddy_list) {
3369 pfn = page_to_pfn(page);
3370 for (i = 0; i < (1UL << order); i++) {
3371 if (!--page_count) {
3372 touch_nmi_watchdog();
3373 page_count = WD_PAGE_COUNT;
3375 swsusp_set_page_free(pfn_to_page(pfn + i));
3379 spin_unlock_irqrestore(&zone->lock, flags);
3381 #endif /* CONFIG_PM */
3383 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3388 if (!free_pcp_prepare(page, order))
3391 migratetype = get_pfnblock_migratetype(page, pfn);
3392 set_pcppage_migratetype(page, migratetype);
3396 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3399 int min_nr_free, max_nr_free;
3401 /* Free everything if batch freeing high-order pages. */
3402 if (unlikely(free_high))
3405 /* Check for PCP disabled or boot pageset */
3406 if (unlikely(high < batch))
3409 /* Leave at least pcp->batch pages on the list */
3410 min_nr_free = batch;
3411 max_nr_free = high - batch;
3414 * Double the number of pages freed each time there is subsequent
3415 * freeing of pages without any allocation.
3417 batch <<= pcp->free_factor;
3418 if (batch < max_nr_free)
3420 batch = clamp(batch, min_nr_free, max_nr_free);
3425 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3428 int high = READ_ONCE(pcp->high);
3430 if (unlikely(!high || free_high))
3433 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3437 * If reclaim is active, limit the number of pages that can be
3438 * stored on pcp lists
3440 return min(READ_ONCE(pcp->batch) << 2, high);
3443 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
3444 struct page *page, int migratetype,
3451 __count_vm_events(PGFREE, 1 << order);
3452 pindex = order_to_pindex(migratetype, order);
3453 list_add(&page->pcp_list, &pcp->lists[pindex]);
3454 pcp->count += 1 << order;
3457 * As high-order pages other than THP's stored on PCP can contribute
3458 * to fragmentation, limit the number stored when PCP is heavily
3459 * freeing without allocation. The remainder after bulk freeing
3460 * stops will be drained from vmstat refresh context.
3462 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3464 high = nr_pcp_high(pcp, zone, free_high);
3465 if (pcp->count >= high) {
3466 int batch = READ_ONCE(pcp->batch);
3468 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3475 void free_unref_page(struct page *page, unsigned int order)
3477 unsigned long flags;
3478 unsigned long __maybe_unused UP_flags;
3479 struct per_cpu_pages *pcp;
3481 unsigned long pfn = page_to_pfn(page);
3484 if (!free_unref_page_prepare(page, pfn, order))
3488 * We only track unmovable, reclaimable and movable on pcp lists.
3489 * Place ISOLATE pages on the isolated list because they are being
3490 * offlined but treat HIGHATOMIC as movable pages so we can get those
3491 * areas back if necessary. Otherwise, we may have to free
3492 * excessively into the page allocator
3494 migratetype = get_pcppage_migratetype(page);
3495 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3496 if (unlikely(is_migrate_isolate(migratetype))) {
3497 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3500 migratetype = MIGRATE_MOVABLE;
3503 zone = page_zone(page);
3504 pcp_trylock_prepare(UP_flags);
3505 pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
3507 free_unref_page_commit(zone, pcp, page, migratetype, order);
3508 pcp_spin_unlock_irqrestore(pcp, flags);
3510 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
3512 pcp_trylock_finish(UP_flags);
3516 * Free a list of 0-order pages
3518 void free_unref_page_list(struct list_head *list)
3520 struct page *page, *next;
3521 struct per_cpu_pages *pcp = NULL;
3522 struct zone *locked_zone = NULL;
3523 unsigned long flags;
3524 int batch_count = 0;
3527 /* Prepare pages for freeing */
3528 list_for_each_entry_safe(page, next, list, lru) {
3529 unsigned long pfn = page_to_pfn(page);
3530 if (!free_unref_page_prepare(page, pfn, 0)) {
3531 list_del(&page->lru);
3536 * Free isolated pages directly to the allocator, see
3537 * comment in free_unref_page.
3539 migratetype = get_pcppage_migratetype(page);
3540 if (unlikely(is_migrate_isolate(migratetype))) {
3541 list_del(&page->lru);
3542 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3547 list_for_each_entry_safe(page, next, list, lru) {
3548 struct zone *zone = page_zone(page);
3550 /* Different zone, different pcp lock. */
3551 if (zone != locked_zone) {
3553 pcp_spin_unlock_irqrestore(pcp, flags);
3556 pcp = pcp_spin_lock_irqsave(locked_zone->per_cpu_pageset, flags);
3560 * Non-isolated types over MIGRATE_PCPTYPES get added
3561 * to the MIGRATE_MOVABLE pcp list.
3563 migratetype = get_pcppage_migratetype(page);
3564 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3565 migratetype = MIGRATE_MOVABLE;
3567 trace_mm_page_free_batched(page);
3568 free_unref_page_commit(zone, pcp, page, migratetype, 0);
3571 * Guard against excessive IRQ disabled times when we get
3572 * a large list of pages to free.
3574 if (++batch_count == SWAP_CLUSTER_MAX) {
3575 pcp_spin_unlock_irqrestore(pcp, flags);
3577 pcp = pcp_spin_lock_irqsave(locked_zone->per_cpu_pageset, flags);
3582 pcp_spin_unlock_irqrestore(pcp, flags);
3586 * split_page takes a non-compound higher-order page, and splits it into
3587 * n (1<<order) sub-pages: page[0..n]
3588 * Each sub-page must be freed individually.
3590 * Note: this is probably too low level an operation for use in drivers.
3591 * Please consult with lkml before using this in your driver.
3593 void split_page(struct page *page, unsigned int order)
3597 VM_BUG_ON_PAGE(PageCompound(page), page);
3598 VM_BUG_ON_PAGE(!page_count(page), page);
3600 for (i = 1; i < (1 << order); i++)
3601 set_page_refcounted(page + i);
3602 split_page_owner(page, 1 << order);
3603 split_page_memcg(page, 1 << order);
3605 EXPORT_SYMBOL_GPL(split_page);
3607 int __isolate_free_page(struct page *page, unsigned int order)
3609 struct zone *zone = page_zone(page);
3610 int mt = get_pageblock_migratetype(page);
3612 if (!is_migrate_isolate(mt)) {
3613 unsigned long watermark;
3615 * Obey watermarks as if the page was being allocated. We can
3616 * emulate a high-order watermark check with a raised order-0
3617 * watermark, because we already know our high-order page
3620 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3621 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3624 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3627 del_page_from_free_list(page, zone, order);
3630 * Set the pageblock if the isolated page is at least half of a
3633 if (order >= pageblock_order - 1) {
3634 struct page *endpage = page + (1 << order) - 1;
3635 for (; page < endpage; page += pageblock_nr_pages) {
3636 int mt = get_pageblock_migratetype(page);
3638 * Only change normal pageblocks (i.e., they can merge
3641 if (migratetype_is_mergeable(mt))
3642 set_pageblock_migratetype(page,
3647 return 1UL << order;
3651 * __putback_isolated_page - Return a now-isolated page back where we got it
3652 * @page: Page that was isolated
3653 * @order: Order of the isolated page
3654 * @mt: The page's pageblock's migratetype
3656 * This function is meant to return a page pulled from the free lists via
3657 * __isolate_free_page back to the free lists they were pulled from.
3659 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3661 struct zone *zone = page_zone(page);
3663 /* zone lock should be held when this function is called */
3664 lockdep_assert_held(&zone->lock);
3666 /* Return isolated page to tail of freelist. */
3667 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3668 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3672 * Update NUMA hit/miss statistics
3674 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3678 enum numa_stat_item local_stat = NUMA_LOCAL;
3680 /* skip numa counters update if numa stats is disabled */
3681 if (!static_branch_likely(&vm_numa_stat_key))
3684 if (zone_to_nid(z) != numa_node_id())
3685 local_stat = NUMA_OTHER;
3687 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3688 __count_numa_events(z, NUMA_HIT, nr_account);
3690 __count_numa_events(z, NUMA_MISS, nr_account);
3691 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3693 __count_numa_events(z, local_stat, nr_account);
3697 static __always_inline
3698 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
3699 unsigned int order, unsigned int alloc_flags,
3703 unsigned long flags;
3707 spin_lock_irqsave(&zone->lock, flags);
3709 * order-0 request can reach here when the pcplist is skipped
3710 * due to non-CMA allocation context. HIGHATOMIC area is
3711 * reserved for high-order atomic allocation, so order-0
3712 * request should skip it.
3714 if (order > 0 && alloc_flags & ALLOC_HARDER)
3715 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3717 page = __rmqueue(zone, order, migratetype, alloc_flags);
3719 spin_unlock_irqrestore(&zone->lock, flags);
3723 __mod_zone_freepage_state(zone, -(1 << order),
3724 get_pcppage_migratetype(page));
3725 spin_unlock_irqrestore(&zone->lock, flags);
3726 } while (check_new_pages(page, order));
3728 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3729 zone_statistics(preferred_zone, zone, 1);
3734 /* Remove page from the per-cpu list, caller must protect the list */
3736 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3738 unsigned int alloc_flags,
3739 struct per_cpu_pages *pcp,
3740 struct list_head *list)
3745 if (list_empty(list)) {
3746 int batch = READ_ONCE(pcp->batch);
3750 * Scale batch relative to order if batch implies
3751 * free pages can be stored on the PCP. Batch can
3752 * be 1 for small zones or for boot pagesets which
3753 * should never store free pages as the pages may
3754 * belong to arbitrary zones.
3757 batch = max(batch >> order, 2);
3758 alloced = rmqueue_bulk(zone, order,
3760 migratetype, alloc_flags);
3762 pcp->count += alloced << order;
3763 if (unlikely(list_empty(list)))
3767 page = list_first_entry(list, struct page, pcp_list);
3768 list_del(&page->pcp_list);
3769 pcp->count -= 1 << order;
3770 } while (check_new_pcp(page, order));
3775 /* Lock and remove page from the per-cpu list */
3776 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3777 struct zone *zone, unsigned int order,
3778 int migratetype, unsigned int alloc_flags)
3780 struct per_cpu_pages *pcp;
3781 struct list_head *list;
3783 unsigned long flags;
3784 unsigned long __maybe_unused UP_flags;
3787 * spin_trylock may fail due to a parallel drain. In the future, the
3788 * trylock will also protect against IRQ reentrancy.
3790 pcp_trylock_prepare(UP_flags);
3791 pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
3793 pcp_trylock_finish(UP_flags);
3798 * On allocation, reduce the number of pages that are batch freed.
3799 * See nr_pcp_free() where free_factor is increased for subsequent
3802 pcp->free_factor >>= 1;
3803 list = &pcp->lists[order_to_pindex(migratetype, order)];
3804 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3805 pcp_spin_unlock_irqrestore(pcp, flags);
3806 pcp_trylock_finish(UP_flags);
3808 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3809 zone_statistics(preferred_zone, zone, 1);
3815 * Allocate a page from the given zone.
3816 * Use pcplists for THP or "cheap" high-order allocations.
3820 * Do not instrument rmqueue() with KMSAN. This function may call
3821 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
3822 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3823 * may call rmqueue() again, which will result in a deadlock.
3825 __no_sanitize_memory
3827 struct page *rmqueue(struct zone *preferred_zone,
3828 struct zone *zone, unsigned int order,
3829 gfp_t gfp_flags, unsigned int alloc_flags,
3835 * We most definitely don't want callers attempting to
3836 * allocate greater than order-1 page units with __GFP_NOFAIL.
3838 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3840 if (likely(pcp_allowed_order(order))) {
3842 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3843 * we need to skip it when CMA area isn't allowed.
3845 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3846 migratetype != MIGRATE_MOVABLE) {
3847 page = rmqueue_pcplist(preferred_zone, zone, order,
3848 migratetype, alloc_flags);
3854 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3858 /* Separate test+clear to avoid unnecessary atomics */
3859 if (unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3860 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3861 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3864 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3868 #ifdef CONFIG_FAIL_PAGE_ALLOC
3871 struct fault_attr attr;
3873 bool ignore_gfp_highmem;
3874 bool ignore_gfp_reclaim;
3876 } fail_page_alloc = {
3877 .attr = FAULT_ATTR_INITIALIZER,
3878 .ignore_gfp_reclaim = true,
3879 .ignore_gfp_highmem = true,
3883 static int __init setup_fail_page_alloc(char *str)
3885 return setup_fault_attr(&fail_page_alloc.attr, str);
3887 __setup("fail_page_alloc=", setup_fail_page_alloc);
3889 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3893 if (order < fail_page_alloc.min_order)
3895 if (gfp_mask & __GFP_NOFAIL)
3897 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3899 if (fail_page_alloc.ignore_gfp_reclaim &&
3900 (gfp_mask & __GFP_DIRECT_RECLAIM))
3903 /* See comment in __should_failslab() */
3904 if (gfp_mask & __GFP_NOWARN)
3905 flags |= FAULT_NOWARN;
3907 return should_fail_ex(&fail_page_alloc.attr, 1 << order, flags);
3910 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3912 static int __init fail_page_alloc_debugfs(void)
3914 umode_t mode = S_IFREG | 0600;
3917 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3918 &fail_page_alloc.attr);
3920 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3921 &fail_page_alloc.ignore_gfp_reclaim);
3922 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3923 &fail_page_alloc.ignore_gfp_highmem);
3924 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3929 late_initcall(fail_page_alloc_debugfs);
3931 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3933 #else /* CONFIG_FAIL_PAGE_ALLOC */
3935 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3940 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3942 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3944 return __should_fail_alloc_page(gfp_mask, order);
3946 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3948 static inline long __zone_watermark_unusable_free(struct zone *z,
3949 unsigned int order, unsigned int alloc_flags)
3951 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3952 long unusable_free = (1 << order) - 1;
3955 * If the caller does not have rights to ALLOC_HARDER then subtract
3956 * the high-atomic reserves. This will over-estimate the size of the
3957 * atomic reserve but it avoids a search.
3959 if (likely(!alloc_harder))
3960 unusable_free += z->nr_reserved_highatomic;
3963 /* If allocation can't use CMA areas don't use free CMA pages */
3964 if (!(alloc_flags & ALLOC_CMA))
3965 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3968 return unusable_free;
3972 * Return true if free base pages are above 'mark'. For high-order checks it
3973 * will return true of the order-0 watermark is reached and there is at least
3974 * one free page of a suitable size. Checking now avoids taking the zone lock
3975 * to check in the allocation paths if no pages are free.
3977 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3978 int highest_zoneidx, unsigned int alloc_flags,
3983 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3985 /* free_pages may go negative - that's OK */
3986 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3988 if (alloc_flags & ALLOC_HIGH)
3991 if (unlikely(alloc_harder)) {
3993 * OOM victims can try even harder than normal ALLOC_HARDER
3994 * users on the grounds that it's definitely going to be in
3995 * the exit path shortly and free memory. Any allocation it
3996 * makes during the free path will be small and short-lived.
3998 if (alloc_flags & ALLOC_OOM)
4005 * Check watermarks for an order-0 allocation request. If these
4006 * are not met, then a high-order request also cannot go ahead
4007 * even if a suitable page happened to be free.
4009 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
4012 /* If this is an order-0 request then the watermark is fine */
4016 /* For a high-order request, check at least one suitable page is free */
4017 for (o = order; o < MAX_ORDER; o++) {
4018 struct free_area *area = &z->free_area[o];
4024 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
4025 if (!free_area_empty(area, mt))
4030 if ((alloc_flags & ALLOC_CMA) &&
4031 !free_area_empty(area, MIGRATE_CMA)) {
4035 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
4041 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
4042 int highest_zoneidx, unsigned int alloc_flags)
4044 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4045 zone_page_state(z, NR_FREE_PAGES));
4048 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
4049 unsigned long mark, int highest_zoneidx,
4050 unsigned int alloc_flags, gfp_t gfp_mask)
4054 free_pages = zone_page_state(z, NR_FREE_PAGES);
4057 * Fast check for order-0 only. If this fails then the reserves
4058 * need to be calculated.
4064 usable_free = free_pages;
4065 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
4067 /* reserved may over estimate high-atomic reserves. */
4068 usable_free -= min(usable_free, reserved);
4069 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
4073 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4077 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
4078 * when checking the min watermark. The min watermark is the
4079 * point where boosting is ignored so that kswapd is woken up
4080 * when below the low watermark.
4082 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
4083 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
4084 mark = z->_watermark[WMARK_MIN];
4085 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
4086 alloc_flags, free_pages);
4092 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
4093 unsigned long mark, int highest_zoneidx)
4095 long free_pages = zone_page_state(z, NR_FREE_PAGES);
4097 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
4098 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
4100 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
4105 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
4107 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4109 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
4110 node_reclaim_distance;
4112 #else /* CONFIG_NUMA */
4113 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4117 #endif /* CONFIG_NUMA */
4120 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4121 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4122 * premature use of a lower zone may cause lowmem pressure problems that
4123 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4124 * probably too small. It only makes sense to spread allocations to avoid
4125 * fragmentation between the Normal and DMA32 zones.
4127 static inline unsigned int
4128 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4130 unsigned int alloc_flags;
4133 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4136 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4138 #ifdef CONFIG_ZONE_DMA32
4142 if (zone_idx(zone) != ZONE_NORMAL)
4146 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4147 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4148 * on UMA that if Normal is populated then so is DMA32.
4150 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4151 if (nr_online_nodes > 1 && !populated_zone(--zone))
4154 alloc_flags |= ALLOC_NOFRAGMENT;
4155 #endif /* CONFIG_ZONE_DMA32 */
4159 /* Must be called after current_gfp_context() which can change gfp_mask */
4160 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4161 unsigned int alloc_flags)
4164 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4165 alloc_flags |= ALLOC_CMA;
4171 * get_page_from_freelist goes through the zonelist trying to allocate
4174 static struct page *
4175 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4176 const struct alloc_context *ac)
4180 struct pglist_data *last_pgdat = NULL;
4181 bool last_pgdat_dirty_ok = false;
4186 * Scan zonelist, looking for a zone with enough free.
4187 * See also __cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
4189 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4190 z = ac->preferred_zoneref;
4191 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4196 if (cpusets_enabled() &&
4197 (alloc_flags & ALLOC_CPUSET) &&
4198 !__cpuset_zone_allowed(zone, gfp_mask))
4201 * When allocating a page cache page for writing, we
4202 * want to get it from a node that is within its dirty
4203 * limit, such that no single node holds more than its
4204 * proportional share of globally allowed dirty pages.
4205 * The dirty limits take into account the node's
4206 * lowmem reserves and high watermark so that kswapd
4207 * should be able to balance it without having to
4208 * write pages from its LRU list.
4210 * XXX: For now, allow allocations to potentially
4211 * exceed the per-node dirty limit in the slowpath
4212 * (spread_dirty_pages unset) before going into reclaim,
4213 * which is important when on a NUMA setup the allowed
4214 * nodes are together not big enough to reach the
4215 * global limit. The proper fix for these situations
4216 * will require awareness of nodes in the
4217 * dirty-throttling and the flusher threads.
