1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/swapops.h>
23 #include <linux/interrupt.h>
24 #include <linux/pagemap.h>
25 #include <linux/jiffies.h>
26 #include <linux/memblock.h>
27 #include <linux/compiler.h>
28 #include <linux/kernel.h>
29 #include <linux/kasan.h>
30 #include <linux/module.h>
31 #include <linux/suspend.h>
32 #include <linux/pagevec.h>
33 #include <linux/blkdev.h>
34 #include <linux/slab.h>
35 #include <linux/ratelimit.h>
36 #include <linux/oom.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/random.h>
49 #include <linux/sort.h>
50 #include <linux/pfn.h>
51 #include <linux/backing-dev.h>
52 #include <linux/fault-inject.h>
53 #include <linux/page-isolation.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/migrate.h>
63 #include <linux/hugetlb.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/mm.h>
66 #include <linux/page_owner.h>
67 #include <linux/page_table_check.h>
68 #include <linux/kthread.h>
69 #include <linux/memcontrol.h>
70 #include <linux/ftrace.h>
71 #include <linux/lockdep.h>
72 #include <linux/nmi.h>
73 #include <linux/psi.h>
74 #include <linux/padata.h>
75 #include <linux/khugepaged.h>
76 #include <linux/buffer_head.h>
77 #include <linux/delayacct.h>
78 #include <asm/sections.h>
79 #include <asm/tlbflush.h>
80 #include <asm/div64.h>
83 #include "page_reporting.h"
85 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
86 typedef int __bitwise fpi_t;
88 /* No special request */
89 #define FPI_NONE ((__force fpi_t)0)
92 * Skip free page reporting notification for the (possibly merged) page.
93 * This does not hinder free page reporting from grabbing the page,
94 * reporting it and marking it "reported" - it only skips notifying
95 * the free page reporting infrastructure about a newly freed page. For
96 * example, used when temporarily pulling a page from a freelist and
97 * putting it back unmodified.
99 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
102 * Place the (possibly merged) page to the tail of the freelist. Will ignore
103 * page shuffling (relevant code - e.g., memory onlining - is expected to
104 * shuffle the whole zone).
106 * Note: No code should rely on this flag for correctness - it's purely
107 * to allow for optimizations when handing back either fresh pages
108 * (memory onlining) or untouched pages (page isolation, free page
111 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
114 * Don't poison memory with KASAN (only for the tag-based modes).
115 * During boot, all non-reserved memblock memory is exposed to page_alloc.
116 * Poisoning all that memory lengthens boot time, especially on systems with
117 * large amount of RAM. This flag is used to skip that poisoning.
118 * This is only done for the tag-based KASAN modes, as those are able to
119 * detect memory corruptions with the memory tags assigned by default.
120 * All memory allocated normally after boot gets poisoned as usual.
122 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
124 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
125 static DEFINE_MUTEX(pcp_batch_high_lock);
126 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
131 static DEFINE_PER_CPU(struct pagesets, pagesets) __maybe_unused = {
132 .lock = INIT_LOCAL_LOCK(lock),
135 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
136 DEFINE_PER_CPU(int, numa_node);
137 EXPORT_PER_CPU_SYMBOL(numa_node);
140 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
142 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
144 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
145 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
146 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
147 * defined in <linux/topology.h>.
149 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
150 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
153 /* work_structs for global per-cpu drains */
156 struct work_struct work;
158 static DEFINE_MUTEX(pcpu_drain_mutex);
159 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
161 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
162 volatile unsigned long latent_entropy __latent_entropy;
163 EXPORT_SYMBOL(latent_entropy);
167 * Array of node states.
169 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
170 [N_POSSIBLE] = NODE_MASK_ALL,
171 [N_ONLINE] = { { [0] = 1UL } },
173 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
174 #ifdef CONFIG_HIGHMEM
175 [N_HIGH_MEMORY] = { { [0] = 1UL } },
177 [N_MEMORY] = { { [0] = 1UL } },
178 [N_CPU] = { { [0] = 1UL } },
181 EXPORT_SYMBOL(node_states);
183 atomic_long_t _totalram_pages __read_mostly;
184 EXPORT_SYMBOL(_totalram_pages);
185 unsigned long totalreserve_pages __read_mostly;
186 unsigned long totalcma_pages __read_mostly;
188 int percpu_pagelist_high_fraction;
189 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
190 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
191 EXPORT_SYMBOL(init_on_alloc);
193 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
194 EXPORT_SYMBOL(init_on_free);
196 static bool _init_on_alloc_enabled_early __read_mostly
197 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
198 static int __init early_init_on_alloc(char *buf)
201 return kstrtobool(buf, &_init_on_alloc_enabled_early);
203 early_param("init_on_alloc", early_init_on_alloc);
205 static bool _init_on_free_enabled_early __read_mostly
206 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
207 static int __init early_init_on_free(char *buf)
209 return kstrtobool(buf, &_init_on_free_enabled_early);
211 early_param("init_on_free", early_init_on_free);
214 * A cached value of the page's pageblock's migratetype, used when the page is
215 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
216 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
217 * Also the migratetype set in the page does not necessarily match the pcplist
218 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
219 * other index - this ensures that it will be put on the correct CMA freelist.
221 static inline int get_pcppage_migratetype(struct page *page)
226 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
228 page->index = migratetype;
231 #ifdef CONFIG_PM_SLEEP
233 * The following functions are used by the suspend/hibernate code to temporarily
234 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
235 * while devices are suspended. To avoid races with the suspend/hibernate code,
236 * they should always be called with system_transition_mutex held
237 * (gfp_allowed_mask also should only be modified with system_transition_mutex
238 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
239 * with that modification).
242 static gfp_t saved_gfp_mask;
244 void pm_restore_gfp_mask(void)
246 WARN_ON(!mutex_is_locked(&system_transition_mutex));
247 if (saved_gfp_mask) {
248 gfp_allowed_mask = saved_gfp_mask;
253 void pm_restrict_gfp_mask(void)
255 WARN_ON(!mutex_is_locked(&system_transition_mutex));
256 WARN_ON(saved_gfp_mask);
257 saved_gfp_mask = gfp_allowed_mask;
258 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
261 bool pm_suspended_storage(void)
263 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
267 #endif /* CONFIG_PM_SLEEP */
269 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
270 unsigned int pageblock_order __read_mostly;
273 static void __free_pages_ok(struct page *page, unsigned int order,
277 * results with 256, 32 in the lowmem_reserve sysctl:
278 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
279 * 1G machine -> (16M dma, 784M normal, 224M high)
280 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
281 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
282 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
284 * TBD: should special case ZONE_DMA32 machines here - in those we normally
285 * don't need any ZONE_NORMAL reservation
287 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
288 #ifdef CONFIG_ZONE_DMA
291 #ifdef CONFIG_ZONE_DMA32
295 #ifdef CONFIG_HIGHMEM
301 static char * const zone_names[MAX_NR_ZONES] = {
302 #ifdef CONFIG_ZONE_DMA
305 #ifdef CONFIG_ZONE_DMA32
309 #ifdef CONFIG_HIGHMEM
313 #ifdef CONFIG_ZONE_DEVICE
318 const char * const migratetype_names[MIGRATE_TYPES] = {
326 #ifdef CONFIG_MEMORY_ISOLATION
331 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
332 [NULL_COMPOUND_DTOR] = NULL,
333 [COMPOUND_PAGE_DTOR] = free_compound_page,
334 #ifdef CONFIG_HUGETLB_PAGE
335 [HUGETLB_PAGE_DTOR] = free_huge_page,
337 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
338 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
342 int min_free_kbytes = 1024;
343 int user_min_free_kbytes = -1;
344 int watermark_boost_factor __read_mostly = 15000;
345 int watermark_scale_factor = 10;
347 static unsigned long nr_kernel_pages __initdata;
348 static unsigned long nr_all_pages __initdata;
349 static unsigned long dma_reserve __initdata;
351 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
352 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
353 static unsigned long required_kernelcore __initdata;
354 static unsigned long required_kernelcore_percent __initdata;
355 static unsigned long required_movablecore __initdata;
356 static unsigned long required_movablecore_percent __initdata;
357 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
358 static bool mirrored_kernelcore __meminitdata;
360 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
362 EXPORT_SYMBOL(movable_zone);
365 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
366 unsigned int nr_online_nodes __read_mostly = 1;
367 EXPORT_SYMBOL(nr_node_ids);
368 EXPORT_SYMBOL(nr_online_nodes);
371 int page_group_by_mobility_disabled __read_mostly;
373 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
375 * During boot we initialize deferred pages on-demand, as needed, but once
376 * page_alloc_init_late() has finished, the deferred pages are all initialized,
377 * and we can permanently disable that path.
379 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
382 * Calling kasan_poison_pages() only after deferred memory initialization
383 * has completed. Poisoning pages during deferred memory init will greatly
384 * lengthen the process and cause problem in large memory systems as the
385 * deferred pages initialization is done with interrupt disabled.
387 * Assuming that there will be no reference to those newly initialized
388 * pages before they are ever allocated, this should have no effect on
389 * KASAN memory tracking as the poison will be properly inserted at page
390 * allocation time. The only corner case is when pages are allocated by
391 * on-demand allocation and then freed again before the deferred pages
392 * initialization is done, but this is not likely to happen.
394 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
396 return static_branch_unlikely(&deferred_pages) ||
397 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
398 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
399 PageSkipKASanPoison(page);
402 /* Returns true if the struct page for the pfn is uninitialised */
403 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
405 int nid = early_pfn_to_nid(pfn);
407 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
414 * Returns true when the remaining initialisation should be deferred until
415 * later in the boot cycle when it can be parallelised.
417 static bool __meminit
418 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
420 static unsigned long prev_end_pfn, nr_initialised;
423 * prev_end_pfn static that contains the end of previous zone
424 * No need to protect because called very early in boot before smp_init.
426 if (prev_end_pfn != end_pfn) {
427 prev_end_pfn = end_pfn;
431 /* Always populate low zones for address-constrained allocations */
432 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
435 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
438 * We start only with one section of pages, more pages are added as
439 * needed until the rest of deferred pages are initialized.
442 if ((nr_initialised > PAGES_PER_SECTION) &&
443 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
444 NODE_DATA(nid)->first_deferred_pfn = pfn;
450 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
452 return (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
453 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
454 PageSkipKASanPoison(page);
457 static inline bool early_page_uninitialised(unsigned long pfn)
462 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
468 /* Return a pointer to the bitmap storing bits affecting a block of pages */
469 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
472 #ifdef CONFIG_SPARSEMEM
473 return section_to_usemap(__pfn_to_section(pfn));
475 return page_zone(page)->pageblock_flags;
476 #endif /* CONFIG_SPARSEMEM */
479 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
481 #ifdef CONFIG_SPARSEMEM
482 pfn &= (PAGES_PER_SECTION-1);
484 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
485 #endif /* CONFIG_SPARSEMEM */
486 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
489 static __always_inline
490 unsigned long __get_pfnblock_flags_mask(const struct page *page,
494 unsigned long *bitmap;
495 unsigned long bitidx, word_bitidx;
498 bitmap = get_pageblock_bitmap(page, pfn);
499 bitidx = pfn_to_bitidx(page, pfn);
500 word_bitidx = bitidx / BITS_PER_LONG;
501 bitidx &= (BITS_PER_LONG-1);
503 word = bitmap[word_bitidx];
504 return (word >> bitidx) & mask;
508 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
509 * @page: The page within the block of interest
510 * @pfn: The target page frame number
511 * @mask: mask of bits that the caller is interested in
513 * Return: pageblock_bits flags
515 unsigned long get_pfnblock_flags_mask(const struct page *page,
516 unsigned long pfn, unsigned long mask)
518 return __get_pfnblock_flags_mask(page, pfn, mask);
521 static __always_inline int get_pfnblock_migratetype(const struct page *page,
524 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
528 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
529 * @page: The page within the block of interest
530 * @flags: The flags to set
531 * @pfn: The target page frame number
532 * @mask: mask of bits that the caller is interested in
534 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
538 unsigned long *bitmap;
539 unsigned long bitidx, word_bitidx;
540 unsigned long old_word, word;
542 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
543 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
545 bitmap = get_pageblock_bitmap(page, pfn);
546 bitidx = pfn_to_bitidx(page, pfn);
547 word_bitidx = bitidx / BITS_PER_LONG;
548 bitidx &= (BITS_PER_LONG-1);
550 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
555 word = READ_ONCE(bitmap[word_bitidx]);
557 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
558 if (word == old_word)
564 void set_pageblock_migratetype(struct page *page, int migratetype)
566 if (unlikely(page_group_by_mobility_disabled &&
567 migratetype < MIGRATE_PCPTYPES))
568 migratetype = MIGRATE_UNMOVABLE;
570 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
571 page_to_pfn(page), MIGRATETYPE_MASK);
574 #ifdef CONFIG_DEBUG_VM
575 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
579 unsigned long pfn = page_to_pfn(page);
580 unsigned long sp, start_pfn;
583 seq = zone_span_seqbegin(zone);
584 start_pfn = zone->zone_start_pfn;
585 sp = zone->spanned_pages;
586 if (!zone_spans_pfn(zone, pfn))
588 } while (zone_span_seqretry(zone, seq));
591 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
592 pfn, zone_to_nid(zone), zone->name,
593 start_pfn, start_pfn + sp);
598 static int page_is_consistent(struct zone *zone, struct page *page)
600 if (zone != page_zone(page))
606 * Temporary debugging check for pages not lying within a given zone.
608 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
610 if (page_outside_zone_boundaries(zone, page))
612 if (!page_is_consistent(zone, page))
618 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
624 static void bad_page(struct page *page, const char *reason)
626 static unsigned long resume;
627 static unsigned long nr_shown;
628 static unsigned long nr_unshown;
631 * Allow a burst of 60 reports, then keep quiet for that minute;
632 * or allow a steady drip of one report per second.
634 if (nr_shown == 60) {
635 if (time_before(jiffies, resume)) {
641 "BUG: Bad page state: %lu messages suppressed\n",
648 resume = jiffies + 60 * HZ;
650 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
651 current->comm, page_to_pfn(page));
652 dump_page(page, reason);
657 /* Leave bad fields for debug, except PageBuddy could make trouble */
658 page_mapcount_reset(page); /* remove PageBuddy */
659 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
662 static inline unsigned int order_to_pindex(int migratetype, int order)
666 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
667 if (order > PAGE_ALLOC_COSTLY_ORDER) {
668 VM_BUG_ON(order != pageblock_order);
669 base = PAGE_ALLOC_COSTLY_ORDER + 1;
672 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
675 return (MIGRATE_PCPTYPES * base) + migratetype;
678 static inline int pindex_to_order(unsigned int pindex)
680 int order = pindex / MIGRATE_PCPTYPES;
682 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
683 if (order > PAGE_ALLOC_COSTLY_ORDER)
684 order = pageblock_order;
686 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
692 static inline bool pcp_allowed_order(unsigned int order)
694 if (order <= PAGE_ALLOC_COSTLY_ORDER)
696 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
697 if (order == pageblock_order)
703 static inline void free_the_page(struct page *page, unsigned int order)
705 if (pcp_allowed_order(order)) /* Via pcp? */
706 free_unref_page(page, order);
708 __free_pages_ok(page, order, FPI_NONE);
712 * Higher-order pages are called "compound pages". They are structured thusly:
714 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
716 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
717 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
719 * The first tail page's ->compound_dtor holds the offset in array of compound
720 * page destructors. See compound_page_dtors.
722 * The first tail page's ->compound_order holds the order of allocation.
723 * This usage means that zero-order pages may not be compound.
726 void free_compound_page(struct page *page)
728 mem_cgroup_uncharge(page_folio(page));
729 free_the_page(page, compound_order(page));
732 static void prep_compound_head(struct page *page, unsigned int order)
734 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
735 set_compound_order(page, order);
736 atomic_set(compound_mapcount_ptr(page), -1);
737 atomic_set(compound_pincount_ptr(page), 0);
740 static void prep_compound_tail(struct page *head, int tail_idx)
742 struct page *p = head + tail_idx;
744 p->mapping = TAIL_MAPPING;
745 set_compound_head(p, head);
748 void prep_compound_page(struct page *page, unsigned int order)
751 int nr_pages = 1 << order;
754 for (i = 1; i < nr_pages; i++)
755 prep_compound_tail(page, i);
757 prep_compound_head(page, order);
760 #ifdef CONFIG_DEBUG_PAGEALLOC
761 unsigned int _debug_guardpage_minorder;
763 bool _debug_pagealloc_enabled_early __read_mostly
764 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
765 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
766 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
767 EXPORT_SYMBOL(_debug_pagealloc_enabled);
769 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
771 static int __init early_debug_pagealloc(char *buf)
773 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
775 early_param("debug_pagealloc", early_debug_pagealloc);
777 static int __init debug_guardpage_minorder_setup(char *buf)
781 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
782 pr_err("Bad debug_guardpage_minorder value\n");
785 _debug_guardpage_minorder = res;
786 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
789 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
791 static inline bool set_page_guard(struct zone *zone, struct page *page,
792 unsigned int order, int migratetype)
794 if (!debug_guardpage_enabled())
797 if (order >= debug_guardpage_minorder())
800 __SetPageGuard(page);
801 INIT_LIST_HEAD(&page->lru);
802 set_page_private(page, order);
803 /* Guard pages are not available for any usage */
804 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
809 static inline void clear_page_guard(struct zone *zone, struct page *page,
810 unsigned int order, int migratetype)
812 if (!debug_guardpage_enabled())
815 __ClearPageGuard(page);
817 set_page_private(page, 0);
818 if (!is_migrate_isolate(migratetype))
819 __mod_zone_freepage_state(zone, (1 << order), migratetype);
822 static inline bool set_page_guard(struct zone *zone, struct page *page,
823 unsigned int order, int migratetype) { return false; }
824 static inline void clear_page_guard(struct zone *zone, struct page *page,
825 unsigned int order, int migratetype) {}
829 * Enable static keys related to various memory debugging and hardening options.
830 * Some override others, and depend on early params that are evaluated in the
831 * order of appearance. So we need to first gather the full picture of what was
832 * enabled, and then make decisions.
834 void init_mem_debugging_and_hardening(void)
836 bool page_poisoning_requested = false;
838 #ifdef CONFIG_PAGE_POISONING
840 * Page poisoning is debug page alloc for some arches. If
841 * either of those options are enabled, enable poisoning.
843 if (page_poisoning_enabled() ||
844 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
845 debug_pagealloc_enabled())) {
846 static_branch_enable(&_page_poisoning_enabled);
847 page_poisoning_requested = true;
851 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
852 page_poisoning_requested) {
853 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
854 "will take precedence over init_on_alloc and init_on_free\n");
855 _init_on_alloc_enabled_early = false;
856 _init_on_free_enabled_early = false;
859 if (_init_on_alloc_enabled_early)
860 static_branch_enable(&init_on_alloc);
862 static_branch_disable(&init_on_alloc);
864 if (_init_on_free_enabled_early)
865 static_branch_enable(&init_on_free);
867 static_branch_disable(&init_on_free);
869 #ifdef CONFIG_DEBUG_PAGEALLOC
870 if (!debug_pagealloc_enabled())
873 static_branch_enable(&_debug_pagealloc_enabled);
875 if (!debug_guardpage_minorder())
878 static_branch_enable(&_debug_guardpage_enabled);
882 static inline void set_buddy_order(struct page *page, unsigned int order)
884 set_page_private(page, order);
885 __SetPageBuddy(page);
889 * This function checks whether a page is free && is the buddy
890 * we can coalesce a page and its buddy if
891 * (a) the buddy is not in a hole (check before calling!) &&
892 * (b) the buddy is in the buddy system &&
893 * (c) a page and its buddy have the same order &&
894 * (d) a page and its buddy are in the same zone.
896 * For recording whether a page is in the buddy system, we set PageBuddy.
897 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
899 * For recording page's order, we use page_private(page).
901 static inline bool page_is_buddy(struct page *page, struct page *buddy,
904 if (!page_is_guard(buddy) && !PageBuddy(buddy))
907 if (buddy_order(buddy) != order)
911 * zone check is done late to avoid uselessly calculating
912 * zone/node ids for pages that could never merge.
914 if (page_zone_id(page) != page_zone_id(buddy))
917 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
922 #ifdef CONFIG_COMPACTION
923 static inline struct capture_control *task_capc(struct zone *zone)
925 struct capture_control *capc = current->capture_control;
927 return unlikely(capc) &&
928 !(current->flags & PF_KTHREAD) &&
930 capc->cc->zone == zone ? capc : NULL;
934 compaction_capture(struct capture_control *capc, struct page *page,
935 int order, int migratetype)
937 if (!capc || order != capc->cc->order)
940 /* Do not accidentally pollute CMA or isolated regions*/
941 if (is_migrate_cma(migratetype) ||
942 is_migrate_isolate(migratetype))
946 * Do not let lower order allocations pollute a movable pageblock.
947 * This might let an unmovable request use a reclaimable pageblock
948 * and vice-versa but no more than normal fallback logic which can
949 * have trouble finding a high-order free page.
951 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
959 static inline struct capture_control *task_capc(struct zone *zone)
965 compaction_capture(struct capture_control *capc, struct page *page,
966 int order, int migratetype)
970 #endif /* CONFIG_COMPACTION */
972 /* Used for pages not on another list */
973 static inline void add_to_free_list(struct page *page, struct zone *zone,
974 unsigned int order, int migratetype)
976 struct free_area *area = &zone->free_area[order];
978 list_add(&page->lru, &area->free_list[migratetype]);
982 /* Used for pages not on another list */
983 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
984 unsigned int order, int migratetype)
986 struct free_area *area = &zone->free_area[order];
988 list_add_tail(&page->lru, &area->free_list[migratetype]);
993 * Used for pages which are on another list. Move the pages to the tail
994 * of the list - so the moved pages won't immediately be considered for
995 * allocation again (e.g., optimization for memory onlining).
997 static inline void move_to_free_list(struct page *page, struct zone *zone,
998 unsigned int order, int migratetype)
1000 struct free_area *area = &zone->free_area[order];
1002 list_move_tail(&page->lru, &area->free_list[migratetype]);
1005 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
1008 /* clear reported state and update reported page count */
1009 if (page_reported(page))
1010 __ClearPageReported(page);
1012 list_del(&page->lru);
1013 __ClearPageBuddy(page);
1014 set_page_private(page, 0);
1015 zone->free_area[order].nr_free--;
1019 * If this is not the largest possible page, check if the buddy
1020 * of the next-highest order is free. If it is, it's possible
1021 * that pages are being freed that will coalesce soon. In case,
1022 * that is happening, add the free page to the tail of the list
1023 * so it's less likely to be used soon and more likely to be merged
1024 * as a higher order page
1027 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1028 struct page *page, unsigned int order)
1030 struct page *higher_page, *higher_buddy;
1031 unsigned long combined_pfn;
1033 if (order >= MAX_ORDER - 2)
1036 combined_pfn = buddy_pfn & pfn;
1037 higher_page = page + (combined_pfn - pfn);
1038 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1039 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1041 return page_is_buddy(higher_page, higher_buddy, order + 1);
1045 * Freeing function for a buddy system allocator.
1047 * The concept of a buddy system is to maintain direct-mapped table
1048 * (containing bit values) for memory blocks of various "orders".
1049 * The bottom level table contains the map for the smallest allocatable
1050 * units of memory (here, pages), and each level above it describes
1051 * pairs of units from the levels below, hence, "buddies".
1052 * At a high level, all that happens here is marking the table entry
1053 * at the bottom level available, and propagating the changes upward
1054 * as necessary, plus some accounting needed to play nicely with other
1055 * parts of the VM system.
1056 * At each level, we keep a list of pages, which are heads of continuous
1057 * free pages of length of (1 << order) and marked with PageBuddy.
1058 * Page's order is recorded in page_private(page) field.
1059 * So when we are allocating or freeing one, we can derive the state of the
1060 * other. That is, if we allocate a small block, and both were
1061 * free, the remainder of the region must be split into blocks.
1062 * If a block is freed, and its buddy is also free, then this
1063 * triggers coalescing into a block of larger size.
1068 static inline void __free_one_page(struct page *page,
1070 struct zone *zone, unsigned int order,
1071 int migratetype, fpi_t fpi_flags)
1073 struct capture_control *capc = task_capc(zone);
1074 unsigned int max_order = pageblock_order;
1075 unsigned long buddy_pfn;
1076 unsigned long combined_pfn;
1080 VM_BUG_ON(!zone_is_initialized(zone));
1081 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1083 VM_BUG_ON(migratetype == -1);
1084 if (likely(!is_migrate_isolate(migratetype)))
1085 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1087 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1088 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1091 while (order < max_order) {
1092 if (compaction_capture(capc, page, order, migratetype)) {
1093 __mod_zone_freepage_state(zone, -(1 << order),
1097 buddy_pfn = __find_buddy_pfn(pfn, order);
1098 buddy = page + (buddy_pfn - pfn);
1100 if (!page_is_buddy(page, buddy, order))
1103 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1104 * merge with it and move up one order.
1106 if (page_is_guard(buddy))
1107 clear_page_guard(zone, buddy, order, migratetype);
1109 del_page_from_free_list(buddy, zone, order);
1110 combined_pfn = buddy_pfn & pfn;
1111 page = page + (combined_pfn - pfn);
1115 if (order < MAX_ORDER - 1) {
1116 /* If we are here, it means order is >= pageblock_order.
1117 * We want to prevent merge between freepages on pageblock
1118 * without fallbacks and normal pageblock. Without this,
1119 * pageblock isolation could cause incorrect freepage or CMA
1120 * accounting or HIGHATOMIC accounting.
1122 * We don't want to hit this code for the more frequent
1123 * low-order merging.
1127 buddy_pfn = __find_buddy_pfn(pfn, order);
1128 buddy = page + (buddy_pfn - pfn);
1129 buddy_mt = get_pageblock_migratetype(buddy);
1131 if (migratetype != buddy_mt
1132 && (!migratetype_is_mergeable(migratetype) ||
1133 !migratetype_is_mergeable(buddy_mt)))
1135 max_order = order + 1;
1136 goto continue_merging;
1140 set_buddy_order(page, order);
1142 if (fpi_flags & FPI_TO_TAIL)
1144 else if (is_shuffle_order(order))
1145 to_tail = shuffle_pick_tail();
1147 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1150 add_to_free_list_tail(page, zone, order, migratetype);
1152 add_to_free_list(page, zone, order, migratetype);
1154 /* Notify page reporting subsystem of freed page */
1155 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1156 page_reporting_notify_free(order);
1160 * A bad page could be due to a number of fields. Instead of multiple branches,
1161 * try and check multiple fields with one check. The caller must do a detailed
1162 * check if necessary.
1164 static inline bool page_expected_state(struct page *page,
1165 unsigned long check_flags)
1167 if (unlikely(atomic_read(&page->_mapcount) != -1))
1170 if (unlikely((unsigned long)page->mapping |
1171 page_ref_count(page) |
1175 (page->flags & check_flags)))
1181 static const char *page_bad_reason(struct page *page, unsigned long flags)
1183 const char *bad_reason = NULL;
1185 if (unlikely(atomic_read(&page->_mapcount) != -1))
1186 bad_reason = "nonzero mapcount";
1187 if (unlikely(page->mapping != NULL))
1188 bad_reason = "non-NULL mapping";
1189 if (unlikely(page_ref_count(page) != 0))
1190 bad_reason = "nonzero _refcount";
1191 if (unlikely(page->flags & flags)) {
1192 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1193 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1195 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1198 if (unlikely(page->memcg_data))
1199 bad_reason = "page still charged to cgroup";
1204 static void check_free_page_bad(struct page *page)
1207 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1210 static inline int check_free_page(struct page *page)
1212 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1215 /* Something has gone sideways, find it */
1216 check_free_page_bad(page);
1220 static int free_tail_pages_check(struct page *head_page, struct page *page)
1225 * We rely page->lru.next never has bit 0 set, unless the page
1226 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1228 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1230 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1234 switch (page - head_page) {
1236 /* the first tail page: ->mapping may be compound_mapcount() */
1237 if (unlikely(compound_mapcount(page))) {
1238 bad_page(page, "nonzero compound_mapcount");
1244 * the second tail page: ->mapping is
1245 * deferred_list.next -- ignore value.
