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 <asm/sections.h>
78 #include <asm/tlbflush.h>
79 #include <asm/div64.h>
82 #include "page_reporting.h"
84 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
85 typedef int __bitwise fpi_t;
87 /* No special request */
88 #define FPI_NONE ((__force fpi_t)0)
91 * Skip free page reporting notification for the (possibly merged) page.
92 * This does not hinder free page reporting from grabbing the page,
93 * reporting it and marking it "reported" - it only skips notifying
94 * the free page reporting infrastructure about a newly freed page. For
95 * example, used when temporarily pulling a page from a freelist and
96 * putting it back unmodified.
98 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
101 * Place the (possibly merged) page to the tail of the freelist. Will ignore
102 * page shuffling (relevant code - e.g., memory onlining - is expected to
103 * shuffle the whole zone).
105 * Note: No code should rely on this flag for correctness - it's purely
106 * to allow for optimizations when handing back either fresh pages
107 * (memory onlining) or untouched pages (page isolation, free page
110 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
113 * Don't poison memory with KASAN (only for the tag-based modes).
114 * During boot, all non-reserved memblock memory is exposed to page_alloc.
115 * Poisoning all that memory lengthens boot time, especially on systems with
116 * large amount of RAM. This flag is used to skip that poisoning.
117 * This is only done for the tag-based KASAN modes, as those are able to
118 * detect memory corruptions with the memory tags assigned by default.
119 * All memory allocated normally after boot gets poisoned as usual.
121 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
123 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
124 static DEFINE_MUTEX(pcp_batch_high_lock);
125 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
130 static DEFINE_PER_CPU(struct pagesets, pagesets) = {
131 .lock = INIT_LOCAL_LOCK(lock),
134 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
135 DEFINE_PER_CPU(int, numa_node);
136 EXPORT_PER_CPU_SYMBOL(numa_node);
139 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
141 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
143 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
144 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
145 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
146 * defined in <linux/topology.h>.
148 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
149 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
152 /* work_structs for global per-cpu drains */
155 struct work_struct work;
157 static DEFINE_MUTEX(pcpu_drain_mutex);
158 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
160 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
161 volatile unsigned long latent_entropy __latent_entropy;
162 EXPORT_SYMBOL(latent_entropy);
166 * Array of node states.
168 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
169 [N_POSSIBLE] = NODE_MASK_ALL,
170 [N_ONLINE] = { { [0] = 1UL } },
172 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
173 #ifdef CONFIG_HIGHMEM
174 [N_HIGH_MEMORY] = { { [0] = 1UL } },
176 [N_MEMORY] = { { [0] = 1UL } },
177 [N_CPU] = { { [0] = 1UL } },
180 EXPORT_SYMBOL(node_states);
182 atomic_long_t _totalram_pages __read_mostly;
183 EXPORT_SYMBOL(_totalram_pages);
184 unsigned long totalreserve_pages __read_mostly;
185 unsigned long totalcma_pages __read_mostly;
187 int percpu_pagelist_high_fraction;
188 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
189 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
190 EXPORT_SYMBOL(init_on_alloc);
192 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
193 EXPORT_SYMBOL(init_on_free);
195 static bool _init_on_alloc_enabled_early __read_mostly
196 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
197 static int __init early_init_on_alloc(char *buf)
200 return kstrtobool(buf, &_init_on_alloc_enabled_early);
202 early_param("init_on_alloc", early_init_on_alloc);
204 static bool _init_on_free_enabled_early __read_mostly
205 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
206 static int __init early_init_on_free(char *buf)
208 return kstrtobool(buf, &_init_on_free_enabled_early);
210 early_param("init_on_free", early_init_on_free);
213 * A cached value of the page's pageblock's migratetype, used when the page is
214 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
215 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
216 * Also the migratetype set in the page does not necessarily match the pcplist
217 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
218 * other index - this ensures that it will be put on the correct CMA freelist.
220 static inline int get_pcppage_migratetype(struct page *page)
225 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
227 page->index = migratetype;
230 #ifdef CONFIG_PM_SLEEP
232 * The following functions are used by the suspend/hibernate code to temporarily
233 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
234 * while devices are suspended. To avoid races with the suspend/hibernate code,
235 * they should always be called with system_transition_mutex held
236 * (gfp_allowed_mask also should only be modified with system_transition_mutex
237 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
238 * with that modification).
241 static gfp_t saved_gfp_mask;
243 void pm_restore_gfp_mask(void)
245 WARN_ON(!mutex_is_locked(&system_transition_mutex));
246 if (saved_gfp_mask) {
247 gfp_allowed_mask = saved_gfp_mask;
252 void pm_restrict_gfp_mask(void)
254 WARN_ON(!mutex_is_locked(&system_transition_mutex));
255 WARN_ON(saved_gfp_mask);
256 saved_gfp_mask = gfp_allowed_mask;
257 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
260 bool pm_suspended_storage(void)
262 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
266 #endif /* CONFIG_PM_SLEEP */
268 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
269 unsigned int pageblock_order __read_mostly;
272 static void __free_pages_ok(struct page *page, unsigned int order,
276 * results with 256, 32 in the lowmem_reserve sysctl:
277 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
278 * 1G machine -> (16M dma, 784M normal, 224M high)
279 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
280 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
281 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
283 * TBD: should special case ZONE_DMA32 machines here - in those we normally
284 * don't need any ZONE_NORMAL reservation
286 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
287 #ifdef CONFIG_ZONE_DMA
290 #ifdef CONFIG_ZONE_DMA32
294 #ifdef CONFIG_HIGHMEM
300 static char * const zone_names[MAX_NR_ZONES] = {
301 #ifdef CONFIG_ZONE_DMA
304 #ifdef CONFIG_ZONE_DMA32
308 #ifdef CONFIG_HIGHMEM
312 #ifdef CONFIG_ZONE_DEVICE
317 const char * const migratetype_names[MIGRATE_TYPES] = {
325 #ifdef CONFIG_MEMORY_ISOLATION
330 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
331 [NULL_COMPOUND_DTOR] = NULL,
332 [COMPOUND_PAGE_DTOR] = free_compound_page,
333 #ifdef CONFIG_HUGETLB_PAGE
334 [HUGETLB_PAGE_DTOR] = free_huge_page,
336 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
337 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
341 int min_free_kbytes = 1024;
342 int user_min_free_kbytes = -1;
343 int watermark_boost_factor __read_mostly = 15000;
344 int watermark_scale_factor = 10;
346 static unsigned long nr_kernel_pages __initdata;
347 static unsigned long nr_all_pages __initdata;
348 static unsigned long dma_reserve __initdata;
350 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
351 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
352 static unsigned long required_kernelcore __initdata;
353 static unsigned long required_kernelcore_percent __initdata;
354 static unsigned long required_movablecore __initdata;
355 static unsigned long required_movablecore_percent __initdata;
356 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
357 static bool mirrored_kernelcore __meminitdata;
359 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
361 EXPORT_SYMBOL(movable_zone);
364 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
365 unsigned int nr_online_nodes __read_mostly = 1;
366 EXPORT_SYMBOL(nr_node_ids);
367 EXPORT_SYMBOL(nr_online_nodes);
370 int page_group_by_mobility_disabled __read_mostly;
372 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
374 * During boot we initialize deferred pages on-demand, as needed, but once
375 * page_alloc_init_late() has finished, the deferred pages are all initialized,
376 * and we can permanently disable that path.
378 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
381 * Calling kasan_poison_pages() only after deferred memory initialization
382 * has completed. Poisoning pages during deferred memory init will greatly
383 * lengthen the process and cause problem in large memory systems as the
384 * deferred pages initialization is done with interrupt disabled.
386 * Assuming that there will be no reference to those newly initialized
387 * pages before they are ever allocated, this should have no effect on
388 * KASAN memory tracking as the poison will be properly inserted at page
389 * allocation time. The only corner case is when pages are allocated by
390 * on-demand allocation and then freed again before the deferred pages
391 * initialization is done, but this is not likely to happen.
393 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
395 return static_branch_unlikely(&deferred_pages) ||
396 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
397 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
398 PageSkipKASanPoison(page);
401 /* Returns true if the struct page for the pfn is uninitialised */
402 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
404 int nid = early_pfn_to_nid(pfn);
406 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
413 * Returns true when the remaining initialisation should be deferred until
414 * later in the boot cycle when it can be parallelised.
416 static bool __meminit
417 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
419 static unsigned long prev_end_pfn, nr_initialised;
422 * prev_end_pfn static that contains the end of previous zone
423 * No need to protect because called very early in boot before smp_init.
425 if (prev_end_pfn != end_pfn) {
426 prev_end_pfn = end_pfn;
430 /* Always populate low zones for address-constrained allocations */
431 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
434 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
437 * We start only with one section of pages, more pages are added as
438 * needed until the rest of deferred pages are initialized.
441 if ((nr_initialised > PAGES_PER_SECTION) &&
442 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
443 NODE_DATA(nid)->first_deferred_pfn = pfn;
449 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
451 return (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
452 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
453 PageSkipKASanPoison(page);
456 static inline bool early_page_uninitialised(unsigned long pfn)
461 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
467 /* Return a pointer to the bitmap storing bits affecting a block of pages */
468 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
471 #ifdef CONFIG_SPARSEMEM
472 return section_to_usemap(__pfn_to_section(pfn));
474 return page_zone(page)->pageblock_flags;
475 #endif /* CONFIG_SPARSEMEM */
478 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
480 #ifdef CONFIG_SPARSEMEM
481 pfn &= (PAGES_PER_SECTION-1);
483 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
484 #endif /* CONFIG_SPARSEMEM */
485 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
488 static __always_inline
489 unsigned long __get_pfnblock_flags_mask(const struct page *page,
493 unsigned long *bitmap;
494 unsigned long bitidx, word_bitidx;
497 bitmap = get_pageblock_bitmap(page, pfn);
498 bitidx = pfn_to_bitidx(page, pfn);
499 word_bitidx = bitidx / BITS_PER_LONG;
500 bitidx &= (BITS_PER_LONG-1);
502 word = bitmap[word_bitidx];
503 return (word >> bitidx) & mask;
507 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
508 * @page: The page within the block of interest
509 * @pfn: The target page frame number
510 * @mask: mask of bits that the caller is interested in
512 * Return: pageblock_bits flags
514 unsigned long get_pfnblock_flags_mask(const struct page *page,
515 unsigned long pfn, unsigned long mask)
517 return __get_pfnblock_flags_mask(page, pfn, mask);
520 static __always_inline int get_pfnblock_migratetype(const struct page *page,
523 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
527 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
528 * @page: The page within the block of interest
529 * @flags: The flags to set
530 * @pfn: The target page frame number
531 * @mask: mask of bits that the caller is interested in
533 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
537 unsigned long *bitmap;
538 unsigned long bitidx, word_bitidx;
539 unsigned long old_word, word;
541 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
542 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
544 bitmap = get_pageblock_bitmap(page, pfn);
545 bitidx = pfn_to_bitidx(page, pfn);
546 word_bitidx = bitidx / BITS_PER_LONG;
547 bitidx &= (BITS_PER_LONG-1);
549 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
554 word = READ_ONCE(bitmap[word_bitidx]);
556 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
557 if (word == old_word)
563 void set_pageblock_migratetype(struct page *page, int migratetype)
565 if (unlikely(page_group_by_mobility_disabled &&
566 migratetype < MIGRATE_PCPTYPES))
567 migratetype = MIGRATE_UNMOVABLE;
569 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
570 page_to_pfn(page), MIGRATETYPE_MASK);
573 #ifdef CONFIG_DEBUG_VM
574 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
578 unsigned long pfn = page_to_pfn(page);
579 unsigned long sp, start_pfn;
582 seq = zone_span_seqbegin(zone);
583 start_pfn = zone->zone_start_pfn;
584 sp = zone->spanned_pages;
585 if (!zone_spans_pfn(zone, pfn))
587 } while (zone_span_seqretry(zone, seq));
590 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
591 pfn, zone_to_nid(zone), zone->name,
592 start_pfn, start_pfn + sp);
597 static int page_is_consistent(struct zone *zone, struct page *page)
599 if (zone != page_zone(page))
605 * Temporary debugging check for pages not lying within a given zone.
607 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
609 if (page_outside_zone_boundaries(zone, page))
611 if (!page_is_consistent(zone, page))
617 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
623 static void bad_page(struct page *page, const char *reason)
625 static unsigned long resume;
626 static unsigned long nr_shown;
627 static unsigned long nr_unshown;
630 * Allow a burst of 60 reports, then keep quiet for that minute;
631 * or allow a steady drip of one report per second.
633 if (nr_shown == 60) {
634 if (time_before(jiffies, resume)) {
640 "BUG: Bad page state: %lu messages suppressed\n",
647 resume = jiffies + 60 * HZ;
649 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
650 current->comm, page_to_pfn(page));
651 dump_page(page, reason);
656 /* Leave bad fields for debug, except PageBuddy could make trouble */
657 page_mapcount_reset(page); /* remove PageBuddy */
658 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
661 static inline unsigned int order_to_pindex(int migratetype, int order)
665 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
666 if (order > PAGE_ALLOC_COSTLY_ORDER) {
667 VM_BUG_ON(order != pageblock_order);
668 base = PAGE_ALLOC_COSTLY_ORDER + 1;
671 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
674 return (MIGRATE_PCPTYPES * base) + migratetype;
677 static inline int pindex_to_order(unsigned int pindex)
679 int order = pindex / MIGRATE_PCPTYPES;
681 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
682 if (order > PAGE_ALLOC_COSTLY_ORDER)
683 order = pageblock_order;
685 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
691 static inline bool pcp_allowed_order(unsigned int order)
693 if (order <= PAGE_ALLOC_COSTLY_ORDER)
695 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
696 if (order == pageblock_order)
702 static inline void free_the_page(struct page *page, unsigned int order)
704 if (pcp_allowed_order(order)) /* Via pcp? */
705 free_unref_page(page, order);
707 __free_pages_ok(page, order, FPI_NONE);
711 * Higher-order pages are called "compound pages". They are structured thusly:
713 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
715 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
716 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
718 * The first tail page's ->compound_dtor holds the offset in array of compound
719 * page destructors. See compound_page_dtors.
721 * The first tail page's ->compound_order holds the order of allocation.
722 * This usage means that zero-order pages may not be compound.
725 void free_compound_page(struct page *page)
727 mem_cgroup_uncharge(page_folio(page));
728 free_the_page(page, compound_order(page));
731 static void prep_compound_head(struct page *page, unsigned int order)
733 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
734 set_compound_order(page, order);
735 atomic_set(compound_mapcount_ptr(page), -1);
736 if (hpage_pincount_available(page))
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 long buddy_pfn;
1075 unsigned long combined_pfn;
1076 unsigned int max_order;
1080 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1082 VM_BUG_ON(!zone_is_initialized(zone));
1083 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1085 VM_BUG_ON(migratetype == -1);
1086 if (likely(!is_migrate_isolate(migratetype)))
1087 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1089 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1090 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1093 while (order < max_order) {
1094 if (compaction_capture(capc, page, order, migratetype)) {
1095 __mod_zone_freepage_state(zone, -(1 << order),
1099 buddy_pfn = __find_buddy_pfn(pfn, order);
1100 buddy = page + (buddy_pfn - pfn);
1102 if (!page_is_buddy(page, buddy, order))
1105 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1106 * merge with it and move up one order.
1108 if (page_is_guard(buddy))
1109 clear_page_guard(zone, buddy, order, migratetype);
1111 del_page_from_free_list(buddy, zone, order);
1112 combined_pfn = buddy_pfn & pfn;
1113 page = page + (combined_pfn - pfn);
1117 if (order < MAX_ORDER - 1) {
1118 /* If we are here, it means order is >= pageblock_order.
1119 * We want to prevent merge between freepages on isolate
1120 * pageblock and normal pageblock. Without this, pageblock
1121 * isolation could cause incorrect freepage or CMA accounting.
1123 * We don't want to hit this code for the more frequent
1124 * low-order merging.
1126 if (unlikely(has_isolate_pageblock(zone))) {
1129 buddy_pfn = __find_buddy_pfn(pfn, order);
1130 buddy = page + (buddy_pfn - pfn);
1131 buddy_mt = get_pageblock_migratetype(buddy);
1133 if (migratetype != buddy_mt
1134 && (is_migrate_isolate(migratetype) ||
1135 is_migrate_isolate(buddy_mt)))
1138 max_order = order + 1;
1139 goto continue_merging;
1143 set_buddy_order(page, order);
1145 if (fpi_flags & FPI_TO_TAIL)
1147 else if (is_shuffle_order(order))
1148 to_tail = shuffle_pick_tail();
1150 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1153 add_to_free_list_tail(page, zone, order, migratetype);
1155 add_to_free_list(page, zone, order, migratetype);
1157 /* Notify page reporting subsystem of freed page */
1158 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1159 page_reporting_notify_free(order);
1163 * A bad page could be due to a number of fields. Instead of multiple branches,
1164 * try and check multiple fields with one check. The caller must do a detailed
1165 * check if necessary.