4219 if (ac->spread_dirty_pages) {
4220 if (last_pgdat != zone->zone_pgdat) {
4221 last_pgdat = zone->zone_pgdat;
4222 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
4225 if (!last_pgdat_dirty_ok)
4229 if (no_fallback && nr_online_nodes > 1 &&
4230 zone != ac->preferred_zoneref->zone) {
4234 * If moving to a remote node, retry but allow
4235 * fragmenting fallbacks. Locality is more important
4236 * than fragmentation avoidance.
4238 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4239 if (zone_to_nid(zone) != local_nid) {
4240 alloc_flags &= ~ALLOC_NOFRAGMENT;
4245 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4246 if (!zone_watermark_fast(zone, order, mark,
4247 ac->highest_zoneidx, alloc_flags,
4251 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4253 * Watermark failed for this zone, but see if we can
4254 * grow this zone if it contains deferred pages.
4256 if (static_branch_unlikely(&deferred_pages)) {
4257 if (_deferred_grow_zone(zone, order))
4261 /* Checked here to keep the fast path fast */
4262 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4263 if (alloc_flags & ALLOC_NO_WATERMARKS)
4266 if (!node_reclaim_enabled() ||
4267 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4270 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4272 case NODE_RECLAIM_NOSCAN:
4275 case NODE_RECLAIM_FULL:
4276 /* scanned but unreclaimable */
4279 /* did we reclaim enough */
4280 if (zone_watermark_ok(zone, order, mark,
4281 ac->highest_zoneidx, alloc_flags))
4289 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4290 gfp_mask, alloc_flags, ac->migratetype);
4292 prep_new_page(page, order, gfp_mask, alloc_flags);
4295 * If this is a high-order atomic allocation then check
4296 * if the pageblock should be reserved for the future
4298 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4299 reserve_highatomic_pageblock(page, zone, order);
4303 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4304 /* Try again if zone has deferred pages */
4305 if (static_branch_unlikely(&deferred_pages)) {
4306 if (_deferred_grow_zone(zone, order))
4314 * It's possible on a UMA machine to get through all zones that are
4315 * fragmented. If avoiding fragmentation, reset and try again.
4318 alloc_flags &= ~ALLOC_NOFRAGMENT;
4325 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4327 unsigned int filter = SHOW_MEM_FILTER_NODES;
4330 * This documents exceptions given to allocations in certain
4331 * contexts that are allowed to allocate outside current's set
4334 if (!(gfp_mask & __GFP_NOMEMALLOC))
4335 if (tsk_is_oom_victim(current) ||
4336 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4337 filter &= ~SHOW_MEM_FILTER_NODES;
4338 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4339 filter &= ~SHOW_MEM_FILTER_NODES;
4341 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
4344 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4346 struct va_format vaf;
4348 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4350 if ((gfp_mask & __GFP_NOWARN) ||
4351 !__ratelimit(&nopage_rs) ||
4352 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4355 va_start(args, fmt);
4358 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4359 current->comm, &vaf, gfp_mask, &gfp_mask,
4360 nodemask_pr_args(nodemask));
4363 cpuset_print_current_mems_allowed();
4366 warn_alloc_show_mem(gfp_mask, nodemask);
4369 static inline struct page *
4370 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4371 unsigned int alloc_flags,
4372 const struct alloc_context *ac)
4376 page = get_page_from_freelist(gfp_mask, order,
4377 alloc_flags|ALLOC_CPUSET, ac);
4379 * fallback to ignore cpuset restriction if our nodes
4383 page = get_page_from_freelist(gfp_mask, order,
4389 static inline struct page *
4390 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4391 const struct alloc_context *ac, unsigned long *did_some_progress)
4393 struct oom_control oc = {
4394 .zonelist = ac->zonelist,
4395 .nodemask = ac->nodemask,
4397 .gfp_mask = gfp_mask,
4402 *did_some_progress = 0;
4405 * Acquire the oom lock. If that fails, somebody else is
4406 * making progress for us.
4408 if (!mutex_trylock(&oom_lock)) {
4409 *did_some_progress = 1;
4410 schedule_timeout_uninterruptible(1);
4415 * Go through the zonelist yet one more time, keep very high watermark
4416 * here, this is only to catch a parallel oom killing, we must fail if
4417 * we're still under heavy pressure. But make sure that this reclaim
4418 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4419 * allocation which will never fail due to oom_lock already held.
4421 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4422 ~__GFP_DIRECT_RECLAIM, order,
4423 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4427 /* Coredumps can quickly deplete all memory reserves */
4428 if (current->flags & PF_DUMPCORE)
4430 /* The OOM killer will not help higher order allocs */
4431 if (order > PAGE_ALLOC_COSTLY_ORDER)
4434 * We have already exhausted all our reclaim opportunities without any
4435 * success so it is time to admit defeat. We will skip the OOM killer
4436 * because it is very likely that the caller has a more reasonable
4437 * fallback than shooting a random task.
4439 * The OOM killer may not free memory on a specific node.
4441 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4443 /* The OOM killer does not needlessly kill tasks for lowmem */
4444 if (ac->highest_zoneidx < ZONE_NORMAL)
4446 if (pm_suspended_storage())
4449 * XXX: GFP_NOFS allocations should rather fail than rely on
4450 * other request to make a forward progress.
4451 * We are in an unfortunate situation where out_of_memory cannot
4452 * do much for this context but let's try it to at least get
4453 * access to memory reserved if the current task is killed (see
4454 * out_of_memory). Once filesystems are ready to handle allocation
4455 * failures more gracefully we should just bail out here.
4458 /* Exhausted what can be done so it's blame time */
4459 if (out_of_memory(&oc) ||
4460 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4461 *did_some_progress = 1;
4464 * Help non-failing allocations by giving them access to memory
4467 if (gfp_mask & __GFP_NOFAIL)
4468 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4469 ALLOC_NO_WATERMARKS, ac);
4472 mutex_unlock(&oom_lock);
4477 * Maximum number of compaction retries with a progress before OOM
4478 * killer is consider as the only way to move forward.
4480 #define MAX_COMPACT_RETRIES 16
4482 #ifdef CONFIG_COMPACTION
4483 /* Try memory compaction for high-order allocations before reclaim */
4484 static struct page *
4485 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4486 unsigned int alloc_flags, const struct alloc_context *ac,
4487 enum compact_priority prio, enum compact_result *compact_result)
4489 struct page *page = NULL;
4490 unsigned long pflags;
4491 unsigned int noreclaim_flag;
4496 psi_memstall_enter(&pflags);
4497 delayacct_compact_start();
4498 noreclaim_flag = memalloc_noreclaim_save();
4500 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4503 memalloc_noreclaim_restore(noreclaim_flag);
4504 psi_memstall_leave(&pflags);
4505 delayacct_compact_end();
4507 if (*compact_result == COMPACT_SKIPPED)
4510 * At least in one zone compaction wasn't deferred or skipped, so let's
4511 * count a compaction stall
4513 count_vm_event(COMPACTSTALL);
4515 /* Prep a captured page if available */
4517 prep_new_page(page, order, gfp_mask, alloc_flags);
4519 /* Try get a page from the freelist if available */
4521 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4524 struct zone *zone = page_zone(page);
4526 zone->compact_blockskip_flush = false;
4527 compaction_defer_reset(zone, order, true);
4528 count_vm_event(COMPACTSUCCESS);
4533 * It's bad if compaction run occurs and fails. The most likely reason
4534 * is that pages exist, but not enough to satisfy watermarks.
4536 count_vm_event(COMPACTFAIL);
4544 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4545 enum compact_result compact_result,
4546 enum compact_priority *compact_priority,
4547 int *compaction_retries)
4549 int max_retries = MAX_COMPACT_RETRIES;
4552 int retries = *compaction_retries;
4553 enum compact_priority priority = *compact_priority;
4558 if (fatal_signal_pending(current))
4561 if (compaction_made_progress(compact_result))
4562 (*compaction_retries)++;
4565 * compaction considers all the zone as desperately out of memory
4566 * so it doesn't really make much sense to retry except when the
4567 * failure could be caused by insufficient priority
4569 if (compaction_failed(compact_result))
4570 goto check_priority;
4573 * compaction was skipped because there are not enough order-0 pages
4574 * to work with, so we retry only if it looks like reclaim can help.
4576 if (compaction_needs_reclaim(compact_result)) {
4577 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4582 * make sure the compaction wasn't deferred or didn't bail out early
4583 * due to locks contention before we declare that we should give up.
4584 * But the next retry should use a higher priority if allowed, so
4585 * we don't just keep bailing out endlessly.
4587 if (compaction_withdrawn(compact_result)) {
4588 goto check_priority;
4592 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4593 * costly ones because they are de facto nofail and invoke OOM
4594 * killer to move on while costly can fail and users are ready
4595 * to cope with that. 1/4 retries is rather arbitrary but we
4596 * would need much more detailed feedback from compaction to
4597 * make a better decision.
4599 if (order > PAGE_ALLOC_COSTLY_ORDER)
4601 if (*compaction_retries <= max_retries) {
4607 * Make sure there are attempts at the highest priority if we exhausted
4608 * all retries or failed at the lower priorities.
4611 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4612 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4614 if (*compact_priority > min_priority) {
4615 (*compact_priority)--;
4616 *compaction_retries = 0;
4620 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4624 static inline struct page *
4625 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4626 unsigned int alloc_flags, const struct alloc_context *ac,
4627 enum compact_priority prio, enum compact_result *compact_result)
4629 *compact_result = COMPACT_SKIPPED;
4634 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4635 enum compact_result compact_result,
4636 enum compact_priority *compact_priority,
4637 int *compaction_retries)
4642 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4646 * There are setups with compaction disabled which would prefer to loop
4647 * inside the allocator rather than hit the oom killer prematurely.
4648 * Let's give them a good hope and keep retrying while the order-0
4649 * watermarks are OK.
4651 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4652 ac->highest_zoneidx, ac->nodemask) {
4653 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4654 ac->highest_zoneidx, alloc_flags))
4659 #endif /* CONFIG_COMPACTION */
4661 #ifdef CONFIG_LOCKDEP
4662 static struct lockdep_map __fs_reclaim_map =
4663 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4665 static bool __need_reclaim(gfp_t gfp_mask)
4667 /* no reclaim without waiting on it */
4668 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4671 /* this guy won't enter reclaim */
4672 if (current->flags & PF_MEMALLOC)
4675 if (gfp_mask & __GFP_NOLOCKDEP)
4681 void __fs_reclaim_acquire(unsigned long ip)
4683 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4686 void __fs_reclaim_release(unsigned long ip)
4688 lock_release(&__fs_reclaim_map, ip);
4691 void fs_reclaim_acquire(gfp_t gfp_mask)
4693 gfp_mask = current_gfp_context(gfp_mask);
4695 if (__need_reclaim(gfp_mask)) {
4696 if (gfp_mask & __GFP_FS)
4697 __fs_reclaim_acquire(_RET_IP_);
4699 #ifdef CONFIG_MMU_NOTIFIER
4700 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4701 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4706 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4708 void fs_reclaim_release(gfp_t gfp_mask)
4710 gfp_mask = current_gfp_context(gfp_mask);
4712 if (__need_reclaim(gfp_mask)) {
4713 if (gfp_mask & __GFP_FS)
4714 __fs_reclaim_release(_RET_IP_);
4717 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4721 * Zonelists may change due to hotplug during allocation. Detect when zonelists
4722 * have been rebuilt so allocation retries. Reader side does not lock and
4723 * retries the allocation if zonelist changes. Writer side is protected by the
4724 * embedded spin_lock.
4726 static DEFINE_SEQLOCK(zonelist_update_seq);
4728 static unsigned int zonelist_iter_begin(void)
4730 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4731 return read_seqbegin(&zonelist_update_seq);
4736 static unsigned int check_retry_zonelist(unsigned int seq)
4738 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4739 return read_seqretry(&zonelist_update_seq, seq);
4744 /* Perform direct synchronous page reclaim */
4745 static unsigned long
4746 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4747 const struct alloc_context *ac)
4749 unsigned int noreclaim_flag;
4750 unsigned long progress;
4754 /* We now go into synchronous reclaim */
4755 cpuset_memory_pressure_bump();
4756 fs_reclaim_acquire(gfp_mask);
4757 noreclaim_flag = memalloc_noreclaim_save();
4759 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4762 memalloc_noreclaim_restore(noreclaim_flag);
4763 fs_reclaim_release(gfp_mask);
4770 /* The really slow allocator path where we enter direct reclaim */
4771 static inline struct page *
4772 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4773 unsigned int alloc_flags, const struct alloc_context *ac,
4774 unsigned long *did_some_progress)
4776 struct page *page = NULL;
4777 unsigned long pflags;
4778 bool drained = false;
4780 psi_memstall_enter(&pflags);
4781 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4782 if (unlikely(!(*did_some_progress)))
4786 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4789 * If an allocation failed after direct reclaim, it could be because
4790 * pages are pinned on the per-cpu lists or in high alloc reserves.
4791 * Shrink them and try again
4793 if (!page && !drained) {
4794 unreserve_highatomic_pageblock(ac, false);
4795 drain_all_pages(NULL);
4800 psi_memstall_leave(&pflags);
4805 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4806 const struct alloc_context *ac)
4810 pg_data_t *last_pgdat = NULL;
4811 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4813 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4815 if (!managed_zone(zone))
4817 if (last_pgdat != zone->zone_pgdat) {
4818 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4819 last_pgdat = zone->zone_pgdat;
4824 static inline unsigned int
4825 gfp_to_alloc_flags(gfp_t gfp_mask)
4827 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4830 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4831 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4832 * to save two branches.
4834 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4835 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4838 * The caller may dip into page reserves a bit more if the caller
4839 * cannot run direct reclaim, or if the caller has realtime scheduling
4840 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4841 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4843 alloc_flags |= (__force int)
4844 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4846 if (gfp_mask & __GFP_ATOMIC) {
4848 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4849 * if it can't schedule.
4851 if (!(gfp_mask & __GFP_NOMEMALLOC))
4852 alloc_flags |= ALLOC_HARDER;
4854 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4855 * comment for __cpuset_node_allowed().
4857 alloc_flags &= ~ALLOC_CPUSET;
4858 } else if (unlikely(rt_task(current)) && in_task())
4859 alloc_flags |= ALLOC_HARDER;
4861 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4866 static bool oom_reserves_allowed(struct task_struct *tsk)
4868 if (!tsk_is_oom_victim(tsk))
4872 * !MMU doesn't have oom reaper so give access to memory reserves
4873 * only to the thread with TIF_MEMDIE set
4875 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4882 * Distinguish requests which really need access to full memory
4883 * reserves from oom victims which can live with a portion of it
4885 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4887 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4889 if (gfp_mask & __GFP_MEMALLOC)
4890 return ALLOC_NO_WATERMARKS;
4891 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4892 return ALLOC_NO_WATERMARKS;
4893 if (!in_interrupt()) {
4894 if (current->flags & PF_MEMALLOC)
4895 return ALLOC_NO_WATERMARKS;
4896 else if (oom_reserves_allowed(current))
4903 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4905 return !!__gfp_pfmemalloc_flags(gfp_mask);
4909 * Checks whether it makes sense to retry the reclaim to make a forward progress
4910 * for the given allocation request.
4912 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4913 * without success, or when we couldn't even meet the watermark if we
4914 * reclaimed all remaining pages on the LRU lists.
4916 * Returns true if a retry is viable or false to enter the oom path.
4919 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4920 struct alloc_context *ac, int alloc_flags,
4921 bool did_some_progress, int *no_progress_loops)
4928 * Costly allocations might have made a progress but this doesn't mean
4929 * their order will become available due to high fragmentation so
4930 * always increment the no progress counter for them
4932 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4933 *no_progress_loops = 0;
4935 (*no_progress_loops)++;
4938 * Make sure we converge to OOM if we cannot make any progress
4939 * several times in the row.
4941 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4942 /* Before OOM, exhaust highatomic_reserve */
4943 return unreserve_highatomic_pageblock(ac, true);
4947 * Keep reclaiming pages while there is a chance this will lead
4948 * somewhere. If none of the target zones can satisfy our allocation
4949 * request even if all reclaimable pages are considered then we are
4950 * screwed and have to go OOM.
4952 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4953 ac->highest_zoneidx, ac->nodemask) {
4954 unsigned long available;
4955 unsigned long reclaimable;
4956 unsigned long min_wmark = min_wmark_pages(zone);
4959 available = reclaimable = zone_reclaimable_pages(zone);
4960 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4963 * Would the allocation succeed if we reclaimed all
4964 * reclaimable pages?
4966 wmark = __zone_watermark_ok(zone, order, min_wmark,
4967 ac->highest_zoneidx, alloc_flags, available);
4968 trace_reclaim_retry_zone(z, order, reclaimable,
4969 available, min_wmark, *no_progress_loops, wmark);
4977 * Memory allocation/reclaim might be called from a WQ context and the
4978 * current implementation of the WQ concurrency control doesn't
4979 * recognize that a particular WQ is congested if the worker thread is
4980 * looping without ever sleeping. Therefore we have to do a short sleep
4981 * here rather than calling cond_resched().
4983 if (current->flags & PF_WQ_WORKER)
4984 schedule_timeout_uninterruptible(1);
4991 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4994 * It's possible that cpuset's mems_allowed and the nodemask from
4995 * mempolicy don't intersect. This should be normally dealt with by
4996 * policy_nodemask(), but it's possible to race with cpuset update in
4997 * such a way the check therein was true, and then it became false
4998 * before we got our cpuset_mems_cookie here.
4999 * This assumes that for all allocations, ac->nodemask can come only
5000 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
5001 * when it does not intersect with the cpuset restrictions) or the
5002 * caller can deal with a violated nodemask.
5004 if (cpusets_enabled() && ac->nodemask &&
5005 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
5006 ac->nodemask = NULL;
5011 * When updating a task's mems_allowed or mempolicy nodemask, it is
5012 * possible to race with parallel threads in such a way that our
5013 * allocation can fail while the mask is being updated. If we are about
5014 * to fail, check if the cpuset changed during allocation and if so,
5017 if (read_mems_allowed_retry(cpuset_mems_cookie))
5023 static inline struct page *
5024 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
5025 struct alloc_context *ac)
5027 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
5028 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
5029 struct page *page = NULL;
5030 unsigned int alloc_flags;
5031 unsigned long did_some_progress;
5032 enum compact_priority compact_priority;
5033 enum compact_result compact_result;
5034 int compaction_retries;
5035 int no_progress_loops;
5036 unsigned int cpuset_mems_cookie;
5037 unsigned int zonelist_iter_cookie;
5041 * We also sanity check to catch abuse of atomic reserves being used by
5042 * callers that are not in atomic context.
5044 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
5045 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
5046 gfp_mask &= ~__GFP_ATOMIC;
5049 compaction_retries = 0;
5050 no_progress_loops = 0;
5051 compact_priority = DEF_COMPACT_PRIORITY;
5052 cpuset_mems_cookie = read_mems_allowed_begin();
5053 zonelist_iter_cookie = zonelist_iter_begin();
5056 * The fast path uses conservative alloc_flags to succeed only until
5057 * kswapd needs to be woken up, and to avoid the cost of setting up
5058 * alloc_flags precisely. So we do that now.