1249 if (page->mapping != TAIL_MAPPING) {
1250 bad_page(page, "corrupted mapping in tail page");
1255 if (unlikely(!PageTail(page))) {
1256 bad_page(page, "PageTail not set");
1259 if (unlikely(compound_head(page) != head_page)) {
1260 bad_page(page, "compound_head not consistent");
1265 page->mapping = NULL;
1266 clear_compound_head(page);
1270 static void kernel_init_free_pages(struct page *page, int numpages, bool zero_tags)
1275 for (i = 0; i < numpages; i++)
1276 tag_clear_highpage(page + i);
1280 /* s390's use of memset() could override KASAN redzones. */
1281 kasan_disable_current();
1282 for (i = 0; i < numpages; i++) {
1283 u8 tag = page_kasan_tag(page + i);
1284 page_kasan_tag_reset(page + i);
1285 clear_highpage(page + i);
1286 page_kasan_tag_set(page + i, tag);
1288 kasan_enable_current();
1291 static __always_inline bool free_pages_prepare(struct page *page,
1292 unsigned int order, bool check_free, fpi_t fpi_flags)
1295 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1297 VM_BUG_ON_PAGE(PageTail(page), page);
1299 trace_mm_page_free(page, order);
1301 if (unlikely(PageHWPoison(page)) && !order) {
1303 * Do not let hwpoison pages hit pcplists/buddy
1304 * Untie memcg state and reset page's owner
1306 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1307 __memcg_kmem_uncharge_page(page, order);
1308 reset_page_owner(page, order);
1309 page_table_check_free(page, order);
1314 * Check tail pages before head page information is cleared to
1315 * avoid checking PageCompound for order-0 pages.
1317 if (unlikely(order)) {
1318 bool compound = PageCompound(page);
1321 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1324 ClearPageDoubleMap(page);
1325 ClearPageHasHWPoisoned(page);
1327 for (i = 1; i < (1 << order); i++) {
1329 bad += free_tail_pages_check(page, page + i);
1330 if (unlikely(check_free_page(page + i))) {
1334 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1337 if (PageMappingFlags(page))
1338 page->mapping = NULL;
1339 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1340 __memcg_kmem_uncharge_page(page, order);
1342 bad += check_free_page(page);
1346 page_cpupid_reset_last(page);
1347 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1348 reset_page_owner(page, order);
1349 page_table_check_free(page, order);
1351 if (!PageHighMem(page)) {
1352 debug_check_no_locks_freed(page_address(page),
1353 PAGE_SIZE << order);
1354 debug_check_no_obj_freed(page_address(page),
1355 PAGE_SIZE << order);
1358 kernel_poison_pages(page, 1 << order);
1361 * As memory initialization might be integrated into KASAN,
1362 * kasan_free_pages and kernel_init_free_pages must be
1363 * kept together to avoid discrepancies in behavior.
1365 * With hardware tag-based KASAN, memory tags must be set before the
1366 * page becomes unavailable via debug_pagealloc or arch_free_page.
1368 if (kasan_has_integrated_init()) {
1369 if (!skip_kasan_poison)
1370 kasan_free_pages(page, order);
1372 bool init = want_init_on_free();
1375 kernel_init_free_pages(page, 1 << order, false);
1376 if (!skip_kasan_poison)
1377 kasan_poison_pages(page, order, init);
1381 * arch_free_page() can make the page's contents inaccessible. s390
1382 * does this. So nothing which can access the page's contents should
1383 * happen after this.
1385 arch_free_page(page, order);
1387 debug_pagealloc_unmap_pages(page, 1 << order);
1392 #ifdef CONFIG_DEBUG_VM
1394 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1395 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1396 * moved from pcp lists to free lists.
1398 static bool free_pcp_prepare(struct page *page, unsigned int order)
1400 return free_pages_prepare(page, order, true, FPI_NONE);
1403 static bool bulkfree_pcp_prepare(struct page *page)
1405 if (debug_pagealloc_enabled_static())
1406 return check_free_page(page);
1412 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1413 * moving from pcp lists to free list in order to reduce overhead. With
1414 * debug_pagealloc enabled, they are checked also immediately when being freed
1417 static bool free_pcp_prepare(struct page *page, unsigned int order)
1419 if (debug_pagealloc_enabled_static())
1420 return free_pages_prepare(page, order, true, FPI_NONE);
1422 return free_pages_prepare(page, order, false, FPI_NONE);
1425 static bool bulkfree_pcp_prepare(struct page *page)
1427 return check_free_page(page);
1429 #endif /* CONFIG_DEBUG_VM */
1432 * Frees a number of pages from the PCP lists
1433 * Assumes all pages on list are in same zone.
1434 * count is the number of pages to free.
1436 static void free_pcppages_bulk(struct zone *zone, int count,
1437 struct per_cpu_pages *pcp,
1441 int max_pindex = NR_PCP_LISTS - 1;
1443 bool isolated_pageblocks;
1447 * Ensure proper count is passed which otherwise would stuck in the
1448 * below while (list_empty(list)) loop.
1450 count = min(pcp->count, count);
1452 /* Ensure requested pindex is drained first. */
1453 pindex = pindex - 1;
1456 * local_lock_irq held so equivalent to spin_lock_irqsave for
1457 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1459 spin_lock(&zone->lock);
1460 isolated_pageblocks = has_isolate_pageblock(zone);
1463 struct list_head *list;
1466 /* Remove pages from lists in a round-robin fashion. */
1468 if (++pindex > max_pindex)
1469 pindex = min_pindex;
1470 list = &pcp->lists[pindex];
1471 if (!list_empty(list))
1474 if (pindex == max_pindex)
1476 if (pindex == min_pindex)
1480 order = pindex_to_order(pindex);
1481 nr_pages = 1 << order;
1482 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1486 page = list_last_entry(list, struct page, lru);
1487 mt = get_pcppage_migratetype(page);
1489 /* must delete to avoid corrupting pcp list */
1490 list_del(&page->lru);
1492 pcp->count -= nr_pages;
1494 if (bulkfree_pcp_prepare(page))
1497 /* MIGRATE_ISOLATE page should not go to pcplists */
1498 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1499 /* Pageblock could have been isolated meanwhile */
1500 if (unlikely(isolated_pageblocks))
1501 mt = get_pageblock_migratetype(page);
1503 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1504 trace_mm_page_pcpu_drain(page, order, mt);
1505 } while (count > 0 && !list_empty(list));
1508 spin_unlock(&zone->lock);
1511 static void free_one_page(struct zone *zone,
1512 struct page *page, unsigned long pfn,
1514 int migratetype, fpi_t fpi_flags)
1516 unsigned long flags;
1518 spin_lock_irqsave(&zone->lock, flags);
1519 if (unlikely(has_isolate_pageblock(zone) ||
1520 is_migrate_isolate(migratetype))) {
1521 migratetype = get_pfnblock_migratetype(page, pfn);
1523 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1524 spin_unlock_irqrestore(&zone->lock, flags);
1527 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1528 unsigned long zone, int nid)
1530 mm_zero_struct_page(page);
1531 set_page_links(page, zone, nid, pfn);
1532 init_page_count(page);
1533 page_mapcount_reset(page);
1534 page_cpupid_reset_last(page);
1535 page_kasan_tag_reset(page);
1537 INIT_LIST_HEAD(&page->lru);
1538 #ifdef WANT_PAGE_VIRTUAL
1539 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1540 if (!is_highmem_idx(zone))
1541 set_page_address(page, __va(pfn << PAGE_SHIFT));
1545 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1546 static void __meminit init_reserved_page(unsigned long pfn)
1551 if (!early_page_uninitialised(pfn))
1554 nid = early_pfn_to_nid(pfn);
1555 pgdat = NODE_DATA(nid);
1557 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1558 struct zone *zone = &pgdat->node_zones[zid];
1560 if (zone_spans_pfn(zone, pfn))
1563 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1566 static inline void init_reserved_page(unsigned long pfn)
1569 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1572 * Initialised pages do not have PageReserved set. This function is
1573 * called for each range allocated by the bootmem allocator and
1574 * marks the pages PageReserved. The remaining valid pages are later
1575 * sent to the buddy page allocator.
1577 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1579 unsigned long start_pfn = PFN_DOWN(start);
1580 unsigned long end_pfn = PFN_UP(end);
1582 for (; start_pfn < end_pfn; start_pfn++) {
1583 if (pfn_valid(start_pfn)) {
1584 struct page *page = pfn_to_page(start_pfn);
1586 init_reserved_page(start_pfn);
1588 /* Avoid false-positive PageTail() */
1589 INIT_LIST_HEAD(&page->lru);
1592 * no need for atomic set_bit because the struct
1593 * page is not visible yet so nobody should
1596 __SetPageReserved(page);
1601 static void __free_pages_ok(struct page *page, unsigned int order,
1604 unsigned long flags;
1606 unsigned long pfn = page_to_pfn(page);
1607 struct zone *zone = page_zone(page);
1609 if (!free_pages_prepare(page, order, true, fpi_flags))
1612 migratetype = get_pfnblock_migratetype(page, pfn);
1614 spin_lock_irqsave(&zone->lock, flags);
1615 if (unlikely(has_isolate_pageblock(zone) ||
1616 is_migrate_isolate(migratetype))) {
1617 migratetype = get_pfnblock_migratetype(page, pfn);
1619 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1620 spin_unlock_irqrestore(&zone->lock, flags);
1622 __count_vm_events(PGFREE, 1 << order);
1625 void __free_pages_core(struct page *page, unsigned int order)
1627 unsigned int nr_pages = 1 << order;
1628 struct page *p = page;
1632 * When initializing the memmap, __init_single_page() sets the refcount
1633 * of all pages to 1 ("allocated"/"not free"). We have to set the
1634 * refcount of all involved pages to 0.
1637 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1639 __ClearPageReserved(p);
1640 set_page_count(p, 0);
1642 __ClearPageReserved(p);
1643 set_page_count(p, 0);
1645 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1648 * Bypass PCP and place fresh pages right to the tail, primarily
1649 * relevant for memory onlining.
1651 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1657 * During memory init memblocks map pfns to nids. The search is expensive and
1658 * this caches recent lookups. The implementation of __early_pfn_to_nid
1659 * treats start/end as pfns.
1661 struct mminit_pfnnid_cache {
1662 unsigned long last_start;
1663 unsigned long last_end;
1667 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1670 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1672 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1673 struct mminit_pfnnid_cache *state)
1675 unsigned long start_pfn, end_pfn;
1678 if (state->last_start <= pfn && pfn < state->last_end)
1679 return state->last_nid;
1681 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1682 if (nid != NUMA_NO_NODE) {
1683 state->last_start = start_pfn;
1684 state->last_end = end_pfn;
1685 state->last_nid = nid;
1691 int __meminit early_pfn_to_nid(unsigned long pfn)
1693 static DEFINE_SPINLOCK(early_pfn_lock);
1696 spin_lock(&early_pfn_lock);
1697 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1699 nid = first_online_node;
1700 spin_unlock(&early_pfn_lock);
1704 #endif /* CONFIG_NUMA */
1706 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1709 if (early_page_uninitialised(pfn))
1711 __free_pages_core(page, order);
1715 * Check that the whole (or subset of) a pageblock given by the interval of
1716 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1717 * with the migration of free compaction scanner.
1719 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1721 * It's possible on some configurations to have a setup like node0 node1 node0
1722 * i.e. it's possible that all pages within a zones range of pages do not
1723 * belong to a single zone. We assume that a border between node0 and node1
1724 * can occur within a single pageblock, but not a node0 node1 node0
1725 * interleaving within a single pageblock. It is therefore sufficient to check
1726 * the first and last page of a pageblock and avoid checking each individual
1727 * page in a pageblock.
1729 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1730 unsigned long end_pfn, struct zone *zone)
1732 struct page *start_page;
1733 struct page *end_page;
1735 /* end_pfn is one past the range we are checking */
1738 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1741 start_page = pfn_to_online_page(start_pfn);
1745 if (page_zone(start_page) != zone)
1748 end_page = pfn_to_page(end_pfn);
1750 /* This gives a shorter code than deriving page_zone(end_page) */
1751 if (page_zone_id(start_page) != page_zone_id(end_page))
1757 void set_zone_contiguous(struct zone *zone)
1759 unsigned long block_start_pfn = zone->zone_start_pfn;
1760 unsigned long block_end_pfn;
1762 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1763 for (; block_start_pfn < zone_end_pfn(zone);
1764 block_start_pfn = block_end_pfn,
1765 block_end_pfn += pageblock_nr_pages) {
1767 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1769 if (!__pageblock_pfn_to_page(block_start_pfn,
1770 block_end_pfn, zone))
1775 /* We confirm that there is no hole */
1776 zone->contiguous = true;
1779 void clear_zone_contiguous(struct zone *zone)
1781 zone->contiguous = false;
1784 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1785 static void __init deferred_free_range(unsigned long pfn,
1786 unsigned long nr_pages)
1794 page = pfn_to_page(pfn);
1796 /* Free a large naturally-aligned chunk if possible */
1797 if (nr_pages == pageblock_nr_pages &&
1798 (pfn & (pageblock_nr_pages - 1)) == 0) {
1799 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1800 __free_pages_core(page, pageblock_order);
1804 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1805 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1806 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1807 __free_pages_core(page, 0);
1811 /* Completion tracking for deferred_init_memmap() threads */
1812 static atomic_t pgdat_init_n_undone __initdata;
1813 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1815 static inline void __init pgdat_init_report_one_done(void)
1817 if (atomic_dec_and_test(&pgdat_init_n_undone))
1818 complete(&pgdat_init_all_done_comp);
1822 * Returns true if page needs to be initialized or freed to buddy allocator.
1824 * First we check if pfn is valid on architectures where it is possible to have
1825 * holes within pageblock_nr_pages. On systems where it is not possible, this
1826 * function is optimized out.
1828 * Then, we check if a current large page is valid by only checking the validity
1831 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1833 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1839 * Free pages to buddy allocator. Try to free aligned pages in
1840 * pageblock_nr_pages sizes.
1842 static void __init deferred_free_pages(unsigned long pfn,
1843 unsigned long end_pfn)
1845 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1846 unsigned long nr_free = 0;
1848 for (; pfn < end_pfn; pfn++) {
1849 if (!deferred_pfn_valid(pfn)) {
1850 deferred_free_range(pfn - nr_free, nr_free);
1852 } else if (!(pfn & nr_pgmask)) {
1853 deferred_free_range(pfn - nr_free, nr_free);
1859 /* Free the last block of pages to allocator */
1860 deferred_free_range(pfn - nr_free, nr_free);
1864 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1865 * by performing it only once every pageblock_nr_pages.
1866 * Return number of pages initialized.
1868 static unsigned long __init deferred_init_pages(struct zone *zone,
1870 unsigned long end_pfn)
1872 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1873 int nid = zone_to_nid(zone);
1874 unsigned long nr_pages = 0;
1875 int zid = zone_idx(zone);
1876 struct page *page = NULL;
1878 for (; pfn < end_pfn; pfn++) {
1879 if (!deferred_pfn_valid(pfn)) {
1882 } else if (!page || !(pfn & nr_pgmask)) {
1883 page = pfn_to_page(pfn);
1887 __init_single_page(page, pfn, zid, nid);
1894 * This function is meant to pre-load the iterator for the zone init.
1895 * Specifically it walks through the ranges until we are caught up to the
1896 * first_init_pfn value and exits there. If we never encounter the value we
1897 * return false indicating there are no valid ranges left.
1900 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1901 unsigned long *spfn, unsigned long *epfn,
1902 unsigned long first_init_pfn)
1907 * Start out by walking through the ranges in this zone that have
1908 * already been initialized. We don't need to do anything with them
1909 * so we just need to flush them out of the system.
1911 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1912 if (*epfn <= first_init_pfn)
1914 if (*spfn < first_init_pfn)
1915 *spfn = first_init_pfn;
1924 * Initialize and free pages. We do it in two loops: first we initialize
1925 * struct page, then free to buddy allocator, because while we are
1926 * freeing pages we can access pages that are ahead (computing buddy
1927 * page in __free_one_page()).
1929 * In order to try and keep some memory in the cache we have the loop
1930 * broken along max page order boundaries. This way we will not cause
1931 * any issues with the buddy page computation.
1933 static unsigned long __init
1934 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1935 unsigned long *end_pfn)
1937 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1938 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1939 unsigned long nr_pages = 0;
1942 /* First we loop through and initialize the page values */
1943 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1946 if (mo_pfn <= *start_pfn)
1949 t = min(mo_pfn, *end_pfn);
1950 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1952 if (mo_pfn < *end_pfn) {
1953 *start_pfn = mo_pfn;
1958 /* Reset values and now loop through freeing pages as needed */
1961 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1967 t = min(mo_pfn, epfn);
1968 deferred_free_pages(spfn, t);
1978 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1981 unsigned long spfn, epfn;
1982 struct zone *zone = arg;
1985 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1988 * Initialize and free pages in MAX_ORDER sized increments so that we
1989 * can avoid introducing any issues with the buddy allocator.
1991 while (spfn < end_pfn) {
1992 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1997 /* An arch may override for more concurrency. */
1999 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2004 /* Initialise remaining memory on a node */
2005 static int __init deferred_init_memmap(void *data)
2007 pg_data_t *pgdat = data;
2008 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2009 unsigned long spfn = 0, epfn = 0;
2010 unsigned long first_init_pfn, flags;
2011 unsigned long start = jiffies;
2013 int zid, max_threads;
2016 /* Bind memory initialisation thread to a local node if possible */
2017 if (!cpumask_empty(cpumask))
2018 set_cpus_allowed_ptr(current, cpumask);
2020 pgdat_resize_lock(pgdat, &flags);
2021 first_init_pfn = pgdat->first_deferred_pfn;
2022 if (first_init_pfn == ULONG_MAX) {
2023 pgdat_resize_unlock(pgdat, &flags);
2024 pgdat_init_report_one_done();
2028 /* Sanity check boundaries */
2029 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2030 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2031 pgdat->first_deferred_pfn = ULONG_MAX;
2034 * Once we unlock here, the zone cannot be grown anymore, thus if an
2035 * interrupt thread must allocate this early in boot, zone must be
2036 * pre-grown prior to start of deferred page initialization.
2038 pgdat_resize_unlock(pgdat, &flags);
2040 /* Only the highest zone is deferred so find it */
2041 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2042 zone = pgdat->node_zones + zid;
2043 if (first_init_pfn < zone_end_pfn(zone))
2047 /* If the zone is empty somebody else may have cleared out the zone */
2048 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2052 max_threads = deferred_page_init_max_threads(cpumask);
2054 while (spfn < epfn) {
2055 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2056 struct padata_mt_job job = {
2057 .thread_fn = deferred_init_memmap_chunk,
2060 .size = epfn_align - spfn,
2061 .align = PAGES_PER_SECTION,
2062 .min_chunk = PAGES_PER_SECTION,
2063 .max_threads = max_threads,
2066 padata_do_multithreaded(&job);
2067 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2071 /* Sanity check that the next zone really is unpopulated */
2072 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2074 pr_info("node %d deferred pages initialised in %ums\n",
2075 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2077 pgdat_init_report_one_done();
2082 * If this zone has deferred pages, try to grow it by initializing enough
2083 * deferred pages to satisfy the allocation specified by order, rounded up to
2084 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2085 * of SECTION_SIZE bytes by initializing struct pages in increments of
2086 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2088 * Return true when zone was grown, otherwise return false. We return true even
2089 * when we grow less than requested, to let the caller decide if there are
2090 * enough pages to satisfy the allocation.
2092 * Note: We use noinline because this function is needed only during boot, and
2093 * it is called from a __ref function _deferred_grow_zone. This way we are
2094 * making sure that it is not inlined into permanent text section.
2096 static noinline bool __init
2097 deferred_grow_zone(struct zone *zone, unsigned int order)
2099 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2100 pg_data_t *pgdat = zone->zone_pgdat;
2101 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2102 unsigned long spfn, epfn, flags;
2103 unsigned long nr_pages = 0;
2106 /* Only the last zone may have deferred pages */
2107 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2110 pgdat_resize_lock(pgdat, &flags);
2113 * If someone grew this zone while we were waiting for spinlock, return
2114 * true, as there might be enough pages already.
2116 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2117 pgdat_resize_unlock(pgdat, &flags);
2121 /* If the zone is empty somebody else may have cleared out the zone */
2122 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2123 first_deferred_pfn)) {
2124 pgdat->first_deferred_pfn = ULONG_MAX;
2125 pgdat_resize_unlock(pgdat, &flags);
2126 /* Retry only once. */
2127 return first_deferred_pfn != ULONG_MAX;
2131 * Initialize and free pages in MAX_ORDER sized increments so
2132 * that we can avoid introducing any issues with the buddy
2135 while (spfn < epfn) {
2136 /* update our first deferred PFN for this section */
2137 first_deferred_pfn = spfn;
2139 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2140 touch_nmi_watchdog();
2142 /* We should only stop along section boundaries */
2143 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2146 /* If our quota has been met we can stop here */
2147 if (nr_pages >= nr_pages_needed)
2151 pgdat->first_deferred_pfn = spfn;
2152 pgdat_resize_unlock(pgdat, &flags);
2154 return nr_pages > 0;
2158 * deferred_grow_zone() is __init, but it is called from
2159 * get_page_from_freelist() during early boot until deferred_pages permanently
2160 * disables this call. This is why we have refdata wrapper to avoid warning,
2161 * and to ensure that the function body gets unloaded.
2164 _deferred_grow_zone(struct zone *zone, unsigned int order)
2166 return deferred_grow_zone(zone, order);
2169 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2171 void __init page_alloc_init_late(void)
2176 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2178 /* There will be num_node_state(N_MEMORY) threads */
2179 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2180 for_each_node_state(nid, N_MEMORY) {
2181 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2184 /* Block until all are initialised */
2185 wait_for_completion(&pgdat_init_all_done_comp);
2188 * We initialized the rest of the deferred pages. Permanently disable
2189 * on-demand struct page initialization.
2191 static_branch_disable(&deferred_pages);
2193 /* Reinit limits that are based on free pages after the kernel is up */
2194 files_maxfiles_init();
2199 /* Discard memblock private memory */
2202 for_each_node_state(nid, N_MEMORY)
2203 shuffle_free_memory(NODE_DATA(nid));
2205 for_each_populated_zone(zone)
2206 set_zone_contiguous(zone);
2210 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2211 void __init init_cma_reserved_pageblock(struct page *page)
2213 unsigned i = pageblock_nr_pages;
2214 struct page *p = page;
2217 __ClearPageReserved(p);
2218 set_page_count(p, 0);
2221 set_pageblock_migratetype(page, MIGRATE_CMA);
2222 set_page_refcounted(page);
2223 __free_pages(page, pageblock_order);
2225 adjust_managed_page_count(page, pageblock_nr_pages);
2226 page_zone(page)->cma_pages += pageblock_nr_pages;
2231 * The order of subdivision here is critical for the IO subsystem.
2232 * Please do not alter this order without good reasons and regression
2233 * testing. Specifically, as large blocks of memory are subdivided,
2234 * the order in which smaller blocks are delivered depends on the order
2235 * they're subdivided in this function. This is the primary factor
2236 * influencing the order in which pages are delivered to the IO
2237 * subsystem according to empirical testing, and this is also justified
2238 * by considering the behavior of a buddy system containing a single
2239 * large block of memory acted on by a series of small allocations.
2240 * This behavior is a critical factor in sglist merging's success.
2244 static inline void expand(struct zone *zone, struct page *page,
2245 int low, int high, int migratetype)
2247 unsigned long size = 1 << high;
2249 while (high > low) {
2252 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2255 * Mark as guard pages (or page), that will allow to
2256 * merge back to allocator when buddy will be freed.
2257 * Corresponding page table entries will not be touched,
2258 * pages will stay not present in virtual address space
2260 if (set_page_guard(zone, &page[size], high, migratetype))
2263 add_to_free_list(&page[size], zone, high, migratetype);
2264 set_buddy_order(&page[size], high);
2268 static void check_new_page_bad(struct page *page)
2270 if (unlikely(page->flags & __PG_HWPOISON)) {
2271 /* Don't complain about hwpoisoned pages */
2272 page_mapcount_reset(page); /* remove PageBuddy */
2277 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2281 * This page is about to be returned from the page allocator
2283 static inline int check_new_page(struct page *page)
2285 if (likely(page_expected_state(page,
2286 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2289 check_new_page_bad(page);
2293 static bool check_new_pages(struct page *page, unsigned int order)
2296 for (i = 0; i < (1 << order); i++) {
2297 struct page *p = page + i;
2299 if (unlikely(check_new_page(p)))
2306 #ifdef CONFIG_DEBUG_VM
2308 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2309 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2310 * also checked when pcp lists are refilled from the free lists.
2312 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2314 if (debug_pagealloc_enabled_static())
2315 return check_new_pages(page, order);
2320 static inline bool check_new_pcp(struct page *page, unsigned int order)
2322 return check_new_pages(page, order);
2326 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2327 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2328 * enabled, they are also checked when being allocated from the pcp lists.
2330 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2332 return check_new_pages(page, order);
2334 static inline bool check_new_pcp(struct page *page, unsigned int order)
2336 if (debug_pagealloc_enabled_static())
2337 return check_new_pages(page, order);
2341 #endif /* CONFIG_DEBUG_VM */
2343 inline void post_alloc_hook(struct page *page, unsigned int order,
2346 set_page_private(page, 0);
2347 set_page_refcounted(page);
2349 arch_alloc_page(page, order);
2350 debug_pagealloc_map_pages(page, 1 << order);
2353 * Page unpoisoning must happen before memory initialization.
2354 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2355 * allocations and the page unpoisoning code will complain.
2357 kernel_unpoison_pages(page, 1 << order);
2360 * As memory initialization might be integrated into KASAN,
2361 * kasan_alloc_pages and kernel_init_free_pages must be
2362 * kept together to avoid discrepancies in behavior.
2364 if (kasan_has_integrated_init()) {
2365 kasan_alloc_pages(page, order, gfp_flags);
2367 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2369 kasan_unpoison_pages(page, order, init);
2371 kernel_init_free_pages(page, 1 << order,
2372 gfp_flags & __GFP_ZEROTAGS);
2375 set_page_owner(page, order, gfp_flags);
2376 page_table_check_alloc(page, order);
2379 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2380 unsigned int alloc_flags)
2382 post_alloc_hook(page, order, gfp_flags);
2384 if (order && (gfp_flags & __GFP_COMP))
2385 prep_compound_page(page, order);
2388 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2389 * allocate the page. The expectation is that the caller is taking
2390 * steps that will free more memory. The caller should avoid the page
2391 * being used for !PFMEMALLOC purposes.
2393 if (alloc_flags & ALLOC_NO_WATERMARKS)
2394 set_page_pfmemalloc(page);
2396 clear_page_pfmemalloc(page);
2400 * Go through the free lists for the given migratetype and remove
2401 * the smallest available page from the freelists
2403 static __always_inline
2404 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2407 unsigned int current_order;
2408 struct free_area *area;
2411 /* Find a page of the appropriate size in the preferred list */
2412 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2413 area = &(zone->free_area[current_order]);
2414 page = get_page_from_free_area(area, migratetype);
2417 del_page_from_free_list(page, zone, current_order);
2418 expand(zone, page, order, current_order, migratetype);
2419 set_pcppage_migratetype(page, migratetype);
2428 * This array describes the order lists are fallen back to when
2429 * the free lists for the desirable migrate type are depleted
2431 * The other migratetypes do not have fallbacks.
2433 static int fallbacks[MIGRATE_TYPES][3] = {
2434 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2435 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2436 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2440 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2443 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2446 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2447 unsigned int order) { return NULL; }
2451 * Move the free pages in a range to the freelist tail of the requested type.
