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 init = want_init_on_free();
1407 VM_BUG_ON_PAGE(PageTail(page), page);
1409 trace_mm_page_free(page, order);
1410 kmsan_free_page(page, order);
1412 if (unlikely(PageHWPoison(page)) && !order) {
1414 * Do not let hwpoison pages hit pcplists/buddy
1415 * Untie memcg state and reset page's owner
1417 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1418 __memcg_kmem_uncharge_page(page, order);
1419 reset_page_owner(page, order);
1420 page_table_check_free(page, order);
1425 * Check tail pages before head page information is cleared to
1426 * avoid checking PageCompound for order-0 pages.
1428 if (unlikely(order)) {
1429 bool compound = PageCompound(page);
1432 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1435 ClearPageDoubleMap(page);
1436 ClearPageHasHWPoisoned(page);
1438 for (i = 1; i < (1 << order); i++) {
1440 bad += free_tail_pages_check(page, page + i);
1441 if (unlikely(free_page_is_bad(page + i))) {
1445 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1448 if (PageMappingFlags(page))
1449 page->mapping = NULL;
1450 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1451 __memcg_kmem_uncharge_page(page, order);
1452 if (check_free && free_page_is_bad(page))
1457 page_cpupid_reset_last(page);
1458 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1459 reset_page_owner(page, order);
1460 page_table_check_free(page, order);
1462 if (!PageHighMem(page)) {
1463 debug_check_no_locks_freed(page_address(page),
1464 PAGE_SIZE << order);
1465 debug_check_no_obj_freed(page_address(page),
1466 PAGE_SIZE << order);
1469 kernel_poison_pages(page, 1 << order);
1472 * As memory initialization might be integrated into KASAN,
1473 * KASAN poisoning and memory initialization code must be
1474 * kept together to avoid discrepancies in behavior.
1476 * With hardware tag-based KASAN, memory tags must be set before the
1477 * page becomes unavailable via debug_pagealloc or arch_free_page.
1479 if (!should_skip_kasan_poison(page, fpi_flags)) {
1480 kasan_poison_pages(page, order, init);
1482 /* Memory is already initialized if KASAN did it internally. */
1483 if (kasan_has_integrated_init())
1487 kernel_init_pages(page, 1 << order);
1490 * arch_free_page() can make the page's contents inaccessible. s390
1491 * does this. So nothing which can access the page's contents should
1492 * happen after this.
1494 arch_free_page(page, order);
1496 debug_pagealloc_unmap_pages(page, 1 << order);
1501 #ifdef CONFIG_DEBUG_VM
1503 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1504 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1505 * moved from pcp lists to free lists.
1507 static bool free_pcp_prepare(struct page *page, unsigned int order)
1509 return free_pages_prepare(page, order, true, FPI_NONE);
1512 /* return true if this page has an inappropriate state */
1513 static bool bulkfree_pcp_prepare(struct page *page)
1515 if (debug_pagealloc_enabled_static())
1516 return free_page_is_bad(page);
1522 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1523 * moving from pcp lists to free list in order to reduce overhead. With
1524 * debug_pagealloc enabled, they are checked also immediately when being freed
1527 static bool free_pcp_prepare(struct page *page, unsigned int order)
1529 if (debug_pagealloc_enabled_static())
1530 return free_pages_prepare(page, order, true, FPI_NONE);
1532 return free_pages_prepare(page, order, false, FPI_NONE);
1535 static bool bulkfree_pcp_prepare(struct page *page)
1537 return free_page_is_bad(page);
1539 #endif /* CONFIG_DEBUG_VM */
1542 * Frees a number of pages from the PCP lists
1543 * Assumes all pages on list are in same zone.
1544 * count is the number of pages to free.
1546 static void free_pcppages_bulk(struct zone *zone, int count,
1547 struct per_cpu_pages *pcp,
1551 int max_pindex = NR_PCP_LISTS - 1;
1553 bool isolated_pageblocks;
1557 * Ensure proper count is passed which otherwise would stuck in the
1558 * below while (list_empty(list)) loop.
1560 count = min(pcp->count, count);
1562 /* Ensure requested pindex is drained first. */
1563 pindex = pindex - 1;
1565 /* Caller must hold IRQ-safe pcp->lock so IRQs are disabled. */
1566 spin_lock(&zone->lock);
1567 isolated_pageblocks = has_isolate_pageblock(zone);
1570 struct list_head *list;
1573 /* Remove pages from lists in a round-robin fashion. */
1575 if (++pindex > max_pindex)
1576 pindex = min_pindex;
1577 list = &pcp->lists[pindex];
1578 if (!list_empty(list))
1581 if (pindex == max_pindex)
1583 if (pindex == min_pindex)
1587 order = pindex_to_order(pindex);
1588 nr_pages = 1 << order;
1592 page = list_last_entry(list, struct page, pcp_list);
1593 mt = get_pcppage_migratetype(page);
1595 /* must delete to avoid corrupting pcp list */
1596 list_del(&page->pcp_list);
1598 pcp->count -= nr_pages;
1600 if (bulkfree_pcp_prepare(page))
1603 /* MIGRATE_ISOLATE page should not go to pcplists */
1604 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1605 /* Pageblock could have been isolated meanwhile */
1606 if (unlikely(isolated_pageblocks))
1607 mt = get_pageblock_migratetype(page);
1609 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1610 trace_mm_page_pcpu_drain(page, order, mt);
1611 } while (count > 0 && !list_empty(list));
1614 spin_unlock(&zone->lock);
1617 static void free_one_page(struct zone *zone,
1618 struct page *page, unsigned long pfn,
1620 int migratetype, fpi_t fpi_flags)
1622 unsigned long flags;
1624 spin_lock_irqsave(&zone->lock, flags);
1625 if (unlikely(has_isolate_pageblock(zone) ||
1626 is_migrate_isolate(migratetype))) {
1627 migratetype = get_pfnblock_migratetype(page, pfn);
1629 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1630 spin_unlock_irqrestore(&zone->lock, flags);
1633 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1634 unsigned long zone, int nid)
1636 mm_zero_struct_page(page);
1637 set_page_links(page, zone, nid, pfn);
1638 init_page_count(page);
1639 page_mapcount_reset(page);
1640 page_cpupid_reset_last(page);
1641 page_kasan_tag_reset(page);
1643 INIT_LIST_HEAD(&page->lru);
1644 #ifdef WANT_PAGE_VIRTUAL
1645 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1646 if (!is_highmem_idx(zone))
1647 set_page_address(page, __va(pfn << PAGE_SHIFT));
1651 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1652 static void __meminit init_reserved_page(unsigned long pfn)
1657 if (!early_page_uninitialised(pfn))
1660 nid = early_pfn_to_nid(pfn);
1661 pgdat = NODE_DATA(nid);
1663 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1664 struct zone *zone = &pgdat->node_zones[zid];
1666 if (zone_spans_pfn(zone, pfn))
1669 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1672 static inline void init_reserved_page(unsigned long pfn)
1675 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1678 * Initialised pages do not have PageReserved set. This function is
1679 * called for each range allocated by the bootmem allocator and
1680 * marks the pages PageReserved. The remaining valid pages are later
1681 * sent to the buddy page allocator.
1683 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1685 unsigned long start_pfn = PFN_DOWN(start);
1686 unsigned long end_pfn = PFN_UP(end);
1688 for (; start_pfn < end_pfn; start_pfn++) {
1689 if (pfn_valid(start_pfn)) {
1690 struct page *page = pfn_to_page(start_pfn);
1692 init_reserved_page(start_pfn);
1694 /* Avoid false-positive PageTail() */
1695 INIT_LIST_HEAD(&page->lru);
1698 * no need for atomic set_bit because the struct
1699 * page is not visible yet so nobody should
1702 __SetPageReserved(page);
1707 static void __free_pages_ok(struct page *page, unsigned int order,
1710 unsigned long flags;
1712 unsigned long pfn = page_to_pfn(page);
1713 struct zone *zone = page_zone(page);
1715 if (!free_pages_prepare(page, order, true, fpi_flags))
1718 migratetype = get_pfnblock_migratetype(page, pfn);
1720 spin_lock_irqsave(&zone->lock, flags);
1721 if (unlikely(has_isolate_pageblock(zone) ||
1722 is_migrate_isolate(migratetype))) {
1723 migratetype = get_pfnblock_migratetype(page, pfn);
1725 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1726 spin_unlock_irqrestore(&zone->lock, flags);
1728 __count_vm_events(PGFREE, 1 << order);
1731 void __free_pages_core(struct page *page, unsigned int order)
1733 unsigned int nr_pages = 1 << order;
1734 struct page *p = page;
1738 * When initializing the memmap, __init_single_page() sets the refcount
1739 * of all pages to 1 ("allocated"/"not free"). We have to set the
1740 * refcount of all involved pages to 0.
1743 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1745 __ClearPageReserved(p);
1746 set_page_count(p, 0);
1748 __ClearPageReserved(p);
1749 set_page_count(p, 0);
1751 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1754 * Bypass PCP and place fresh pages right to the tail, primarily
1755 * relevant for memory onlining.
1757 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1763 * During memory init memblocks map pfns to nids. The search is expensive and
1764 * this caches recent lookups. The implementation of __early_pfn_to_nid
1765 * treats start/end as pfns.
1767 struct mminit_pfnnid_cache {
1768 unsigned long last_start;
1769 unsigned long last_end;
1773 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1776 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1778 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1779 struct mminit_pfnnid_cache *state)
1781 unsigned long start_pfn, end_pfn;
1784 if (state->last_start <= pfn && pfn < state->last_end)
1785 return state->last_nid;
1787 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1788 if (nid != NUMA_NO_NODE) {
1789 state->last_start = start_pfn;
1790 state->last_end = end_pfn;
1791 state->last_nid = nid;
1797 int __meminit early_pfn_to_nid(unsigned long pfn)
1799 static DEFINE_SPINLOCK(early_pfn_lock);
1802 spin_lock(&early_pfn_lock);
1803 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1805 nid = first_online_node;
1806 spin_unlock(&early_pfn_lock);
1810 #endif /* CONFIG_NUMA */
1812 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1815 if (early_page_uninitialised(pfn))
1817 if (!kmsan_memblock_free_pages(page, order)) {
1818 /* KMSAN will take care of these pages. */
1821 __free_pages_core(page, order);
1825 * Check that the whole (or subset of) a pageblock given by the interval of
1826 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1827 * with the migration of free compaction scanner.
1829 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1831 * It's possible on some configurations to have a setup like node0 node1 node0
1832 * i.e. it's possible that all pages within a zones range of pages do not
1833 * belong to a single zone. We assume that a border between node0 and node1
1834 * can occur within a single pageblock, but not a node0 node1 node0
1835 * interleaving within a single pageblock. It is therefore sufficient to check
1836 * the first and last page of a pageblock and avoid checking each individual
1837 * page in a pageblock.
1839 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1840 unsigned long end_pfn, struct zone *zone)
1842 struct page *start_page;
1843 struct page *end_page;
1845 /* end_pfn is one past the range we are checking */
1848 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1851 start_page = pfn_to_online_page(start_pfn);
1855 if (page_zone(start_page) != zone)
1858 end_page = pfn_to_page(end_pfn);
1860 /* This gives a shorter code than deriving page_zone(end_page) */
1861 if (page_zone_id(start_page) != page_zone_id(end_page))
1867 void set_zone_contiguous(struct zone *zone)
1869 unsigned long block_start_pfn = zone->zone_start_pfn;
1870 unsigned long block_end_pfn;
1872 block_end_pfn = pageblock_end_pfn(block_start_pfn);
1873 for (; block_start_pfn < zone_end_pfn(zone);
1874 block_start_pfn = block_end_pfn,
1875 block_end_pfn += pageblock_nr_pages) {
1877 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1879 if (!__pageblock_pfn_to_page(block_start_pfn,
1880 block_end_pfn, zone))
1885 /* We confirm that there is no hole */
1886 zone->contiguous = true;
1889 void clear_zone_contiguous(struct zone *zone)
1891 zone->contiguous = false;
1894 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1895 static void __init deferred_free_range(unsigned long pfn,
1896 unsigned long nr_pages)
1904 page = pfn_to_page(pfn);
1906 /* Free a large naturally-aligned chunk if possible */
1907 if (nr_pages == pageblock_nr_pages && pageblock_aligned(pfn)) {
1908 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1909 __free_pages_core(page, pageblock_order);
1913 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1914 if (pageblock_aligned(pfn))
1915 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1916 __free_pages_core(page, 0);
1920 /* Completion tracking for deferred_init_memmap() threads */
1921 static atomic_t pgdat_init_n_undone __initdata;
1922 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1924 static inline void __init pgdat_init_report_one_done(void)
1926 if (atomic_dec_and_test(&pgdat_init_n_undone))
1927 complete(&pgdat_init_all_done_comp);
1931 * Returns true if page needs to be initialized or freed to buddy allocator.
1933 * We check if a current large page is valid by only checking the validity
1936 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1938 if (pageblock_aligned(pfn) && !pfn_valid(pfn))
1944 * Free pages to buddy allocator. Try to free aligned pages in
1945 * pageblock_nr_pages sizes.
1947 static void __init deferred_free_pages(unsigned long pfn,
1948 unsigned long end_pfn)
1950 unsigned long nr_free = 0;
1952 for (; pfn < end_pfn; pfn++) {
1953 if (!deferred_pfn_valid(pfn)) {
1954 deferred_free_range(pfn - nr_free, nr_free);
1956 } else if (pageblock_aligned(pfn)) {
1957 deferred_free_range(pfn - nr_free, nr_free);
1963 /* Free the last block of pages to allocator */
1964 deferred_free_range(pfn - nr_free, nr_free);
1968 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1969 * by performing it only once every pageblock_nr_pages.
1970 * Return number of pages initialized.
1972 static unsigned long __init deferred_init_pages(struct zone *zone,
1974 unsigned long end_pfn)
1976 int nid = zone_to_nid(zone);
1977 unsigned long nr_pages = 0;
1978 int zid = zone_idx(zone);
1979 struct page *page = NULL;
1981 for (; pfn < end_pfn; pfn++) {
1982 if (!deferred_pfn_valid(pfn)) {
1985 } else if (!page || pageblock_aligned(pfn)) {
1986 page = pfn_to_page(pfn);
1990 __init_single_page(page, pfn, zid, nid);
1997 * This function is meant to pre-load the iterator for the zone init.
1998 * Specifically it walks through the ranges until we are caught up to the
1999 * first_init_pfn value and exits there. If we never encounter the value we
2000 * return false indicating there are no valid ranges left.
2003 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
2004 unsigned long *spfn, unsigned long *epfn,
2005 unsigned long first_init_pfn)
2010 * Start out by walking through the ranges in this zone that have
2011 * already been initialized. We don't need to do anything with them
2012 * so we just need to flush them out of the system.
2014 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
2015 if (*epfn <= first_init_pfn)
2017 if (*spfn < first_init_pfn)
2018 *spfn = first_init_pfn;
2027 * Initialize and free pages. We do it in two loops: first we initialize
2028 * struct page, then free to buddy allocator, because while we are
2029 * freeing pages we can access pages that are ahead (computing buddy
2030 * page in __free_one_page()).
2032 * In order to try and keep some memory in the cache we have the loop
2033 * broken along max page order boundaries. This way we will not cause
2034 * any issues with the buddy page computation.
2036 static unsigned long __init
2037 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
2038 unsigned long *end_pfn)
2040 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
2041 unsigned long spfn = *start_pfn, epfn = *end_pfn;
2042 unsigned long nr_pages = 0;
2045 /* First we loop through and initialize the page values */
2046 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
2049 if (mo_pfn <= *start_pfn)
2052 t = min(mo_pfn, *end_pfn);
2053 nr_pages += deferred_init_pages(zone, *start_pfn, t);
2055 if (mo_pfn < *end_pfn) {
2056 *start_pfn = mo_pfn;
2061 /* Reset values and now loop through freeing pages as needed */
2064 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2070 t = min(mo_pfn, epfn);
2071 deferred_free_pages(spfn, t);
2081 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2084 unsigned long spfn, epfn;
2085 struct zone *zone = arg;
2088 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2091 * Initialize and free pages in MAX_ORDER sized increments so that we
2092 * can avoid introducing any issues with the buddy allocator.
2094 while (spfn < end_pfn) {
2095 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2100 /* An arch may override for more concurrency. */
2102 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2107 /* Initialise remaining memory on a node */
2108 static int __init deferred_init_memmap(void *data)
2110 pg_data_t *pgdat = data;
2111 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2112 unsigned long spfn = 0, epfn = 0;
2113 unsigned long first_init_pfn, flags;
2114 unsigned long start = jiffies;
2116 int zid, max_threads;
2119 /* Bind memory initialisation thread to a local node if possible */
2120 if (!cpumask_empty(cpumask))
2121 set_cpus_allowed_ptr(current, cpumask);
2123 pgdat_resize_lock(pgdat, &flags);
2124 first_init_pfn = pgdat->first_deferred_pfn;
2125 if (first_init_pfn == ULONG_MAX) {
2126 pgdat_resize_unlock(pgdat, &flags);
2127 pgdat_init_report_one_done();
2131 /* Sanity check boundaries */
2132 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2133 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2134 pgdat->first_deferred_pfn = ULONG_MAX;
2137 * Once we unlock here, the zone cannot be grown anymore, thus if an
2138 * interrupt thread must allocate this early in boot, zone must be
2139 * pre-grown prior to start of deferred page initialization.
2141 pgdat_resize_unlock(pgdat, &flags);
2143 /* Only the highest zone is deferred so find it */
2144 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2145 zone = pgdat->node_zones + zid;
2146 if (first_init_pfn < zone_end_pfn(zone))
2150 /* If the zone is empty somebody else may have cleared out the zone */
2151 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2155 max_threads = deferred_page_init_max_threads(cpumask);
2157 while (spfn < epfn) {
2158 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2159 struct padata_mt_job job = {
2160 .thread_fn = deferred_init_memmap_chunk,
2163 .size = epfn_align - spfn,
2164 .align = PAGES_PER_SECTION,
2165 .min_chunk = PAGES_PER_SECTION,
2166 .max_threads = max_threads,
2169 padata_do_multithreaded(&job);
2170 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2174 /* Sanity check that the next zone really is unpopulated */
2175 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2177 pr_info("node %d deferred pages initialised in %ums\n",
2178 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2180 pgdat_init_report_one_done();
2185 * If this zone has deferred pages, try to grow it by initializing enough
2186 * deferred pages to satisfy the allocation specified by order, rounded up to
2187 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2188 * of SECTION_SIZE bytes by initializing struct pages in increments of
2189 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2191 * Return true when zone was grown, otherwise return false. We return true even
2192 * when we grow less than requested, to let the caller decide if there are
2193 * enough pages to satisfy the allocation.
2195 * Note: We use noinline because this function is needed only during boot, and
2196 * it is called from a __ref function _deferred_grow_zone. This way we are
2197 * making sure that it is not inlined into permanent text section.
2199 static noinline bool __init
2200 deferred_grow_zone(struct zone *zone, unsigned int order)
2202 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2203 pg_data_t *pgdat = zone->zone_pgdat;
2204 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2205 unsigned long spfn, epfn, flags;
2206 unsigned long nr_pages = 0;
2209 /* Only the last zone may have deferred pages */
2210 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2213 pgdat_resize_lock(pgdat, &flags);
2216 * If someone grew this zone while we were waiting for spinlock, return
2217 * true, as there might be enough pages already.
2219 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2220 pgdat_resize_unlock(pgdat, &flags);
2224 /* If the zone is empty somebody else may have cleared out the zone */
2225 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2226 first_deferred_pfn)) {
2227 pgdat->first_deferred_pfn = ULONG_MAX;
2228 pgdat_resize_unlock(pgdat, &flags);
2229 /* Retry only once. */
2230 return first_deferred_pfn != ULONG_MAX;
2234 * Initialize and free pages in MAX_ORDER sized increments so
2235 * that we can avoid introducing any issues with the buddy
2238 while (spfn < epfn) {
2239 /* update our first deferred PFN for this section */
2240 first_deferred_pfn = spfn;
2242 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2243 touch_nmi_watchdog();
2245 /* We should only stop along section boundaries */
2246 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2249 /* If our quota has been met we can stop here */
2250 if (nr_pages >= nr_pages_needed)
2254 pgdat->first_deferred_pfn = spfn;
2255 pgdat_resize_unlock(pgdat, &flags);
2257 return nr_pages > 0;
2261 * deferred_grow_zone() is __init, but it is called from
2262 * get_page_from_freelist() during early boot until deferred_pages permanently
2263 * disables this call. This is why we have refdata wrapper to avoid warning,
2264 * and to ensure that the function body gets unloaded.
2267 _deferred_grow_zone(struct zone *zone, unsigned int order)
2269 return deferred_grow_zone(zone, order);
2272 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2274 void __init page_alloc_init_late(void)
2279 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2281 /* There will be num_node_state(N_MEMORY) threads */
2282 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2283 for_each_node_state(nid, N_MEMORY) {
2284 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2287 /* Block until all are initialised */
2288 wait_for_completion(&pgdat_init_all_done_comp);
2291 * We initialized the rest of the deferred pages. Permanently disable
2292 * on-demand struct page initialization.
2294 static_branch_disable(&deferred_pages);
2296 /* Reinit limits that are based on free pages after the kernel is up */
2297 files_maxfiles_init();
2302 /* Discard memblock private memory */
2305 for_each_node_state(nid, N_MEMORY)
2306 shuffle_free_memory(NODE_DATA(nid));
2308 for_each_populated_zone(zone)
2309 set_zone_contiguous(zone);
2313 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2314 void __init init_cma_reserved_pageblock(struct page *page)
2316 unsigned i = pageblock_nr_pages;
2317 struct page *p = page;
2320 __ClearPageReserved(p);
2321 set_page_count(p, 0);
2324 set_pageblock_migratetype(page, MIGRATE_CMA);
2325 set_page_refcounted(page);
2326 __free_pages(page, pageblock_order);
2328 adjust_managed_page_count(page, pageblock_nr_pages);
2329 page_zone(page)->cma_pages += pageblock_nr_pages;
2334 * The order of subdivision here is critical for the IO subsystem.
2335 * Please do not alter this order without good reasons and regression
2336 * testing. Specifically, as large blocks of memory are subdivided,
2337 * the order in which smaller blocks are delivered depends on the order
2338 * they're subdivided in this function. This is the primary factor
2339 * influencing the order in which pages are delivered to the IO
2340 * subsystem according to empirical testing, and this is also justified
2341 * by considering the behavior of a buddy system containing a single
2342 * large block of memory acted on by a series of small allocations.
2343 * This behavior is a critical factor in sglist merging's success.
2347 static inline void expand(struct zone *zone, struct page *page,
2348 int low, int high, int migratetype)
2350 unsigned long size = 1 << high;
2352 while (high > low) {
2355 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2358 * Mark as guard pages (or page), that will allow to
2359 * merge back to allocator when buddy will be freed.
2360 * Corresponding page table entries will not be touched,
2361 * pages will stay not present in virtual address space
2363 if (set_page_guard(zone, &page[size], high, migratetype))
2366 add_to_free_list(&page[size], zone, high, migratetype);
2367 set_buddy_order(&page[size], high);
2371 static void check_new_page_bad(struct page *page)
2373 if (unlikely(page->flags & __PG_HWPOISON)) {
2374 /* Don't complain about hwpoisoned pages */
2375 page_mapcount_reset(page); /* remove PageBuddy */
2380 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2384 * This page is about to be returned from the page allocator
2386 static inline int check_new_page(struct page *page)
2388 if (likely(page_expected_state(page,
2389 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2392 check_new_page_bad(page);
2396 static bool check_new_pages(struct page *page, unsigned int order)
2399 for (i = 0; i < (1 << order); i++) {
2400 struct page *p = page + i;
2402 if (unlikely(check_new_page(p)))
2409 #ifdef CONFIG_DEBUG_VM
2411 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2412 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2413 * also checked when pcp lists are refilled from the free lists.
2415 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2417 if (debug_pagealloc_enabled_static())
2418 return check_new_pages(page, order);
2423 static inline bool check_new_pcp(struct page *page, unsigned int order)
2425 return check_new_pages(page, order);
2429 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2430 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2431 * enabled, they are also checked when being allocated from the pcp lists.
2433 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2435 return check_new_pages(page, order);
2437 static inline bool check_new_pcp(struct page *page, unsigned int order)
2439 if (debug_pagealloc_enabled_static())
2440 return check_new_pages(page, order);
2444 #endif /* CONFIG_DEBUG_VM */
2446 static inline bool should_skip_kasan_unpoison(gfp_t flags)
2448 /* Don't skip if a software KASAN mode is enabled. */
2449 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
2450 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
2453 /* Skip, if hardware tag-based KASAN is not enabled. */
2454 if (!kasan_hw_tags_enabled())
2458 * With hardware tag-based KASAN enabled, skip if this has been
2459 * requested via __GFP_SKIP_KASAN_UNPOISON.