5060 alloc_flags = gfp_to_alloc_flags(gfp_mask);
5063 * We need to recalculate the starting point for the zonelist iterator
5064 * because we might have used different nodemask in the fast path, or
5065 * there was a cpuset modification and we are retrying - otherwise we
5066 * could end up iterating over non-eligible zones endlessly.
5068 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5069 ac->highest_zoneidx, ac->nodemask);
5070 if (!ac->preferred_zoneref->zone)
5074 * Check for insane configurations where the cpuset doesn't contain
5075 * any suitable zone to satisfy the request - e.g. non-movable
5076 * GFP_HIGHUSER allocations from MOVABLE nodes only.
5078 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
5079 struct zoneref *z = first_zones_zonelist(ac->zonelist,
5080 ac->highest_zoneidx,
5081 &cpuset_current_mems_allowed);
5086 if (alloc_flags & ALLOC_KSWAPD)
5087 wake_all_kswapds(order, gfp_mask, ac);
5090 * The adjusted alloc_flags might result in immediate success, so try
5093 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5098 * For costly allocations, try direct compaction first, as it's likely
5099 * that we have enough base pages and don't need to reclaim. For non-
5100 * movable high-order allocations, do that as well, as compaction will
5101 * try prevent permanent fragmentation by migrating from blocks of the
5103 * Don't try this for allocations that are allowed to ignore
5104 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
5106 if (can_direct_reclaim &&
5108 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
5109 && !gfp_pfmemalloc_allowed(gfp_mask)) {
5110 page = __alloc_pages_direct_compact(gfp_mask, order,
5112 INIT_COMPACT_PRIORITY,
5118 * Checks for costly allocations with __GFP_NORETRY, which
5119 * includes some THP page fault allocations
5121 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
5123 * If allocating entire pageblock(s) and compaction
5124 * failed because all zones are below low watermarks
5125 * or is prohibited because it recently failed at this
5126 * order, fail immediately unless the allocator has
5127 * requested compaction and reclaim retry.
5130 * - potentially very expensive because zones are far
5131 * below their low watermarks or this is part of very
5132 * bursty high order allocations,
5133 * - not guaranteed to help because isolate_freepages()
5134 * may not iterate over freed pages as part of its
5136 * - unlikely to make entire pageblocks free on its
5139 if (compact_result == COMPACT_SKIPPED ||
5140 compact_result == COMPACT_DEFERRED)
5144 * Looks like reclaim/compaction is worth trying, but
5145 * sync compaction could be very expensive, so keep
5146 * using async compaction.
5148 compact_priority = INIT_COMPACT_PRIORITY;
5153 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5154 if (alloc_flags & ALLOC_KSWAPD)
5155 wake_all_kswapds(order, gfp_mask, ac);
5157 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5159 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
5160 (alloc_flags & ALLOC_KSWAPD);
5163 * Reset the nodemask and zonelist iterators if memory policies can be
5164 * ignored. These allocations are high priority and system rather than
5167 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5168 ac->nodemask = NULL;
5169 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5170 ac->highest_zoneidx, ac->nodemask);
5173 /* Attempt with potentially adjusted zonelist and alloc_flags */
5174 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5178 /* Caller is not willing to reclaim, we can't balance anything */
5179 if (!can_direct_reclaim)
5182 /* Avoid recursion of direct reclaim */
5183 if (current->flags & PF_MEMALLOC)
5186 /* Try direct reclaim and then allocating */
5187 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5188 &did_some_progress);
5192 /* Try direct compaction and then allocating */
5193 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5194 compact_priority, &compact_result);
5198 /* Do not loop if specifically requested */
5199 if (gfp_mask & __GFP_NORETRY)
5203 * Do not retry costly high order allocations unless they are
5204 * __GFP_RETRY_MAYFAIL
5206 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5209 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5210 did_some_progress > 0, &no_progress_loops))
5214 * It doesn't make any sense to retry for the compaction if the order-0
5215 * reclaim is not able to make any progress because the current
5216 * implementation of the compaction depends on the sufficient amount
5217 * of free memory (see __compaction_suitable)
5219 if (did_some_progress > 0 &&
5220 should_compact_retry(ac, order, alloc_flags,
5221 compact_result, &compact_priority,
5222 &compaction_retries))
5227 * Deal with possible cpuset update races or zonelist updates to avoid
5228 * a unnecessary OOM kill.
5230 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5231 check_retry_zonelist(zonelist_iter_cookie))
5234 /* Reclaim has failed us, start killing things */
5235 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5239 /* Avoid allocations with no watermarks from looping endlessly */
5240 if (tsk_is_oom_victim(current) &&
5241 (alloc_flags & ALLOC_OOM ||
5242 (gfp_mask & __GFP_NOMEMALLOC)))
5245 /* Retry as long as the OOM killer is making progress */
5246 if (did_some_progress) {
5247 no_progress_loops = 0;
5253 * Deal with possible cpuset update races or zonelist updates to avoid
5254 * a unnecessary OOM kill.
5256 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5257 check_retry_zonelist(zonelist_iter_cookie))
5261 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5264 if (gfp_mask & __GFP_NOFAIL) {
5266 * All existing users of the __GFP_NOFAIL are blockable, so warn
5267 * of any new users that actually require GFP_NOWAIT
5269 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
5273 * PF_MEMALLOC request from this context is rather bizarre
5274 * because we cannot reclaim anything and only can loop waiting
5275 * for somebody to do a work for us
5277 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
5280 * non failing costly orders are a hard requirement which we
5281 * are not prepared for much so let's warn about these users
5282 * so that we can identify them and convert them to something
5285 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
5288 * Help non-failing allocations by giving them access to memory
5289 * reserves but do not use ALLOC_NO_WATERMARKS because this
5290 * could deplete whole memory reserves which would just make
5291 * the situation worse
5293 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5301 warn_alloc(gfp_mask, ac->nodemask,
5302 "page allocation failure: order:%u", order);
5307 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5308 int preferred_nid, nodemask_t *nodemask,
5309 struct alloc_context *ac, gfp_t *alloc_gfp,
5310 unsigned int *alloc_flags)
5312 ac->highest_zoneidx = gfp_zone(gfp_mask);
5313 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5314 ac->nodemask = nodemask;
5315 ac->migratetype = gfp_migratetype(gfp_mask);
5317 if (cpusets_enabled()) {
5318 *alloc_gfp |= __GFP_HARDWALL;
5320 * When we are in the interrupt context, it is irrelevant
5321 * to the current task context. It means that any node ok.
5323 if (in_task() && !ac->nodemask)
5324 ac->nodemask = &cpuset_current_mems_allowed;
5326 *alloc_flags |= ALLOC_CPUSET;
5329 might_alloc(gfp_mask);
5331 if (should_fail_alloc_page(gfp_mask, order))
5334 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5336 /* Dirty zone balancing only done in the fast path */
5337 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5340 * The preferred zone is used for statistics but crucially it is
5341 * also used as the starting point for the zonelist iterator. It
5342 * may get reset for allocations that ignore memory policies.
5344 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5345 ac->highest_zoneidx, ac->nodemask);
5351 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5352 * @gfp: GFP flags for the allocation
5353 * @preferred_nid: The preferred NUMA node ID to allocate from
5354 * @nodemask: Set of nodes to allocate from, may be NULL
5355 * @nr_pages: The number of pages desired on the list or array
5356 * @page_list: Optional list to store the allocated pages
5357 * @page_array: Optional array to store the pages
5359 * This is a batched version of the page allocator that attempts to
5360 * allocate nr_pages quickly. Pages are added to page_list if page_list
5361 * is not NULL, otherwise it is assumed that the page_array is valid.
5363 * For lists, nr_pages is the number of pages that should be allocated.
5365 * For arrays, only NULL elements are populated with pages and nr_pages
5366 * is the maximum number of pages that will be stored in the array.
5368 * Returns the number of pages on the list or array.
5370 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5371 nodemask_t *nodemask, int nr_pages,
5372 struct list_head *page_list,
5373 struct page **page_array)
5376 unsigned long flags;
5377 unsigned long __maybe_unused UP_flags;
5380 struct per_cpu_pages *pcp;
5381 struct list_head *pcp_list;
5382 struct alloc_context ac;
5384 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5385 int nr_populated = 0, nr_account = 0;
5388 * Skip populated array elements to determine if any pages need
5389 * to be allocated before disabling IRQs.
5391 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5394 /* No pages requested? */
5395 if (unlikely(nr_pages <= 0))
5398 /* Already populated array? */
5399 if (unlikely(page_array && nr_pages - nr_populated == 0))
5402 /* Bulk allocator does not support memcg accounting. */
5403 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5406 /* Use the single page allocator for one page. */
5407 if (nr_pages - nr_populated == 1)
5410 #ifdef CONFIG_PAGE_OWNER
5412 * PAGE_OWNER may recurse into the allocator to allocate space to
5413 * save the stack with pagesets.lock held. Releasing/reacquiring
5414 * removes much of the performance benefit of bulk allocation so
5415 * force the caller to allocate one page at a time as it'll have
5416 * similar performance to added complexity to the bulk allocator.
5418 if (static_branch_unlikely(&page_owner_inited))
5422 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5423 gfp &= gfp_allowed_mask;
5425 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5429 /* Find an allowed local zone that meets the low watermark. */
5430 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5433 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5434 !__cpuset_zone_allowed(zone, gfp)) {
5438 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5439 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5443 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5444 if (zone_watermark_fast(zone, 0, mark,
5445 zonelist_zone_idx(ac.preferred_zoneref),
5446 alloc_flags, gfp)) {
5452 * If there are no allowed local zones that meets the watermarks then
5453 * try to allocate a single page and reclaim if necessary.
5455 if (unlikely(!zone))
5458 /* Is a parallel drain in progress? */
5459 pcp_trylock_prepare(UP_flags);
5460 pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
5464 /* Attempt the batch allocation */
5465 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5466 while (nr_populated < nr_pages) {
5468 /* Skip existing pages */
5469 if (page_array && page_array[nr_populated]) {
5474 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5476 if (unlikely(!page)) {
5477 /* Try and allocate at least one page */
5479 pcp_spin_unlock_irqrestore(pcp, flags);
5486 prep_new_page(page, 0, gfp, 0);
5488 list_add(&page->lru, page_list);
5490 page_array[nr_populated] = page;
5494 pcp_spin_unlock_irqrestore(pcp, flags);
5495 pcp_trylock_finish(UP_flags);
5497 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5498 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5501 return nr_populated;
5504 pcp_trylock_finish(UP_flags);
5507 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5510 list_add(&page->lru, page_list);
5512 page_array[nr_populated] = page;
5518 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5521 * This is the 'heart' of the zoned buddy allocator.
5523 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5524 nodemask_t *nodemask)
5527 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5528 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5529 struct alloc_context ac = { };
5532 * There are several places where we assume that the order value is sane
5533 * so bail out early if the request is out of bound.
5535 if (WARN_ON_ONCE_GFP(order >= MAX_ORDER, gfp))
5538 gfp &= gfp_allowed_mask;
5540 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5541 * resp. GFP_NOIO which has to be inherited for all allocation requests
5542 * from a particular context which has been marked by
5543 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5544 * movable zones are not used during allocation.
5546 gfp = current_gfp_context(gfp);
5548 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5549 &alloc_gfp, &alloc_flags))
5553 * Forbid the first pass from falling back to types that fragment
5554 * memory until all local zones are considered.
5556 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5558 /* First allocation attempt */
5559 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5564 ac.spread_dirty_pages = false;
5567 * Restore the original nodemask if it was potentially replaced with
5568 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5570 ac.nodemask = nodemask;
5572 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5575 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5576 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5577 __free_pages(page, order);
5581 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5582 kmsan_alloc_page(page, order, alloc_gfp);
5586 EXPORT_SYMBOL(__alloc_pages);
5588 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5589 nodemask_t *nodemask)
5591 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5592 preferred_nid, nodemask);
5594 if (page && order > 1)
5595 prep_transhuge_page(page);
5596 return (struct folio *)page;
5598 EXPORT_SYMBOL(__folio_alloc);
5601 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5602 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5603 * you need to access high mem.
5605 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5609 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5612 return (unsigned long) page_address(page);
5614 EXPORT_SYMBOL(__get_free_pages);
5616 unsigned long get_zeroed_page(gfp_t gfp_mask)
5618 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5620 EXPORT_SYMBOL(get_zeroed_page);
5623 * __free_pages - Free pages allocated with alloc_pages().
5624 * @page: The page pointer returned from alloc_pages().
5625 * @order: The order of the allocation.
5627 * This function can free multi-page allocations that are not compound
5628 * pages. It does not check that the @order passed in matches that of
5629 * the allocation, so it is easy to leak memory. Freeing more memory
5630 * than was allocated will probably emit a warning.
5632 * If the last reference to this page is speculative, it will be released
5633 * by put_page() which only frees the first page of a non-compound
5634 * allocation. To prevent the remaining pages from being leaked, we free
5635 * the subsequent pages here. If you want to use the page's reference
5636 * count to decide when to free the allocation, you should allocate a
5637 * compound page, and use put_page() instead of __free_pages().
5639 * Context: May be called in interrupt context or while holding a normal
5640 * spinlock, but not in NMI context or while holding a raw spinlock.
5642 void __free_pages(struct page *page, unsigned int order)
5644 /* get PageHead before we drop reference */
5645 int head = PageHead(page);
5647 if (put_page_testzero(page))
5648 free_the_page(page, order);
5651 free_the_page(page + (1 << order), order);
5653 EXPORT_SYMBOL(__free_pages);
5655 void free_pages(unsigned long addr, unsigned int order)
5658 VM_BUG_ON(!virt_addr_valid((void *)addr));
5659 __free_pages(virt_to_page((void *)addr), order);
5663 EXPORT_SYMBOL(free_pages);
5667 * An arbitrary-length arbitrary-offset area of memory which resides
5668 * within a 0 or higher order page. Multiple fragments within that page
5669 * are individually refcounted, in the page's reference counter.
5671 * The page_frag functions below provide a simple allocation framework for
5672 * page fragments. This is used by the network stack and network device
5673 * drivers to provide a backing region of memory for use as either an
5674 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5676 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5679 struct page *page = NULL;
5680 gfp_t gfp = gfp_mask;
5682 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5683 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5685 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5686 PAGE_FRAG_CACHE_MAX_ORDER);
5687 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5689 if (unlikely(!page))
5690 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5692 nc->va = page ? page_address(page) : NULL;
5697 void __page_frag_cache_drain(struct page *page, unsigned int count)
5699 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5701 if (page_ref_sub_and_test(page, count))
5702 free_the_page(page, compound_order(page));
5704 EXPORT_SYMBOL(__page_frag_cache_drain);
5706 void *page_frag_alloc_align(struct page_frag_cache *nc,
5707 unsigned int fragsz, gfp_t gfp_mask,
5708 unsigned int align_mask)
5710 unsigned int size = PAGE_SIZE;
5714 if (unlikely(!nc->va)) {
5716 page = __page_frag_cache_refill(nc, gfp_mask);
5720 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5721 /* if size can vary use size else just use PAGE_SIZE */
5724 /* Even if we own the page, we do not use atomic_set().
5725 * This would break get_page_unless_zero() users.
5727 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5729 /* reset page count bias and offset to start of new frag */
5730 nc->pfmemalloc = page_is_pfmemalloc(page);
5731 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5735 offset = nc->offset - fragsz;
5736 if (unlikely(offset < 0)) {
5737 page = virt_to_page(nc->va);
5739 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5742 if (unlikely(nc->pfmemalloc)) {
5743 free_the_page(page, compound_order(page));
5747 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5748 /* if size can vary use size else just use PAGE_SIZE */
5751 /* OK, page count is 0, we can safely set it */
5752 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5754 /* reset page count bias and offset to start of new frag */
5755 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5756 offset = size - fragsz;
5757 if (unlikely(offset < 0)) {
5759 * The caller is trying to allocate a fragment
5760 * with fragsz > PAGE_SIZE but the cache isn't big
5761 * enough to satisfy the request, this may
5762 * happen in low memory conditions.
5763 * We don't release the cache page because
5764 * it could make memory pressure worse
5765 * so we simply return NULL here.
5772 offset &= align_mask;
5773 nc->offset = offset;
5775 return nc->va + offset;
5777 EXPORT_SYMBOL(page_frag_alloc_align);
5780 * Frees a page fragment allocated out of either a compound or order 0 page.
5782 void page_frag_free(void *addr)
5784 struct page *page = virt_to_head_page(addr);
5786 if (unlikely(put_page_testzero(page)))
5787 free_the_page(page, compound_order(page));
5789 EXPORT_SYMBOL(page_frag_free);
5791 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5795 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
5796 struct page *page = virt_to_page((void *)addr);
5797 struct page *last = page + nr;
5799 split_page_owner(page, 1 << order);
5800 split_page_memcg(page, 1 << order);
5801 while (page < --last)
5802 set_page_refcounted(last);
5804 last = page + (1UL << order);
5805 for (page += nr; page < last; page++)
5806 __free_pages_ok(page, 0, FPI_TO_TAIL);
5808 return (void *)addr;
5812 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5813 * @size: the number of bytes to allocate
5814 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5816 * This function is similar to alloc_pages(), except that it allocates the
5817 * minimum number of pages to satisfy the request. alloc_pages() can only
5818 * allocate memory in power-of-two pages.
5820 * This function is also limited by MAX_ORDER.
5822 * Memory allocated by this function must be released by free_pages_exact().
5824 * Return: pointer to the allocated area or %NULL in case of error.
5826 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5828 unsigned int order = get_order(size);
5831 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5832 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5834 addr = __get_free_pages(gfp_mask, order);
5835 return make_alloc_exact(addr, order, size);
5837 EXPORT_SYMBOL(alloc_pages_exact);
5840 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5842 * @nid: the preferred node ID where memory should be allocated
5843 * @size: the number of bytes to allocate
5844 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5846 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5849 * Return: pointer to the allocated area or %NULL in case of error.
5851 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5853 unsigned int order = get_order(size);
5856 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5857 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5859 p = alloc_pages_node(nid, gfp_mask, order);
5862 return make_alloc_exact((unsigned long)page_address(p), order, size);
5866 * free_pages_exact - release memory allocated via alloc_pages_exact()
5867 * @virt: the value returned by alloc_pages_exact.
5868 * @size: size of allocation, same value as passed to alloc_pages_exact().
5870 * Release the memory allocated by a previous call to alloc_pages_exact.
5872 void free_pages_exact(void *virt, size_t size)
5874 unsigned long addr = (unsigned long)virt;
5875 unsigned long end = addr + PAGE_ALIGN(size);
5877 while (addr < end) {
5882 EXPORT_SYMBOL(free_pages_exact);
5885 * nr_free_zone_pages - count number of pages beyond high watermark
5886 * @offset: The zone index of the highest zone
5888 * nr_free_zone_pages() counts the number of pages which are beyond the
5889 * high watermark within all zones at or below a given zone index. For each
5890 * zone, the number of pages is calculated as:
5892 * nr_free_zone_pages = managed_pages - high_pages
5894 * Return: number of pages beyond high watermark.
5896 static unsigned long nr_free_zone_pages(int offset)
5901 /* Just pick one node, since fallback list is circular */
5902 unsigned long sum = 0;
5904 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5906 for_each_zone_zonelist(zone, z, zonelist, offset) {
5907 unsigned long size = zone_managed_pages(zone);
5908 unsigned long high = high_wmark_pages(zone);
5917 * nr_free_buffer_pages - count number of pages beyond high watermark
5919 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5920 * watermark within ZONE_DMA and ZONE_NORMAL.