1167 static inline bool page_expected_state(struct page *page,
1168 unsigned long check_flags)
1170 if (unlikely(atomic_read(&page->_mapcount) != -1))
1173 if (unlikely((unsigned long)page->mapping |
1174 page_ref_count(page) |
1178 (page->flags & check_flags)))
1184 static const char *page_bad_reason(struct page *page, unsigned long flags)
1186 const char *bad_reason = NULL;
1188 if (unlikely(atomic_read(&page->_mapcount) != -1))
1189 bad_reason = "nonzero mapcount";
1190 if (unlikely(page->mapping != NULL))
1191 bad_reason = "non-NULL mapping";
1192 if (unlikely(page_ref_count(page) != 0))
1193 bad_reason = "nonzero _refcount";
1194 if (unlikely(page->flags & flags)) {
1195 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1196 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1198 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1201 if (unlikely(page->memcg_data))
1202 bad_reason = "page still charged to cgroup";
1207 static void check_free_page_bad(struct page *page)
1210 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1213 static inline int check_free_page(struct page *page)
1215 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1218 /* Something has gone sideways, find it */
1219 check_free_page_bad(page);
1223 static int free_tail_pages_check(struct page *head_page, struct page *page)
1228 * We rely page->lru.next never has bit 0 set, unless the page
1229 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1231 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1233 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1237 switch (page - head_page) {
1239 /* the first tail page: ->mapping may be compound_mapcount() */
1240 if (unlikely(compound_mapcount(page))) {
1241 bad_page(page, "nonzero compound_mapcount");
1247 * the second tail page: ->mapping is
1248 * deferred_list.next -- ignore value.
1252 if (page->mapping != TAIL_MAPPING) {
1253 bad_page(page, "corrupted mapping in tail page");
1258 if (unlikely(!PageTail(page))) {
1259 bad_page(page, "PageTail not set");
1262 if (unlikely(compound_head(page) != head_page)) {
1263 bad_page(page, "compound_head not consistent");
1268 page->mapping = NULL;
1269 clear_compound_head(page);
1273 static void kernel_init_free_pages(struct page *page, int numpages, bool zero_tags)
1278 for (i = 0; i < numpages; i++)
1279 tag_clear_highpage(page + i);
1283 /* s390's use of memset() could override KASAN redzones. */
1284 kasan_disable_current();
1285 for (i = 0; i < numpages; i++) {
1286 u8 tag = page_kasan_tag(page + i);
1287 page_kasan_tag_reset(page + i);
1288 clear_highpage(page + i);
1289 page_kasan_tag_set(page + i, tag);
1291 kasan_enable_current();
1294 static __always_inline bool free_pages_prepare(struct page *page,
1295 unsigned int order, bool check_free, fpi_t fpi_flags)
1298 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1300 VM_BUG_ON_PAGE(PageTail(page), page);
1302 trace_mm_page_free(page, order);
1304 if (unlikely(PageHWPoison(page)) && !order) {
1306 * Do not let hwpoison pages hit pcplists/buddy
1307 * Untie memcg state and reset page's owner
1309 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1310 __memcg_kmem_uncharge_page(page, order);
1311 reset_page_owner(page, order);
1312 page_table_check_free(page, order);
1317 * Check tail pages before head page information is cleared to
1318 * avoid checking PageCompound for order-0 pages.
1320 if (unlikely(order)) {
1321 bool compound = PageCompound(page);
1324 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1327 ClearPageDoubleMap(page);
1328 ClearPageHasHWPoisoned(page);
1330 for (i = 1; i < (1 << order); i++) {
1332 bad += free_tail_pages_check(page, page + i);
1333 if (unlikely(check_free_page(page + i))) {
1337 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1340 if (PageMappingFlags(page))
1341 page->mapping = NULL;
1342 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1343 __memcg_kmem_uncharge_page(page, order);
1345 bad += check_free_page(page);
1349 page_cpupid_reset_last(page);
1350 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1351 reset_page_owner(page, order);
1352 page_table_check_free(page, order);
1354 if (!PageHighMem(page)) {
1355 debug_check_no_locks_freed(page_address(page),
1356 PAGE_SIZE << order);
1357 debug_check_no_obj_freed(page_address(page),
1358 PAGE_SIZE << order);
1361 kernel_poison_pages(page, 1 << order);
1364 * As memory initialization might be integrated into KASAN,
1365 * kasan_free_pages and kernel_init_free_pages must be
1366 * kept together to avoid discrepancies in behavior.
1368 * With hardware tag-based KASAN, memory tags must be set before the
1369 * page becomes unavailable via debug_pagealloc or arch_free_page.
1371 if (kasan_has_integrated_init()) {
1372 if (!skip_kasan_poison)
1373 kasan_free_pages(page, order);
1375 bool init = want_init_on_free();
1378 kernel_init_free_pages(page, 1 << order, false);
1379 if (!skip_kasan_poison)
1380 kasan_poison_pages(page, order, init);
1384 * arch_free_page() can make the page's contents inaccessible. s390
1385 * does this. So nothing which can access the page's contents should
1386 * happen after this.
1388 arch_free_page(page, order);
1390 debug_pagealloc_unmap_pages(page, 1 << order);
1395 #ifdef CONFIG_DEBUG_VM
1397 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1398 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1399 * moved from pcp lists to free lists.
1401 static bool free_pcp_prepare(struct page *page, unsigned int order)
1403 return free_pages_prepare(page, order, true, FPI_NONE);
1406 static bool bulkfree_pcp_prepare(struct page *page)
1408 if (debug_pagealloc_enabled_static())
1409 return check_free_page(page);
1415 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1416 * moving from pcp lists to free list in order to reduce overhead. With
1417 * debug_pagealloc enabled, they are checked also immediately when being freed
1420 static bool free_pcp_prepare(struct page *page, unsigned int order)
1422 if (debug_pagealloc_enabled_static())
1423 return free_pages_prepare(page, order, true, FPI_NONE);
1425 return free_pages_prepare(page, order, false, FPI_NONE);
1428 static bool bulkfree_pcp_prepare(struct page *page)
1430 return check_free_page(page);
1432 #endif /* CONFIG_DEBUG_VM */
1434 static inline void prefetch_buddy(struct page *page)
1436 unsigned long pfn = page_to_pfn(page);
1437 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1438 struct page *buddy = page + (buddy_pfn - pfn);
1444 * Frees a number of pages from the PCP lists
1445 * Assumes all pages on list are in same zone.
1446 * count is the number of pages to free.
1448 static void free_pcppages_bulk(struct zone *zone, int count,
1449 struct per_cpu_pages *pcp)
1455 int prefetch_nr = READ_ONCE(pcp->batch);
1456 bool isolated_pageblocks;
1457 struct page *page, *tmp;
1461 * Ensure proper count is passed which otherwise would stuck in the
1462 * below while (list_empty(list)) loop.
1464 count = min(pcp->count, count);
1466 struct list_head *list;
1469 * Remove pages from lists in a round-robin fashion. A
1470 * batch_free count is maintained that is incremented when an
1471 * empty list is encountered. This is so more pages are freed
1472 * off fuller lists instead of spinning excessively around empty
1477 if (++pindex == NR_PCP_LISTS)
1479 list = &pcp->lists[pindex];
1480 } while (list_empty(list));
1482 /* This is the only non-empty list. Free them all. */
1483 if (batch_free == NR_PCP_LISTS)
1486 order = pindex_to_order(pindex);
1487 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1489 page = list_last_entry(list, struct page, lru);
1490 /* must delete to avoid corrupting pcp list */
1491 list_del(&page->lru);
1492 nr_freed += 1 << order;
1493 count -= 1 << order;
1495 if (bulkfree_pcp_prepare(page))
1498 /* Encode order with the migratetype */
1499 page->index <<= NR_PCP_ORDER_WIDTH;
1500 page->index |= order;
1502 list_add_tail(&page->lru, &head);
1505 * We are going to put the page back to the global
1506 * pool, prefetch its buddy to speed up later access
1507 * under zone->lock. It is believed the overhead of
1508 * an additional test and calculating buddy_pfn here
1509 * can be offset by reduced memory latency later. To
1510 * avoid excessive prefetching due to large count, only
1511 * prefetch buddy for the first pcp->batch nr of pages.
1514 prefetch_buddy(page);
1517 } while (count > 0 && --batch_free && !list_empty(list));
1519 pcp->count -= nr_freed;
1522 * local_lock_irq held so equivalent to spin_lock_irqsave for
1523 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1525 spin_lock(&zone->lock);
1526 isolated_pageblocks = has_isolate_pageblock(zone);
1529 * Use safe version since after __free_one_page(),
1530 * page->lru.next will not point to original list.
1532 list_for_each_entry_safe(page, tmp, &head, lru) {
1533 int mt = get_pcppage_migratetype(page);
1535 /* mt has been encoded with the order (see above) */
1536 order = mt & NR_PCP_ORDER_MASK;
1537 mt >>= NR_PCP_ORDER_WIDTH;
1539 /* MIGRATE_ISOLATE page should not go to pcplists */
1540 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1541 /* Pageblock could have been isolated meanwhile */
1542 if (unlikely(isolated_pageblocks))
1543 mt = get_pageblock_migratetype(page);
1545 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1546 trace_mm_page_pcpu_drain(page, order, mt);
1548 spin_unlock(&zone->lock);
1551 static void free_one_page(struct zone *zone,
1552 struct page *page, unsigned long pfn,
1554 int migratetype, fpi_t fpi_flags)
1556 unsigned long flags;
1558 spin_lock_irqsave(&zone->lock, flags);
1559 if (unlikely(has_isolate_pageblock(zone) ||
1560 is_migrate_isolate(migratetype))) {
1561 migratetype = get_pfnblock_migratetype(page, pfn);
1563 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1564 spin_unlock_irqrestore(&zone->lock, flags);
1567 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1568 unsigned long zone, int nid)
1570 mm_zero_struct_page(page);
1571 set_page_links(page, zone, nid, pfn);
1572 init_page_count(page);
1573 page_mapcount_reset(page);
1574 page_cpupid_reset_last(page);
1575 page_kasan_tag_reset(page);
1577 INIT_LIST_HEAD(&page->lru);
1578 #ifdef WANT_PAGE_VIRTUAL
1579 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1580 if (!is_highmem_idx(zone))
1581 set_page_address(page, __va(pfn << PAGE_SHIFT));
1585 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1586 static void __meminit init_reserved_page(unsigned long pfn)
1591 if (!early_page_uninitialised(pfn))
1594 nid = early_pfn_to_nid(pfn);
1595 pgdat = NODE_DATA(nid);
1597 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1598 struct zone *zone = &pgdat->node_zones[zid];
1600 if (zone_spans_pfn(zone, pfn))
1603 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1606 static inline void init_reserved_page(unsigned long pfn)
1609 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1612 * Initialised pages do not have PageReserved set. This function is
1613 * called for each range allocated by the bootmem allocator and
1614 * marks the pages PageReserved. The remaining valid pages are later
1615 * sent to the buddy page allocator.
1617 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1619 unsigned long start_pfn = PFN_DOWN(start);
1620 unsigned long end_pfn = PFN_UP(end);
1622 for (; start_pfn < end_pfn; start_pfn++) {
1623 if (pfn_valid(start_pfn)) {
1624 struct page *page = pfn_to_page(start_pfn);
1626 init_reserved_page(start_pfn);
1628 /* Avoid false-positive PageTail() */
1629 INIT_LIST_HEAD(&page->lru);
1632 * no need for atomic set_bit because the struct
1633 * page is not visible yet so nobody should
1636 __SetPageReserved(page);
1641 static void __free_pages_ok(struct page *page, unsigned int order,
1644 unsigned long flags;
1646 unsigned long pfn = page_to_pfn(page);
1647 struct zone *zone = page_zone(page);
1649 if (!free_pages_prepare(page, order, true, fpi_flags))
1652 migratetype = get_pfnblock_migratetype(page, pfn);
1654 spin_lock_irqsave(&zone->lock, flags);
1655 if (unlikely(has_isolate_pageblock(zone) ||
1656 is_migrate_isolate(migratetype))) {
1657 migratetype = get_pfnblock_migratetype(page, pfn);
1659 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1660 spin_unlock_irqrestore(&zone->lock, flags);
1662 __count_vm_events(PGFREE, 1 << order);
1665 void __free_pages_core(struct page *page, unsigned int order)
1667 unsigned int nr_pages = 1 << order;
1668 struct page *p = page;
1672 * When initializing the memmap, __init_single_page() sets the refcount
1673 * of all pages to 1 ("allocated"/"not free"). We have to set the
1674 * refcount of all involved pages to 0.
1677 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1679 __ClearPageReserved(p);
1680 set_page_count(p, 0);
1682 __ClearPageReserved(p);
1683 set_page_count(p, 0);
1685 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1688 * Bypass PCP and place fresh pages right to the tail, primarily
1689 * relevant for memory onlining.
1691 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1697 * During memory init memblocks map pfns to nids. The search is expensive and
1698 * this caches recent lookups. The implementation of __early_pfn_to_nid
1699 * treats start/end as pfns.
1701 struct mminit_pfnnid_cache {
1702 unsigned long last_start;
1703 unsigned long last_end;
1707 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1710 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1712 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1713 struct mminit_pfnnid_cache *state)
1715 unsigned long start_pfn, end_pfn;
1718 if (state->last_start <= pfn && pfn < state->last_end)
1719 return state->last_nid;
1721 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1722 if (nid != NUMA_NO_NODE) {
1723 state->last_start = start_pfn;
1724 state->last_end = end_pfn;
1725 state->last_nid = nid;
1731 int __meminit early_pfn_to_nid(unsigned long pfn)
1733 static DEFINE_SPINLOCK(early_pfn_lock);
1736 spin_lock(&early_pfn_lock);
1737 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1739 nid = first_online_node;
1740 spin_unlock(&early_pfn_lock);
1744 #endif /* CONFIG_NUMA */
1746 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1749 if (early_page_uninitialised(pfn))
1751 __free_pages_core(page, order);
1755 * Check that the whole (or subset of) a pageblock given by the interval of
1756 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1757 * with the migration of free compaction scanner.
1759 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1761 * It's possible on some configurations to have a setup like node0 node1 node0
1762 * i.e. it's possible that all pages within a zones range of pages do not
1763 * belong to a single zone. We assume that a border between node0 and node1
1764 * can occur within a single pageblock, but not a node0 node1 node0
1765 * interleaving within a single pageblock. It is therefore sufficient to check
1766 * the first and last page of a pageblock and avoid checking each individual
1767 * page in a pageblock.
1769 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1770 unsigned long end_pfn, struct zone *zone)
1772 struct page *start_page;
1773 struct page *end_page;
1775 /* end_pfn is one past the range we are checking */
1778 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1781 start_page = pfn_to_online_page(start_pfn);
1785 if (page_zone(start_page) != zone)
1788 end_page = pfn_to_page(end_pfn);
1790 /* This gives a shorter code than deriving page_zone(end_page) */
1791 if (page_zone_id(start_page) != page_zone_id(end_page))
1797 void set_zone_contiguous(struct zone *zone)
1799 unsigned long block_start_pfn = zone->zone_start_pfn;
1800 unsigned long block_end_pfn;
1802 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1803 for (; block_start_pfn < zone_end_pfn(zone);
1804 block_start_pfn = block_end_pfn,
1805 block_end_pfn += pageblock_nr_pages) {
1807 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1809 if (!__pageblock_pfn_to_page(block_start_pfn,
1810 block_end_pfn, zone))
1815 /* We confirm that there is no hole */
1816 zone->contiguous = true;
1819 void clear_zone_contiguous(struct zone *zone)
1821 zone->contiguous = false;
1824 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1825 static void __init deferred_free_range(unsigned long pfn,
1826 unsigned long nr_pages)
1834 page = pfn_to_page(pfn);
1836 /* Free a large naturally-aligned chunk if possible */
1837 if (nr_pages == pageblock_nr_pages &&
1838 (pfn & (pageblock_nr_pages - 1)) == 0) {
1839 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1840 __free_pages_core(page, pageblock_order);
1844 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1845 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1846 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1847 __free_pages_core(page, 0);
1851 /* Completion tracking for deferred_init_memmap() threads */
1852 static atomic_t pgdat_init_n_undone __initdata;
1853 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1855 static inline void __init pgdat_init_report_one_done(void)
1857 if (atomic_dec_and_test(&pgdat_init_n_undone))
1858 complete(&pgdat_init_all_done_comp);
1862 * Returns true if page needs to be initialized or freed to buddy allocator.
1864 * First we check if pfn is valid on architectures where it is possible to have
1865 * holes within pageblock_nr_pages. On systems where it is not possible, this
1866 * function is optimized out.
1868 * Then, we check if a current large page is valid by only checking the validity
1871 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1873 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1879 * Free pages to buddy allocator. Try to free aligned pages in
1880 * pageblock_nr_pages sizes.
1882 static void __init deferred_free_pages(unsigned long pfn,
1883 unsigned long end_pfn)
1885 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1886 unsigned long nr_free = 0;
1888 for (; pfn < end_pfn; pfn++) {
1889 if (!deferred_pfn_valid(pfn)) {
1890 deferred_free_range(pfn - nr_free, nr_free);
1892 } else if (!(pfn & nr_pgmask)) {
1893 deferred_free_range(pfn - nr_free, nr_free);
1899 /* Free the last block of pages to allocator */
1900 deferred_free_range(pfn - nr_free, nr_free);
1904 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1905 * by performing it only once every pageblock_nr_pages.
1906 * Return number of pages initialized.
1908 static unsigned long __init deferred_init_pages(struct zone *zone,
1910 unsigned long end_pfn)
1912 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1913 int nid = zone_to_nid(zone);
1914 unsigned long nr_pages = 0;
1915 int zid = zone_idx(zone);
1916 struct page *page = NULL;
1918 for (; pfn < end_pfn; pfn++) {
1919 if (!deferred_pfn_valid(pfn)) {
1922 } else if (!page || !(pfn & nr_pgmask)) {
1923 page = pfn_to_page(pfn);
1927 __init_single_page(page, pfn, zid, nid);
1934 * This function is meant to pre-load the iterator for the zone init.