2461 return flags & __GFP_SKIP_KASAN_UNPOISON;
2464 static inline bool should_skip_init(gfp_t flags)
2466 /* Don't skip, if hardware tag-based KASAN is not enabled. */
2467 if (!kasan_hw_tags_enabled())
2470 /* For hardware tag-based KASAN, skip if requested. */
2471 return (flags & __GFP_SKIP_ZERO);
2474 inline void post_alloc_hook(struct page *page, unsigned int order,
2477 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
2478 !should_skip_init(gfp_flags);
2479 bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
2482 set_page_private(page, 0);
2483 set_page_refcounted(page);
2485 arch_alloc_page(page, order);
2486 debug_pagealloc_map_pages(page, 1 << order);
2489 * Page unpoisoning must happen before memory initialization.
2490 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2491 * allocations and the page unpoisoning code will complain.
2493 kernel_unpoison_pages(page, 1 << order);
2496 * As memory initialization might be integrated into KASAN,
2497 * KASAN unpoisoning and memory initializion code must be
2498 * kept together to avoid discrepancies in behavior.
2502 * If memory tags should be zeroed (which happens only when memory
2503 * should be initialized as well).
2506 /* Initialize both memory and tags. */
2507 for (i = 0; i != 1 << order; ++i)
2508 tag_clear_highpage(page + i);
2510 /* Note that memory is already initialized by the loop above. */
2513 if (!should_skip_kasan_unpoison(gfp_flags)) {
2514 /* Unpoison shadow memory or set memory tags. */
2515 kasan_unpoison_pages(page, order, init);
2517 /* Note that memory is already initialized by KASAN. */
2518 if (kasan_has_integrated_init())
2521 /* Ensure page_address() dereferencing does not fault. */
2522 for (i = 0; i != 1 << order; ++i)
2523 page_kasan_tag_reset(page + i);
2525 /* If memory is still not initialized, do it now. */
2527 kernel_init_pages(page, 1 << order);
2528 /* Propagate __GFP_SKIP_KASAN_POISON to page flags. */
2529 if (kasan_hw_tags_enabled() && (gfp_flags & __GFP_SKIP_KASAN_POISON))
2530 SetPageSkipKASanPoison(page);
2532 set_page_owner(page, order, gfp_flags);
2533 page_table_check_alloc(page, order);
2536 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2537 unsigned int alloc_flags)
2539 post_alloc_hook(page, order, gfp_flags);
2541 if (order && (gfp_flags & __GFP_COMP))
2542 prep_compound_page(page, order);
2545 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2546 * allocate the page. The expectation is that the caller is taking
2547 * steps that will free more memory. The caller should avoid the page
2548 * being used for !PFMEMALLOC purposes.
2550 if (alloc_flags & ALLOC_NO_WATERMARKS)
2551 set_page_pfmemalloc(page);
2553 clear_page_pfmemalloc(page);
2557 * Go through the free lists for the given migratetype and remove
2558 * the smallest available page from the freelists
2560 static __always_inline
2561 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2564 unsigned int current_order;
2565 struct free_area *area;
2568 /* Find a page of the appropriate size in the preferred list */
2569 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2570 area = &(zone->free_area[current_order]);
2571 page = get_page_from_free_area(area, migratetype);
2574 del_page_from_free_list(page, zone, current_order);
2575 expand(zone, page, order, current_order, migratetype);
2576 set_pcppage_migratetype(page, migratetype);
2577 trace_mm_page_alloc_zone_locked(page, order, migratetype,
2578 pcp_allowed_order(order) &&
2579 migratetype < MIGRATE_PCPTYPES);
2588 * This array describes the order lists are fallen back to when
2589 * the free lists for the desirable migrate type are depleted
2591 * The other migratetypes do not have fallbacks.
2593 static int fallbacks[MIGRATE_TYPES][3] = {
2594 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2595 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2596 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2600 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2603 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2606 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2607 unsigned int order) { return NULL; }
2611 * Move the free pages in a range to the freelist tail of the requested type.
2612 * Note that start_page and end_pages are not aligned on a pageblock
2613 * boundary. If alignment is required, use move_freepages_block()
2615 static int move_freepages(struct zone *zone,
2616 unsigned long start_pfn, unsigned long end_pfn,
2617 int migratetype, int *num_movable)
2622 int pages_moved = 0;
2624 for (pfn = start_pfn; pfn <= end_pfn;) {
2625 page = pfn_to_page(pfn);
2626 if (!PageBuddy(page)) {
2628 * We assume that pages that could be isolated for
2629 * migration are movable. But we don't actually try
2630 * isolating, as that would be expensive.
2633 (PageLRU(page) || __PageMovable(page)))
2639 /* Make sure we are not inadvertently changing nodes */
2640 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2641 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2643 order = buddy_order(page);
2644 move_to_free_list(page, zone, order, migratetype);
2646 pages_moved += 1 << order;
2652 int move_freepages_block(struct zone *zone, struct page *page,
2653 int migratetype, int *num_movable)
2655 unsigned long start_pfn, end_pfn, pfn;
2660 pfn = page_to_pfn(page);
2661 start_pfn = pageblock_start_pfn(pfn);
2662 end_pfn = pageblock_end_pfn(pfn) - 1;
2664 /* Do not cross zone boundaries */
2665 if (!zone_spans_pfn(zone, start_pfn))
2667 if (!zone_spans_pfn(zone, end_pfn))
2670 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2674 static void change_pageblock_range(struct page *pageblock_page,
2675 int start_order, int migratetype)
2677 int nr_pageblocks = 1 << (start_order - pageblock_order);
2679 while (nr_pageblocks--) {
2680 set_pageblock_migratetype(pageblock_page, migratetype);
2681 pageblock_page += pageblock_nr_pages;
2686 * When we are falling back to another migratetype during allocation, try to
2687 * steal extra free pages from the same pageblocks to satisfy further
2688 * allocations, instead of polluting multiple pageblocks.
2690 * If we are stealing a relatively large buddy page, it is likely there will
2691 * be more free pages in the pageblock, so try to steal them all. For
2692 * reclaimable and unmovable allocations, we steal regardless of page size,
2693 * as fragmentation caused by those allocations polluting movable pageblocks
2694 * is worse than movable allocations stealing from unmovable and reclaimable
2697 static bool can_steal_fallback(unsigned int order, int start_mt)
2700 * Leaving this order check is intended, although there is
2701 * relaxed order check in next check. The reason is that
2702 * we can actually steal whole pageblock if this condition met,
2703 * but, below check doesn't guarantee it and that is just heuristic
2704 * so could be changed anytime.
2706 if (order >= pageblock_order)
2709 if (order >= pageblock_order / 2 ||
2710 start_mt == MIGRATE_RECLAIMABLE ||
2711 start_mt == MIGRATE_UNMOVABLE ||
2712 page_group_by_mobility_disabled)
2718 static inline bool boost_watermark(struct zone *zone)
2720 unsigned long max_boost;
2722 if (!watermark_boost_factor)
2725 * Don't bother in zones that are unlikely to produce results.
2726 * On small machines, including kdump capture kernels running
2727 * in a small area, boosting the watermark can cause an out of
2728 * memory situation immediately.
2730 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2733 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2734 watermark_boost_factor, 10000);
2737 * high watermark may be uninitialised if fragmentation occurs
2738 * very early in boot so do not boost. We do not fall
2739 * through and boost by pageblock_nr_pages as failing
2740 * allocations that early means that reclaim is not going
2741 * to help and it may even be impossible to reclaim the
2742 * boosted watermark resulting in a hang.
2747 max_boost = max(pageblock_nr_pages, max_boost);
2749 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2756 * This function implements actual steal behaviour. If order is large enough,
2757 * we can steal whole pageblock. If not, we first move freepages in this
2758 * pageblock to our migratetype and determine how many already-allocated pages
2759 * are there in the pageblock with a compatible migratetype. If at least half
2760 * of pages are free or compatible, we can change migratetype of the pageblock
2761 * itself, so pages freed in the future will be put on the correct free list.
2763 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2764 unsigned int alloc_flags, int start_type, bool whole_block)
2766 unsigned int current_order = buddy_order(page);
2767 int free_pages, movable_pages, alike_pages;
2770 old_block_type = get_pageblock_migratetype(page);
2773 * This can happen due to races and we want to prevent broken
2774 * highatomic accounting.
2776 if (is_migrate_highatomic(old_block_type))
2779 /* Take ownership for orders >= pageblock_order */
2780 if (current_order >= pageblock_order) {
2781 change_pageblock_range(page, current_order, start_type);
2786 * Boost watermarks to increase reclaim pressure to reduce the
2787 * likelihood of future fallbacks. Wake kswapd now as the node
2788 * may be balanced overall and kswapd will not wake naturally.
2790 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2791 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2793 /* We are not allowed to try stealing from the whole block */
2797 free_pages = move_freepages_block(zone, page, start_type,
2800 * Determine how many pages are compatible with our allocation.
2801 * For movable allocation, it's the number of movable pages which
2802 * we just obtained. For other types it's a bit more tricky.
2804 if (start_type == MIGRATE_MOVABLE) {
2805 alike_pages = movable_pages;
2808 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2809 * to MOVABLE pageblock, consider all non-movable pages as
2810 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2811 * vice versa, be conservative since we can't distinguish the
2812 * exact migratetype of non-movable pages.
2814 if (old_block_type == MIGRATE_MOVABLE)
2815 alike_pages = pageblock_nr_pages
2816 - (free_pages + movable_pages);
2821 /* moving whole block can fail due to zone boundary conditions */
2826 * If a sufficient number of pages in the block are either free or of
2827 * comparable migratability as our allocation, claim the whole block.
2829 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2830 page_group_by_mobility_disabled)
2831 set_pageblock_migratetype(page, start_type);
2836 move_to_free_list(page, zone, current_order, start_type);
2840 * Check whether there is a suitable fallback freepage with requested order.
2841 * If only_stealable is true, this function returns fallback_mt only if
2842 * we can steal other freepages all together. This would help to reduce
2843 * fragmentation due to mixed migratetype pages in one pageblock.
2845 int find_suitable_fallback(struct free_area *area, unsigned int order,
2846 int migratetype, bool only_stealable, bool *can_steal)
2851 if (area->nr_free == 0)
2856 fallback_mt = fallbacks[migratetype][i];
2857 if (fallback_mt == MIGRATE_TYPES)
2860 if (free_area_empty(area, fallback_mt))
2863 if (can_steal_fallback(order, migratetype))
2866 if (!only_stealable)
2877 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2878 * there are no empty page blocks that contain a page with a suitable order
2880 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2881 unsigned int alloc_order)
2884 unsigned long max_managed, flags;
2887 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2888 * Check is race-prone but harmless.
2890 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2891 if (zone->nr_reserved_highatomic >= max_managed)
2894 spin_lock_irqsave(&zone->lock, flags);
2896 /* Recheck the nr_reserved_highatomic limit under the lock */
2897 if (zone->nr_reserved_highatomic >= max_managed)
2901 mt = get_pageblock_migratetype(page);
2902 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2903 if (migratetype_is_mergeable(mt)) {
2904 zone->nr_reserved_highatomic += pageblock_nr_pages;
2905 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2906 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2910 spin_unlock_irqrestore(&zone->lock, flags);
2914 * Used when an allocation is about to fail under memory pressure. This
2915 * potentially hurts the reliability of high-order allocations when under
2916 * intense memory pressure but failed atomic allocations should be easier
2917 * to recover from than an OOM.
2919 * If @force is true, try to unreserve a pageblock even though highatomic
2920 * pageblock is exhausted.
2922 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2925 struct zonelist *zonelist = ac->zonelist;
2926 unsigned long flags;
2933 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2936 * Preserve at least one pageblock unless memory pressure
2939 if (!force && zone->nr_reserved_highatomic <=
2943 spin_lock_irqsave(&zone->lock, flags);
2944 for (order = 0; order < MAX_ORDER; order++) {
2945 struct free_area *area = &(zone->free_area[order]);
2947 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2952 * In page freeing path, migratetype change is racy so
2953 * we can counter several free pages in a pageblock
2954 * in this loop although we changed the pageblock type
2955 * from highatomic to ac->migratetype. So we should
2956 * adjust the count once.
2958 if (is_migrate_highatomic_page(page)) {
2960 * It should never happen but changes to
2961 * locking could inadvertently allow a per-cpu
2962 * drain to add pages to MIGRATE_HIGHATOMIC
2963 * while unreserving so be safe and watch for
2966 zone->nr_reserved_highatomic -= min(
2968 zone->nr_reserved_highatomic);
2972 * Convert to ac->migratetype and avoid the normal
2973 * pageblock stealing heuristics. Minimally, the caller
2974 * is doing the work and needs the pages. More
2975 * importantly, if the block was always converted to
2976 * MIGRATE_UNMOVABLE or another type then the number
2977 * of pageblocks that cannot be completely freed
2980 set_pageblock_migratetype(page, ac->migratetype);
2981 ret = move_freepages_block(zone, page, ac->migratetype,
2984 spin_unlock_irqrestore(&zone->lock, flags);
2988 spin_unlock_irqrestore(&zone->lock, flags);
2995 * Try finding a free buddy page on the fallback list and put it on the free
2996 * list of requested migratetype, possibly along with other pages from the same
2997 * block, depending on fragmentation avoidance heuristics. Returns true if
2998 * fallback was found so that __rmqueue_smallest() can grab it.
3000 * The use of signed ints for order and current_order is a deliberate
3001 * deviation from the rest of this file, to make the for loop
3002 * condition simpler.
3004 static __always_inline bool
3005 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
3006 unsigned int alloc_flags)
3008 struct free_area *area;
3010 int min_order = order;
3016 * Do not steal pages from freelists belonging to other pageblocks
3017 * i.e. orders < pageblock_order. If there are no local zones free,
3018 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
3020 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
3021 min_order = pageblock_order;
3024 * Find the largest available free page in the other list. This roughly
3025 * approximates finding the pageblock with the most free pages, which
3026 * would be too costly to do exactly.
3028 for (current_order = MAX_ORDER - 1; current_order >= min_order;
3030 area = &(zone->free_area[current_order]);
3031 fallback_mt = find_suitable_fallback(area, current_order,
3032 start_migratetype, false, &can_steal);
3033 if (fallback_mt == -1)
3037 * We cannot steal all free pages from the pageblock and the
3038 * requested migratetype is movable. In that case it's better to
3039 * steal and split the smallest available page instead of the
3040 * largest available page, because even if the next movable
3041 * allocation falls back into a different pageblock than this
3042 * one, it won't cause permanent fragmentation.
3044 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
3045 && current_order > order)
3054 for (current_order = order; current_order < MAX_ORDER;
3056 area = &(zone->free_area[current_order]);
3057 fallback_mt = find_suitable_fallback(area, current_order,
3058 start_migratetype, false, &can_steal);
3059 if (fallback_mt != -1)
3064 * This should not happen - we already found a suitable fallback
3065 * when looking for the largest page.
3067 VM_BUG_ON(current_order == MAX_ORDER);
3070 page = get_page_from_free_area(area, fallback_mt);
3072 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
3075 trace_mm_page_alloc_extfrag(page, order, current_order,
3076 start_migratetype, fallback_mt);
3083 * Do the hard work of removing an element from the buddy allocator.
3084 * Call me with the zone->lock already held.
3086 static __always_inline struct page *
3087 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
3088 unsigned int alloc_flags)
3092 if (IS_ENABLED(CONFIG_CMA)) {
3094 * Balance movable allocations between regular and CMA areas by
3095 * allocating from CMA when over half of the zone's free memory
3096 * is in the CMA area.
3098 if (alloc_flags & ALLOC_CMA &&
3099 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3100 zone_page_state(zone, NR_FREE_PAGES) / 2) {
3101 page = __rmqueue_cma_fallback(zone, order);
3107 page = __rmqueue_smallest(zone, order, migratetype);
3108 if (unlikely(!page)) {
3109 if (alloc_flags & ALLOC_CMA)
3110 page = __rmqueue_cma_fallback(zone, order);
3112 if (!page && __rmqueue_fallback(zone, order, migratetype,
3120 * Obtain a specified number of elements from the buddy allocator, all under
3121 * a single hold of the lock, for efficiency. Add them to the supplied list.
3122 * Returns the number of new pages which were placed at *list.
3124 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3125 unsigned long count, struct list_head *list,
3126 int migratetype, unsigned int alloc_flags)
3128 int i, allocated = 0;
3130 /* Caller must hold IRQ-safe pcp->lock so IRQs are disabled. */
3131 spin_lock(&zone->lock);
3132 for (i = 0; i < count; ++i) {
3133 struct page *page = __rmqueue(zone, order, migratetype,
3135 if (unlikely(page == NULL))
3138 if (unlikely(check_pcp_refill(page, order)))
3142 * Split buddy pages returned by expand() are received here in
3143 * physical page order. The page is added to the tail of
3144 * caller's list. From the callers perspective, the linked list
3145 * is ordered by page number under some conditions. This is
3146 * useful for IO devices that can forward direction from the
3147 * head, thus also in the physical page order. This is useful
3148 * for IO devices that can merge IO requests if the physical
3149 * pages are ordered properly.
3151 list_add_tail(&page->pcp_list, list);
3153 if (is_migrate_cma(get_pcppage_migratetype(page)))
3154 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3159 * i pages were removed from the buddy list even if some leak due
3160 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3161 * on i. Do not confuse with 'allocated' which is the number of
3162 * pages added to the pcp list.
3164 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3165 spin_unlock(&zone->lock);
3171 * Called from the vmstat counter updater to drain pagesets of this
3172 * currently executing processor on remote nodes after they have
3175 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3177 int to_drain, batch;
3179 batch = READ_ONCE(pcp->batch);
3180 to_drain = min(pcp->count, batch);
3182 unsigned long flags;
3185 * free_pcppages_bulk expects IRQs disabled for zone->lock
3186 * so even though pcp->lock is not intended to be IRQ-safe,
3187 * it's needed in this context.
3189 spin_lock_irqsave(&pcp->lock, flags);
3190 free_pcppages_bulk(zone, to_drain, pcp, 0);
3191 spin_unlock_irqrestore(&pcp->lock, flags);
3197 * Drain pcplists of the indicated processor and zone.
3199 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3201 struct per_cpu_pages *pcp;
3203 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3205 unsigned long flags;
3207 /* See drain_zone_pages on why this is disabling IRQs */
3208 spin_lock_irqsave(&pcp->lock, flags);
3209 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3210 spin_unlock_irqrestore(&pcp->lock, flags);
3215 * Drain pcplists of all zones on the indicated processor.
3217 static void drain_pages(unsigned int cpu)
3221 for_each_populated_zone(zone) {
3222 drain_pages_zone(cpu, zone);
3227 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3229 void drain_local_pages(struct zone *zone)
3231 int cpu = smp_processor_id();
3234 drain_pages_zone(cpu, zone);
3240 * The implementation of drain_all_pages(), exposing an extra parameter to
3241 * drain on all cpus.
3243 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3244 * not empty. The check for non-emptiness can however race with a free to
3245 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3246 * that need the guarantee that every CPU has drained can disable the
3247 * optimizing racy check.
3249 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3254 * Allocate in the BSS so we won't require allocation in
3255 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3257 static cpumask_t cpus_with_pcps;
3260 * Do not drain if one is already in progress unless it's specific to
3261 * a zone. Such callers are primarily CMA and memory hotplug and need
3262 * the drain to be complete when the call returns.
3264 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3267 mutex_lock(&pcpu_drain_mutex);
3271 * We don't care about racing with CPU hotplug event
3272 * as offline notification will cause the notified
3273 * cpu to drain that CPU pcps and on_each_cpu_mask
3274 * disables preemption as part of its processing
3276 for_each_online_cpu(cpu) {
3277 struct per_cpu_pages *pcp;
3279 bool has_pcps = false;
3281 if (force_all_cpus) {
3283 * The pcp.count check is racy, some callers need a
3284 * guarantee that no cpu is missed.
3288 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3292 for_each_populated_zone(z) {
3293 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3302 cpumask_set_cpu(cpu, &cpus_with_pcps);
3304 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3307 for_each_cpu(cpu, &cpus_with_pcps) {
3309 drain_pages_zone(cpu, zone);
3314 mutex_unlock(&pcpu_drain_mutex);
3318 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3320 * When zone parameter is non-NULL, spill just the single zone's pages.
3322 void drain_all_pages(struct zone *zone)
3324 __drain_all_pages(zone, false);
3327 #ifdef CONFIG_HIBERNATION
3330 * Touch the watchdog for every WD_PAGE_COUNT pages.
3332 #define WD_PAGE_COUNT (128*1024)
3334 void mark_free_pages(struct zone *zone)
3336 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3337 unsigned long flags;
3338 unsigned int order, t;
3341 if (zone_is_empty(zone))
3344 spin_lock_irqsave(&zone->lock, flags);
3346 max_zone_pfn = zone_end_pfn(zone);
3347 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3348 if (pfn_valid(pfn)) {
3349 page = pfn_to_page(pfn);
3351 if (!--page_count) {
3352 touch_nmi_watchdog();
3353 page_count = WD_PAGE_COUNT;
3356 if (page_zone(page) != zone)
3359 if (!swsusp_page_is_forbidden(page))
3360 swsusp_unset_page_free(page);
3363 for_each_migratetype_order(order, t) {
3364 list_for_each_entry(page,
3365 &zone->free_area[order].free_list[t], buddy_list) {
3368 pfn = page_to_pfn(page);
3369 for (i = 0; i < (1UL << order); i++) {
3370 if (!--page_count) {
3371 touch_nmi_watchdog();
3372 page_count = WD_PAGE_COUNT;
3374 swsusp_set_page_free(pfn_to_page(pfn + i));
3378 spin_unlock_irqrestore(&zone->lock, flags);
3380 #endif /* CONFIG_PM */
3382 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3387 if (!free_pcp_prepare(page, order))
3390 migratetype = get_pfnblock_migratetype(page, pfn);
3391 set_pcppage_migratetype(page, migratetype);
3395 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3398 int min_nr_free, max_nr_free;
3400 /* Free everything if batch freeing high-order pages. */
3401 if (unlikely(free_high))
3404 /* Check for PCP disabled or boot pageset */
3405 if (unlikely(high < batch))
3408 /* Leave at least pcp->batch pages on the list */
3409 min_nr_free = batch;
3410 max_nr_free = high - batch;
3413 * Double the number of pages freed each time there is subsequent
3414 * freeing of pages without any allocation.
3416 batch <<= pcp->free_factor;
3417 if (batch < max_nr_free)
3419 batch = clamp(batch, min_nr_free, max_nr_free);
3424 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3427 int high = READ_ONCE(pcp->high);
3429 if (unlikely(!high || free_high))
3432 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3436 * If reclaim is active, limit the number of pages that can be
3437 * stored on pcp lists
3439 return min(READ_ONCE(pcp->batch) << 2, high);
3442 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
3443 struct page *page, int migratetype,
3450 __count_vm_events(PGFREE, 1 << order);
3451 pindex = order_to_pindex(migratetype, order);
3452 list_add(&page->pcp_list, &pcp->lists[pindex]);
3453 pcp->count += 1 << order;
3456 * As high-order pages other than THP's stored on PCP can contribute
3457 * to fragmentation, limit the number stored when PCP is heavily
3458 * freeing without allocation. The remainder after bulk freeing
3459 * stops will be drained from vmstat refresh context.
3461 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3463 high = nr_pcp_high(pcp, zone, free_high);
3464 if (pcp->count >= high) {
3465 int batch = READ_ONCE(pcp->batch);
3467 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3474 void free_unref_page(struct page *page, unsigned int order)
3476 unsigned long flags;
3477 unsigned long __maybe_unused UP_flags;
3478 struct per_cpu_pages *pcp;
3480 unsigned long pfn = page_to_pfn(page);
3483 if (!free_unref_page_prepare(page, pfn, order))
3487 * We only track unmovable, reclaimable and movable on pcp lists.
3488 * Place ISOLATE pages on the isolated list because they are being
3489 * offlined but treat HIGHATOMIC as movable pages so we can get those
3490 * areas back if necessary. Otherwise, we may have to free
3491 * excessively into the page allocator
3493 migratetype = get_pcppage_migratetype(page);
3494 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3495 if (unlikely(is_migrate_isolate(migratetype))) {
3496 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3499 migratetype = MIGRATE_MOVABLE;
3502 zone = page_zone(page);
3503 pcp_trylock_prepare(UP_flags);
3504 pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
3506 free_unref_page_commit(zone, pcp, page, migratetype, order);
3507 pcp_spin_unlock_irqrestore(pcp, flags);
3509 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
3511 pcp_trylock_finish(UP_flags);
3515 * Free a list of 0-order pages
3517 void free_unref_page_list(struct list_head *list)
3519 struct page *page, *next;
3520 struct per_cpu_pages *pcp = NULL;
3521 struct zone *locked_zone = NULL;
3522 unsigned long flags;
3523 int batch_count = 0;
3526 /* Prepare pages for freeing */
3527 list_for_each_entry_safe(page, next, list, lru) {
3528 unsigned long pfn = page_to_pfn(page);
3529 if (!free_unref_page_prepare(page, pfn, 0)) {
3530 list_del(&page->lru);
3535 * Free isolated pages directly to the allocator, see
3536 * comment in free_unref_page.