5922 * Return: number of pages beyond high watermark within ZONE_DMA and
5925 unsigned long nr_free_buffer_pages(void)
5927 return nr_free_zone_pages(gfp_zone(GFP_USER));
5929 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5931 static inline void show_node(struct zone *zone)
5933 if (IS_ENABLED(CONFIG_NUMA))
5934 printk("Node %d ", zone_to_nid(zone));
5937 long si_mem_available(void)
5940 unsigned long pagecache;
5941 unsigned long wmark_low = 0;
5942 unsigned long pages[NR_LRU_LISTS];
5943 unsigned long reclaimable;
5947 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5948 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5951 wmark_low += low_wmark_pages(zone);
5954 * Estimate the amount of memory available for userspace allocations,
5955 * without causing swapping or OOM.
5957 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5960 * Not all the page cache can be freed, otherwise the system will
5961 * start swapping or thrashing. Assume at least half of the page
5962 * cache, or the low watermark worth of cache, needs to stay.
5964 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5965 pagecache -= min(pagecache / 2, wmark_low);
5966 available += pagecache;
5969 * Part of the reclaimable slab and other kernel memory consists of
5970 * items that are in use, and cannot be freed. Cap this estimate at the
5973 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5974 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5975 available += reclaimable - min(reclaimable / 2, wmark_low);
5981 EXPORT_SYMBOL_GPL(si_mem_available);
5983 void si_meminfo(struct sysinfo *val)
5985 val->totalram = totalram_pages();
5986 val->sharedram = global_node_page_state(NR_SHMEM);
5987 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5988 val->bufferram = nr_blockdev_pages();
5989 val->totalhigh = totalhigh_pages();
5990 val->freehigh = nr_free_highpages();
5991 val->mem_unit = PAGE_SIZE;
5994 EXPORT_SYMBOL(si_meminfo);
5997 void si_meminfo_node(struct sysinfo *val, int nid)
5999 int zone_type; /* needs to be signed */
6000 unsigned long managed_pages = 0;
6001 unsigned long managed_highpages = 0;
6002 unsigned long free_highpages = 0;
6003 pg_data_t *pgdat = NODE_DATA(nid);
6005 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
6006 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
6007 val->totalram = managed_pages;
6008 val->sharedram = node_page_state(pgdat, NR_SHMEM);
6009 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
6010 #ifdef CONFIG_HIGHMEM
6011 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
6012 struct zone *zone = &pgdat->node_zones[zone_type];
6014 if (is_highmem(zone)) {
6015 managed_highpages += zone_managed_pages(zone);
6016 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
6019 val->totalhigh = managed_highpages;
6020 val->freehigh = free_highpages;
6022 val->totalhigh = managed_highpages;
6023 val->freehigh = free_highpages;
6025 val->mem_unit = PAGE_SIZE;
6030 * Determine whether the node should be displayed or not, depending on whether
6031 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
6033 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
6035 if (!(flags & SHOW_MEM_FILTER_NODES))
6039 * no node mask - aka implicit memory numa policy. Do not bother with
6040 * the synchronization - read_mems_allowed_begin - because we do not
6041 * have to be precise here.
6044 nodemask = &cpuset_current_mems_allowed;
6046 return !node_isset(nid, *nodemask);
6049 #define K(x) ((x) << (PAGE_SHIFT-10))
6051 static void show_migration_types(unsigned char type)
6053 static const char types[MIGRATE_TYPES] = {
6054 [MIGRATE_UNMOVABLE] = 'U',
6055 [MIGRATE_MOVABLE] = 'M',
6056 [MIGRATE_RECLAIMABLE] = 'E',
6057 [MIGRATE_HIGHATOMIC] = 'H',
6059 [MIGRATE_CMA] = 'C',
6061 #ifdef CONFIG_MEMORY_ISOLATION
6062 [MIGRATE_ISOLATE] = 'I',
6065 char tmp[MIGRATE_TYPES + 1];
6069 for (i = 0; i < MIGRATE_TYPES; i++) {
6070 if (type & (1 << i))
6075 printk(KERN_CONT "(%s) ", tmp);
6078 static bool node_has_managed_zones(pg_data_t *pgdat, int max_zone_idx)
6081 for (zone_idx = 0; zone_idx <= max_zone_idx; zone_idx++)
6082 if (zone_managed_pages(pgdat->node_zones + zone_idx))
6088 * Show free area list (used inside shift_scroll-lock stuff)
6089 * We also calculate the percentage fragmentation. We do this by counting the
6090 * memory on each free list with the exception of the first item on the list.
6093 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
6096 void __show_free_areas(unsigned int filter, nodemask_t *nodemask, int max_zone_idx)
6098 unsigned long free_pcp = 0;
6103 for_each_populated_zone(zone) {
6104 if (zone_idx(zone) > max_zone_idx)
6106 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6109 for_each_online_cpu(cpu)
6110 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6113 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
6114 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
6115 " unevictable:%lu dirty:%lu writeback:%lu\n"
6116 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
6117 " mapped:%lu shmem:%lu pagetables:%lu\n"
6118 " sec_pagetables:%lu bounce:%lu\n"
6119 " kernel_misc_reclaimable:%lu\n"
6120 " free:%lu free_pcp:%lu free_cma:%lu\n",
6121 global_node_page_state(NR_ACTIVE_ANON),
6122 global_node_page_state(NR_INACTIVE_ANON),
6123 global_node_page_state(NR_ISOLATED_ANON),
6124 global_node_page_state(NR_ACTIVE_FILE),
6125 global_node_page_state(NR_INACTIVE_FILE),
6126 global_node_page_state(NR_ISOLATED_FILE),
6127 global_node_page_state(NR_UNEVICTABLE),
6128 global_node_page_state(NR_FILE_DIRTY),
6129 global_node_page_state(NR_WRITEBACK),
6130 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
6131 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
6132 global_node_page_state(NR_FILE_MAPPED),
6133 global_node_page_state(NR_SHMEM),
6134 global_node_page_state(NR_PAGETABLE),
6135 global_node_page_state(NR_SECONDARY_PAGETABLE),
6136 global_zone_page_state(NR_BOUNCE),
6137 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
6138 global_zone_page_state(NR_FREE_PAGES),
6140 global_zone_page_state(NR_FREE_CMA_PAGES));
6142 for_each_online_pgdat(pgdat) {
6143 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
6145 if (!node_has_managed_zones(pgdat, max_zone_idx))
6149 " active_anon:%lukB"
6150 " inactive_anon:%lukB"
6151 " active_file:%lukB"
6152 " inactive_file:%lukB"
6153 " unevictable:%lukB"
6154 " isolated(anon):%lukB"
6155 " isolated(file):%lukB"
6160 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6162 " shmem_pmdmapped: %lukB"
6165 " writeback_tmp:%lukB"
6166 " kernel_stack:%lukB"
6167 #ifdef CONFIG_SHADOW_CALL_STACK
6168 " shadow_call_stack:%lukB"
6171 " sec_pagetables:%lukB"
6172 " all_unreclaimable? %s"
6175 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
6176 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
6177 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
6178 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
6179 K(node_page_state(pgdat, NR_UNEVICTABLE)),
6180 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
6181 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
6182 K(node_page_state(pgdat, NR_FILE_MAPPED)),
6183 K(node_page_state(pgdat, NR_FILE_DIRTY)),
6184 K(node_page_state(pgdat, NR_WRITEBACK)),
6185 K(node_page_state(pgdat, NR_SHMEM)),
6186 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6187 K(node_page_state(pgdat, NR_SHMEM_THPS)),
6188 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6189 K(node_page_state(pgdat, NR_ANON_THPS)),
6191 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6192 node_page_state(pgdat, NR_KERNEL_STACK_KB),
6193 #ifdef CONFIG_SHADOW_CALL_STACK
6194 node_page_state(pgdat, NR_KERNEL_SCS_KB),
6196 K(node_page_state(pgdat, NR_PAGETABLE)),
6197 K(node_page_state(pgdat, NR_SECONDARY_PAGETABLE)),
6198 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6202 for_each_populated_zone(zone) {
6205 if (zone_idx(zone) > max_zone_idx)
6207 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6211 for_each_online_cpu(cpu)
6212 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6222 " reserved_highatomic:%luKB"
6223 " active_anon:%lukB"
6224 " inactive_anon:%lukB"
6225 " active_file:%lukB"
6226 " inactive_file:%lukB"
6227 " unevictable:%lukB"
6228 " writepending:%lukB"
6238 K(zone_page_state(zone, NR_FREE_PAGES)),
6239 K(zone->watermark_boost),
6240 K(min_wmark_pages(zone)),
6241 K(low_wmark_pages(zone)),
6242 K(high_wmark_pages(zone)),
6243 K(zone->nr_reserved_highatomic),
6244 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6245 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6246 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6247 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6248 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6249 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6250 K(zone->present_pages),
6251 K(zone_managed_pages(zone)),
6252 K(zone_page_state(zone, NR_MLOCK)),
6253 K(zone_page_state(zone, NR_BOUNCE)),
6255 K(this_cpu_read(zone->per_cpu_pageset->count)),
6256 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6257 printk("lowmem_reserve[]:");
6258 for (i = 0; i < MAX_NR_ZONES; i++)
6259 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6260 printk(KERN_CONT "\n");
6263 for_each_populated_zone(zone) {
6265 unsigned long nr[MAX_ORDER], flags, total = 0;
6266 unsigned char types[MAX_ORDER];
6268 if (zone_idx(zone) > max_zone_idx)
6270 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6273 printk(KERN_CONT "%s: ", zone->name);
6275 spin_lock_irqsave(&zone->lock, flags);
6276 for (order = 0; order < MAX_ORDER; order++) {
6277 struct free_area *area = &zone->free_area[order];
6280 nr[order] = area->nr_free;
6281 total += nr[order] << order;
6284 for (type = 0; type < MIGRATE_TYPES; type++) {
6285 if (!free_area_empty(area, type))
6286 types[order] |= 1 << type;
6289 spin_unlock_irqrestore(&zone->lock, flags);
6290 for (order = 0; order < MAX_ORDER; order++) {
6291 printk(KERN_CONT "%lu*%lukB ",
6292 nr[order], K(1UL) << order);
6294 show_migration_types(types[order]);
6296 printk(KERN_CONT "= %lukB\n", K(total));
6299 for_each_online_node(nid) {
6300 if (show_mem_node_skip(filter, nid, nodemask))
6302 hugetlb_show_meminfo_node(nid);
6305 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6307 show_swap_cache_info();
6310 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6312 zoneref->zone = zone;
6313 zoneref->zone_idx = zone_idx(zone);
6317 * Builds allocation fallback zone lists.
6319 * Add all populated zones of a node to the zonelist.
6321 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6324 enum zone_type zone_type = MAX_NR_ZONES;
6329 zone = pgdat->node_zones + zone_type;
6330 if (populated_zone(zone)) {
6331 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6332 check_highest_zone(zone_type);
6334 } while (zone_type);
6341 static int __parse_numa_zonelist_order(char *s)
6344 * We used to support different zonelists modes but they turned
6345 * out to be just not useful. Let's keep the warning in place
6346 * if somebody still use the cmd line parameter so that we do
6347 * not fail it silently
6349 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6350 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6356 char numa_zonelist_order[] = "Node";
6359 * sysctl handler for numa_zonelist_order
6361 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6362 void *buffer, size_t *length, loff_t *ppos)
6365 return __parse_numa_zonelist_order(buffer);
6366 return proc_dostring(table, write, buffer, length, ppos);
6370 static int node_load[MAX_NUMNODES];
6373 * find_next_best_node - find the next node that should appear in a given node's fallback list
6374 * @node: node whose fallback list we're appending
6375 * @used_node_mask: nodemask_t of already used nodes
6377 * We use a number of factors to determine which is the next node that should
6378 * appear on a given node's fallback list. The node should not have appeared
6379 * already in @node's fallback list, and it should be the next closest node
6380 * according to the distance array (which contains arbitrary distance values
6381 * from each node to each node in the system), and should also prefer nodes
6382 * with no CPUs, since presumably they'll have very little allocation pressure
6383 * on them otherwise.
6385 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6387 int find_next_best_node(int node, nodemask_t *used_node_mask)
6390 int min_val = INT_MAX;
6391 int best_node = NUMA_NO_NODE;
6393 /* Use the local node if we haven't already */
6394 if (!node_isset(node, *used_node_mask)) {
6395 node_set(node, *used_node_mask);
6399 for_each_node_state(n, N_MEMORY) {
6401 /* Don't want a node to appear more than once */
6402 if (node_isset(n, *used_node_mask))
6405 /* Use the distance array to find the distance */
6406 val = node_distance(node, n);
6408 /* Penalize nodes under us ("prefer the next node") */
6411 /* Give preference to headless and unused nodes */
6412 if (!cpumask_empty(cpumask_of_node(n)))
6413 val += PENALTY_FOR_NODE_WITH_CPUS;
6415 /* Slight preference for less loaded node */
6416 val *= MAX_NUMNODES;
6417 val += node_load[n];
6419 if (val < min_val) {
6426 node_set(best_node, *used_node_mask);
6433 * Build zonelists ordered by node and zones within node.
6434 * This results in maximum locality--normal zone overflows into local
6435 * DMA zone, if any--but risks exhausting DMA zone.
6437 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6440 struct zoneref *zonerefs;
6443 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6445 for (i = 0; i < nr_nodes; i++) {
6448 pg_data_t *node = NODE_DATA(node_order[i]);
6450 nr_zones = build_zonerefs_node(node, zonerefs);
6451 zonerefs += nr_zones;
6453 zonerefs->zone = NULL;
6454 zonerefs->zone_idx = 0;
6458 * Build gfp_thisnode zonelists
6460 static void build_thisnode_zonelists(pg_data_t *pgdat)
6462 struct zoneref *zonerefs;
6465 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6466 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6467 zonerefs += nr_zones;
6468 zonerefs->zone = NULL;
6469 zonerefs->zone_idx = 0;
6473 * Build zonelists ordered by zone and nodes within zones.
6474 * This results in conserving DMA zone[s] until all Normal memory is
6475 * exhausted, but results in overflowing to remote node while memory
6476 * may still exist in local DMA zone.
6479 static void build_zonelists(pg_data_t *pgdat)
6481 static int node_order[MAX_NUMNODES];
6482 int node, nr_nodes = 0;
6483 nodemask_t used_mask = NODE_MASK_NONE;
6484 int local_node, prev_node;
6486 /* NUMA-aware ordering of nodes */
6487 local_node = pgdat->node_id;
6488 prev_node = local_node;
6490 memset(node_order, 0, sizeof(node_order));
6491 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6493 * We don't want to pressure a particular node.
6494 * So adding penalty to the first node in same
6495 * distance group to make it round-robin.
6497 if (node_distance(local_node, node) !=
6498 node_distance(local_node, prev_node))
6499 node_load[node] += 1;
6501 node_order[nr_nodes++] = node;
6505 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6506 build_thisnode_zonelists(pgdat);
6507 pr_info("Fallback order for Node %d: ", local_node);
6508 for (node = 0; node < nr_nodes; node++)
6509 pr_cont("%d ", node_order[node]);
6513 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6515 * Return node id of node used for "local" allocations.
6516 * I.e., first node id of first zone in arg node's generic zonelist.
6517 * Used for initializing percpu 'numa_mem', which is used primarily
6518 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6520 int local_memory_node(int node)
6524 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6525 gfp_zone(GFP_KERNEL),
6527 return zone_to_nid(z->zone);
6531 static void setup_min_unmapped_ratio(void);
6532 static void setup_min_slab_ratio(void);
6533 #else /* CONFIG_NUMA */
6535 static void build_zonelists(pg_data_t *pgdat)
6537 int node, local_node;
6538 struct zoneref *zonerefs;
6541 local_node = pgdat->node_id;
6543 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6544 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6545 zonerefs += nr_zones;
6548 * Now we build the zonelist so that it contains the zones
6549 * of all the other nodes.
6550 * We don't want to pressure a particular node, so when
6551 * building the zones for node N, we make sure that the
6552 * zones coming right after the local ones are those from
6553 * node N+1 (modulo N)
6555 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6556 if (!node_online(node))
6558 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6559 zonerefs += nr_zones;
6561 for (node = 0; node < local_node; node++) {
6562 if (!node_online(node))
6564 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6565 zonerefs += nr_zones;
6568 zonerefs->zone = NULL;
6569 zonerefs->zone_idx = 0;
6572 #endif /* CONFIG_NUMA */
6575 * Boot pageset table. One per cpu which is going to be used for all
6576 * zones and all nodes. The parameters will be set in such a way
6577 * that an item put on a list will immediately be handed over to
6578 * the buddy list. This is safe since pageset manipulation is done
6579 * with interrupts disabled.
6581 * The boot_pagesets must be kept even after bootup is complete for
6582 * unused processors and/or zones. They do play a role for bootstrapping
6583 * hotplugged processors.
6585 * zoneinfo_show() and maybe other functions do
6586 * not check if the processor is online before following the pageset pointer.
6587 * Other parts of the kernel may not check if the zone is available.
6589 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6590 /* These effectively disable the pcplists in the boot pageset completely */
6591 #define BOOT_PAGESET_HIGH 0
6592 #define BOOT_PAGESET_BATCH 1
6593 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6594 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6595 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6597 static void __build_all_zonelists(void *data)
6600 int __maybe_unused cpu;
6601 pg_data_t *self = data;
6603 write_seqlock(&zonelist_update_seq);
6606 memset(node_load, 0, sizeof(node_load));
6610 * This node is hotadded and no memory is yet present. So just
6611 * building zonelists is fine - no need to touch other nodes.
6613 if (self && !node_online(self->node_id)) {
6614 build_zonelists(self);
6617 * All possible nodes have pgdat preallocated
6620 for_each_node(nid) {
6621 pg_data_t *pgdat = NODE_DATA(nid);
6623 build_zonelists(pgdat);
6626 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6628 * We now know the "local memory node" for each node--
6629 * i.e., the node of the first zone in the generic zonelist.
6630 * Set up numa_mem percpu variable for on-line cpus. During
6631 * boot, only the boot cpu should be on-line; we'll init the
6632 * secondary cpus' numa_mem as they come on-line. During
6633 * node/memory hotplug, we'll fixup all on-line cpus.
6635 for_each_online_cpu(cpu)
6636 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6640 write_sequnlock(&zonelist_update_seq);
6643 static noinline void __init
6644 build_all_zonelists_init(void)
6648 __build_all_zonelists(NULL);
6651 * Initialize the boot_pagesets that are going to be used
6652 * for bootstrapping processors. The real pagesets for
6653 * each zone will be allocated later when the per cpu
6654 * allocator is available.
6656 * boot_pagesets are used also for bootstrapping offline
6657 * cpus if the system is already booted because the pagesets
6658 * are needed to initialize allocators on a specific cpu too.
6659 * F.e. the percpu allocator needs the page allocator which
6660 * needs the percpu allocator in order to allocate its pagesets
6661 * (a chicken-egg dilemma).