1935 * Specifically it walks through the ranges until we are caught up to the
1936 * first_init_pfn value and exits there. If we never encounter the value we
1937 * return false indicating there are no valid ranges left.
1940 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1941 unsigned long *spfn, unsigned long *epfn,
1942 unsigned long first_init_pfn)
1947 * Start out by walking through the ranges in this zone that have
1948 * already been initialized. We don't need to do anything with them
1949 * so we just need to flush them out of the system.
1951 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1952 if (*epfn <= first_init_pfn)
1954 if (*spfn < first_init_pfn)
1955 *spfn = first_init_pfn;
1964 * Initialize and free pages. We do it in two loops: first we initialize
1965 * struct page, then free to buddy allocator, because while we are
1966 * freeing pages we can access pages that are ahead (computing buddy
1967 * page in __free_one_page()).
1969 * In order to try and keep some memory in the cache we have the loop
1970 * broken along max page order boundaries. This way we will not cause
1971 * any issues with the buddy page computation.
1973 static unsigned long __init
1974 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1975 unsigned long *end_pfn)
1977 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1978 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1979 unsigned long nr_pages = 0;
1982 /* First we loop through and initialize the page values */
1983 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1986 if (mo_pfn <= *start_pfn)
1989 t = min(mo_pfn, *end_pfn);
1990 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1992 if (mo_pfn < *end_pfn) {
1993 *start_pfn = mo_pfn;
1998 /* Reset values and now loop through freeing pages as needed */
2001 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2007 t = min(mo_pfn, epfn);
2008 deferred_free_pages(spfn, t);
2018 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2021 unsigned long spfn, epfn;
2022 struct zone *zone = arg;
2025 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2028 * Initialize and free pages in MAX_ORDER sized increments so that we
2029 * can avoid introducing any issues with the buddy allocator.
2031 while (spfn < end_pfn) {
2032 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2037 /* An arch may override for more concurrency. */
2039 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2044 /* Initialise remaining memory on a node */
2045 static int __init deferred_init_memmap(void *data)
2047 pg_data_t *pgdat = data;
2048 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2049 unsigned long spfn = 0, epfn = 0;
2050 unsigned long first_init_pfn, flags;
2051 unsigned long start = jiffies;
2053 int zid, max_threads;
2056 /* Bind memory initialisation thread to a local node if possible */
2057 if (!cpumask_empty(cpumask))
2058 set_cpus_allowed_ptr(current, cpumask);
2060 pgdat_resize_lock(pgdat, &flags);
2061 first_init_pfn = pgdat->first_deferred_pfn;
2062 if (first_init_pfn == ULONG_MAX) {
2063 pgdat_resize_unlock(pgdat, &flags);
2064 pgdat_init_report_one_done();
2068 /* Sanity check boundaries */
2069 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2070 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2071 pgdat->first_deferred_pfn = ULONG_MAX;
2074 * Once we unlock here, the zone cannot be grown anymore, thus if an
2075 * interrupt thread must allocate this early in boot, zone must be
2076 * pre-grown prior to start of deferred page initialization.
2078 pgdat_resize_unlock(pgdat, &flags);
2080 /* Only the highest zone is deferred so find it */
2081 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2082 zone = pgdat->node_zones + zid;
2083 if (first_init_pfn < zone_end_pfn(zone))
2087 /* If the zone is empty somebody else may have cleared out the zone */
2088 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2092 max_threads = deferred_page_init_max_threads(cpumask);
2094 while (spfn < epfn) {
2095 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2096 struct padata_mt_job job = {
2097 .thread_fn = deferred_init_memmap_chunk,
2100 .size = epfn_align - spfn,
2101 .align = PAGES_PER_SECTION,
2102 .min_chunk = PAGES_PER_SECTION,
2103 .max_threads = max_threads,
2106 padata_do_multithreaded(&job);
2107 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2111 /* Sanity check that the next zone really is unpopulated */
2112 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2114 pr_info("node %d deferred pages initialised in %ums\n",
2115 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2117 pgdat_init_report_one_done();
2122 * If this zone has deferred pages, try to grow it by initializing enough
2123 * deferred pages to satisfy the allocation specified by order, rounded up to
2124 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2125 * of SECTION_SIZE bytes by initializing struct pages in increments of
2126 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2128 * Return true when zone was grown, otherwise return false. We return true even
2129 * when we grow less than requested, to let the caller decide if there are
2130 * enough pages to satisfy the allocation.
2132 * Note: We use noinline because this function is needed only during boot, and
2133 * it is called from a __ref function _deferred_grow_zone. This way we are
2134 * making sure that it is not inlined into permanent text section.
2136 static noinline bool __init
2137 deferred_grow_zone(struct zone *zone, unsigned int order)
2139 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2140 pg_data_t *pgdat = zone->zone_pgdat;
2141 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2142 unsigned long spfn, epfn, flags;
2143 unsigned long nr_pages = 0;
2146 /* Only the last zone may have deferred pages */
2147 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2150 pgdat_resize_lock(pgdat, &flags);
2153 * If someone grew this zone while we were waiting for spinlock, return
2154 * true, as there might be enough pages already.
2156 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2157 pgdat_resize_unlock(pgdat, &flags);
2161 /* If the zone is empty somebody else may have cleared out the zone */
2162 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2163 first_deferred_pfn)) {
2164 pgdat->first_deferred_pfn = ULONG_MAX;
2165 pgdat_resize_unlock(pgdat, &flags);
2166 /* Retry only once. */
2167 return first_deferred_pfn != ULONG_MAX;
2171 * Initialize and free pages in MAX_ORDER sized increments so
2172 * that we can avoid introducing any issues with the buddy
2175 while (spfn < epfn) {
2176 /* update our first deferred PFN for this section */
2177 first_deferred_pfn = spfn;
2179 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2180 touch_nmi_watchdog();
2182 /* We should only stop along section boundaries */
2183 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2186 /* If our quota has been met we can stop here */
2187 if (nr_pages >= nr_pages_needed)
2191 pgdat->first_deferred_pfn = spfn;
2192 pgdat_resize_unlock(pgdat, &flags);
2194 return nr_pages > 0;
2198 * deferred_grow_zone() is __init, but it is called from
2199 * get_page_from_freelist() during early boot until deferred_pages permanently
2200 * disables this call. This is why we have refdata wrapper to avoid warning,
2201 * and to ensure that the function body gets unloaded.
2204 _deferred_grow_zone(struct zone *zone, unsigned int order)
2206 return deferred_grow_zone(zone, order);
2209 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2211 void __init page_alloc_init_late(void)
2216 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2218 /* There will be num_node_state(N_MEMORY) threads */
2219 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2220 for_each_node_state(nid, N_MEMORY) {
2221 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2224 /* Block until all are initialised */
2225 wait_for_completion(&pgdat_init_all_done_comp);
2228 * We initialized the rest of the deferred pages. Permanently disable
2229 * on-demand struct page initialization.
2231 static_branch_disable(&deferred_pages);
2233 /* Reinit limits that are based on free pages after the kernel is up */
2234 files_maxfiles_init();
2239 /* Discard memblock private memory */
2242 for_each_node_state(nid, N_MEMORY)
2243 shuffle_free_memory(NODE_DATA(nid));
2245 for_each_populated_zone(zone)
2246 set_zone_contiguous(zone);
2250 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2251 void __init init_cma_reserved_pageblock(struct page *page)
2253 unsigned i = pageblock_nr_pages;
2254 struct page *p = page;
2257 __ClearPageReserved(p);
2258 set_page_count(p, 0);
2261 set_pageblock_migratetype(page, MIGRATE_CMA);
2263 if (pageblock_order >= MAX_ORDER) {
2264 i = pageblock_nr_pages;
2267 set_page_refcounted(p);
2268 __free_pages(p, MAX_ORDER - 1);
2269 p += MAX_ORDER_NR_PAGES;
2270 } while (i -= MAX_ORDER_NR_PAGES);
2272 set_page_refcounted(page);
2273 __free_pages(page, pageblock_order);
2276 adjust_managed_page_count(page, pageblock_nr_pages);
2277 page_zone(page)->cma_pages += pageblock_nr_pages;
2282 * The order of subdivision here is critical for the IO subsystem.
2283 * Please do not alter this order without good reasons and regression
2284 * testing. Specifically, as large blocks of memory are subdivided,
2285 * the order in which smaller blocks are delivered depends on the order
2286 * they're subdivided in this function. This is the primary factor
2287 * influencing the order in which pages are delivered to the IO
2288 * subsystem according to empirical testing, and this is also justified
2289 * by considering the behavior of a buddy system containing a single
2290 * large block of memory acted on by a series of small allocations.
2291 * This behavior is a critical factor in sglist merging's success.
2295 static inline void expand(struct zone *zone, struct page *page,
2296 int low, int high, int migratetype)
2298 unsigned long size = 1 << high;
2300 while (high > low) {
2303 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2306 * Mark as guard pages (or page), that will allow to
2307 * merge back to allocator when buddy will be freed.
2308 * Corresponding page table entries will not be touched,
2309 * pages will stay not present in virtual address space
2311 if (set_page_guard(zone, &page[size], high, migratetype))
2314 add_to_free_list(&page[size], zone, high, migratetype);
2315 set_buddy_order(&page[size], high);
2319 static void check_new_page_bad(struct page *page)
2321 if (unlikely(page->flags & __PG_HWPOISON)) {
2322 /* Don't complain about hwpoisoned pages */
2323 page_mapcount_reset(page); /* remove PageBuddy */
2328 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2332 * This page is about to be returned from the page allocator
2334 static inline int check_new_page(struct page *page)
2336 if (likely(page_expected_state(page,
2337 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2340 check_new_page_bad(page);
2344 #ifdef CONFIG_DEBUG_VM
2346 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2347 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2348 * also checked when pcp lists are refilled from the free lists.
2350 static inline bool check_pcp_refill(struct page *page)
2352 if (debug_pagealloc_enabled_static())
2353 return check_new_page(page);
2358 static inline bool check_new_pcp(struct page *page)
2360 return check_new_page(page);
2364 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2365 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2366 * enabled, they are also checked when being allocated from the pcp lists.
2368 static inline bool check_pcp_refill(struct page *page)
2370 return check_new_page(page);
2372 static inline bool check_new_pcp(struct page *page)
2374 if (debug_pagealloc_enabled_static())
2375 return check_new_page(page);
2379 #endif /* CONFIG_DEBUG_VM */
2381 static bool check_new_pages(struct page *page, unsigned int order)
2384 for (i = 0; i < (1 << order); i++) {
2385 struct page *p = page + i;
2387 if (unlikely(check_new_page(p)))
2394 inline void post_alloc_hook(struct page *page, unsigned int order,
2397 set_page_private(page, 0);
2398 set_page_refcounted(page);
2400 arch_alloc_page(page, order);
2401 debug_pagealloc_map_pages(page, 1 << order);
2404 * Page unpoisoning must happen before memory initialization.
2405 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2406 * allocations and the page unpoisoning code will complain.
2408 kernel_unpoison_pages(page, 1 << order);
2411 * As memory initialization might be integrated into KASAN,
2412 * kasan_alloc_pages and kernel_init_free_pages must be
2413 * kept together to avoid discrepancies in behavior.
2415 if (kasan_has_integrated_init()) {
2416 kasan_alloc_pages(page, order, gfp_flags);
2418 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2420 kasan_unpoison_pages(page, order, init);
2422 kernel_init_free_pages(page, 1 << order,
2423 gfp_flags & __GFP_ZEROTAGS);
2426 set_page_owner(page, order, gfp_flags);
2427 page_table_check_alloc(page, order);
2430 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2431 unsigned int alloc_flags)
2433 post_alloc_hook(page, order, gfp_flags);
2435 if (order && (gfp_flags & __GFP_COMP))
2436 prep_compound_page(page, order);
2439 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2440 * allocate the page. The expectation is that the caller is taking
2441 * steps that will free more memory. The caller should avoid the page
2442 * being used for !PFMEMALLOC purposes.
2444 if (alloc_flags & ALLOC_NO_WATERMARKS)
2445 set_page_pfmemalloc(page);
2447 clear_page_pfmemalloc(page);
2451 * Go through the free lists for the given migratetype and remove
2452 * the smallest available page from the freelists
2454 static __always_inline
2455 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2458 unsigned int current_order;
2459 struct free_area *area;
2462 /* Find a page of the appropriate size in the preferred list */
2463 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2464 area = &(zone->free_area[current_order]);
2465 page = get_page_from_free_area(area, migratetype);
2468 del_page_from_free_list(page, zone, current_order);
2469 expand(zone, page, order, current_order, migratetype);
2470 set_pcppage_migratetype(page, migratetype);
2479 * This array describes the order lists are fallen back to when
2480 * the free lists for the desirable migrate type are depleted
2482 static int fallbacks[MIGRATE_TYPES][3] = {
2483 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2484 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2485 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2487 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2489 #ifdef CONFIG_MEMORY_ISOLATION
2490 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2495 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2498 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2501 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2502 unsigned int order) { return NULL; }
2506 * Move the free pages in a range to the freelist tail of the requested type.
2507 * Note that start_page and end_pages are not aligned on a pageblock
2508 * boundary. If alignment is required, use move_freepages_block()
2510 static int move_freepages(struct zone *zone,
2511 unsigned long start_pfn, unsigned long end_pfn,
2512 int migratetype, int *num_movable)
2517 int pages_moved = 0;
2519 for (pfn = start_pfn; pfn <= end_pfn;) {
2520 page = pfn_to_page(pfn);
2521 if (!PageBuddy(page)) {
2523 * We assume that pages that could be isolated for
2524 * migration are movable. But we don't actually try
2525 * isolating, as that would be expensive.
2528 (PageLRU(page) || __PageMovable(page)))
2534 /* Make sure we are not inadvertently changing nodes */
2535 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2536 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2538 order = buddy_order(page);
2539 move_to_free_list(page, zone, order, migratetype);
2541 pages_moved += 1 << order;
2547 int move_freepages_block(struct zone *zone, struct page *page,
2548 int migratetype, int *num_movable)
2550 unsigned long start_pfn, end_pfn, pfn;
2555 pfn = page_to_pfn(page);
2556 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2557 end_pfn = start_pfn + pageblock_nr_pages - 1;
2559 /* Do not cross zone boundaries */
2560 if (!zone_spans_pfn(zone, start_pfn))
2562 if (!zone_spans_pfn(zone, end_pfn))
2565 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2569 static void change_pageblock_range(struct page *pageblock_page,
2570 int start_order, int migratetype)
2572 int nr_pageblocks = 1 << (start_order - pageblock_order);
2574 while (nr_pageblocks--) {
2575 set_pageblock_migratetype(pageblock_page, migratetype);
2576 pageblock_page += pageblock_nr_pages;
2581 * When we are falling back to another migratetype during allocation, try to
2582 * steal extra free pages from the same pageblocks to satisfy further
2583 * allocations, instead of polluting multiple pageblocks.
2585 * If we are stealing a relatively large buddy page, it is likely there will
2586 * be more free pages in the pageblock, so try to steal them all. For
2587 * reclaimable and unmovable allocations, we steal regardless of page size,
2588 * as fragmentation caused by those allocations polluting movable pageblocks
2589 * is worse than movable allocations stealing from unmovable and reclaimable
2592 static bool can_steal_fallback(unsigned int order, int start_mt)
2595 * Leaving this order check is intended, although there is
2596 * relaxed order check in next check. The reason is that
2597 * we can actually steal whole pageblock if this condition met,
2598 * but, below check doesn't guarantee it and that is just heuristic
2599 * so could be changed anytime.
2601 if (order >= pageblock_order)
2604 if (order >= pageblock_order / 2 ||
2605 start_mt == MIGRATE_RECLAIMABLE ||
2606 start_mt == MIGRATE_UNMOVABLE ||
2607 page_group_by_mobility_disabled)
2613 static inline bool boost_watermark(struct zone *zone)
2615 unsigned long max_boost;
2617 if (!watermark_boost_factor)
2620 * Don't bother in zones that are unlikely to produce results.
2621 * On small machines, including kdump capture kernels running
2622 * in a small area, boosting the watermark can cause an out of
2623 * memory situation immediately.
2625 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2628 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2629 watermark_boost_factor, 10000);
2632 * high watermark may be uninitialised if fragmentation occurs
2633 * very early in boot so do not boost. We do not fall
2634 * through and boost by pageblock_nr_pages as failing
2635 * allocations that early means that reclaim is not going
2636 * to help and it may even be impossible to reclaim the
2637 * boosted watermark resulting in a hang.
2642 max_boost = max(pageblock_nr_pages, max_boost);
2644 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2651 * This function implements actual steal behaviour. If order is large enough,
2652 * we can steal whole pageblock. If not, we first move freepages in this
2653 * pageblock to our migratetype and determine how many already-allocated pages
2654 * are there in the pageblock with a compatible migratetype. If at least half
2655 * of pages are free or compatible, we can change migratetype of the pageblock
2656 * itself, so pages freed in the future will be put on the correct free list.
2658 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2659 unsigned int alloc_flags, int start_type, bool whole_block)
2661 unsigned int current_order = buddy_order(page);
2662 int free_pages, movable_pages, alike_pages;
2665 old_block_type = get_pageblock_migratetype(page);
2668 * This can happen due to races and we want to prevent broken
2669 * highatomic accounting.
2671 if (is_migrate_highatomic(old_block_type))
2674 /* Take ownership for orders >= pageblock_order */
2675 if (current_order >= pageblock_order) {
2676 change_pageblock_range(page, current_order, start_type);
2681 * Boost watermarks to increase reclaim pressure to reduce the
2682 * likelihood of future fallbacks. Wake kswapd now as the node
2683 * may be balanced overall and kswapd will not wake naturally.