3538 migratetype = get_pcppage_migratetype(page);
3539 if (unlikely(is_migrate_isolate(migratetype))) {
3540 list_del(&page->lru);
3541 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3546 list_for_each_entry_safe(page, next, list, lru) {
3547 struct zone *zone = page_zone(page);
3549 /* Different zone, different pcp lock. */
3550 if (zone != locked_zone) {
3552 pcp_spin_unlock_irqrestore(pcp, flags);
3555 pcp = pcp_spin_lock_irqsave(locked_zone->per_cpu_pageset, flags);
3559 * Non-isolated types over MIGRATE_PCPTYPES get added
3560 * to the MIGRATE_MOVABLE pcp list.
3562 migratetype = get_pcppage_migratetype(page);
3563 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3564 migratetype = MIGRATE_MOVABLE;
3566 trace_mm_page_free_batched(page);
3567 free_unref_page_commit(zone, pcp, page, migratetype, 0);
3570 * Guard against excessive IRQ disabled times when we get
3571 * a large list of pages to free.
3573 if (++batch_count == SWAP_CLUSTER_MAX) {
3574 pcp_spin_unlock_irqrestore(pcp, flags);
3576 pcp = pcp_spin_lock_irqsave(locked_zone->per_cpu_pageset, flags);
3581 pcp_spin_unlock_irqrestore(pcp, flags);
3585 * split_page takes a non-compound higher-order page, and splits it into
3586 * n (1<<order) sub-pages: page[0..n]
3587 * Each sub-page must be freed individually.
3589 * Note: this is probably too low level an operation for use in drivers.
3590 * Please consult with lkml before using this in your driver.
3592 void split_page(struct page *page, unsigned int order)
3596 VM_BUG_ON_PAGE(PageCompound(page), page);
3597 VM_BUG_ON_PAGE(!page_count(page), page);
3599 for (i = 1; i < (1 << order); i++)
3600 set_page_refcounted(page + i);
3601 split_page_owner(page, 1 << order);
3602 split_page_memcg(page, 1 << order);
3604 EXPORT_SYMBOL_GPL(split_page);
3606 int __isolate_free_page(struct page *page, unsigned int order)
3608 struct zone *zone = page_zone(page);
3609 int mt = get_pageblock_migratetype(page);
3611 if (!is_migrate_isolate(mt)) {
3612 unsigned long watermark;
3614 * Obey watermarks as if the page was being allocated. We can
3615 * emulate a high-order watermark check with a raised order-0
3616 * watermark, because we already know our high-order page
3619 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3620 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3623 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3626 del_page_from_free_list(page, zone, order);
3629 * Set the pageblock if the isolated page is at least half of a
3632 if (order >= pageblock_order - 1) {
3633 struct page *endpage = page + (1 << order) - 1;
3634 for (; page < endpage; page += pageblock_nr_pages) {
3635 int mt = get_pageblock_migratetype(page);
3637 * Only change normal pageblocks (i.e., they can merge
3640 if (migratetype_is_mergeable(mt))
3641 set_pageblock_migratetype(page,
3646 return 1UL << order;
3650 * __putback_isolated_page - Return a now-isolated page back where we got it
3651 * @page: Page that was isolated
3652 * @order: Order of the isolated page
3653 * @mt: The page's pageblock's migratetype
3655 * This function is meant to return a page pulled from the free lists via
3656 * __isolate_free_page back to the free lists they were pulled from.
3658 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3660 struct zone *zone = page_zone(page);
3662 /* zone lock should be held when this function is called */
3663 lockdep_assert_held(&zone->lock);
3665 /* Return isolated page to tail of freelist. */
3666 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3667 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3671 * Update NUMA hit/miss statistics
3673 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3677 enum numa_stat_item local_stat = NUMA_LOCAL;
3679 /* skip numa counters update if numa stats is disabled */
3680 if (!static_branch_likely(&vm_numa_stat_key))
3683 if (zone_to_nid(z) != numa_node_id())
3684 local_stat = NUMA_OTHER;
3686 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3687 __count_numa_events(z, NUMA_HIT, nr_account);
3689 __count_numa_events(z, NUMA_MISS, nr_account);
3690 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3692 __count_numa_events(z, local_stat, nr_account);
3696 static __always_inline
3697 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
3698 unsigned int order, unsigned int alloc_flags,
3702 unsigned long flags;
3706 spin_lock_irqsave(&zone->lock, flags);
3708 * order-0 request can reach here when the pcplist is skipped
3709 * due to non-CMA allocation context. HIGHATOMIC area is
3710 * reserved for high-order atomic allocation, so order-0
3711 * request should skip it.
3713 if (order > 0 && alloc_flags & ALLOC_HARDER)
3714 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3716 page = __rmqueue(zone, order, migratetype, alloc_flags);
3718 spin_unlock_irqrestore(&zone->lock, flags);
3722 __mod_zone_freepage_state(zone, -(1 << order),
3723 get_pcppage_migratetype(page));
3724 spin_unlock_irqrestore(&zone->lock, flags);
3725 } while (check_new_pages(page, order));
3727 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3728 zone_statistics(preferred_zone, zone, 1);
3733 /* Remove page from the per-cpu list, caller must protect the list */
3735 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3737 unsigned int alloc_flags,
3738 struct per_cpu_pages *pcp,
3739 struct list_head *list)
3744 if (list_empty(list)) {
3745 int batch = READ_ONCE(pcp->batch);
3749 * Scale batch relative to order if batch implies
3750 * free pages can be stored on the PCP. Batch can
3751 * be 1 for small zones or for boot pagesets which
3752 * should never store free pages as the pages may
3753 * belong to arbitrary zones.
3756 batch = max(batch >> order, 2);
3757 alloced = rmqueue_bulk(zone, order,
3759 migratetype, alloc_flags);
3761 pcp->count += alloced << order;
3762 if (unlikely(list_empty(list)))
3766 page = list_first_entry(list, struct page, pcp_list);
3767 list_del(&page->pcp_list);
3768 pcp->count -= 1 << order;
3769 } while (check_new_pcp(page, order));
3774 /* Lock and remove page from the per-cpu list */
3775 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3776 struct zone *zone, unsigned int order,
3777 int migratetype, unsigned int alloc_flags)
3779 struct per_cpu_pages *pcp;
3780 struct list_head *list;
3782 unsigned long flags;
3783 unsigned long __maybe_unused UP_flags;
3786 * spin_trylock may fail due to a parallel drain. In the future, the
3787 * trylock will also protect against IRQ reentrancy.
3789 pcp_trylock_prepare(UP_flags);
3790 pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
3792 pcp_trylock_finish(UP_flags);
3797 * On allocation, reduce the number of pages that are batch freed.
3798 * See nr_pcp_free() where free_factor is increased for subsequent
3801 pcp->free_factor >>= 1;
3802 list = &pcp->lists[order_to_pindex(migratetype, order)];
3803 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3804 pcp_spin_unlock_irqrestore(pcp, flags);
3805 pcp_trylock_finish(UP_flags);
3807 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3808 zone_statistics(preferred_zone, zone, 1);
3814 * Allocate a page from the given zone.
3815 * Use pcplists for THP or "cheap" high-order allocations.
3819 * Do not instrument rmqueue() with KMSAN. This function may call
3820 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
3821 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3822 * may call rmqueue() again, which will result in a deadlock.
3824 __no_sanitize_memory
3826 struct page *rmqueue(struct zone *preferred_zone,
3827 struct zone *zone, unsigned int order,
3828 gfp_t gfp_flags, unsigned int alloc_flags,
3834 * We most definitely don't want callers attempting to
3835 * allocate greater than order-1 page units with __GFP_NOFAIL.
3837 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3839 if (likely(pcp_allowed_order(order))) {
3841 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3842 * we need to skip it when CMA area isn't allowed.
3844 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3845 migratetype != MIGRATE_MOVABLE) {
3846 page = rmqueue_pcplist(preferred_zone, zone, order,
3847 migratetype, alloc_flags);
3853 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3857 /* Separate test+clear to avoid unnecessary atomics */
3858 if (unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3859 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3860 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3863 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3867 #ifdef CONFIG_FAIL_PAGE_ALLOC
3870 struct fault_attr attr;
3872 bool ignore_gfp_highmem;
3873 bool ignore_gfp_reclaim;
3875 } fail_page_alloc = {
3876 .attr = FAULT_ATTR_INITIALIZER,
3877 .ignore_gfp_reclaim = true,
3878 .ignore_gfp_highmem = true,
3882 static int __init setup_fail_page_alloc(char *str)
3884 return setup_fault_attr(&fail_page_alloc.attr, str);
3886 __setup("fail_page_alloc=", setup_fail_page_alloc);
3888 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3892 if (order < fail_page_alloc.min_order)
3894 if (gfp_mask & __GFP_NOFAIL)
3896 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3898 if (fail_page_alloc.ignore_gfp_reclaim &&
3899 (gfp_mask & __GFP_DIRECT_RECLAIM))
3902 /* See comment in __should_failslab() */
3903 if (gfp_mask & __GFP_NOWARN)
3904 flags |= FAULT_NOWARN;
3906 return should_fail_ex(&fail_page_alloc.attr, 1 << order, flags);
3909 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3911 static int __init fail_page_alloc_debugfs(void)
3913 umode_t mode = S_IFREG | 0600;
3916 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3917 &fail_page_alloc.attr);
3919 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3920 &fail_page_alloc.ignore_gfp_reclaim);
3921 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3922 &fail_page_alloc.ignore_gfp_highmem);
3923 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3928 late_initcall(fail_page_alloc_debugfs);
3930 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3932 #else /* CONFIG_FAIL_PAGE_ALLOC */
3934 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3939 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3941 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3943 return __should_fail_alloc_page(gfp_mask, order);
3945 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3947 static inline long __zone_watermark_unusable_free(struct zone *z,
3948 unsigned int order, unsigned int alloc_flags)
3950 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3951 long unusable_free = (1 << order) - 1;
3954 * If the caller does not have rights to ALLOC_HARDER then subtract
3955 * the high-atomic reserves. This will over-estimate the size of the
3956 * atomic reserve but it avoids a search.
3958 if (likely(!alloc_harder))
3959 unusable_free += z->nr_reserved_highatomic;
3962 /* If allocation can't use CMA areas don't use free CMA pages */
3963 if (!(alloc_flags & ALLOC_CMA))
3964 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3967 return unusable_free;
3971 * Return true if free base pages are above 'mark'. For high-order checks it
3972 * will return true of the order-0 watermark is reached and there is at least
3973 * one free page of a suitable size. Checking now avoids taking the zone lock
3974 * to check in the allocation paths if no pages are free.
3976 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3977 int highest_zoneidx, unsigned int alloc_flags,
3982 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3984 /* free_pages may go negative - that's OK */
3985 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3987 if (alloc_flags & ALLOC_HIGH)
3990 if (unlikely(alloc_harder)) {
3992 * OOM victims can try even harder than normal ALLOC_HARDER
3993 * users on the grounds that it's definitely going to be in
3994 * the exit path shortly and free memory. Any allocation it
3995 * makes during the free path will be small and short-lived.
3997 if (alloc_flags & ALLOC_OOM)
4004 * Check watermarks for an order-0 allocation request. If these
4005 * are not met, then a high-order request also cannot go ahead
4006 * even if a suitable page happened to be free.
4008 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
4011 /* If this is an order-0 request then the watermark is fine */
4015 /* For a high-order request, check at least one suitable page is free */
4016 for (o = order; o < MAX_ORDER; o++) {
4017 struct free_area *area = &z->free_area[o];
4023 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
4024 if (!free_area_empty(area, mt))
4029 if ((alloc_flags & ALLOC_CMA) &&
4030 !free_area_empty(area, MIGRATE_CMA)) {
4034 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
4040 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
4041 int highest_zoneidx, unsigned int alloc_flags)
4043 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4044 zone_page_state(z, NR_FREE_PAGES));
4047 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
4048 unsigned long mark, int highest_zoneidx,
4049 unsigned int alloc_flags, gfp_t gfp_mask)
4053 free_pages = zone_page_state(z, NR_FREE_PAGES);
4056 * Fast check for order-0 only. If this fails then the reserves
4057 * need to be calculated.
4063 usable_free = free_pages;
4064 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
4066 /* reserved may over estimate high-atomic reserves. */
4067 usable_free -= min(usable_free, reserved);
4068 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
4072 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4076 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
4077 * when checking the min watermark. The min watermark is the
4078 * point where boosting is ignored so that kswapd is woken up
4079 * when below the low watermark.
4081 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
4082 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
4083 mark = z->_watermark[WMARK_MIN];
4084 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
4085 alloc_flags, free_pages);
4091 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
4092 unsigned long mark, int highest_zoneidx)
4094 long free_pages = zone_page_state(z, NR_FREE_PAGES);
4096 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
4097 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
4099 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
4104 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
4106 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4108 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
4109 node_reclaim_distance;
4111 #else /* CONFIG_NUMA */
4112 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4116 #endif /* CONFIG_NUMA */
4119 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4120 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4121 * premature use of a lower zone may cause lowmem pressure problems that
4122 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4123 * probably too small. It only makes sense to spread allocations to avoid
4124 * fragmentation between the Normal and DMA32 zones.
4126 static inline unsigned int
4127 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4129 unsigned int alloc_flags;
4132 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4135 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4137 #ifdef CONFIG_ZONE_DMA32
4141 if (zone_idx(zone) != ZONE_NORMAL)
4145 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4146 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4147 * on UMA that if Normal is populated then so is DMA32.
4149 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4150 if (nr_online_nodes > 1 && !populated_zone(--zone))
4153 alloc_flags |= ALLOC_NOFRAGMENT;
4154 #endif /* CONFIG_ZONE_DMA32 */
4158 /* Must be called after current_gfp_context() which can change gfp_mask */
4159 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4160 unsigned int alloc_flags)
4163 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4164 alloc_flags |= ALLOC_CMA;
4170 * get_page_from_freelist goes through the zonelist trying to allocate
4173 static struct page *
4174 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4175 const struct alloc_context *ac)
4179 struct pglist_data *last_pgdat = NULL;
4180 bool last_pgdat_dirty_ok = false;
4185 * Scan zonelist, looking for a zone with enough free.
4186 * See also __cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
4188 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4189 z = ac->preferred_zoneref;
4190 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4195 if (cpusets_enabled() &&
4196 (alloc_flags & ALLOC_CPUSET) &&
4197 !__cpuset_zone_allowed(zone, gfp_mask))
4200 * When allocating a page cache page for writing, we
4201 * want to get it from a node that is within its dirty
4202 * limit, such that no single node holds more than its
4203 * proportional share of globally allowed dirty pages.
4204 * The dirty limits take into account the node's
4205 * lowmem reserves and high watermark so that kswapd
4206 * should be able to balance it without having to
4207 * write pages from its LRU list.
4209 * XXX: For now, allow allocations to potentially
4210 * exceed the per-node dirty limit in the slowpath
4211 * (spread_dirty_pages unset) before going into reclaim,
4212 * which is important when on a NUMA setup the allowed
4213 * nodes are together not big enough to reach the
4214 * global limit. The proper fix for these situations
4215 * will require awareness of nodes in the
4216 * dirty-throttling and the flusher threads.
4218 if (ac->spread_dirty_pages) {
4219 if (last_pgdat != zone->zone_pgdat) {
4220 last_pgdat = zone->zone_pgdat;
4221 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
4224 if (!last_pgdat_dirty_ok)
4228 if (no_fallback && nr_online_nodes > 1 &&
4229 zone != ac->preferred_zoneref->zone) {
4233 * If moving to a remote node, retry but allow
4234 * fragmenting fallbacks. Locality is more important
4235 * than fragmentation avoidance.
4237 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4238 if (zone_to_nid(zone) != local_nid) {
4239 alloc_flags &= ~ALLOC_NOFRAGMENT;
4244 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4245 if (!zone_watermark_fast(zone, order, mark,
4246 ac->highest_zoneidx, alloc_flags,
4250 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4252 * Watermark failed for this zone, but see if we can
4253 * grow this zone if it contains deferred pages.
4255 if (static_branch_unlikely(&deferred_pages)) {
4256 if (_deferred_grow_zone(zone, order))
4260 /* Checked here to keep the fast path fast */
4261 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4262 if (alloc_flags & ALLOC_NO_WATERMARKS)
4265 if (!node_reclaim_enabled() ||
4266 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4269 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4271 case NODE_RECLAIM_NOSCAN:
4274 case NODE_RECLAIM_FULL:
4275 /* scanned but unreclaimable */
4278 /* did we reclaim enough */
4279 if (zone_watermark_ok(zone, order, mark,
4280 ac->highest_zoneidx, alloc_flags))
4288 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4289 gfp_mask, alloc_flags, ac->migratetype);
4291 prep_new_page(page, order, gfp_mask, alloc_flags);
4294 * If this is a high-order atomic allocation then check
4295 * if the pageblock should be reserved for the future
4297 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4298 reserve_highatomic_pageblock(page, zone, order);
4302 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4303 /* Try again if zone has deferred pages */
4304 if (static_branch_unlikely(&deferred_pages)) {
4305 if (_deferred_grow_zone(zone, order))
4313 * It's possible on a UMA machine to get through all zones that are
4314 * fragmented. If avoiding fragmentation, reset and try again.
4317 alloc_flags &= ~ALLOC_NOFRAGMENT;
4324 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4326 unsigned int filter = SHOW_MEM_FILTER_NODES;
4329 * This documents exceptions given to allocations in certain
4330 * contexts that are allowed to allocate outside current's set
4333 if (!(gfp_mask & __GFP_NOMEMALLOC))
4334 if (tsk_is_oom_victim(current) ||
4335 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4336 filter &= ~SHOW_MEM_FILTER_NODES;
4337 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4338 filter &= ~SHOW_MEM_FILTER_NODES;
4340 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
4343 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4345 struct va_format vaf;
4347 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4349 if ((gfp_mask & __GFP_NOWARN) ||
4350 !__ratelimit(&nopage_rs) ||
4351 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4354 va_start(args, fmt);
4357 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4358 current->comm, &vaf, gfp_mask, &gfp_mask,
4359 nodemask_pr_args(nodemask));
4362 cpuset_print_current_mems_allowed();
4365 warn_alloc_show_mem(gfp_mask, nodemask);
4368 static inline struct page *
4369 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4370 unsigned int alloc_flags,
4371 const struct alloc_context *ac)
4375 page = get_page_from_freelist(gfp_mask, order,
4376 alloc_flags|ALLOC_CPUSET, ac);
4378 * fallback to ignore cpuset restriction if our nodes
4382 page = get_page_from_freelist(gfp_mask, order,
4388 static inline struct page *
4389 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4390 const struct alloc_context *ac, unsigned long *did_some_progress)
4392 struct oom_control oc = {
4393 .zonelist = ac->zonelist,
4394 .nodemask = ac->nodemask,
4396 .gfp_mask = gfp_mask,
4401 *did_some_progress = 0;
4404 * Acquire the oom lock. If that fails, somebody else is
4405 * making progress for us.
4407 if (!mutex_trylock(&oom_lock)) {
4408 *did_some_progress = 1;
4409 schedule_timeout_uninterruptible(1);
4414 * Go through the zonelist yet one more time, keep very high watermark
4415 * here, this is only to catch a parallel oom killing, we must fail if
4416 * we're still under heavy pressure. But make sure that this reclaim
4417 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4418 * allocation which will never fail due to oom_lock already held.
4420 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4421 ~__GFP_DIRECT_RECLAIM, order,
4422 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4426 /* Coredumps can quickly deplete all memory reserves */
4427 if (current->flags & PF_DUMPCORE)
4429 /* The OOM killer will not help higher order allocs */
4430 if (order > PAGE_ALLOC_COSTLY_ORDER)
4433 * We have already exhausted all our reclaim opportunities without any
4434 * success so it is time to admit defeat. We will skip the OOM killer
4435 * because it is very likely that the caller has a more reasonable
4436 * fallback than shooting a random task.
4438 * The OOM killer may not free memory on a specific node.
4440 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4442 /* The OOM killer does not needlessly kill tasks for lowmem */
4443 if (ac->highest_zoneidx < ZONE_NORMAL)
4445 if (pm_suspended_storage())
4448 * XXX: GFP_NOFS allocations should rather fail than rely on
4449 * other request to make a forward progress.
4450 * We are in an unfortunate situation where out_of_memory cannot
4451 * do much for this context but let's try it to at least get
4452 * access to memory reserved if the current task is killed (see
4453 * out_of_memory). Once filesystems are ready to handle allocation
4454 * failures more gracefully we should just bail out here.
4457 /* Exhausted what can be done so it's blame time */
4458 if (out_of_memory(&oc) ||
4459 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4460 *did_some_progress = 1;
4463 * Help non-failing allocations by giving them access to memory
4466 if (gfp_mask & __GFP_NOFAIL)
4467 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4468 ALLOC_NO_WATERMARKS, ac);
4471 mutex_unlock(&oom_lock);
4476 * Maximum number of compaction retries with a progress before OOM
4477 * killer is consider as the only way to move forward.
4479 #define MAX_COMPACT_RETRIES 16
4481 #ifdef CONFIG_COMPACTION
4482 /* Try memory compaction for high-order allocations before reclaim */
4483 static struct page *
4484 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4485 unsigned int alloc_flags, const struct alloc_context *ac,
4486 enum compact_priority prio, enum compact_result *compact_result)
4488 struct page *page = NULL;
4489 unsigned long pflags;
4490 unsigned int noreclaim_flag;
4495 psi_memstall_enter(&pflags);
4496 delayacct_compact_start();
4497 noreclaim_flag = memalloc_noreclaim_save();
4499 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4502 memalloc_noreclaim_restore(noreclaim_flag);
4503 psi_memstall_leave(&pflags);
4504 delayacct_compact_end();
4506 if (*compact_result == COMPACT_SKIPPED)
4509 * At least in one zone compaction wasn't deferred or skipped, so let's
4510 * count a compaction stall
4512 count_vm_event(COMPACTSTALL);
4514 /* Prep a captured page if available */
4516 prep_new_page(page, order, gfp_mask, alloc_flags);
4518 /* Try get a page from the freelist if available */
4520 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4523 struct zone *zone = page_zone(page);
4525 zone->compact_blockskip_flush = false;
4526 compaction_defer_reset(zone, order, true);
4527 count_vm_event(COMPACTSUCCESS);
4532 * It's bad if compaction run occurs and fails. The most likely reason
4533 * is that pages exist, but not enough to satisfy watermarks.
4535 count_vm_event(COMPACTFAIL);
4543 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4544 enum compact_result compact_result,
4545 enum compact_priority *compact_priority,
4546 int *compaction_retries)
4548 int max_retries = MAX_COMPACT_RETRIES;
4551 int retries = *compaction_retries;
4552 enum compact_priority priority = *compact_priority;
4557 if (fatal_signal_pending(current))
4560 if (compaction_made_progress(compact_result))
4561 (*compaction_retries)++;
4564 * compaction considers all the zone as desperately out of memory
4565 * so it doesn't really make much sense to retry except when the
4566 * failure could be caused by insufficient priority
4568 if (compaction_failed(compact_result))
4569 goto check_priority;
4572 * compaction was skipped because there are not enough order-0 pages
4573 * to work with, so we retry only if it looks like reclaim can help.
4575 if (compaction_needs_reclaim(compact_result)) {
4576 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4581 * make sure the compaction wasn't deferred or didn't bail out early
4582 * due to locks contention before we declare that we should give up.
4583 * But the next retry should use a higher priority if allowed, so
4584 * we don't just keep bailing out endlessly.
4586 if (compaction_withdrawn(compact_result)) {
4587 goto check_priority;
4591 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4592 * costly ones because they are de facto nofail and invoke OOM
4593 * killer to move on while costly can fail and users are ready
4594 * to cope with that. 1/4 retries is rather arbitrary but we
4595 * would need much more detailed feedback from compaction to
4596 * make a better decision.
4598 if (order > PAGE_ALLOC_COSTLY_ORDER)
4600 if (*compaction_retries <= max_retries) {
4606 * Make sure there are attempts at the highest priority if we exhausted
4607 * all retries or failed at the lower priorities.
4610 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4611 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4613 if (*compact_priority > min_priority) {
4614 (*compact_priority)--;
4615 *compaction_retries = 0;
4619 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4623 static inline struct page *
4624 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4625 unsigned int alloc_flags, const struct alloc_context *ac,
4626 enum compact_priority prio, enum compact_result *compact_result)
4628 *compact_result = COMPACT_SKIPPED;
4633 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4634 enum compact_result compact_result,
4635 enum compact_priority *compact_priority,
4636 int *compaction_retries)
4641 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4645 * There are setups with compaction disabled which would prefer to loop
4646 * inside the allocator rather than hit the oom killer prematurely.