6663 for_each_possible_cpu(cpu)
6664 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6666 mminit_verify_zonelist();
6667 cpuset_init_current_mems_allowed();
6671 * unless system_state == SYSTEM_BOOTING.
6673 * __ref due to call of __init annotated helper build_all_zonelists_init
6674 * [protected by SYSTEM_BOOTING].
6676 void __ref build_all_zonelists(pg_data_t *pgdat)
6678 unsigned long vm_total_pages;
6680 if (system_state == SYSTEM_BOOTING) {
6681 build_all_zonelists_init();
6683 __build_all_zonelists(pgdat);
6684 /* cpuset refresh routine should be here */
6686 /* Get the number of free pages beyond high watermark in all zones. */
6687 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6689 * Disable grouping by mobility if the number of pages in the
6690 * system is too low to allow the mechanism to work. It would be
6691 * more accurate, but expensive to check per-zone. This check is
6692 * made on memory-hotadd so a system can start with mobility
6693 * disabled and enable it later
6695 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6696 page_group_by_mobility_disabled = 1;
6698 page_group_by_mobility_disabled = 0;
6700 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6702 page_group_by_mobility_disabled ? "off" : "on",
6705 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6709 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6710 static bool __meminit
6711 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6713 static struct memblock_region *r;
6715 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6716 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6717 for_each_mem_region(r) {
6718 if (*pfn < memblock_region_memory_end_pfn(r))
6722 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6723 memblock_is_mirror(r)) {
6724 *pfn = memblock_region_memory_end_pfn(r);
6732 * Initially all pages are reserved - free ones are freed
6733 * up by memblock_free_all() once the early boot process is
6734 * done. Non-atomic initialization, single-pass.
6736 * All aligned pageblocks are initialized to the specified migratetype
6737 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6738 * zone stats (e.g., nr_isolate_pageblock) are touched.
6740 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6741 unsigned long start_pfn, unsigned long zone_end_pfn,
6742 enum meminit_context context,
6743 struct vmem_altmap *altmap, int migratetype)
6745 unsigned long pfn, end_pfn = start_pfn + size;
6748 if (highest_memmap_pfn < end_pfn - 1)
6749 highest_memmap_pfn = end_pfn - 1;
6751 #ifdef CONFIG_ZONE_DEVICE
6753 * Honor reservation requested by the driver for this ZONE_DEVICE
6754 * memory. We limit the total number of pages to initialize to just
6755 * those that might contain the memory mapping. We will defer the
6756 * ZONE_DEVICE page initialization until after we have released
6759 if (zone == ZONE_DEVICE) {
6763 if (start_pfn == altmap->base_pfn)
6764 start_pfn += altmap->reserve;
6765 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6769 for (pfn = start_pfn; pfn < end_pfn; ) {
6771 * There can be holes in boot-time mem_map[]s handed to this
6772 * function. They do not exist on hotplugged memory.
6774 if (context == MEMINIT_EARLY) {
6775 if (overlap_memmap_init(zone, &pfn))
6777 if (defer_init(nid, pfn, zone_end_pfn))
6781 page = pfn_to_page(pfn);
6782 __init_single_page(page, pfn, zone, nid);
6783 if (context == MEMINIT_HOTPLUG)
6784 __SetPageReserved(page);
6787 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6788 * such that unmovable allocations won't be scattered all
6789 * over the place during system boot.
6791 if (pageblock_aligned(pfn)) {
6792 set_pageblock_migratetype(page, migratetype);
6799 #ifdef CONFIG_ZONE_DEVICE
6800 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6801 unsigned long zone_idx, int nid,
6802 struct dev_pagemap *pgmap)
6805 __init_single_page(page, pfn, zone_idx, nid);
6808 * Mark page reserved as it will need to wait for onlining
6809 * phase for it to be fully associated with a zone.
6811 * We can use the non-atomic __set_bit operation for setting
6812 * the flag as we are still initializing the pages.
6814 __SetPageReserved(page);
6817 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6818 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6819 * ever freed or placed on a driver-private list.
6821 page->pgmap = pgmap;
6822 page->zone_device_data = NULL;
6825 * Mark the block movable so that blocks are reserved for
6826 * movable at startup. This will force kernel allocations
6827 * to reserve their blocks rather than leaking throughout
6828 * the address space during boot when many long-lived
6829 * kernel allocations are made.
6831 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6832 * because this is done early in section_activate()
6834 if (pageblock_aligned(pfn)) {
6835 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6840 * ZONE_DEVICE pages are released directly to the driver page allocator
6841 * which will set the page count to 1 when allocating the page.
6843 if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
6844 pgmap->type == MEMORY_DEVICE_COHERENT)
6845 set_page_count(page, 0);
6849 * With compound page geometry and when struct pages are stored in ram most
6850 * tail pages are reused. Consequently, the amount of unique struct pages to
6851 * initialize is a lot smaller that the total amount of struct pages being
6852 * mapped. This is a paired / mild layering violation with explicit knowledge
6853 * of how the sparse_vmemmap internals handle compound pages in the lack
6854 * of an altmap. See vmemmap_populate_compound_pages().
6856 static inline unsigned long compound_nr_pages(struct vmem_altmap *altmap,
6857 unsigned long nr_pages)
6859 return is_power_of_2(sizeof(struct page)) &&
6860 !altmap ? 2 * (PAGE_SIZE / sizeof(struct page)) : nr_pages;
6863 static void __ref memmap_init_compound(struct page *head,
6864 unsigned long head_pfn,
6865 unsigned long zone_idx, int nid,
6866 struct dev_pagemap *pgmap,
6867 unsigned long nr_pages)
6869 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6870 unsigned int order = pgmap->vmemmap_shift;
6872 __SetPageHead(head);
6873 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6874 struct page *page = pfn_to_page(pfn);
6876 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6877 prep_compound_tail(head, pfn - head_pfn);
6878 set_page_count(page, 0);
6881 * The first tail page stores compound_mapcount_ptr() and
6882 * compound_order() and the second tail page stores
6883 * compound_pincount_ptr(). Call prep_compound_head() after
6884 * the first and second tail pages have been initialized to
6885 * not have the data overwritten.
6887 if (pfn == head_pfn + 2)
6888 prep_compound_head(head, order);
6892 void __ref memmap_init_zone_device(struct zone *zone,
6893 unsigned long start_pfn,
6894 unsigned long nr_pages,
6895 struct dev_pagemap *pgmap)
6897 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6898 struct pglist_data *pgdat = zone->zone_pgdat;
6899 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6900 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6901 unsigned long zone_idx = zone_idx(zone);
6902 unsigned long start = jiffies;
6903 int nid = pgdat->node_id;
6905 if (WARN_ON_ONCE(!pgmap || zone_idx != ZONE_DEVICE))
6909 * The call to memmap_init should have already taken care
6910 * of the pages reserved for the memmap, so we can just jump to
6911 * the end of that region and start processing the device pages.
6914 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6915 nr_pages = end_pfn - start_pfn;
6918 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6919 struct page *page = pfn_to_page(pfn);
6921 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6923 if (pfns_per_compound == 1)
6926 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6927 compound_nr_pages(altmap, pfns_per_compound));
6930 pr_info("%s initialised %lu pages in %ums\n", __func__,
6931 nr_pages, jiffies_to_msecs(jiffies - start));
6935 static void __meminit zone_init_free_lists(struct zone *zone)
6937 unsigned int order, t;
6938 for_each_migratetype_order(order, t) {
6939 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6940 zone->free_area[order].nr_free = 0;
6945 * Only struct pages that correspond to ranges defined by memblock.memory
6946 * are zeroed and initialized by going through __init_single_page() during
6947 * memmap_init_zone_range().
6949 * But, there could be struct pages that correspond to holes in
6950 * memblock.memory. This can happen because of the following reasons:
6951 * - physical memory bank size is not necessarily the exact multiple of the
6952 * arbitrary section size
6953 * - early reserved memory may not be listed in memblock.memory
6954 * - memory layouts defined with memmap= kernel parameter may not align
6955 * nicely with memmap sections
6957 * Explicitly initialize those struct pages so that:
6958 * - PG_Reserved is set
6959 * - zone and node links point to zone and node that span the page if the
6960 * hole is in the middle of a zone
6961 * - zone and node links point to adjacent zone/node if the hole falls on
6962 * the zone boundary; the pages in such holes will be prepended to the
6963 * zone/node above the hole except for the trailing pages in the last
6964 * section that will be appended to the zone/node below.
6966 static void __init init_unavailable_range(unsigned long spfn,
6973 for (pfn = spfn; pfn < epfn; pfn++) {
6974 if (!pfn_valid(pageblock_start_pfn(pfn))) {
6975 pfn = pageblock_end_pfn(pfn) - 1;
6978 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6979 __SetPageReserved(pfn_to_page(pfn));
6984 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6985 node, zone_names[zone], pgcnt);
6988 static void __init memmap_init_zone_range(struct zone *zone,
6989 unsigned long start_pfn,
6990 unsigned long end_pfn,
6991 unsigned long *hole_pfn)
6993 unsigned long zone_start_pfn = zone->zone_start_pfn;
6994 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6995 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6997 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6998 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
7000 if (start_pfn >= end_pfn)
7003 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
7004 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
7006 if (*hole_pfn < start_pfn)
7007 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
7009 *hole_pfn = end_pfn;
7012 static void __init memmap_init(void)
7014 unsigned long start_pfn, end_pfn;
7015 unsigned long hole_pfn = 0;
7016 int i, j, zone_id = 0, nid;
7018 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7019 struct pglist_data *node = NODE_DATA(nid);
7021 for (j = 0; j < MAX_NR_ZONES; j++) {
7022 struct zone *zone = node->node_zones + j;
7024 if (!populated_zone(zone))
7027 memmap_init_zone_range(zone, start_pfn, end_pfn,
7033 #ifdef CONFIG_SPARSEMEM
7035 * Initialize the memory map for hole in the range [memory_end,
7037 * Append the pages in this hole to the highest zone in the last
7039 * The call to init_unavailable_range() is outside the ifdef to
7040 * silence the compiler warining about zone_id set but not used;
7041 * for FLATMEM it is a nop anyway
7043 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
7044 if (hole_pfn < end_pfn)
7046 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
7049 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
7050 phys_addr_t min_addr, int nid, bool exact_nid)
7055 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
7056 MEMBLOCK_ALLOC_ACCESSIBLE,
7059 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
7060 MEMBLOCK_ALLOC_ACCESSIBLE,
7063 if (ptr && size > 0)
7064 page_init_poison(ptr, size);
7069 static int zone_batchsize(struct zone *zone)
7075 * The number of pages to batch allocate is either ~0.1%
7076 * of the zone or 1MB, whichever is smaller. The batch
7077 * size is striking a balance between allocation latency
7078 * and zone lock contention.
7080 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
7081 batch /= 4; /* We effectively *= 4 below */
7086 * Clamp the batch to a 2^n - 1 value. Having a power
7087 * of 2 value was found to be more likely to have
7088 * suboptimal cache aliasing properties in some cases.
7090 * For example if 2 tasks are alternately allocating
7091 * batches of pages, one task can end up with a lot
7092 * of pages of one half of the possible page colors
7093 * and the other with pages of the other colors.
7095 batch = rounddown_pow_of_two(batch + batch/2) - 1;
7100 /* The deferral and batching of frees should be suppressed under NOMMU
7103 * The problem is that NOMMU needs to be able to allocate large chunks
7104 * of contiguous memory as there's no hardware page translation to
7105 * assemble apparent contiguous memory from discontiguous pages.
7107 * Queueing large contiguous runs of pages for batching, however,
7108 * causes the pages to actually be freed in smaller chunks. As there
7109 * can be a significant delay between the individual batches being
7110 * recycled, this leads to the once large chunks of space being
7111 * fragmented and becoming unavailable for high-order allocations.
7117 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
7122 unsigned long total_pages;
7124 if (!percpu_pagelist_high_fraction) {
7126 * By default, the high value of the pcp is based on the zone
7127 * low watermark so that if they are full then background
7128 * reclaim will not be started prematurely.
7130 total_pages = low_wmark_pages(zone);
7133 * If percpu_pagelist_high_fraction is configured, the high
7134 * value is based on a fraction of the managed pages in the
7137 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
7141 * Split the high value across all online CPUs local to the zone. Note
7142 * that early in boot that CPUs may not be online yet and that during
7143 * CPU hotplug that the cpumask is not yet updated when a CPU is being
7144 * onlined. For memory nodes that have no CPUs, split pcp->high across
7145 * all online CPUs to mitigate the risk that reclaim is triggered
7146 * prematurely due to pages stored on pcp lists.
7148 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
7150 nr_split_cpus = num_online_cpus();
7151 high = total_pages / nr_split_cpus;
7154 * Ensure high is at least batch*4. The multiple is based on the
7155 * historical relationship between high and batch.
7157 high = max(high, batch << 2);
7166 * pcp->high and pcp->batch values are related and generally batch is lower
7167 * than high. They are also related to pcp->count such that count is lower
7168 * than high, and as soon as it reaches high, the pcplist is flushed.
7170 * However, guaranteeing these relations at all times would require e.g. write
7171 * barriers here but also careful usage of read barriers at the read side, and
7172 * thus be prone to error and bad for performance. Thus the update only prevents
7173 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
7174 * can cope with those fields changing asynchronously, and fully trust only the
7175 * pcp->count field on the local CPU with interrupts disabled.
7177 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
7178 * outside of boot time (or some other assurance that no concurrent updaters
7181 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
7182 unsigned long batch)
7184 WRITE_ONCE(pcp->batch, batch);
7185 WRITE_ONCE(pcp->high, high);
7188 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
7192 memset(pcp, 0, sizeof(*pcp));
7193 memset(pzstats, 0, sizeof(*pzstats));
7195 spin_lock_init(&pcp->lock);
7196 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
7197 INIT_LIST_HEAD(&pcp->lists[pindex]);
7200 * Set batch and high values safe for a boot pageset. A true percpu
7201 * pageset's initialization will update them subsequently. Here we don't
7202 * need to be as careful as pageset_update() as nobody can access the
7205 pcp->high = BOOT_PAGESET_HIGH;
7206 pcp->batch = BOOT_PAGESET_BATCH;
7207 pcp->free_factor = 0;
7210 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
7211 unsigned long batch)
7213 struct per_cpu_pages *pcp;
7216 for_each_possible_cpu(cpu) {
7217 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7218 pageset_update(pcp, high, batch);
7223 * Calculate and set new high and batch values for all per-cpu pagesets of a
7224 * zone based on the zone's size.
7226 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
7228 int new_high, new_batch;
7230 new_batch = max(1, zone_batchsize(zone));
7231 new_high = zone_highsize(zone, new_batch, cpu_online);
7233 if (zone->pageset_high == new_high &&
7234 zone->pageset_batch == new_batch)
7237 zone->pageset_high = new_high;
7238 zone->pageset_batch = new_batch;
7240 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7243 void __meminit setup_zone_pageset(struct zone *zone)
7247 /* Size may be 0 on !SMP && !NUMA */
7248 if (sizeof(struct per_cpu_zonestat) > 0)
7249 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7251 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7252 for_each_possible_cpu(cpu) {
7253 struct per_cpu_pages *pcp;
7254 struct per_cpu_zonestat *pzstats;
7256 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7257 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7258 per_cpu_pages_init(pcp, pzstats);
7261 zone_set_pageset_high_and_batch(zone, 0);
7265 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7266 * page high values need to be recalculated.
7268 static void zone_pcp_update(struct zone *zone, int cpu_online)
7270 mutex_lock(&pcp_batch_high_lock);
7271 zone_set_pageset_high_and_batch(zone, cpu_online);
7272 mutex_unlock(&pcp_batch_high_lock);
7276 * Allocate per cpu pagesets and initialize them.
7277 * Before this call only boot pagesets were available.
7279 void __init setup_per_cpu_pageset(void)
7281 struct pglist_data *pgdat;
7283 int __maybe_unused cpu;
7285 for_each_populated_zone(zone)
7286 setup_zone_pageset(zone);
7290 * Unpopulated zones continue using the boot pagesets.
7291 * The numa stats for these pagesets need to be reset.
7292 * Otherwise, they will end up skewing the stats of
7293 * the nodes these zones are associated with.
7295 for_each_possible_cpu(cpu) {
7296 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7297 memset(pzstats->vm_numa_event, 0,
7298 sizeof(pzstats->vm_numa_event));
7302 for_each_online_pgdat(pgdat)
7303 pgdat->per_cpu_nodestats =
7304 alloc_percpu(struct per_cpu_nodestat);
7307 static __meminit void zone_pcp_init(struct zone *zone)
7310 * per cpu subsystem is not up at this point. The following code
7311 * relies on the ability of the linker to provide the
7312 * offset of a (static) per cpu variable into the per cpu area.
7314 zone->per_cpu_pageset = &boot_pageset;
7315 zone->per_cpu_zonestats = &boot_zonestats;
7316 zone->pageset_high = BOOT_PAGESET_HIGH;
7317 zone->pageset_batch = BOOT_PAGESET_BATCH;
7319 if (populated_zone(zone))
7320 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7321 zone->present_pages, zone_batchsize(zone));
7324 void __meminit init_currently_empty_zone(struct zone *zone,
7325 unsigned long zone_start_pfn,
7328 struct pglist_data *pgdat = zone->zone_pgdat;
7329 int zone_idx = zone_idx(zone) + 1;
7331 if (zone_idx > pgdat->nr_zones)
7332 pgdat->nr_zones = zone_idx;
7334 zone->zone_start_pfn = zone_start_pfn;
7336 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7337 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7339 (unsigned long)zone_idx(zone),
7340 zone_start_pfn, (zone_start_pfn + size));
7342 zone_init_free_lists(zone);
7343 zone->initialized = 1;
7347 * get_pfn_range_for_nid - Return the start and end page frames for a node
7348 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7349 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7350 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7352 * It returns the start and end page frame of a node based on information
7353 * provided by memblock_set_node(). If called for a node
7354 * with no available memory, a warning is printed and the start and end
7357 void __init get_pfn_range_for_nid(unsigned int nid,
7358 unsigned long *start_pfn, unsigned long *end_pfn)
7360 unsigned long this_start_pfn, this_end_pfn;
7366 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7367 *start_pfn = min(*start_pfn, this_start_pfn);
7368 *end_pfn = max(*end_pfn, this_end_pfn);
7371 if (*start_pfn == -1UL)
7376 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7377 * assumption is made that zones within a node are ordered in monotonic
7378 * increasing memory addresses so that the "highest" populated zone is used
7380 static void __init find_usable_zone_for_movable(void)
7383 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7384 if (zone_index == ZONE_MOVABLE)
7387 if (arch_zone_highest_possible_pfn[zone_index] >
7388 arch_zone_lowest_possible_pfn[zone_index])
7392 VM_BUG_ON(zone_index == -1);
7393 movable_zone = zone_index;
7397 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7398 * because it is sized independent of architecture. Unlike the other zones,
7399 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7400 * in each node depending on the size of each node and how evenly kernelcore
7401 * is distributed. This helper function adjusts the zone ranges
7402 * provided by the architecture for a given node by using the end of the
7403 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7404 * zones within a node are in order of monotonic increases memory addresses
7406 static void __init adjust_zone_range_for_zone_movable(int nid,
7407 unsigned long zone_type,
7408 unsigned long node_start_pfn,
7409 unsigned long node_end_pfn,
7410 unsigned long *zone_start_pfn,
7411 unsigned long *zone_end_pfn)
7413 /* Only adjust if ZONE_MOVABLE is on this node */
7414 if (zone_movable_pfn[nid]) {
7415 /* Size ZONE_MOVABLE */
7416 if (zone_type == ZONE_MOVABLE) {
7417 *zone_start_pfn = zone_movable_pfn[nid];
7418 *zone_end_pfn = min(node_end_pfn,
7419 arch_zone_highest_possible_pfn[movable_zone]);
7421 /* Adjust for ZONE_MOVABLE starting within this range */
7422 } else if (!mirrored_kernelcore &&
7423 *zone_start_pfn < zone_movable_pfn[nid] &&
7424 *zone_end_pfn > zone_movable_pfn[nid]) {
7425 *zone_end_pfn = zone_movable_pfn[nid];
7427 /* Check if this whole range is within ZONE_MOVABLE */
7428 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7429 *zone_start_pfn = *zone_end_pfn;
7434 * Return the number of pages a zone spans in a node, including holes
7435 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7437 static unsigned long __init zone_spanned_pages_in_node(int nid,
7438 unsigned long zone_type,
7439 unsigned long node_start_pfn,
7440 unsigned long node_end_pfn,
7441 unsigned long *zone_start_pfn,
7442 unsigned long *zone_end_pfn)
7444 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7445 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7446 /* When hotadd a new node from cpu_up(), the node should be empty */
7447 if (!node_start_pfn && !node_end_pfn)
7450 /* Get the start and end of the zone */
7451 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7452 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7453 adjust_zone_range_for_zone_movable(nid, zone_type,
7454 node_start_pfn, node_end_pfn,
7455 zone_start_pfn, zone_end_pfn);
7457 /* Check that this node has pages within the zone's required range */
7458 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7461 /* Move the zone boundaries inside the node if necessary */
7462 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7463 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7465 /* Return the spanned pages */
7466 return *zone_end_pfn - *zone_start_pfn;
7470 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7471 * then all holes in the requested range will be accounted for.