2685 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2686 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2688 /* We are not allowed to try stealing from the whole block */
2692 free_pages = move_freepages_block(zone, page, start_type,
2695 * Determine how many pages are compatible with our allocation.
2696 * For movable allocation, it's the number of movable pages which
2697 * we just obtained. For other types it's a bit more tricky.
2699 if (start_type == MIGRATE_MOVABLE) {
2700 alike_pages = movable_pages;
2703 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2704 * to MOVABLE pageblock, consider all non-movable pages as
2705 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2706 * vice versa, be conservative since we can't distinguish the
2707 * exact migratetype of non-movable pages.
2709 if (old_block_type == MIGRATE_MOVABLE)
2710 alike_pages = pageblock_nr_pages
2711 - (free_pages + movable_pages);
2716 /* moving whole block can fail due to zone boundary conditions */
2721 * If a sufficient number of pages in the block are either free or of
2722 * comparable migratability as our allocation, claim the whole block.
2724 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2725 page_group_by_mobility_disabled)
2726 set_pageblock_migratetype(page, start_type);
2731 move_to_free_list(page, zone, current_order, start_type);
2735 * Check whether there is a suitable fallback freepage with requested order.
2736 * If only_stealable is true, this function returns fallback_mt only if
2737 * we can steal other freepages all together. This would help to reduce
2738 * fragmentation due to mixed migratetype pages in one pageblock.
2740 int find_suitable_fallback(struct free_area *area, unsigned int order,
2741 int migratetype, bool only_stealable, bool *can_steal)
2746 if (area->nr_free == 0)
2751 fallback_mt = fallbacks[migratetype][i];
2752 if (fallback_mt == MIGRATE_TYPES)
2755 if (free_area_empty(area, fallback_mt))
2758 if (can_steal_fallback(order, migratetype))
2761 if (!only_stealable)
2772 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2773 * there are no empty page blocks that contain a page with a suitable order
2775 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2776 unsigned int alloc_order)
2779 unsigned long max_managed, flags;
2782 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2783 * Check is race-prone but harmless.
2785 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2786 if (zone->nr_reserved_highatomic >= max_managed)
2789 spin_lock_irqsave(&zone->lock, flags);
2791 /* Recheck the nr_reserved_highatomic limit under the lock */
2792 if (zone->nr_reserved_highatomic >= max_managed)
2796 mt = get_pageblock_migratetype(page);
2797 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2798 && !is_migrate_cma(mt)) {
2799 zone->nr_reserved_highatomic += pageblock_nr_pages;
2800 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2801 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2805 spin_unlock_irqrestore(&zone->lock, flags);
2809 * Used when an allocation is about to fail under memory pressure. This
2810 * potentially hurts the reliability of high-order allocations when under
2811 * intense memory pressure but failed atomic allocations should be easier
2812 * to recover from than an OOM.
2814 * If @force is true, try to unreserve a pageblock even though highatomic
2815 * pageblock is exhausted.
2817 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2820 struct zonelist *zonelist = ac->zonelist;
2821 unsigned long flags;
2828 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2831 * Preserve at least one pageblock unless memory pressure
2834 if (!force && zone->nr_reserved_highatomic <=
2838 spin_lock_irqsave(&zone->lock, flags);
2839 for (order = 0; order < MAX_ORDER; order++) {
2840 struct free_area *area = &(zone->free_area[order]);
2842 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2847 * In page freeing path, migratetype change is racy so
2848 * we can counter several free pages in a pageblock
2849 * in this loop although we changed the pageblock type
2850 * from highatomic to ac->migratetype. So we should
2851 * adjust the count once.
2853 if (is_migrate_highatomic_page(page)) {
2855 * It should never happen but changes to
2856 * locking could inadvertently allow a per-cpu
2857 * drain to add pages to MIGRATE_HIGHATOMIC
2858 * while unreserving so be safe and watch for
2861 zone->nr_reserved_highatomic -= min(
2863 zone->nr_reserved_highatomic);
2867 * Convert to ac->migratetype and avoid the normal
2868 * pageblock stealing heuristics. Minimally, the caller
2869 * is doing the work and needs the pages. More
2870 * importantly, if the block was always converted to
2871 * MIGRATE_UNMOVABLE or another type then the number
2872 * of pageblocks that cannot be completely freed
2875 set_pageblock_migratetype(page, ac->migratetype);
2876 ret = move_freepages_block(zone, page, ac->migratetype,
2879 spin_unlock_irqrestore(&zone->lock, flags);
2883 spin_unlock_irqrestore(&zone->lock, flags);
2890 * Try finding a free buddy page on the fallback list and put it on the free
2891 * list of requested migratetype, possibly along with other pages from the same
2892 * block, depending on fragmentation avoidance heuristics. Returns true if
2893 * fallback was found so that __rmqueue_smallest() can grab it.
2895 * The use of signed ints for order and current_order is a deliberate
2896 * deviation from the rest of this file, to make the for loop
2897 * condition simpler.
2899 static __always_inline bool
2900 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2901 unsigned int alloc_flags)
2903 struct free_area *area;
2905 int min_order = order;
2911 * Do not steal pages from freelists belonging to other pageblocks
2912 * i.e. orders < pageblock_order. If there are no local zones free,
2913 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2915 if (alloc_flags & ALLOC_NOFRAGMENT)
2916 min_order = pageblock_order;
2919 * Find the largest available free page in the other list. This roughly
2920 * approximates finding the pageblock with the most free pages, which
2921 * would be too costly to do exactly.
2923 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2925 area = &(zone->free_area[current_order]);
2926 fallback_mt = find_suitable_fallback(area, current_order,
2927 start_migratetype, false, &can_steal);
2928 if (fallback_mt == -1)
2932 * We cannot steal all free pages from the pageblock and the
2933 * requested migratetype is movable. In that case it's better to
2934 * steal and split the smallest available page instead of the
2935 * largest available page, because even if the next movable
2936 * allocation falls back into a different pageblock than this
2937 * one, it won't cause permanent fragmentation.
2939 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2940 && current_order > order)
2949 for (current_order = order; current_order < MAX_ORDER;
2951 area = &(zone->free_area[current_order]);
2952 fallback_mt = find_suitable_fallback(area, current_order,
2953 start_migratetype, false, &can_steal);
2954 if (fallback_mt != -1)
2959 * This should not happen - we already found a suitable fallback
2960 * when looking for the largest page.
2962 VM_BUG_ON(current_order == MAX_ORDER);
2965 page = get_page_from_free_area(area, fallback_mt);
2967 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2970 trace_mm_page_alloc_extfrag(page, order, current_order,
2971 start_migratetype, fallback_mt);
2978 * Do the hard work of removing an element from the buddy allocator.
2979 * Call me with the zone->lock already held.
2981 static __always_inline struct page *
2982 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2983 unsigned int alloc_flags)
2987 if (IS_ENABLED(CONFIG_CMA)) {
2989 * Balance movable allocations between regular and CMA areas by
2990 * allocating from CMA when over half of the zone's free memory
2991 * is in the CMA area.
2993 if (alloc_flags & ALLOC_CMA &&
2994 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2995 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2996 page = __rmqueue_cma_fallback(zone, order);
3002 page = __rmqueue_smallest(zone, order, migratetype);
3003 if (unlikely(!page)) {
3004 if (alloc_flags & ALLOC_CMA)
3005 page = __rmqueue_cma_fallback(zone, order);
3007 if (!page && __rmqueue_fallback(zone, order, migratetype,
3013 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3018 * Obtain a specified number of elements from the buddy allocator, all under
3019 * a single hold of the lock, for efficiency. Add them to the supplied list.
3020 * Returns the number of new pages which were placed at *list.
3022 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3023 unsigned long count, struct list_head *list,
3024 int migratetype, unsigned int alloc_flags)
3026 int i, allocated = 0;
3029 * local_lock_irq held so equivalent to spin_lock_irqsave for
3030 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3032 spin_lock(&zone->lock);
3033 for (i = 0; i < count; ++i) {
3034 struct page *page = __rmqueue(zone, order, migratetype,
3036 if (unlikely(page == NULL))
3039 if (unlikely(check_pcp_refill(page)))
3043 * Split buddy pages returned by expand() are received here in
3044 * physical page order. The page is added to the tail of
3045 * caller's list. From the callers perspective, the linked list
3046 * is ordered by page number under some conditions. This is
3047 * useful for IO devices that can forward direction from the
3048 * head, thus also in the physical page order. This is useful
3049 * for IO devices that can merge IO requests if the physical
3050 * pages are ordered properly.
3052 list_add_tail(&page->lru, list);
3054 if (is_migrate_cma(get_pcppage_migratetype(page)))
3055 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3060 * i pages were removed from the buddy list even if some leak due
3061 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3062 * on i. Do not confuse with 'allocated' which is the number of
3063 * pages added to the pcp list.
3065 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3066 spin_unlock(&zone->lock);
3072 * Called from the vmstat counter updater to drain pagesets of this
3073 * currently executing processor on remote nodes after they have
3076 * Note that this function must be called with the thread pinned to
3077 * a single processor.
3079 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3081 unsigned long flags;
3082 int to_drain, batch;
3084 local_lock_irqsave(&pagesets.lock, flags);
3085 batch = READ_ONCE(pcp->batch);
3086 to_drain = min(pcp->count, batch);
3088 free_pcppages_bulk(zone, to_drain, pcp);
3089 local_unlock_irqrestore(&pagesets.lock, flags);
3094 * Drain pcplists of the indicated processor and zone.
3096 * The processor must either be the current processor and the
3097 * thread pinned to the current processor or a processor that
3100 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3102 unsigned long flags;
3103 struct per_cpu_pages *pcp;
3105 local_lock_irqsave(&pagesets.lock, flags);
3107 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3109 free_pcppages_bulk(zone, pcp->count, pcp);
3111 local_unlock_irqrestore(&pagesets.lock, flags);
3115 * Drain pcplists of all zones on the indicated processor.
3117 * The processor must either be the current processor and the
3118 * thread pinned to the current processor or a processor that
3121 static void drain_pages(unsigned int cpu)
3125 for_each_populated_zone(zone) {
3126 drain_pages_zone(cpu, zone);
3131 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3133 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3134 * the single zone's pages.
3136 void drain_local_pages(struct zone *zone)
3138 int cpu = smp_processor_id();
3141 drain_pages_zone(cpu, zone);
3146 static void drain_local_pages_wq(struct work_struct *work)
3148 struct pcpu_drain *drain;
3150 drain = container_of(work, struct pcpu_drain, work);
3153 * drain_all_pages doesn't use proper cpu hotplug protection so
3154 * we can race with cpu offline when the WQ can move this from
3155 * a cpu pinned worker to an unbound one. We can operate on a different
3156 * cpu which is alright but we also have to make sure to not move to
3160 drain_local_pages(drain->zone);
3165 * The implementation of drain_all_pages(), exposing an extra parameter to
3166 * drain on all cpus.
3168 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3169 * not empty. The check for non-emptiness can however race with a free to
3170 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3171 * that need the guarantee that every CPU has drained can disable the
3172 * optimizing racy check.
3174 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3179 * Allocate in the BSS so we won't require allocation in
3180 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3182 static cpumask_t cpus_with_pcps;
3185 * Make sure nobody triggers this path before mm_percpu_wq is fully
3188 if (WARN_ON_ONCE(!mm_percpu_wq))
3192 * Do not drain if one is already in progress unless it's specific to
3193 * a zone. Such callers are primarily CMA and memory hotplug and need
3194 * the drain to be complete when the call returns.
3196 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3199 mutex_lock(&pcpu_drain_mutex);
3203 * We don't care about racing with CPU hotplug event
3204 * as offline notification will cause the notified
3205 * cpu to drain that CPU pcps and on_each_cpu_mask
3206 * disables preemption as part of its processing
3208 for_each_online_cpu(cpu) {
3209 struct per_cpu_pages *pcp;
3211 bool has_pcps = false;
3213 if (force_all_cpus) {
3215 * The pcp.count check is racy, some callers need a
3216 * guarantee that no cpu is missed.
3220 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3224 for_each_populated_zone(z) {
3225 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3234 cpumask_set_cpu(cpu, &cpus_with_pcps);
3236 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3239 for_each_cpu(cpu, &cpus_with_pcps) {
3240 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3243 INIT_WORK(&drain->work, drain_local_pages_wq);
3244 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3246 for_each_cpu(cpu, &cpus_with_pcps)
3247 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3249 mutex_unlock(&pcpu_drain_mutex);
3253 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3255 * When zone parameter is non-NULL, spill just the single zone's pages.
3257 * Note that this can be extremely slow as the draining happens in a workqueue.
3259 void drain_all_pages(struct zone *zone)
3261 __drain_all_pages(zone, false);
3264 #ifdef CONFIG_HIBERNATION
3267 * Touch the watchdog for every WD_PAGE_COUNT pages.
3269 #define WD_PAGE_COUNT (128*1024)
3271 void mark_free_pages(struct zone *zone)
3273 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3274 unsigned long flags;
3275 unsigned int order, t;
3278 if (zone_is_empty(zone))
3281 spin_lock_irqsave(&zone->lock, flags);
3283 max_zone_pfn = zone_end_pfn(zone);
3284 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3285 if (pfn_valid(pfn)) {
3286 page = pfn_to_page(pfn);
3288 if (!--page_count) {
3289 touch_nmi_watchdog();
3290 page_count = WD_PAGE_COUNT;
3293 if (page_zone(page) != zone)
3296 if (!swsusp_page_is_forbidden(page))
3297 swsusp_unset_page_free(page);
3300 for_each_migratetype_order(order, t) {
3301 list_for_each_entry(page,
3302 &zone->free_area[order].free_list[t], lru) {
3305 pfn = page_to_pfn(page);
3306 for (i = 0; i < (1UL << order); i++) {
3307 if (!--page_count) {
3308 touch_nmi_watchdog();
3309 page_count = WD_PAGE_COUNT;
3311 swsusp_set_page_free(pfn_to_page(pfn + i));
3315 spin_unlock_irqrestore(&zone->lock, flags);
3317 #endif /* CONFIG_PM */
3319 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3324 if (!free_pcp_prepare(page, order))
3327 migratetype = get_pfnblock_migratetype(page, pfn);
3328 set_pcppage_migratetype(page, migratetype);
3332 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch)
3334 int min_nr_free, max_nr_free;
3336 /* Check for PCP disabled or boot pageset */
3337 if (unlikely(high < batch))
3340 /* Leave at least pcp->batch pages on the list */
3341 min_nr_free = batch;
3342 max_nr_free = high - batch;
3345 * Double the number of pages freed each time there is subsequent
3346 * freeing of pages without any allocation.
3348 batch <<= pcp->free_factor;
3349 if (batch < max_nr_free)
3351 batch = clamp(batch, min_nr_free, max_nr_free);
3356 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone)
3358 int high = READ_ONCE(pcp->high);
3360 if (unlikely(!high))
3363 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3367 * If reclaim is active, limit the number of pages that can be
3368 * stored on pcp lists
3370 return min(READ_ONCE(pcp->batch) << 2, high);
3373 static void free_unref_page_commit(struct page *page, unsigned long pfn,
3374 int migratetype, unsigned int order)
3376 struct zone *zone = page_zone(page);
3377 struct per_cpu_pages *pcp;
3381 __count_vm_event(PGFREE);
3382 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3383 pindex = order_to_pindex(migratetype, order);
3384 list_add(&page->lru, &pcp->lists[pindex]);
3385 pcp->count += 1 << order;
3386 high = nr_pcp_high(pcp, zone);
3387 if (pcp->count >= high) {
3388 int batch = READ_ONCE(pcp->batch);
3390 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch), pcp);
3397 void free_unref_page(struct page *page, unsigned int order)
3399 unsigned long flags;
3400 unsigned long pfn = page_to_pfn(page);
3403 if (!free_unref_page_prepare(page, pfn, order))
3407 * We only track unmovable, reclaimable and movable on pcp lists.
3408 * Place ISOLATE pages on the isolated list because they are being
3409 * offlined but treat HIGHATOMIC as movable pages so we can get those
3410 * areas back if necessary. Otherwise, we may have to free
3411 * excessively into the page allocator
3413 migratetype = get_pcppage_migratetype(page);
3414 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3415 if (unlikely(is_migrate_isolate(migratetype))) {
3416 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3419 migratetype = MIGRATE_MOVABLE;
3422 local_lock_irqsave(&pagesets.lock, flags);
3423 free_unref_page_commit(page, pfn, migratetype, order);
3424 local_unlock_irqrestore(&pagesets.lock, flags);
3428 * Free a list of 0-order pages
3430 void free_unref_page_list(struct list_head *list)
3432 struct page *page, *next;
3433 unsigned long flags, pfn;
3434 int batch_count = 0;
3437 /* Prepare pages for freeing */
3438 list_for_each_entry_safe(page, next, list, lru) {
3439 pfn = page_to_pfn(page);
3440 if (!free_unref_page_prepare(page, pfn, 0)) {
3441 list_del(&page->lru);
3446 * Free isolated pages directly to the allocator, see
3447 * comment in free_unref_page.
3449 migratetype = get_pcppage_migratetype(page);
3450 if (unlikely(is_migrate_isolate(migratetype))) {
3451 list_del(&page->lru);
3452 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3456 set_page_private(page, pfn);
3459 local_lock_irqsave(&pagesets.lock, flags);
3460 list_for_each_entry_safe(page, next, list, lru) {
3461 pfn = page_private(page);
3462 set_page_private(page, 0);
3465 * Non-isolated types over MIGRATE_PCPTYPES get added
3466 * to the MIGRATE_MOVABLE pcp list.