2452 * Note that start_page and end_pages are not aligned on a pageblock
2453 * boundary. If alignment is required, use move_freepages_block()
2455 static int move_freepages(struct zone *zone,
2456 unsigned long start_pfn, unsigned long end_pfn,
2457 int migratetype, int *num_movable)
2462 int pages_moved = 0;
2464 for (pfn = start_pfn; pfn <= end_pfn;) {
2465 page = pfn_to_page(pfn);
2466 if (!PageBuddy(page)) {
2468 * We assume that pages that could be isolated for
2469 * migration are movable. But we don't actually try
2470 * isolating, as that would be expensive.
2473 (PageLRU(page) || __PageMovable(page)))
2479 /* Make sure we are not inadvertently changing nodes */
2480 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2481 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2483 order = buddy_order(page);
2484 move_to_free_list(page, zone, order, migratetype);
2486 pages_moved += 1 << order;
2492 int move_freepages_block(struct zone *zone, struct page *page,
2493 int migratetype, int *num_movable)
2495 unsigned long start_pfn, end_pfn, pfn;
2500 pfn = page_to_pfn(page);
2501 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2502 end_pfn = start_pfn + pageblock_nr_pages - 1;
2504 /* Do not cross zone boundaries */
2505 if (!zone_spans_pfn(zone, start_pfn))
2507 if (!zone_spans_pfn(zone, end_pfn))
2510 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2514 static void change_pageblock_range(struct page *pageblock_page,
2515 int start_order, int migratetype)
2517 int nr_pageblocks = 1 << (start_order - pageblock_order);
2519 while (nr_pageblocks--) {
2520 set_pageblock_migratetype(pageblock_page, migratetype);
2521 pageblock_page += pageblock_nr_pages;
2526 * When we are falling back to another migratetype during allocation, try to
2527 * steal extra free pages from the same pageblocks to satisfy further
2528 * allocations, instead of polluting multiple pageblocks.
2530 * If we are stealing a relatively large buddy page, it is likely there will
2531 * be more free pages in the pageblock, so try to steal them all. For
2532 * reclaimable and unmovable allocations, we steal regardless of page size,
2533 * as fragmentation caused by those allocations polluting movable pageblocks
2534 * is worse than movable allocations stealing from unmovable and reclaimable
2537 static bool can_steal_fallback(unsigned int order, int start_mt)
2540 * Leaving this order check is intended, although there is
2541 * relaxed order check in next check. The reason is that
2542 * we can actually steal whole pageblock if this condition met,
2543 * but, below check doesn't guarantee it and that is just heuristic
2544 * so could be changed anytime.
2546 if (order >= pageblock_order)
2549 if (order >= pageblock_order / 2 ||
2550 start_mt == MIGRATE_RECLAIMABLE ||
2551 start_mt == MIGRATE_UNMOVABLE ||
2552 page_group_by_mobility_disabled)
2558 static inline bool boost_watermark(struct zone *zone)
2560 unsigned long max_boost;
2562 if (!watermark_boost_factor)
2565 * Don't bother in zones that are unlikely to produce results.
2566 * On small machines, including kdump capture kernels running
2567 * in a small area, boosting the watermark can cause an out of
2568 * memory situation immediately.
2570 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2573 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2574 watermark_boost_factor, 10000);
2577 * high watermark may be uninitialised if fragmentation occurs
2578 * very early in boot so do not boost. We do not fall
2579 * through and boost by pageblock_nr_pages as failing
2580 * allocations that early means that reclaim is not going
2581 * to help and it may even be impossible to reclaim the
2582 * boosted watermark resulting in a hang.
2587 max_boost = max(pageblock_nr_pages, max_boost);
2589 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2596 * This function implements actual steal behaviour. If order is large enough,
2597 * we can steal whole pageblock. If not, we first move freepages in this
2598 * pageblock to our migratetype and determine how many already-allocated pages
2599 * are there in the pageblock with a compatible migratetype. If at least half
2600 * of pages are free or compatible, we can change migratetype of the pageblock
2601 * itself, so pages freed in the future will be put on the correct free list.
2603 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2604 unsigned int alloc_flags, int start_type, bool whole_block)
2606 unsigned int current_order = buddy_order(page);
2607 int free_pages, movable_pages, alike_pages;
2610 old_block_type = get_pageblock_migratetype(page);
2613 * This can happen due to races and we want to prevent broken
2614 * highatomic accounting.
2616 if (is_migrate_highatomic(old_block_type))
2619 /* Take ownership for orders >= pageblock_order */
2620 if (current_order >= pageblock_order) {
2621 change_pageblock_range(page, current_order, start_type);
2626 * Boost watermarks to increase reclaim pressure to reduce the
2627 * likelihood of future fallbacks. Wake kswapd now as the node
2628 * may be balanced overall and kswapd will not wake naturally.
2630 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2631 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2633 /* We are not allowed to try stealing from the whole block */
2637 free_pages = move_freepages_block(zone, page, start_type,
2640 * Determine how many pages are compatible with our allocation.
2641 * For movable allocation, it's the number of movable pages which
2642 * we just obtained. For other types it's a bit more tricky.
2644 if (start_type == MIGRATE_MOVABLE) {
2645 alike_pages = movable_pages;
2648 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2649 * to MOVABLE pageblock, consider all non-movable pages as
2650 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2651 * vice versa, be conservative since we can't distinguish the
2652 * exact migratetype of non-movable pages.
2654 if (old_block_type == MIGRATE_MOVABLE)
2655 alike_pages = pageblock_nr_pages
2656 - (free_pages + movable_pages);
2661 /* moving whole block can fail due to zone boundary conditions */
2666 * If a sufficient number of pages in the block are either free or of
2667 * comparable migratability as our allocation, claim the whole block.
2669 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2670 page_group_by_mobility_disabled)
2671 set_pageblock_migratetype(page, start_type);
2676 move_to_free_list(page, zone, current_order, start_type);
2680 * Check whether there is a suitable fallback freepage with requested order.
2681 * If only_stealable is true, this function returns fallback_mt only if
2682 * we can steal other freepages all together. This would help to reduce
2683 * fragmentation due to mixed migratetype pages in one pageblock.
2685 int find_suitable_fallback(struct free_area *area, unsigned int order,
2686 int migratetype, bool only_stealable, bool *can_steal)
2691 if (area->nr_free == 0)
2696 fallback_mt = fallbacks[migratetype][i];
2697 if (fallback_mt == MIGRATE_TYPES)
2700 if (free_area_empty(area, fallback_mt))
2703 if (can_steal_fallback(order, migratetype))
2706 if (!only_stealable)
2717 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2718 * there are no empty page blocks that contain a page with a suitable order
2720 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2721 unsigned int alloc_order)
2724 unsigned long max_managed, flags;
2727 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2728 * Check is race-prone but harmless.
2730 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2731 if (zone->nr_reserved_highatomic >= max_managed)
2734 spin_lock_irqsave(&zone->lock, flags);
2736 /* Recheck the nr_reserved_highatomic limit under the lock */
2737 if (zone->nr_reserved_highatomic >= max_managed)
2741 mt = get_pageblock_migratetype(page);
2742 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2743 if (migratetype_is_mergeable(mt)) {
2744 zone->nr_reserved_highatomic += pageblock_nr_pages;
2745 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2746 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2750 spin_unlock_irqrestore(&zone->lock, flags);
2754 * Used when an allocation is about to fail under memory pressure. This
2755 * potentially hurts the reliability of high-order allocations when under
2756 * intense memory pressure but failed atomic allocations should be easier
2757 * to recover from than an OOM.
2759 * If @force is true, try to unreserve a pageblock even though highatomic
2760 * pageblock is exhausted.
2762 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2765 struct zonelist *zonelist = ac->zonelist;
2766 unsigned long flags;
2773 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2776 * Preserve at least one pageblock unless memory pressure
2779 if (!force && zone->nr_reserved_highatomic <=
2783 spin_lock_irqsave(&zone->lock, flags);
2784 for (order = 0; order < MAX_ORDER; order++) {
2785 struct free_area *area = &(zone->free_area[order]);
2787 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2792 * In page freeing path, migratetype change is racy so
2793 * we can counter several free pages in a pageblock
2794 * in this loop although we changed the pageblock type
2795 * from highatomic to ac->migratetype. So we should
2796 * adjust the count once.
2798 if (is_migrate_highatomic_page(page)) {
2800 * It should never happen but changes to
2801 * locking could inadvertently allow a per-cpu
2802 * drain to add pages to MIGRATE_HIGHATOMIC
2803 * while unreserving so be safe and watch for
2806 zone->nr_reserved_highatomic -= min(
2808 zone->nr_reserved_highatomic);
2812 * Convert to ac->migratetype and avoid the normal
2813 * pageblock stealing heuristics. Minimally, the caller
2814 * is doing the work and needs the pages. More
2815 * importantly, if the block was always converted to
2816 * MIGRATE_UNMOVABLE or another type then the number
2817 * of pageblocks that cannot be completely freed
2820 set_pageblock_migratetype(page, ac->migratetype);
2821 ret = move_freepages_block(zone, page, ac->migratetype,
2824 spin_unlock_irqrestore(&zone->lock, flags);
2828 spin_unlock_irqrestore(&zone->lock, flags);
2835 * Try finding a free buddy page on the fallback list and put it on the free
2836 * list of requested migratetype, possibly along with other pages from the same
2837 * block, depending on fragmentation avoidance heuristics. Returns true if
2838 * fallback was found so that __rmqueue_smallest() can grab it.
2840 * The use of signed ints for order and current_order is a deliberate
2841 * deviation from the rest of this file, to make the for loop
2842 * condition simpler.
2844 static __always_inline bool
2845 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2846 unsigned int alloc_flags)
2848 struct free_area *area;
2850 int min_order = order;
2856 * Do not steal pages from freelists belonging to other pageblocks
2857 * i.e. orders < pageblock_order. If there are no local zones free,
2858 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2860 if (alloc_flags & ALLOC_NOFRAGMENT)
2861 min_order = pageblock_order;
2864 * Find the largest available free page in the other list. This roughly
2865 * approximates finding the pageblock with the most free pages, which
2866 * would be too costly to do exactly.
2868 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2870 area = &(zone->free_area[current_order]);
2871 fallback_mt = find_suitable_fallback(area, current_order,
2872 start_migratetype, false, &can_steal);
2873 if (fallback_mt == -1)
2877 * We cannot steal all free pages from the pageblock and the
2878 * requested migratetype is movable. In that case it's better to
2879 * steal and split the smallest available page instead of the
2880 * largest available page, because even if the next movable
2881 * allocation falls back into a different pageblock than this
2882 * one, it won't cause permanent fragmentation.
2884 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2885 && current_order > order)
2894 for (current_order = order; current_order < MAX_ORDER;
2896 area = &(zone->free_area[current_order]);
2897 fallback_mt = find_suitable_fallback(area, current_order,
2898 start_migratetype, false, &can_steal);
2899 if (fallback_mt != -1)
2904 * This should not happen - we already found a suitable fallback
2905 * when looking for the largest page.
2907 VM_BUG_ON(current_order == MAX_ORDER);
2910 page = get_page_from_free_area(area, fallback_mt);
2912 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2915 trace_mm_page_alloc_extfrag(page, order, current_order,
2916 start_migratetype, fallback_mt);
2923 * Do the hard work of removing an element from the buddy allocator.
2924 * Call me with the zone->lock already held.
2926 static __always_inline struct page *
2927 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2928 unsigned int alloc_flags)
2932 if (IS_ENABLED(CONFIG_CMA)) {
2934 * Balance movable allocations between regular and CMA areas by
2935 * allocating from CMA when over half of the zone's free memory
2936 * is in the CMA area.
2938 if (alloc_flags & ALLOC_CMA &&
2939 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2940 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2941 page = __rmqueue_cma_fallback(zone, order);
2947 page = __rmqueue_smallest(zone, order, migratetype);
2948 if (unlikely(!page)) {
2949 if (alloc_flags & ALLOC_CMA)
2950 page = __rmqueue_cma_fallback(zone, order);
2952 if (!page && __rmqueue_fallback(zone, order, migratetype,
2958 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2963 * Obtain a specified number of elements from the buddy allocator, all under
2964 * a single hold of the lock, for efficiency. Add them to the supplied list.
2965 * Returns the number of new pages which were placed at *list.
2967 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2968 unsigned long count, struct list_head *list,
2969 int migratetype, unsigned int alloc_flags)
2971 int i, allocated = 0;
2974 * local_lock_irq held so equivalent to spin_lock_irqsave for
2975 * both PREEMPT_RT and non-PREEMPT_RT configurations.
2977 spin_lock(&zone->lock);
2978 for (i = 0; i < count; ++i) {
2979 struct page *page = __rmqueue(zone, order, migratetype,
2981 if (unlikely(page == NULL))
2984 if (unlikely(check_pcp_refill(page, order)))
2988 * Split buddy pages returned by expand() are received here in
2989 * physical page order. The page is added to the tail of
2990 * caller's list. From the callers perspective, the linked list
2991 * is ordered by page number under some conditions. This is
2992 * useful for IO devices that can forward direction from the
2993 * head, thus also in the physical page order. This is useful
2994 * for IO devices that can merge IO requests if the physical
2995 * pages are ordered properly.
2997 list_add_tail(&page->lru, list);
2999 if (is_migrate_cma(get_pcppage_migratetype(page)))
3000 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3005 * i pages were removed from the buddy list even if some leak due
3006 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3007 * on i. Do not confuse with 'allocated' which is the number of
3008 * pages added to the pcp list.
3010 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3011 spin_unlock(&zone->lock);
3017 * Called from the vmstat counter updater to drain pagesets of this
3018 * currently executing processor on remote nodes after they have
3021 * Note that this function must be called with the thread pinned to
3022 * a single processor.
3024 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3026 unsigned long flags;
3027 int to_drain, batch;
3029 local_lock_irqsave(&pagesets.lock, flags);
3030 batch = READ_ONCE(pcp->batch);
3031 to_drain = min(pcp->count, batch);
3033 free_pcppages_bulk(zone, to_drain, pcp, 0);
3034 local_unlock_irqrestore(&pagesets.lock, flags);
3039 * Drain pcplists of the indicated processor and zone.
3041 * The processor must either be the current processor and the
3042 * thread pinned to the current processor or a processor that
3045 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3047 unsigned long flags;
3048 struct per_cpu_pages *pcp;
3050 local_lock_irqsave(&pagesets.lock, flags);
3052 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3054 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3056 local_unlock_irqrestore(&pagesets.lock, flags);
3060 * Drain pcplists of all zones on the indicated processor.
3062 * The processor must either be the current processor and the
3063 * thread pinned to the current processor or a processor that
3066 static void drain_pages(unsigned int cpu)
3070 for_each_populated_zone(zone) {
3071 drain_pages_zone(cpu, zone);
3076 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3078 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3079 * the single zone's pages.
3081 void drain_local_pages(struct zone *zone)
3083 int cpu = smp_processor_id();
3086 drain_pages_zone(cpu, zone);
3091 static void drain_local_pages_wq(struct work_struct *work)
3093 struct pcpu_drain *drain;
3095 drain = container_of(work, struct pcpu_drain, work);
3098 * drain_all_pages doesn't use proper cpu hotplug protection so
3099 * we can race with cpu offline when the WQ can move this from
3100 * a cpu pinned worker to an unbound one. We can operate on a different
3101 * cpu which is alright but we also have to make sure to not move to
3105 drain_local_pages(drain->zone);
3110 * The implementation of drain_all_pages(), exposing an extra parameter to
3111 * drain on all cpus.
3113 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3114 * not empty. The check for non-emptiness can however race with a free to
3115 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3116 * that need the guarantee that every CPU has drained can disable the
3117 * optimizing racy check.
3119 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3124 * Allocate in the BSS so we won't require allocation in
3125 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3127 static cpumask_t cpus_with_pcps;
3130 * Make sure nobody triggers this path before mm_percpu_wq is fully
3133 if (WARN_ON_ONCE(!mm_percpu_wq))
3137 * Do not drain if one is already in progress unless it's specific to
3138 * a zone. Such callers are primarily CMA and memory hotplug and need
3139 * the drain to be complete when the call returns.
3141 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3144 mutex_lock(&pcpu_drain_mutex);
3148 * We don't care about racing with CPU hotplug event
3149 * as offline notification will cause the notified
3150 * cpu to drain that CPU pcps and on_each_cpu_mask
3151 * disables preemption as part of its processing
3153 for_each_online_cpu(cpu) {
3154 struct per_cpu_pages *pcp;
3156 bool has_pcps = false;
3158 if (force_all_cpus) {
3160 * The pcp.count check is racy, some callers need a
3161 * guarantee that no cpu is missed.
3165 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3169 for_each_populated_zone(z) {
3170 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3179 cpumask_set_cpu(cpu, &cpus_with_pcps);
3181 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3184 for_each_cpu(cpu, &cpus_with_pcps) {
3185 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3188 INIT_WORK(&drain->work, drain_local_pages_wq);
3189 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3191 for_each_cpu(cpu, &cpus_with_pcps)
3192 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3194 mutex_unlock(&pcpu_drain_mutex);
3198 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3200 * When zone parameter is non-NULL, spill just the single zone's pages.
3202 * Note that this can be extremely slow as the draining happens in a workqueue.
3204 void drain_all_pages(struct zone *zone)
3206 __drain_all_pages(zone, false);
3209 #ifdef CONFIG_HIBERNATION
3212 * Touch the watchdog for every WD_PAGE_COUNT pages.
3214 #define WD_PAGE_COUNT (128*1024)
3216 void mark_free_pages(struct zone *zone)
3218 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3219 unsigned long flags;
3220 unsigned int order, t;
3223 if (zone_is_empty(zone))
3226 spin_lock_irqsave(&zone->lock, flags);
3228 max_zone_pfn = zone_end_pfn(zone);
3229 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3230 if (pfn_valid(pfn)) {
3231 page = pfn_to_page(pfn);
3233 if (!--page_count) {
3234 touch_nmi_watchdog();
3235 page_count = WD_PAGE_COUNT;
3238 if (page_zone(page) != zone)
3241 if (!swsusp_page_is_forbidden(page))
3242 swsusp_unset_page_free(page);
3245 for_each_migratetype_order(order, t) {
3246 list_for_each_entry(page,
3247 &zone->free_area[order].free_list[t], lru) {
3250 pfn = page_to_pfn(page);
3251 for (i = 0; i < (1UL << order); i++) {
3252 if (!--page_count) {
3253 touch_nmi_watchdog();
3254 page_count = WD_PAGE_COUNT;
3256 swsusp_set_page_free(pfn_to_page(pfn + i));
3260 spin_unlock_irqrestore(&zone->lock, flags);
3262 #endif /* CONFIG_PM */
3264 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3269 if (!free_pcp_prepare(page, order))
3272 migratetype = get_pfnblock_migratetype(page, pfn);
3273 set_pcppage_migratetype(page, migratetype);
3277 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3280 int min_nr_free, max_nr_free;
3282 /* Free everything if batch freeing high-order pages. */
3283 if (unlikely(free_high))
3286 /* Check for PCP disabled or boot pageset */
3287 if (unlikely(high < batch))
3290 /* Leave at least pcp->batch pages on the list */
3291 min_nr_free = batch;
3292 max_nr_free = high - batch;
3295 * Double the number of pages freed each time there is subsequent
3296 * freeing of pages without any allocation.
3298 batch <<= pcp->free_factor;
3299 if (batch < max_nr_free)
3301 batch = clamp(batch, min_nr_free, max_nr_free);
3306 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3309 int high = READ_ONCE(pcp->high);
3311 if (unlikely(!high || free_high))
3314 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3318 * If reclaim is active, limit the number of pages that can be
3319 * stored on pcp lists
3321 return min(READ_ONCE(pcp->batch) << 2, high);
3324 static void free_unref_page_commit(struct page *page, int migratetype,
3327 struct zone *zone = page_zone(page);
3328 struct per_cpu_pages *pcp;
3333 __count_vm_event(PGFREE);
3334 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3335 pindex = order_to_pindex(migratetype, order);
3336 list_add(&page->lru, &pcp->lists[pindex]);
3337 pcp->count += 1 << order;
3340 * As high-order pages other than THP's stored on PCP can contribute
3341 * to fragmentation, limit the number stored when PCP is heavily
3342 * freeing without allocation. The remainder after bulk freeing
3343 * stops will be drained from vmstat refresh context.
3345 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3347 high = nr_pcp_high(pcp, zone, free_high);
3348 if (pcp->count >= high) {
3349 int batch = READ_ONCE(pcp->batch);
3351 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3358 void free_unref_page(struct page *page, unsigned int order)
3360 unsigned long flags;
3361 unsigned long pfn = page_to_pfn(page);
3364 if (!free_unref_page_prepare(page, pfn, order))
3368 * We only track unmovable, reclaimable and movable on pcp lists.
3369 * Place ISOLATE pages on the isolated list because they are being
3370 * offlined but treat HIGHATOMIC as movable pages so we can get those
3371 * areas back if necessary. Otherwise, we may have to free
3372 * excessively into the page allocator
3374 migratetype = get_pcppage_migratetype(page);
3375 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3376 if (unlikely(is_migrate_isolate(migratetype))) {
3377 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3380 migratetype = MIGRATE_MOVABLE;
3383 local_lock_irqsave(&pagesets.lock, flags);
3384 free_unref_page_commit(page, migratetype, order);
3385 local_unlock_irqrestore(&pagesets.lock, flags);
3389 * Free a list of 0-order pages
3391 void free_unref_page_list(struct list_head *list)
3393 struct page *page, *next;
3394 unsigned long flags;
3395 int batch_count = 0;
3398 /* Prepare pages for freeing */
3399 list_for_each_entry_safe(page, next, list, lru) {
3400 unsigned long pfn = page_to_pfn(page);
3401 if (!free_unref_page_prepare(page, pfn, 0)) {
3402 list_del(&page->lru);
3407 * Free isolated pages directly to the allocator, see
3408 * comment in free_unref_page.
3410 migratetype = get_pcppage_migratetype(page);
3411 if (unlikely(is_migrate_isolate(migratetype))) {
3412 list_del(&page->lru);
3413 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3418 local_lock_irqsave(&pagesets.lock, flags);
3419 list_for_each_entry_safe(page, next, list, lru) {
3421 * Non-isolated types over MIGRATE_PCPTYPES get added
3422 * to the MIGRATE_MOVABLE pcp list.
3424 migratetype = get_pcppage_migratetype(page);
3425 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3426 migratetype = MIGRATE_MOVABLE;
3428 trace_mm_page_free_batched(page);
3429 free_unref_page_commit(page, migratetype, 0);
3432 * Guard against excessive IRQ disabled times when we get
3433 * a large list of pages to free.
3435 if (++batch_count == SWAP_CLUSTER_MAX) {
3436 local_unlock_irqrestore(&pagesets.lock, flags);
3438 local_lock_irqsave(&pagesets.lock, flags);
3441 local_unlock_irqrestore(&pagesets.lock, flags);
3445 * split_page takes a non-compound higher-order page, and splits it into
3446 * n (1<<order) sub-pages: page[0..n]
3447 * Each sub-page must be freed individually.
3449 * Note: this is probably too low level an operation for use in drivers.
3450 * Please consult with lkml before using this in your driver.
3452 void split_page(struct page *page, unsigned int order)
3456 VM_BUG_ON_PAGE(PageCompound(page), page);
3457 VM_BUG_ON_PAGE(!page_count(page), page);
3459 for (i = 1; i < (1 << order); i++)
3460 set_page_refcounted(page + i);
3461 split_page_owner(page, 1 << order);
3462 split_page_memcg(page, 1 << order);
3464 EXPORT_SYMBOL_GPL(split_page);
3466 int __isolate_free_page(struct page *page, unsigned int order)
3468 unsigned long watermark;
3472 BUG_ON(!PageBuddy(page));
3474 zone = page_zone(page);
3475 mt = get_pageblock_migratetype(page);
3477 if (!is_migrate_isolate(mt)) {
3479 * Obey watermarks as if the page was being allocated. We can
3480 * emulate a high-order watermark check with a raised order-0
3481 * watermark, because we already know our high-order page
3484 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3485 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3488 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3491 /* Remove page from free list */
3493 del_page_from_free_list(page, zone, order);
3496 * Set the pageblock if the isolated page is at least half of a
3499 if (order >= pageblock_order - 1) {
3500 struct page *endpage = page + (1 << order) - 1;
3501 for (; page < endpage; page += pageblock_nr_pages) {
3502 int mt = get_pageblock_migratetype(page);
3504 * Only change normal pageblocks (i.e., they can merge
3507 if (migratetype_is_mergeable(mt))
3508 set_pageblock_migratetype(page,
3514 return 1UL << order;
3518 * __putback_isolated_page - Return a now-isolated page back where we got it
3519 * @page: Page that was isolated
3520 * @order: Order of the isolated page
3521 * @mt: The page's pageblock's migratetype
3523 * This function is meant to return a page pulled from the free lists via
3524 * __isolate_free_page back to the free lists they were pulled from.
3526 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3528 struct zone *zone = page_zone(page);
3530 /* zone lock should be held when this function is called */
3531 lockdep_assert_held(&zone->lock);
3533 /* Return isolated page to tail of freelist. */
3534 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3535 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3539 * Update NUMA hit/miss statistics
3541 * Must be called with interrupts disabled.
3543 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3547 enum numa_stat_item local_stat = NUMA_LOCAL;
3549 /* skip numa counters update if numa stats is disabled */
3550 if (!static_branch_likely(&vm_numa_stat_key))
3553 if (zone_to_nid(z) != numa_node_id())
3554 local_stat = NUMA_OTHER;
3556 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3557 __count_numa_events(z, NUMA_HIT, nr_account);
3559 __count_numa_events(z, NUMA_MISS, nr_account);
3560 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3562 __count_numa_events(z, local_stat, nr_account);
3566 /* Remove page from the per-cpu list, caller must protect the list */
3568 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3570 unsigned int alloc_flags,
3571 struct per_cpu_pages *pcp,
3572 struct list_head *list)
3577 if (list_empty(list)) {
3578 int batch = READ_ONCE(pcp->batch);
3582 * Scale batch relative to order if batch implies
3583 * free pages can be stored on the PCP. Batch can
3584 * be 1 for small zones or for boot pagesets which
3585 * should never store free pages as the pages may
3586 * belong to arbitrary zones.
3589 batch = max(batch >> order, 2);
3590 alloced = rmqueue_bulk(zone, order,
3592 migratetype, alloc_flags);
3594 pcp->count += alloced << order;
3595 if (unlikely(list_empty(list)))
3599 page = list_first_entry(list, struct page, lru);
3600 list_del(&page->lru);
3601 pcp->count -= 1 << order;
3602 } while (check_new_pcp(page, order));
3607 /* Lock and remove page from the per-cpu list */
3608 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3609 struct zone *zone, unsigned int order,
3610 gfp_t gfp_flags, int migratetype,
3611 unsigned int alloc_flags)
3613 struct per_cpu_pages *pcp;
3614 struct list_head *list;
3616 unsigned long flags;
3618 local_lock_irqsave(&pagesets.lock, flags);
3621 * On allocation, reduce the number of pages that are batch freed.
3622 * See nr_pcp_free() where free_factor is increased for subsequent
3625 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3626 pcp->free_factor >>= 1;
3627 list = &pcp->lists[order_to_pindex(migratetype, order)];
3628 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3629 local_unlock_irqrestore(&pagesets.lock, flags);
3631 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3632 zone_statistics(preferred_zone, zone, 1);
3638 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3641 struct page *rmqueue(struct zone *preferred_zone,
3642 struct zone *zone, unsigned int order,
3643 gfp_t gfp_flags, unsigned int alloc_flags,
3646 unsigned long flags;
3649 if (likely(pcp_allowed_order(order))) {
3651 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3652 * we need to skip it when CMA area isn't allowed.