4647 * Let's give them a good hope and keep retrying while the order-0
4648 * watermarks are OK.
4650 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4651 ac->highest_zoneidx, ac->nodemask) {
4652 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4653 ac->highest_zoneidx, alloc_flags))
4658 #endif /* CONFIG_COMPACTION */
4660 #ifdef CONFIG_LOCKDEP
4661 static struct lockdep_map __fs_reclaim_map =
4662 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4664 static bool __need_reclaim(gfp_t gfp_mask)
4666 /* no reclaim without waiting on it */
4667 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4670 /* this guy won't enter reclaim */
4671 if (current->flags & PF_MEMALLOC)
4674 if (gfp_mask & __GFP_NOLOCKDEP)
4680 void __fs_reclaim_acquire(unsigned long ip)
4682 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4685 void __fs_reclaim_release(unsigned long ip)
4687 lock_release(&__fs_reclaim_map, ip);
4690 void fs_reclaim_acquire(gfp_t gfp_mask)
4692 gfp_mask = current_gfp_context(gfp_mask);
4694 if (__need_reclaim(gfp_mask)) {
4695 if (gfp_mask & __GFP_FS)
4696 __fs_reclaim_acquire(_RET_IP_);
4698 #ifdef CONFIG_MMU_NOTIFIER
4699 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4700 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4705 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4707 void fs_reclaim_release(gfp_t gfp_mask)
4709 gfp_mask = current_gfp_context(gfp_mask);
4711 if (__need_reclaim(gfp_mask)) {
4712 if (gfp_mask & __GFP_FS)
4713 __fs_reclaim_release(_RET_IP_);
4716 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4720 * Zonelists may change due to hotplug during allocation. Detect when zonelists
4721 * have been rebuilt so allocation retries. Reader side does not lock and
4722 * retries the allocation if zonelist changes. Writer side is protected by the
4723 * embedded spin_lock.
4725 static DEFINE_SEQLOCK(zonelist_update_seq);
4727 static unsigned int zonelist_iter_begin(void)
4729 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4730 return read_seqbegin(&zonelist_update_seq);
4735 static unsigned int check_retry_zonelist(unsigned int seq)
4737 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4738 return read_seqretry(&zonelist_update_seq, seq);
4743 /* Perform direct synchronous page reclaim */
4744 static unsigned long
4745 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4746 const struct alloc_context *ac)
4748 unsigned int noreclaim_flag;
4749 unsigned long progress;
4753 /* We now go into synchronous reclaim */
4754 cpuset_memory_pressure_bump();
4755 fs_reclaim_acquire(gfp_mask);
4756 noreclaim_flag = memalloc_noreclaim_save();
4758 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4761 memalloc_noreclaim_restore(noreclaim_flag);
4762 fs_reclaim_release(gfp_mask);
4769 /* The really slow allocator path where we enter direct reclaim */
4770 static inline struct page *
4771 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4772 unsigned int alloc_flags, const struct alloc_context *ac,
4773 unsigned long *did_some_progress)
4775 struct page *page = NULL;
4776 unsigned long pflags;
4777 bool drained = false;
4779 psi_memstall_enter(&pflags);
4780 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4781 if (unlikely(!(*did_some_progress)))
4785 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4788 * If an allocation failed after direct reclaim, it could be because
4789 * pages are pinned on the per-cpu lists or in high alloc reserves.
4790 * Shrink them and try again
4792 if (!page && !drained) {
4793 unreserve_highatomic_pageblock(ac, false);
4794 drain_all_pages(NULL);
4799 psi_memstall_leave(&pflags);
4804 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4805 const struct alloc_context *ac)
4809 pg_data_t *last_pgdat = NULL;
4810 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4812 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4814 if (!managed_zone(zone))
4816 if (last_pgdat != zone->zone_pgdat) {
4817 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4818 last_pgdat = zone->zone_pgdat;
4823 static inline unsigned int
4824 gfp_to_alloc_flags(gfp_t gfp_mask)
4826 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4829 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4830 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4831 * to save two branches.
4833 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4834 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4837 * The caller may dip into page reserves a bit more if the caller
4838 * cannot run direct reclaim, or if the caller has realtime scheduling
4839 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4840 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4842 alloc_flags |= (__force int)
4843 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4845 if (gfp_mask & __GFP_ATOMIC) {
4847 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4848 * if it can't schedule.
4850 if (!(gfp_mask & __GFP_NOMEMALLOC))
4851 alloc_flags |= ALLOC_HARDER;
4853 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4854 * comment for __cpuset_node_allowed().
4856 alloc_flags &= ~ALLOC_CPUSET;
4857 } else if (unlikely(rt_task(current)) && in_task())
4858 alloc_flags |= ALLOC_HARDER;
4860 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4865 static bool oom_reserves_allowed(struct task_struct *tsk)
4867 if (!tsk_is_oom_victim(tsk))
4871 * !MMU doesn't have oom reaper so give access to memory reserves
4872 * only to the thread with TIF_MEMDIE set
4874 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4881 * Distinguish requests which really need access to full memory
4882 * reserves from oom victims which can live with a portion of it
4884 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4886 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4888 if (gfp_mask & __GFP_MEMALLOC)
4889 return ALLOC_NO_WATERMARKS;
4890 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4891 return ALLOC_NO_WATERMARKS;
4892 if (!in_interrupt()) {
4893 if (current->flags & PF_MEMALLOC)
4894 return ALLOC_NO_WATERMARKS;
4895 else if (oom_reserves_allowed(current))
4902 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4904 return !!__gfp_pfmemalloc_flags(gfp_mask);
4908 * Checks whether it makes sense to retry the reclaim to make a forward progress
4909 * for the given allocation request.
4911 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4912 * without success, or when we couldn't even meet the watermark if we
4913 * reclaimed all remaining pages on the LRU lists.
4915 * Returns true if a retry is viable or false to enter the oom path.
4918 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4919 struct alloc_context *ac, int alloc_flags,
4920 bool did_some_progress, int *no_progress_loops)
4927 * Costly allocations might have made a progress but this doesn't mean
4928 * their order will become available due to high fragmentation so
4929 * always increment the no progress counter for them
4931 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4932 *no_progress_loops = 0;
4934 (*no_progress_loops)++;
4937 * Make sure we converge to OOM if we cannot make any progress
4938 * several times in the row.
4940 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4941 /* Before OOM, exhaust highatomic_reserve */
4942 return unreserve_highatomic_pageblock(ac, true);
4946 * Keep reclaiming pages while there is a chance this will lead
4947 * somewhere. If none of the target zones can satisfy our allocation
4948 * request even if all reclaimable pages are considered then we are
4949 * screwed and have to go OOM.
4951 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4952 ac->highest_zoneidx, ac->nodemask) {
4953 unsigned long available;
4954 unsigned long reclaimable;
4955 unsigned long min_wmark = min_wmark_pages(zone);
4958 available = reclaimable = zone_reclaimable_pages(zone);
4959 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4962 * Would the allocation succeed if we reclaimed all
4963 * reclaimable pages?
4965 wmark = __zone_watermark_ok(zone, order, min_wmark,
4966 ac->highest_zoneidx, alloc_flags, available);
4967 trace_reclaim_retry_zone(z, order, reclaimable,
4968 available, min_wmark, *no_progress_loops, wmark);
4976 * Memory allocation/reclaim might be called from a WQ context and the
4977 * current implementation of the WQ concurrency control doesn't
4978 * recognize that a particular WQ is congested if the worker thread is
4979 * looping without ever sleeping. Therefore we have to do a short sleep
4980 * here rather than calling cond_resched().
4982 if (current->flags & PF_WQ_WORKER)
4983 schedule_timeout_uninterruptible(1);
4990 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4993 * It's possible that cpuset's mems_allowed and the nodemask from
4994 * mempolicy don't intersect. This should be normally dealt with by
4995 * policy_nodemask(), but it's possible to race with cpuset update in
4996 * such a way the check therein was true, and then it became false
4997 * before we got our cpuset_mems_cookie here.
4998 * This assumes that for all allocations, ac->nodemask can come only
4999 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
5000 * when it does not intersect with the cpuset restrictions) or the
5001 * caller can deal with a violated nodemask.
5003 if (cpusets_enabled() && ac->nodemask &&
5004 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
5005 ac->nodemask = NULL;
5010 * When updating a task's mems_allowed or mempolicy nodemask, it is
5011 * possible to race with parallel threads in such a way that our
5012 * allocation can fail while the mask is being updated. If we are about
5013 * to fail, check if the cpuset changed during allocation and if so,
5016 if (read_mems_allowed_retry(cpuset_mems_cookie))
5022 static inline struct page *
5023 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
5024 struct alloc_context *ac)
5026 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
5027 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
5028 struct page *page = NULL;
5029 unsigned int alloc_flags;
5030 unsigned long did_some_progress;
5031 enum compact_priority compact_priority;
5032 enum compact_result compact_result;
5033 int compaction_retries;
5034 int no_progress_loops;
5035 unsigned int cpuset_mems_cookie;
5036 unsigned int zonelist_iter_cookie;
5040 * We also sanity check to catch abuse of atomic reserves being used by
5041 * callers that are not in atomic context.
5043 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
5044 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
5045 gfp_mask &= ~__GFP_ATOMIC;
5048 compaction_retries = 0;
5049 no_progress_loops = 0;
5050 compact_priority = DEF_COMPACT_PRIORITY;
5051 cpuset_mems_cookie = read_mems_allowed_begin();
5052 zonelist_iter_cookie = zonelist_iter_begin();
5055 * The fast path uses conservative alloc_flags to succeed only until
5056 * kswapd needs to be woken up, and to avoid the cost of setting up
5057 * alloc_flags precisely. So we do that now.
5059 alloc_flags = gfp_to_alloc_flags(gfp_mask);
5062 * We need to recalculate the starting point for the zonelist iterator
5063 * because we might have used different nodemask in the fast path, or
5064 * there was a cpuset modification and we are retrying - otherwise we
5065 * could end up iterating over non-eligible zones endlessly.
5067 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5068 ac->highest_zoneidx, ac->nodemask);
5069 if (!ac->preferred_zoneref->zone)
5073 * Check for insane configurations where the cpuset doesn't contain
5074 * any suitable zone to satisfy the request - e.g. non-movable
5075 * GFP_HIGHUSER allocations from MOVABLE nodes only.
5077 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
5078 struct zoneref *z = first_zones_zonelist(ac->zonelist,
5079 ac->highest_zoneidx,
5080 &cpuset_current_mems_allowed);
5085 if (alloc_flags & ALLOC_KSWAPD)
5086 wake_all_kswapds(order, gfp_mask, ac);
5089 * The adjusted alloc_flags might result in immediate success, so try
5092 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5097 * For costly allocations, try direct compaction first, as it's likely
5098 * that we have enough base pages and don't need to reclaim. For non-
5099 * movable high-order allocations, do that as well, as compaction will
5100 * try prevent permanent fragmentation by migrating from blocks of the
5102 * Don't try this for allocations that are allowed to ignore
5103 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
5105 if (can_direct_reclaim &&
5107 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
5108 && !gfp_pfmemalloc_allowed(gfp_mask)) {
5109 page = __alloc_pages_direct_compact(gfp_mask, order,
5111 INIT_COMPACT_PRIORITY,
5117 * Checks for costly allocations with __GFP_NORETRY, which
5118 * includes some THP page fault allocations
5120 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
5122 * If allocating entire pageblock(s) and compaction
5123 * failed because all zones are below low watermarks
5124 * or is prohibited because it recently failed at this
5125 * order, fail immediately unless the allocator has
5126 * requested compaction and reclaim retry.
5129 * - potentially very expensive because zones are far
5130 * below their low watermarks or this is part of very
5131 * bursty high order allocations,
5132 * - not guaranteed to help because isolate_freepages()
5133 * may not iterate over freed pages as part of its
5135 * - unlikely to make entire pageblocks free on its
5138 if (compact_result == COMPACT_SKIPPED ||
5139 compact_result == COMPACT_DEFERRED)
5143 * Looks like reclaim/compaction is worth trying, but
5144 * sync compaction could be very expensive, so keep
5145 * using async compaction.
5147 compact_priority = INIT_COMPACT_PRIORITY;
5152 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5153 if (alloc_flags & ALLOC_KSWAPD)
5154 wake_all_kswapds(order, gfp_mask, ac);
5156 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5158 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
5159 (alloc_flags & ALLOC_KSWAPD);
5162 * Reset the nodemask and zonelist iterators if memory policies can be
5163 * ignored. These allocations are high priority and system rather than
5166 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5167 ac->nodemask = NULL;
5168 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5169 ac->highest_zoneidx, ac->nodemask);
5172 /* Attempt with potentially adjusted zonelist and alloc_flags */
5173 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5177 /* Caller is not willing to reclaim, we can't balance anything */
5178 if (!can_direct_reclaim)
5181 /* Avoid recursion of direct reclaim */
5182 if (current->flags & PF_MEMALLOC)
5185 /* Try direct reclaim and then allocating */
5186 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5187 &did_some_progress);
5191 /* Try direct compaction and then allocating */
5192 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5193 compact_priority, &compact_result);
5197 /* Do not loop if specifically requested */
5198 if (gfp_mask & __GFP_NORETRY)
5202 * Do not retry costly high order allocations unless they are
5203 * __GFP_RETRY_MAYFAIL
5205 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5208 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5209 did_some_progress > 0, &no_progress_loops))
5213 * It doesn't make any sense to retry for the compaction if the order-0
5214 * reclaim is not able to make any progress because the current
5215 * implementation of the compaction depends on the sufficient amount
5216 * of free memory (see __compaction_suitable)
5218 if (did_some_progress > 0 &&
5219 should_compact_retry(ac, order, alloc_flags,
5220 compact_result, &compact_priority,
5221 &compaction_retries))
5226 * Deal with possible cpuset update races or zonelist updates to avoid
5227 * a unnecessary OOM kill.
5229 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5230 check_retry_zonelist(zonelist_iter_cookie))
5233 /* Reclaim has failed us, start killing things */
5234 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5238 /* Avoid allocations with no watermarks from looping endlessly */
5239 if (tsk_is_oom_victim(current) &&
5240 (alloc_flags & ALLOC_OOM ||
5241 (gfp_mask & __GFP_NOMEMALLOC)))
5244 /* Retry as long as the OOM killer is making progress */
5245 if (did_some_progress) {
5246 no_progress_loops = 0;
5252 * Deal with possible cpuset update races or zonelist updates to avoid
5253 * a unnecessary OOM kill.
5255 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5256 check_retry_zonelist(zonelist_iter_cookie))
5260 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5263 if (gfp_mask & __GFP_NOFAIL) {
5265 * All existing users of the __GFP_NOFAIL are blockable, so warn
5266 * of any new users that actually require GFP_NOWAIT
5268 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
5272 * PF_MEMALLOC request from this context is rather bizarre
5273 * because we cannot reclaim anything and only can loop waiting
5274 * for somebody to do a work for us
5276 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
5279 * non failing costly orders are a hard requirement which we
5280 * are not prepared for much so let's warn about these users
5281 * so that we can identify them and convert them to something
5284 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
5287 * Help non-failing allocations by giving them access to memory
5288 * reserves but do not use ALLOC_NO_WATERMARKS because this
5289 * could deplete whole memory reserves which would just make
5290 * the situation worse
5292 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5300 warn_alloc(gfp_mask, ac->nodemask,
5301 "page allocation failure: order:%u", order);
5306 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5307 int preferred_nid, nodemask_t *nodemask,
5308 struct alloc_context *ac, gfp_t *alloc_gfp,
5309 unsigned int *alloc_flags)
5311 ac->highest_zoneidx = gfp_zone(gfp_mask);
5312 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5313 ac->nodemask = nodemask;
5314 ac->migratetype = gfp_migratetype(gfp_mask);
5316 if (cpusets_enabled()) {
5317 *alloc_gfp |= __GFP_HARDWALL;
5319 * When we are in the interrupt context, it is irrelevant
5320 * to the current task context. It means that any node ok.
5322 if (in_task() && !ac->nodemask)
5323 ac->nodemask = &cpuset_current_mems_allowed;
5325 *alloc_flags |= ALLOC_CPUSET;
5328 might_alloc(gfp_mask);
5330 if (should_fail_alloc_page(gfp_mask, order))
5333 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5335 /* Dirty zone balancing only done in the fast path */
5336 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5339 * The preferred zone is used for statistics but crucially it is
5340 * also used as the starting point for the zonelist iterator. It
5341 * may get reset for allocations that ignore memory policies.
5343 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5344 ac->highest_zoneidx, ac->nodemask);
5350 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5351 * @gfp: GFP flags for the allocation
5352 * @preferred_nid: The preferred NUMA node ID to allocate from
5353 * @nodemask: Set of nodes to allocate from, may be NULL
5354 * @nr_pages: The number of pages desired on the list or array
5355 * @page_list: Optional list to store the allocated pages
5356 * @page_array: Optional array to store the pages
5358 * This is a batched version of the page allocator that attempts to
5359 * allocate nr_pages quickly. Pages are added to page_list if page_list
5360 * is not NULL, otherwise it is assumed that the page_array is valid.
5362 * For lists, nr_pages is the number of pages that should be allocated.
5364 * For arrays, only NULL elements are populated with pages and nr_pages
5365 * is the maximum number of pages that will be stored in the array.
5367 * Returns the number of pages on the list or array.
5369 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5370 nodemask_t *nodemask, int nr_pages,
5371 struct list_head *page_list,
5372 struct page **page_array)
5375 unsigned long flags;
5376 unsigned long __maybe_unused UP_flags;
5379 struct per_cpu_pages *pcp;
5380 struct list_head *pcp_list;
5381 struct alloc_context ac;
5383 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5384 int nr_populated = 0, nr_account = 0;
5387 * Skip populated array elements to determine if any pages need
5388 * to be allocated before disabling IRQs.
5390 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5393 /* No pages requested? */
5394 if (unlikely(nr_pages <= 0))
5397 /* Already populated array? */
5398 if (unlikely(page_array && nr_pages - nr_populated == 0))
5401 /* Bulk allocator does not support memcg accounting. */
5402 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5405 /* Use the single page allocator for one page. */
5406 if (nr_pages - nr_populated == 1)
5409 #ifdef CONFIG_PAGE_OWNER
5411 * PAGE_OWNER may recurse into the allocator to allocate space to
5412 * save the stack with pagesets.lock held. Releasing/reacquiring
5413 * removes much of the performance benefit of bulk allocation so
5414 * force the caller to allocate one page at a time as it'll have
5415 * similar performance to added complexity to the bulk allocator.
5417 if (static_branch_unlikely(&page_owner_inited))
5421 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5422 gfp &= gfp_allowed_mask;
5424 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5428 /* Find an allowed local zone that meets the low watermark. */
5429 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5432 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5433 !__cpuset_zone_allowed(zone, gfp)) {
5437 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5438 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5442 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5443 if (zone_watermark_fast(zone, 0, mark,
5444 zonelist_zone_idx(ac.preferred_zoneref),
5445 alloc_flags, gfp)) {
5451 * If there are no allowed local zones that meets the watermarks then
5452 * try to allocate a single page and reclaim if necessary.
5454 if (unlikely(!zone))
5457 /* Is a parallel drain in progress? */
5458 pcp_trylock_prepare(UP_flags);
5459 pcp = pcp_spin_trylock_irqsave(zone->per_cpu_pageset, flags);
5463 /* Attempt the batch allocation */
5464 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5465 while (nr_populated < nr_pages) {
5467 /* Skip existing pages */
5468 if (page_array && page_array[nr_populated]) {
5473 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5475 if (unlikely(!page)) {
5476 /* Try and allocate at least one page */
5478 pcp_spin_unlock_irqrestore(pcp, flags);
5485 prep_new_page(page, 0, gfp, 0);
5487 list_add(&page->lru, page_list);
5489 page_array[nr_populated] = page;
5493 pcp_spin_unlock_irqrestore(pcp, flags);
5494 pcp_trylock_finish(UP_flags);
5496 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5497 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5500 return nr_populated;
5503 pcp_trylock_finish(UP_flags);
5506 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5509 list_add(&page->lru, page_list);
5511 page_array[nr_populated] = page;
5517 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5520 * This is the 'heart' of the zoned buddy allocator.
5522 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5523 nodemask_t *nodemask)
5526 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5527 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5528 struct alloc_context ac = { };
5531 * There are several places where we assume that the order value is sane
5532 * so bail out early if the request is out of bound.
5534 if (WARN_ON_ONCE_GFP(order >= MAX_ORDER, gfp))
5537 gfp &= gfp_allowed_mask;
5539 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5540 * resp. GFP_NOIO which has to be inherited for all allocation requests
5541 * from a particular context which has been marked by
5542 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5543 * movable zones are not used during allocation.
5545 gfp = current_gfp_context(gfp);
5547 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5548 &alloc_gfp, &alloc_flags))
5552 * Forbid the first pass from falling back to types that fragment
5553 * memory until all local zones are considered.
5555 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5557 /* First allocation attempt */
5558 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5563 ac.spread_dirty_pages = false;
5566 * Restore the original nodemask if it was potentially replaced with
5567 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5569 ac.nodemask = nodemask;
5571 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5574 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5575 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5576 __free_pages(page, order);
5580 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5581 kmsan_alloc_page(page, order, alloc_gfp);
5585 EXPORT_SYMBOL(__alloc_pages);
5587 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5588 nodemask_t *nodemask)
5590 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5591 preferred_nid, nodemask);
5593 if (page && order > 1)
5594 prep_transhuge_page(page);
5595 return (struct folio *)page;
5597 EXPORT_SYMBOL(__folio_alloc);
5600 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5601 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5602 * you need to access high mem.
5604 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5608 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5611 return (unsigned long) page_address(page);
5613 EXPORT_SYMBOL(__get_free_pages);
5615 unsigned long get_zeroed_page(gfp_t gfp_mask)
5617 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5619 EXPORT_SYMBOL(get_zeroed_page);
5622 * __free_pages - Free pages allocated with alloc_pages().
5623 * @page: The page pointer returned from alloc_pages().
5624 * @order: The order of the allocation.
5626 * This function can free multi-page allocations that are not compound
5627 * pages. It does not check that the @order passed in matches that of
5628 * the allocation, so it is easy to leak memory. Freeing more memory
5629 * than was allocated will probably emit a warning.
5631 * If the last reference to this page is speculative, it will be released
5632 * by put_page() which only frees the first page of a non-compound
5633 * allocation. To prevent the remaining pages from being leaked, we free
5634 * the subsequent pages here. If you want to use the page's reference
5635 * count to decide when to free the allocation, you should allocate a
5636 * compound page, and use put_page() instead of __free_pages().
5638 * Context: May be called in interrupt context or while holding a normal
5639 * spinlock, but not in NMI context or while holding a raw spinlock.
5641 void __free_pages(struct page *page, unsigned int order)
5643 /* get PageHead before we drop reference */
5644 int head = PageHead(page);
5646 if (put_page_testzero(page))
5647 free_the_page(page, order);
5650 free_the_page(page + (1 << order), order);
5652 EXPORT_SYMBOL(__free_pages);
5654 void free_pages(unsigned long addr, unsigned int order)
5657 VM_BUG_ON(!virt_addr_valid((void *)addr));
5658 __free_pages(virt_to_page((void *)addr), order);
5662 EXPORT_SYMBOL(free_pages);
5666 * An arbitrary-length arbitrary-offset area of memory which resides
5667 * within a 0 or higher order page. Multiple fragments within that page
5668 * are individually refcounted, in the page's reference counter.
5670 * The page_frag functions below provide a simple allocation framework for
5671 * page fragments. This is used by the network stack and network device
5672 * drivers to provide a backing region of memory for use as either an
5673 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5675 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5678 struct page *page = NULL;
5679 gfp_t gfp = gfp_mask;
5681 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5682 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5684 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5685 PAGE_FRAG_CACHE_MAX_ORDER);
5686 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5688 if (unlikely(!page))
5689 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5691 nc->va = page ? page_address(page) : NULL;
5696 void __page_frag_cache_drain(struct page *page, unsigned int count)
5698 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5700 if (page_ref_sub_and_test(page, count))
5701 free_the_page(page, compound_order(page));
5703 EXPORT_SYMBOL(__page_frag_cache_drain);
5705 void *page_frag_alloc_align(struct page_frag_cache *nc,
5706 unsigned int fragsz, gfp_t gfp_mask,
5707 unsigned int align_mask)
5709 unsigned int size = PAGE_SIZE;
5713 if (unlikely(!nc->va)) {
5715 page = __page_frag_cache_refill(nc, gfp_mask);
5719 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5720 /* if size can vary use size else just use PAGE_SIZE */
5723 /* Even if we own the page, we do not use atomic_set().
5724 * This would break get_page_unless_zero() users.