7473 unsigned long __init __absent_pages_in_range(int nid,
7474 unsigned long range_start_pfn,
7475 unsigned long range_end_pfn)
7477 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7478 unsigned long start_pfn, end_pfn;
7481 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7482 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7483 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7484 nr_absent -= end_pfn - start_pfn;
7490 * absent_pages_in_range - Return number of page frames in holes within a range
7491 * @start_pfn: The start PFN to start searching for holes
7492 * @end_pfn: The end PFN to stop searching for holes
7494 * Return: the number of pages frames in memory holes within a range.
7496 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7497 unsigned long end_pfn)
7499 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7502 /* Return the number of page frames in holes in a zone on a node */
7503 static unsigned long __init zone_absent_pages_in_node(int nid,
7504 unsigned long zone_type,
7505 unsigned long node_start_pfn,
7506 unsigned long node_end_pfn)
7508 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7509 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7510 unsigned long zone_start_pfn, zone_end_pfn;
7511 unsigned long nr_absent;
7513 /* When hotadd a new node from cpu_up(), the node should be empty */
7514 if (!node_start_pfn && !node_end_pfn)
7517 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7518 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7520 adjust_zone_range_for_zone_movable(nid, zone_type,
7521 node_start_pfn, node_end_pfn,
7522 &zone_start_pfn, &zone_end_pfn);
7523 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7526 * ZONE_MOVABLE handling.
7527 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7530 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7531 unsigned long start_pfn, end_pfn;
7532 struct memblock_region *r;
7534 for_each_mem_region(r) {
7535 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7536 zone_start_pfn, zone_end_pfn);
7537 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7538 zone_start_pfn, zone_end_pfn);
7540 if (zone_type == ZONE_MOVABLE &&
7541 memblock_is_mirror(r))
7542 nr_absent += end_pfn - start_pfn;
7544 if (zone_type == ZONE_NORMAL &&
7545 !memblock_is_mirror(r))
7546 nr_absent += end_pfn - start_pfn;
7553 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7554 unsigned long node_start_pfn,
7555 unsigned long node_end_pfn)
7557 unsigned long realtotalpages = 0, totalpages = 0;
7560 for (i = 0; i < MAX_NR_ZONES; i++) {
7561 struct zone *zone = pgdat->node_zones + i;
7562 unsigned long zone_start_pfn, zone_end_pfn;
7563 unsigned long spanned, absent;
7564 unsigned long size, real_size;
7566 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7571 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7576 real_size = size - absent;
7579 zone->zone_start_pfn = zone_start_pfn;
7581 zone->zone_start_pfn = 0;
7582 zone->spanned_pages = size;
7583 zone->present_pages = real_size;
7584 #if defined(CONFIG_MEMORY_HOTPLUG)
7585 zone->present_early_pages = real_size;
7589 realtotalpages += real_size;
7592 pgdat->node_spanned_pages = totalpages;
7593 pgdat->node_present_pages = realtotalpages;
7594 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7597 #ifndef CONFIG_SPARSEMEM
7599 * Calculate the size of the zone->blockflags rounded to an unsigned long
7600 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7601 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7602 * round what is now in bits to nearest long in bits, then return it in
7605 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7607 unsigned long usemapsize;
7609 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7610 usemapsize = roundup(zonesize, pageblock_nr_pages);
7611 usemapsize = usemapsize >> pageblock_order;
7612 usemapsize *= NR_PAGEBLOCK_BITS;
7613 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7615 return usemapsize / 8;
7618 static void __ref setup_usemap(struct zone *zone)
7620 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7621 zone->spanned_pages);
7622 zone->pageblock_flags = NULL;
7624 zone->pageblock_flags =
7625 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7627 if (!zone->pageblock_flags)
7628 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7629 usemapsize, zone->name, zone_to_nid(zone));
7633 static inline void setup_usemap(struct zone *zone) {}
7634 #endif /* CONFIG_SPARSEMEM */
7636 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7638 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7639 void __init set_pageblock_order(void)
7641 unsigned int order = MAX_ORDER - 1;
7643 /* Check that pageblock_nr_pages has not already been setup */
7644 if (pageblock_order)
7647 /* Don't let pageblocks exceed the maximum allocation granularity. */
7648 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7649 order = HUGETLB_PAGE_ORDER;
7652 * Assume the largest contiguous order of interest is a huge page.
7653 * This value may be variable depending on boot parameters on IA64 and
7656 pageblock_order = order;
7658 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7661 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7662 * is unused as pageblock_order is set at compile-time. See
7663 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7666 void __init set_pageblock_order(void)
7670 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7672 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7673 unsigned long present_pages)
7675 unsigned long pages = spanned_pages;
7678 * Provide a more accurate estimation if there are holes within
7679 * the zone and SPARSEMEM is in use. If there are holes within the
7680 * zone, each populated memory region may cost us one or two extra
7681 * memmap pages due to alignment because memmap pages for each
7682 * populated regions may not be naturally aligned on page boundary.
7683 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7685 if (spanned_pages > present_pages + (present_pages >> 4) &&
7686 IS_ENABLED(CONFIG_SPARSEMEM))
7687 pages = present_pages;
7689 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7692 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7693 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7695 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7697 spin_lock_init(&ds_queue->split_queue_lock);
7698 INIT_LIST_HEAD(&ds_queue->split_queue);
7699 ds_queue->split_queue_len = 0;
7702 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7705 #ifdef CONFIG_COMPACTION
7706 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7708 init_waitqueue_head(&pgdat->kcompactd_wait);
7711 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7714 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7718 pgdat_resize_init(pgdat);
7719 pgdat_kswapd_lock_init(pgdat);
7721 pgdat_init_split_queue(pgdat);
7722 pgdat_init_kcompactd(pgdat);
7724 init_waitqueue_head(&pgdat->kswapd_wait);
7725 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7727 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7728 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7730 pgdat_page_ext_init(pgdat);
7731 lruvec_init(&pgdat->__lruvec);
7734 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7735 unsigned long remaining_pages)
7737 atomic_long_set(&zone->managed_pages, remaining_pages);
7738 zone_set_nid(zone, nid);
7739 zone->name = zone_names[idx];
7740 zone->zone_pgdat = NODE_DATA(nid);
7741 spin_lock_init(&zone->lock);
7742 zone_seqlock_init(zone);
7743 zone_pcp_init(zone);
7747 * Set up the zone data structures
7748 * - init pgdat internals
7749 * - init all zones belonging to this node
7751 * NOTE: this function is only called during memory hotplug
7753 #ifdef CONFIG_MEMORY_HOTPLUG
7754 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7756 int nid = pgdat->node_id;
7760 pgdat_init_internals(pgdat);
7762 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7763 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7766 * Reset the nr_zones, order and highest_zoneidx before reuse.
7767 * Note that kswapd will init kswapd_highest_zoneidx properly
7768 * when it starts in the near future.
7770 pgdat->nr_zones = 0;
7771 pgdat->kswapd_order = 0;
7772 pgdat->kswapd_highest_zoneidx = 0;
7773 pgdat->node_start_pfn = 0;
7774 for_each_online_cpu(cpu) {
7775 struct per_cpu_nodestat *p;
7777 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7778 memset(p, 0, sizeof(*p));
7781 for (z = 0; z < MAX_NR_ZONES; z++)
7782 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7787 * Set up the zone data structures:
7788 * - mark all pages reserved
7789 * - mark all memory queues empty
7790 * - clear the memory bitmaps
7792 * NOTE: pgdat should get zeroed by caller.
7793 * NOTE: this function is only called during early init.
7795 static void __init free_area_init_core(struct pglist_data *pgdat)
7798 int nid = pgdat->node_id;
7800 pgdat_init_internals(pgdat);
7801 pgdat->per_cpu_nodestats = &boot_nodestats;
7803 for (j = 0; j < MAX_NR_ZONES; j++) {
7804 struct zone *zone = pgdat->node_zones + j;
7805 unsigned long size, freesize, memmap_pages;
7807 size = zone->spanned_pages;
7808 freesize = zone->present_pages;
7811 * Adjust freesize so that it accounts for how much memory
7812 * is used by this zone for memmap. This affects the watermark
7813 * and per-cpu initialisations
7815 memmap_pages = calc_memmap_size(size, freesize);
7816 if (!is_highmem_idx(j)) {
7817 if (freesize >= memmap_pages) {
7818 freesize -= memmap_pages;
7820 pr_debug(" %s zone: %lu pages used for memmap\n",
7821 zone_names[j], memmap_pages);
7823 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7824 zone_names[j], memmap_pages, freesize);
7827 /* Account for reserved pages */
7828 if (j == 0 && freesize > dma_reserve) {
7829 freesize -= dma_reserve;
7830 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7833 if (!is_highmem_idx(j))
7834 nr_kernel_pages += freesize;
7835 /* Charge for highmem memmap if there are enough kernel pages */
7836 else if (nr_kernel_pages > memmap_pages * 2)
7837 nr_kernel_pages -= memmap_pages;
7838 nr_all_pages += freesize;
7841 * Set an approximate value for lowmem here, it will be adjusted
7842 * when the bootmem allocator frees pages into the buddy system.
7843 * And all highmem pages will be managed by the buddy system.
7845 zone_init_internals(zone, j, nid, freesize);
7850 set_pageblock_order();
7852 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7856 #ifdef CONFIG_FLATMEM
7857 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7859 unsigned long __maybe_unused start = 0;
7860 unsigned long __maybe_unused offset = 0;
7862 /* Skip empty nodes */
7863 if (!pgdat->node_spanned_pages)
7866 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7867 offset = pgdat->node_start_pfn - start;
7868 /* ia64 gets its own node_mem_map, before this, without bootmem */
7869 if (!pgdat->node_mem_map) {
7870 unsigned long size, end;
7874 * The zone's endpoints aren't required to be MAX_ORDER
7875 * aligned but the node_mem_map endpoints must be in order
7876 * for the buddy allocator to function correctly.
7878 end = pgdat_end_pfn(pgdat);
7879 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7880 size = (end - start) * sizeof(struct page);
7881 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7882 pgdat->node_id, false);
7884 panic("Failed to allocate %ld bytes for node %d memory map\n",
7885 size, pgdat->node_id);
7886 pgdat->node_mem_map = map + offset;
7888 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7889 __func__, pgdat->node_id, (unsigned long)pgdat,
7890 (unsigned long)pgdat->node_mem_map);
7893 * With no DISCONTIG, the global mem_map is just set as node 0's
7895 if (pgdat == NODE_DATA(0)) {
7896 mem_map = NODE_DATA(0)->node_mem_map;
7897 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7903 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7904 #endif /* CONFIG_FLATMEM */
7906 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7907 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7909 pgdat->first_deferred_pfn = ULONG_MAX;
7912 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7915 static void __init free_area_init_node(int nid)
7917 pg_data_t *pgdat = NODE_DATA(nid);
7918 unsigned long start_pfn = 0;
7919 unsigned long end_pfn = 0;
7921 /* pg_data_t should be reset to zero when it's allocated */
7922 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7924 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7926 pgdat->node_id = nid;
7927 pgdat->node_start_pfn = start_pfn;
7928 pgdat->per_cpu_nodestats = NULL;
7930 if (start_pfn != end_pfn) {
7931 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7932 (u64)start_pfn << PAGE_SHIFT,
7933 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7935 pr_info("Initmem setup node %d as memoryless\n", nid);
7938 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7940 alloc_node_mem_map(pgdat);
7941 pgdat_set_deferred_range(pgdat);
7943 free_area_init_core(pgdat);
7946 static void __init free_area_init_memoryless_node(int nid)
7948 free_area_init_node(nid);
7951 #if MAX_NUMNODES > 1
7953 * Figure out the number of possible node ids.
7955 void __init setup_nr_node_ids(void)
7957 unsigned int highest;
7959 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7960 nr_node_ids = highest + 1;
7965 * node_map_pfn_alignment - determine the maximum internode alignment
7967 * This function should be called after node map is populated and sorted.
7968 * It calculates the maximum power of two alignment which can distinguish
7971 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7972 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7973 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7974 * shifted, 1GiB is enough and this function will indicate so.
7976 * This is used to test whether pfn -> nid mapping of the chosen memory
7977 * model has fine enough granularity to avoid incorrect mapping for the
7978 * populated node map.
7980 * Return: the determined alignment in pfn's. 0 if there is no alignment
7981 * requirement (single node).
7983 unsigned long __init node_map_pfn_alignment(void)
7985 unsigned long accl_mask = 0, last_end = 0;
7986 unsigned long start, end, mask;
7987 int last_nid = NUMA_NO_NODE;
7990 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7991 if (!start || last_nid < 0 || last_nid == nid) {
7998 * Start with a mask granular enough to pin-point to the
7999 * start pfn and tick off bits one-by-one until it becomes
8000 * too coarse to separate the current node from the last.
8002 mask = ~((1 << __ffs(start)) - 1);
8003 while (mask && last_end <= (start & (mask << 1)))
8006 /* accumulate all internode masks */
8010 /* convert mask to number of pages */
8011 return ~accl_mask + 1;
8015 * early_calculate_totalpages()
8016 * Sum pages in active regions for movable zone.
8017 * Populate N_MEMORY for calculating usable_nodes.
8019 static unsigned long __init early_calculate_totalpages(void)
8021 unsigned long totalpages = 0;
8022 unsigned long start_pfn, end_pfn;
8025 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8026 unsigned long pages = end_pfn - start_pfn;
8028 totalpages += pages;
8030 node_set_state(nid, N_MEMORY);
8036 * Find the PFN the Movable zone begins in each node. Kernel memory
8037 * is spread evenly between nodes as long as the nodes have enough
8038 * memory. When they don't, some nodes will have more kernelcore than
8041 static void __init find_zone_movable_pfns_for_nodes(void)
8044 unsigned long usable_startpfn;
8045 unsigned long kernelcore_node, kernelcore_remaining;
8046 /* save the state before borrow the nodemask */
8047 nodemask_t saved_node_state = node_states[N_MEMORY];
8048 unsigned long totalpages = early_calculate_totalpages();
8049 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
8050 struct memblock_region *r;
8052 /* Need to find movable_zone earlier when movable_node is specified. */
8053 find_usable_zone_for_movable();
8056 * If movable_node is specified, ignore kernelcore and movablecore
8059 if (movable_node_is_enabled()) {
8060 for_each_mem_region(r) {
8061 if (!memblock_is_hotpluggable(r))
8064 nid = memblock_get_region_node(r);
8066 usable_startpfn = PFN_DOWN(r->base);
8067 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
8068 min(usable_startpfn, zone_movable_pfn[nid]) :
8076 * If kernelcore=mirror is specified, ignore movablecore option
8078 if (mirrored_kernelcore) {
8079 bool mem_below_4gb_not_mirrored = false;
8081 for_each_mem_region(r) {
8082 if (memblock_is_mirror(r))
8085 nid = memblock_get_region_node(r);
8087 usable_startpfn = memblock_region_memory_base_pfn(r);
8089 if (usable_startpfn < PHYS_PFN(SZ_4G)) {
8090 mem_below_4gb_not_mirrored = true;
8094 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
8095 min(usable_startpfn, zone_movable_pfn[nid]) :
8099 if (mem_below_4gb_not_mirrored)
8100 pr_warn("This configuration results in unmirrored kernel memory.\n");
8106 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
8107 * amount of necessary memory.
8109 if (required_kernelcore_percent)
8110 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
8112 if (required_movablecore_percent)
8113 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
8117 * If movablecore= was specified, calculate what size of
8118 * kernelcore that corresponds so that memory usable for
8119 * any allocation type is evenly spread. If both kernelcore
8120 * and movablecore are specified, then the value of kernelcore
8121 * will be used for required_kernelcore if it's greater than
8122 * what movablecore would have allowed.
8124 if (required_movablecore) {
8125 unsigned long corepages;
8128 * Round-up so that ZONE_MOVABLE is at least as large as what
8129 * was requested by the user
8131 required_movablecore =
8132 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
8133 required_movablecore = min(totalpages, required_movablecore);
8134 corepages = totalpages - required_movablecore;
8136 required_kernelcore = max(required_kernelcore, corepages);
8140 * If kernelcore was not specified or kernelcore size is larger
8141 * than totalpages, there is no ZONE_MOVABLE.