3468 migratetype = get_pcppage_migratetype(page);
3469 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3470 migratetype = MIGRATE_MOVABLE;
3472 trace_mm_page_free_batched(page);
3473 free_unref_page_commit(page, pfn, migratetype, 0);
3476 * Guard against excessive IRQ disabled times when we get
3477 * a large list of pages to free.
3479 if (++batch_count == SWAP_CLUSTER_MAX) {
3480 local_unlock_irqrestore(&pagesets.lock, flags);
3482 local_lock_irqsave(&pagesets.lock, flags);
3485 local_unlock_irqrestore(&pagesets.lock, flags);
3489 * split_page takes a non-compound higher-order page, and splits it into
3490 * n (1<<order) sub-pages: page[0..n]
3491 * Each sub-page must be freed individually.
3493 * Note: this is probably too low level an operation for use in drivers.
3494 * Please consult with lkml before using this in your driver.
3496 void split_page(struct page *page, unsigned int order)
3500 VM_BUG_ON_PAGE(PageCompound(page), page);
3501 VM_BUG_ON_PAGE(!page_count(page), page);
3503 for (i = 1; i < (1 << order); i++)
3504 set_page_refcounted(page + i);
3505 split_page_owner(page, 1 << order);
3506 split_page_memcg(page, 1 << order);
3508 EXPORT_SYMBOL_GPL(split_page);
3510 int __isolate_free_page(struct page *page, unsigned int order)
3512 unsigned long watermark;
3516 BUG_ON(!PageBuddy(page));
3518 zone = page_zone(page);
3519 mt = get_pageblock_migratetype(page);
3521 if (!is_migrate_isolate(mt)) {
3523 * Obey watermarks as if the page was being allocated. We can
3524 * emulate a high-order watermark check with a raised order-0
3525 * watermark, because we already know our high-order page
3528 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3529 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3532 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3535 /* Remove page from free list */
3537 del_page_from_free_list(page, zone, order);
3540 * Set the pageblock if the isolated page is at least half of a
3543 if (order >= pageblock_order - 1) {
3544 struct page *endpage = page + (1 << order) - 1;
3545 for (; page < endpage; page += pageblock_nr_pages) {
3546 int mt = get_pageblock_migratetype(page);
3547 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3548 && !is_migrate_highatomic(mt))
3549 set_pageblock_migratetype(page,
3555 return 1UL << order;
3559 * __putback_isolated_page - Return a now-isolated page back where we got it
3560 * @page: Page that was isolated
3561 * @order: Order of the isolated page
3562 * @mt: The page's pageblock's migratetype
3564 * This function is meant to return a page pulled from the free lists via
3565 * __isolate_free_page back to the free lists they were pulled from.
3567 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3569 struct zone *zone = page_zone(page);
3571 /* zone lock should be held when this function is called */
3572 lockdep_assert_held(&zone->lock);
3574 /* Return isolated page to tail of freelist. */
3575 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3576 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3580 * Update NUMA hit/miss statistics
3582 * Must be called with interrupts disabled.
3584 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3588 enum numa_stat_item local_stat = NUMA_LOCAL;
3590 /* skip numa counters update if numa stats is disabled */
3591 if (!static_branch_likely(&vm_numa_stat_key))
3594 if (zone_to_nid(z) != numa_node_id())
3595 local_stat = NUMA_OTHER;
3597 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3598 __count_numa_events(z, NUMA_HIT, nr_account);
3600 __count_numa_events(z, NUMA_MISS, nr_account);
3601 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3603 __count_numa_events(z, local_stat, nr_account);
3607 /* Remove page from the per-cpu list, caller must protect the list */
3609 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3611 unsigned int alloc_flags,
3612 struct per_cpu_pages *pcp,
3613 struct list_head *list)
3618 if (list_empty(list)) {
3619 int batch = READ_ONCE(pcp->batch);
3623 * Scale batch relative to order if batch implies
3624 * free pages can be stored on the PCP. Batch can
3625 * be 1 for small zones or for boot pagesets which
3626 * should never store free pages as the pages may
3627 * belong to arbitrary zones.
3630 batch = max(batch >> order, 2);
3631 alloced = rmqueue_bulk(zone, order,
3633 migratetype, alloc_flags);
3635 pcp->count += alloced << order;
3636 if (unlikely(list_empty(list)))
3640 page = list_first_entry(list, struct page, lru);
3641 list_del(&page->lru);
3642 pcp->count -= 1 << order;
3643 } while (check_new_pcp(page));
3648 /* Lock and remove page from the per-cpu list */
3649 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3650 struct zone *zone, unsigned int order,
3651 gfp_t gfp_flags, int migratetype,
3652 unsigned int alloc_flags)
3654 struct per_cpu_pages *pcp;
3655 struct list_head *list;
3657 unsigned long flags;
3659 local_lock_irqsave(&pagesets.lock, flags);
3662 * On allocation, reduce the number of pages that are batch freed.
3663 * See nr_pcp_free() where free_factor is increased for subsequent
3666 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3667 pcp->free_factor >>= 1;
3668 list = &pcp->lists[order_to_pindex(migratetype, order)];
3669 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3670 local_unlock_irqrestore(&pagesets.lock, flags);
3672 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3673 zone_statistics(preferred_zone, zone, 1);
3679 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3682 struct page *rmqueue(struct zone *preferred_zone,
3683 struct zone *zone, unsigned int order,
3684 gfp_t gfp_flags, unsigned int alloc_flags,
3687 unsigned long flags;
3690 if (likely(pcp_allowed_order(order))) {
3692 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3693 * we need to skip it when CMA area isn't allowed.
3695 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3696 migratetype != MIGRATE_MOVABLE) {
3697 page = rmqueue_pcplist(preferred_zone, zone, order,
3698 gfp_flags, migratetype, alloc_flags);
3704 * We most definitely don't want callers attempting to
3705 * allocate greater than order-1 page units with __GFP_NOFAIL.
3707 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3708 spin_lock_irqsave(&zone->lock, flags);
3713 * order-0 request can reach here when the pcplist is skipped
3714 * due to non-CMA allocation context. HIGHATOMIC area is
3715 * reserved for high-order atomic allocation, so order-0
3716 * request should skip it.
3718 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3719 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3721 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3724 page = __rmqueue(zone, order, migratetype, alloc_flags);
3725 } while (page && check_new_pages(page, order));
3729 __mod_zone_freepage_state(zone, -(1 << order),
3730 get_pcppage_migratetype(page));
3731 spin_unlock_irqrestore(&zone->lock, flags);
3733 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3734 zone_statistics(preferred_zone, zone, 1);
3737 /* Separate test+clear to avoid unnecessary atomics */
3738 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3739 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3740 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3743 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3747 spin_unlock_irqrestore(&zone->lock, flags);
3751 #ifdef CONFIG_FAIL_PAGE_ALLOC
3754 struct fault_attr attr;
3756 bool ignore_gfp_highmem;
3757 bool ignore_gfp_reclaim;
3759 } fail_page_alloc = {
3760 .attr = FAULT_ATTR_INITIALIZER,
3761 .ignore_gfp_reclaim = true,
3762 .ignore_gfp_highmem = true,
3766 static int __init setup_fail_page_alloc(char *str)
3768 return setup_fault_attr(&fail_page_alloc.attr, str);
3770 __setup("fail_page_alloc=", setup_fail_page_alloc);
3772 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3774 if (order < fail_page_alloc.min_order)
3776 if (gfp_mask & __GFP_NOFAIL)
3778 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3780 if (fail_page_alloc.ignore_gfp_reclaim &&
3781 (gfp_mask & __GFP_DIRECT_RECLAIM))
3784 return should_fail(&fail_page_alloc.attr, 1 << order);
3787 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3789 static int __init fail_page_alloc_debugfs(void)
3791 umode_t mode = S_IFREG | 0600;
3794 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3795 &fail_page_alloc.attr);
3797 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3798 &fail_page_alloc.ignore_gfp_reclaim);
3799 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3800 &fail_page_alloc.ignore_gfp_highmem);
3801 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3806 late_initcall(fail_page_alloc_debugfs);
3808 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3810 #else /* CONFIG_FAIL_PAGE_ALLOC */
3812 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3817 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3819 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3821 return __should_fail_alloc_page(gfp_mask, order);
3823 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3825 static inline long __zone_watermark_unusable_free(struct zone *z,
3826 unsigned int order, unsigned int alloc_flags)
3828 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3829 long unusable_free = (1 << order) - 1;
3832 * If the caller does not have rights to ALLOC_HARDER then subtract
3833 * the high-atomic reserves. This will over-estimate the size of the
3834 * atomic reserve but it avoids a search.
3836 if (likely(!alloc_harder))
3837 unusable_free += z->nr_reserved_highatomic;
3840 /* If allocation can't use CMA areas don't use free CMA pages */
3841 if (!(alloc_flags & ALLOC_CMA))
3842 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3845 return unusable_free;
3849 * Return true if free base pages are above 'mark'. For high-order checks it
3850 * will return true of the order-0 watermark is reached and there is at least
3851 * one free page of a suitable size. Checking now avoids taking the zone lock
3852 * to check in the allocation paths if no pages are free.
3854 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3855 int highest_zoneidx, unsigned int alloc_flags,
3860 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3862 /* free_pages may go negative - that's OK */
3863 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3865 if (alloc_flags & ALLOC_HIGH)
3868 if (unlikely(alloc_harder)) {
3870 * OOM victims can try even harder than normal ALLOC_HARDER
3871 * users on the grounds that it's definitely going to be in
3872 * the exit path shortly and free memory. Any allocation it
3873 * makes during the free path will be small and short-lived.
3875 if (alloc_flags & ALLOC_OOM)
3882 * Check watermarks for an order-0 allocation request. If these
3883 * are not met, then a high-order request also cannot go ahead
3884 * even if a suitable page happened to be free.
3886 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3889 /* If this is an order-0 request then the watermark is fine */
3893 /* For a high-order request, check at least one suitable page is free */
3894 for (o = order; o < MAX_ORDER; o++) {
3895 struct free_area *area = &z->free_area[o];
3901 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3902 if (!free_area_empty(area, mt))
3907 if ((alloc_flags & ALLOC_CMA) &&
3908 !free_area_empty(area, MIGRATE_CMA)) {
3912 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3918 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3919 int highest_zoneidx, unsigned int alloc_flags)
3921 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3922 zone_page_state(z, NR_FREE_PAGES));
3925 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3926 unsigned long mark, int highest_zoneidx,
3927 unsigned int alloc_flags, gfp_t gfp_mask)
3931 free_pages = zone_page_state(z, NR_FREE_PAGES);
3934 * Fast check for order-0 only. If this fails then the reserves
3935 * need to be calculated.
3940 fast_free = free_pages;
3941 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3942 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3946 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3950 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3951 * when checking the min watermark. The min watermark is the
3952 * point where boosting is ignored so that kswapd is woken up
3953 * when below the low watermark.
3955 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3956 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3957 mark = z->_watermark[WMARK_MIN];
3958 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3959 alloc_flags, free_pages);
3965 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3966 unsigned long mark, int highest_zoneidx)
3968 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3970 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3971 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3973 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3978 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3980 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3982 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3983 node_reclaim_distance;
3985 #else /* CONFIG_NUMA */
3986 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3990 #endif /* CONFIG_NUMA */
3993 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3994 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3995 * premature use of a lower zone may cause lowmem pressure problems that
3996 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3997 * probably too small. It only makes sense to spread allocations to avoid
3998 * fragmentation between the Normal and DMA32 zones.
4000 static inline unsigned int
4001 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4003 unsigned int alloc_flags;
4006 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4009 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4011 #ifdef CONFIG_ZONE_DMA32
4015 if (zone_idx(zone) != ZONE_NORMAL)
4019 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4020 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4021 * on UMA that if Normal is populated then so is DMA32.
4023 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4024 if (nr_online_nodes > 1 && !populated_zone(--zone))
4027 alloc_flags |= ALLOC_NOFRAGMENT;
4028 #endif /* CONFIG_ZONE_DMA32 */
4032 /* Must be called after current_gfp_context() which can change gfp_mask */
4033 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4034 unsigned int alloc_flags)
4037 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4038 alloc_flags |= ALLOC_CMA;
4044 * get_page_from_freelist goes through the zonelist trying to allocate
4047 static struct page *
4048 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4049 const struct alloc_context *ac)
4053 struct pglist_data *last_pgdat_dirty_limit = NULL;
4058 * Scan zonelist, looking for a zone with enough free.
4059 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4061 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4062 z = ac->preferred_zoneref;
4063 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4068 if (cpusets_enabled() &&
4069 (alloc_flags & ALLOC_CPUSET) &&
4070 !__cpuset_zone_allowed(zone, gfp_mask))
4073 * When allocating a page cache page for writing, we
4074 * want to get it from a node that is within its dirty
4075 * limit, such that no single node holds more than its
4076 * proportional share of globally allowed dirty pages.
4077 * The dirty limits take into account the node's
4078 * lowmem reserves and high watermark so that kswapd
4079 * should be able to balance it without having to
4080 * write pages from its LRU list.
4082 * XXX: For now, allow allocations to potentially
4083 * exceed the per-node dirty limit in the slowpath
4084 * (spread_dirty_pages unset) before going into reclaim,
4085 * which is important when on a NUMA setup the allowed
4086 * nodes are together not big enough to reach the
4087 * global limit. The proper fix for these situations
4088 * will require awareness of nodes in the
4089 * dirty-throttling and the flusher threads.
4091 if (ac->spread_dirty_pages) {
4092 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4095 if (!node_dirty_ok(zone->zone_pgdat)) {
4096 last_pgdat_dirty_limit = zone->zone_pgdat;
4101 if (no_fallback && nr_online_nodes > 1 &&
4102 zone != ac->preferred_zoneref->zone) {
4106 * If moving to a remote node, retry but allow
4107 * fragmenting fallbacks. Locality is more important
4108 * than fragmentation avoidance.
4110 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4111 if (zone_to_nid(zone) != local_nid) {
4112 alloc_flags &= ~ALLOC_NOFRAGMENT;
4117 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4118 if (!zone_watermark_fast(zone, order, mark,
4119 ac->highest_zoneidx, alloc_flags,
4123 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4125 * Watermark failed for this zone, but see if we can
4126 * grow this zone if it contains deferred pages.
4128 if (static_branch_unlikely(&deferred_pages)) {
4129 if (_deferred_grow_zone(zone, order))
4133 /* Checked here to keep the fast path fast */
4134 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4135 if (alloc_flags & ALLOC_NO_WATERMARKS)
4138 if (!node_reclaim_enabled() ||
4139 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4142 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4144 case NODE_RECLAIM_NOSCAN:
4147 case NODE_RECLAIM_FULL:
4148 /* scanned but unreclaimable */
4151 /* did we reclaim enough */
4152 if (zone_watermark_ok(zone, order, mark,
4153 ac->highest_zoneidx, alloc_flags))
4161 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4162 gfp_mask, alloc_flags, ac->migratetype);
4164 prep_new_page(page, order, gfp_mask, alloc_flags);
4167 * If this is a high-order atomic allocation then check
4168 * if the pageblock should be reserved for the future
4170 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4171 reserve_highatomic_pageblock(page, zone, order);
4175 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4176 /* Try again if zone has deferred pages */
4177 if (static_branch_unlikely(&deferred_pages)) {
4178 if (_deferred_grow_zone(zone, order))
4186 * It's possible on a UMA machine to get through all zones that are
4187 * fragmented. If avoiding fragmentation, reset and try again.
4190 alloc_flags &= ~ALLOC_NOFRAGMENT;
4197 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4199 unsigned int filter = SHOW_MEM_FILTER_NODES;
4202 * This documents exceptions given to allocations in certain
4203 * contexts that are allowed to allocate outside current's set
4206 if (!(gfp_mask & __GFP_NOMEMALLOC))
4207 if (tsk_is_oom_victim(current) ||
4208 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4209 filter &= ~SHOW_MEM_FILTER_NODES;
4210 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4211 filter &= ~SHOW_MEM_FILTER_NODES;
4213 show_mem(filter, nodemask);
4216 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4218 struct va_format vaf;
4220 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4222 if ((gfp_mask & __GFP_NOWARN) ||
4223 !__ratelimit(&nopage_rs) ||
4224 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4227 va_start(args, fmt);
4230 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4231 current->comm, &vaf, gfp_mask, &gfp_mask,
4232 nodemask_pr_args(nodemask));
4235 cpuset_print_current_mems_allowed();
4238 warn_alloc_show_mem(gfp_mask, nodemask);
4241 static inline struct page *
4242 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4243 unsigned int alloc_flags,
4244 const struct alloc_context *ac)
4248 page = get_page_from_freelist(gfp_mask, order,
4249 alloc_flags|ALLOC_CPUSET, ac);
4251 * fallback to ignore cpuset restriction if our nodes
4255 page = get_page_from_freelist(gfp_mask, order,
4261 static inline struct page *
4262 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4263 const struct alloc_context *ac, unsigned long *did_some_progress)
4265 struct oom_control oc = {
4266 .zonelist = ac->zonelist,
4267 .nodemask = ac->nodemask,
4269 .gfp_mask = gfp_mask,
4274 *did_some_progress = 0;
4277 * Acquire the oom lock. If that fails, somebody else is
4278 * making progress for us.