5726 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5728 /* reset page count bias and offset to start of new frag */
5729 nc->pfmemalloc = page_is_pfmemalloc(page);
5730 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5734 offset = nc->offset - fragsz;
5735 if (unlikely(offset < 0)) {
5736 page = virt_to_page(nc->va);
5738 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5741 if (unlikely(nc->pfmemalloc)) {
5742 free_the_page(page, compound_order(page));
5746 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5747 /* if size can vary use size else just use PAGE_SIZE */
5750 /* OK, page count is 0, we can safely set it */
5751 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5753 /* reset page count bias and offset to start of new frag */
5754 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5755 offset = size - fragsz;
5756 if (unlikely(offset < 0)) {
5758 * The caller is trying to allocate a fragment
5759 * with fragsz > PAGE_SIZE but the cache isn't big
5760 * enough to satisfy the request, this may
5761 * happen in low memory conditions.
5762 * We don't release the cache page because
5763 * it could make memory pressure worse
5764 * so we simply return NULL here.
5771 offset &= align_mask;
5772 nc->offset = offset;
5774 return nc->va + offset;
5776 EXPORT_SYMBOL(page_frag_alloc_align);
5779 * Frees a page fragment allocated out of either a compound or order 0 page.
5781 void page_frag_free(void *addr)
5783 struct page *page = virt_to_head_page(addr);
5785 if (unlikely(put_page_testzero(page)))
5786 free_the_page(page, compound_order(page));
5788 EXPORT_SYMBOL(page_frag_free);
5790 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5794 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
5795 struct page *page = virt_to_page((void *)addr);
5796 struct page *last = page + nr;
5798 split_page_owner(page, 1 << order);
5799 split_page_memcg(page, 1 << order);
5800 while (page < --last)
5801 set_page_refcounted(last);
5803 last = page + (1UL << order);
5804 for (page += nr; page < last; page++)
5805 __free_pages_ok(page, 0, FPI_TO_TAIL);
5807 return (void *)addr;
5811 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5812 * @size: the number of bytes to allocate
5813 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5815 * This function is similar to alloc_pages(), except that it allocates the
5816 * minimum number of pages to satisfy the request. alloc_pages() can only
5817 * allocate memory in power-of-two pages.
5819 * This function is also limited by MAX_ORDER.
5821 * Memory allocated by this function must be released by free_pages_exact().
5823 * Return: pointer to the allocated area or %NULL in case of error.
5825 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5827 unsigned int order = get_order(size);
5830 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5831 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5833 addr = __get_free_pages(gfp_mask, order);
5834 return make_alloc_exact(addr, order, size);
5836 EXPORT_SYMBOL(alloc_pages_exact);
5839 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5841 * @nid: the preferred node ID where memory should be allocated
5842 * @size: the number of bytes to allocate
5843 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5845 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5848 * Return: pointer to the allocated area or %NULL in case of error.
5850 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5852 unsigned int order = get_order(size);
5855 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5856 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5858 p = alloc_pages_node(nid, gfp_mask, order);
5861 return make_alloc_exact((unsigned long)page_address(p), order, size);
5865 * free_pages_exact - release memory allocated via alloc_pages_exact()
5866 * @virt: the value returned by alloc_pages_exact.
5867 * @size: size of allocation, same value as passed to alloc_pages_exact().
5869 * Release the memory allocated by a previous call to alloc_pages_exact.
5871 void free_pages_exact(void *virt, size_t size)
5873 unsigned long addr = (unsigned long)virt;
5874 unsigned long end = addr + PAGE_ALIGN(size);
5876 while (addr < end) {
5881 EXPORT_SYMBOL(free_pages_exact);
5884 * nr_free_zone_pages - count number of pages beyond high watermark
5885 * @offset: The zone index of the highest zone
5887 * nr_free_zone_pages() counts the number of pages which are beyond the
5888 * high watermark within all zones at or below a given zone index. For each
5889 * zone, the number of pages is calculated as:
5891 * nr_free_zone_pages = managed_pages - high_pages
5893 * Return: number of pages beyond high watermark.
5895 static unsigned long nr_free_zone_pages(int offset)
5900 /* Just pick one node, since fallback list is circular */
5901 unsigned long sum = 0;
5903 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5905 for_each_zone_zonelist(zone, z, zonelist, offset) {
5906 unsigned long size = zone_managed_pages(zone);
5907 unsigned long high = high_wmark_pages(zone);
5916 * nr_free_buffer_pages - count number of pages beyond high watermark
5918 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5919 * watermark within ZONE_DMA and ZONE_NORMAL.
5921 * Return: number of pages beyond high watermark within ZONE_DMA and
5924 unsigned long nr_free_buffer_pages(void)
5926 return nr_free_zone_pages(gfp_zone(GFP_USER));
5928 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5930 static inline void show_node(struct zone *zone)
5932 if (IS_ENABLED(CONFIG_NUMA))
5933 printk("Node %d ", zone_to_nid(zone));
5936 long si_mem_available(void)
5939 unsigned long pagecache;
5940 unsigned long wmark_low = 0;
5941 unsigned long pages[NR_LRU_LISTS];
5942 unsigned long reclaimable;
5946 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5947 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5950 wmark_low += low_wmark_pages(zone);
5953 * Estimate the amount of memory available for userspace allocations,
5954 * without causing swapping or OOM.
5956 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5959 * Not all the page cache can be freed, otherwise the system will
5960 * start swapping or thrashing. Assume at least half of the page
5961 * cache, or the low watermark worth of cache, needs to stay.
5963 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5964 pagecache -= min(pagecache / 2, wmark_low);
5965 available += pagecache;
5968 * Part of the reclaimable slab and other kernel memory consists of
5969 * items that are in use, and cannot be freed. Cap this estimate at the
5972 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5973 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5974 available += reclaimable - min(reclaimable / 2, wmark_low);
5980 EXPORT_SYMBOL_GPL(si_mem_available);
5982 void si_meminfo(struct sysinfo *val)
5984 val->totalram = totalram_pages();
5985 val->sharedram = global_node_page_state(NR_SHMEM);
5986 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5987 val->bufferram = nr_blockdev_pages();
5988 val->totalhigh = totalhigh_pages();
5989 val->freehigh = nr_free_highpages();
5990 val->mem_unit = PAGE_SIZE;
5993 EXPORT_SYMBOL(si_meminfo);
5996 void si_meminfo_node(struct sysinfo *val, int nid)
5998 int zone_type; /* needs to be signed */
5999 unsigned long managed_pages = 0;
6000 unsigned long managed_highpages = 0;
6001 unsigned long free_highpages = 0;
6002 pg_data_t *pgdat = NODE_DATA(nid);
6004 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
6005 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
6006 val->totalram = managed_pages;
6007 val->sharedram = node_page_state(pgdat, NR_SHMEM);
6008 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
6009 #ifdef CONFIG_HIGHMEM
6010 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
6011 struct zone *zone = &pgdat->node_zones[zone_type];
6013 if (is_highmem(zone)) {
6014 managed_highpages += zone_managed_pages(zone);
6015 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
6018 val->totalhigh = managed_highpages;
6019 val->freehigh = free_highpages;
6021 val->totalhigh = managed_highpages;
6022 val->freehigh = free_highpages;
6024 val->mem_unit = PAGE_SIZE;
6029 * Determine whether the node should be displayed or not, depending on whether
6030 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
6032 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
6034 if (!(flags & SHOW_MEM_FILTER_NODES))
6038 * no node mask - aka implicit memory numa policy. Do not bother with
6039 * the synchronization - read_mems_allowed_begin - because we do not
6040 * have to be precise here.
6043 nodemask = &cpuset_current_mems_allowed;
6045 return !node_isset(nid, *nodemask);
6048 #define K(x) ((x) << (PAGE_SHIFT-10))
6050 static void show_migration_types(unsigned char type)
6052 static const char types[MIGRATE_TYPES] = {
6053 [MIGRATE_UNMOVABLE] = 'U',
6054 [MIGRATE_MOVABLE] = 'M',
6055 [MIGRATE_RECLAIMABLE] = 'E',
6056 [MIGRATE_HIGHATOMIC] = 'H',
6058 [MIGRATE_CMA] = 'C',
6060 #ifdef CONFIG_MEMORY_ISOLATION
6061 [MIGRATE_ISOLATE] = 'I',
6064 char tmp[MIGRATE_TYPES + 1];
6068 for (i = 0; i < MIGRATE_TYPES; i++) {
6069 if (type & (1 << i))
6074 printk(KERN_CONT "(%s) ", tmp);
6077 static bool node_has_managed_zones(pg_data_t *pgdat, int max_zone_idx)
6080 for (zone_idx = 0; zone_idx <= max_zone_idx; zone_idx++)
6081 if (zone_managed_pages(pgdat->node_zones + zone_idx))
6087 * Show free area list (used inside shift_scroll-lock stuff)
6088 * We also calculate the percentage fragmentation. We do this by counting the
6089 * memory on each free list with the exception of the first item on the list.
6092 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
6095 void __show_free_areas(unsigned int filter, nodemask_t *nodemask, int max_zone_idx)
6097 unsigned long free_pcp = 0;
6102 for_each_populated_zone(zone) {
6103 if (zone_idx(zone) > max_zone_idx)
6105 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6108 for_each_online_cpu(cpu)
6109 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6112 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
6113 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
6114 " unevictable:%lu dirty:%lu writeback:%lu\n"
6115 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
6116 " mapped:%lu shmem:%lu pagetables:%lu\n"
6117 " sec_pagetables:%lu bounce:%lu\n"
6118 " kernel_misc_reclaimable:%lu\n"
6119 " free:%lu free_pcp:%lu free_cma:%lu\n",
6120 global_node_page_state(NR_ACTIVE_ANON),
6121 global_node_page_state(NR_INACTIVE_ANON),
6122 global_node_page_state(NR_ISOLATED_ANON),
6123 global_node_page_state(NR_ACTIVE_FILE),
6124 global_node_page_state(NR_INACTIVE_FILE),
6125 global_node_page_state(NR_ISOLATED_FILE),
6126 global_node_page_state(NR_UNEVICTABLE),
6127 global_node_page_state(NR_FILE_DIRTY),
6128 global_node_page_state(NR_WRITEBACK),
6129 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
6130 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
6131 global_node_page_state(NR_FILE_MAPPED),
6132 global_node_page_state(NR_SHMEM),
6133 global_node_page_state(NR_PAGETABLE),
6134 global_node_page_state(NR_SECONDARY_PAGETABLE),
6135 global_zone_page_state(NR_BOUNCE),
6136 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
6137 global_zone_page_state(NR_FREE_PAGES),
6139 global_zone_page_state(NR_FREE_CMA_PAGES));
6141 for_each_online_pgdat(pgdat) {
6142 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
6144 if (!node_has_managed_zones(pgdat, max_zone_idx))
6148 " active_anon:%lukB"
6149 " inactive_anon:%lukB"
6150 " active_file:%lukB"
6151 " inactive_file:%lukB"
6152 " unevictable:%lukB"
6153 " isolated(anon):%lukB"
6154 " isolated(file):%lukB"
6159 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6161 " shmem_pmdmapped: %lukB"
6164 " writeback_tmp:%lukB"
6165 " kernel_stack:%lukB"
6166 #ifdef CONFIG_SHADOW_CALL_STACK
6167 " shadow_call_stack:%lukB"
6170 " sec_pagetables:%lukB"
6171 " all_unreclaimable? %s"
6174 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
6175 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
6176 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
6177 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
6178 K(node_page_state(pgdat, NR_UNEVICTABLE)),
6179 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
6180 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
6181 K(node_page_state(pgdat, NR_FILE_MAPPED)),
6182 K(node_page_state(pgdat, NR_FILE_DIRTY)),
6183 K(node_page_state(pgdat, NR_WRITEBACK)),
6184 K(node_page_state(pgdat, NR_SHMEM)),
6185 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6186 K(node_page_state(pgdat, NR_SHMEM_THPS)),
6187 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6188 K(node_page_state(pgdat, NR_ANON_THPS)),
6190 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6191 node_page_state(pgdat, NR_KERNEL_STACK_KB),
6192 #ifdef CONFIG_SHADOW_CALL_STACK
6193 node_page_state(pgdat, NR_KERNEL_SCS_KB),
6195 K(node_page_state(pgdat, NR_PAGETABLE)),
6196 K(node_page_state(pgdat, NR_SECONDARY_PAGETABLE)),
6197 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6201 for_each_populated_zone(zone) {
6204 if (zone_idx(zone) > max_zone_idx)
6206 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6210 for_each_online_cpu(cpu)
6211 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6221 " reserved_highatomic:%luKB"
6222 " active_anon:%lukB"
6223 " inactive_anon:%lukB"
6224 " active_file:%lukB"
6225 " inactive_file:%lukB"
6226 " unevictable:%lukB"
6227 " writepending:%lukB"
6237 K(zone_page_state(zone, NR_FREE_PAGES)),
6238 K(zone->watermark_boost),
6239 K(min_wmark_pages(zone)),
6240 K(low_wmark_pages(zone)),
6241 K(high_wmark_pages(zone)),
6242 K(zone->nr_reserved_highatomic),
6243 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6244 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6245 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6246 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6247 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6248 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6249 K(zone->present_pages),
6250 K(zone_managed_pages(zone)),
6251 K(zone_page_state(zone, NR_MLOCK)),
6252 K(zone_page_state(zone, NR_BOUNCE)),
6254 K(this_cpu_read(zone->per_cpu_pageset->count)),
6255 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6256 printk("lowmem_reserve[]:");
6257 for (i = 0; i < MAX_NR_ZONES; i++)
6258 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6259 printk(KERN_CONT "\n");
6262 for_each_populated_zone(zone) {
6264 unsigned long nr[MAX_ORDER], flags, total = 0;
6265 unsigned char types[MAX_ORDER];
6267 if (zone_idx(zone) > max_zone_idx)
6269 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6272 printk(KERN_CONT "%s: ", zone->name);
6274 spin_lock_irqsave(&zone->lock, flags);
6275 for (order = 0; order < MAX_ORDER; order++) {
6276 struct free_area *area = &zone->free_area[order];
6279 nr[order] = area->nr_free;
6280 total += nr[order] << order;
6283 for (type = 0; type < MIGRATE_TYPES; type++) {
6284 if (!free_area_empty(area, type))
6285 types[order] |= 1 << type;
6288 spin_unlock_irqrestore(&zone->lock, flags);
6289 for (order = 0; order < MAX_ORDER; order++) {
6290 printk(KERN_CONT "%lu*%lukB ",
6291 nr[order], K(1UL) << order);
6293 show_migration_types(types[order]);
6295 printk(KERN_CONT "= %lukB\n", K(total));
6298 for_each_online_node(nid) {
6299 if (show_mem_node_skip(filter, nid, nodemask))
6301 hugetlb_show_meminfo_node(nid);
6304 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6306 show_swap_cache_info();
6309 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6311 zoneref->zone = zone;
6312 zoneref->zone_idx = zone_idx(zone);
6316 * Builds allocation fallback zone lists.
6318 * Add all populated zones of a node to the zonelist.
6320 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6323 enum zone_type zone_type = MAX_NR_ZONES;
6328 zone = pgdat->node_zones + zone_type;
6329 if (populated_zone(zone)) {
6330 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6331 check_highest_zone(zone_type);
6333 } while (zone_type);
6340 static int __parse_numa_zonelist_order(char *s)
6343 * We used to support different zonelists modes but they turned
6344 * out to be just not useful. Let's keep the warning in place
6345 * if somebody still use the cmd line parameter so that we do
6346 * not fail it silently
6348 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6349 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6355 char numa_zonelist_order[] = "Node";
6358 * sysctl handler for numa_zonelist_order
6360 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6361 void *buffer, size_t *length, loff_t *ppos)
6364 return __parse_numa_zonelist_order(buffer);
6365 return proc_dostring(table, write, buffer, length, ppos);
6369 static int node_load[MAX_NUMNODES];
6372 * find_next_best_node - find the next node that should appear in a given node's fallback list
6373 * @node: node whose fallback list we're appending
6374 * @used_node_mask: nodemask_t of already used nodes
6376 * We use a number of factors to determine which is the next node that should
6377 * appear on a given node's fallback list. The node should not have appeared
6378 * already in @node's fallback list, and it should be the next closest node
6379 * according to the distance array (which contains arbitrary distance values
6380 * from each node to each node in the system), and should also prefer nodes
6381 * with no CPUs, since presumably they'll have very little allocation pressure
6382 * on them otherwise.
6384 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6386 int find_next_best_node(int node, nodemask_t *used_node_mask)
6389 int min_val = INT_MAX;
6390 int best_node = NUMA_NO_NODE;
6392 /* Use the local node if we haven't already */
6393 if (!node_isset(node, *used_node_mask)) {
6394 node_set(node, *used_node_mask);
6398 for_each_node_state(n, N_MEMORY) {
6400 /* Don't want a node to appear more than once */
6401 if (node_isset(n, *used_node_mask))
6404 /* Use the distance array to find the distance */
6405 val = node_distance(node, n);
6407 /* Penalize nodes under us ("prefer the next node") */
6410 /* Give preference to headless and unused nodes */
6411 if (!cpumask_empty(cpumask_of_node(n)))
6412 val += PENALTY_FOR_NODE_WITH_CPUS;
6414 /* Slight preference for less loaded node */
6415 val *= MAX_NUMNODES;
6416 val += node_load[n];
6418 if (val < min_val) {
6425 node_set(best_node, *used_node_mask);
6432 * Build zonelists ordered by node and zones within node.
6433 * This results in maximum locality--normal zone overflows into local
6434 * DMA zone, if any--but risks exhausting DMA zone.
6436 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6439 struct zoneref *zonerefs;
6442 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6444 for (i = 0; i < nr_nodes; i++) {
6447 pg_data_t *node = NODE_DATA(node_order[i]);
6449 nr_zones = build_zonerefs_node(node, zonerefs);
6450 zonerefs += nr_zones;
6452 zonerefs->zone = NULL;
6453 zonerefs->zone_idx = 0;
6457 * Build gfp_thisnode zonelists
6459 static void build_thisnode_zonelists(pg_data_t *pgdat)
6461 struct zoneref *zonerefs;
6464 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6465 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6466 zonerefs += nr_zones;
6467 zonerefs->zone = NULL;
6468 zonerefs->zone_idx = 0;
6472 * Build zonelists ordered by zone and nodes within zones.
6473 * This results in conserving DMA zone[s] until all Normal memory is
6474 * exhausted, but results in overflowing to remote node while memory
6475 * may still exist in local DMA zone.
6478 static void build_zonelists(pg_data_t *pgdat)
6480 static int node_order[MAX_NUMNODES];
6481 int node, nr_nodes = 0;
6482 nodemask_t used_mask = NODE_MASK_NONE;
6483 int local_node, prev_node;
6485 /* NUMA-aware ordering of nodes */
6486 local_node = pgdat->node_id;
6487 prev_node = local_node;
6489 memset(node_order, 0, sizeof(node_order));
6490 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6492 * We don't want to pressure a particular node.
6493 * So adding penalty to the first node in same
6494 * distance group to make it round-robin.
6496 if (node_distance(local_node, node) !=
6497 node_distance(local_node, prev_node))
6498 node_load[node] += 1;
6500 node_order[nr_nodes++] = node;
6504 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6505 build_thisnode_zonelists(pgdat);
6506 pr_info("Fallback order for Node %d: ", local_node);
6507 for (node = 0; node < nr_nodes; node++)
6508 pr_cont("%d ", node_order[node]);
6512 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6514 * Return node id of node used for "local" allocations.
6515 * I.e., first node id of first zone in arg node's generic zonelist.
6516 * Used for initializing percpu 'numa_mem', which is used primarily
6517 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6519 int local_memory_node(int node)
6523 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6524 gfp_zone(GFP_KERNEL),
6526 return zone_to_nid(z->zone);
6530 static void setup_min_unmapped_ratio(void);
6531 static void setup_min_slab_ratio(void);
6532 #else /* CONFIG_NUMA */
6534 static void build_zonelists(pg_data_t *pgdat)
6536 int node, local_node;
6537 struct zoneref *zonerefs;
6540 local_node = pgdat->node_id;
6542 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6543 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6544 zonerefs += nr_zones;
6547 * Now we build the zonelist so that it contains the zones
6548 * of all the other nodes.
6549 * We don't want to pressure a particular node, so when
6550 * building the zones for node N, we make sure that the
6551 * zones coming right after the local ones are those from
6552 * node N+1 (modulo N)
6554 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6555 if (!node_online(node))
6557 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6558 zonerefs += nr_zones;
6560 for (node = 0; node < local_node; node++) {
6561 if (!node_online(node))
6563 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6564 zonerefs += nr_zones;
6567 zonerefs->zone = NULL;
6568 zonerefs->zone_idx = 0;
6571 #endif /* CONFIG_NUMA */
6574 * Boot pageset table. One per cpu which is going to be used for all
6575 * zones and all nodes. The parameters will be set in such a way
6576 * that an item put on a list will immediately be handed over to
6577 * the buddy list. This is safe since pageset manipulation is done
6578 * with interrupts disabled.
6580 * The boot_pagesets must be kept even after bootup is complete for
6581 * unused processors and/or zones. They do play a role for bootstrapping
6582 * hotplugged processors.
6584 * zoneinfo_show() and maybe other functions do
6585 * not check if the processor is online before following the pageset pointer.
6586 * Other parts of the kernel may not check if the zone is available.
6588 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6589 /* These effectively disable the pcplists in the boot pageset completely */
6590 #define BOOT_PAGESET_HIGH 0
6591 #define BOOT_PAGESET_BATCH 1
6592 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6593 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6594 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6596 static void __build_all_zonelists(void *data)
6599 int __maybe_unused cpu;
6600 pg_data_t *self = data;
6602 write_seqlock(&zonelist_update_seq);
6605 memset(node_load, 0, sizeof(node_load));
6609 * This node is hotadded and no memory is yet present. So just
6610 * building zonelists is fine - no need to touch other nodes.
6612 if (self && !node_online(self->node_id)) {
6613 build_zonelists(self);
6616 * All possible nodes have pgdat preallocated
6619 for_each_node(nid) {
6620 pg_data_t *pgdat = NODE_DATA(nid);
6622 build_zonelists(pgdat);
6625 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6627 * We now know the "local memory node" for each node--
6628 * i.e., the node of the first zone in the generic zonelist.
6629 * Set up numa_mem percpu variable for on-line cpus. During
6630 * boot, only the boot cpu should be on-line; we'll init the
6631 * secondary cpus' numa_mem as they come on-line. During
6632 * node/memory hotplug, we'll fixup all on-line cpus.
6634 for_each_online_cpu(cpu)
6635 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6639 write_sequnlock(&zonelist_update_seq);
6642 static noinline void __init
6643 build_all_zonelists_init(void)
6647 __build_all_zonelists(NULL);
6650 * Initialize the boot_pagesets that are going to be used
6651 * for bootstrapping processors. The real pagesets for
6652 * each zone will be allocated later when the per cpu
6653 * allocator is available.
6655 * boot_pagesets are used also for bootstrapping offline
6656 * cpus if the system is already booted because the pagesets
6657 * are needed to initialize allocators on a specific cpu too.
6658 * F.e. the percpu allocator needs the page allocator which
6659 * needs the percpu allocator in order to allocate its pagesets
6660 * (a chicken-egg dilemma).
6662 for_each_possible_cpu(cpu)
6663 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6665 mminit_verify_zonelist();
6666 cpuset_init_current_mems_allowed();
6670 * unless system_state == SYSTEM_BOOTING.
6672 * __ref due to call of __init annotated helper build_all_zonelists_init
6673 * [protected by SYSTEM_BOOTING].
6675 void __ref build_all_zonelists(pg_data_t *pgdat)
6677 unsigned long vm_total_pages;
6679 if (system_state == SYSTEM_BOOTING) {
6680 build_all_zonelists_init();
6682 __build_all_zonelists(pgdat);
6683 /* cpuset refresh routine should be here */
6685 /* Get the number of free pages beyond high watermark in all zones. */
6686 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6688 * Disable grouping by mobility if the number of pages in the
6689 * system is too low to allow the mechanism to work. It would be
6690 * more accurate, but expensive to check per-zone. This check is
6691 * made on memory-hotadd so a system can start with mobility
6692 * disabled and enable it later
6694 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6695 page_group_by_mobility_disabled = 1;
6697 page_group_by_mobility_disabled = 0;
6699 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6701 page_group_by_mobility_disabled ? "off" : "on",
6704 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6708 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6709 static bool __meminit
6710 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6712 static struct memblock_region *r;
6714 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6715 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6716 for_each_mem_region(r) {
6717 if (*pfn < memblock_region_memory_end_pfn(r))
6721 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6722 memblock_is_mirror(r)) {
6723 *pfn = memblock_region_memory_end_pfn(r);
6731 * Initially all pages are reserved - free ones are freed
6732 * up by memblock_free_all() once the early boot process is
6733 * done. Non-atomic initialization, single-pass.
6735 * All aligned pageblocks are initialized to the specified migratetype
6736 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6737 * zone stats (e.g., nr_isolate_pageblock) are touched.