3654 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3655 migratetype != MIGRATE_MOVABLE) {
3656 page = rmqueue_pcplist(preferred_zone, zone, order,
3657 gfp_flags, migratetype, alloc_flags);
3663 * We most definitely don't want callers attempting to
3664 * allocate greater than order-1 page units with __GFP_NOFAIL.
3666 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3670 spin_lock_irqsave(&zone->lock, flags);
3672 * order-0 request can reach here when the pcplist is skipped
3673 * due to non-CMA allocation context. HIGHATOMIC area is
3674 * reserved for high-order atomic allocation, so order-0
3675 * request should skip it.
3677 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3678 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3680 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3683 page = __rmqueue(zone, order, migratetype, alloc_flags);
3687 __mod_zone_freepage_state(zone, -(1 << order),
3688 get_pcppage_migratetype(page));
3689 spin_unlock_irqrestore(&zone->lock, flags);
3690 } while (check_new_pages(page, order));
3692 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3693 zone_statistics(preferred_zone, zone, 1);
3696 /* Separate test+clear to avoid unnecessary atomics */
3697 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3698 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3699 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3702 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3706 spin_unlock_irqrestore(&zone->lock, flags);
3710 #ifdef CONFIG_FAIL_PAGE_ALLOC
3713 struct fault_attr attr;
3715 bool ignore_gfp_highmem;
3716 bool ignore_gfp_reclaim;
3718 } fail_page_alloc = {
3719 .attr = FAULT_ATTR_INITIALIZER,
3720 .ignore_gfp_reclaim = true,
3721 .ignore_gfp_highmem = true,
3725 static int __init setup_fail_page_alloc(char *str)
3727 return setup_fault_attr(&fail_page_alloc.attr, str);
3729 __setup("fail_page_alloc=", setup_fail_page_alloc);
3731 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3733 if (order < fail_page_alloc.min_order)
3735 if (gfp_mask & __GFP_NOFAIL)
3737 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3739 if (fail_page_alloc.ignore_gfp_reclaim &&
3740 (gfp_mask & __GFP_DIRECT_RECLAIM))
3743 return should_fail(&fail_page_alloc.attr, 1 << order);
3746 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3748 static int __init fail_page_alloc_debugfs(void)
3750 umode_t mode = S_IFREG | 0600;
3753 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3754 &fail_page_alloc.attr);
3756 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3757 &fail_page_alloc.ignore_gfp_reclaim);
3758 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3759 &fail_page_alloc.ignore_gfp_highmem);
3760 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3765 late_initcall(fail_page_alloc_debugfs);
3767 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3769 #else /* CONFIG_FAIL_PAGE_ALLOC */
3771 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3776 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3778 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3780 return __should_fail_alloc_page(gfp_mask, order);
3782 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3784 static inline long __zone_watermark_unusable_free(struct zone *z,
3785 unsigned int order, unsigned int alloc_flags)
3787 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3788 long unusable_free = (1 << order) - 1;
3791 * If the caller does not have rights to ALLOC_HARDER then subtract
3792 * the high-atomic reserves. This will over-estimate the size of the
3793 * atomic reserve but it avoids a search.
3795 if (likely(!alloc_harder))
3796 unusable_free += z->nr_reserved_highatomic;
3799 /* If allocation can't use CMA areas don't use free CMA pages */
3800 if (!(alloc_flags & ALLOC_CMA))
3801 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3804 return unusable_free;
3808 * Return true if free base pages are above 'mark'. For high-order checks it
3809 * will return true of the order-0 watermark is reached and there is at least
3810 * one free page of a suitable size. Checking now avoids taking the zone lock
3811 * to check in the allocation paths if no pages are free.
3813 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3814 int highest_zoneidx, unsigned int alloc_flags,
3819 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3821 /* free_pages may go negative - that's OK */
3822 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3824 if (alloc_flags & ALLOC_HIGH)
3827 if (unlikely(alloc_harder)) {
3829 * OOM victims can try even harder than normal ALLOC_HARDER
3830 * users on the grounds that it's definitely going to be in
3831 * the exit path shortly and free memory. Any allocation it
3832 * makes during the free path will be small and short-lived.
3834 if (alloc_flags & ALLOC_OOM)
3841 * Check watermarks for an order-0 allocation request. If these
3842 * are not met, then a high-order request also cannot go ahead
3843 * even if a suitable page happened to be free.
3845 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3848 /* If this is an order-0 request then the watermark is fine */
3852 /* For a high-order request, check at least one suitable page is free */
3853 for (o = order; o < MAX_ORDER; o++) {
3854 struct free_area *area = &z->free_area[o];
3860 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3861 if (!free_area_empty(area, mt))
3866 if ((alloc_flags & ALLOC_CMA) &&
3867 !free_area_empty(area, MIGRATE_CMA)) {
3871 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3877 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3878 int highest_zoneidx, unsigned int alloc_flags)
3880 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3881 zone_page_state(z, NR_FREE_PAGES));
3884 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3885 unsigned long mark, int highest_zoneidx,
3886 unsigned int alloc_flags, gfp_t gfp_mask)
3890 free_pages = zone_page_state(z, NR_FREE_PAGES);
3893 * Fast check for order-0 only. If this fails then the reserves
3894 * need to be calculated.
3899 fast_free = free_pages;
3900 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3901 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3905 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3909 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3910 * when checking the min watermark. The min watermark is the
3911 * point where boosting is ignored so that kswapd is woken up
3912 * when below the low watermark.
3914 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3915 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3916 mark = z->_watermark[WMARK_MIN];
3917 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3918 alloc_flags, free_pages);
3924 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3925 unsigned long mark, int highest_zoneidx)
3927 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3929 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3930 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3932 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3937 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3939 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3941 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3942 node_reclaim_distance;
3944 #else /* CONFIG_NUMA */
3945 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3949 #endif /* CONFIG_NUMA */
3952 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3953 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3954 * premature use of a lower zone may cause lowmem pressure problems that
3955 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3956 * probably too small. It only makes sense to spread allocations to avoid
3957 * fragmentation between the Normal and DMA32 zones.
3959 static inline unsigned int
3960 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3962 unsigned int alloc_flags;
3965 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3968 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3970 #ifdef CONFIG_ZONE_DMA32
3974 if (zone_idx(zone) != ZONE_NORMAL)
3978 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3979 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3980 * on UMA that if Normal is populated then so is DMA32.
3982 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3983 if (nr_online_nodes > 1 && !populated_zone(--zone))
3986 alloc_flags |= ALLOC_NOFRAGMENT;
3987 #endif /* CONFIG_ZONE_DMA32 */
3991 /* Must be called after current_gfp_context() which can change gfp_mask */
3992 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3993 unsigned int alloc_flags)
3996 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3997 alloc_flags |= ALLOC_CMA;
4003 * get_page_from_freelist goes through the zonelist trying to allocate
4006 static struct page *
4007 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4008 const struct alloc_context *ac)
4012 struct pglist_data *last_pgdat_dirty_limit = NULL;
4017 * Scan zonelist, looking for a zone with enough free.
4018 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4020 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4021 z = ac->preferred_zoneref;
4022 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4027 if (cpusets_enabled() &&
4028 (alloc_flags & ALLOC_CPUSET) &&
4029 !__cpuset_zone_allowed(zone, gfp_mask))
4032 * When allocating a page cache page for writing, we
4033 * want to get it from a node that is within its dirty
4034 * limit, such that no single node holds more than its
4035 * proportional share of globally allowed dirty pages.
4036 * The dirty limits take into account the node's
4037 * lowmem reserves and high watermark so that kswapd
4038 * should be able to balance it without having to
4039 * write pages from its LRU list.
4041 * XXX: For now, allow allocations to potentially
4042 * exceed the per-node dirty limit in the slowpath
4043 * (spread_dirty_pages unset) before going into reclaim,
4044 * which is important when on a NUMA setup the allowed
4045 * nodes are together not big enough to reach the
4046 * global limit. The proper fix for these situations
4047 * will require awareness of nodes in the
4048 * dirty-throttling and the flusher threads.
4050 if (ac->spread_dirty_pages) {
4051 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4054 if (!node_dirty_ok(zone->zone_pgdat)) {
4055 last_pgdat_dirty_limit = zone->zone_pgdat;
4060 if (no_fallback && nr_online_nodes > 1 &&
4061 zone != ac->preferred_zoneref->zone) {
4065 * If moving to a remote node, retry but allow
4066 * fragmenting fallbacks. Locality is more important
4067 * than fragmentation avoidance.
4069 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4070 if (zone_to_nid(zone) != local_nid) {
4071 alloc_flags &= ~ALLOC_NOFRAGMENT;
4076 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4077 if (!zone_watermark_fast(zone, order, mark,
4078 ac->highest_zoneidx, alloc_flags,
4082 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4084 * Watermark failed for this zone, but see if we can
4085 * grow this zone if it contains deferred pages.
4087 if (static_branch_unlikely(&deferred_pages)) {
4088 if (_deferred_grow_zone(zone, order))
4092 /* Checked here to keep the fast path fast */
4093 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4094 if (alloc_flags & ALLOC_NO_WATERMARKS)
4097 if (!node_reclaim_enabled() ||
4098 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4101 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4103 case NODE_RECLAIM_NOSCAN:
4106 case NODE_RECLAIM_FULL:
4107 /* scanned but unreclaimable */
4110 /* did we reclaim enough */
4111 if (zone_watermark_ok(zone, order, mark,
4112 ac->highest_zoneidx, alloc_flags))
4120 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4121 gfp_mask, alloc_flags, ac->migratetype);
4123 prep_new_page(page, order, gfp_mask, alloc_flags);
4126 * If this is a high-order atomic allocation then check
4127 * if the pageblock should be reserved for the future
4129 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4130 reserve_highatomic_pageblock(page, zone, order);
4134 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4135 /* Try again if zone has deferred pages */
4136 if (static_branch_unlikely(&deferred_pages)) {
4137 if (_deferred_grow_zone(zone, order))
4145 * It's possible on a UMA machine to get through all zones that are
4146 * fragmented. If avoiding fragmentation, reset and try again.
4149 alloc_flags &= ~ALLOC_NOFRAGMENT;
4156 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4158 unsigned int filter = SHOW_MEM_FILTER_NODES;
4161 * This documents exceptions given to allocations in certain
4162 * contexts that are allowed to allocate outside current's set
4165 if (!(gfp_mask & __GFP_NOMEMALLOC))
4166 if (tsk_is_oom_victim(current) ||
4167 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4168 filter &= ~SHOW_MEM_FILTER_NODES;
4169 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4170 filter &= ~SHOW_MEM_FILTER_NODES;
4172 show_mem(filter, nodemask);
4175 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4177 struct va_format vaf;
4179 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4181 if ((gfp_mask & __GFP_NOWARN) ||
4182 !__ratelimit(&nopage_rs) ||
4183 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4186 va_start(args, fmt);
4189 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4190 current->comm, &vaf, gfp_mask, &gfp_mask,
4191 nodemask_pr_args(nodemask));
4194 cpuset_print_current_mems_allowed();
4197 warn_alloc_show_mem(gfp_mask, nodemask);
4200 static inline struct page *
4201 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4202 unsigned int alloc_flags,
4203 const struct alloc_context *ac)
4207 page = get_page_from_freelist(gfp_mask, order,
4208 alloc_flags|ALLOC_CPUSET, ac);
4210 * fallback to ignore cpuset restriction if our nodes
4214 page = get_page_from_freelist(gfp_mask, order,
4220 static inline struct page *
4221 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4222 const struct alloc_context *ac, unsigned long *did_some_progress)
4224 struct oom_control oc = {
4225 .zonelist = ac->zonelist,
4226 .nodemask = ac->nodemask,
4228 .gfp_mask = gfp_mask,
4233 *did_some_progress = 0;
4236 * Acquire the oom lock. If that fails, somebody else is
4237 * making progress for us.
4239 if (!mutex_trylock(&oom_lock)) {
4240 *did_some_progress = 1;
4241 schedule_timeout_uninterruptible(1);
4246 * Go through the zonelist yet one more time, keep very high watermark
4247 * here, this is only to catch a parallel oom killing, we must fail if
4248 * we're still under heavy pressure. But make sure that this reclaim
4249 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4250 * allocation which will never fail due to oom_lock already held.
4252 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4253 ~__GFP_DIRECT_RECLAIM, order,
4254 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4258 /* Coredumps can quickly deplete all memory reserves */
4259 if (current->flags & PF_DUMPCORE)
4261 /* The OOM killer will not help higher order allocs */
4262 if (order > PAGE_ALLOC_COSTLY_ORDER)
4265 * We have already exhausted all our reclaim opportunities without any
4266 * success so it is time to admit defeat. We will skip the OOM killer
4267 * because it is very likely that the caller has a more reasonable
4268 * fallback than shooting a random task.
4270 * The OOM killer may not free memory on a specific node.
4272 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4274 /* The OOM killer does not needlessly kill tasks for lowmem */
4275 if (ac->highest_zoneidx < ZONE_NORMAL)
4277 if (pm_suspended_storage())
4280 * XXX: GFP_NOFS allocations should rather fail than rely on
4281 * other request to make a forward progress.
4282 * We are in an unfortunate situation where out_of_memory cannot
4283 * do much for this context but let's try it to at least get
4284 * access to memory reserved if the current task is killed (see
4285 * out_of_memory). Once filesystems are ready to handle allocation
4286 * failures more gracefully we should just bail out here.
4289 /* Exhausted what can be done so it's blame time */
4290 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4291 *did_some_progress = 1;
4294 * Help non-failing allocations by giving them access to memory
4297 if (gfp_mask & __GFP_NOFAIL)
4298 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4299 ALLOC_NO_WATERMARKS, ac);
4302 mutex_unlock(&oom_lock);
4307 * Maximum number of compaction retries with a progress before OOM
4308 * killer is consider as the only way to move forward.
4310 #define MAX_COMPACT_RETRIES 16
4312 #ifdef CONFIG_COMPACTION
4313 /* Try memory compaction for high-order allocations before reclaim */
4314 static struct page *
4315 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4316 unsigned int alloc_flags, const struct alloc_context *ac,
4317 enum compact_priority prio, enum compact_result *compact_result)
4319 struct page *page = NULL;
4320 unsigned long pflags;
4321 unsigned int noreclaim_flag;
4326 psi_memstall_enter(&pflags);
4327 delayacct_compact_start();
4328 noreclaim_flag = memalloc_noreclaim_save();
4330 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4333 memalloc_noreclaim_restore(noreclaim_flag);
4334 psi_memstall_leave(&pflags);
4335 delayacct_compact_end();
4337 if (*compact_result == COMPACT_SKIPPED)
4340 * At least in one zone compaction wasn't deferred or skipped, so let's
4341 * count a compaction stall
4343 count_vm_event(COMPACTSTALL);
4345 /* Prep a captured page if available */
4347 prep_new_page(page, order, gfp_mask, alloc_flags);
4349 /* Try get a page from the freelist if available */
4351 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4354 struct zone *zone = page_zone(page);
4356 zone->compact_blockskip_flush = false;
4357 compaction_defer_reset(zone, order, true);
4358 count_vm_event(COMPACTSUCCESS);
4363 * It's bad if compaction run occurs and fails. The most likely reason
4364 * is that pages exist, but not enough to satisfy watermarks.
4366 count_vm_event(COMPACTFAIL);
4374 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4375 enum compact_result compact_result,
4376 enum compact_priority *compact_priority,
4377 int *compaction_retries)
4379 int max_retries = MAX_COMPACT_RETRIES;
4382 int retries = *compaction_retries;
4383 enum compact_priority priority = *compact_priority;
4388 if (fatal_signal_pending(current))
4391 if (compaction_made_progress(compact_result))
4392 (*compaction_retries)++;
4395 * compaction considers all the zone as desperately out of memory
4396 * so it doesn't really make much sense to retry except when the
4397 * failure could be caused by insufficient priority
4399 if (compaction_failed(compact_result))
4400 goto check_priority;
4403 * compaction was skipped because there are not enough order-0 pages
4404 * to work with, so we retry only if it looks like reclaim can help.
4406 if (compaction_needs_reclaim(compact_result)) {
4407 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4412 * make sure the compaction wasn't deferred or didn't bail out early
4413 * due to locks contention before we declare that we should give up.
4414 * But the next retry should use a higher priority if allowed, so
4415 * we don't just keep bailing out endlessly.
4417 if (compaction_withdrawn(compact_result)) {
4418 goto check_priority;
4422 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4423 * costly ones because they are de facto nofail and invoke OOM
4424 * killer to move on while costly can fail and users are ready
4425 * to cope with that. 1/4 retries is rather arbitrary but we
4426 * would need much more detailed feedback from compaction to
4427 * make a better decision.
4429 if (order > PAGE_ALLOC_COSTLY_ORDER)
4431 if (*compaction_retries <= max_retries) {
4437 * Make sure there are attempts at the highest priority if we exhausted
4438 * all retries or failed at the lower priorities.
4441 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4442 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4444 if (*compact_priority > min_priority) {
4445 (*compact_priority)--;
4446 *compaction_retries = 0;
4450 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4454 static inline struct page *
4455 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4456 unsigned int alloc_flags, const struct alloc_context *ac,
4457 enum compact_priority prio, enum compact_result *compact_result)
4459 *compact_result = COMPACT_SKIPPED;
4464 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4465 enum compact_result compact_result,
4466 enum compact_priority *compact_priority,
4467 int *compaction_retries)
4472 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4476 * There are setups with compaction disabled which would prefer to loop
4477 * inside the allocator rather than hit the oom killer prematurely.
4478 * Let's give them a good hope and keep retrying while the order-0
4479 * watermarks are OK.
4481 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4482 ac->highest_zoneidx, ac->nodemask) {
4483 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4484 ac->highest_zoneidx, alloc_flags))
4489 #endif /* CONFIG_COMPACTION */
4491 #ifdef CONFIG_LOCKDEP
4492 static struct lockdep_map __fs_reclaim_map =
4493 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4495 static bool __need_reclaim(gfp_t gfp_mask)
4497 /* no reclaim without waiting on it */
4498 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4501 /* this guy won't enter reclaim */
4502 if (current->flags & PF_MEMALLOC)
4505 if (gfp_mask & __GFP_NOLOCKDEP)
4511 void __fs_reclaim_acquire(unsigned long ip)
4513 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4516 void __fs_reclaim_release(unsigned long ip)
4518 lock_release(&__fs_reclaim_map, ip);
4521 void fs_reclaim_acquire(gfp_t gfp_mask)
4523 gfp_mask = current_gfp_context(gfp_mask);
4525 if (__need_reclaim(gfp_mask)) {
4526 if (gfp_mask & __GFP_FS)
4527 __fs_reclaim_acquire(_RET_IP_);
4529 #ifdef CONFIG_MMU_NOTIFIER
4530 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4531 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4536 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4538 void fs_reclaim_release(gfp_t gfp_mask)
4540 gfp_mask = current_gfp_context(gfp_mask);
4542 if (__need_reclaim(gfp_mask)) {
4543 if (gfp_mask & __GFP_FS)
4544 __fs_reclaim_release(_RET_IP_);
4547 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4550 /* Perform direct synchronous page reclaim */
4551 static unsigned long
4552 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4553 const struct alloc_context *ac)
4555 unsigned int noreclaim_flag;
4556 unsigned long progress;
4560 /* We now go into synchronous reclaim */
4561 cpuset_memory_pressure_bump();
4562 fs_reclaim_acquire(gfp_mask);
4563 noreclaim_flag = memalloc_noreclaim_save();
4565 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4568 memalloc_noreclaim_restore(noreclaim_flag);
4569 fs_reclaim_release(gfp_mask);
4576 /* The really slow allocator path where we enter direct reclaim */
4577 static inline struct page *
4578 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4579 unsigned int alloc_flags, const struct alloc_context *ac,
4580 unsigned long *did_some_progress)
4582 struct page *page = NULL;
4583 unsigned long pflags;
4584 bool drained = false;
4586 psi_memstall_enter(&pflags);
4587 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4588 if (unlikely(!(*did_some_progress)))
4592 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4595 * If an allocation failed after direct reclaim, it could be because
4596 * pages are pinned on the per-cpu lists or in high alloc reserves.
4597 * Shrink them and try again
4599 if (!page && !drained) {
4600 unreserve_highatomic_pageblock(ac, false);
4601 drain_all_pages(NULL);
4606 psi_memstall_leave(&pflags);
4611 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4612 const struct alloc_context *ac)
4616 pg_data_t *last_pgdat = NULL;
4617 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4619 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4621 if (last_pgdat != zone->zone_pgdat)
4622 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4623 last_pgdat = zone->zone_pgdat;
4627 static inline unsigned int
4628 gfp_to_alloc_flags(gfp_t gfp_mask)
4630 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4633 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4634 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4635 * to save two branches.
4637 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4638 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4641 * The caller may dip into page reserves a bit more if the caller
4642 * cannot run direct reclaim, or if the caller has realtime scheduling
4643 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4644 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4646 alloc_flags |= (__force int)
4647 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4649 if (gfp_mask & __GFP_ATOMIC) {
4651 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4652 * if it can't schedule.
4654 if (!(gfp_mask & __GFP_NOMEMALLOC))
4655 alloc_flags |= ALLOC_HARDER;
4657 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4658 * comment for __cpuset_node_allowed().
4660 alloc_flags &= ~ALLOC_CPUSET;
4661 } else if (unlikely(rt_task(current)) && in_task())
4662 alloc_flags |= ALLOC_HARDER;
4664 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4669 static bool oom_reserves_allowed(struct task_struct *tsk)
4671 if (!tsk_is_oom_victim(tsk))
4675 * !MMU doesn't have oom reaper so give access to memory reserves
4676 * only to the thread with TIF_MEMDIE set
4678 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4685 * Distinguish requests which really need access to full memory
4686 * reserves from oom victims which can live with a portion of it
4688 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4690 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4692 if (gfp_mask & __GFP_MEMALLOC)
4693 return ALLOC_NO_WATERMARKS;
4694 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4695 return ALLOC_NO_WATERMARKS;
4696 if (!in_interrupt()) {
4697 if (current->flags & PF_MEMALLOC)
4698 return ALLOC_NO_WATERMARKS;
4699 else if (oom_reserves_allowed(current))
4706 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4708 return !!__gfp_pfmemalloc_flags(gfp_mask);
4712 * Checks whether it makes sense to retry the reclaim to make a forward progress
4713 * for the given allocation request.
4715 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4716 * without success, or when we couldn't even meet the watermark if we
4717 * reclaimed all remaining pages on the LRU lists.
4719 * Returns true if a retry is viable or false to enter the oom path.
4722 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4723 struct alloc_context *ac, int alloc_flags,
4724 bool did_some_progress, int *no_progress_loops)
4731 * Costly allocations might have made a progress but this doesn't mean
4732 * their order will become available due to high fragmentation so
4733 * always increment the no progress counter for them
4735 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4736 *no_progress_loops = 0;
4738 (*no_progress_loops)++;
4741 * Make sure we converge to OOM if we cannot make any progress
4742 * several times in the row.
4744 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4745 /* Before OOM, exhaust highatomic_reserve */
4746 return unreserve_highatomic_pageblock(ac, true);
4750 * Keep reclaiming pages while there is a chance this will lead
4751 * somewhere. If none of the target zones can satisfy our allocation
4752 * request even if all reclaimable pages are considered then we are
4753 * screwed and have to go OOM.
4755 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4756 ac->highest_zoneidx, ac->nodemask) {
4757 unsigned long available;
4758 unsigned long reclaimable;
4759 unsigned long min_wmark = min_wmark_pages(zone);
4762 available = reclaimable = zone_reclaimable_pages(zone);
4763 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4766 * Would the allocation succeed if we reclaimed all
4767 * reclaimable pages?
4769 wmark = __zone_watermark_ok(zone, order, min_wmark,
4770 ac->highest_zoneidx, alloc_flags, available);
4771 trace_reclaim_retry_zone(z, order, reclaimable,
4772 available, min_wmark, *no_progress_loops, wmark);
4780 * Memory allocation/reclaim might be called from a WQ context and the
4781 * current implementation of the WQ concurrency control doesn't
4782 * recognize that a particular WQ is congested if the worker thread is
4783 * looping without ever sleeping. Therefore we have to do a short sleep
4784 * here rather than calling cond_resched().
4786 if (current->flags & PF_WQ_WORKER)
4787 schedule_timeout_uninterruptible(1);
4794 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4797 * It's possible that cpuset's mems_allowed and the nodemask from
4798 * mempolicy don't intersect. This should be normally dealt with by
4799 * policy_nodemask(), but it's possible to race with cpuset update in
4800 * such a way the check therein was true, and then it became false
4801 * before we got our cpuset_mems_cookie here.
4802 * This assumes that for all allocations, ac->nodemask can come only
4803 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4804 * when it does not intersect with the cpuset restrictions) or the
4805 * caller can deal with a violated nodemask.
4807 if (cpusets_enabled() && ac->nodemask &&
4808 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4809 ac->nodemask = NULL;
4814 * When updating a task's mems_allowed or mempolicy nodemask, it is
4815 * possible to race with parallel threads in such a way that our
4816 * allocation can fail while the mask is being updated. If we are about
4817 * to fail, check if the cpuset changed during allocation and if so,
4820 if (read_mems_allowed_retry(cpuset_mems_cookie))
4826 static inline struct page *
4827 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4828 struct alloc_context *ac)
4830 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4831 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4832 struct page *page = NULL;
4833 unsigned int alloc_flags;
4834 unsigned long did_some_progress;
4835 enum compact_priority compact_priority;
4836 enum compact_result compact_result;
4837 int compaction_retries;
4838 int no_progress_loops;
4839 unsigned int cpuset_mems_cookie;
4843 * We also sanity check to catch abuse of atomic reserves being used by
4844 * callers that are not in atomic context.
4846 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4847 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4848 gfp_mask &= ~__GFP_ATOMIC;
4851 compaction_retries = 0;
4852 no_progress_loops = 0;
4853 compact_priority = DEF_COMPACT_PRIORITY;
4854 cpuset_mems_cookie = read_mems_allowed_begin();
4857 * The fast path uses conservative alloc_flags to succeed only until
4858 * kswapd needs to be woken up, and to avoid the cost of setting up
4859 * alloc_flags precisely. So we do that now.
4861 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4864 * We need to recalculate the starting point for the zonelist iterator
4865 * because we might have used different nodemask in the fast path, or
4866 * there was a cpuset modification and we are retrying - otherwise we
4867 * could end up iterating over non-eligible zones endlessly.
4869 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4870 ac->highest_zoneidx, ac->nodemask);
4871 if (!ac->preferred_zoneref->zone)
4875 * Check for insane configurations where the cpuset doesn't contain
4876 * any suitable zone to satisfy the request - e.g. non-movable
4877 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4879 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4880 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4881 ac->highest_zoneidx,
4882 &cpuset_current_mems_allowed);
4887 if (alloc_flags & ALLOC_KSWAPD)
4888 wake_all_kswapds(order, gfp_mask, ac);
4891 * The adjusted alloc_flags might result in immediate success, so try
4894 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4899 * For costly allocations, try direct compaction first, as it's likely
4900 * that we have enough base pages and don't need to reclaim. For non-
4901 * movable high-order allocations, do that as well, as compaction will
4902 * try prevent permanent fragmentation by migrating from blocks of the
4904 * Don't try this for allocations that are allowed to ignore
4905 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4907 if (can_direct_reclaim &&
4909 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4910 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4911 page = __alloc_pages_direct_compact(gfp_mask, order,
4913 INIT_COMPACT_PRIORITY,
4919 * Checks for costly allocations with __GFP_NORETRY, which
4920 * includes some THP page fault allocations
4922 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4924 * If allocating entire pageblock(s) and compaction
4925 * failed because all zones are below low watermarks
4926 * or is prohibited because it recently failed at this
4927 * order, fail immediately unless the allocator has
4928 * requested compaction and reclaim retry.