4280 if (!mutex_trylock(&oom_lock)) {
4281 *did_some_progress = 1;
4282 schedule_timeout_uninterruptible(1);
4287 * Go through the zonelist yet one more time, keep very high watermark
4288 * here, this is only to catch a parallel oom killing, we must fail if
4289 * we're still under heavy pressure. But make sure that this reclaim
4290 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4291 * allocation which will never fail due to oom_lock already held.
4293 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4294 ~__GFP_DIRECT_RECLAIM, order,
4295 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4299 /* Coredumps can quickly deplete all memory reserves */
4300 if (current->flags & PF_DUMPCORE)
4302 /* The OOM killer will not help higher order allocs */
4303 if (order > PAGE_ALLOC_COSTLY_ORDER)
4306 * We have already exhausted all our reclaim opportunities without any
4307 * success so it is time to admit defeat. We will skip the OOM killer
4308 * because it is very likely that the caller has a more reasonable
4309 * fallback than shooting a random task.
4311 * The OOM killer may not free memory on a specific node.
4313 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4315 /* The OOM killer does not needlessly kill tasks for lowmem */
4316 if (ac->highest_zoneidx < ZONE_NORMAL)
4318 if (pm_suspended_storage())
4321 * XXX: GFP_NOFS allocations should rather fail than rely on
4322 * other request to make a forward progress.
4323 * We are in an unfortunate situation where out_of_memory cannot
4324 * do much for this context but let's try it to at least get
4325 * access to memory reserved if the current task is killed (see
4326 * out_of_memory). Once filesystems are ready to handle allocation
4327 * failures more gracefully we should just bail out here.
4330 /* Exhausted what can be done so it's blame time */
4331 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4332 *did_some_progress = 1;
4335 * Help non-failing allocations by giving them access to memory
4338 if (gfp_mask & __GFP_NOFAIL)
4339 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4340 ALLOC_NO_WATERMARKS, ac);
4343 mutex_unlock(&oom_lock);
4348 * Maximum number of compaction retries with a progress before OOM
4349 * killer is consider as the only way to move forward.
4351 #define MAX_COMPACT_RETRIES 16
4353 #ifdef CONFIG_COMPACTION
4354 /* Try memory compaction for high-order allocations before reclaim */
4355 static struct page *
4356 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4357 unsigned int alloc_flags, const struct alloc_context *ac,
4358 enum compact_priority prio, enum compact_result *compact_result)
4360 struct page *page = NULL;
4361 unsigned long pflags;
4362 unsigned int noreclaim_flag;
4367 psi_memstall_enter(&pflags);
4368 noreclaim_flag = memalloc_noreclaim_save();
4370 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4373 memalloc_noreclaim_restore(noreclaim_flag);
4374 psi_memstall_leave(&pflags);
4376 if (*compact_result == COMPACT_SKIPPED)
4379 * At least in one zone compaction wasn't deferred or skipped, so let's
4380 * count a compaction stall
4382 count_vm_event(COMPACTSTALL);
4384 /* Prep a captured page if available */
4386 prep_new_page(page, order, gfp_mask, alloc_flags);
4388 /* Try get a page from the freelist if available */
4390 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4393 struct zone *zone = page_zone(page);
4395 zone->compact_blockskip_flush = false;
4396 compaction_defer_reset(zone, order, true);
4397 count_vm_event(COMPACTSUCCESS);
4402 * It's bad if compaction run occurs and fails. The most likely reason
4403 * is that pages exist, but not enough to satisfy watermarks.
4405 count_vm_event(COMPACTFAIL);
4413 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4414 enum compact_result compact_result,
4415 enum compact_priority *compact_priority,
4416 int *compaction_retries)
4418 int max_retries = MAX_COMPACT_RETRIES;
4421 int retries = *compaction_retries;
4422 enum compact_priority priority = *compact_priority;
4427 if (fatal_signal_pending(current))
4430 if (compaction_made_progress(compact_result))
4431 (*compaction_retries)++;
4434 * compaction considers all the zone as desperately out of memory
4435 * so it doesn't really make much sense to retry except when the
4436 * failure could be caused by insufficient priority
4438 if (compaction_failed(compact_result))
4439 goto check_priority;
4442 * compaction was skipped because there are not enough order-0 pages
4443 * to work with, so we retry only if it looks like reclaim can help.
4445 if (compaction_needs_reclaim(compact_result)) {
4446 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4451 * make sure the compaction wasn't deferred or didn't bail out early
4452 * due to locks contention before we declare that we should give up.
4453 * But the next retry should use a higher priority if allowed, so
4454 * we don't just keep bailing out endlessly.
4456 if (compaction_withdrawn(compact_result)) {
4457 goto check_priority;
4461 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4462 * costly ones because they are de facto nofail and invoke OOM
4463 * killer to move on while costly can fail and users are ready
4464 * to cope with that. 1/4 retries is rather arbitrary but we
4465 * would need much more detailed feedback from compaction to
4466 * make a better decision.
4468 if (order > PAGE_ALLOC_COSTLY_ORDER)
4470 if (*compaction_retries <= max_retries) {
4476 * Make sure there are attempts at the highest priority if we exhausted
4477 * all retries or failed at the lower priorities.
4480 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4481 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4483 if (*compact_priority > min_priority) {
4484 (*compact_priority)--;
4485 *compaction_retries = 0;
4489 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4493 static inline struct page *
4494 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4495 unsigned int alloc_flags, const struct alloc_context *ac,
4496 enum compact_priority prio, enum compact_result *compact_result)
4498 *compact_result = COMPACT_SKIPPED;
4503 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4504 enum compact_result compact_result,
4505 enum compact_priority *compact_priority,
4506 int *compaction_retries)
4511 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4515 * There are setups with compaction disabled which would prefer to loop
4516 * inside the allocator rather than hit the oom killer prematurely.
4517 * Let's give them a good hope and keep retrying while the order-0
4518 * watermarks are OK.
4520 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4521 ac->highest_zoneidx, ac->nodemask) {
4522 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4523 ac->highest_zoneidx, alloc_flags))
4528 #endif /* CONFIG_COMPACTION */
4530 #ifdef CONFIG_LOCKDEP
4531 static struct lockdep_map __fs_reclaim_map =
4532 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4534 static bool __need_reclaim(gfp_t gfp_mask)
4536 /* no reclaim without waiting on it */
4537 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4540 /* this guy won't enter reclaim */
4541 if (current->flags & PF_MEMALLOC)
4544 if (gfp_mask & __GFP_NOLOCKDEP)
4550 void __fs_reclaim_acquire(unsigned long ip)
4552 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4555 void __fs_reclaim_release(unsigned long ip)
4557 lock_release(&__fs_reclaim_map, ip);
4560 void fs_reclaim_acquire(gfp_t gfp_mask)
4562 gfp_mask = current_gfp_context(gfp_mask);
4564 if (__need_reclaim(gfp_mask)) {
4565 if (gfp_mask & __GFP_FS)
4566 __fs_reclaim_acquire(_RET_IP_);
4568 #ifdef CONFIG_MMU_NOTIFIER
4569 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4570 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4575 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4577 void fs_reclaim_release(gfp_t gfp_mask)
4579 gfp_mask = current_gfp_context(gfp_mask);
4581 if (__need_reclaim(gfp_mask)) {
4582 if (gfp_mask & __GFP_FS)
4583 __fs_reclaim_release(_RET_IP_);
4586 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4589 /* Perform direct synchronous page reclaim */
4590 static unsigned long
4591 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4592 const struct alloc_context *ac)
4594 unsigned int noreclaim_flag;
4595 unsigned long pflags, progress;
4599 /* We now go into synchronous reclaim */
4600 cpuset_memory_pressure_bump();
4601 psi_memstall_enter(&pflags);
4602 fs_reclaim_acquire(gfp_mask);
4603 noreclaim_flag = memalloc_noreclaim_save();
4605 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4608 memalloc_noreclaim_restore(noreclaim_flag);
4609 fs_reclaim_release(gfp_mask);
4610 psi_memstall_leave(&pflags);
4617 /* The really slow allocator path where we enter direct reclaim */
4618 static inline struct page *
4619 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4620 unsigned int alloc_flags, const struct alloc_context *ac,
4621 unsigned long *did_some_progress)
4623 struct page *page = NULL;
4624 bool drained = false;
4626 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4627 if (unlikely(!(*did_some_progress)))
4631 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4634 * If an allocation failed after direct reclaim, it could be because
4635 * pages are pinned on the per-cpu lists or in high alloc reserves.
4636 * Shrink them and try again
4638 if (!page && !drained) {
4639 unreserve_highatomic_pageblock(ac, false);
4640 drain_all_pages(NULL);
4648 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4649 const struct alloc_context *ac)
4653 pg_data_t *last_pgdat = NULL;
4654 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4656 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4658 if (last_pgdat != zone->zone_pgdat)
4659 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4660 last_pgdat = zone->zone_pgdat;
4664 static inline unsigned int
4665 gfp_to_alloc_flags(gfp_t gfp_mask)
4667 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4670 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4671 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4672 * to save two branches.
4674 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4675 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4678 * The caller may dip into page reserves a bit more if the caller
4679 * cannot run direct reclaim, or if the caller has realtime scheduling
4680 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4681 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4683 alloc_flags |= (__force int)
4684 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4686 if (gfp_mask & __GFP_ATOMIC) {
4688 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4689 * if it can't schedule.
4691 if (!(gfp_mask & __GFP_NOMEMALLOC))
4692 alloc_flags |= ALLOC_HARDER;
4694 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4695 * comment for __cpuset_node_allowed().
4697 alloc_flags &= ~ALLOC_CPUSET;
4698 } else if (unlikely(rt_task(current)) && in_task())
4699 alloc_flags |= ALLOC_HARDER;
4701 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4706 static bool oom_reserves_allowed(struct task_struct *tsk)
4708 if (!tsk_is_oom_victim(tsk))
4712 * !MMU doesn't have oom reaper so give access to memory reserves
4713 * only to the thread with TIF_MEMDIE set
4715 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4722 * Distinguish requests which really need access to full memory
4723 * reserves from oom victims which can live with a portion of it
4725 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4727 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4729 if (gfp_mask & __GFP_MEMALLOC)
4730 return ALLOC_NO_WATERMARKS;
4731 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4732 return ALLOC_NO_WATERMARKS;
4733 if (!in_interrupt()) {
4734 if (current->flags & PF_MEMALLOC)
4735 return ALLOC_NO_WATERMARKS;
4736 else if (oom_reserves_allowed(current))
4743 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4745 return !!__gfp_pfmemalloc_flags(gfp_mask);
4749 * Checks whether it makes sense to retry the reclaim to make a forward progress
4750 * for the given allocation request.
4752 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4753 * without success, or when we couldn't even meet the watermark if we
4754 * reclaimed all remaining pages on the LRU lists.
4756 * Returns true if a retry is viable or false to enter the oom path.
4759 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4760 struct alloc_context *ac, int alloc_flags,
4761 bool did_some_progress, int *no_progress_loops)
4768 * Costly allocations might have made a progress but this doesn't mean
4769 * their order will become available due to high fragmentation so
4770 * always increment the no progress counter for them
4772 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4773 *no_progress_loops = 0;
4775 (*no_progress_loops)++;
4778 * Make sure we converge to OOM if we cannot make any progress
4779 * several times in the row.
4781 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4782 /* Before OOM, exhaust highatomic_reserve */
4783 return unreserve_highatomic_pageblock(ac, true);
4787 * Keep reclaiming pages while there is a chance this will lead
4788 * somewhere. If none of the target zones can satisfy our allocation
4789 * request even if all reclaimable pages are considered then we are
4790 * screwed and have to go OOM.
4792 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4793 ac->highest_zoneidx, ac->nodemask) {
4794 unsigned long available;
4795 unsigned long reclaimable;
4796 unsigned long min_wmark = min_wmark_pages(zone);
4799 available = reclaimable = zone_reclaimable_pages(zone);
4800 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4803 * Would the allocation succeed if we reclaimed all
4804 * reclaimable pages?
4806 wmark = __zone_watermark_ok(zone, order, min_wmark,
4807 ac->highest_zoneidx, alloc_flags, available);
4808 trace_reclaim_retry_zone(z, order, reclaimable,
4809 available, min_wmark, *no_progress_loops, wmark);
4817 * Memory allocation/reclaim might be called from a WQ context and the
4818 * current implementation of the WQ concurrency control doesn't
4819 * recognize that a particular WQ is congested if the worker thread is
4820 * looping without ever sleeping. Therefore we have to do a short sleep
4821 * here rather than calling cond_resched().
4823 if (current->flags & PF_WQ_WORKER)
4824 schedule_timeout_uninterruptible(1);
4831 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4834 * It's possible that cpuset's mems_allowed and the nodemask from
4835 * mempolicy don't intersect. This should be normally dealt with by
4836 * policy_nodemask(), but it's possible to race with cpuset update in
4837 * such a way the check therein was true, and then it became false
4838 * before we got our cpuset_mems_cookie here.
4839 * This assumes that for all allocations, ac->nodemask can come only
4840 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4841 * when it does not intersect with the cpuset restrictions) or the
4842 * caller can deal with a violated nodemask.
4844 if (cpusets_enabled() && ac->nodemask &&
4845 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4846 ac->nodemask = NULL;
4851 * When updating a task's mems_allowed or mempolicy nodemask, it is
4852 * possible to race with parallel threads in such a way that our
4853 * allocation can fail while the mask is being updated. If we are about
4854 * to fail, check if the cpuset changed during allocation and if so,
4857 if (read_mems_allowed_retry(cpuset_mems_cookie))
4863 static inline struct page *
4864 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4865 struct alloc_context *ac)
4867 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4868 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4869 struct page *page = NULL;
4870 unsigned int alloc_flags;
4871 unsigned long did_some_progress;
4872 enum compact_priority compact_priority;
4873 enum compact_result compact_result;
4874 int compaction_retries;
4875 int no_progress_loops;
4876 unsigned int cpuset_mems_cookie;
4880 * We also sanity check to catch abuse of atomic reserves being used by
4881 * callers that are not in atomic context.
4883 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4884 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4885 gfp_mask &= ~__GFP_ATOMIC;
4888 compaction_retries = 0;
4889 no_progress_loops = 0;
4890 compact_priority = DEF_COMPACT_PRIORITY;
4891 cpuset_mems_cookie = read_mems_allowed_begin();
4894 * The fast path uses conservative alloc_flags to succeed only until
4895 * kswapd needs to be woken up, and to avoid the cost of setting up
4896 * alloc_flags precisely. So we do that now.
4898 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4901 * We need to recalculate the starting point for the zonelist iterator
4902 * because we might have used different nodemask in the fast path, or
4903 * there was a cpuset modification and we are retrying - otherwise we
4904 * could end up iterating over non-eligible zones endlessly.
4906 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4907 ac->highest_zoneidx, ac->nodemask);
4908 if (!ac->preferred_zoneref->zone)
4912 * Check for insane configurations where the cpuset doesn't contain
4913 * any suitable zone to satisfy the request - e.g. non-movable
4914 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4916 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4917 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4918 ac->highest_zoneidx,
4919 &cpuset_current_mems_allowed);
4924 if (alloc_flags & ALLOC_KSWAPD)
4925 wake_all_kswapds(order, gfp_mask, ac);
4928 * The adjusted alloc_flags might result in immediate success, so try
4931 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4936 * For costly allocations, try direct compaction first, as it's likely
4937 * that we have enough base pages and don't need to reclaim. For non-
4938 * movable high-order allocations, do that as well, as compaction will
4939 * try prevent permanent fragmentation by migrating from blocks of the
4941 * Don't try this for allocations that are allowed to ignore
4942 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4944 if (can_direct_reclaim &&
4946 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4947 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4948 page = __alloc_pages_direct_compact(gfp_mask, order,
4950 INIT_COMPACT_PRIORITY,
4956 * Checks for costly allocations with __GFP_NORETRY, which
4957 * includes some THP page fault allocations
4959 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4961 * If allocating entire pageblock(s) and compaction
4962 * failed because all zones are below low watermarks
4963 * or is prohibited because it recently failed at this
4964 * order, fail immediately unless the allocator has
4965 * requested compaction and reclaim retry.
4968 * - potentially very expensive because zones are far
4969 * below their low watermarks or this is part of very
4970 * bursty high order allocations,
4971 * - not guaranteed to help because isolate_freepages()
4972 * may not iterate over freed pages as part of its
4974 * - unlikely to make entire pageblocks free on its
4977 if (compact_result == COMPACT_SKIPPED ||
4978 compact_result == COMPACT_DEFERRED)
4982 * Looks like reclaim/compaction is worth trying, but
4983 * sync compaction could be very expensive, so keep
4984 * using async compaction.