6739 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6740 unsigned long start_pfn, unsigned long zone_end_pfn,
6741 enum meminit_context context,
6742 struct vmem_altmap *altmap, int migratetype)
6744 unsigned long pfn, end_pfn = start_pfn + size;
6747 if (highest_memmap_pfn < end_pfn - 1)
6748 highest_memmap_pfn = end_pfn - 1;
6750 #ifdef CONFIG_ZONE_DEVICE
6752 * Honor reservation requested by the driver for this ZONE_DEVICE
6753 * memory. We limit the total number of pages to initialize to just
6754 * those that might contain the memory mapping. We will defer the
6755 * ZONE_DEVICE page initialization until after we have released
6758 if (zone == ZONE_DEVICE) {
6762 if (start_pfn == altmap->base_pfn)
6763 start_pfn += altmap->reserve;
6764 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6768 for (pfn = start_pfn; pfn < end_pfn; ) {
6770 * There can be holes in boot-time mem_map[]s handed to this
6771 * function. They do not exist on hotplugged memory.
6773 if (context == MEMINIT_EARLY) {
6774 if (overlap_memmap_init(zone, &pfn))
6776 if (defer_init(nid, pfn, zone_end_pfn))
6780 page = pfn_to_page(pfn);
6781 __init_single_page(page, pfn, zone, nid);
6782 if (context == MEMINIT_HOTPLUG)
6783 __SetPageReserved(page);
6786 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6787 * such that unmovable allocations won't be scattered all
6788 * over the place during system boot.
6790 if (pageblock_aligned(pfn)) {
6791 set_pageblock_migratetype(page, migratetype);
6798 #ifdef CONFIG_ZONE_DEVICE
6799 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6800 unsigned long zone_idx, int nid,
6801 struct dev_pagemap *pgmap)
6804 __init_single_page(page, pfn, zone_idx, nid);
6807 * Mark page reserved as it will need to wait for onlining
6808 * phase for it to be fully associated with a zone.
6810 * We can use the non-atomic __set_bit operation for setting
6811 * the flag as we are still initializing the pages.
6813 __SetPageReserved(page);
6816 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6817 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6818 * ever freed or placed on a driver-private list.
6820 page->pgmap = pgmap;
6821 page->zone_device_data = NULL;
6824 * Mark the block movable so that blocks are reserved for
6825 * movable at startup. This will force kernel allocations
6826 * to reserve their blocks rather than leaking throughout
6827 * the address space during boot when many long-lived
6828 * kernel allocations are made.
6830 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6831 * because this is done early in section_activate()
6833 if (pageblock_aligned(pfn)) {
6834 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6839 * ZONE_DEVICE pages are released directly to the driver page allocator
6840 * which will set the page count to 1 when allocating the page.
6842 if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
6843 pgmap->type == MEMORY_DEVICE_COHERENT)
6844 set_page_count(page, 0);
6848 * With compound page geometry and when struct pages are stored in ram most
6849 * tail pages are reused. Consequently, the amount of unique struct pages to
6850 * initialize is a lot smaller that the total amount of struct pages being
6851 * mapped. This is a paired / mild layering violation with explicit knowledge
6852 * of how the sparse_vmemmap internals handle compound pages in the lack
6853 * of an altmap. See vmemmap_populate_compound_pages().
6855 static inline unsigned long compound_nr_pages(struct vmem_altmap *altmap,
6856 unsigned long nr_pages)
6858 return is_power_of_2(sizeof(struct page)) &&
6859 !altmap ? 2 * (PAGE_SIZE / sizeof(struct page)) : nr_pages;
6862 static void __ref memmap_init_compound(struct page *head,
6863 unsigned long head_pfn,
6864 unsigned long zone_idx, int nid,
6865 struct dev_pagemap *pgmap,
6866 unsigned long nr_pages)
6868 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6869 unsigned int order = pgmap->vmemmap_shift;
6871 __SetPageHead(head);
6872 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6873 struct page *page = pfn_to_page(pfn);
6875 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6876 prep_compound_tail(head, pfn - head_pfn);
6877 set_page_count(page, 0);
6880 * The first tail page stores compound_mapcount_ptr() and
6881 * compound_order() and the second tail page stores
6882 * compound_pincount_ptr(). Call prep_compound_head() after
6883 * the first and second tail pages have been initialized to
6884 * not have the data overwritten.
6886 if (pfn == head_pfn + 2)
6887 prep_compound_head(head, order);
6891 void __ref memmap_init_zone_device(struct zone *zone,
6892 unsigned long start_pfn,
6893 unsigned long nr_pages,
6894 struct dev_pagemap *pgmap)
6896 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6897 struct pglist_data *pgdat = zone->zone_pgdat;
6898 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6899 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6900 unsigned long zone_idx = zone_idx(zone);
6901 unsigned long start = jiffies;
6902 int nid = pgdat->node_id;
6904 if (WARN_ON_ONCE(!pgmap || zone_idx != ZONE_DEVICE))
6908 * The call to memmap_init should have already taken care
6909 * of the pages reserved for the memmap, so we can just jump to
6910 * the end of that region and start processing the device pages.
6913 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6914 nr_pages = end_pfn - start_pfn;
6917 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6918 struct page *page = pfn_to_page(pfn);
6920 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6922 if (pfns_per_compound == 1)
6925 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6926 compound_nr_pages(altmap, pfns_per_compound));
6929 pr_info("%s initialised %lu pages in %ums\n", __func__,
6930 nr_pages, jiffies_to_msecs(jiffies - start));
6934 static void __meminit zone_init_free_lists(struct zone *zone)
6936 unsigned int order, t;
6937 for_each_migratetype_order(order, t) {
6938 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6939 zone->free_area[order].nr_free = 0;
6944 * Only struct pages that correspond to ranges defined by memblock.memory
6945 * are zeroed and initialized by going through __init_single_page() during
6946 * memmap_init_zone_range().
6948 * But, there could be struct pages that correspond to holes in
6949 * memblock.memory. This can happen because of the following reasons:
6950 * - physical memory bank size is not necessarily the exact multiple of the
6951 * arbitrary section size
6952 * - early reserved memory may not be listed in memblock.memory
6953 * - memory layouts defined with memmap= kernel parameter may not align
6954 * nicely with memmap sections
6956 * Explicitly initialize those struct pages so that:
6957 * - PG_Reserved is set
6958 * - zone and node links point to zone and node that span the page if the
6959 * hole is in the middle of a zone
6960 * - zone and node links point to adjacent zone/node if the hole falls on
6961 * the zone boundary; the pages in such holes will be prepended to the
6962 * zone/node above the hole except for the trailing pages in the last
6963 * section that will be appended to the zone/node below.
6965 static void __init init_unavailable_range(unsigned long spfn,
6972 for (pfn = spfn; pfn < epfn; pfn++) {
6973 if (!pfn_valid(pageblock_start_pfn(pfn))) {
6974 pfn = pageblock_end_pfn(pfn) - 1;
6977 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6978 __SetPageReserved(pfn_to_page(pfn));
6983 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6984 node, zone_names[zone], pgcnt);
6987 static void __init memmap_init_zone_range(struct zone *zone,
6988 unsigned long start_pfn,
6989 unsigned long end_pfn,
6990 unsigned long *hole_pfn)
6992 unsigned long zone_start_pfn = zone->zone_start_pfn;
6993 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6994 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6996 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6997 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6999 if (start_pfn >= end_pfn)
7002 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
7003 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
7005 if (*hole_pfn < start_pfn)
7006 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
7008 *hole_pfn = end_pfn;
7011 static void __init memmap_init(void)
7013 unsigned long start_pfn, end_pfn;
7014 unsigned long hole_pfn = 0;
7015 int i, j, zone_id = 0, nid;
7017 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7018 struct pglist_data *node = NODE_DATA(nid);
7020 for (j = 0; j < MAX_NR_ZONES; j++) {
7021 struct zone *zone = node->node_zones + j;
7023 if (!populated_zone(zone))
7026 memmap_init_zone_range(zone, start_pfn, end_pfn,
7032 #ifdef CONFIG_SPARSEMEM
7034 * Initialize the memory map for hole in the range [memory_end,
7036 * Append the pages in this hole to the highest zone in the last
7038 * The call to init_unavailable_range() is outside the ifdef to
7039 * silence the compiler warining about zone_id set but not used;
7040 * for FLATMEM it is a nop anyway
7042 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
7043 if (hole_pfn < end_pfn)
7045 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
7048 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
7049 phys_addr_t min_addr, int nid, bool exact_nid)
7054 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
7055 MEMBLOCK_ALLOC_ACCESSIBLE,
7058 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
7059 MEMBLOCK_ALLOC_ACCESSIBLE,
7062 if (ptr && size > 0)
7063 page_init_poison(ptr, size);
7068 static int zone_batchsize(struct zone *zone)
7074 * The number of pages to batch allocate is either ~0.1%
7075 * of the zone or 1MB, whichever is smaller. The batch
7076 * size is striking a balance between allocation latency
7077 * and zone lock contention.
7079 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
7080 batch /= 4; /* We effectively *= 4 below */
7085 * Clamp the batch to a 2^n - 1 value. Having a power
7086 * of 2 value was found to be more likely to have
7087 * suboptimal cache aliasing properties in some cases.
7089 * For example if 2 tasks are alternately allocating
7090 * batches of pages, one task can end up with a lot
7091 * of pages of one half of the possible page colors
7092 * and the other with pages of the other colors.
7094 batch = rounddown_pow_of_two(batch + batch/2) - 1;
7099 /* The deferral and batching of frees should be suppressed under NOMMU
7102 * The problem is that NOMMU needs to be able to allocate large chunks
7103 * of contiguous memory as there's no hardware page translation to
7104 * assemble apparent contiguous memory from discontiguous pages.
7106 * Queueing large contiguous runs of pages for batching, however,
7107 * causes the pages to actually be freed in smaller chunks. As there
7108 * can be a significant delay between the individual batches being
7109 * recycled, this leads to the once large chunks of space being
7110 * fragmented and becoming unavailable for high-order allocations.
7116 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
7121 unsigned long total_pages;
7123 if (!percpu_pagelist_high_fraction) {
7125 * By default, the high value of the pcp is based on the zone
7126 * low watermark so that if they are full then background
7127 * reclaim will not be started prematurely.
7129 total_pages = low_wmark_pages(zone);
7132 * If percpu_pagelist_high_fraction is configured, the high
7133 * value is based on a fraction of the managed pages in the
7136 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
7140 * Split the high value across all online CPUs local to the zone. Note
7141 * that early in boot that CPUs may not be online yet and that during
7142 * CPU hotplug that the cpumask is not yet updated when a CPU is being
7143 * onlined. For memory nodes that have no CPUs, split pcp->high across
7144 * all online CPUs to mitigate the risk that reclaim is triggered
7145 * prematurely due to pages stored on pcp lists.
7147 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
7149 nr_split_cpus = num_online_cpus();
7150 high = total_pages / nr_split_cpus;
7153 * Ensure high is at least batch*4. The multiple is based on the
7154 * historical relationship between high and batch.
7156 high = max(high, batch << 2);
7165 * pcp->high and pcp->batch values are related and generally batch is lower
7166 * than high. They are also related to pcp->count such that count is lower
7167 * than high, and as soon as it reaches high, the pcplist is flushed.
7169 * However, guaranteeing these relations at all times would require e.g. write
7170 * barriers here but also careful usage of read barriers at the read side, and
7171 * thus be prone to error and bad for performance. Thus the update only prevents
7172 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
7173 * can cope with those fields changing asynchronously, and fully trust only the
7174 * pcp->count field on the local CPU with interrupts disabled.
7176 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
7177 * outside of boot time (or some other assurance that no concurrent updaters
7180 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
7181 unsigned long batch)
7183 WRITE_ONCE(pcp->batch, batch);
7184 WRITE_ONCE(pcp->high, high);
7187 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
7191 memset(pcp, 0, sizeof(*pcp));
7192 memset(pzstats, 0, sizeof(*pzstats));
7194 spin_lock_init(&pcp->lock);
7195 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
7196 INIT_LIST_HEAD(&pcp->lists[pindex]);
7199 * Set batch and high values safe for a boot pageset. A true percpu
7200 * pageset's initialization will update them subsequently. Here we don't
7201 * need to be as careful as pageset_update() as nobody can access the
7204 pcp->high = BOOT_PAGESET_HIGH;
7205 pcp->batch = BOOT_PAGESET_BATCH;
7206 pcp->free_factor = 0;
7209 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
7210 unsigned long batch)
7212 struct per_cpu_pages *pcp;
7215 for_each_possible_cpu(cpu) {
7216 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7217 pageset_update(pcp, high, batch);
7222 * Calculate and set new high and batch values for all per-cpu pagesets of a
7223 * zone based on the zone's size.
7225 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
7227 int new_high, new_batch;
7229 new_batch = max(1, zone_batchsize(zone));
7230 new_high = zone_highsize(zone, new_batch, cpu_online);
7232 if (zone->pageset_high == new_high &&
7233 zone->pageset_batch == new_batch)
7236 zone->pageset_high = new_high;
7237 zone->pageset_batch = new_batch;
7239 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7242 void __meminit setup_zone_pageset(struct zone *zone)
7246 /* Size may be 0 on !SMP && !NUMA */
7247 if (sizeof(struct per_cpu_zonestat) > 0)
7248 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7250 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7251 for_each_possible_cpu(cpu) {
7252 struct per_cpu_pages *pcp;
7253 struct per_cpu_zonestat *pzstats;
7255 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7256 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7257 per_cpu_pages_init(pcp, pzstats);
7260 zone_set_pageset_high_and_batch(zone, 0);
7264 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7265 * page high values need to be recalculated.
7267 static void zone_pcp_update(struct zone *zone, int cpu_online)
7269 mutex_lock(&pcp_batch_high_lock);
7270 zone_set_pageset_high_and_batch(zone, cpu_online);
7271 mutex_unlock(&pcp_batch_high_lock);
7275 * Allocate per cpu pagesets and initialize them.
7276 * Before this call only boot pagesets were available.
7278 void __init setup_per_cpu_pageset(void)
7280 struct pglist_data *pgdat;
7282 int __maybe_unused cpu;
7284 for_each_populated_zone(zone)
7285 setup_zone_pageset(zone);
7289 * Unpopulated zones continue using the boot pagesets.
7290 * The numa stats for these pagesets need to be reset.
7291 * Otherwise, they will end up skewing the stats of
7292 * the nodes these zones are associated with.
7294 for_each_possible_cpu(cpu) {
7295 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7296 memset(pzstats->vm_numa_event, 0,
7297 sizeof(pzstats->vm_numa_event));
7301 for_each_online_pgdat(pgdat)
7302 pgdat->per_cpu_nodestats =
7303 alloc_percpu(struct per_cpu_nodestat);
7306 static __meminit void zone_pcp_init(struct zone *zone)
7309 * per cpu subsystem is not up at this point. The following code
7310 * relies on the ability of the linker to provide the
7311 * offset of a (static) per cpu variable into the per cpu area.
7313 zone->per_cpu_pageset = &boot_pageset;
7314 zone->per_cpu_zonestats = &boot_zonestats;
7315 zone->pageset_high = BOOT_PAGESET_HIGH;
7316 zone->pageset_batch = BOOT_PAGESET_BATCH;
7318 if (populated_zone(zone))
7319 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7320 zone->present_pages, zone_batchsize(zone));
7323 void __meminit init_currently_empty_zone(struct zone *zone,
7324 unsigned long zone_start_pfn,
7327 struct pglist_data *pgdat = zone->zone_pgdat;
7328 int zone_idx = zone_idx(zone) + 1;
7330 if (zone_idx > pgdat->nr_zones)
7331 pgdat->nr_zones = zone_idx;
7333 zone->zone_start_pfn = zone_start_pfn;
7335 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7336 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7338 (unsigned long)zone_idx(zone),
7339 zone_start_pfn, (zone_start_pfn + size));
7341 zone_init_free_lists(zone);
7342 zone->initialized = 1;
7346 * get_pfn_range_for_nid - Return the start and end page frames for a node
7347 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7348 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7349 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7351 * It returns the start and end page frame of a node based on information
7352 * provided by memblock_set_node(). If called for a node
7353 * with no available memory, a warning is printed and the start and end
7356 void __init get_pfn_range_for_nid(unsigned int nid,
7357 unsigned long *start_pfn, unsigned long *end_pfn)
7359 unsigned long this_start_pfn, this_end_pfn;
7365 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7366 *start_pfn = min(*start_pfn, this_start_pfn);
7367 *end_pfn = max(*end_pfn, this_end_pfn);
7370 if (*start_pfn == -1UL)
7375 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7376 * assumption is made that zones within a node are ordered in monotonic
7377 * increasing memory addresses so that the "highest" populated zone is used
7379 static void __init find_usable_zone_for_movable(void)
7382 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7383 if (zone_index == ZONE_MOVABLE)
7386 if (arch_zone_highest_possible_pfn[zone_index] >
7387 arch_zone_lowest_possible_pfn[zone_index])
7391 VM_BUG_ON(zone_index == -1);
7392 movable_zone = zone_index;
7396 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7397 * because it is sized independent of architecture. Unlike the other zones,
7398 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7399 * in each node depending on the size of each node and how evenly kernelcore
7400 * is distributed. This helper function adjusts the zone ranges
7401 * provided by the architecture for a given node by using the end of the
7402 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7403 * zones within a node are in order of monotonic increases memory addresses
7405 static void __init adjust_zone_range_for_zone_movable(int nid,
7406 unsigned long zone_type,
7407 unsigned long node_start_pfn,
7408 unsigned long node_end_pfn,
7409 unsigned long *zone_start_pfn,
7410 unsigned long *zone_end_pfn)
7412 /* Only adjust if ZONE_MOVABLE is on this node */
7413 if (zone_movable_pfn[nid]) {
7414 /* Size ZONE_MOVABLE */
7415 if (zone_type == ZONE_MOVABLE) {
7416 *zone_start_pfn = zone_movable_pfn[nid];
7417 *zone_end_pfn = min(node_end_pfn,
7418 arch_zone_highest_possible_pfn[movable_zone]);
7420 /* Adjust for ZONE_MOVABLE starting within this range */
7421 } else if (!mirrored_kernelcore &&
7422 *zone_start_pfn < zone_movable_pfn[nid] &&
7423 *zone_end_pfn > zone_movable_pfn[nid]) {
7424 *zone_end_pfn = zone_movable_pfn[nid];
7426 /* Check if this whole range is within ZONE_MOVABLE */
7427 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7428 *zone_start_pfn = *zone_end_pfn;
7433 * Return the number of pages a zone spans in a node, including holes
7434 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7436 static unsigned long __init zone_spanned_pages_in_node(int nid,
7437 unsigned long zone_type,
7438 unsigned long node_start_pfn,
7439 unsigned long node_end_pfn,
7440 unsigned long *zone_start_pfn,
7441 unsigned long *zone_end_pfn)
7443 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7444 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7445 /* When hotadd a new node from cpu_up(), the node should be empty */
7446 if (!node_start_pfn && !node_end_pfn)
7449 /* Get the start and end of the zone */
7450 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7451 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7452 adjust_zone_range_for_zone_movable(nid, zone_type,
7453 node_start_pfn, node_end_pfn,
7454 zone_start_pfn, zone_end_pfn);
7456 /* Check that this node has pages within the zone's required range */
7457 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7460 /* Move the zone boundaries inside the node if necessary */
7461 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7462 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7464 /* Return the spanned pages */
7465 return *zone_end_pfn - *zone_start_pfn;
7469 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7470 * then all holes in the requested range will be accounted for.
7472 unsigned long __init __absent_pages_in_range(int nid,
7473 unsigned long range_start_pfn,
7474 unsigned long range_end_pfn)
7476 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7477 unsigned long start_pfn, end_pfn;
7480 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7481 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7482 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7483 nr_absent -= end_pfn - start_pfn;
7489 * absent_pages_in_range - Return number of page frames in holes within a range
7490 * @start_pfn: The start PFN to start searching for holes
7491 * @end_pfn: The end PFN to stop searching for holes
7493 * Return: the number of pages frames in memory holes within a range.
7495 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7496 unsigned long end_pfn)
7498 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7501 /* Return the number of page frames in holes in a zone on a node */
7502 static unsigned long __init zone_absent_pages_in_node(int nid,
7503 unsigned long zone_type,
7504 unsigned long node_start_pfn,
7505 unsigned long node_end_pfn)
7507 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7508 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7509 unsigned long zone_start_pfn, zone_end_pfn;
7510 unsigned long nr_absent;
7512 /* When hotadd a new node from cpu_up(), the node should be empty */
7513 if (!node_start_pfn && !node_end_pfn)
7516 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7517 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7519 adjust_zone_range_for_zone_movable(nid, zone_type,
7520 node_start_pfn, node_end_pfn,
7521 &zone_start_pfn, &zone_end_pfn);
7522 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7525 * ZONE_MOVABLE handling.
7526 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7529 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7530 unsigned long start_pfn, end_pfn;
7531 struct memblock_region *r;
7533 for_each_mem_region(r) {
7534 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7535 zone_start_pfn, zone_end_pfn);
7536 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7537 zone_start_pfn, zone_end_pfn);
7539 if (zone_type == ZONE_MOVABLE &&
7540 memblock_is_mirror(r))
7541 nr_absent += end_pfn - start_pfn;
7543 if (zone_type == ZONE_NORMAL &&
7544 !memblock_is_mirror(r))
7545 nr_absent += end_pfn - start_pfn;
7552 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7553 unsigned long node_start_pfn,
7554 unsigned long node_end_pfn)
7556 unsigned long realtotalpages = 0, totalpages = 0;
7559 for (i = 0; i < MAX_NR_ZONES; i++) {
7560 struct zone *zone = pgdat->node_zones + i;
7561 unsigned long zone_start_pfn, zone_end_pfn;
7562 unsigned long spanned, absent;
7563 unsigned long size, real_size;
7565 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7570 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7575 real_size = size - absent;
7578 zone->zone_start_pfn = zone_start_pfn;
7580 zone->zone_start_pfn = 0;
7581 zone->spanned_pages = size;
7582 zone->present_pages = real_size;
7583 #if defined(CONFIG_MEMORY_HOTPLUG)
7584 zone->present_early_pages = real_size;
7588 realtotalpages += real_size;
7591 pgdat->node_spanned_pages = totalpages;
7592 pgdat->node_present_pages = realtotalpages;
7593 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7596 #ifndef CONFIG_SPARSEMEM
7598 * Calculate the size of the zone->blockflags rounded to an unsigned long
7599 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7600 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7601 * round what is now in bits to nearest long in bits, then return it in
7604 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7606 unsigned long usemapsize;
7608 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7609 usemapsize = roundup(zonesize, pageblock_nr_pages);
7610 usemapsize = usemapsize >> pageblock_order;
7611 usemapsize *= NR_PAGEBLOCK_BITS;
7612 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7614 return usemapsize / 8;
7617 static void __ref setup_usemap(struct zone *zone)
7619 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7620 zone->spanned_pages);
7621 zone->pageblock_flags = NULL;
7623 zone->pageblock_flags =
7624 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7626 if (!zone->pageblock_flags)
7627 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7628 usemapsize, zone->name, zone_to_nid(zone));
7632 static inline void setup_usemap(struct zone *zone) {}
7633 #endif /* CONFIG_SPARSEMEM */
7635 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7637 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7638 void __init set_pageblock_order(void)
7640 unsigned int order = MAX_ORDER - 1;
7642 /* Check that pageblock_nr_pages has not already been setup */
7643 if (pageblock_order)
7646 /* Don't let pageblocks exceed the maximum allocation granularity. */
7647 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7648 order = HUGETLB_PAGE_ORDER;
7651 * Assume the largest contiguous order of interest is a huge page.
7652 * This value may be variable depending on boot parameters on IA64 and
7655 pageblock_order = order;
7657 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7660 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7661 * is unused as pageblock_order is set at compile-time. See
7662 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7665 void __init set_pageblock_order(void)
7669 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7671 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7672 unsigned long present_pages)
7674 unsigned long pages = spanned_pages;
7677 * Provide a more accurate estimation if there are holes within
7678 * the zone and SPARSEMEM is in use. If there are holes within the
7679 * zone, each populated memory region may cost us one or two extra
7680 * memmap pages due to alignment because memmap pages for each
7681 * populated regions may not be naturally aligned on page boundary.