4931 * - potentially very expensive because zones are far
4932 * below their low watermarks or this is part of very
4933 * bursty high order allocations,
4934 * - not guaranteed to help because isolate_freepages()
4935 * may not iterate over freed pages as part of its
4937 * - unlikely to make entire pageblocks free on its
4940 if (compact_result == COMPACT_SKIPPED ||
4941 compact_result == COMPACT_DEFERRED)
4945 * Looks like reclaim/compaction is worth trying, but
4946 * sync compaction could be very expensive, so keep
4947 * using async compaction.
4949 compact_priority = INIT_COMPACT_PRIORITY;
4954 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4955 if (alloc_flags & ALLOC_KSWAPD)
4956 wake_all_kswapds(order, gfp_mask, ac);
4958 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4960 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
4963 * Reset the nodemask and zonelist iterators if memory policies can be
4964 * ignored. These allocations are high priority and system rather than
4967 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4968 ac->nodemask = NULL;
4969 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4970 ac->highest_zoneidx, ac->nodemask);
4973 /* Attempt with potentially adjusted zonelist and alloc_flags */
4974 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4978 /* Caller is not willing to reclaim, we can't balance anything */
4979 if (!can_direct_reclaim)
4982 /* Avoid recursion of direct reclaim */
4983 if (current->flags & PF_MEMALLOC)
4986 /* Try direct reclaim and then allocating */
4987 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4988 &did_some_progress);
4992 /* Try direct compaction and then allocating */
4993 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4994 compact_priority, &compact_result);
4998 /* Do not loop if specifically requested */
4999 if (gfp_mask & __GFP_NORETRY)
5003 * Do not retry costly high order allocations unless they are
5004 * __GFP_RETRY_MAYFAIL
5006 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5009 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5010 did_some_progress > 0, &no_progress_loops))
5014 * It doesn't make any sense to retry for the compaction if the order-0
5015 * reclaim is not able to make any progress because the current
5016 * implementation of the compaction depends on the sufficient amount
5017 * of free memory (see __compaction_suitable)
5019 if (did_some_progress > 0 &&
5020 should_compact_retry(ac, order, alloc_flags,
5021 compact_result, &compact_priority,
5022 &compaction_retries))
5026 /* Deal with possible cpuset update races before we start OOM killing */
5027 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5030 /* Reclaim has failed us, start killing things */
5031 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5035 /* Avoid allocations with no watermarks from looping endlessly */
5036 if (tsk_is_oom_victim(current) &&
5037 (alloc_flags & ALLOC_OOM ||
5038 (gfp_mask & __GFP_NOMEMALLOC)))
5041 /* Retry as long as the OOM killer is making progress */
5042 if (did_some_progress) {
5043 no_progress_loops = 0;
5048 /* Deal with possible cpuset update races before we fail */
5049 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5053 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5056 if (gfp_mask & __GFP_NOFAIL) {
5058 * All existing users of the __GFP_NOFAIL are blockable, so warn
5059 * of any new users that actually require GFP_NOWAIT
5061 if (WARN_ON_ONCE(!can_direct_reclaim))
5065 * PF_MEMALLOC request from this context is rather bizarre
5066 * because we cannot reclaim anything and only can loop waiting
5067 * for somebody to do a work for us
5069 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5072 * non failing costly orders are a hard requirement which we
5073 * are not prepared for much so let's warn about these users
5074 * so that we can identify them and convert them to something
5077 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5080 * Help non-failing allocations by giving them access to memory
5081 * reserves but do not use ALLOC_NO_WATERMARKS because this
5082 * could deplete whole memory reserves which would just make
5083 * the situation worse
5085 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5093 warn_alloc(gfp_mask, ac->nodemask,
5094 "page allocation failure: order:%u", order);
5099 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5100 int preferred_nid, nodemask_t *nodemask,
5101 struct alloc_context *ac, gfp_t *alloc_gfp,
5102 unsigned int *alloc_flags)
5104 ac->highest_zoneidx = gfp_zone(gfp_mask);
5105 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5106 ac->nodemask = nodemask;
5107 ac->migratetype = gfp_migratetype(gfp_mask);
5109 if (cpusets_enabled()) {
5110 *alloc_gfp |= __GFP_HARDWALL;
5112 * When we are in the interrupt context, it is irrelevant
5113 * to the current task context. It means that any node ok.
5115 if (in_task() && !ac->nodemask)
5116 ac->nodemask = &cpuset_current_mems_allowed;
5118 *alloc_flags |= ALLOC_CPUSET;
5121 fs_reclaim_acquire(gfp_mask);
5122 fs_reclaim_release(gfp_mask);
5124 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5126 if (should_fail_alloc_page(gfp_mask, order))
5129 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5131 /* Dirty zone balancing only done in the fast path */
5132 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5135 * The preferred zone is used for statistics but crucially it is
5136 * also used as the starting point for the zonelist iterator. It
5137 * may get reset for allocations that ignore memory policies.
5139 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5140 ac->highest_zoneidx, ac->nodemask);
5146 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5147 * @gfp: GFP flags for the allocation
5148 * @preferred_nid: The preferred NUMA node ID to allocate from
5149 * @nodemask: Set of nodes to allocate from, may be NULL
5150 * @nr_pages: The number of pages desired on the list or array
5151 * @page_list: Optional list to store the allocated pages
5152 * @page_array: Optional array to store the pages
5154 * This is a batched version of the page allocator that attempts to
5155 * allocate nr_pages quickly. Pages are added to page_list if page_list
5156 * is not NULL, otherwise it is assumed that the page_array is valid.
5158 * For lists, nr_pages is the number of pages that should be allocated.
5160 * For arrays, only NULL elements are populated with pages and nr_pages
5161 * is the maximum number of pages that will be stored in the array.
5163 * Returns the number of pages on the list or array.
5165 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5166 nodemask_t *nodemask, int nr_pages,
5167 struct list_head *page_list,
5168 struct page **page_array)
5171 unsigned long flags;
5174 struct per_cpu_pages *pcp;
5175 struct list_head *pcp_list;
5176 struct alloc_context ac;
5178 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5179 int nr_populated = 0, nr_account = 0;
5182 * Skip populated array elements to determine if any pages need
5183 * to be allocated before disabling IRQs.
5185 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5188 /* No pages requested? */
5189 if (unlikely(nr_pages <= 0))
5192 /* Already populated array? */
5193 if (unlikely(page_array && nr_pages - nr_populated == 0))
5196 /* Bulk allocator does not support memcg accounting. */
5197 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5200 /* Use the single page allocator for one page. */
5201 if (nr_pages - nr_populated == 1)
5204 #ifdef CONFIG_PAGE_OWNER
5206 * PAGE_OWNER may recurse into the allocator to allocate space to
5207 * save the stack with pagesets.lock held. Releasing/reacquiring
5208 * removes much of the performance benefit of bulk allocation so
5209 * force the caller to allocate one page at a time as it'll have
5210 * similar performance to added complexity to the bulk allocator.
5212 if (static_branch_unlikely(&page_owner_inited))
5216 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5217 gfp &= gfp_allowed_mask;
5219 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5223 /* Find an allowed local zone that meets the low watermark. */
5224 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5227 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5228 !__cpuset_zone_allowed(zone, gfp)) {
5232 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5233 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5237 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5238 if (zone_watermark_fast(zone, 0, mark,
5239 zonelist_zone_idx(ac.preferred_zoneref),
5240 alloc_flags, gfp)) {
5246 * If there are no allowed local zones that meets the watermarks then
5247 * try to allocate a single page and reclaim if necessary.
5249 if (unlikely(!zone))
5252 /* Attempt the batch allocation */
5253 local_lock_irqsave(&pagesets.lock, flags);
5254 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5255 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5257 while (nr_populated < nr_pages) {
5259 /* Skip existing pages */
5260 if (page_array && page_array[nr_populated]) {
5265 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5267 if (unlikely(!page)) {
5268 /* Try and get at least one page */
5275 prep_new_page(page, 0, gfp, 0);
5277 list_add(&page->lru, page_list);
5279 page_array[nr_populated] = page;
5283 local_unlock_irqrestore(&pagesets.lock, flags);
5285 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5286 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5289 return nr_populated;
5292 local_unlock_irqrestore(&pagesets.lock, flags);
5295 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5298 list_add(&page->lru, page_list);
5300 page_array[nr_populated] = page;
5306 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5309 * This is the 'heart' of the zoned buddy allocator.
5311 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5312 nodemask_t *nodemask)
5315 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5316 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5317 struct alloc_context ac = { };
5320 * There are several places where we assume that the order value is sane
5321 * so bail out early if the request is out of bound.
5323 if (unlikely(order >= MAX_ORDER)) {
5324 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5328 gfp &= gfp_allowed_mask;
5330 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5331 * resp. GFP_NOIO which has to be inherited for all allocation requests
5332 * from a particular context which has been marked by
5333 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5334 * movable zones are not used during allocation.
5336 gfp = current_gfp_context(gfp);
5338 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5339 &alloc_gfp, &alloc_flags))
5343 * Forbid the first pass from falling back to types that fragment
5344 * memory until all local zones are considered.
5346 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5348 /* First allocation attempt */
5349 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5354 ac.spread_dirty_pages = false;
5357 * Restore the original nodemask if it was potentially replaced with
5358 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5360 ac.nodemask = nodemask;
5362 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5365 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5366 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5367 __free_pages(page, order);
5371 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5375 EXPORT_SYMBOL(__alloc_pages);
5377 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5378 nodemask_t *nodemask)
5380 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5381 preferred_nid, nodemask);
5383 if (page && order > 1)
5384 prep_transhuge_page(page);
5385 return (struct folio *)page;
5387 EXPORT_SYMBOL(__folio_alloc);
5390 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5391 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5392 * you need to access high mem.
5394 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5398 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5401 return (unsigned long) page_address(page);
5403 EXPORT_SYMBOL(__get_free_pages);
5405 unsigned long get_zeroed_page(gfp_t gfp_mask)
5407 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5409 EXPORT_SYMBOL(get_zeroed_page);
5412 * __free_pages - Free pages allocated with alloc_pages().
5413 * @page: The page pointer returned from alloc_pages().
5414 * @order: The order of the allocation.
5416 * This function can free multi-page allocations that are not compound
5417 * pages. It does not check that the @order passed in matches that of
5418 * the allocation, so it is easy to leak memory. Freeing more memory
5419 * than was allocated will probably emit a warning.
5421 * If the last reference to this page is speculative, it will be released
5422 * by put_page() which only frees the first page of a non-compound
5423 * allocation. To prevent the remaining pages from being leaked, we free
5424 * the subsequent pages here. If you want to use the page's reference
5425 * count to decide when to free the allocation, you should allocate a
5426 * compound page, and use put_page() instead of __free_pages().
5428 * Context: May be called in interrupt context or while holding a normal
5429 * spinlock, but not in NMI context or while holding a raw spinlock.
5431 void __free_pages(struct page *page, unsigned int order)
5433 if (put_page_testzero(page))
5434 free_the_page(page, order);
5435 else if (!PageHead(page))
5437 free_the_page(page + (1 << order), order);
5439 EXPORT_SYMBOL(__free_pages);
5441 void free_pages(unsigned long addr, unsigned int order)
5444 VM_BUG_ON(!virt_addr_valid((void *)addr));
5445 __free_pages(virt_to_page((void *)addr), order);
5449 EXPORT_SYMBOL(free_pages);
5453 * An arbitrary-length arbitrary-offset area of memory which resides
5454 * within a 0 or higher order page. Multiple fragments within that page
5455 * are individually refcounted, in the page's reference counter.
5457 * The page_frag functions below provide a simple allocation framework for
5458 * page fragments. This is used by the network stack and network device
5459 * drivers to provide a backing region of memory for use as either an
5460 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5462 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5465 struct page *page = NULL;
5466 gfp_t gfp = gfp_mask;
5468 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5469 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5471 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5472 PAGE_FRAG_CACHE_MAX_ORDER);
5473 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5475 if (unlikely(!page))
5476 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5478 nc->va = page ? page_address(page) : NULL;
5483 void __page_frag_cache_drain(struct page *page, unsigned int count)
5485 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5487 if (page_ref_sub_and_test(page, count))
5488 free_the_page(page, compound_order(page));
5490 EXPORT_SYMBOL(__page_frag_cache_drain);
5492 void *page_frag_alloc_align(struct page_frag_cache *nc,
5493 unsigned int fragsz, gfp_t gfp_mask,
5494 unsigned int align_mask)
5496 unsigned int size = PAGE_SIZE;
5500 if (unlikely(!nc->va)) {
5502 page = __page_frag_cache_refill(nc, gfp_mask);
5506 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5507 /* if size can vary use size else just use PAGE_SIZE */
5510 /* Even if we own the page, we do not use atomic_set().
5511 * This would break get_page_unless_zero() users.
5513 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5515 /* reset page count bias and offset to start of new frag */
5516 nc->pfmemalloc = page_is_pfmemalloc(page);
5517 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5521 offset = nc->offset - fragsz;
5522 if (unlikely(offset < 0)) {
5523 page = virt_to_page(nc->va);
5525 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5528 if (unlikely(nc->pfmemalloc)) {
5529 free_the_page(page, compound_order(page));
5533 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5534 /* if size can vary use size else just use PAGE_SIZE */
5537 /* OK, page count is 0, we can safely set it */
5538 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5540 /* reset page count bias and offset to start of new frag */
5541 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5542 offset = size - fragsz;
5546 offset &= align_mask;
5547 nc->offset = offset;
5549 return nc->va + offset;
5551 EXPORT_SYMBOL(page_frag_alloc_align);
5554 * Frees a page fragment allocated out of either a compound or order 0 page.
5556 void page_frag_free(void *addr)
5558 struct page *page = virt_to_head_page(addr);
5560 if (unlikely(put_page_testzero(page)))
5561 free_the_page(page, compound_order(page));
5563 EXPORT_SYMBOL(page_frag_free);
5565 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5569 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5570 unsigned long used = addr + PAGE_ALIGN(size);
5572 split_page(virt_to_page((void *)addr), order);
5573 while (used < alloc_end) {
5578 return (void *)addr;
5582 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5583 * @size: the number of bytes to allocate
5584 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5586 * This function is similar to alloc_pages(), except that it allocates the
5587 * minimum number of pages to satisfy the request. alloc_pages() can only
5588 * allocate memory in power-of-two pages.
5590 * This function is also limited by MAX_ORDER.
5592 * Memory allocated by this function must be released by free_pages_exact().
5594 * Return: pointer to the allocated area or %NULL in case of error.
5596 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5598 unsigned int order = get_order(size);
5601 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5602 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5604 addr = __get_free_pages(gfp_mask, order);
5605 return make_alloc_exact(addr, order, size);
5607 EXPORT_SYMBOL(alloc_pages_exact);
5610 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5612 * @nid: the preferred node ID where memory should be allocated
5613 * @size: the number of bytes to allocate
5614 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5616 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5619 * Return: pointer to the allocated area or %NULL in case of error.
5621 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5623 unsigned int order = get_order(size);
5626 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5627 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5629 p = alloc_pages_node(nid, gfp_mask, order);
5632 return make_alloc_exact((unsigned long)page_address(p), order, size);
5636 * free_pages_exact - release memory allocated via alloc_pages_exact()
5637 * @virt: the value returned by alloc_pages_exact.
5638 * @size: size of allocation, same value as passed to alloc_pages_exact().
5640 * Release the memory allocated by a previous call to alloc_pages_exact.
5642 void free_pages_exact(void *virt, size_t size)
5644 unsigned long addr = (unsigned long)virt;
5645 unsigned long end = addr + PAGE_ALIGN(size);
5647 while (addr < end) {
5652 EXPORT_SYMBOL(free_pages_exact);
5655 * nr_free_zone_pages - count number of pages beyond high watermark
5656 * @offset: The zone index of the highest zone
5658 * nr_free_zone_pages() counts the number of pages which are beyond the
5659 * high watermark within all zones at or below a given zone index. For each
5660 * zone, the number of pages is calculated as:
5662 * nr_free_zone_pages = managed_pages - high_pages
5664 * Return: number of pages beyond high watermark.
5666 static unsigned long nr_free_zone_pages(int offset)
5671 /* Just pick one node, since fallback list is circular */
5672 unsigned long sum = 0;
5674 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5676 for_each_zone_zonelist(zone, z, zonelist, offset) {
5677 unsigned long size = zone_managed_pages(zone);
5678 unsigned long high = high_wmark_pages(zone);
5687 * nr_free_buffer_pages - count number of pages beyond high watermark
5689 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5690 * watermark within ZONE_DMA and ZONE_NORMAL.
5692 * Return: number of pages beyond high watermark within ZONE_DMA and
5695 unsigned long nr_free_buffer_pages(void)
5697 return nr_free_zone_pages(gfp_zone(GFP_USER));
5699 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5701 static inline void show_node(struct zone *zone)
5703 if (IS_ENABLED(CONFIG_NUMA))
5704 printk("Node %d ", zone_to_nid(zone));
5707 long si_mem_available(void)
5710 unsigned long pagecache;
5711 unsigned long wmark_low = 0;
5712 unsigned long pages[NR_LRU_LISTS];
5713 unsigned long reclaimable;
5717 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5718 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5721 wmark_low += low_wmark_pages(zone);
5724 * Estimate the amount of memory available for userspace allocations,
5725 * without causing swapping.
5727 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5730 * Not all the page cache can be freed, otherwise the system will
5731 * start swapping. Assume at least half of the page cache, or the
5732 * low watermark worth of cache, needs to stay.
5734 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5735 pagecache -= min(pagecache / 2, wmark_low);
5736 available += pagecache;
5739 * Part of the reclaimable slab and other kernel memory consists of
5740 * items that are in use, and cannot be freed. Cap this estimate at the
5743 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5744 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5745 available += reclaimable - min(reclaimable / 2, wmark_low);
5751 EXPORT_SYMBOL_GPL(si_mem_available);
5753 void si_meminfo(struct sysinfo *val)
5755 val->totalram = totalram_pages();
5756 val->sharedram = global_node_page_state(NR_SHMEM);
5757 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5758 val->bufferram = nr_blockdev_pages();
5759 val->totalhigh = totalhigh_pages();
5760 val->freehigh = nr_free_highpages();
5761 val->mem_unit = PAGE_SIZE;
5764 EXPORT_SYMBOL(si_meminfo);
5767 void si_meminfo_node(struct sysinfo *val, int nid)
5769 int zone_type; /* needs to be signed */
5770 unsigned long managed_pages = 0;
5771 unsigned long managed_highpages = 0;
5772 unsigned long free_highpages = 0;
5773 pg_data_t *pgdat = NODE_DATA(nid);
5775 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5776 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5777 val->totalram = managed_pages;
5778 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5779 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5780 #ifdef CONFIG_HIGHMEM
5781 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5782 struct zone *zone = &pgdat->node_zones[zone_type];
5784 if (is_highmem(zone)) {
5785 managed_highpages += zone_managed_pages(zone);
5786 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5789 val->totalhigh = managed_highpages;
5790 val->freehigh = free_highpages;
5792 val->totalhigh = managed_highpages;
5793 val->freehigh = free_highpages;
5795 val->mem_unit = PAGE_SIZE;
5800 * Determine whether the node should be displayed or not, depending on whether
5801 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5803 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5805 if (!(flags & SHOW_MEM_FILTER_NODES))
5809 * no node mask - aka implicit memory numa policy. Do not bother with
5810 * the synchronization - read_mems_allowed_begin - because we do not
5811 * have to be precise here.
5814 nodemask = &cpuset_current_mems_allowed;
5816 return !node_isset(nid, *nodemask);
5819 #define K(x) ((x) << (PAGE_SHIFT-10))
5821 static void show_migration_types(unsigned char type)
5823 static const char types[MIGRATE_TYPES] = {
5824 [MIGRATE_UNMOVABLE] = 'U',
5825 [MIGRATE_MOVABLE] = 'M',
5826 [MIGRATE_RECLAIMABLE] = 'E',
5827 [MIGRATE_HIGHATOMIC] = 'H',
5829 [MIGRATE_CMA] = 'C',
5831 #ifdef CONFIG_MEMORY_ISOLATION
5832 [MIGRATE_ISOLATE] = 'I',
5835 char tmp[MIGRATE_TYPES + 1];
5839 for (i = 0; i < MIGRATE_TYPES; i++) {
5840 if (type & (1 << i))
5845 printk(KERN_CONT "(%s) ", tmp);
5849 * Show free area list (used inside shift_scroll-lock stuff)
5850 * We also calculate the percentage fragmentation. We do this by counting the
5851 * memory on each free list with the exception of the first item on the list.
5854 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5857 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5859 unsigned long free_pcp = 0;
5864 for_each_populated_zone(zone) {
5865 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5868 for_each_online_cpu(cpu)
5869 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5872 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5873 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5874 " unevictable:%lu dirty:%lu writeback:%lu\n"
5875 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5876 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5877 " kernel_misc_reclaimable:%lu\n"
5878 " free:%lu free_pcp:%lu free_cma:%lu\n",
5879 global_node_page_state(NR_ACTIVE_ANON),
5880 global_node_page_state(NR_INACTIVE_ANON),
5881 global_node_page_state(NR_ISOLATED_ANON),
5882 global_node_page_state(NR_ACTIVE_FILE),
5883 global_node_page_state(NR_INACTIVE_FILE),
5884 global_node_page_state(NR_ISOLATED_FILE),
5885 global_node_page_state(NR_UNEVICTABLE),
5886 global_node_page_state(NR_FILE_DIRTY),
5887 global_node_page_state(NR_WRITEBACK),
5888 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5889 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5890 global_node_page_state(NR_FILE_MAPPED),
5891 global_node_page_state(NR_SHMEM),
5892 global_node_page_state(NR_PAGETABLE),
5893 global_zone_page_state(NR_BOUNCE),
5894 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5895 global_zone_page_state(NR_FREE_PAGES),
5897 global_zone_page_state(NR_FREE_CMA_PAGES));
5899 for_each_online_pgdat(pgdat) {
5900 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5904 " active_anon:%lukB"
5905 " inactive_anon:%lukB"
5906 " active_file:%lukB"
5907 " inactive_file:%lukB"
5908 " unevictable:%lukB"
5909 " isolated(anon):%lukB"
5910 " isolated(file):%lukB"
5915 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5917 " shmem_pmdmapped: %lukB"
5920 " writeback_tmp:%lukB"
5921 " kernel_stack:%lukB"
5922 #ifdef CONFIG_SHADOW_CALL_STACK
5923 " shadow_call_stack:%lukB"
5926 " all_unreclaimable? %s"
5929 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5930 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5931 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5932 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5933 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5934 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5935 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5936 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5937 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5938 K(node_page_state(pgdat, NR_WRITEBACK)),
5939 K(node_page_state(pgdat, NR_SHMEM)),
5940 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5941 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5942 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5943 K(node_page_state(pgdat, NR_ANON_THPS)),
5945 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5946 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5947 #ifdef CONFIG_SHADOW_CALL_STACK
5948 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5950 K(node_page_state(pgdat, NR_PAGETABLE)),
5951 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5955 for_each_populated_zone(zone) {
5958 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5962 for_each_online_cpu(cpu)
5963 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5973 " reserved_highatomic:%luKB"
5974 " active_anon:%lukB"
5975 " inactive_anon:%lukB"
5976 " active_file:%lukB"
5977 " inactive_file:%lukB"
5978 " unevictable:%lukB"
5979 " writepending:%lukB"
5989 K(zone_page_state(zone, NR_FREE_PAGES)),
5990 K(zone->watermark_boost),
5991 K(min_wmark_pages(zone)),
5992 K(low_wmark_pages(zone)),
5993 K(high_wmark_pages(zone)),
5994 K(zone->nr_reserved_highatomic),
5995 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5996 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5997 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5998 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5999 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6000 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6001 K(zone->present_pages),
6002 K(zone_managed_pages(zone)),
6003 K(zone_page_state(zone, NR_MLOCK)),
6004 K(zone_page_state(zone, NR_BOUNCE)),
6006 K(this_cpu_read(zone->per_cpu_pageset->count)),
6007 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6008 printk("lowmem_reserve[]:");
6009 for (i = 0; i < MAX_NR_ZONES; i++)
6010 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6011 printk(KERN_CONT "\n");
6014 for_each_populated_zone(zone) {
6016 unsigned long nr[MAX_ORDER], flags, total = 0;
6017 unsigned char types[MAX_ORDER];
6019 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6022 printk(KERN_CONT "%s: ", zone->name);
6024 spin_lock_irqsave(&zone->lock, flags);
6025 for (order = 0; order < MAX_ORDER; order++) {
6026 struct free_area *area = &zone->free_area[order];
6029 nr[order] = area->nr_free;
6030 total += nr[order] << order;
6033 for (type = 0; type < MIGRATE_TYPES; type++) {
6034 if (!free_area_empty(area, type))
6035 types[order] |= 1 << type;
6038 spin_unlock_irqrestore(&zone->lock, flags);
6039 for (order = 0; order < MAX_ORDER; order++) {
6040 printk(KERN_CONT "%lu*%lukB ",
6041 nr[order], K(1UL) << order);
6043 show_migration_types(types[order]);
6045 printk(KERN_CONT "= %lukB\n", K(total));
6048 hugetlb_show_meminfo();
6050 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6052 show_swap_cache_info();
6055 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6057 zoneref->zone = zone;
6058 zoneref->zone_idx = zone_idx(zone);
6062 * Builds allocation fallback zone lists.
6064 * Add all populated zones of a node to the zonelist.
6066 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6069 enum zone_type zone_type = MAX_NR_ZONES;
6074 zone = pgdat->node_zones + zone_type;
6075 if (managed_zone(zone)) {
6076 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6077 check_highest_zone(zone_type);
6079 } while (zone_type);
6086 static int __parse_numa_zonelist_order(char *s)
6089 * We used to support different zonelists modes but they turned
6090 * out to be just not useful. Let's keep the warning in place
6091 * if somebody still use the cmd line parameter so that we do
6092 * not fail it silently
6094 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6095 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6101 char numa_zonelist_order[] = "Node";
6104 * sysctl handler for numa_zonelist_order
6106 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6107 void *buffer, size_t *length, loff_t *ppos)
6110 return __parse_numa_zonelist_order(buffer);
6111 return proc_dostring(table, write, buffer, length, ppos);
6115 #define MAX_NODE_LOAD (nr_online_nodes)
6116 static int node_load[MAX_NUMNODES];
6119 * find_next_best_node - find the next node that should appear in a given node's fallback list
6120 * @node: node whose fallback list we're appending
6121 * @used_node_mask: nodemask_t of already used nodes
6123 * We use a number of factors to determine which is the next node that should
6124 * appear on a given node's fallback list. The node should not have appeared
6125 * already in @node's fallback list, and it should be the next closest node
6126 * according to the distance array (which contains arbitrary distance values
6127 * from each node to each node in the system), and should also prefer nodes
6128 * with no CPUs, since presumably they'll have very little allocation pressure
6129 * on them otherwise.
6131 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6133 int find_next_best_node(int node, nodemask_t *used_node_mask)
6136 int min_val = INT_MAX;
6137 int best_node = NUMA_NO_NODE;
6139 /* Use the local node if we haven't already */
6140 if (!node_isset(node, *used_node_mask)) {
6141 node_set(node, *used_node_mask);
6145 for_each_node_state(n, N_MEMORY) {
6147 /* Don't want a node to appear more than once */
6148 if (node_isset(n, *used_node_mask))
6151 /* Use the distance array to find the distance */
6152 val = node_distance(node, n);
6154 /* Penalize nodes under us ("prefer the next node") */
6157 /* Give preference to headless and unused nodes */
6158 if (!cpumask_empty(cpumask_of_node(n)))
6159 val += PENALTY_FOR_NODE_WITH_CPUS;
6161 /* Slight preference for less loaded node */
6162 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6163 val += node_load[n];
6165 if (val < min_val) {
6172 node_set(best_node, *used_node_mask);
6179 * Build zonelists ordered by node and zones within node.
6180 * This results in maximum locality--normal zone overflows into local
6181 * DMA zone, if any--but risks exhausting DMA zone.