4986 compact_priority = INIT_COMPACT_PRIORITY;
4991 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4992 if (alloc_flags & ALLOC_KSWAPD)
4993 wake_all_kswapds(order, gfp_mask, ac);
4995 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4997 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
5000 * Reset the nodemask and zonelist iterators if memory policies can be
5001 * ignored. These allocations are high priority and system rather than
5004 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5005 ac->nodemask = NULL;
5006 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5007 ac->highest_zoneidx, ac->nodemask);
5010 /* Attempt with potentially adjusted zonelist and alloc_flags */
5011 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5015 /* Caller is not willing to reclaim, we can't balance anything */
5016 if (!can_direct_reclaim)
5019 /* Avoid recursion of direct reclaim */
5020 if (current->flags & PF_MEMALLOC)
5023 /* Try direct reclaim and then allocating */
5024 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5025 &did_some_progress);
5029 /* Try direct compaction and then allocating */
5030 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5031 compact_priority, &compact_result);
5035 /* Do not loop if specifically requested */
5036 if (gfp_mask & __GFP_NORETRY)
5040 * Do not retry costly high order allocations unless they are
5041 * __GFP_RETRY_MAYFAIL
5043 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5046 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5047 did_some_progress > 0, &no_progress_loops))
5051 * It doesn't make any sense to retry for the compaction if the order-0
5052 * reclaim is not able to make any progress because the current
5053 * implementation of the compaction depends on the sufficient amount
5054 * of free memory (see __compaction_suitable)
5056 if (did_some_progress > 0 &&
5057 should_compact_retry(ac, order, alloc_flags,
5058 compact_result, &compact_priority,
5059 &compaction_retries))
5063 /* Deal with possible cpuset update races before we start OOM killing */
5064 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5067 /* Reclaim has failed us, start killing things */
5068 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5072 /* Avoid allocations with no watermarks from looping endlessly */
5073 if (tsk_is_oom_victim(current) &&
5074 (alloc_flags & ALLOC_OOM ||
5075 (gfp_mask & __GFP_NOMEMALLOC)))
5078 /* Retry as long as the OOM killer is making progress */
5079 if (did_some_progress) {
5080 no_progress_loops = 0;
5085 /* Deal with possible cpuset update races before we fail */
5086 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5090 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5093 if (gfp_mask & __GFP_NOFAIL) {
5095 * All existing users of the __GFP_NOFAIL are blockable, so warn
5096 * of any new users that actually require GFP_NOWAIT
5098 if (WARN_ON_ONCE(!can_direct_reclaim))
5102 * PF_MEMALLOC request from this context is rather bizarre
5103 * because we cannot reclaim anything and only can loop waiting
5104 * for somebody to do a work for us
5106 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5109 * non failing costly orders are a hard requirement which we
5110 * are not prepared for much so let's warn about these users
5111 * so that we can identify them and convert them to something
5114 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5117 * Help non-failing allocations by giving them access to memory
5118 * reserves but do not use ALLOC_NO_WATERMARKS because this
5119 * could deplete whole memory reserves which would just make
5120 * the situation worse
5122 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5130 warn_alloc(gfp_mask, ac->nodemask,
5131 "page allocation failure: order:%u", order);
5136 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5137 int preferred_nid, nodemask_t *nodemask,
5138 struct alloc_context *ac, gfp_t *alloc_gfp,
5139 unsigned int *alloc_flags)
5141 ac->highest_zoneidx = gfp_zone(gfp_mask);
5142 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5143 ac->nodemask = nodemask;
5144 ac->migratetype = gfp_migratetype(gfp_mask);
5146 if (cpusets_enabled()) {
5147 *alloc_gfp |= __GFP_HARDWALL;
5149 * When we are in the interrupt context, it is irrelevant
5150 * to the current task context. It means that any node ok.
5152 if (in_task() && !ac->nodemask)
5153 ac->nodemask = &cpuset_current_mems_allowed;
5155 *alloc_flags |= ALLOC_CPUSET;
5158 fs_reclaim_acquire(gfp_mask);
5159 fs_reclaim_release(gfp_mask);
5161 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5163 if (should_fail_alloc_page(gfp_mask, order))
5166 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5168 /* Dirty zone balancing only done in the fast path */
5169 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5172 * The preferred zone is used for statistics but crucially it is
5173 * also used as the starting point for the zonelist iterator. It
5174 * may get reset for allocations that ignore memory policies.
5176 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5177 ac->highest_zoneidx, ac->nodemask);
5183 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5184 * @gfp: GFP flags for the allocation
5185 * @preferred_nid: The preferred NUMA node ID to allocate from
5186 * @nodemask: Set of nodes to allocate from, may be NULL
5187 * @nr_pages: The number of pages desired on the list or array
5188 * @page_list: Optional list to store the allocated pages
5189 * @page_array: Optional array to store the pages
5191 * This is a batched version of the page allocator that attempts to
5192 * allocate nr_pages quickly. Pages are added to page_list if page_list
5193 * is not NULL, otherwise it is assumed that the page_array is valid.
5195 * For lists, nr_pages is the number of pages that should be allocated.
5197 * For arrays, only NULL elements are populated with pages and nr_pages
5198 * is the maximum number of pages that will be stored in the array.
5200 * Returns the number of pages on the list or array.
5202 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5203 nodemask_t *nodemask, int nr_pages,
5204 struct list_head *page_list,
5205 struct page **page_array)
5208 unsigned long flags;
5211 struct per_cpu_pages *pcp;
5212 struct list_head *pcp_list;
5213 struct alloc_context ac;
5215 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5216 int nr_populated = 0, nr_account = 0;
5219 * Skip populated array elements to determine if any pages need
5220 * to be allocated before disabling IRQs.
5222 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5225 /* No pages requested? */
5226 if (unlikely(nr_pages <= 0))
5229 /* Already populated array? */
5230 if (unlikely(page_array && nr_pages - nr_populated == 0))
5233 /* Bulk allocator does not support memcg accounting. */
5234 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5237 /* Use the single page allocator for one page. */
5238 if (nr_pages - nr_populated == 1)
5241 #ifdef CONFIG_PAGE_OWNER
5243 * PAGE_OWNER may recurse into the allocator to allocate space to
5244 * save the stack with pagesets.lock held. Releasing/reacquiring
5245 * removes much of the performance benefit of bulk allocation so
5246 * force the caller to allocate one page at a time as it'll have
5247 * similar performance to added complexity to the bulk allocator.
5249 if (static_branch_unlikely(&page_owner_inited))
5253 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5254 gfp &= gfp_allowed_mask;
5256 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5260 /* Find an allowed local zone that meets the low watermark. */
5261 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5264 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5265 !__cpuset_zone_allowed(zone, gfp)) {
5269 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5270 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5274 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5275 if (zone_watermark_fast(zone, 0, mark,
5276 zonelist_zone_idx(ac.preferred_zoneref),
5277 alloc_flags, gfp)) {
5283 * If there are no allowed local zones that meets the watermarks then
5284 * try to allocate a single page and reclaim if necessary.
5286 if (unlikely(!zone))
5289 /* Attempt the batch allocation */
5290 local_lock_irqsave(&pagesets.lock, flags);
5291 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5292 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5294 while (nr_populated < nr_pages) {
5296 /* Skip existing pages */
5297 if (page_array && page_array[nr_populated]) {
5302 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5304 if (unlikely(!page)) {
5305 /* Try and get at least one page */
5312 prep_new_page(page, 0, gfp, 0);
5314 list_add(&page->lru, page_list);
5316 page_array[nr_populated] = page;
5320 local_unlock_irqrestore(&pagesets.lock, flags);
5322 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5323 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5326 return nr_populated;
5329 local_unlock_irqrestore(&pagesets.lock, flags);
5332 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5335 list_add(&page->lru, page_list);
5337 page_array[nr_populated] = page;
5343 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5346 * This is the 'heart' of the zoned buddy allocator.
5348 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5349 nodemask_t *nodemask)
5352 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5353 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5354 struct alloc_context ac = { };
5357 * There are several places where we assume that the order value is sane
5358 * so bail out early if the request is out of bound.
5360 if (unlikely(order >= MAX_ORDER)) {
5361 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5365 gfp &= gfp_allowed_mask;
5367 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5368 * resp. GFP_NOIO which has to be inherited for all allocation requests
5369 * from a particular context which has been marked by
5370 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5371 * movable zones are not used during allocation.
5373 gfp = current_gfp_context(gfp);
5375 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5376 &alloc_gfp, &alloc_flags))
5380 * Forbid the first pass from falling back to types that fragment
5381 * memory until all local zones are considered.
5383 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5385 /* First allocation attempt */
5386 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5391 ac.spread_dirty_pages = false;
5394 * Restore the original nodemask if it was potentially replaced with
5395 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5397 ac.nodemask = nodemask;
5399 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5402 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5403 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5404 __free_pages(page, order);
5408 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5412 EXPORT_SYMBOL(__alloc_pages);
5414 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5415 nodemask_t *nodemask)
5417 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5418 preferred_nid, nodemask);
5420 if (page && order > 1)
5421 prep_transhuge_page(page);
5422 return (struct folio *)page;
5424 EXPORT_SYMBOL(__folio_alloc);
5427 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5428 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5429 * you need to access high mem.
5431 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5435 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5438 return (unsigned long) page_address(page);
5440 EXPORT_SYMBOL(__get_free_pages);
5442 unsigned long get_zeroed_page(gfp_t gfp_mask)
5444 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5446 EXPORT_SYMBOL(get_zeroed_page);
5449 * __free_pages - Free pages allocated with alloc_pages().
5450 * @page: The page pointer returned from alloc_pages().
5451 * @order: The order of the allocation.
5453 * This function can free multi-page allocations that are not compound
5454 * pages. It does not check that the @order passed in matches that of
5455 * the allocation, so it is easy to leak memory. Freeing more memory
5456 * than was allocated will probably emit a warning.
5458 * If the last reference to this page is speculative, it will be released
5459 * by put_page() which only frees the first page of a non-compound
5460 * allocation. To prevent the remaining pages from being leaked, we free
5461 * the subsequent pages here. If you want to use the page's reference
5462 * count to decide when to free the allocation, you should allocate a
5463 * compound page, and use put_page() instead of __free_pages().
5465 * Context: May be called in interrupt context or while holding a normal
5466 * spinlock, but not in NMI context or while holding a raw spinlock.
5468 void __free_pages(struct page *page, unsigned int order)
5470 if (put_page_testzero(page))
5471 free_the_page(page, order);
5472 else if (!PageHead(page))
5474 free_the_page(page + (1 << order), order);
5476 EXPORT_SYMBOL(__free_pages);
5478 void free_pages(unsigned long addr, unsigned int order)
5481 VM_BUG_ON(!virt_addr_valid((void *)addr));
5482 __free_pages(virt_to_page((void *)addr), order);
5486 EXPORT_SYMBOL(free_pages);
5490 * An arbitrary-length arbitrary-offset area of memory which resides
5491 * within a 0 or higher order page. Multiple fragments within that page
5492 * are individually refcounted, in the page's reference counter.
5494 * The page_frag functions below provide a simple allocation framework for
5495 * page fragments. This is used by the network stack and network device
5496 * drivers to provide a backing region of memory for use as either an
5497 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5499 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5502 struct page *page = NULL;
5503 gfp_t gfp = gfp_mask;
5505 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5506 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5508 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5509 PAGE_FRAG_CACHE_MAX_ORDER);
5510 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5512 if (unlikely(!page))
5513 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5515 nc->va = page ? page_address(page) : NULL;
5520 void __page_frag_cache_drain(struct page *page, unsigned int count)
5522 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5524 if (page_ref_sub_and_test(page, count))
5525 free_the_page(page, compound_order(page));
5527 EXPORT_SYMBOL(__page_frag_cache_drain);
5529 void *page_frag_alloc_align(struct page_frag_cache *nc,
5530 unsigned int fragsz, gfp_t gfp_mask,
5531 unsigned int align_mask)
5533 unsigned int size = PAGE_SIZE;
5537 if (unlikely(!nc->va)) {
5539 page = __page_frag_cache_refill(nc, gfp_mask);
5543 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5544 /* if size can vary use size else just use PAGE_SIZE */
5547 /* Even if we own the page, we do not use atomic_set().
5548 * This would break get_page_unless_zero() users.
5550 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5552 /* reset page count bias and offset to start of new frag */
5553 nc->pfmemalloc = page_is_pfmemalloc(page);
5554 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5558 offset = nc->offset - fragsz;
5559 if (unlikely(offset < 0)) {
5560 page = virt_to_page(nc->va);
5562 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5565 if (unlikely(nc->pfmemalloc)) {
5566 free_the_page(page, compound_order(page));
5570 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5571 /* if size can vary use size else just use PAGE_SIZE */
5574 /* OK, page count is 0, we can safely set it */
5575 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5577 /* reset page count bias and offset to start of new frag */
5578 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5579 offset = size - fragsz;
5583 offset &= align_mask;
5584 nc->offset = offset;
5586 return nc->va + offset;
5588 EXPORT_SYMBOL(page_frag_alloc_align);
5591 * Frees a page fragment allocated out of either a compound or order 0 page.
5593 void page_frag_free(void *addr)
5595 struct page *page = virt_to_head_page(addr);
5597 if (unlikely(put_page_testzero(page)))
5598 free_the_page(page, compound_order(page));
5600 EXPORT_SYMBOL(page_frag_free);
5602 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5606 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5607 unsigned long used = addr + PAGE_ALIGN(size);
5609 split_page(virt_to_page((void *)addr), order);
5610 while (used < alloc_end) {
5615 return (void *)addr;
5619 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5620 * @size: the number of bytes to allocate
5621 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5623 * This function is similar to alloc_pages(), except that it allocates the
5624 * minimum number of pages to satisfy the request. alloc_pages() can only
5625 * allocate memory in power-of-two pages.
5627 * This function is also limited by MAX_ORDER.
5629 * Memory allocated by this function must be released by free_pages_exact().
5631 * Return: pointer to the allocated area or %NULL in case of error.
5633 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5635 unsigned int order = get_order(size);
5638 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5639 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5641 addr = __get_free_pages(gfp_mask, order);
5642 return make_alloc_exact(addr, order, size);
5644 EXPORT_SYMBOL(alloc_pages_exact);
5647 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5649 * @nid: the preferred node ID where memory should be allocated
5650 * @size: the number of bytes to allocate
5651 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5653 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5656 * Return: pointer to the allocated area or %NULL in case of error.
5658 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5660 unsigned int order = get_order(size);
5663 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5664 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5666 p = alloc_pages_node(nid, gfp_mask, order);
5669 return make_alloc_exact((unsigned long)page_address(p), order, size);
5673 * free_pages_exact - release memory allocated via alloc_pages_exact()
5674 * @virt: the value returned by alloc_pages_exact.
5675 * @size: size of allocation, same value as passed to alloc_pages_exact().
5677 * Release the memory allocated by a previous call to alloc_pages_exact.
5679 void free_pages_exact(void *virt, size_t size)
5681 unsigned long addr = (unsigned long)virt;
5682 unsigned long end = addr + PAGE_ALIGN(size);
5684 while (addr < end) {
5689 EXPORT_SYMBOL(free_pages_exact);
5692 * nr_free_zone_pages - count number of pages beyond high watermark
5693 * @offset: The zone index of the highest zone
5695 * nr_free_zone_pages() counts the number of pages which are beyond the
5696 * high watermark within all zones at or below a given zone index. For each
5697 * zone, the number of pages is calculated as:
5699 * nr_free_zone_pages = managed_pages - high_pages
5701 * Return: number of pages beyond high watermark.
5703 static unsigned long nr_free_zone_pages(int offset)
5708 /* Just pick one node, since fallback list is circular */
5709 unsigned long sum = 0;
5711 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5713 for_each_zone_zonelist(zone, z, zonelist, offset) {
5714 unsigned long size = zone_managed_pages(zone);
5715 unsigned long high = high_wmark_pages(zone);
5724 * nr_free_buffer_pages - count number of pages beyond high watermark
5726 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5727 * watermark within ZONE_DMA and ZONE_NORMAL.
5729 * Return: number of pages beyond high watermark within ZONE_DMA and
5732 unsigned long nr_free_buffer_pages(void)
5734 return nr_free_zone_pages(gfp_zone(GFP_USER));
5736 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5738 static inline void show_node(struct zone *zone)
5740 if (IS_ENABLED(CONFIG_NUMA))
5741 printk("Node %d ", zone_to_nid(zone));
5744 long si_mem_available(void)
5747 unsigned long pagecache;
5748 unsigned long wmark_low = 0;
5749 unsigned long pages[NR_LRU_LISTS];
5750 unsigned long reclaimable;
5754 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5755 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5758 wmark_low += low_wmark_pages(zone);
5761 * Estimate the amount of memory available for userspace allocations,
5762 * without causing swapping.
5764 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5767 * Not all the page cache can be freed, otherwise the system will
5768 * start swapping. Assume at least half of the page cache, or the
5769 * low watermark worth of cache, needs to stay.
5771 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5772 pagecache -= min(pagecache / 2, wmark_low);
5773 available += pagecache;
5776 * Part of the reclaimable slab and other kernel memory consists of
5777 * items that are in use, and cannot be freed. Cap this estimate at the
5780 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5781 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5782 available += reclaimable - min(reclaimable / 2, wmark_low);
5788 EXPORT_SYMBOL_GPL(si_mem_available);
5790 void si_meminfo(struct sysinfo *val)
5792 val->totalram = totalram_pages();
5793 val->sharedram = global_node_page_state(NR_SHMEM);
5794 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5795 val->bufferram = nr_blockdev_pages();
5796 val->totalhigh = totalhigh_pages();
5797 val->freehigh = nr_free_highpages();
5798 val->mem_unit = PAGE_SIZE;
5801 EXPORT_SYMBOL(si_meminfo);
5804 void si_meminfo_node(struct sysinfo *val, int nid)
5806 int zone_type; /* needs to be signed */
5807 unsigned long managed_pages = 0;
5808 unsigned long managed_highpages = 0;
5809 unsigned long free_highpages = 0;
5810 pg_data_t *pgdat = NODE_DATA(nid);
5812 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5813 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5814 val->totalram = managed_pages;
5815 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5816 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5817 #ifdef CONFIG_HIGHMEM
5818 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5819 struct zone *zone = &pgdat->node_zones[zone_type];
5821 if (is_highmem(zone)) {
5822 managed_highpages += zone_managed_pages(zone);
5823 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5826 val->totalhigh = managed_highpages;
5827 val->freehigh = free_highpages;
5829 val->totalhigh = managed_highpages;
5830 val->freehigh = free_highpages;
5832 val->mem_unit = PAGE_SIZE;
5837 * Determine whether the node should be displayed or not, depending on whether
5838 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5840 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5842 if (!(flags & SHOW_MEM_FILTER_NODES))
5846 * no node mask - aka implicit memory numa policy. Do not bother with
5847 * the synchronization - read_mems_allowed_begin - because we do not
5848 * have to be precise here.