7682 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7684 if (spanned_pages > present_pages + (present_pages >> 4) &&
7685 IS_ENABLED(CONFIG_SPARSEMEM))
7686 pages = present_pages;
7688 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7691 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7692 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7694 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7696 spin_lock_init(&ds_queue->split_queue_lock);
7697 INIT_LIST_HEAD(&ds_queue->split_queue);
7698 ds_queue->split_queue_len = 0;
7701 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7704 #ifdef CONFIG_COMPACTION
7705 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7707 init_waitqueue_head(&pgdat->kcompactd_wait);
7710 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7713 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7717 pgdat_resize_init(pgdat);
7718 pgdat_kswapd_lock_init(pgdat);
7720 pgdat_init_split_queue(pgdat);
7721 pgdat_init_kcompactd(pgdat);
7723 init_waitqueue_head(&pgdat->kswapd_wait);
7724 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7726 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7727 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7729 pgdat_page_ext_init(pgdat);
7730 lruvec_init(&pgdat->__lruvec);
7733 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7734 unsigned long remaining_pages)
7736 atomic_long_set(&zone->managed_pages, remaining_pages);
7737 zone_set_nid(zone, nid);
7738 zone->name = zone_names[idx];
7739 zone->zone_pgdat = NODE_DATA(nid);
7740 spin_lock_init(&zone->lock);
7741 zone_seqlock_init(zone);
7742 zone_pcp_init(zone);
7746 * Set up the zone data structures
7747 * - init pgdat internals
7748 * - init all zones belonging to this node
7750 * NOTE: this function is only called during memory hotplug
7752 #ifdef CONFIG_MEMORY_HOTPLUG
7753 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7755 int nid = pgdat->node_id;
7759 pgdat_init_internals(pgdat);
7761 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7762 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7765 * Reset the nr_zones, order and highest_zoneidx before reuse.
7766 * Note that kswapd will init kswapd_highest_zoneidx properly
7767 * when it starts in the near future.
7769 pgdat->nr_zones = 0;
7770 pgdat->kswapd_order = 0;
7771 pgdat->kswapd_highest_zoneidx = 0;
7772 pgdat->node_start_pfn = 0;
7773 for_each_online_cpu(cpu) {
7774 struct per_cpu_nodestat *p;
7776 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7777 memset(p, 0, sizeof(*p));
7780 for (z = 0; z < MAX_NR_ZONES; z++)
7781 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7786 * Set up the zone data structures:
7787 * - mark all pages reserved
7788 * - mark all memory queues empty
7789 * - clear the memory bitmaps
7791 * NOTE: pgdat should get zeroed by caller.
7792 * NOTE: this function is only called during early init.
7794 static void __init free_area_init_core(struct pglist_data *pgdat)
7797 int nid = pgdat->node_id;
7799 pgdat_init_internals(pgdat);
7800 pgdat->per_cpu_nodestats = &boot_nodestats;
7802 for (j = 0; j < MAX_NR_ZONES; j++) {
7803 struct zone *zone = pgdat->node_zones + j;
7804 unsigned long size, freesize, memmap_pages;
7806 size = zone->spanned_pages;
7807 freesize = zone->present_pages;
7810 * Adjust freesize so that it accounts for how much memory
7811 * is used by this zone for memmap. This affects the watermark
7812 * and per-cpu initialisations
7814 memmap_pages = calc_memmap_size(size, freesize);
7815 if (!is_highmem_idx(j)) {
7816 if (freesize >= memmap_pages) {
7817 freesize -= memmap_pages;
7819 pr_debug(" %s zone: %lu pages used for memmap\n",
7820 zone_names[j], memmap_pages);
7822 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7823 zone_names[j], memmap_pages, freesize);
7826 /* Account for reserved pages */
7827 if (j == 0 && freesize > dma_reserve) {
7828 freesize -= dma_reserve;
7829 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7832 if (!is_highmem_idx(j))
7833 nr_kernel_pages += freesize;
7834 /* Charge for highmem memmap if there are enough kernel pages */
7835 else if (nr_kernel_pages > memmap_pages * 2)
7836 nr_kernel_pages -= memmap_pages;
7837 nr_all_pages += freesize;
7840 * Set an approximate value for lowmem here, it will be adjusted
7841 * when the bootmem allocator frees pages into the buddy system.
7842 * And all highmem pages will be managed by the buddy system.
7844 zone_init_internals(zone, j, nid, freesize);
7849 set_pageblock_order();
7851 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7855 #ifdef CONFIG_FLATMEM
7856 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7858 unsigned long __maybe_unused start = 0;
7859 unsigned long __maybe_unused offset = 0;
7861 /* Skip empty nodes */
7862 if (!pgdat->node_spanned_pages)
7865 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7866 offset = pgdat->node_start_pfn - start;
7867 /* ia64 gets its own node_mem_map, before this, without bootmem */
7868 if (!pgdat->node_mem_map) {
7869 unsigned long size, end;
7873 * The zone's endpoints aren't required to be MAX_ORDER
7874 * aligned but the node_mem_map endpoints must be in order
7875 * for the buddy allocator to function correctly.
7877 end = pgdat_end_pfn(pgdat);
7878 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7879 size = (end - start) * sizeof(struct page);
7880 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7881 pgdat->node_id, false);
7883 panic("Failed to allocate %ld bytes for node %d memory map\n",
7884 size, pgdat->node_id);
7885 pgdat->node_mem_map = map + offset;
7887 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7888 __func__, pgdat->node_id, (unsigned long)pgdat,
7889 (unsigned long)pgdat->node_mem_map);
7892 * With no DISCONTIG, the global mem_map is just set as node 0's
7894 if (pgdat == NODE_DATA(0)) {
7895 mem_map = NODE_DATA(0)->node_mem_map;
7896 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7902 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7903 #endif /* CONFIG_FLATMEM */
7905 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7906 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7908 pgdat->first_deferred_pfn = ULONG_MAX;
7911 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7914 static void __init free_area_init_node(int nid)
7916 pg_data_t *pgdat = NODE_DATA(nid);
7917 unsigned long start_pfn = 0;
7918 unsigned long end_pfn = 0;
7920 /* pg_data_t should be reset to zero when it's allocated */
7921 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7923 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7925 pgdat->node_id = nid;
7926 pgdat->node_start_pfn = start_pfn;
7927 pgdat->per_cpu_nodestats = NULL;
7929 if (start_pfn != end_pfn) {
7930 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7931 (u64)start_pfn << PAGE_SHIFT,
7932 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7934 pr_info("Initmem setup node %d as memoryless\n", nid);
7937 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7939 alloc_node_mem_map(pgdat);
7940 pgdat_set_deferred_range(pgdat);
7942 free_area_init_core(pgdat);
7945 static void __init free_area_init_memoryless_node(int nid)
7947 free_area_init_node(nid);
7950 #if MAX_NUMNODES > 1
7952 * Figure out the number of possible node ids.
7954 void __init setup_nr_node_ids(void)
7956 unsigned int highest;
7958 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7959 nr_node_ids = highest + 1;
7964 * node_map_pfn_alignment - determine the maximum internode alignment
7966 * This function should be called after node map is populated and sorted.
7967 * It calculates the maximum power of two alignment which can distinguish
7970 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7971 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7972 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7973 * shifted, 1GiB is enough and this function will indicate so.
7975 * This is used to test whether pfn -> nid mapping of the chosen memory
7976 * model has fine enough granularity to avoid incorrect mapping for the
7977 * populated node map.
7979 * Return: the determined alignment in pfn's. 0 if there is no alignment
7980 * requirement (single node).
7982 unsigned long __init node_map_pfn_alignment(void)
7984 unsigned long accl_mask = 0, last_end = 0;
7985 unsigned long start, end, mask;
7986 int last_nid = NUMA_NO_NODE;
7989 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7990 if (!start || last_nid < 0 || last_nid == nid) {
7997 * Start with a mask granular enough to pin-point to the
7998 * start pfn and tick off bits one-by-one until it becomes
7999 * too coarse to separate the current node from the last.
8001 mask = ~((1 << __ffs(start)) - 1);
8002 while (mask && last_end <= (start & (mask << 1)))
8005 /* accumulate all internode masks */
8009 /* convert mask to number of pages */
8010 return ~accl_mask + 1;
8014 * early_calculate_totalpages()
8015 * Sum pages in active regions for movable zone.
8016 * Populate N_MEMORY for calculating usable_nodes.
8018 static unsigned long __init early_calculate_totalpages(void)
8020 unsigned long totalpages = 0;
8021 unsigned long start_pfn, end_pfn;
8024 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8025 unsigned long pages = end_pfn - start_pfn;
8027 totalpages += pages;
8029 node_set_state(nid, N_MEMORY);
8035 * Find the PFN the Movable zone begins in each node. Kernel memory
8036 * is spread evenly between nodes as long as the nodes have enough
8037 * memory. When they don't, some nodes will have more kernelcore than
8040 static void __init find_zone_movable_pfns_for_nodes(void)
8043 unsigned long usable_startpfn;
8044 unsigned long kernelcore_node, kernelcore_remaining;
8045 /* save the state before borrow the nodemask */
8046 nodemask_t saved_node_state = node_states[N_MEMORY];
8047 unsigned long totalpages = early_calculate_totalpages();
8048 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
8049 struct memblock_region *r;
8051 /* Need to find movable_zone earlier when movable_node is specified. */
8052 find_usable_zone_for_movable();
8055 * If movable_node is specified, ignore kernelcore and movablecore
8058 if (movable_node_is_enabled()) {
8059 for_each_mem_region(r) {
8060 if (!memblock_is_hotpluggable(r))
8063 nid = memblock_get_region_node(r);
8065 usable_startpfn = PFN_DOWN(r->base);
8066 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
8067 min(usable_startpfn, zone_movable_pfn[nid]) :
8075 * If kernelcore=mirror is specified, ignore movablecore option
8077 if (mirrored_kernelcore) {
8078 bool mem_below_4gb_not_mirrored = false;
8080 for_each_mem_region(r) {
8081 if (memblock_is_mirror(r))
8084 nid = memblock_get_region_node(r);
8086 usable_startpfn = memblock_region_memory_base_pfn(r);
8088 if (usable_startpfn < PHYS_PFN(SZ_4G)) {
8089 mem_below_4gb_not_mirrored = true;
8093 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
8094 min(usable_startpfn, zone_movable_pfn[nid]) :
8098 if (mem_below_4gb_not_mirrored)
8099 pr_warn("This configuration results in unmirrored kernel memory.\n");
8105 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
8106 * amount of necessary memory.
8108 if (required_kernelcore_percent)
8109 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
8111 if (required_movablecore_percent)
8112 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
8116 * If movablecore= was specified, calculate what size of
8117 * kernelcore that corresponds so that memory usable for
8118 * any allocation type is evenly spread. If both kernelcore
8119 * and movablecore are specified, then the value of kernelcore
8120 * will be used for required_kernelcore if it's greater than
8121 * what movablecore would have allowed.
8123 if (required_movablecore) {
8124 unsigned long corepages;
8127 * Round-up so that ZONE_MOVABLE is at least as large as what
8128 * was requested by the user
8130 required_movablecore =
8131 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
8132 required_movablecore = min(totalpages, required_movablecore);
8133 corepages = totalpages - required_movablecore;
8135 required_kernelcore = max(required_kernelcore, corepages);
8139 * If kernelcore was not specified or kernelcore size is larger
8140 * than totalpages, there is no ZONE_MOVABLE.
8142 if (!required_kernelcore || required_kernelcore >= totalpages)
8145 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
8146 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
8149 /* Spread kernelcore memory as evenly as possible throughout nodes */
8150 kernelcore_node = required_kernelcore / usable_nodes;
8151 for_each_node_state(nid, N_MEMORY) {
8152 unsigned long start_pfn, end_pfn;
8155 * Recalculate kernelcore_node if the division per node
8156 * now exceeds what is necessary to satisfy the requested
8157 * amount of memory for the kernel
8159 if (required_kernelcore < kernelcore_node)
8160 kernelcore_node = required_kernelcore / usable_nodes;
8163 * As the map is walked, we track how much memory is usable
8164 * by the kernel using kernelcore_remaining. When it is
8165 * 0, the rest of the node is usable by ZONE_MOVABLE
8167 kernelcore_remaining = kernelcore_node;
8169 /* Go through each range of PFNs within this node */
8170 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
8171 unsigned long size_pages;
8173 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
8174 if (start_pfn >= end_pfn)
8177 /* Account for what is only usable for kernelcore */
8178 if (start_pfn < usable_startpfn) {
8179 unsigned long kernel_pages;
8180 kernel_pages = min(end_pfn, usable_startpfn)
8183 kernelcore_remaining -= min(kernel_pages,
8184 kernelcore_remaining);
8185 required_kernelcore -= min(kernel_pages,
8186 required_kernelcore);
8188 /* Continue if range is now fully accounted */
8189 if (end_pfn <= usable_startpfn) {
8192 * Push zone_movable_pfn to the end so
8193 * that if we have to rebalance
8194 * kernelcore across nodes, we will
8195 * not double account here
8197 zone_movable_pfn[nid] = end_pfn;
8200 start_pfn = usable_startpfn;
8204 * The usable PFN range for ZONE_MOVABLE is from
8205 * start_pfn->end_pfn. Calculate size_pages as the
8206 * number of pages used as kernelcore
8208 size_pages = end_pfn - start_pfn;
8209 if (size_pages > kernelcore_remaining)
8210 size_pages = kernelcore_remaining;
8211 zone_movable_pfn[nid] = start_pfn + size_pages;
8214 * Some kernelcore has been met, update counts and
8215 * break if the kernelcore for this node has been
8218 required_kernelcore -= min(required_kernelcore,
8220 kernelcore_remaining -= size_pages;
8221 if (!kernelcore_remaining)
8227 * If there is still required_kernelcore, we do another pass with one
8228 * less node in the count. This will push zone_movable_pfn[nid] further
8229 * along on the nodes that still have memory until kernelcore is
8233 if (usable_nodes && required_kernelcore > usable_nodes)
8237 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
8238 for (nid = 0; nid < MAX_NUMNODES; nid++) {
8239 unsigned long start_pfn, end_pfn;
8241 zone_movable_pfn[nid] =
8242 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
8244 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
8245 if (zone_movable_pfn[nid] >= end_pfn)
8246 zone_movable_pfn[nid] = 0;
8250 /* restore the node_state */
8251 node_states[N_MEMORY] = saved_node_state;
8254 /* Any regular or high memory on that node ? */
8255 static void check_for_memory(pg_data_t *pgdat, int nid)
8257 enum zone_type zone_type;
8259 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8260 struct zone *zone = &pgdat->node_zones[zone_type];
8261 if (populated_zone(zone)) {
8262 if (IS_ENABLED(CONFIG_HIGHMEM))
8263 node_set_state(nid, N_HIGH_MEMORY);
8264 if (zone_type <= ZONE_NORMAL)
8265 node_set_state(nid, N_NORMAL_MEMORY);
8272 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8273 * such cases we allow max_zone_pfn sorted in the descending order
8275 bool __weak arch_has_descending_max_zone_pfns(void)
8281 * free_area_init - Initialise all pg_data_t and zone data
8282 * @max_zone_pfn: an array of max PFNs for each zone
8284 * This will call free_area_init_node() for each active node in the system.
8285 * Using the page ranges provided by memblock_set_node(), the size of each
8286 * zone in each node and their holes is calculated. If the maximum PFN
8287 * between two adjacent zones match, it is assumed that the zone is empty.
8288 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8289 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8290 * starts where the previous one ended. For example, ZONE_DMA32 starts
8291 * at arch_max_dma_pfn.
8293 void __init free_area_init(unsigned long *max_zone_pfn)
8295 unsigned long start_pfn, end_pfn;
8299 /* Record where the zone boundaries are */
8300 memset(arch_zone_lowest_possible_pfn, 0,
8301 sizeof(arch_zone_lowest_possible_pfn));
8302 memset(arch_zone_highest_possible_pfn, 0,
8303 sizeof(arch_zone_highest_possible_pfn));
8305 start_pfn = PHYS_PFN(memblock_start_of_DRAM());
8306 descending = arch_has_descending_max_zone_pfns();
8308 for (i = 0; i < MAX_NR_ZONES; i++) {
8310 zone = MAX_NR_ZONES - i - 1;
8314 if (zone == ZONE_MOVABLE)
8317 end_pfn = max(max_zone_pfn[zone], start_pfn);
8318 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8319 arch_zone_highest_possible_pfn[zone] = end_pfn;
8321 start_pfn = end_pfn;
8324 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8325 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8326 find_zone_movable_pfns_for_nodes();
8328 /* Print out the zone ranges */
8329 pr_info("Zone ranges:\n");
8330 for (i = 0; i < MAX_NR_ZONES; i++) {
8331 if (i == ZONE_MOVABLE)
8333 pr_info(" %-8s ", zone_names[i]);
8334 if (arch_zone_lowest_possible_pfn[i] ==
8335 arch_zone_highest_possible_pfn[i])
8338 pr_cont("[mem %#018Lx-%#018Lx]\n",
8339 (u64)arch_zone_lowest_possible_pfn[i]
8341 ((u64)arch_zone_highest_possible_pfn[i]
8342 << PAGE_SHIFT) - 1);
8345 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8346 pr_info("Movable zone start for each node\n");
8347 for (i = 0; i < MAX_NUMNODES; i++) {
8348 if (zone_movable_pfn[i])
8349 pr_info(" Node %d: %#018Lx\n", i,
8350 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8354 * Print out the early node map, and initialize the
8355 * subsection-map relative to active online memory ranges to
8356 * enable future "sub-section" extensions of the memory map.
8358 pr_info("Early memory node ranges\n");
8359 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8360 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8361 (u64)start_pfn << PAGE_SHIFT,
8362 ((u64)end_pfn << PAGE_SHIFT) - 1);
8363 subsection_map_init(start_pfn, end_pfn - start_pfn);
8366 /* Initialise every node */
8367 mminit_verify_pageflags_layout();
8368 setup_nr_node_ids();
8369 for_each_node(nid) {
8372 if (!node_online(nid)) {
8373 pr_info("Initializing node %d as memoryless\n", nid);
8375 /* Allocator not initialized yet */
8376 pgdat = arch_alloc_nodedata(nid);
8378 pr_err("Cannot allocate %zuB for node %d.\n",
8379 sizeof(*pgdat), nid);
8382 arch_refresh_nodedata(nid, pgdat);
8383 free_area_init_memoryless_node(nid);
8386 * We do not want to confuse userspace by sysfs
8387 * files/directories for node without any memory
8388 * attached to it, so this node is not marked as
8389 * N_MEMORY and not marked online so that no sysfs
8390 * hierarchy will be created via register_one_node for
8391 * it. The pgdat will get fully initialized by
8392 * hotadd_init_pgdat() when memory is hotplugged into
8398 pgdat = NODE_DATA(nid);
8399 free_area_init_node(nid);
8401 /* Any memory on that node */
8402 if (pgdat->node_present_pages)
8403 node_set_state(nid, N_MEMORY);
8404 check_for_memory(pgdat, nid);
8410 static int __init cmdline_parse_core(char *p, unsigned long *core,
8411 unsigned long *percent)
8413 unsigned long long coremem;
8419 /* Value may be a percentage of total memory, otherwise bytes */
8420 coremem = simple_strtoull(p, &endptr, 0);
8421 if (*endptr == '%') {
8422 /* Paranoid check for percent values greater than 100 */
8423 WARN_ON(coremem > 100);
8427 coremem = memparse(p, &p);
8428 /* Paranoid check that UL is enough for the coremem value */
8429 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8431 *core = coremem >> PAGE_SHIFT;
8438 * kernelcore=size sets the amount of memory for use for allocations that
8439 * cannot be reclaimed or migrated.
8441 static int __init cmdline_parse_kernelcore(char *p)
8443 /* parse kernelcore=mirror */
8444 if (parse_option_str(p, "mirror")) {
8445 mirrored_kernelcore = true;
8449 return cmdline_parse_core(p, &required_kernelcore,
8450 &required_kernelcore_percent);
8454 * movablecore=size sets the amount of memory for use for allocations that
8455 * can be reclaimed or migrated.
8457 static int __init cmdline_parse_movablecore(char *p)
8459 return cmdline_parse_core(p, &required_movablecore,
8460 &required_movablecore_percent);
8463 early_param("kernelcore", cmdline_parse_kernelcore);
8464 early_param("movablecore", cmdline_parse_movablecore);
8466 void adjust_managed_page_count(struct page *page, long count)
8468 atomic_long_add(count, &page_zone(page)->managed_pages);
8469 totalram_pages_add(count);
8470 #ifdef CONFIG_HIGHMEM
8471 if (PageHighMem(page))
8472 totalhigh_pages_add(count);
8475 EXPORT_SYMBOL(adjust_managed_page_count);
8477 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8480 unsigned long pages = 0;
8482 start = (void *)PAGE_ALIGN((unsigned long)start);
8483 end = (void *)((unsigned long)end & PAGE_MASK);
8484 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8485 struct page *page = virt_to_page(pos);
8486 void *direct_map_addr;
8489 * 'direct_map_addr' might be different from 'pos'
8490 * because some architectures' virt_to_page()
8491 * work with aliases. Getting the direct map
8492 * address ensures that we get a _writeable_
8493 * alias for the memset().
8495 direct_map_addr = page_address(page);
8497 * Perform a kasan-unchecked memset() since this memory
8498 * has not been initialized.
8500 direct_map_addr = kasan_reset_tag(direct_map_addr);
8501 if ((unsigned int)poison <= 0xFF)
8502 memset(direct_map_addr, poison, PAGE_SIZE);
8504 free_reserved_page(page);
8508 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8513 void __init mem_init_print_info(void)
8515 unsigned long physpages, codesize, datasize, rosize, bss_size;
8516 unsigned long init_code_size, init_data_size;
8518 physpages = get_num_physpages();
8519 codesize = _etext - _stext;
8520 datasize = _edata - _sdata;
8521 rosize = __end_rodata - __start_rodata;
8522 bss_size = __bss_stop - __bss_start;
8523 init_data_size = __init_end - __init_begin;
8524 init_code_size = _einittext - _sinittext;
8527 * Detect special cases and adjust section sizes accordingly:
8528 * 1) .init.* may be embedded into .data sections
8529 * 2) .init.text.* may be out of [__init_begin, __init_end],
8530 * please refer to arch/tile/kernel/vmlinux.lds.S.
8531 * 3) .rodata.* may be embedded into .text or .data sections.
8533 #define adj_init_size(start, end, size, pos, adj) \
8535 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8539 adj_init_size(__init_begin, __init_end, init_data_size,
8540 _sinittext, init_code_size);
8541 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8542 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8543 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8544 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8546 #undef adj_init_size
8548 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8549 #ifdef CONFIG_HIGHMEM
8553 K(nr_free_pages()), K(physpages),
8554 codesize / SZ_1K, datasize / SZ_1K, rosize / SZ_1K,
8555 (init_data_size + init_code_size) / SZ_1K, bss_size / SZ_1K,
8556 K(physpages - totalram_pages() - totalcma_pages),
8558 #ifdef CONFIG_HIGHMEM
8559 , K(totalhigh_pages())
8565 * set_dma_reserve - set the specified number of pages reserved in the first zone
8566 * @new_dma_reserve: The number of pages to mark reserved
8568 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8569 * In the DMA zone, a significant percentage may be consumed by kernel image
8570 * and other unfreeable allocations which can skew the watermarks badly. This
8571 * function may optionally be used to account for unfreeable pages in the
8572 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8573 * smaller per-cpu batchsize.
8575 void __init set_dma_reserve(unsigned long new_dma_reserve)
8577 dma_reserve = new_dma_reserve;
8580 static int page_alloc_cpu_dead(unsigned int cpu)
8584 lru_add_drain_cpu(cpu);
8585 mlock_page_drain_remote(cpu);
8589 * Spill the event counters of the dead processor
8590 * into the current processors event counters.
8591 * This artificially elevates the count of the current
8594 vm_events_fold_cpu(cpu);
8597 * Zero the differential counters of the dead processor
8598 * so that the vm statistics are consistent.
8600 * This is only okay since the processor is dead and cannot
8601 * race with what we are doing.
8603 cpu_vm_stats_fold(cpu);
8605 for_each_populated_zone(zone)
8606 zone_pcp_update(zone, 0);
8611 static int page_alloc_cpu_online(unsigned int cpu)
8615 for_each_populated_zone(zone)
8616 zone_pcp_update(zone, 1);
8621 int hashdist = HASHDIST_DEFAULT;
8623 static int __init set_hashdist(char *str)
8627 hashdist = simple_strtoul(str, &str, 0);
8630 __setup("hashdist=", set_hashdist);
8633 void __init page_alloc_init(void)
8638 if (num_node_state(N_MEMORY) == 1)
8642 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8643 "mm/page_alloc:pcp",
8644 page_alloc_cpu_online,
8645 page_alloc_cpu_dead);
8650 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8651 * or min_free_kbytes changes.
8653 static void calculate_totalreserve_pages(void)
8655 struct pglist_data *pgdat;
8656 unsigned long reserve_pages = 0;
8657 enum zone_type i, j;
8659 for_each_online_pgdat(pgdat) {
8661 pgdat->totalreserve_pages = 0;
8663 for (i = 0; i < MAX_NR_ZONES; i++) {
8664 struct zone *zone = pgdat->node_zones + i;
8666 unsigned long managed_pages = zone_managed_pages(zone);
8668 /* Find valid and maximum lowmem_reserve in the zone */
8669 for (j = i; j < MAX_NR_ZONES; j++) {
8670 if (zone->lowmem_reserve[j] > max)
8671 max = zone->lowmem_reserve[j];
8674 /* we treat the high watermark as reserved pages. */
8675 max += high_wmark_pages(zone);
8677 if (max > managed_pages)
8678 max = managed_pages;
8680 pgdat->totalreserve_pages += max;
8682 reserve_pages += max;
8685 totalreserve_pages = reserve_pages;
8689 * setup_per_zone_lowmem_reserve - called whenever
8690 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8691 * has a correct pages reserved value, so an adequate number of
8692 * pages are left in the zone after a successful __alloc_pages().