8143 if (!required_kernelcore || required_kernelcore >= totalpages)
8146 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
8147 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
8150 /* Spread kernelcore memory as evenly as possible throughout nodes */
8151 kernelcore_node = required_kernelcore / usable_nodes;
8152 for_each_node_state(nid, N_MEMORY) {
8153 unsigned long start_pfn, end_pfn;
8156 * Recalculate kernelcore_node if the division per node
8157 * now exceeds what is necessary to satisfy the requested
8158 * amount of memory for the kernel
8160 if (required_kernelcore < kernelcore_node)
8161 kernelcore_node = required_kernelcore / usable_nodes;
8164 * As the map is walked, we track how much memory is usable
8165 * by the kernel using kernelcore_remaining. When it is
8166 * 0, the rest of the node is usable by ZONE_MOVABLE
8168 kernelcore_remaining = kernelcore_node;
8170 /* Go through each range of PFNs within this node */
8171 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
8172 unsigned long size_pages;
8174 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
8175 if (start_pfn >= end_pfn)
8178 /* Account for what is only usable for kernelcore */
8179 if (start_pfn < usable_startpfn) {
8180 unsigned long kernel_pages;
8181 kernel_pages = min(end_pfn, usable_startpfn)
8184 kernelcore_remaining -= min(kernel_pages,
8185 kernelcore_remaining);
8186 required_kernelcore -= min(kernel_pages,
8187 required_kernelcore);
8189 /* Continue if range is now fully accounted */
8190 if (end_pfn <= usable_startpfn) {
8193 * Push zone_movable_pfn to the end so
8194 * that if we have to rebalance
8195 * kernelcore across nodes, we will
8196 * not double account here
8198 zone_movable_pfn[nid] = end_pfn;
8201 start_pfn = usable_startpfn;
8205 * The usable PFN range for ZONE_MOVABLE is from
8206 * start_pfn->end_pfn. Calculate size_pages as the
8207 * number of pages used as kernelcore
8209 size_pages = end_pfn - start_pfn;
8210 if (size_pages > kernelcore_remaining)
8211 size_pages = kernelcore_remaining;
8212 zone_movable_pfn[nid] = start_pfn + size_pages;
8215 * Some kernelcore has been met, update counts and
8216 * break if the kernelcore for this node has been
8219 required_kernelcore -= min(required_kernelcore,
8221 kernelcore_remaining -= size_pages;
8222 if (!kernelcore_remaining)
8228 * If there is still required_kernelcore, we do another pass with one
8229 * less node in the count. This will push zone_movable_pfn[nid] further
8230 * along on the nodes that still have memory until kernelcore is
8234 if (usable_nodes && required_kernelcore > usable_nodes)
8238 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
8239 for (nid = 0; nid < MAX_NUMNODES; nid++) {
8240 unsigned long start_pfn, end_pfn;
8242 zone_movable_pfn[nid] =
8243 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
8245 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
8246 if (zone_movable_pfn[nid] >= end_pfn)
8247 zone_movable_pfn[nid] = 0;
8251 /* restore the node_state */
8252 node_states[N_MEMORY] = saved_node_state;
8255 /* Any regular or high memory on that node ? */
8256 static void check_for_memory(pg_data_t *pgdat, int nid)
8258 enum zone_type zone_type;
8260 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8261 struct zone *zone = &pgdat->node_zones[zone_type];
8262 if (populated_zone(zone)) {
8263 if (IS_ENABLED(CONFIG_HIGHMEM))
8264 node_set_state(nid, N_HIGH_MEMORY);
8265 if (zone_type <= ZONE_NORMAL)
8266 node_set_state(nid, N_NORMAL_MEMORY);
8273 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8274 * such cases we allow max_zone_pfn sorted in the descending order
8276 bool __weak arch_has_descending_max_zone_pfns(void)
8282 * free_area_init - Initialise all pg_data_t and zone data
8283 * @max_zone_pfn: an array of max PFNs for each zone
8285 * This will call free_area_init_node() for each active node in the system.
8286 * Using the page ranges provided by memblock_set_node(), the size of each
8287 * zone in each node and their holes is calculated. If the maximum PFN
8288 * between two adjacent zones match, it is assumed that the zone is empty.
8289 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8290 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8291 * starts where the previous one ended. For example, ZONE_DMA32 starts
8292 * at arch_max_dma_pfn.
8294 void __init free_area_init(unsigned long *max_zone_pfn)
8296 unsigned long start_pfn, end_pfn;
8300 /* Record where the zone boundaries are */
8301 memset(arch_zone_lowest_possible_pfn, 0,
8302 sizeof(arch_zone_lowest_possible_pfn));
8303 memset(arch_zone_highest_possible_pfn, 0,
8304 sizeof(arch_zone_highest_possible_pfn));
8306 start_pfn = PHYS_PFN(memblock_start_of_DRAM());
8307 descending = arch_has_descending_max_zone_pfns();
8309 for (i = 0; i < MAX_NR_ZONES; i++) {
8311 zone = MAX_NR_ZONES - i - 1;
8315 if (zone == ZONE_MOVABLE)
8318 end_pfn = max(max_zone_pfn[zone], start_pfn);
8319 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8320 arch_zone_highest_possible_pfn[zone] = end_pfn;
8322 start_pfn = end_pfn;
8325 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8326 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8327 find_zone_movable_pfns_for_nodes();
8329 /* Print out the zone ranges */
8330 pr_info("Zone ranges:\n");
8331 for (i = 0; i < MAX_NR_ZONES; i++) {
8332 if (i == ZONE_MOVABLE)
8334 pr_info(" %-8s ", zone_names[i]);
8335 if (arch_zone_lowest_possible_pfn[i] ==
8336 arch_zone_highest_possible_pfn[i])
8339 pr_cont("[mem %#018Lx-%#018Lx]\n",
8340 (u64)arch_zone_lowest_possible_pfn[i]
8342 ((u64)arch_zone_highest_possible_pfn[i]
8343 << PAGE_SHIFT) - 1);
8346 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8347 pr_info("Movable zone start for each node\n");
8348 for (i = 0; i < MAX_NUMNODES; i++) {
8349 if (zone_movable_pfn[i])
8350 pr_info(" Node %d: %#018Lx\n", i,
8351 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8355 * Print out the early node map, and initialize the
8356 * subsection-map relative to active online memory ranges to
8357 * enable future "sub-section" extensions of the memory map.
8359 pr_info("Early memory node ranges\n");
8360 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8361 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8362 (u64)start_pfn << PAGE_SHIFT,
8363 ((u64)end_pfn << PAGE_SHIFT) - 1);
8364 subsection_map_init(start_pfn, end_pfn - start_pfn);
8367 /* Initialise every node */
8368 mminit_verify_pageflags_layout();
8369 setup_nr_node_ids();
8370 for_each_node(nid) {
8373 if (!node_online(nid)) {
8374 pr_info("Initializing node %d as memoryless\n", nid);
8376 /* Allocator not initialized yet */
8377 pgdat = arch_alloc_nodedata(nid);
8379 pr_err("Cannot allocate %zuB for node %d.\n",
8380 sizeof(*pgdat), nid);
8383 arch_refresh_nodedata(nid, pgdat);
8384 free_area_init_memoryless_node(nid);
8387 * We do not want to confuse userspace by sysfs
8388 * files/directories for node without any memory
8389 * attached to it, so this node is not marked as
8390 * N_MEMORY and not marked online so that no sysfs
8391 * hierarchy will be created via register_one_node for
8392 * it. The pgdat will get fully initialized by
8393 * hotadd_init_pgdat() when memory is hotplugged into
8399 pgdat = NODE_DATA(nid);
8400 free_area_init_node(nid);
8402 /* Any memory on that node */
8403 if (pgdat->node_present_pages)
8404 node_set_state(nid, N_MEMORY);
8405 check_for_memory(pgdat, nid);
8411 static int __init cmdline_parse_core(char *p, unsigned long *core,
8412 unsigned long *percent)
8414 unsigned long long coremem;
8420 /* Value may be a percentage of total memory, otherwise bytes */
8421 coremem = simple_strtoull(p, &endptr, 0);
8422 if (*endptr == '%') {
8423 /* Paranoid check for percent values greater than 100 */
8424 WARN_ON(coremem > 100);
8428 coremem = memparse(p, &p);
8429 /* Paranoid check that UL is enough for the coremem value */
8430 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8432 *core = coremem >> PAGE_SHIFT;
8439 * kernelcore=size sets the amount of memory for use for allocations that
8440 * cannot be reclaimed or migrated.
8442 static int __init cmdline_parse_kernelcore(char *p)
8444 /* parse kernelcore=mirror */
8445 if (parse_option_str(p, "mirror")) {
8446 mirrored_kernelcore = true;
8450 return cmdline_parse_core(p, &required_kernelcore,
8451 &required_kernelcore_percent);
8455 * movablecore=size sets the amount of memory for use for allocations that
8456 * can be reclaimed or migrated.
8458 static int __init cmdline_parse_movablecore(char *p)
8460 return cmdline_parse_core(p, &required_movablecore,
8461 &required_movablecore_percent);
8464 early_param("kernelcore", cmdline_parse_kernelcore);
8465 early_param("movablecore", cmdline_parse_movablecore);
8467 void adjust_managed_page_count(struct page *page, long count)
8469 atomic_long_add(count, &page_zone(page)->managed_pages);
8470 totalram_pages_add(count);
8471 #ifdef CONFIG_HIGHMEM
8472 if (PageHighMem(page))
8473 totalhigh_pages_add(count);
8476 EXPORT_SYMBOL(adjust_managed_page_count);
8478 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8481 unsigned long pages = 0;
8483 start = (void *)PAGE_ALIGN((unsigned long)start);
8484 end = (void *)((unsigned long)end & PAGE_MASK);
8485 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8486 struct page *page = virt_to_page(pos);
8487 void *direct_map_addr;
8490 * 'direct_map_addr' might be different from 'pos'
8491 * because some architectures' virt_to_page()
8492 * work with aliases. Getting the direct map
8493 * address ensures that we get a _writeable_
8494 * alias for the memset().
8496 direct_map_addr = page_address(page);
8498 * Perform a kasan-unchecked memset() since this memory
8499 * has not been initialized.
8501 direct_map_addr = kasan_reset_tag(direct_map_addr);
8502 if ((unsigned int)poison <= 0xFF)
8503 memset(direct_map_addr, poison, PAGE_SIZE);
8505 free_reserved_page(page);
8509 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8514 void __init mem_init_print_info(void)
8516 unsigned long physpages, codesize, datasize, rosize, bss_size;
8517 unsigned long init_code_size, init_data_size;
8519 physpages = get_num_physpages();
8520 codesize = _etext - _stext;
8521 datasize = _edata - _sdata;
8522 rosize = __end_rodata - __start_rodata;
8523 bss_size = __bss_stop - __bss_start;
8524 init_data_size = __init_end - __init_begin;
8525 init_code_size = _einittext - _sinittext;
8528 * Detect special cases and adjust section sizes accordingly:
8529 * 1) .init.* may be embedded into .data sections
8530 * 2) .init.text.* may be out of [__init_begin, __init_end],
8531 * please refer to arch/tile/kernel/vmlinux.lds.S.
8532 * 3) .rodata.* may be embedded into .text or .data sections.
8534 #define adj_init_size(start, end, size, pos, adj) \
8536 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8540 adj_init_size(__init_begin, __init_end, init_data_size,
8541 _sinittext, init_code_size);
8542 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8543 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8544 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8545 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8547 #undef adj_init_size
8549 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8550 #ifdef CONFIG_HIGHMEM
8554 K(nr_free_pages()), K(physpages),
8555 codesize / SZ_1K, datasize / SZ_1K, rosize / SZ_1K,
8556 (init_data_size + init_code_size) / SZ_1K, bss_size / SZ_1K,
8557 K(physpages - totalram_pages() - totalcma_pages),
8559 #ifdef CONFIG_HIGHMEM
8560 , K(totalhigh_pages())
8566 * set_dma_reserve - set the specified number of pages reserved in the first zone
8567 * @new_dma_reserve: The number of pages to mark reserved
8569 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8570 * In the DMA zone, a significant percentage may be consumed by kernel image
8571 * and other unfreeable allocations which can skew the watermarks badly. This
8572 * function may optionally be used to account for unfreeable pages in the
8573 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8574 * smaller per-cpu batchsize.
8576 void __init set_dma_reserve(unsigned long new_dma_reserve)
8578 dma_reserve = new_dma_reserve;
8581 static int page_alloc_cpu_dead(unsigned int cpu)
8585 lru_add_drain_cpu(cpu);
8586 mlock_page_drain_remote(cpu);
8590 * Spill the event counters of the dead processor
8591 * into the current processors event counters.
8592 * This artificially elevates the count of the current
8595 vm_events_fold_cpu(cpu);
8598 * Zero the differential counters of the dead processor
8599 * so that the vm statistics are consistent.
8601 * This is only okay since the processor is dead and cannot
8602 * race with what we are doing.
8604 cpu_vm_stats_fold(cpu);
8606 for_each_populated_zone(zone)
8607 zone_pcp_update(zone, 0);
8612 static int page_alloc_cpu_online(unsigned int cpu)
8616 for_each_populated_zone(zone)
8617 zone_pcp_update(zone, 1);
8622 int hashdist = HASHDIST_DEFAULT;
8624 static int __init set_hashdist(char *str)
8628 hashdist = simple_strtoul(str, &str, 0);
8631 __setup("hashdist=", set_hashdist);
8634 void __init page_alloc_init(void)
8639 if (num_node_state(N_MEMORY) == 1)
8643 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8644 "mm/page_alloc:pcp",
8645 page_alloc_cpu_online,
8646 page_alloc_cpu_dead);
8651 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8652 * or min_free_kbytes changes.
8654 static void calculate_totalreserve_pages(void)
8656 struct pglist_data *pgdat;
8657 unsigned long reserve_pages = 0;
8658 enum zone_type i, j;
8660 for_each_online_pgdat(pgdat) {
8662 pgdat->totalreserve_pages = 0;
8664 for (i = 0; i < MAX_NR_ZONES; i++) {
8665 struct zone *zone = pgdat->node_zones + i;
8667 unsigned long managed_pages = zone_managed_pages(zone);
8669 /* Find valid and maximum lowmem_reserve in the zone */
8670 for (j = i; j < MAX_NR_ZONES; j++) {
8671 if (zone->lowmem_reserve[j] > max)
8672 max = zone->lowmem_reserve[j];
8675 /* we treat the high watermark as reserved pages. */
8676 max += high_wmark_pages(zone);
8678 if (max > managed_pages)
8679 max = managed_pages;
8681 pgdat->totalreserve_pages += max;
8683 reserve_pages += max;
8686 totalreserve_pages = reserve_pages;
8690 * setup_per_zone_lowmem_reserve - called whenever
8691 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8692 * has a correct pages reserved value, so an adequate number of
8693 * pages are left in the zone after a successful __alloc_pages().
8695 static void setup_per_zone_lowmem_reserve(void)
8697 struct pglist_data *pgdat;
8698 enum zone_type i, j;
8700 for_each_online_pgdat(pgdat) {
8701 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8702 struct zone *zone = &pgdat->node_zones[i];
8703 int ratio = sysctl_lowmem_reserve_ratio[i];
8704 bool clear = !ratio || !zone_managed_pages(zone);
8705 unsigned long managed_pages = 0;
8707 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8708 struct zone *upper_zone = &pgdat->node_zones[j];
8710 managed_pages += zone_managed_pages(upper_zone);
8713 zone->lowmem_reserve[j] = 0;
8715 zone->lowmem_reserve[j] = managed_pages / ratio;
8720 /* update totalreserve_pages */
8721 calculate_totalreserve_pages();
8724 static void __setup_per_zone_wmarks(void)
8726 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8727 unsigned long lowmem_pages = 0;
8729 unsigned long flags;
8731 /* Calculate total number of !ZONE_HIGHMEM pages */
8732 for_each_zone(zone) {
8733 if (!is_highmem(zone))
8734 lowmem_pages += zone_managed_pages(zone);
8737 for_each_zone(zone) {
8740 spin_lock_irqsave(&zone->lock, flags);
8741 tmp = (u64)pages_min * zone_managed_pages(zone);
8742 do_div(tmp, lowmem_pages);
8743 if (is_highmem(zone)) {
8745 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8746 * need highmem pages, so cap pages_min to a small
8749 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8750 * deltas control async page reclaim, and so should
8751 * not be capped for highmem.
8753 unsigned long min_pages;
8755 min_pages = zone_managed_pages(zone) / 1024;
8756 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8757 zone->_watermark[WMARK_MIN] = min_pages;
8760 * If it's a lowmem zone, reserve a number of pages
8761 * proportionate to the zone's size.
8763 zone->_watermark[WMARK_MIN] = tmp;
8767 * Set the kswapd watermarks distance according to the
8768 * scale factor in proportion to available memory, but
8769 * ensure a minimum size on small systems.
8771 tmp = max_t(u64, tmp >> 2,
8772 mult_frac(zone_managed_pages(zone),
8773 watermark_scale_factor, 10000));
8775 zone->watermark_boost = 0;
8776 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8777 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8778 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8780 spin_unlock_irqrestore(&zone->lock, flags);
8783 /* update totalreserve_pages */
8784 calculate_totalreserve_pages();
8788 * setup_per_zone_wmarks - called when min_free_kbytes changes
8789 * or when memory is hot-{added|removed}
8791 * Ensures that the watermark[min,low,high] values for each zone are set
8792 * correctly with respect to min_free_kbytes.
8794 void setup_per_zone_wmarks(void)
8797 static DEFINE_SPINLOCK(lock);
8800 __setup_per_zone_wmarks();
8804 * The watermark size have changed so update the pcpu batch
8805 * and high limits or the limits may be inappropriate.
8808 zone_pcp_update(zone, 0);
8812 * Initialise min_free_kbytes.
8814 * For small machines we want it small (128k min). For large machines
8815 * we want it large (256MB max). But it is not linear, because network
8816 * bandwidth does not increase linearly with machine size. We use
8818 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8819 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8835 void calculate_min_free_kbytes(void)
8837 unsigned long lowmem_kbytes;
8838 int new_min_free_kbytes;
8840 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8841 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8843 if (new_min_free_kbytes > user_min_free_kbytes)
8844 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8846 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8847 new_min_free_kbytes, user_min_free_kbytes);
8851 int __meminit init_per_zone_wmark_min(void)
8853 calculate_min_free_kbytes();
8854 setup_per_zone_wmarks();
8855 refresh_zone_stat_thresholds();
8856 setup_per_zone_lowmem_reserve();
8859 setup_min_unmapped_ratio();
8860 setup_min_slab_ratio();
8863 khugepaged_min_free_kbytes_update();
8867 postcore_initcall(init_per_zone_wmark_min)
8870 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8871 * that we can call two helper functions whenever min_free_kbytes
8874 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8875 void *buffer, size_t *length, loff_t *ppos)
8879 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8884 user_min_free_kbytes = min_free_kbytes;
8885 setup_per_zone_wmarks();
8890 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8891 void *buffer, size_t *length, loff_t *ppos)
8895 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8900 setup_per_zone_wmarks();
8906 static void setup_min_unmapped_ratio(void)
8911 for_each_online_pgdat(pgdat)
8912 pgdat->min_unmapped_pages = 0;
8915 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8916 sysctl_min_unmapped_ratio) / 100;
8920 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8921 void *buffer, size_t *length, loff_t *ppos)
8925 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8929 setup_min_unmapped_ratio();
8934 static void setup_min_slab_ratio(void)
8939 for_each_online_pgdat(pgdat)
8940 pgdat->min_slab_pages = 0;
8943 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8944 sysctl_min_slab_ratio) / 100;
8947 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8948 void *buffer, size_t *length, loff_t *ppos)
8952 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8956 setup_min_slab_ratio();
8963 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8964 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8965 * whenever sysctl_lowmem_reserve_ratio changes.
8967 * The reserve ratio obviously has absolutely no relation with the
8968 * minimum watermarks. The lowmem reserve ratio can only make sense
8969 * if in function of the boot time zone sizes.