6183 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6186 struct zoneref *zonerefs;
6189 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6191 for (i = 0; i < nr_nodes; i++) {
6194 pg_data_t *node = NODE_DATA(node_order[i]);
6196 nr_zones = build_zonerefs_node(node, zonerefs);
6197 zonerefs += nr_zones;
6199 zonerefs->zone = NULL;
6200 zonerefs->zone_idx = 0;
6204 * Build gfp_thisnode zonelists
6206 static void build_thisnode_zonelists(pg_data_t *pgdat)
6208 struct zoneref *zonerefs;
6211 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6212 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6213 zonerefs += nr_zones;
6214 zonerefs->zone = NULL;
6215 zonerefs->zone_idx = 0;
6219 * Build zonelists ordered by zone and nodes within zones.
6220 * This results in conserving DMA zone[s] until all Normal memory is
6221 * exhausted, but results in overflowing to remote node while memory
6222 * may still exist in local DMA zone.
6225 static void build_zonelists(pg_data_t *pgdat)
6227 static int node_order[MAX_NUMNODES];
6228 int node, load, nr_nodes = 0;
6229 nodemask_t used_mask = NODE_MASK_NONE;
6230 int local_node, prev_node;
6232 /* NUMA-aware ordering of nodes */
6233 local_node = pgdat->node_id;
6234 load = nr_online_nodes;
6235 prev_node = local_node;
6237 memset(node_order, 0, sizeof(node_order));
6238 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6240 * We don't want to pressure a particular node.
6241 * So adding penalty to the first node in same
6242 * distance group to make it round-robin.
6244 if (node_distance(local_node, node) !=
6245 node_distance(local_node, prev_node))
6246 node_load[node] += load;
6248 node_order[nr_nodes++] = node;
6253 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6254 build_thisnode_zonelists(pgdat);
6255 pr_info("Fallback order for Node %d: ", local_node);
6256 for (node = 0; node < nr_nodes; node++)
6257 pr_cont("%d ", node_order[node]);
6261 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6263 * Return node id of node used for "local" allocations.
6264 * I.e., first node id of first zone in arg node's generic zonelist.
6265 * Used for initializing percpu 'numa_mem', which is used primarily
6266 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6268 int local_memory_node(int node)
6272 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6273 gfp_zone(GFP_KERNEL),
6275 return zone_to_nid(z->zone);
6279 static void setup_min_unmapped_ratio(void);
6280 static void setup_min_slab_ratio(void);
6281 #else /* CONFIG_NUMA */
6283 static void build_zonelists(pg_data_t *pgdat)
6285 int node, local_node;
6286 struct zoneref *zonerefs;
6289 local_node = pgdat->node_id;
6291 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6292 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6293 zonerefs += nr_zones;
6296 * Now we build the zonelist so that it contains the zones
6297 * of all the other nodes.
6298 * We don't want to pressure a particular node, so when
6299 * building the zones for node N, we make sure that the
6300 * zones coming right after the local ones are those from
6301 * node N+1 (modulo N)
6303 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6304 if (!node_online(node))
6306 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6307 zonerefs += nr_zones;
6309 for (node = 0; node < local_node; node++) {
6310 if (!node_online(node))
6312 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6313 zonerefs += nr_zones;
6316 zonerefs->zone = NULL;
6317 zonerefs->zone_idx = 0;
6320 #endif /* CONFIG_NUMA */
6323 * Boot pageset table. One per cpu which is going to be used for all
6324 * zones and all nodes. The parameters will be set in such a way
6325 * that an item put on a list will immediately be handed over to
6326 * the buddy list. This is safe since pageset manipulation is done
6327 * with interrupts disabled.
6329 * The boot_pagesets must be kept even after bootup is complete for
6330 * unused processors and/or zones. They do play a role for bootstrapping
6331 * hotplugged processors.
6333 * zoneinfo_show() and maybe other functions do
6334 * not check if the processor is online before following the pageset pointer.
6335 * Other parts of the kernel may not check if the zone is available.
6337 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6338 /* These effectively disable the pcplists in the boot pageset completely */
6339 #define BOOT_PAGESET_HIGH 0
6340 #define BOOT_PAGESET_BATCH 1
6341 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6342 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6343 DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6345 static void __build_all_zonelists(void *data)
6348 int __maybe_unused cpu;
6349 pg_data_t *self = data;
6350 static DEFINE_SPINLOCK(lock);
6355 memset(node_load, 0, sizeof(node_load));
6359 * This node is hotadded and no memory is yet present. So just
6360 * building zonelists is fine - no need to touch other nodes.
6362 if (self && !node_online(self->node_id)) {
6363 build_zonelists(self);
6366 * All possible nodes have pgdat preallocated
6369 for_each_node(nid) {
6370 pg_data_t *pgdat = NODE_DATA(nid);
6372 build_zonelists(pgdat);
6375 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6377 * We now know the "local memory node" for each node--
6378 * i.e., the node of the first zone in the generic zonelist.
6379 * Set up numa_mem percpu variable for on-line cpus. During
6380 * boot, only the boot cpu should be on-line; we'll init the
6381 * secondary cpus' numa_mem as they come on-line. During
6382 * node/memory hotplug, we'll fixup all on-line cpus.
6384 for_each_online_cpu(cpu)
6385 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6392 static noinline void __init
6393 build_all_zonelists_init(void)
6397 __build_all_zonelists(NULL);
6400 * Initialize the boot_pagesets that are going to be used
6401 * for bootstrapping processors. The real pagesets for
6402 * each zone will be allocated later when the per cpu
6403 * allocator is available.
6405 * boot_pagesets are used also for bootstrapping offline
6406 * cpus if the system is already booted because the pagesets
6407 * are needed to initialize allocators on a specific cpu too.
6408 * F.e. the percpu allocator needs the page allocator which
6409 * needs the percpu allocator in order to allocate its pagesets
6410 * (a chicken-egg dilemma).
6412 for_each_possible_cpu(cpu)
6413 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6415 mminit_verify_zonelist();
6416 cpuset_init_current_mems_allowed();
6420 * unless system_state == SYSTEM_BOOTING.
6422 * __ref due to call of __init annotated helper build_all_zonelists_init
6423 * [protected by SYSTEM_BOOTING].
6425 void __ref build_all_zonelists(pg_data_t *pgdat)
6427 unsigned long vm_total_pages;
6429 if (system_state == SYSTEM_BOOTING) {
6430 build_all_zonelists_init();
6432 __build_all_zonelists(pgdat);
6433 /* cpuset refresh routine should be here */
6435 /* Get the number of free pages beyond high watermark in all zones. */
6436 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6438 * Disable grouping by mobility if the number of pages in the
6439 * system is too low to allow the mechanism to work. It would be
6440 * more accurate, but expensive to check per-zone. This check is
6441 * made on memory-hotadd so a system can start with mobility
6442 * disabled and enable it later
6444 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6445 page_group_by_mobility_disabled = 1;
6447 page_group_by_mobility_disabled = 0;
6449 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6451 page_group_by_mobility_disabled ? "off" : "on",
6454 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6458 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6459 static bool __meminit
6460 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6462 static struct memblock_region *r;
6464 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6465 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6466 for_each_mem_region(r) {
6467 if (*pfn < memblock_region_memory_end_pfn(r))
6471 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6472 memblock_is_mirror(r)) {
6473 *pfn = memblock_region_memory_end_pfn(r);
6481 * Initially all pages are reserved - free ones are freed
6482 * up by memblock_free_all() once the early boot process is
6483 * done. Non-atomic initialization, single-pass.
6485 * All aligned pageblocks are initialized to the specified migratetype
6486 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6487 * zone stats (e.g., nr_isolate_pageblock) are touched.
6489 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6490 unsigned long start_pfn, unsigned long zone_end_pfn,
6491 enum meminit_context context,
6492 struct vmem_altmap *altmap, int migratetype)
6494 unsigned long pfn, end_pfn = start_pfn + size;
6497 if (highest_memmap_pfn < end_pfn - 1)
6498 highest_memmap_pfn = end_pfn - 1;
6500 #ifdef CONFIG_ZONE_DEVICE
6502 * Honor reservation requested by the driver for this ZONE_DEVICE
6503 * memory. We limit the total number of pages to initialize to just
6504 * those that might contain the memory mapping. We will defer the
6505 * ZONE_DEVICE page initialization until after we have released
6508 if (zone == ZONE_DEVICE) {
6512 if (start_pfn == altmap->base_pfn)
6513 start_pfn += altmap->reserve;
6514 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6518 for (pfn = start_pfn; pfn < end_pfn; ) {
6520 * There can be holes in boot-time mem_map[]s handed to this
6521 * function. They do not exist on hotplugged memory.
6523 if (context == MEMINIT_EARLY) {
6524 if (overlap_memmap_init(zone, &pfn))
6526 if (defer_init(nid, pfn, zone_end_pfn))
6530 page = pfn_to_page(pfn);
6531 __init_single_page(page, pfn, zone, nid);
6532 if (context == MEMINIT_HOTPLUG)
6533 __SetPageReserved(page);
6536 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6537 * such that unmovable allocations won't be scattered all
6538 * over the place during system boot.
6540 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6541 set_pageblock_migratetype(page, migratetype);
6548 #ifdef CONFIG_ZONE_DEVICE
6549 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6550 unsigned long zone_idx, int nid,
6551 struct dev_pagemap *pgmap)
6554 __init_single_page(page, pfn, zone_idx, nid);
6557 * Mark page reserved as it will need to wait for onlining
6558 * phase for it to be fully associated with a zone.
6560 * We can use the non-atomic __set_bit operation for setting
6561 * the flag as we are still initializing the pages.
6563 __SetPageReserved(page);
6566 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6567 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6568 * ever freed or placed on a driver-private list.
6570 page->pgmap = pgmap;
6571 page->zone_device_data = NULL;
6574 * Mark the block movable so that blocks are reserved for
6575 * movable at startup. This will force kernel allocations
6576 * to reserve their blocks rather than leaking throughout
6577 * the address space during boot when many long-lived
6578 * kernel allocations are made.
6580 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6581 * because this is done early in section_activate()
6583 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6584 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6589 static void __ref memmap_init_compound(struct page *head,
6590 unsigned long head_pfn,
6591 unsigned long zone_idx, int nid,
6592 struct dev_pagemap *pgmap,
6593 unsigned long nr_pages)
6595 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6596 unsigned int order = pgmap->vmemmap_shift;
6598 __SetPageHead(head);
6599 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6600 struct page *page = pfn_to_page(pfn);
6602 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6603 prep_compound_tail(head, pfn - head_pfn);
6604 set_page_count(page, 0);
6607 * The first tail page stores compound_mapcount_ptr() and
6608 * compound_order() and the second tail page stores
6609 * compound_pincount_ptr(). Call prep_compound_head() after
6610 * the first and second tail pages have been initialized to
6611 * not have the data overwritten.
6613 if (pfn == head_pfn + 2)
6614 prep_compound_head(head, order);
6618 void __ref memmap_init_zone_device(struct zone *zone,
6619 unsigned long start_pfn,
6620 unsigned long nr_pages,
6621 struct dev_pagemap *pgmap)
6623 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6624 struct pglist_data *pgdat = zone->zone_pgdat;
6625 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6626 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6627 unsigned long zone_idx = zone_idx(zone);
6628 unsigned long start = jiffies;
6629 int nid = pgdat->node_id;
6631 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6635 * The call to memmap_init should have already taken care
6636 * of the pages reserved for the memmap, so we can just jump to
6637 * the end of that region and start processing the device pages.
6640 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6641 nr_pages = end_pfn - start_pfn;
6644 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6645 struct page *page = pfn_to_page(pfn);
6647 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6649 if (pfns_per_compound == 1)
6652 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6656 pr_info("%s initialised %lu pages in %ums\n", __func__,
6657 nr_pages, jiffies_to_msecs(jiffies - start));
6661 static void __meminit zone_init_free_lists(struct zone *zone)
6663 unsigned int order, t;
6664 for_each_migratetype_order(order, t) {
6665 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6666 zone->free_area[order].nr_free = 0;
6671 * Only struct pages that correspond to ranges defined by memblock.memory
6672 * are zeroed and initialized by going through __init_single_page() during
6673 * memmap_init_zone_range().
6675 * But, there could be struct pages that correspond to holes in
6676 * memblock.memory. This can happen because of the following reasons:
6677 * - physical memory bank size is not necessarily the exact multiple of the
6678 * arbitrary section size
6679 * - early reserved memory may not be listed in memblock.memory
6680 * - memory layouts defined with memmap= kernel parameter may not align
6681 * nicely with memmap sections
6683 * Explicitly initialize those struct pages so that:
6684 * - PG_Reserved is set
6685 * - zone and node links point to zone and node that span the page if the
6686 * hole is in the middle of a zone
6687 * - zone and node links point to adjacent zone/node if the hole falls on
6688 * the zone boundary; the pages in such holes will be prepended to the
6689 * zone/node above the hole except for the trailing pages in the last
6690 * section that will be appended to the zone/node below.
6692 static void __init init_unavailable_range(unsigned long spfn,
6699 for (pfn = spfn; pfn < epfn; pfn++) {
6700 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6701 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6702 + pageblock_nr_pages - 1;
6705 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6706 __SetPageReserved(pfn_to_page(pfn));
6711 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6712 node, zone_names[zone], pgcnt);
6715 static void __init memmap_init_zone_range(struct zone *zone,
6716 unsigned long start_pfn,
6717 unsigned long end_pfn,
6718 unsigned long *hole_pfn)
6720 unsigned long zone_start_pfn = zone->zone_start_pfn;
6721 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6722 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6724 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6725 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6727 if (start_pfn >= end_pfn)
6730 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6731 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6733 if (*hole_pfn < start_pfn)
6734 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6736 *hole_pfn = end_pfn;
6739 static void __init memmap_init(void)
6741 unsigned long start_pfn, end_pfn;
6742 unsigned long hole_pfn = 0;
6743 int i, j, zone_id = 0, nid;
6745 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6746 struct pglist_data *node = NODE_DATA(nid);
6748 for (j = 0; j < MAX_NR_ZONES; j++) {
6749 struct zone *zone = node->node_zones + j;
6751 if (!populated_zone(zone))
6754 memmap_init_zone_range(zone, start_pfn, end_pfn,
6760 #ifdef CONFIG_SPARSEMEM
6762 * Initialize the memory map for hole in the range [memory_end,
6764 * Append the pages in this hole to the highest zone in the last
6766 * The call to init_unavailable_range() is outside the ifdef to
6767 * silence the compiler warining about zone_id set but not used;
6768 * for FLATMEM it is a nop anyway
6770 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6771 if (hole_pfn < end_pfn)
6773 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6776 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6777 phys_addr_t min_addr, int nid, bool exact_nid)
6782 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6783 MEMBLOCK_ALLOC_ACCESSIBLE,
6786 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6787 MEMBLOCK_ALLOC_ACCESSIBLE,
6790 if (ptr && size > 0)
6791 page_init_poison(ptr, size);
6796 static int zone_batchsize(struct zone *zone)
6802 * The number of pages to batch allocate is either ~0.1%
6803 * of the zone or 1MB, whichever is smaller. The batch
6804 * size is striking a balance between allocation latency
6805 * and zone lock contention.
6807 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6808 batch /= 4; /* We effectively *= 4 below */
6813 * Clamp the batch to a 2^n - 1 value. Having a power
6814 * of 2 value was found to be more likely to have
6815 * suboptimal cache aliasing properties in some cases.
6817 * For example if 2 tasks are alternately allocating
6818 * batches of pages, one task can end up with a lot
6819 * of pages of one half of the possible page colors
6820 * and the other with pages of the other colors.
6822 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6827 /* The deferral and batching of frees should be suppressed under NOMMU
6830 * The problem is that NOMMU needs to be able to allocate large chunks
6831 * of contiguous memory as there's no hardware page translation to
6832 * assemble apparent contiguous memory from discontiguous pages.
6834 * Queueing large contiguous runs of pages for batching, however,
6835 * causes the pages to actually be freed in smaller chunks. As there
6836 * can be a significant delay between the individual batches being
6837 * recycled, this leads to the once large chunks of space being
6838 * fragmented and becoming unavailable for high-order allocations.
6844 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6849 unsigned long total_pages;
6851 if (!percpu_pagelist_high_fraction) {
6853 * By default, the high value of the pcp is based on the zone
6854 * low watermark so that if they are full then background
6855 * reclaim will not be started prematurely.
6857 total_pages = low_wmark_pages(zone);
6860 * If percpu_pagelist_high_fraction is configured, the high
6861 * value is based on a fraction of the managed pages in the
6864 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6868 * Split the high value across all online CPUs local to the zone. Note
6869 * that early in boot that CPUs may not be online yet and that during
6870 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6871 * onlined. For memory nodes that have no CPUs, split pcp->high across
6872 * all online CPUs to mitigate the risk that reclaim is triggered
6873 * prematurely due to pages stored on pcp lists.
6875 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6877 nr_split_cpus = num_online_cpus();
6878 high = total_pages / nr_split_cpus;
6881 * Ensure high is at least batch*4. The multiple is based on the
6882 * historical relationship between high and batch.
6884 high = max(high, batch << 2);
6893 * pcp->high and pcp->batch values are related and generally batch is lower
6894 * than high. They are also related to pcp->count such that count is lower
6895 * than high, and as soon as it reaches high, the pcplist is flushed.
6897 * However, guaranteeing these relations at all times would require e.g. write
6898 * barriers here but also careful usage of read barriers at the read side, and
6899 * thus be prone to error and bad for performance. Thus the update only prevents
6900 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6901 * can cope with those fields changing asynchronously, and fully trust only the
6902 * pcp->count field on the local CPU with interrupts disabled.
6904 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6905 * outside of boot time (or some other assurance that no concurrent updaters
6908 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6909 unsigned long batch)
6911 WRITE_ONCE(pcp->batch, batch);
6912 WRITE_ONCE(pcp->high, high);
6915 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6919 memset(pcp, 0, sizeof(*pcp));
6920 memset(pzstats, 0, sizeof(*pzstats));
6922 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6923 INIT_LIST_HEAD(&pcp->lists[pindex]);
6926 * Set batch and high values safe for a boot pageset. A true percpu
6927 * pageset's initialization will update them subsequently. Here we don't
6928 * need to be as careful as pageset_update() as nobody can access the
6931 pcp->high = BOOT_PAGESET_HIGH;
6932 pcp->batch = BOOT_PAGESET_BATCH;
6933 pcp->free_factor = 0;
6936 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6937 unsigned long batch)
6939 struct per_cpu_pages *pcp;
6942 for_each_possible_cpu(cpu) {
6943 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6944 pageset_update(pcp, high, batch);
6949 * Calculate and set new high and batch values for all per-cpu pagesets of a
6950 * zone based on the zone's size.
6952 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6954 int new_high, new_batch;
6956 new_batch = max(1, zone_batchsize(zone));
6957 new_high = zone_highsize(zone, new_batch, cpu_online);
6959 if (zone->pageset_high == new_high &&
6960 zone->pageset_batch == new_batch)
6963 zone->pageset_high = new_high;
6964 zone->pageset_batch = new_batch;
6966 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6969 void __meminit setup_zone_pageset(struct zone *zone)
6973 /* Size may be 0 on !SMP && !NUMA */
6974 if (sizeof(struct per_cpu_zonestat) > 0)
6975 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6977 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6978 for_each_possible_cpu(cpu) {
6979 struct per_cpu_pages *pcp;
6980 struct per_cpu_zonestat *pzstats;
6982 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6983 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6984 per_cpu_pages_init(pcp, pzstats);
6987 zone_set_pageset_high_and_batch(zone, 0);
6991 * Allocate per cpu pagesets and initialize them.
6992 * Before this call only boot pagesets were available.
6994 void __init setup_per_cpu_pageset(void)
6996 struct pglist_data *pgdat;
6998 int __maybe_unused cpu;
7000 for_each_populated_zone(zone)
7001 setup_zone_pageset(zone);
7005 * Unpopulated zones continue using the boot pagesets.
7006 * The numa stats for these pagesets need to be reset.
7007 * Otherwise, they will end up skewing the stats of
7008 * the nodes these zones are associated with.
7010 for_each_possible_cpu(cpu) {
7011 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7012 memset(pzstats->vm_numa_event, 0,
7013 sizeof(pzstats->vm_numa_event));
7017 for_each_online_pgdat(pgdat)
7018 pgdat->per_cpu_nodestats =
7019 alloc_percpu(struct per_cpu_nodestat);
7022 static __meminit void zone_pcp_init(struct zone *zone)
7025 * per cpu subsystem is not up at this point. The following code
7026 * relies on the ability of the linker to provide the
7027 * offset of a (static) per cpu variable into the per cpu area.
7029 zone->per_cpu_pageset = &boot_pageset;
7030 zone->per_cpu_zonestats = &boot_zonestats;
7031 zone->pageset_high = BOOT_PAGESET_HIGH;
7032 zone->pageset_batch = BOOT_PAGESET_BATCH;
7034 if (populated_zone(zone))
7035 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7036 zone->present_pages, zone_batchsize(zone));
7039 void __meminit init_currently_empty_zone(struct zone *zone,
7040 unsigned long zone_start_pfn,
7043 struct pglist_data *pgdat = zone->zone_pgdat;
7044 int zone_idx = zone_idx(zone) + 1;
7046 if (zone_idx > pgdat->nr_zones)
7047 pgdat->nr_zones = zone_idx;
7049 zone->zone_start_pfn = zone_start_pfn;
7051 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7052 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7054 (unsigned long)zone_idx(zone),
7055 zone_start_pfn, (zone_start_pfn + size));
7057 zone_init_free_lists(zone);
7058 zone->initialized = 1;
7062 * get_pfn_range_for_nid - Return the start and end page frames for a node
7063 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7064 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7065 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7067 * It returns the start and end page frame of a node based on information
7068 * provided by memblock_set_node(). If called for a node
7069 * with no available memory, a warning is printed and the start and end
7072 void __init get_pfn_range_for_nid(unsigned int nid,
7073 unsigned long *start_pfn, unsigned long *end_pfn)
7075 unsigned long this_start_pfn, this_end_pfn;
7081 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7082 *start_pfn = min(*start_pfn, this_start_pfn);
7083 *end_pfn = max(*end_pfn, this_end_pfn);
7086 if (*start_pfn == -1UL)
7091 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7092 * assumption is made that zones within a node are ordered in monotonic
7093 * increasing memory addresses so that the "highest" populated zone is used
7095 static void __init find_usable_zone_for_movable(void)
7098 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7099 if (zone_index == ZONE_MOVABLE)
7102 if (arch_zone_highest_possible_pfn[zone_index] >
7103 arch_zone_lowest_possible_pfn[zone_index])
7107 VM_BUG_ON(zone_index == -1);
7108 movable_zone = zone_index;
7112 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7113 * because it is sized independent of architecture. Unlike the other zones,
7114 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7115 * in each node depending on the size of each node and how evenly kernelcore
7116 * is distributed. This helper function adjusts the zone ranges
7117 * provided by the architecture for a given node by using the end of the
7118 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7119 * zones within a node are in order of monotonic increases memory addresses
7121 static void __init adjust_zone_range_for_zone_movable(int nid,
7122 unsigned long zone_type,
7123 unsigned long node_start_pfn,
7124 unsigned long node_end_pfn,
7125 unsigned long *zone_start_pfn,
7126 unsigned long *zone_end_pfn)
7128 /* Only adjust if ZONE_MOVABLE is on this node */
7129 if (zone_movable_pfn[nid]) {
7130 /* Size ZONE_MOVABLE */
7131 if (zone_type == ZONE_MOVABLE) {
7132 *zone_start_pfn = zone_movable_pfn[nid];
7133 *zone_end_pfn = min(node_end_pfn,
7134 arch_zone_highest_possible_pfn[movable_zone]);
7136 /* Adjust for ZONE_MOVABLE starting within this range */
7137 } else if (!mirrored_kernelcore &&
7138 *zone_start_pfn < zone_movable_pfn[nid] &&
7139 *zone_end_pfn > zone_movable_pfn[nid]) {
7140 *zone_end_pfn = zone_movable_pfn[nid];
7142 /* Check if this whole range is within ZONE_MOVABLE */
7143 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7144 *zone_start_pfn = *zone_end_pfn;
7149 * Return the number of pages a zone spans in a node, including holes
7150 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7152 static unsigned long __init zone_spanned_pages_in_node(int nid,
7153 unsigned long zone_type,
7154 unsigned long node_start_pfn,
7155 unsigned long node_end_pfn,
7156 unsigned long *zone_start_pfn,
7157 unsigned long *zone_end_pfn)
7159 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7160 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7161 /* When hotadd a new node from cpu_up(), the node should be empty */
7162 if (!node_start_pfn && !node_end_pfn)
7165 /* Get the start and end of the zone */
7166 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7167 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7168 adjust_zone_range_for_zone_movable(nid, zone_type,
7169 node_start_pfn, node_end_pfn,
7170 zone_start_pfn, zone_end_pfn);
7172 /* Check that this node has pages within the zone's required range */
7173 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7176 /* Move the zone boundaries inside the node if necessary */
7177 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7178 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7180 /* Return the spanned pages */
7181 return *zone_end_pfn - *zone_start_pfn;
7185 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7186 * then all holes in the requested range will be accounted for.
7188 unsigned long __init __absent_pages_in_range(int nid,
7189 unsigned long range_start_pfn,
7190 unsigned long range_end_pfn)
7192 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7193 unsigned long start_pfn, end_pfn;
7196 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7197 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7198 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7199 nr_absent -= end_pfn - start_pfn;
7205 * absent_pages_in_range - Return number of page frames in holes within a range
7206 * @start_pfn: The start PFN to start searching for holes
7207 * @end_pfn: The end PFN to stop searching for holes
7209 * Return: the number of pages frames in memory holes within a range.
7211 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7212 unsigned long end_pfn)
7214 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7217 /* Return the number of page frames in holes in a zone on a node */
7218 static unsigned long __init zone_absent_pages_in_node(int nid,
7219 unsigned long zone_type,
7220 unsigned long node_start_pfn,
7221 unsigned long node_end_pfn)
7223 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7224 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7225 unsigned long zone_start_pfn, zone_end_pfn;
7226 unsigned long nr_absent;
7228 /* When hotadd a new node from cpu_up(), the node should be empty */
7229 if (!node_start_pfn && !node_end_pfn)
7232 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7233 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7235 adjust_zone_range_for_zone_movable(nid, zone_type,
7236 node_start_pfn, node_end_pfn,
7237 &zone_start_pfn, &zone_end_pfn);
7238 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7241 * ZONE_MOVABLE handling.
7242 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7245 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7246 unsigned long start_pfn, end_pfn;
7247 struct memblock_region *r;
7249 for_each_mem_region(r) {
7250 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7251 zone_start_pfn, zone_end_pfn);
7252 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7253 zone_start_pfn, zone_end_pfn);
7255 if (zone_type == ZONE_MOVABLE &&
7256 memblock_is_mirror(r))
7257 nr_absent += end_pfn - start_pfn;
7259 if (zone_type == ZONE_NORMAL &&
7260 !memblock_is_mirror(r))
7261 nr_absent += end_pfn - start_pfn;
7268 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7269 unsigned long node_start_pfn,
7270 unsigned long node_end_pfn)
7272 unsigned long realtotalpages = 0, totalpages = 0;
7275 for (i = 0; i < MAX_NR_ZONES; i++) {
7276 struct zone *zone = pgdat->node_zones + i;
7277 unsigned long zone_start_pfn, zone_end_pfn;
7278 unsigned long spanned, absent;
7279 unsigned long size, real_size;
7281 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7286 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7291 real_size = size - absent;
7294 zone->zone_start_pfn = zone_start_pfn;
7296 zone->zone_start_pfn = 0;
7297 zone->spanned_pages = size;
7298 zone->present_pages = real_size;
7299 #if defined(CONFIG_MEMORY_HOTPLUG)
7300 zone->present_early_pages = real_size;
7304 realtotalpages += real_size;
7307 pgdat->node_spanned_pages = totalpages;
7308 pgdat->node_present_pages = realtotalpages;
7309 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7312 #ifndef CONFIG_SPARSEMEM
7314 * Calculate the size of the zone->blockflags rounded to an unsigned long
7315 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7316 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7317 * round what is now in bits to nearest long in bits, then return it in
7320 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7322 unsigned long usemapsize;
7324 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7325 usemapsize = roundup(zonesize, pageblock_nr_pages);
7326 usemapsize = usemapsize >> pageblock_order;
7327 usemapsize *= NR_PAGEBLOCK_BITS;
7328 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7330 return usemapsize / 8;
7333 static void __ref setup_usemap(struct zone *zone)
7335 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7336 zone->spanned_pages);
7337 zone->pageblock_flags = NULL;
7339 zone->pageblock_flags =
7340 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7342 if (!zone->pageblock_flags)
7343 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7344 usemapsize, zone->name, zone_to_nid(zone));
7348 static inline void setup_usemap(struct zone *zone) {}
7349 #endif /* CONFIG_SPARSEMEM */
7351 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7353 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7354 void __init set_pageblock_order(void)
7356 unsigned int order = MAX_ORDER - 1;
7358 /* Check that pageblock_nr_pages has not already been setup */
7359 if (pageblock_order)
7362 /* Don't let pageblocks exceed the maximum allocation granularity. */
7363 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7364 order = HUGETLB_PAGE_ORDER;
7367 * Assume the largest contiguous order of interest is a huge page.