8694 static void setup_per_zone_lowmem_reserve(void)
8696 struct pglist_data *pgdat;
8697 enum zone_type i, j;
8699 for_each_online_pgdat(pgdat) {
8700 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8701 struct zone *zone = &pgdat->node_zones[i];
8702 int ratio = sysctl_lowmem_reserve_ratio[i];
8703 bool clear = !ratio || !zone_managed_pages(zone);
8704 unsigned long managed_pages = 0;
8706 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8707 struct zone *upper_zone = &pgdat->node_zones[j];
8709 managed_pages += zone_managed_pages(upper_zone);
8712 zone->lowmem_reserve[j] = 0;
8714 zone->lowmem_reserve[j] = managed_pages / ratio;
8719 /* update totalreserve_pages */
8720 calculate_totalreserve_pages();
8723 static void __setup_per_zone_wmarks(void)
8725 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8726 unsigned long lowmem_pages = 0;
8728 unsigned long flags;
8730 /* Calculate total number of !ZONE_HIGHMEM pages */
8731 for_each_zone(zone) {
8732 if (!is_highmem(zone))
8733 lowmem_pages += zone_managed_pages(zone);
8736 for_each_zone(zone) {
8739 spin_lock_irqsave(&zone->lock, flags);
8740 tmp = (u64)pages_min * zone_managed_pages(zone);
8741 do_div(tmp, lowmem_pages);
8742 if (is_highmem(zone)) {
8744 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8745 * need highmem pages, so cap pages_min to a small
8748 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8749 * deltas control async page reclaim, and so should
8750 * not be capped for highmem.
8752 unsigned long min_pages;
8754 min_pages = zone_managed_pages(zone) / 1024;
8755 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8756 zone->_watermark[WMARK_MIN] = min_pages;
8759 * If it's a lowmem zone, reserve a number of pages
8760 * proportionate to the zone's size.
8762 zone->_watermark[WMARK_MIN] = tmp;
8766 * Set the kswapd watermarks distance according to the
8767 * scale factor in proportion to available memory, but
8768 * ensure a minimum size on small systems.
8770 tmp = max_t(u64, tmp >> 2,
8771 mult_frac(zone_managed_pages(zone),
8772 watermark_scale_factor, 10000));
8774 zone->watermark_boost = 0;
8775 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8776 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8777 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8779 spin_unlock_irqrestore(&zone->lock, flags);
8782 /* update totalreserve_pages */
8783 calculate_totalreserve_pages();
8787 * setup_per_zone_wmarks - called when min_free_kbytes changes
8788 * or when memory is hot-{added|removed}
8790 * Ensures that the watermark[min,low,high] values for each zone are set
8791 * correctly with respect to min_free_kbytes.
8793 void setup_per_zone_wmarks(void)
8796 static DEFINE_SPINLOCK(lock);
8799 __setup_per_zone_wmarks();
8803 * The watermark size have changed so update the pcpu batch
8804 * and high limits or the limits may be inappropriate.
8807 zone_pcp_update(zone, 0);
8811 * Initialise min_free_kbytes.
8813 * For small machines we want it small (128k min). For large machines
8814 * we want it large (256MB max). But it is not linear, because network
8815 * bandwidth does not increase linearly with machine size. We use
8817 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8818 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8834 void calculate_min_free_kbytes(void)
8836 unsigned long lowmem_kbytes;
8837 int new_min_free_kbytes;
8839 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8840 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8842 if (new_min_free_kbytes > user_min_free_kbytes)
8843 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8845 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8846 new_min_free_kbytes, user_min_free_kbytes);
8850 int __meminit init_per_zone_wmark_min(void)
8852 calculate_min_free_kbytes();
8853 setup_per_zone_wmarks();
8854 refresh_zone_stat_thresholds();
8855 setup_per_zone_lowmem_reserve();
8858 setup_min_unmapped_ratio();
8859 setup_min_slab_ratio();
8862 khugepaged_min_free_kbytes_update();
8866 postcore_initcall(init_per_zone_wmark_min)
8869 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8870 * that we can call two helper functions whenever min_free_kbytes
8873 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8874 void *buffer, size_t *length, loff_t *ppos)
8878 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8883 user_min_free_kbytes = min_free_kbytes;
8884 setup_per_zone_wmarks();
8889 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8890 void *buffer, size_t *length, loff_t *ppos)
8894 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8899 setup_per_zone_wmarks();
8905 static void setup_min_unmapped_ratio(void)
8910 for_each_online_pgdat(pgdat)
8911 pgdat->min_unmapped_pages = 0;
8914 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8915 sysctl_min_unmapped_ratio) / 100;
8919 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8920 void *buffer, size_t *length, loff_t *ppos)
8924 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8928 setup_min_unmapped_ratio();
8933 static void setup_min_slab_ratio(void)
8938 for_each_online_pgdat(pgdat)
8939 pgdat->min_slab_pages = 0;
8942 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8943 sysctl_min_slab_ratio) / 100;
8946 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8947 void *buffer, size_t *length, loff_t *ppos)
8951 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8955 setup_min_slab_ratio();
8962 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8963 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8964 * whenever sysctl_lowmem_reserve_ratio changes.
8966 * The reserve ratio obviously has absolutely no relation with the
8967 * minimum watermarks. The lowmem reserve ratio can only make sense
8968 * if in function of the boot time zone sizes.
8970 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8971 void *buffer, size_t *length, loff_t *ppos)
8975 proc_dointvec_minmax(table, write, buffer, length, ppos);
8977 for (i = 0; i < MAX_NR_ZONES; i++) {
8978 if (sysctl_lowmem_reserve_ratio[i] < 1)
8979 sysctl_lowmem_reserve_ratio[i] = 0;
8982 setup_per_zone_lowmem_reserve();
8987 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8988 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8989 * pagelist can have before it gets flushed back to buddy allocator.
8991 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8992 int write, void *buffer, size_t *length, loff_t *ppos)
8995 int old_percpu_pagelist_high_fraction;
8998 mutex_lock(&pcp_batch_high_lock);
8999 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
9001 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
9002 if (!write || ret < 0)
9005 /* Sanity checking to avoid pcp imbalance */
9006 if (percpu_pagelist_high_fraction &&
9007 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
9008 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
9014 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
9017 for_each_populated_zone(zone)
9018 zone_set_pageset_high_and_batch(zone, 0);
9020 mutex_unlock(&pcp_batch_high_lock);
9024 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
9026 * Returns the number of pages that arch has reserved but
9027 * is not known to alloc_large_system_hash().
9029 static unsigned long __init arch_reserved_kernel_pages(void)
9036 * Adaptive scale is meant to reduce sizes of hash tables on large memory
9037 * machines. As memory size is increased the scale is also increased but at
9038 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
9039 * quadruples the scale is increased by one, which means the size of hash table
9040 * only doubles, instead of quadrupling as well.
9041 * Because 32-bit systems cannot have large physical memory, where this scaling
9042 * makes sense, it is disabled on such platforms.
9044 #if __BITS_PER_LONG > 32
9045 #define ADAPT_SCALE_BASE (64ul << 30)
9046 #define ADAPT_SCALE_SHIFT 2
9047 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
9051 * allocate a large system hash table from bootmem
9052 * - it is assumed that the hash table must contain an exact power-of-2
9053 * quantity of entries
9054 * - limit is the number of hash buckets, not the total allocation size
9056 void *__init alloc_large_system_hash(const char *tablename,
9057 unsigned long bucketsize,
9058 unsigned long numentries,
9061 unsigned int *_hash_shift,
9062 unsigned int *_hash_mask,
9063 unsigned long low_limit,
9064 unsigned long high_limit)
9066 unsigned long long max = high_limit;
9067 unsigned long log2qty, size;
9073 /* allow the kernel cmdline to have a say */
9075 /* round applicable memory size up to nearest megabyte */
9076 numentries = nr_kernel_pages;
9077 numentries -= arch_reserved_kernel_pages();
9079 /* It isn't necessary when PAGE_SIZE >= 1MB */
9080 if (PAGE_SIZE < SZ_1M)
9081 numentries = round_up(numentries, SZ_1M / PAGE_SIZE);
9083 #if __BITS_PER_LONG > 32
9085 unsigned long adapt;
9087 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
9088 adapt <<= ADAPT_SCALE_SHIFT)
9093 /* limit to 1 bucket per 2^scale bytes of low memory */
9094 if (scale > PAGE_SHIFT)
9095 numentries >>= (scale - PAGE_SHIFT);
9097 numentries <<= (PAGE_SHIFT - scale);
9099 /* Make sure we've got at least a 0-order allocation.. */
9100 if (unlikely(flags & HASH_SMALL)) {
9101 /* Makes no sense without HASH_EARLY */
9102 WARN_ON(!(flags & HASH_EARLY));
9103 if (!(numentries >> *_hash_shift)) {
9104 numentries = 1UL << *_hash_shift;
9105 BUG_ON(!numentries);
9107 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
9108 numentries = PAGE_SIZE / bucketsize;
9110 numentries = roundup_pow_of_two(numentries);
9112 /* limit allocation size to 1/16 total memory by default */
9114 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
9115 do_div(max, bucketsize);
9117 max = min(max, 0x80000000ULL);
9119 if (numentries < low_limit)
9120 numentries = low_limit;
9121 if (numentries > max)
9124 log2qty = ilog2(numentries);
9126 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
9129 size = bucketsize << log2qty;
9130 if (flags & HASH_EARLY) {
9131 if (flags & HASH_ZERO)
9132 table = memblock_alloc(size, SMP_CACHE_BYTES);
9134 table = memblock_alloc_raw(size,
9136 } else if (get_order(size) >= MAX_ORDER || hashdist) {
9137 table = vmalloc_huge(size, gfp_flags);
9140 huge = is_vm_area_hugepages(table);
9143 * If bucketsize is not a power-of-two, we may free
9144 * some pages at the end of hash table which
9145 * alloc_pages_exact() automatically does
9147 table = alloc_pages_exact(size, gfp_flags);
9148 kmemleak_alloc(table, size, 1, gfp_flags);
9150 } while (!table && size > PAGE_SIZE && --log2qty);
9153 panic("Failed to allocate %s hash table\n", tablename);
9155 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
9156 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
9157 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
9160 *_hash_shift = log2qty;
9162 *_hash_mask = (1 << log2qty) - 1;
9167 #ifdef CONFIG_CONTIG_ALLOC
9168 #if defined(CONFIG_DYNAMIC_DEBUG) || \
9169 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
9170 /* Usage: See admin-guide/dynamic-debug-howto.rst */
9171 static void alloc_contig_dump_pages(struct list_head *page_list)
9173 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9175 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9179 list_for_each_entry(page, page_list, lru)
9180 dump_page(page, "migration failure");
9184 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9189 /* [start, end) must belong to a single zone. */
9190 int __alloc_contig_migrate_range(struct compact_control *cc,
9191 unsigned long start, unsigned long end)
9193 /* This function is based on compact_zone() from compaction.c. */
9194 unsigned int nr_reclaimed;
9195 unsigned long pfn = start;
9196 unsigned int tries = 0;
9198 struct migration_target_control mtc = {
9199 .nid = zone_to_nid(cc->zone),
9200 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9203 lru_cache_disable();
9205 while (pfn < end || !list_empty(&cc->migratepages)) {
9206 if (fatal_signal_pending(current)) {
9211 if (list_empty(&cc->migratepages)) {
9212 cc->nr_migratepages = 0;
9213 ret = isolate_migratepages_range(cc, pfn, end);
9214 if (ret && ret != -EAGAIN)
9216 pfn = cc->migrate_pfn;
9218 } else if (++tries == 5) {
9223 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9225 cc->nr_migratepages -= nr_reclaimed;
9227 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9228 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9231 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9232 * to retry again over this error, so do the same here.
9240 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
9241 alloc_contig_dump_pages(&cc->migratepages);
9242 putback_movable_pages(&cc->migratepages);
9249 * alloc_contig_range() -- tries to allocate given range of pages
9250 * @start: start PFN to allocate
9251 * @end: one-past-the-last PFN to allocate
9252 * @migratetype: migratetype of the underlying pageblocks (either
9253 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9254 * in range must have the same migratetype and it must
9255 * be either of the two.
9256 * @gfp_mask: GFP mask to use during compaction
9258 * The PFN range does not have to be pageblock aligned. The PFN range must
9259 * belong to a single zone.
9261 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9262 * pageblocks in the range. Once isolated, the pageblocks should not
9263 * be modified by others.
9265 * Return: zero on success or negative error code. On success all
9266 * pages which PFN is in [start, end) are allocated for the caller and
9267 * need to be freed with free_contig_range().
9269 int alloc_contig_range(unsigned long start, unsigned long end,
9270 unsigned migratetype, gfp_t gfp_mask)
9272 unsigned long outer_start, outer_end;
9276 struct compact_control cc = {
9277 .nr_migratepages = 0,
9279 .zone = page_zone(pfn_to_page(start)),
9280 .mode = MIGRATE_SYNC,
9281 .ignore_skip_hint = true,
9282 .no_set_skip_hint = true,
9283 .gfp_mask = current_gfp_context(gfp_mask),
9284 .alloc_contig = true,
9286 INIT_LIST_HEAD(&cc.migratepages);
9289 * What we do here is we mark all pageblocks in range as
9290 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9291 * have different sizes, and due to the way page allocator
9292 * work, start_isolate_page_range() has special handlings for this.
9294 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9295 * migrate the pages from an unaligned range (ie. pages that
9296 * we are interested in). This will put all the pages in
9297 * range back to page allocator as MIGRATE_ISOLATE.
9299 * When this is done, we take the pages in range from page
9300 * allocator removing them from the buddy system. This way
9301 * page allocator will never consider using them.
9303 * This lets us mark the pageblocks back as
9304 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9305 * aligned range but not in the unaligned, original range are
9306 * put back to page allocator so that buddy can use them.
9309 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
9313 drain_all_pages(cc.zone);
9316 * In case of -EBUSY, we'd like to know which page causes problem.
9317 * So, just fall through. test_pages_isolated() has a tracepoint
9318 * which will report the busy page.
9320 * It is possible that busy pages could become available before
9321 * the call to test_pages_isolated, and the range will actually be
9322 * allocated. So, if we fall through be sure to clear ret so that
9323 * -EBUSY is not accidentally used or returned to caller.
9325 ret = __alloc_contig_migrate_range(&cc, start, end);
9326 if (ret && ret != -EBUSY)
9331 * Pages from [start, end) are within a pageblock_nr_pages
9332 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9333 * more, all pages in [start, end) are free in page allocator.
9334 * What we are going to do is to allocate all pages from
9335 * [start, end) (that is remove them from page allocator).
9337 * The only problem is that pages at the beginning and at the
9338 * end of interesting range may be not aligned with pages that
9339 * page allocator holds, ie. they can be part of higher order
9340 * pages. Because of this, we reserve the bigger range and
9341 * once this is done free the pages we are not interested in.
9343 * We don't have to hold zone->lock here because the pages are
9344 * isolated thus they won't get removed from buddy.
9348 outer_start = start;
9349 while (!PageBuddy(pfn_to_page(outer_start))) {
9350 if (++order >= MAX_ORDER) {
9351 outer_start = start;
9354 outer_start &= ~0UL << order;
9357 if (outer_start != start) {
9358 order = buddy_order(pfn_to_page(outer_start));
9361 * outer_start page could be small order buddy page and
9362 * it doesn't include start page. Adjust outer_start
9363 * in this case to report failed page properly
9364 * on tracepoint in test_pages_isolated()
9366 if (outer_start + (1UL << order) <= start)
9367 outer_start = start;
9370 /* Make sure the range is really isolated. */
9371 if (test_pages_isolated(outer_start, end, 0)) {
9376 /* Grab isolated pages from freelists. */
9377 outer_end = isolate_freepages_range(&cc, outer_start, end);
9383 /* Free head and tail (if any) */
9384 if (start != outer_start)
9385 free_contig_range(outer_start, start - outer_start);
9386 if (end != outer_end)
9387 free_contig_range(end, outer_end - end);
9390 undo_isolate_page_range(start, end, migratetype);
9393 EXPORT_SYMBOL(alloc_contig_range);
9395 static int __alloc_contig_pages(unsigned long start_pfn,
9396 unsigned long nr_pages, gfp_t gfp_mask)
9398 unsigned long end_pfn = start_pfn + nr_pages;
9400 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9404 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9405 unsigned long nr_pages)
9407 unsigned long i, end_pfn = start_pfn + nr_pages;
9410 for (i = start_pfn; i < end_pfn; i++) {
9411 page = pfn_to_online_page(i);
9415 if (page_zone(page) != z)
9418 if (PageReserved(page))
9424 static bool zone_spans_last_pfn(const struct zone *zone,
9425 unsigned long start_pfn, unsigned long nr_pages)
9427 unsigned long last_pfn = start_pfn + nr_pages - 1;
9429 return zone_spans_pfn(zone, last_pfn);
9433 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9434 * @nr_pages: Number of contiguous pages to allocate
9435 * @gfp_mask: GFP mask to limit search and used during compaction
9437 * @nodemask: Mask for other possible nodes
9439 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9440 * on an applicable zonelist to find a contiguous pfn range which can then be
9441 * tried for allocation with alloc_contig_range(). This routine is intended
9442 * for allocation requests which can not be fulfilled with the buddy allocator.
9444 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9445 * power of two, then allocated range is also guaranteed to be aligned to same
9446 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9448 * Allocated pages can be freed with free_contig_range() or by manually calling
9449 * __free_page() on each allocated page.
9451 * Return: pointer to contiguous pages on success, or NULL if not successful.
9453 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9454 int nid, nodemask_t *nodemask)
9456 unsigned long ret, pfn, flags;
9457 struct zonelist *zonelist;
9461 zonelist = node_zonelist(nid, gfp_mask);
9462 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9463 gfp_zone(gfp_mask), nodemask) {
9464 spin_lock_irqsave(&zone->lock, flags);
9466 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9467 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9468 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9470 * We release the zone lock here because
9471 * alloc_contig_range() will also lock the zone
9472 * at some point. If there's an allocation
9473 * spinning on this lock, it may win the race
9474 * and cause alloc_contig_range() to fail...
9476 spin_unlock_irqrestore(&zone->lock, flags);
9477 ret = __alloc_contig_pages(pfn, nr_pages,
9480 return pfn_to_page(pfn);
9481 spin_lock_irqsave(&zone->lock, flags);
9485 spin_unlock_irqrestore(&zone->lock, flags);
9489 #endif /* CONFIG_CONTIG_ALLOC */
9491 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9493 unsigned long count = 0;
9495 for (; nr_pages--; pfn++) {
9496 struct page *page = pfn_to_page(pfn);
9498 count += page_count(page) != 1;
9501 WARN(count != 0, "%lu pages are still in use!\n", count);
9503 EXPORT_SYMBOL(free_contig_range);
9506 * Effectively disable pcplists for the zone by setting the high limit to 0
9507 * and draining all cpus. A concurrent page freeing on another CPU that's about
9508 * to put the page on pcplist will either finish before the drain and the page
9509 * will be drained, or observe the new high limit and skip the pcplist.
9511 * Must be paired with a call to zone_pcp_enable().
9513 void zone_pcp_disable(struct zone *zone)
9515 mutex_lock(&pcp_batch_high_lock);
9516 __zone_set_pageset_high_and_batch(zone, 0, 1);
9517 __drain_all_pages(zone, true);
9520 void zone_pcp_enable(struct zone *zone)
9522 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9523 mutex_unlock(&pcp_batch_high_lock);
9526 void zone_pcp_reset(struct zone *zone)
9529 struct per_cpu_zonestat *pzstats;
9531 if (zone->per_cpu_pageset != &boot_pageset) {
9532 for_each_online_cpu(cpu) {
9533 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9534 drain_zonestat(zone, pzstats);
9536 free_percpu(zone->per_cpu_pageset);
9537 zone->per_cpu_pageset = &boot_pageset;
9538 if (zone->per_cpu_zonestats != &boot_zonestats) {
9539 free_percpu(zone->per_cpu_zonestats);
9540 zone->per_cpu_zonestats = &boot_zonestats;
9545 #ifdef CONFIG_MEMORY_HOTREMOVE
9547 * All pages in the range must be in a single zone, must not contain holes,
9548 * must span full sections, and must be isolated before calling this function.
9550 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9552 unsigned long pfn = start_pfn;
9556 unsigned long flags;
9558 offline_mem_sections(pfn, end_pfn);
9559 zone = page_zone(pfn_to_page(pfn));
9560 spin_lock_irqsave(&zone->lock, flags);
9561 while (pfn < end_pfn) {
9562 page = pfn_to_page(pfn);
9564 * The HWPoisoned page may be not in buddy system, and
9565 * page_count() is not 0.
9567 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9572 * At this point all remaining PageOffline() pages have a
9573 * reference count of 0 and can simply be skipped.
9575 if (PageOffline(page)) {
9576 BUG_ON(page_count(page));
9577 BUG_ON(PageBuddy(page));
9582 BUG_ON(page_count(page));
9583 BUG_ON(!PageBuddy(page));
9584 order = buddy_order(page);
9585 del_page_from_free_list(page, zone, order);
9586 pfn += (1 << order);
9588 spin_unlock_irqrestore(&zone->lock, flags);
9593 * This function returns a stable result only if called under zone lock.
9595 bool is_free_buddy_page(struct page *page)
9597 unsigned long pfn = page_to_pfn(page);
9600 for (order = 0; order < MAX_ORDER; order++) {
9601 struct page *page_head = page - (pfn & ((1 << order) - 1));
9603 if (PageBuddy(page_head) &&
9604 buddy_order_unsafe(page_head) >= order)
9608 return order < MAX_ORDER;
9610 EXPORT_SYMBOL(is_free_buddy_page);
9612 #ifdef CONFIG_MEMORY_FAILURE
9614 * Break down a higher-order page in sub-pages, and keep our target out of
9617 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9618 struct page *target, int low, int high,
9621 unsigned long size = 1 << high;
9622 struct page *current_buddy, *next_page;
9624 while (high > low) {
9628 if (target >= &page[size]) {
9629 next_page = page + size;
9630 current_buddy = page;
9633 current_buddy = page + size;
9636 if (set_page_guard(zone, current_buddy, high, migratetype))
9639 if (current_buddy != target) {
9640 add_to_free_list(current_buddy, zone, high, migratetype);
9641 set_buddy_order(current_buddy, high);
9648 * Take a page that will be marked as poisoned off the buddy allocator.
9650 bool take_page_off_buddy(struct page *page)
9652 struct zone *zone = page_zone(page);
9653 unsigned long pfn = page_to_pfn(page);
9654 unsigned long flags;
9658 spin_lock_irqsave(&zone->lock, flags);
9659 for (order = 0; order < MAX_ORDER; order++) {
9660 struct page *page_head = page - (pfn & ((1 << order) - 1));
9661 int page_order = buddy_order(page_head);
9663 if (PageBuddy(page_head) && page_order >= order) {
9664 unsigned long pfn_head = page_to_pfn(page_head);
9665 int migratetype = get_pfnblock_migratetype(page_head,
9668 del_page_from_free_list(page_head, zone, page_order);
9669 break_down_buddy_pages(zone, page_head, page, 0,
9670 page_order, migratetype);
9671 SetPageHWPoisonTakenOff(page);
9672 if (!is_migrate_isolate(migratetype))
9673 __mod_zone_freepage_state(zone, -1, migratetype);
9677 if (page_count(page_head) > 0)
9680 spin_unlock_irqrestore(&zone->lock, flags);
9685 * Cancel takeoff done by take_page_off_buddy().
9687 bool put_page_back_buddy(struct page *page)
9689 struct zone *zone = page_zone(page);
9690 unsigned long pfn = page_to_pfn(page);
9691 unsigned long flags;
9692 int migratetype = get_pfnblock_migratetype(page, pfn);
9695 spin_lock_irqsave(&zone->lock, flags);
9696 if (put_page_testzero(page)) {
9697 ClearPageHWPoisonTakenOff(page);
9698 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9699 if (TestClearPageHWPoison(page)) {
9703 spin_unlock_irqrestore(&zone->lock, flags);
9709 #ifdef CONFIG_ZONE_DMA
9710 bool has_managed_dma(void)
9712 struct pglist_data *pgdat;
9714 for_each_online_pgdat(pgdat) {
9715 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9717 if (managed_zone(zone))
9722 #endif /* CONFIG_ZONE_DMA */