8971 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8972 void *buffer, size_t *length, loff_t *ppos)
8976 proc_dointvec_minmax(table, write, buffer, length, ppos);
8978 for (i = 0; i < MAX_NR_ZONES; i++) {
8979 if (sysctl_lowmem_reserve_ratio[i] < 1)
8980 sysctl_lowmem_reserve_ratio[i] = 0;
8983 setup_per_zone_lowmem_reserve();
8988 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8989 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8990 * pagelist can have before it gets flushed back to buddy allocator.
8992 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8993 int write, void *buffer, size_t *length, loff_t *ppos)
8996 int old_percpu_pagelist_high_fraction;
8999 mutex_lock(&pcp_batch_high_lock);
9000 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
9002 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
9003 if (!write || ret < 0)
9006 /* Sanity checking to avoid pcp imbalance */
9007 if (percpu_pagelist_high_fraction &&
9008 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
9009 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
9015 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
9018 for_each_populated_zone(zone)
9019 zone_set_pageset_high_and_batch(zone, 0);
9021 mutex_unlock(&pcp_batch_high_lock);
9025 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
9027 * Returns the number of pages that arch has reserved but
9028 * is not known to alloc_large_system_hash().
9030 static unsigned long __init arch_reserved_kernel_pages(void)
9037 * Adaptive scale is meant to reduce sizes of hash tables on large memory
9038 * machines. As memory size is increased the scale is also increased but at
9039 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
9040 * quadruples the scale is increased by one, which means the size of hash table
9041 * only doubles, instead of quadrupling as well.
9042 * Because 32-bit systems cannot have large physical memory, where this scaling
9043 * makes sense, it is disabled on such platforms.
9045 #if __BITS_PER_LONG > 32
9046 #define ADAPT_SCALE_BASE (64ul << 30)
9047 #define ADAPT_SCALE_SHIFT 2
9048 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
9052 * allocate a large system hash table from bootmem
9053 * - it is assumed that the hash table must contain an exact power-of-2
9054 * quantity of entries
9055 * - limit is the number of hash buckets, not the total allocation size
9057 void *__init alloc_large_system_hash(const char *tablename,
9058 unsigned long bucketsize,
9059 unsigned long numentries,
9062 unsigned int *_hash_shift,
9063 unsigned int *_hash_mask,
9064 unsigned long low_limit,
9065 unsigned long high_limit)
9067 unsigned long long max = high_limit;
9068 unsigned long log2qty, size;
9074 /* allow the kernel cmdline to have a say */
9076 /* round applicable memory size up to nearest megabyte */
9077 numentries = nr_kernel_pages;
9078 numentries -= arch_reserved_kernel_pages();
9080 /* It isn't necessary when PAGE_SIZE >= 1MB */
9081 if (PAGE_SIZE < SZ_1M)
9082 numentries = round_up(numentries, SZ_1M / PAGE_SIZE);
9084 #if __BITS_PER_LONG > 32
9086 unsigned long adapt;
9088 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
9089 adapt <<= ADAPT_SCALE_SHIFT)
9094 /* limit to 1 bucket per 2^scale bytes of low memory */
9095 if (scale > PAGE_SHIFT)
9096 numentries >>= (scale - PAGE_SHIFT);
9098 numentries <<= (PAGE_SHIFT - scale);
9100 /* Make sure we've got at least a 0-order allocation.. */
9101 if (unlikely(flags & HASH_SMALL)) {
9102 /* Makes no sense without HASH_EARLY */
9103 WARN_ON(!(flags & HASH_EARLY));
9104 if (!(numentries >> *_hash_shift)) {
9105 numentries = 1UL << *_hash_shift;
9106 BUG_ON(!numentries);
9108 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
9109 numentries = PAGE_SIZE / bucketsize;
9111 numentries = roundup_pow_of_two(numentries);
9113 /* limit allocation size to 1/16 total memory by default */
9115 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
9116 do_div(max, bucketsize);
9118 max = min(max, 0x80000000ULL);
9120 if (numentries < low_limit)
9121 numentries = low_limit;
9122 if (numentries > max)
9125 log2qty = ilog2(numentries);
9127 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
9130 size = bucketsize << log2qty;
9131 if (flags & HASH_EARLY) {
9132 if (flags & HASH_ZERO)
9133 table = memblock_alloc(size, SMP_CACHE_BYTES);
9135 table = memblock_alloc_raw(size,
9137 } else if (get_order(size) >= MAX_ORDER || hashdist) {
9138 table = vmalloc_huge(size, gfp_flags);
9141 huge = is_vm_area_hugepages(table);
9144 * If bucketsize is not a power-of-two, we may free
9145 * some pages at the end of hash table which
9146 * alloc_pages_exact() automatically does
9148 table = alloc_pages_exact(size, gfp_flags);
9149 kmemleak_alloc(table, size, 1, gfp_flags);
9151 } while (!table && size > PAGE_SIZE && --log2qty);
9154 panic("Failed to allocate %s hash table\n", tablename);
9156 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
9157 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
9158 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
9161 *_hash_shift = log2qty;
9163 *_hash_mask = (1 << log2qty) - 1;
9168 #ifdef CONFIG_CONTIG_ALLOC
9169 #if defined(CONFIG_DYNAMIC_DEBUG) || \
9170 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
9171 /* Usage: See admin-guide/dynamic-debug-howto.rst */
9172 static void alloc_contig_dump_pages(struct list_head *page_list)
9174 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9176 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9180 list_for_each_entry(page, page_list, lru)
9181 dump_page(page, "migration failure");
9185 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9190 /* [start, end) must belong to a single zone. */
9191 int __alloc_contig_migrate_range(struct compact_control *cc,
9192 unsigned long start, unsigned long end)
9194 /* This function is based on compact_zone() from compaction.c. */
9195 unsigned int nr_reclaimed;
9196 unsigned long pfn = start;
9197 unsigned int tries = 0;
9199 struct migration_target_control mtc = {
9200 .nid = zone_to_nid(cc->zone),
9201 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9204 lru_cache_disable();
9206 while (pfn < end || !list_empty(&cc->migratepages)) {
9207 if (fatal_signal_pending(current)) {
9212 if (list_empty(&cc->migratepages)) {
9213 cc->nr_migratepages = 0;
9214 ret = isolate_migratepages_range(cc, pfn, end);
9215 if (ret && ret != -EAGAIN)
9217 pfn = cc->migrate_pfn;
9219 } else if (++tries == 5) {
9224 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9226 cc->nr_migratepages -= nr_reclaimed;
9228 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9229 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9232 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9233 * to retry again over this error, so do the same here.
9241 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
9242 alloc_contig_dump_pages(&cc->migratepages);
9243 putback_movable_pages(&cc->migratepages);
9250 * alloc_contig_range() -- tries to allocate given range of pages
9251 * @start: start PFN to allocate
9252 * @end: one-past-the-last PFN to allocate
9253 * @migratetype: migratetype of the underlying pageblocks (either
9254 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9255 * in range must have the same migratetype and it must
9256 * be either of the two.
9257 * @gfp_mask: GFP mask to use during compaction
9259 * The PFN range does not have to be pageblock aligned. The PFN range must
9260 * belong to a single zone.
9262 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9263 * pageblocks in the range. Once isolated, the pageblocks should not
9264 * be modified by others.
9266 * Return: zero on success or negative error code. On success all
9267 * pages which PFN is in [start, end) are allocated for the caller and
9268 * need to be freed with free_contig_range().
9270 int alloc_contig_range(unsigned long start, unsigned long end,
9271 unsigned migratetype, gfp_t gfp_mask)
9273 unsigned long outer_start, outer_end;
9277 struct compact_control cc = {
9278 .nr_migratepages = 0,
9280 .zone = page_zone(pfn_to_page(start)),
9281 .mode = MIGRATE_SYNC,
9282 .ignore_skip_hint = true,
9283 .no_set_skip_hint = true,
9284 .gfp_mask = current_gfp_context(gfp_mask),
9285 .alloc_contig = true,
9287 INIT_LIST_HEAD(&cc.migratepages);
9290 * What we do here is we mark all pageblocks in range as
9291 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9292 * have different sizes, and due to the way page allocator
9293 * work, start_isolate_page_range() has special handlings for this.
9295 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9296 * migrate the pages from an unaligned range (ie. pages that
9297 * we are interested in). This will put all the pages in
9298 * range back to page allocator as MIGRATE_ISOLATE.
9300 * When this is done, we take the pages in range from page
9301 * allocator removing them from the buddy system. This way
9302 * page allocator will never consider using them.
9304 * This lets us mark the pageblocks back as
9305 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9306 * aligned range but not in the unaligned, original range are
9307 * put back to page allocator so that buddy can use them.
9310 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
9314 drain_all_pages(cc.zone);
9317 * In case of -EBUSY, we'd like to know which page causes problem.
9318 * So, just fall through. test_pages_isolated() has a tracepoint
9319 * which will report the busy page.
9321 * It is possible that busy pages could become available before
9322 * the call to test_pages_isolated, and the range will actually be
9323 * allocated. So, if we fall through be sure to clear ret so that
9324 * -EBUSY is not accidentally used or returned to caller.
9326 ret = __alloc_contig_migrate_range(&cc, start, end);
9327 if (ret && ret != -EBUSY)
9332 * Pages from [start, end) are within a pageblock_nr_pages
9333 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9334 * more, all pages in [start, end) are free in page allocator.
9335 * What we are going to do is to allocate all pages from
9336 * [start, end) (that is remove them from page allocator).
9338 * The only problem is that pages at the beginning and at the
9339 * end of interesting range may be not aligned with pages that
9340 * page allocator holds, ie. they can be part of higher order
9341 * pages. Because of this, we reserve the bigger range and
9342 * once this is done free the pages we are not interested in.
9344 * We don't have to hold zone->lock here because the pages are
9345 * isolated thus they won't get removed from buddy.
9349 outer_start = start;
9350 while (!PageBuddy(pfn_to_page(outer_start))) {
9351 if (++order >= MAX_ORDER) {
9352 outer_start = start;
9355 outer_start &= ~0UL << order;
9358 if (outer_start != start) {
9359 order = buddy_order(pfn_to_page(outer_start));
9362 * outer_start page could be small order buddy page and
9363 * it doesn't include start page. Adjust outer_start
9364 * in this case to report failed page properly
9365 * on tracepoint in test_pages_isolated()
9367 if (outer_start + (1UL << order) <= start)
9368 outer_start = start;
9371 /* Make sure the range is really isolated. */
9372 if (test_pages_isolated(outer_start, end, 0)) {
9377 /* Grab isolated pages from freelists. */
9378 outer_end = isolate_freepages_range(&cc, outer_start, end);
9384 /* Free head and tail (if any) */
9385 if (start != outer_start)
9386 free_contig_range(outer_start, start - outer_start);
9387 if (end != outer_end)
9388 free_contig_range(end, outer_end - end);
9391 undo_isolate_page_range(start, end, migratetype);
9394 EXPORT_SYMBOL(alloc_contig_range);
9396 static int __alloc_contig_pages(unsigned long start_pfn,
9397 unsigned long nr_pages, gfp_t gfp_mask)
9399 unsigned long end_pfn = start_pfn + nr_pages;
9401 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9405 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9406 unsigned long nr_pages)
9408 unsigned long i, end_pfn = start_pfn + nr_pages;
9411 for (i = start_pfn; i < end_pfn; i++) {
9412 page = pfn_to_online_page(i);
9416 if (page_zone(page) != z)
9419 if (PageReserved(page))
9428 static bool zone_spans_last_pfn(const struct zone *zone,
9429 unsigned long start_pfn, unsigned long nr_pages)
9431 unsigned long last_pfn = start_pfn + nr_pages - 1;
9433 return zone_spans_pfn(zone, last_pfn);
9437 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9438 * @nr_pages: Number of contiguous pages to allocate
9439 * @gfp_mask: GFP mask to limit search and used during compaction
9441 * @nodemask: Mask for other possible nodes
9443 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9444 * on an applicable zonelist to find a contiguous pfn range which can then be
9445 * tried for allocation with alloc_contig_range(). This routine is intended
9446 * for allocation requests which can not be fulfilled with the buddy allocator.
9448 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9449 * power of two, then allocated range is also guaranteed to be aligned to same
9450 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9452 * Allocated pages can be freed with free_contig_range() or by manually calling
9453 * __free_page() on each allocated page.
9455 * Return: pointer to contiguous pages on success, or NULL if not successful.
9457 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9458 int nid, nodemask_t *nodemask)
9460 unsigned long ret, pfn, flags;
9461 struct zonelist *zonelist;
9465 zonelist = node_zonelist(nid, gfp_mask);
9466 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9467 gfp_zone(gfp_mask), nodemask) {
9468 spin_lock_irqsave(&zone->lock, flags);
9470 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9471 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9472 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9474 * We release the zone lock here because
9475 * alloc_contig_range() will also lock the zone
9476 * at some point. If there's an allocation
9477 * spinning on this lock, it may win the race
9478 * and cause alloc_contig_range() to fail...
9480 spin_unlock_irqrestore(&zone->lock, flags);
9481 ret = __alloc_contig_pages(pfn, nr_pages,
9484 return pfn_to_page(pfn);
9485 spin_lock_irqsave(&zone->lock, flags);
9489 spin_unlock_irqrestore(&zone->lock, flags);
9493 #endif /* CONFIG_CONTIG_ALLOC */
9495 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9497 unsigned long count = 0;
9499 for (; nr_pages--; pfn++) {
9500 struct page *page = pfn_to_page(pfn);
9502 count += page_count(page) != 1;
9505 WARN(count != 0, "%lu pages are still in use!\n", count);
9507 EXPORT_SYMBOL(free_contig_range);
9510 * Effectively disable pcplists for the zone by setting the high limit to 0
9511 * and draining all cpus. A concurrent page freeing on another CPU that's about
9512 * to put the page on pcplist will either finish before the drain and the page
9513 * will be drained, or observe the new high limit and skip the pcplist.
9515 * Must be paired with a call to zone_pcp_enable().
9517 void zone_pcp_disable(struct zone *zone)
9519 mutex_lock(&pcp_batch_high_lock);
9520 __zone_set_pageset_high_and_batch(zone, 0, 1);
9521 __drain_all_pages(zone, true);
9524 void zone_pcp_enable(struct zone *zone)
9526 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9527 mutex_unlock(&pcp_batch_high_lock);
9530 void zone_pcp_reset(struct zone *zone)
9533 struct per_cpu_zonestat *pzstats;
9535 if (zone->per_cpu_pageset != &boot_pageset) {
9536 for_each_online_cpu(cpu) {
9537 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9538 drain_zonestat(zone, pzstats);
9540 free_percpu(zone->per_cpu_pageset);
9541 zone->per_cpu_pageset = &boot_pageset;
9542 if (zone->per_cpu_zonestats != &boot_zonestats) {
9543 free_percpu(zone->per_cpu_zonestats);
9544 zone->per_cpu_zonestats = &boot_zonestats;
9549 #ifdef CONFIG_MEMORY_HOTREMOVE
9551 * All pages in the range must be in a single zone, must not contain holes,
9552 * must span full sections, and must be isolated before calling this function.
9554 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9556 unsigned long pfn = start_pfn;
9560 unsigned long flags;
9562 offline_mem_sections(pfn, end_pfn);
9563 zone = page_zone(pfn_to_page(pfn));
9564 spin_lock_irqsave(&zone->lock, flags);
9565 while (pfn < end_pfn) {
9566 page = pfn_to_page(pfn);
9568 * The HWPoisoned page may be not in buddy system, and
9569 * page_count() is not 0.
9571 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9576 * At this point all remaining PageOffline() pages have a
9577 * reference count of 0 and can simply be skipped.
9579 if (PageOffline(page)) {
9580 BUG_ON(page_count(page));
9581 BUG_ON(PageBuddy(page));
9586 BUG_ON(page_count(page));
9587 BUG_ON(!PageBuddy(page));
9588 order = buddy_order(page);
9589 del_page_from_free_list(page, zone, order);
9590 pfn += (1 << order);
9592 spin_unlock_irqrestore(&zone->lock, flags);
9597 * This function returns a stable result only if called under zone lock.
9599 bool is_free_buddy_page(struct page *page)
9601 unsigned long pfn = page_to_pfn(page);
9604 for (order = 0; order < MAX_ORDER; order++) {
9605 struct page *page_head = page - (pfn & ((1 << order) - 1));
9607 if (PageBuddy(page_head) &&
9608 buddy_order_unsafe(page_head) >= order)
9612 return order < MAX_ORDER;
9614 EXPORT_SYMBOL(is_free_buddy_page);
9616 #ifdef CONFIG_MEMORY_FAILURE
9618 * Break down a higher-order page in sub-pages, and keep our target out of
9621 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9622 struct page *target, int low, int high,
9625 unsigned long size = 1 << high;
9626 struct page *current_buddy, *next_page;
9628 while (high > low) {
9632 if (target >= &page[size]) {
9633 next_page = page + size;
9634 current_buddy = page;
9637 current_buddy = page + size;
9640 if (set_page_guard(zone, current_buddy, high, migratetype))
9643 if (current_buddy != target) {
9644 add_to_free_list(current_buddy, zone, high, migratetype);
9645 set_buddy_order(current_buddy, high);
9652 * Take a page that will be marked as poisoned off the buddy allocator.
9654 bool take_page_off_buddy(struct page *page)
9656 struct zone *zone = page_zone(page);
9657 unsigned long pfn = page_to_pfn(page);
9658 unsigned long flags;
9662 spin_lock_irqsave(&zone->lock, flags);
9663 for (order = 0; order < MAX_ORDER; order++) {
9664 struct page *page_head = page - (pfn & ((1 << order) - 1));
9665 int page_order = buddy_order(page_head);
9667 if (PageBuddy(page_head) && page_order >= order) {
9668 unsigned long pfn_head = page_to_pfn(page_head);
9669 int migratetype = get_pfnblock_migratetype(page_head,
9672 del_page_from_free_list(page_head, zone, page_order);
9673 break_down_buddy_pages(zone, page_head, page, 0,
9674 page_order, migratetype);
9675 SetPageHWPoisonTakenOff(page);
9676 if (!is_migrate_isolate(migratetype))
9677 __mod_zone_freepage_state(zone, -1, migratetype);
9681 if (page_count(page_head) > 0)
9684 spin_unlock_irqrestore(&zone->lock, flags);
9689 * Cancel takeoff done by take_page_off_buddy().
9691 bool put_page_back_buddy(struct page *page)
9693 struct zone *zone = page_zone(page);
9694 unsigned long pfn = page_to_pfn(page);
9695 unsigned long flags;
9696 int migratetype = get_pfnblock_migratetype(page, pfn);
9699 spin_lock_irqsave(&zone->lock, flags);
9700 if (put_page_testzero(page)) {
9701 ClearPageHWPoisonTakenOff(page);
9702 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9703 if (TestClearPageHWPoison(page)) {
9707 spin_unlock_irqrestore(&zone->lock, flags);
9713 #ifdef CONFIG_ZONE_DMA
9714 bool has_managed_dma(void)
9716 struct pglist_data *pgdat;
9718 for_each_online_pgdat(pgdat) {
9719 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9721 if (managed_zone(zone))
9726 #endif /* CONFIG_ZONE_DMA */