7368 * This value may be variable depending on boot parameters on IA64 and
7371 pageblock_order = order;
7373 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7376 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7377 * is unused as pageblock_order is set at compile-time. See
7378 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7381 void __init set_pageblock_order(void)
7385 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7387 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7388 unsigned long present_pages)
7390 unsigned long pages = spanned_pages;
7393 * Provide a more accurate estimation if there are holes within
7394 * the zone and SPARSEMEM is in use. If there are holes within the
7395 * zone, each populated memory region may cost us one or two extra
7396 * memmap pages due to alignment because memmap pages for each
7397 * populated regions may not be naturally aligned on page boundary.
7398 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7400 if (spanned_pages > present_pages + (present_pages >> 4) &&
7401 IS_ENABLED(CONFIG_SPARSEMEM))
7402 pages = present_pages;
7404 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7407 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7408 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7410 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7412 spin_lock_init(&ds_queue->split_queue_lock);
7413 INIT_LIST_HEAD(&ds_queue->split_queue);
7414 ds_queue->split_queue_len = 0;
7417 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7420 #ifdef CONFIG_COMPACTION
7421 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7423 init_waitqueue_head(&pgdat->kcompactd_wait);
7426 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7429 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7433 pgdat_resize_init(pgdat);
7435 pgdat_init_split_queue(pgdat);
7436 pgdat_init_kcompactd(pgdat);
7438 init_waitqueue_head(&pgdat->kswapd_wait);
7439 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7441 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7442 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7444 pgdat_page_ext_init(pgdat);
7445 lruvec_init(&pgdat->__lruvec);
7448 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7449 unsigned long remaining_pages)
7451 atomic_long_set(&zone->managed_pages, remaining_pages);
7452 zone_set_nid(zone, nid);
7453 zone->name = zone_names[idx];
7454 zone->zone_pgdat = NODE_DATA(nid);
7455 spin_lock_init(&zone->lock);
7456 zone_seqlock_init(zone);
7457 zone_pcp_init(zone);
7461 * Set up the zone data structures
7462 * - init pgdat internals
7463 * - init all zones belonging to this node
7465 * NOTE: this function is only called during memory hotplug
7467 #ifdef CONFIG_MEMORY_HOTPLUG
7468 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7470 int nid = pgdat->node_id;
7474 pgdat_init_internals(pgdat);
7476 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7477 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7480 * Reset the nr_zones, order and highest_zoneidx before reuse.
7481 * Note that kswapd will init kswapd_highest_zoneidx properly
7482 * when it starts in the near future.
7484 pgdat->nr_zones = 0;
7485 pgdat->kswapd_order = 0;
7486 pgdat->kswapd_highest_zoneidx = 0;
7487 pgdat->node_start_pfn = 0;
7488 for_each_online_cpu(cpu) {
7489 struct per_cpu_nodestat *p;
7491 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7492 memset(p, 0, sizeof(*p));
7495 for (z = 0; z < MAX_NR_ZONES; z++)
7496 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7501 * Set up the zone data structures:
7502 * - mark all pages reserved
7503 * - mark all memory queues empty
7504 * - clear the memory bitmaps
7506 * NOTE: pgdat should get zeroed by caller.
7507 * NOTE: this function is only called during early init.
7509 static void __init free_area_init_core(struct pglist_data *pgdat)
7512 int nid = pgdat->node_id;
7514 pgdat_init_internals(pgdat);
7515 pgdat->per_cpu_nodestats = &boot_nodestats;
7517 for (j = 0; j < MAX_NR_ZONES; j++) {
7518 struct zone *zone = pgdat->node_zones + j;
7519 unsigned long size, freesize, memmap_pages;
7521 size = zone->spanned_pages;
7522 freesize = zone->present_pages;
7525 * Adjust freesize so that it accounts for how much memory
7526 * is used by this zone for memmap. This affects the watermark
7527 * and per-cpu initialisations
7529 memmap_pages = calc_memmap_size(size, freesize);
7530 if (!is_highmem_idx(j)) {
7531 if (freesize >= memmap_pages) {
7532 freesize -= memmap_pages;
7534 pr_debug(" %s zone: %lu pages used for memmap\n",
7535 zone_names[j], memmap_pages);
7537 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7538 zone_names[j], memmap_pages, freesize);
7541 /* Account for reserved pages */
7542 if (j == 0 && freesize > dma_reserve) {
7543 freesize -= dma_reserve;
7544 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7547 if (!is_highmem_idx(j))
7548 nr_kernel_pages += freesize;
7549 /* Charge for highmem memmap if there are enough kernel pages */
7550 else if (nr_kernel_pages > memmap_pages * 2)
7551 nr_kernel_pages -= memmap_pages;
7552 nr_all_pages += freesize;
7555 * Set an approximate value for lowmem here, it will be adjusted
7556 * when the bootmem allocator frees pages into the buddy system.
7557 * And all highmem pages will be managed by the buddy system.
7559 zone_init_internals(zone, j, nid, freesize);
7564 set_pageblock_order();
7566 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7570 #ifdef CONFIG_FLATMEM
7571 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7573 unsigned long __maybe_unused start = 0;
7574 unsigned long __maybe_unused offset = 0;
7576 /* Skip empty nodes */
7577 if (!pgdat->node_spanned_pages)
7580 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7581 offset = pgdat->node_start_pfn - start;
7582 /* ia64 gets its own node_mem_map, before this, without bootmem */
7583 if (!pgdat->node_mem_map) {
7584 unsigned long size, end;
7588 * The zone's endpoints aren't required to be MAX_ORDER
7589 * aligned but the node_mem_map endpoints must be in order
7590 * for the buddy allocator to function correctly.
7592 end = pgdat_end_pfn(pgdat);
7593 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7594 size = (end - start) * sizeof(struct page);
7595 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7596 pgdat->node_id, false);
7598 panic("Failed to allocate %ld bytes for node %d memory map\n",
7599 size, pgdat->node_id);
7600 pgdat->node_mem_map = map + offset;
7602 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7603 __func__, pgdat->node_id, (unsigned long)pgdat,
7604 (unsigned long)pgdat->node_mem_map);
7607 * With no DISCONTIG, the global mem_map is just set as node 0's
7609 if (pgdat == NODE_DATA(0)) {
7610 mem_map = NODE_DATA(0)->node_mem_map;
7611 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7617 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7618 #endif /* CONFIG_FLATMEM */
7620 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7621 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7623 pgdat->first_deferred_pfn = ULONG_MAX;
7626 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7629 static void __init free_area_init_node(int nid)
7631 pg_data_t *pgdat = NODE_DATA(nid);
7632 unsigned long start_pfn = 0;
7633 unsigned long end_pfn = 0;
7635 /* pg_data_t should be reset to zero when it's allocated */
7636 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7638 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7640 pgdat->node_id = nid;
7641 pgdat->node_start_pfn = start_pfn;
7642 pgdat->per_cpu_nodestats = NULL;
7644 if (start_pfn != end_pfn) {
7645 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7646 (u64)start_pfn << PAGE_SHIFT,
7647 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7649 pr_info("Initmem setup node %d as memoryless\n", nid);
7652 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7654 alloc_node_mem_map(pgdat);
7655 pgdat_set_deferred_range(pgdat);
7657 free_area_init_core(pgdat);
7660 static void __init free_area_init_memoryless_node(int nid)
7662 free_area_init_node(nid);
7665 #if MAX_NUMNODES > 1
7667 * Figure out the number of possible node ids.
7669 void __init setup_nr_node_ids(void)
7671 unsigned int highest;
7673 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7674 nr_node_ids = highest + 1;
7679 * node_map_pfn_alignment - determine the maximum internode alignment
7681 * This function should be called after node map is populated and sorted.
7682 * It calculates the maximum power of two alignment which can distinguish
7685 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7686 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7687 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7688 * shifted, 1GiB is enough and this function will indicate so.
7690 * This is used to test whether pfn -> nid mapping of the chosen memory
7691 * model has fine enough granularity to avoid incorrect mapping for the
7692 * populated node map.
7694 * Return: the determined alignment in pfn's. 0 if there is no alignment
7695 * requirement (single node).
7697 unsigned long __init node_map_pfn_alignment(void)
7699 unsigned long accl_mask = 0, last_end = 0;
7700 unsigned long start, end, mask;
7701 int last_nid = NUMA_NO_NODE;
7704 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7705 if (!start || last_nid < 0 || last_nid == nid) {
7712 * Start with a mask granular enough to pin-point to the
7713 * start pfn and tick off bits one-by-one until it becomes
7714 * too coarse to separate the current node from the last.
7716 mask = ~((1 << __ffs(start)) - 1);
7717 while (mask && last_end <= (start & (mask << 1)))
7720 /* accumulate all internode masks */
7724 /* convert mask to number of pages */
7725 return ~accl_mask + 1;
7729 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7731 * Return: the minimum PFN based on information provided via
7732 * memblock_set_node().
7734 unsigned long __init find_min_pfn_with_active_regions(void)
7736 return PHYS_PFN(memblock_start_of_DRAM());
7740 * early_calculate_totalpages()
7741 * Sum pages in active regions for movable zone.
7742 * Populate N_MEMORY for calculating usable_nodes.
7744 static unsigned long __init early_calculate_totalpages(void)
7746 unsigned long totalpages = 0;
7747 unsigned long start_pfn, end_pfn;
7750 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7751 unsigned long pages = end_pfn - start_pfn;
7753 totalpages += pages;
7755 node_set_state(nid, N_MEMORY);
7761 * Find the PFN the Movable zone begins in each node. Kernel memory
7762 * is spread evenly between nodes as long as the nodes have enough
7763 * memory. When they don't, some nodes will have more kernelcore than
7766 static void __init find_zone_movable_pfns_for_nodes(void)
7769 unsigned long usable_startpfn;
7770 unsigned long kernelcore_node, kernelcore_remaining;
7771 /* save the state before borrow the nodemask */
7772 nodemask_t saved_node_state = node_states[N_MEMORY];
7773 unsigned long totalpages = early_calculate_totalpages();
7774 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7775 struct memblock_region *r;
7777 /* Need to find movable_zone earlier when movable_node is specified. */
7778 find_usable_zone_for_movable();
7781 * If movable_node is specified, ignore kernelcore and movablecore
7784 if (movable_node_is_enabled()) {
7785 for_each_mem_region(r) {
7786 if (!memblock_is_hotpluggable(r))
7789 nid = memblock_get_region_node(r);
7791 usable_startpfn = PFN_DOWN(r->base);
7792 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7793 min(usable_startpfn, zone_movable_pfn[nid]) :
7801 * If kernelcore=mirror is specified, ignore movablecore option
7803 if (mirrored_kernelcore) {
7804 bool mem_below_4gb_not_mirrored = false;
7806 for_each_mem_region(r) {
7807 if (memblock_is_mirror(r))
7810 nid = memblock_get_region_node(r);
7812 usable_startpfn = memblock_region_memory_base_pfn(r);
7814 if (usable_startpfn < 0x100000) {
7815 mem_below_4gb_not_mirrored = true;
7819 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7820 min(usable_startpfn, zone_movable_pfn[nid]) :
7824 if (mem_below_4gb_not_mirrored)
7825 pr_warn("This configuration results in unmirrored kernel memory.\n");
7831 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7832 * amount of necessary memory.
7834 if (required_kernelcore_percent)
7835 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7837 if (required_movablecore_percent)
7838 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7842 * If movablecore= was specified, calculate what size of
7843 * kernelcore that corresponds so that memory usable for
7844 * any allocation type is evenly spread. If both kernelcore
7845 * and movablecore are specified, then the value of kernelcore
7846 * will be used for required_kernelcore if it's greater than
7847 * what movablecore would have allowed.
7849 if (required_movablecore) {
7850 unsigned long corepages;
7853 * Round-up so that ZONE_MOVABLE is at least as large as what
7854 * was requested by the user
7856 required_movablecore =
7857 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7858 required_movablecore = min(totalpages, required_movablecore);
7859 corepages = totalpages - required_movablecore;
7861 required_kernelcore = max(required_kernelcore, corepages);
7865 * If kernelcore was not specified or kernelcore size is larger
7866 * than totalpages, there is no ZONE_MOVABLE.
7868 if (!required_kernelcore || required_kernelcore >= totalpages)
7871 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7872 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7875 /* Spread kernelcore memory as evenly as possible throughout nodes */
7876 kernelcore_node = required_kernelcore / usable_nodes;
7877 for_each_node_state(nid, N_MEMORY) {
7878 unsigned long start_pfn, end_pfn;
7881 * Recalculate kernelcore_node if the division per node
7882 * now exceeds what is necessary to satisfy the requested
7883 * amount of memory for the kernel
7885 if (required_kernelcore < kernelcore_node)
7886 kernelcore_node = required_kernelcore / usable_nodes;
7889 * As the map is walked, we track how much memory is usable
7890 * by the kernel using kernelcore_remaining. When it is
7891 * 0, the rest of the node is usable by ZONE_MOVABLE
7893 kernelcore_remaining = kernelcore_node;
7895 /* Go through each range of PFNs within this node */
7896 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7897 unsigned long size_pages;
7899 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7900 if (start_pfn >= end_pfn)
7903 /* Account for what is only usable for kernelcore */
7904 if (start_pfn < usable_startpfn) {
7905 unsigned long kernel_pages;
7906 kernel_pages = min(end_pfn, usable_startpfn)
7909 kernelcore_remaining -= min(kernel_pages,
7910 kernelcore_remaining);
7911 required_kernelcore -= min(kernel_pages,
7912 required_kernelcore);
7914 /* Continue if range is now fully accounted */
7915 if (end_pfn <= usable_startpfn) {
7918 * Push zone_movable_pfn to the end so
7919 * that if we have to rebalance
7920 * kernelcore across nodes, we will
7921 * not double account here
7923 zone_movable_pfn[nid] = end_pfn;
7926 start_pfn = usable_startpfn;
7930 * The usable PFN range for ZONE_MOVABLE is from
7931 * start_pfn->end_pfn. Calculate size_pages as the
7932 * number of pages used as kernelcore
7934 size_pages = end_pfn - start_pfn;
7935 if (size_pages > kernelcore_remaining)
7936 size_pages = kernelcore_remaining;
7937 zone_movable_pfn[nid] = start_pfn + size_pages;
7940 * Some kernelcore has been met, update counts and
7941 * break if the kernelcore for this node has been
7944 required_kernelcore -= min(required_kernelcore,
7946 kernelcore_remaining -= size_pages;
7947 if (!kernelcore_remaining)
7953 * If there is still required_kernelcore, we do another pass with one
7954 * less node in the count. This will push zone_movable_pfn[nid] further
7955 * along on the nodes that still have memory until kernelcore is
7959 if (usable_nodes && required_kernelcore > usable_nodes)
7963 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7964 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7965 unsigned long start_pfn, end_pfn;
7967 zone_movable_pfn[nid] =
7968 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7970 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7971 if (zone_movable_pfn[nid] >= end_pfn)
7972 zone_movable_pfn[nid] = 0;
7976 /* restore the node_state */
7977 node_states[N_MEMORY] = saved_node_state;
7980 /* Any regular or high memory on that node ? */
7981 static void check_for_memory(pg_data_t *pgdat, int nid)
7983 enum zone_type zone_type;
7985 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7986 struct zone *zone = &pgdat->node_zones[zone_type];
7987 if (populated_zone(zone)) {
7988 if (IS_ENABLED(CONFIG_HIGHMEM))
7989 node_set_state(nid, N_HIGH_MEMORY);
7990 if (zone_type <= ZONE_NORMAL)
7991 node_set_state(nid, N_NORMAL_MEMORY);
7998 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7999 * such cases we allow max_zone_pfn sorted in the descending order
8001 bool __weak arch_has_descending_max_zone_pfns(void)
8007 * free_area_init - Initialise all pg_data_t and zone data
8008 * @max_zone_pfn: an array of max PFNs for each zone
8010 * This will call free_area_init_node() for each active node in the system.
8011 * Using the page ranges provided by memblock_set_node(), the size of each
8012 * zone in each node and their holes is calculated. If the maximum PFN
8013 * between two adjacent zones match, it is assumed that the zone is empty.
8014 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8015 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8016 * starts where the previous one ended. For example, ZONE_DMA32 starts
8017 * at arch_max_dma_pfn.
8019 void __init free_area_init(unsigned long *max_zone_pfn)
8021 unsigned long start_pfn, end_pfn;
8025 /* Record where the zone boundaries are */
8026 memset(arch_zone_lowest_possible_pfn, 0,
8027 sizeof(arch_zone_lowest_possible_pfn));
8028 memset(arch_zone_highest_possible_pfn, 0,
8029 sizeof(arch_zone_highest_possible_pfn));
8031 start_pfn = find_min_pfn_with_active_regions();
8032 descending = arch_has_descending_max_zone_pfns();
8034 for (i = 0; i < MAX_NR_ZONES; i++) {
8036 zone = MAX_NR_ZONES - i - 1;
8040 if (zone == ZONE_MOVABLE)
8043 end_pfn = max(max_zone_pfn[zone], start_pfn);
8044 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8045 arch_zone_highest_possible_pfn[zone] = end_pfn;
8047 start_pfn = end_pfn;
8050 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8051 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8052 find_zone_movable_pfns_for_nodes();
8054 /* Print out the zone ranges */
8055 pr_info("Zone ranges:\n");
8056 for (i = 0; i < MAX_NR_ZONES; i++) {
8057 if (i == ZONE_MOVABLE)
8059 pr_info(" %-8s ", zone_names[i]);
8060 if (arch_zone_lowest_possible_pfn[i] ==
8061 arch_zone_highest_possible_pfn[i])
8064 pr_cont("[mem %#018Lx-%#018Lx]\n",
8065 (u64)arch_zone_lowest_possible_pfn[i]
8067 ((u64)arch_zone_highest_possible_pfn[i]
8068 << PAGE_SHIFT) - 1);
8071 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8072 pr_info("Movable zone start for each node\n");
8073 for (i = 0; i < MAX_NUMNODES; i++) {
8074 if (zone_movable_pfn[i])
8075 pr_info(" Node %d: %#018Lx\n", i,
8076 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8080 * Print out the early node map, and initialize the
8081 * subsection-map relative to active online memory ranges to
8082 * enable future "sub-section" extensions of the memory map.
8084 pr_info("Early memory node ranges\n");
8085 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8086 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8087 (u64)start_pfn << PAGE_SHIFT,
8088 ((u64)end_pfn << PAGE_SHIFT) - 1);
8089 subsection_map_init(start_pfn, end_pfn - start_pfn);
8092 /* Initialise every node */
8093 mminit_verify_pageflags_layout();
8094 setup_nr_node_ids();
8095 for_each_node(nid) {
8098 if (!node_online(nid)) {
8099 pr_info("Initializing node %d as memoryless\n", nid);
8101 /* Allocator not initialized yet */
8102 pgdat = arch_alloc_nodedata(nid);
8104 pr_err("Cannot allocate %zuB for node %d.\n",
8105 sizeof(*pgdat), nid);
8108 arch_refresh_nodedata(nid, pgdat);
8109 free_area_init_memoryless_node(nid);
8112 * We do not want to confuse userspace by sysfs
8113 * files/directories for node without any memory
8114 * attached to it, so this node is not marked as
8115 * N_MEMORY and not marked online so that no sysfs
8116 * hierarchy will be created via register_one_node for
8117 * it. The pgdat will get fully initialized by
8118 * hotadd_init_pgdat() when memory is hotplugged into
8124 pgdat = NODE_DATA(nid);
8125 free_area_init_node(nid);
8127 /* Any memory on that node */
8128 if (pgdat->node_present_pages)
8129 node_set_state(nid, N_MEMORY);
8130 check_for_memory(pgdat, nid);
8136 static int __init cmdline_parse_core(char *p, unsigned long *core,
8137 unsigned long *percent)
8139 unsigned long long coremem;
8145 /* Value may be a percentage of total memory, otherwise bytes */
8146 coremem = simple_strtoull(p, &endptr, 0);
8147 if (*endptr == '%') {
8148 /* Paranoid check for percent values greater than 100 */
8149 WARN_ON(coremem > 100);
8153 coremem = memparse(p, &p);
8154 /* Paranoid check that UL is enough for the coremem value */
8155 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8157 *core = coremem >> PAGE_SHIFT;
8164 * kernelcore=size sets the amount of memory for use for allocations that
8165 * cannot be reclaimed or migrated.
8167 static int __init cmdline_parse_kernelcore(char *p)
8169 /* parse kernelcore=mirror */
8170 if (parse_option_str(p, "mirror")) {
8171 mirrored_kernelcore = true;
8175 return cmdline_parse_core(p, &required_kernelcore,
8176 &required_kernelcore_percent);
8180 * movablecore=size sets the amount of memory for use for allocations that
8181 * can be reclaimed or migrated.
8183 static int __init cmdline_parse_movablecore(char *p)
8185 return cmdline_parse_core(p, &required_movablecore,
8186 &required_movablecore_percent);
8189 early_param("kernelcore", cmdline_parse_kernelcore);
8190 early_param("movablecore", cmdline_parse_movablecore);
8192 void adjust_managed_page_count(struct page *page, long count)
8194 atomic_long_add(count, &page_zone(page)->managed_pages);
8195 totalram_pages_add(count);
8196 #ifdef CONFIG_HIGHMEM
8197 if (PageHighMem(page))
8198 totalhigh_pages_add(count);
8201 EXPORT_SYMBOL(adjust_managed_page_count);
8203 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8206 unsigned long pages = 0;
8208 start = (void *)PAGE_ALIGN((unsigned long)start);
8209 end = (void *)((unsigned long)end & PAGE_MASK);
8210 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8211 struct page *page = virt_to_page(pos);
8212 void *direct_map_addr;
8215 * 'direct_map_addr' might be different from 'pos'
8216 * because some architectures' virt_to_page()
8217 * work with aliases. Getting the direct map
8218 * address ensures that we get a _writeable_
8219 * alias for the memset().
8221 direct_map_addr = page_address(page);
8223 * Perform a kasan-unchecked memset() since this memory
8224 * has not been initialized.
8226 direct_map_addr = kasan_reset_tag(direct_map_addr);
8227 if ((unsigned int)poison <= 0xFF)
8228 memset(direct_map_addr, poison, PAGE_SIZE);
8230 free_reserved_page(page);
8234 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8239 void __init mem_init_print_info(void)
8241 unsigned long physpages, codesize, datasize, rosize, bss_size;
8242 unsigned long init_code_size, init_data_size;
8244 physpages = get_num_physpages();
8245 codesize = _etext - _stext;
8246 datasize = _edata - _sdata;
8247 rosize = __end_rodata - __start_rodata;
8248 bss_size = __bss_stop - __bss_start;
8249 init_data_size = __init_end - __init_begin;
8250 init_code_size = _einittext - _sinittext;
8253 * Detect special cases and adjust section sizes accordingly:
8254 * 1) .init.* may be embedded into .data sections
8255 * 2) .init.text.* may be out of [__init_begin, __init_end],
8256 * please refer to arch/tile/kernel/vmlinux.lds.S.
8257 * 3) .rodata.* may be embedded into .text or .data sections.
8259 #define adj_init_size(start, end, size, pos, adj) \
8261 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8265 adj_init_size(__init_begin, __init_end, init_data_size,
8266 _sinittext, init_code_size);
8267 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8268 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8269 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8270 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8272 #undef adj_init_size
8274 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8275 #ifdef CONFIG_HIGHMEM
8279 K(nr_free_pages()), K(physpages),
8280 codesize >> 10, datasize >> 10, rosize >> 10,
8281 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8282 K(physpages - totalram_pages() - totalcma_pages),
8284 #ifdef CONFIG_HIGHMEM
8285 , K(totalhigh_pages())
8291 * set_dma_reserve - set the specified number of pages reserved in the first zone
8292 * @new_dma_reserve: The number of pages to mark reserved
8294 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8295 * In the DMA zone, a significant percentage may be consumed by kernel image
8296 * and other unfreeable allocations which can skew the watermarks badly. This
8297 * function may optionally be used to account for unfreeable pages in the
8298 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8299 * smaller per-cpu batchsize.
8301 void __init set_dma_reserve(unsigned long new_dma_reserve)
8303 dma_reserve = new_dma_reserve;
8306 static int page_alloc_cpu_dead(unsigned int cpu)
8310 lru_add_drain_cpu(cpu);
8314 * Spill the event counters of the dead processor
8315 * into the current processors event counters.
8316 * This artificially elevates the count of the current
8319 vm_events_fold_cpu(cpu);
8322 * Zero the differential counters of the dead processor
8323 * so that the vm statistics are consistent.
8325 * This is only okay since the processor is dead and cannot
8326 * race with what we are doing.
8328 cpu_vm_stats_fold(cpu);
8330 for_each_populated_zone(zone)
8331 zone_pcp_update(zone, 0);
8336 static int page_alloc_cpu_online(unsigned int cpu)
8340 for_each_populated_zone(zone)
8341 zone_pcp_update(zone, 1);
8346 int hashdist = HASHDIST_DEFAULT;
8348 static int __init set_hashdist(char *str)
8352 hashdist = simple_strtoul(str, &str, 0);
8355 __setup("hashdist=", set_hashdist);
8358 void __init page_alloc_init(void)
8363 if (num_node_state(N_MEMORY) == 1)
8367 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8368 "mm/page_alloc:pcp",
8369 page_alloc_cpu_online,
8370 page_alloc_cpu_dead);
8375 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8376 * or min_free_kbytes changes.
8378 static void calculate_totalreserve_pages(void)
8380 struct pglist_data *pgdat;
8381 unsigned long reserve_pages = 0;
8382 enum zone_type i, j;
8384 for_each_online_pgdat(pgdat) {
8386 pgdat->totalreserve_pages = 0;
8388 for (i = 0; i < MAX_NR_ZONES; i++) {
8389 struct zone *zone = pgdat->node_zones + i;
8391 unsigned long managed_pages = zone_managed_pages(zone);
8393 /* Find valid and maximum lowmem_reserve in the zone */
8394 for (j = i; j < MAX_NR_ZONES; j++) {
8395 if (zone->lowmem_reserve[j] > max)
8396 max = zone->lowmem_reserve[j];
8399 /* we treat the high watermark as reserved pages. */
8400 max += high_wmark_pages(zone);
8402 if (max > managed_pages)
8403 max = managed_pages;
8405 pgdat->totalreserve_pages += max;
8407 reserve_pages += max;
8410 totalreserve_pages = reserve_pages;
8414 * setup_per_zone_lowmem_reserve - called whenever
8415 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8416 * has a correct pages reserved value, so an adequate number of
8417 * pages are left in the zone after a successful __alloc_pages().
8419 static void setup_per_zone_lowmem_reserve(void)
8421 struct pglist_data *pgdat;
8422 enum zone_type i, j;
8424 for_each_online_pgdat(pgdat) {
8425 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8426 struct zone *zone = &pgdat->node_zones[i];
8427 int ratio = sysctl_lowmem_reserve_ratio[i];
8428 bool clear = !ratio || !zone_managed_pages(zone);
8429 unsigned long managed_pages = 0;
8431 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8432 struct zone *upper_zone = &pgdat->node_zones[j];
8434 managed_pages += zone_managed_pages(upper_zone);
8437 zone->lowmem_reserve[j] = 0;
8439 zone->lowmem_reserve[j] = managed_pages / ratio;
8444 /* update totalreserve_pages */
8445 calculate_totalreserve_pages();
8448 static void __setup_per_zone_wmarks(void)
8450 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8451 unsigned long lowmem_pages = 0;
8453 unsigned long flags;
8455 /* Calculate total number of !ZONE_HIGHMEM pages */
8456 for_each_zone(zone) {
8457 if (!is_highmem(zone))
8458 lowmem_pages += zone_managed_pages(zone);
8461 for_each_zone(zone) {
8464 spin_lock_irqsave(&zone->lock, flags);
8465 tmp = (u64)pages_min * zone_managed_pages(zone);
8466 do_div(tmp, lowmem_pages);
8467 if (is_highmem(zone)) {
8469 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8470 * need highmem pages, so cap pages_min to a small
8473 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8474 * deltas control async page reclaim, and so should
8475 * not be capped for highmem.
8477 unsigned long min_pages;
8479 min_pages = zone_managed_pages(zone) / 1024;
8480 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8481 zone->_watermark[WMARK_MIN] = min_pages;
8484 * If it's a lowmem zone, reserve a number of pages
8485 * proportionate to the zone's size.
8487 zone->_watermark[WMARK_MIN] = tmp;
8491 * Set the kswapd watermarks distance according to the
8492 * scale factor in proportion to available memory, but
8493 * ensure a minimum size on small systems.
8495 tmp = max_t(u64, tmp >> 2,
8496 mult_frac(zone_managed_pages(zone),
8497 watermark_scale_factor, 10000));
8499 zone->watermark_boost = 0;
8500 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8501 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8502 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8504 spin_unlock_irqrestore(&zone->lock, flags);
8507 /* update totalreserve_pages */
8508 calculate_totalreserve_pages();
8512 * setup_per_zone_wmarks - called when min_free_kbytes changes
8513 * or when memory is hot-{added|removed}
8515 * Ensures that the watermark[min,low,high] values for each zone are set
8516 * correctly with respect to min_free_kbytes.
8518 void setup_per_zone_wmarks(void)
8521 static DEFINE_SPINLOCK(lock);
8524 __setup_per_zone_wmarks();
8528 * The watermark size have changed so update the pcpu batch
8529 * and high limits or the limits may be inappropriate.
8532 zone_pcp_update(zone, 0);
8536 * Initialise min_free_kbytes.
8538 * For small machines we want it small (128k min). For large machines
8539 * we want it large (256MB max). But it is not linear, because network
8540 * bandwidth does not increase linearly with machine size. We use
8542 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8543 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8559 void calculate_min_free_kbytes(void)
8561 unsigned long lowmem_kbytes;
8562 int new_min_free_kbytes;
8564 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8565 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8567 if (new_min_free_kbytes > user_min_free_kbytes)
8568 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8570 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8571 new_min_free_kbytes, user_min_free_kbytes);
8575 int __meminit init_per_zone_wmark_min(void)
8577 calculate_min_free_kbytes();
8578 setup_per_zone_wmarks();
8579 refresh_zone_stat_thresholds();
8580 setup_per_zone_lowmem_reserve();
8583 setup_min_unmapped_ratio();
8584 setup_min_slab_ratio();
8587 khugepaged_min_free_kbytes_update();
8591 postcore_initcall(init_per_zone_wmark_min)
8594 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8595 * that we can call two helper functions whenever min_free_kbytes
8598 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8599 void *buffer, size_t *length, loff_t *ppos)
8603 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8608 user_min_free_kbytes = min_free_kbytes;
8609 setup_per_zone_wmarks();
8614 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8615 void *buffer, size_t *length, loff_t *ppos)
8619 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8624 setup_per_zone_wmarks();
8630 static void setup_min_unmapped_ratio(void)
8635 for_each_online_pgdat(pgdat)
8636 pgdat->min_unmapped_pages = 0;
8639 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8640 sysctl_min_unmapped_ratio) / 100;
8644 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8645 void *buffer, size_t *length, loff_t *ppos)
8649 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8653 setup_min_unmapped_ratio();
8658 static void setup_min_slab_ratio(void)
8663 for_each_online_pgdat(pgdat)
8664 pgdat->min_slab_pages = 0;
8667 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8668 sysctl_min_slab_ratio) / 100;
8671 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8672 void *buffer, size_t *length, loff_t *ppos)
8676 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8680 setup_min_slab_ratio();
8687 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8688 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8689 * whenever sysctl_lowmem_reserve_ratio changes.
8691 * The reserve ratio obviously has absolutely no relation with the
8692 * minimum watermarks. The lowmem reserve ratio can only make sense
8693 * if in function of the boot time zone sizes.
8695 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8696 void *buffer, size_t *length, loff_t *ppos)
8700 proc_dointvec_minmax(table, write, buffer, length, ppos);
8702 for (i = 0; i < MAX_NR_ZONES; i++) {
8703 if (sysctl_lowmem_reserve_ratio[i] < 1)
8704 sysctl_lowmem_reserve_ratio[i] = 0;
8707 setup_per_zone_lowmem_reserve();
8712 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8713 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8714 * pagelist can have before it gets flushed back to buddy allocator.
8716 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8717 int write, void *buffer, size_t *length, loff_t *ppos)
8720 int old_percpu_pagelist_high_fraction;
8723 mutex_lock(&pcp_batch_high_lock);
8724 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8726 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8727 if (!write || ret < 0)
8730 /* Sanity checking to avoid pcp imbalance */
8731 if (percpu_pagelist_high_fraction &&
8732 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8733 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8739 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8742 for_each_populated_zone(zone)
8743 zone_set_pageset_high_and_batch(zone, 0);
8745 mutex_unlock(&pcp_batch_high_lock);
8749 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8751 * Returns the number of pages that arch has reserved but
8752 * is not known to alloc_large_system_hash().
8754 static unsigned long __init arch_reserved_kernel_pages(void)
8761 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8762 * machines. As memory size is increased the scale is also increased but at
8763 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8764 * quadruples the scale is increased by one, which means the size of hash table
8765 * only doubles, instead of quadrupling as well.
8766 * Because 32-bit systems cannot have large physical memory, where this scaling
8767 * makes sense, it is disabled on such platforms.
8769 #if __BITS_PER_LONG > 32
8770 #define ADAPT_SCALE_BASE (64ul << 30)
8771 #define ADAPT_SCALE_SHIFT 2
8772 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8776 * allocate a large system hash table from bootmem
8777 * - it is assumed that the hash table must contain an exact power-of-2
8778 * quantity of entries
8779 * - limit is the number of hash buckets, not the total allocation size
8781 void *__init alloc_large_system_hash(const char *tablename,
8782 unsigned long bucketsize,
8783 unsigned long numentries,
8786 unsigned int *_hash_shift,
8787 unsigned int *_hash_mask,
8788 unsigned long low_limit,
8789 unsigned long high_limit)
8791 unsigned long long max = high_limit;
8792 unsigned long log2qty, size;
8798 /* allow the kernel cmdline to have a say */
8800 /* round applicable memory size up to nearest megabyte */
8801 numentries = nr_kernel_pages;
8802 numentries -= arch_reserved_kernel_pages();
8804 /* It isn't necessary when PAGE_SIZE >= 1MB */
8805 if (PAGE_SHIFT < 20)
8806 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8808 #if __BITS_PER_LONG > 32
8810 unsigned long adapt;
8812 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8813 adapt <<= ADAPT_SCALE_SHIFT)
8818 /* limit to 1 bucket per 2^scale bytes of low memory */
8819 if (scale > PAGE_SHIFT)
8820 numentries >>= (scale - PAGE_SHIFT);
8822 numentries <<= (PAGE_SHIFT - scale);
8824 /* Make sure we've got at least a 0-order allocation.. */
8825 if (unlikely(flags & HASH_SMALL)) {
8826 /* Makes no sense without HASH_EARLY */
8827 WARN_ON(!(flags & HASH_EARLY));
8828 if (!(numentries >> *_hash_shift)) {
8829 numentries = 1UL << *_hash_shift;
8830 BUG_ON(!numentries);
8832 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8833 numentries = PAGE_SIZE / bucketsize;
8835 numentries = roundup_pow_of_two(numentries);
8837 /* limit allocation size to 1/16 total memory by default */
8839 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8840 do_div(max, bucketsize);
8842 max = min(max, 0x80000000ULL);
8844 if (numentries < low_limit)
8845 numentries = low_limit;
8846 if (numentries > max)
8849 log2qty = ilog2(numentries);
8851 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8854 size = bucketsize << log2qty;
8855 if (flags & HASH_EARLY) {
8856 if (flags & HASH_ZERO)
8857 table = memblock_alloc(size, SMP_CACHE_BYTES);
8859 table = memblock_alloc_raw(size,
8861 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8862 table = __vmalloc(size, gfp_flags);
8865 huge = is_vm_area_hugepages(table);
8868 * If bucketsize is not a power-of-two, we may free
8869 * some pages at the end of hash table which
8870 * alloc_pages_exact() automatically does
8872 table = alloc_pages_exact(size, gfp_flags);
8873 kmemleak_alloc(table, size, 1, gfp_flags);
8875 } while (!table && size > PAGE_SIZE && --log2qty);
8878 panic("Failed to allocate %s hash table\n", tablename);
8880 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8881 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8882 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8885 *_hash_shift = log2qty;
8887 *_hash_mask = (1 << log2qty) - 1;
8893 * This function checks whether pageblock includes unmovable pages or not.
8895 * PageLRU check without isolation or lru_lock could race so that
8896 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8897 * check without lock_page also may miss some movable non-lru pages at
8898 * race condition. So you can't expect this function should be exact.
8900 * Returns a page without holding a reference. If the caller wants to
8901 * dereference that page (e.g., dumping), it has to make sure that it
8902 * cannot get removed (e.g., via memory unplug) concurrently.
8905 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8906 int migratetype, int flags)
8908 unsigned long iter = 0;
8909 unsigned long pfn = page_to_pfn(page);
8910 unsigned long offset = pfn % pageblock_nr_pages;
8912 if (is_migrate_cma_page(page)) {
8914 * CMA allocations (alloc_contig_range) really need to mark
8915 * isolate CMA pageblocks even when they are not movable in fact
8916 * so consider them movable here.
8918 if (is_migrate_cma(migratetype))
8924 for (; iter < pageblock_nr_pages - offset; iter++) {
8925 page = pfn_to_page(pfn + iter);
8928 * Both, bootmem allocations and memory holes are marked
8929 * PG_reserved and are unmovable. We can even have unmovable
8930 * allocations inside ZONE_MOVABLE, for example when
8931 * specifying "movablecore".
8933 if (PageReserved(page))
8937 * If the zone is movable and we have ruled out all reserved
8938 * pages then it should be reasonably safe to assume the rest
8941 if (zone_idx(zone) == ZONE_MOVABLE)
8945 * Hugepages are not in LRU lists, but they're movable.
8946 * THPs are on the LRU, but need to be counted as #small pages.
8947 * We need not scan over tail pages because we don't
8948 * handle each tail page individually in migration.
8950 if (PageHuge(page) || PageTransCompound(page)) {
8951 struct page *head = compound_head(page);
8952 unsigned int skip_pages;
8954 if (PageHuge(page)) {
8955 if (!hugepage_migration_supported(page_hstate(head)))
8957 } else if (!PageLRU(head) && !__PageMovable(head)) {
8961 skip_pages = compound_nr(head) - (page - head);
8962 iter += skip_pages - 1;
8967 * We can't use page_count without pin a page
8968 * because another CPU can free compound page.
8969 * This check already skips compound tails of THP
8970 * because their page->_refcount is zero at all time.
8972 if (!page_ref_count(page)) {
8973 if (PageBuddy(page))
8974 iter += (1 << buddy_order(page)) - 1;
8979 * The HWPoisoned page may be not in buddy system, and
8980 * page_count() is not 0.
8982 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8986 * We treat all PageOffline() pages as movable when offlining
8987 * to give drivers a chance to decrement their reference count
8988 * in MEM_GOING_OFFLINE in order to indicate that these pages
8989 * can be offlined as there are no direct references anymore.
8990 * For actually unmovable PageOffline() where the driver does
8991 * not support this, we will fail later when trying to actually
8992 * move these pages that still have a reference count > 0.
8993 * (false negatives in this function only)
8995 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8998 if (__PageMovable(page) || PageLRU(page))
9002 * If there are RECLAIMABLE pages, we need to check
9003 * it. But now, memory offline itself doesn't call
9004 * shrink_node_slabs() and it still to be fixed.
9011 #ifdef CONFIG_CONTIG_ALLOC
9012 static unsigned long pfn_max_align_down(unsigned long pfn)
9014 return ALIGN_DOWN(pfn, MAX_ORDER_NR_PAGES);
9017 static unsigned long pfn_max_align_up(unsigned long pfn)
9019 return ALIGN(pfn, MAX_ORDER_NR_PAGES);
9022 #if defined(CONFIG_DYNAMIC_DEBUG) || \
9023 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
9024 /* Usage: See admin-guide/dynamic-debug-howto.rst */
9025 static void alloc_contig_dump_pages(struct list_head *page_list)
9027 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9029 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9033 list_for_each_entry(page, page_list, lru)
9034 dump_page(page, "migration failure");
9038 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9043 /* [start, end) must belong to a single zone. */
9044 static int __alloc_contig_migrate_range(struct compact_control *cc,
9045 unsigned long start, unsigned long end)
9047 /* This function is based on compact_zone() from compaction.c. */
9048 unsigned int nr_reclaimed;
9049 unsigned long pfn = start;
9050 unsigned int tries = 0;
9052 struct migration_target_control mtc = {
9053 .nid = zone_to_nid(cc->zone),
9054 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9057 lru_cache_disable();
9059 while (pfn < end || !list_empty(&cc->migratepages)) {
9060 if (fatal_signal_pending(current)) {
9065 if (list_empty(&cc->migratepages)) {
9066 cc->nr_migratepages = 0;
9067 ret = isolate_migratepages_range(cc, pfn, end);
9068 if (ret && ret != -EAGAIN)
9070 pfn = cc->migrate_pfn;
9072 } else if (++tries == 5) {
9077 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9079 cc->nr_migratepages -= nr_reclaimed;
9081 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9082 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9085 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9086 * to retry again over this error, so do the same here.
9095 alloc_contig_dump_pages(&cc->migratepages);
9096 putback_movable_pages(&cc->migratepages);
9103 * alloc_contig_range() -- tries to allocate given range of pages
9104 * @start: start PFN to allocate
9105 * @end: one-past-the-last PFN to allocate
9106 * @migratetype: migratetype of the underlying pageblocks (either
9107 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9108 * in range must have the same migratetype and it must
9109 * be either of the two.
9110 * @gfp_mask: GFP mask to use during compaction
9112 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9113 * aligned. The PFN range must belong to a single zone.
9115 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9116 * pageblocks in the range. Once isolated, the pageblocks should not
9117 * be modified by others.
9119 * Return: zero on success or negative error code. On success all
9120 * pages which PFN is in [start, end) are allocated for the caller and
9121 * need to be freed with free_contig_range().
9123 int alloc_contig_range(unsigned long start, unsigned long end,
9124 unsigned migratetype, gfp_t gfp_mask)
9126 unsigned long outer_start, outer_end;
9130 struct compact_control cc = {
9131 .nr_migratepages = 0,
9133 .zone = page_zone(pfn_to_page(start)),
9134 .mode = MIGRATE_SYNC,
9135 .ignore_skip_hint = true,
9136 .no_set_skip_hint = true,
9137 .gfp_mask = current_gfp_context(gfp_mask),
9138 .alloc_contig = true,
9140 INIT_LIST_HEAD(&cc.migratepages);
9143 * What we do here is we mark all pageblocks in range as
9144 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9145 * have different sizes, and due to the way page allocator
9146 * work, we align the range to biggest of the two pages so
9147 * that page allocator won't try to merge buddies from
9148 * different pageblocks and change MIGRATE_ISOLATE to some
9149 * other migration type.
9151 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9152 * migrate the pages from an unaligned range (ie. pages that
9153 * we are interested in). This will put all the pages in
9154 * range back to page allocator as MIGRATE_ISOLATE.
9156 * When this is done, we take the pages in range from page
9157 * allocator removing them from the buddy system. This way
9158 * page allocator will never consider using them.
9160 * This lets us mark the pageblocks back as
9161 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9162 * aligned range but not in the unaligned, original range are
9163 * put back to page allocator so that buddy can use them.
9166 ret = start_isolate_page_range(pfn_max_align_down(start),
9167 pfn_max_align_up(end), migratetype, 0);
9171 drain_all_pages(cc.zone);
9174 * In case of -EBUSY, we'd like to know which page causes problem.
9175 * So, just fall through. test_pages_isolated() has a tracepoint
9176 * which will report the busy page.
9178 * It is possible that busy pages could become available before
9179 * the call to test_pages_isolated, and the range will actually be
9180 * allocated. So, if we fall through be sure to clear ret so that
9181 * -EBUSY is not accidentally used or returned to caller.
9183 ret = __alloc_contig_migrate_range(&cc, start, end);
9184 if (ret && ret != -EBUSY)
9189 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9190 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9191 * more, all pages in [start, end) are free in page allocator.
9192 * What we are going to do is to allocate all pages from
9193 * [start, end) (that is remove them from page allocator).
9195 * The only problem is that pages at the beginning and at the
9196 * end of interesting range may be not aligned with pages that
9197 * page allocator holds, ie. they can be part of higher order
9198 * pages. Because of this, we reserve the bigger range and
9199 * once this is done free the pages we are not interested in.
9201 * We don't have to hold zone->lock here because the pages are
9202 * isolated thus they won't get removed from buddy.
9206 outer_start = start;
9207 while (!PageBuddy(pfn_to_page(outer_start))) {
9208 if (++order >= MAX_ORDER) {
9209 outer_start = start;
9212 outer_start &= ~0UL << order;
9215 if (outer_start != start) {
9216 order = buddy_order(pfn_to_page(outer_start));
9219 * outer_start page could be small order buddy page and
9220 * it doesn't include start page. Adjust outer_start
9221 * in this case to report failed page properly
9222 * on tracepoint in test_pages_isolated()
9224 if (outer_start + (1UL << order) <= start)
9225 outer_start = start;
9228 /* Make sure the range is really isolated. */
9229 if (test_pages_isolated(outer_start, end, 0)) {
9234 /* Grab isolated pages from freelists. */
9235 outer_end = isolate_freepages_range(&cc, outer_start, end);
9241 /* Free head and tail (if any) */
9242 if (start != outer_start)
9243 free_contig_range(outer_start, start - outer_start);
9244 if (end != outer_end)
9245 free_contig_range(end, outer_end - end);
9248 undo_isolate_page_range(pfn_max_align_down(start),
9249 pfn_max_align_up(end), migratetype);
9252 EXPORT_SYMBOL(alloc_contig_range);
9254 static int __alloc_contig_pages(unsigned long start_pfn,
9255 unsigned long nr_pages, gfp_t gfp_mask)
9257 unsigned long end_pfn = start_pfn + nr_pages;
9259 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9263 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9264 unsigned long nr_pages)
9266 unsigned long i, end_pfn = start_pfn + nr_pages;
9269 for (i = start_pfn; i < end_pfn; i++) {
9270 page = pfn_to_online_page(i);
9274 if (page_zone(page) != z)
9277 if (PageReserved(page))
9283 static bool zone_spans_last_pfn(const struct zone *zone,
9284 unsigned long start_pfn, unsigned long nr_pages)
9286 unsigned long last_pfn = start_pfn + nr_pages - 1;
9288 return zone_spans_pfn(zone, last_pfn);
9292 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9293 * @nr_pages: Number of contiguous pages to allocate
9294 * @gfp_mask: GFP mask to limit search and used during compaction
9296 * @nodemask: Mask for other possible nodes
9298 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9299 * on an applicable zonelist to find a contiguous pfn range which can then be
9300 * tried for allocation with alloc_contig_range(). This routine is intended
9301 * for allocation requests which can not be fulfilled with the buddy allocator.
9303 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9304 * power of two, then allocated range is also guaranteed to be aligned to same
9305 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9307 * Allocated pages can be freed with free_contig_range() or by manually calling
9308 * __free_page() on each allocated page.
9310 * Return: pointer to contiguous pages on success, or NULL if not successful.
9312 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9313 int nid, nodemask_t *nodemask)
9315 unsigned long ret, pfn, flags;
9316 struct zonelist *zonelist;
9320 zonelist = node_zonelist(nid, gfp_mask);
9321 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9322 gfp_zone(gfp_mask), nodemask) {
9323 spin_lock_irqsave(&zone->lock, flags);
9325 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9326 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9327 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9329 * We release the zone lock here because
9330 * alloc_contig_range() will also lock the zone
9331 * at some point. If there's an allocation
9332 * spinning on this lock, it may win the race
9333 * and cause alloc_contig_range() to fail...
9335 spin_unlock_irqrestore(&zone->lock, flags);
9336 ret = __alloc_contig_pages(pfn, nr_pages,
9339 return pfn_to_page(pfn);
9340 spin_lock_irqsave(&zone->lock, flags);
9344 spin_unlock_irqrestore(&zone->lock, flags);
9348 #endif /* CONFIG_CONTIG_ALLOC */
9350 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9352 unsigned long count = 0;
9354 for (; nr_pages--; pfn++) {
9355 struct page *page = pfn_to_page(pfn);
9357 count += page_count(page) != 1;
9360 WARN(count != 0, "%lu pages are still in use!\n", count);
9362 EXPORT_SYMBOL(free_contig_range);
9365 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9366 * page high values need to be recalculated.
9368 void zone_pcp_update(struct zone *zone, int cpu_online)
9370 mutex_lock(&pcp_batch_high_lock);
9371 zone_set_pageset_high_and_batch(zone, cpu_online);
9372 mutex_unlock(&pcp_batch_high_lock);
9376 * Effectively disable pcplists for the zone by setting the high limit to 0
9377 * and draining all cpus. A concurrent page freeing on another CPU that's about
9378 * to put the page on pcplist will either finish before the drain and the page
9379 * will be drained, or observe the new high limit and skip the pcplist.
9381 * Must be paired with a call to zone_pcp_enable().
9383 void zone_pcp_disable(struct zone *zone)
9385 mutex_lock(&pcp_batch_high_lock);
9386 __zone_set_pageset_high_and_batch(zone, 0, 1);
9387 __drain_all_pages(zone, true);
9390 void zone_pcp_enable(struct zone *zone)
9392 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9393 mutex_unlock(&pcp_batch_high_lock);
9396 void zone_pcp_reset(struct zone *zone)
9399 struct per_cpu_zonestat *pzstats;
9401 if (zone->per_cpu_pageset != &boot_pageset) {
9402 for_each_online_cpu(cpu) {
9403 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9404 drain_zonestat(zone, pzstats);
9406 free_percpu(zone->per_cpu_pageset);
9407 free_percpu(zone->per_cpu_zonestats);
9408 zone->per_cpu_pageset = &boot_pageset;
9409 zone->per_cpu_zonestats = &boot_zonestats;
9413 #ifdef CONFIG_MEMORY_HOTREMOVE
9415 * All pages in the range must be in a single zone, must not contain holes,
9416 * must span full sections, and must be isolated before calling this function.
9418 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9420 unsigned long pfn = start_pfn;
9424 unsigned long flags;
9426 offline_mem_sections(pfn, end_pfn);
9427 zone = page_zone(pfn_to_page(pfn));
9428 spin_lock_irqsave(&zone->lock, flags);
9429 while (pfn < end_pfn) {
9430 page = pfn_to_page(pfn);
9432 * The HWPoisoned page may be not in buddy system, and
9433 * page_count() is not 0.
9435 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9440 * At this point all remaining PageOffline() pages have a
9441 * reference count of 0 and can simply be skipped.
9443 if (PageOffline(page)) {
9444 BUG_ON(page_count(page));
9445 BUG_ON(PageBuddy(page));
9450 BUG_ON(page_count(page));
9451 BUG_ON(!PageBuddy(page));
9452 order = buddy_order(page);
9453 del_page_from_free_list(page, zone, order);
9454 pfn += (1 << order);
9456 spin_unlock_irqrestore(&zone->lock, flags);
9461 * This function returns a stable result only if called under zone lock.
9463 bool is_free_buddy_page(struct page *page)
9465 unsigned long pfn = page_to_pfn(page);
9468 for (order = 0; order < MAX_ORDER; order++) {
9469 struct page *page_head = page - (pfn & ((1 << order) - 1));
9471 if (PageBuddy(page_head) &&
9472 buddy_order_unsafe(page_head) >= order)
9476 return order < MAX_ORDER;
9478 EXPORT_SYMBOL(is_free_buddy_page);
9480 #ifdef CONFIG_MEMORY_FAILURE
9482 * Break down a higher-order page in sub-pages, and keep our target out of
9485 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9486 struct page *target, int low, int high,
9489 unsigned long size = 1 << high;
9490 struct page *current_buddy, *next_page;
9492 while (high > low) {
9496 if (target >= &page[size]) {
9497 next_page = page + size;
9498 current_buddy = page;
9501 current_buddy = page + size;
9504 if (set_page_guard(zone, current_buddy, high, migratetype))
9507 if (current_buddy != target) {
9508 add_to_free_list(current_buddy, zone, high, migratetype);
9509 set_buddy_order(current_buddy, high);
9516 * Take a page that will be marked as poisoned off the buddy allocator.
9518 bool take_page_off_buddy(struct page *page)
9520 struct zone *zone = page_zone(page);
9521 unsigned long pfn = page_to_pfn(page);
9522 unsigned long flags;
9526 spin_lock_irqsave(&zone->lock, flags);
9527 for (order = 0; order < MAX_ORDER; order++) {
9528 struct page *page_head = page - (pfn & ((1 << order) - 1));
9529 int page_order = buddy_order(page_head);
9531 if (PageBuddy(page_head) && page_order >= order) {
9532 unsigned long pfn_head = page_to_pfn(page_head);
9533 int migratetype = get_pfnblock_migratetype(page_head,
9536 del_page_from_free_list(page_head, zone, page_order);
9537 break_down_buddy_pages(zone, page_head, page, 0,
9538 page_order, migratetype);
9539 SetPageHWPoisonTakenOff(page);
9540 if (!is_migrate_isolate(migratetype))
9541 __mod_zone_freepage_state(zone, -1, migratetype);
9545 if (page_count(page_head) > 0)
9548 spin_unlock_irqrestore(&zone->lock, flags);
9553 * Cancel takeoff done by take_page_off_buddy().
9555 bool put_page_back_buddy(struct page *page)
9557 struct zone *zone = page_zone(page);
9558 unsigned long pfn = page_to_pfn(page);
9559 unsigned long flags;
9560 int migratetype = get_pfnblock_migratetype(page, pfn);
9563 spin_lock_irqsave(&zone->lock, flags);
9564 if (put_page_testzero(page)) {
9565 ClearPageHWPoisonTakenOff(page);
9566 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9567 if (TestClearPageHWPoison(page)) {
9568 num_poisoned_pages_dec();
9572 spin_unlock_irqrestore(&zone->lock, flags);
9578 #ifdef CONFIG_ZONE_DMA
9579 bool has_managed_dma(void)
9581 struct pglist_data *pgdat;
9583 for_each_online_pgdat(pgdat) {
9584 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9586 if (managed_zone(zone))
9591 #endif /* CONFIG_ZONE_DMA */