5851 nodemask = &cpuset_current_mems_allowed;
5853 return !node_isset(nid, *nodemask);
5856 #define K(x) ((x) << (PAGE_SHIFT-10))
5858 static void show_migration_types(unsigned char type)
5860 static const char types[MIGRATE_TYPES] = {
5861 [MIGRATE_UNMOVABLE] = 'U',
5862 [MIGRATE_MOVABLE] = 'M',
5863 [MIGRATE_RECLAIMABLE] = 'E',
5864 [MIGRATE_HIGHATOMIC] = 'H',
5866 [MIGRATE_CMA] = 'C',
5868 #ifdef CONFIG_MEMORY_ISOLATION
5869 [MIGRATE_ISOLATE] = 'I',
5872 char tmp[MIGRATE_TYPES + 1];
5876 for (i = 0; i < MIGRATE_TYPES; i++) {
5877 if (type & (1 << i))
5882 printk(KERN_CONT "(%s) ", tmp);
5886 * Show free area list (used inside shift_scroll-lock stuff)
5887 * We also calculate the percentage fragmentation. We do this by counting the
5888 * memory on each free list with the exception of the first item on the list.
5891 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5894 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5896 unsigned long free_pcp = 0;
5901 for_each_populated_zone(zone) {
5902 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5905 for_each_online_cpu(cpu)
5906 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5909 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5910 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5911 " unevictable:%lu dirty:%lu writeback:%lu\n"
5912 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5913 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5914 " kernel_misc_reclaimable:%lu\n"
5915 " free:%lu free_pcp:%lu free_cma:%lu\n",
5916 global_node_page_state(NR_ACTIVE_ANON),
5917 global_node_page_state(NR_INACTIVE_ANON),
5918 global_node_page_state(NR_ISOLATED_ANON),
5919 global_node_page_state(NR_ACTIVE_FILE),
5920 global_node_page_state(NR_INACTIVE_FILE),
5921 global_node_page_state(NR_ISOLATED_FILE),
5922 global_node_page_state(NR_UNEVICTABLE),
5923 global_node_page_state(NR_FILE_DIRTY),
5924 global_node_page_state(NR_WRITEBACK),
5925 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5926 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5927 global_node_page_state(NR_FILE_MAPPED),
5928 global_node_page_state(NR_SHMEM),
5929 global_node_page_state(NR_PAGETABLE),
5930 global_zone_page_state(NR_BOUNCE),
5931 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5932 global_zone_page_state(NR_FREE_PAGES),
5934 global_zone_page_state(NR_FREE_CMA_PAGES));
5936 for_each_online_pgdat(pgdat) {
5937 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5941 " active_anon:%lukB"
5942 " inactive_anon:%lukB"
5943 " active_file:%lukB"
5944 " inactive_file:%lukB"
5945 " unevictable:%lukB"
5946 " isolated(anon):%lukB"
5947 " isolated(file):%lukB"
5952 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5954 " shmem_pmdmapped: %lukB"
5957 " writeback_tmp:%lukB"
5958 " kernel_stack:%lukB"
5959 #ifdef CONFIG_SHADOW_CALL_STACK
5960 " shadow_call_stack:%lukB"
5963 " all_unreclaimable? %s"
5966 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5967 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5968 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5969 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5970 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5971 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5972 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5973 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5974 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5975 K(node_page_state(pgdat, NR_WRITEBACK)),
5976 K(node_page_state(pgdat, NR_SHMEM)),
5977 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5978 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5979 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5980 K(node_page_state(pgdat, NR_ANON_THPS)),
5982 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5983 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5984 #ifdef CONFIG_SHADOW_CALL_STACK
5985 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5987 K(node_page_state(pgdat, NR_PAGETABLE)),
5988 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5992 for_each_populated_zone(zone) {
5995 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5999 for_each_online_cpu(cpu)
6000 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6010 " reserved_highatomic:%luKB"
6011 " active_anon:%lukB"
6012 " inactive_anon:%lukB"
6013 " active_file:%lukB"
6014 " inactive_file:%lukB"
6015 " unevictable:%lukB"
6016 " writepending:%lukB"
6026 K(zone_page_state(zone, NR_FREE_PAGES)),
6027 K(zone->watermark_boost),
6028 K(min_wmark_pages(zone)),
6029 K(low_wmark_pages(zone)),
6030 K(high_wmark_pages(zone)),
6031 K(zone->nr_reserved_highatomic),
6032 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6033 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6034 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6035 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6036 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6037 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6038 K(zone->present_pages),
6039 K(zone_managed_pages(zone)),
6040 K(zone_page_state(zone, NR_MLOCK)),
6041 K(zone_page_state(zone, NR_BOUNCE)),
6043 K(this_cpu_read(zone->per_cpu_pageset->count)),
6044 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6045 printk("lowmem_reserve[]:");
6046 for (i = 0; i < MAX_NR_ZONES; i++)
6047 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6048 printk(KERN_CONT "\n");
6051 for_each_populated_zone(zone) {
6053 unsigned long nr[MAX_ORDER], flags, total = 0;
6054 unsigned char types[MAX_ORDER];
6056 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6059 printk(KERN_CONT "%s: ", zone->name);
6061 spin_lock_irqsave(&zone->lock, flags);
6062 for (order = 0; order < MAX_ORDER; order++) {
6063 struct free_area *area = &zone->free_area[order];
6066 nr[order] = area->nr_free;
6067 total += nr[order] << order;
6070 for (type = 0; type < MIGRATE_TYPES; type++) {
6071 if (!free_area_empty(area, type))
6072 types[order] |= 1 << type;
6075 spin_unlock_irqrestore(&zone->lock, flags);
6076 for (order = 0; order < MAX_ORDER; order++) {
6077 printk(KERN_CONT "%lu*%lukB ",
6078 nr[order], K(1UL) << order);
6080 show_migration_types(types[order]);
6082 printk(KERN_CONT "= %lukB\n", K(total));
6085 hugetlb_show_meminfo();
6087 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6089 show_swap_cache_info();
6092 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6094 zoneref->zone = zone;
6095 zoneref->zone_idx = zone_idx(zone);
6099 * Builds allocation fallback zone lists.
6101 * Add all populated zones of a node to the zonelist.
6103 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6106 enum zone_type zone_type = MAX_NR_ZONES;
6111 zone = pgdat->node_zones + zone_type;
6112 if (managed_zone(zone)) {
6113 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6114 check_highest_zone(zone_type);
6116 } while (zone_type);
6123 static int __parse_numa_zonelist_order(char *s)
6126 * We used to support different zonelists modes but they turned
6127 * out to be just not useful. Let's keep the warning in place
6128 * if somebody still use the cmd line parameter so that we do
6129 * not fail it silently
6131 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6132 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6138 char numa_zonelist_order[] = "Node";
6141 * sysctl handler for numa_zonelist_order
6143 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6144 void *buffer, size_t *length, loff_t *ppos)
6147 return __parse_numa_zonelist_order(buffer);
6148 return proc_dostring(table, write, buffer, length, ppos);
6152 #define MAX_NODE_LOAD (nr_online_nodes)
6153 static int node_load[MAX_NUMNODES];
6156 * find_next_best_node - find the next node that should appear in a given node's fallback list
6157 * @node: node whose fallback list we're appending
6158 * @used_node_mask: nodemask_t of already used nodes
6160 * We use a number of factors to determine which is the next node that should
6161 * appear on a given node's fallback list. The node should not have appeared
6162 * already in @node's fallback list, and it should be the next closest node
6163 * according to the distance array (which contains arbitrary distance values
6164 * from each node to each node in the system), and should also prefer nodes
6165 * with no CPUs, since presumably they'll have very little allocation pressure
6166 * on them otherwise.
6168 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6170 int find_next_best_node(int node, nodemask_t *used_node_mask)
6173 int min_val = INT_MAX;
6174 int best_node = NUMA_NO_NODE;
6176 /* Use the local node if we haven't already */
6177 if (!node_isset(node, *used_node_mask)) {
6178 node_set(node, *used_node_mask);
6182 for_each_node_state(n, N_MEMORY) {
6184 /* Don't want a node to appear more than once */
6185 if (node_isset(n, *used_node_mask))
6188 /* Use the distance array to find the distance */
6189 val = node_distance(node, n);
6191 /* Penalize nodes under us ("prefer the next node") */
6194 /* Give preference to headless and unused nodes */
6195 if (!cpumask_empty(cpumask_of_node(n)))
6196 val += PENALTY_FOR_NODE_WITH_CPUS;
6198 /* Slight preference for less loaded node */
6199 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6200 val += node_load[n];
6202 if (val < min_val) {
6209 node_set(best_node, *used_node_mask);
6216 * Build zonelists ordered by node and zones within node.
6217 * This results in maximum locality--normal zone overflows into local
6218 * DMA zone, if any--but risks exhausting DMA zone.
6220 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6223 struct zoneref *zonerefs;
6226 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6228 for (i = 0; i < nr_nodes; i++) {
6231 pg_data_t *node = NODE_DATA(node_order[i]);
6233 nr_zones = build_zonerefs_node(node, zonerefs);
6234 zonerefs += nr_zones;
6236 zonerefs->zone = NULL;
6237 zonerefs->zone_idx = 0;
6241 * Build gfp_thisnode zonelists
6243 static void build_thisnode_zonelists(pg_data_t *pgdat)
6245 struct zoneref *zonerefs;
6248 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6249 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6250 zonerefs += nr_zones;
6251 zonerefs->zone = NULL;
6252 zonerefs->zone_idx = 0;
6256 * Build zonelists ordered by zone and nodes within zones.
6257 * This results in conserving DMA zone[s] until all Normal memory is
6258 * exhausted, but results in overflowing to remote node while memory
6259 * may still exist in local DMA zone.
6262 static void build_zonelists(pg_data_t *pgdat)
6264 static int node_order[MAX_NUMNODES];
6265 int node, load, nr_nodes = 0;
6266 nodemask_t used_mask = NODE_MASK_NONE;
6267 int local_node, prev_node;
6269 /* NUMA-aware ordering of nodes */
6270 local_node = pgdat->node_id;
6271 load = nr_online_nodes;
6272 prev_node = local_node;
6274 memset(node_order, 0, sizeof(node_order));
6275 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6277 * We don't want to pressure a particular node.
6278 * So adding penalty to the first node in same
6279 * distance group to make it round-robin.
6281 if (node_distance(local_node, node) !=
6282 node_distance(local_node, prev_node))
6283 node_load[node] += load;
6285 node_order[nr_nodes++] = node;
6290 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6291 build_thisnode_zonelists(pgdat);
6292 pr_info("Fallback order for Node %d: ", local_node);
6293 for (node = 0; node < nr_nodes; node++)
6294 pr_cont("%d ", node_order[node]);
6298 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6300 * Return node id of node used for "local" allocations.
6301 * I.e., first node id of first zone in arg node's generic zonelist.
6302 * Used for initializing percpu 'numa_mem', which is used primarily
6303 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6305 int local_memory_node(int node)
6309 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6310 gfp_zone(GFP_KERNEL),
6312 return zone_to_nid(z->zone);
6316 static void setup_min_unmapped_ratio(void);
6317 static void setup_min_slab_ratio(void);
6318 #else /* CONFIG_NUMA */
6320 static void build_zonelists(pg_data_t *pgdat)
6322 int node, local_node;
6323 struct zoneref *zonerefs;
6326 local_node = pgdat->node_id;
6328 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6329 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6330 zonerefs += nr_zones;
6333 * Now we build the zonelist so that it contains the zones
6334 * of all the other nodes.
6335 * We don't want to pressure a particular node, so when
6336 * building the zones for node N, we make sure that the
6337 * zones coming right after the local ones are those from
6338 * node N+1 (modulo N)
6340 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6341 if (!node_online(node))
6343 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6344 zonerefs += nr_zones;
6346 for (node = 0; node < local_node; node++) {
6347 if (!node_online(node))
6349 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6350 zonerefs += nr_zones;
6353 zonerefs->zone = NULL;
6354 zonerefs->zone_idx = 0;
6357 #endif /* CONFIG_NUMA */
6360 * Boot pageset table. One per cpu which is going to be used for all
6361 * zones and all nodes. The parameters will be set in such a way
6362 * that an item put on a list will immediately be handed over to
6363 * the buddy list. This is safe since pageset manipulation is done
6364 * with interrupts disabled.
6366 * The boot_pagesets must be kept even after bootup is complete for
6367 * unused processors and/or zones. They do play a role for bootstrapping
6368 * hotplugged processors.
6370 * zoneinfo_show() and maybe other functions do
6371 * not check if the processor is online before following the pageset pointer.
6372 * Other parts of the kernel may not check if the zone is available.
6374 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6375 /* These effectively disable the pcplists in the boot pageset completely */
6376 #define BOOT_PAGESET_HIGH 0
6377 #define BOOT_PAGESET_BATCH 1
6378 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6379 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6380 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6382 static void __build_all_zonelists(void *data)
6385 int __maybe_unused cpu;
6386 pg_data_t *self = data;
6387 static DEFINE_SPINLOCK(lock);
6392 memset(node_load, 0, sizeof(node_load));
6396 * This node is hotadded and no memory is yet present. So just
6397 * building zonelists is fine - no need to touch other nodes.
6399 if (self && !node_online(self->node_id)) {
6400 build_zonelists(self);
6402 for_each_online_node(nid) {
6403 pg_data_t *pgdat = NODE_DATA(nid);
6405 build_zonelists(pgdat);
6408 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6410 * We now know the "local memory node" for each node--
6411 * i.e., the node of the first zone in the generic zonelist.
6412 * Set up numa_mem percpu variable for on-line cpus. During
6413 * boot, only the boot cpu should be on-line; we'll init the
6414 * secondary cpus' numa_mem as they come on-line. During
6415 * node/memory hotplug, we'll fixup all on-line cpus.
6417 for_each_online_cpu(cpu)
6418 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6425 static noinline void __init
6426 build_all_zonelists_init(void)
6430 __build_all_zonelists(NULL);
6433 * Initialize the boot_pagesets that are going to be used
6434 * for bootstrapping processors. The real pagesets for
6435 * each zone will be allocated later when the per cpu
6436 * allocator is available.
6438 * boot_pagesets are used also for bootstrapping offline
6439 * cpus if the system is already booted because the pagesets
6440 * are needed to initialize allocators on a specific cpu too.
6441 * F.e. the percpu allocator needs the page allocator which
6442 * needs the percpu allocator in order to allocate its pagesets
6443 * (a chicken-egg dilemma).
6445 for_each_possible_cpu(cpu)
6446 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6448 mminit_verify_zonelist();
6449 cpuset_init_current_mems_allowed();
6453 * unless system_state == SYSTEM_BOOTING.
6455 * __ref due to call of __init annotated helper build_all_zonelists_init
6456 * [protected by SYSTEM_BOOTING].
6458 void __ref build_all_zonelists(pg_data_t *pgdat)
6460 unsigned long vm_total_pages;
6462 if (system_state == SYSTEM_BOOTING) {
6463 build_all_zonelists_init();
6465 __build_all_zonelists(pgdat);
6466 /* cpuset refresh routine should be here */
6468 /* Get the number of free pages beyond high watermark in all zones. */
6469 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6471 * Disable grouping by mobility if the number of pages in the
6472 * system is too low to allow the mechanism to work. It would be
6473 * more accurate, but expensive to check per-zone. This check is
6474 * made on memory-hotadd so a system can start with mobility
6475 * disabled and enable it later
6477 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6478 page_group_by_mobility_disabled = 1;
6480 page_group_by_mobility_disabled = 0;
6482 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6484 page_group_by_mobility_disabled ? "off" : "on",
6487 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6491 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6492 static bool __meminit
6493 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6495 static struct memblock_region *r;
6497 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6498 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6499 for_each_mem_region(r) {
6500 if (*pfn < memblock_region_memory_end_pfn(r))
6504 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6505 memblock_is_mirror(r)) {
6506 *pfn = memblock_region_memory_end_pfn(r);
6514 * Initially all pages are reserved - free ones are freed
6515 * up by memblock_free_all() once the early boot process is
6516 * done. Non-atomic initialization, single-pass.
6518 * All aligned pageblocks are initialized to the specified migratetype
6519 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6520 * zone stats (e.g., nr_isolate_pageblock) are touched.
6522 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6523 unsigned long start_pfn, unsigned long zone_end_pfn,
6524 enum meminit_context context,
6525 struct vmem_altmap *altmap, int migratetype)
6527 unsigned long pfn, end_pfn = start_pfn + size;
6530 if (highest_memmap_pfn < end_pfn - 1)
6531 highest_memmap_pfn = end_pfn - 1;
6533 #ifdef CONFIG_ZONE_DEVICE
6535 * Honor reservation requested by the driver for this ZONE_DEVICE
6536 * memory. We limit the total number of pages to initialize to just
6537 * those that might contain the memory mapping. We will defer the
6538 * ZONE_DEVICE page initialization until after we have released
6541 if (zone == ZONE_DEVICE) {
6545 if (start_pfn == altmap->base_pfn)
6546 start_pfn += altmap->reserve;
6547 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6551 for (pfn = start_pfn; pfn < end_pfn; ) {
6553 * There can be holes in boot-time mem_map[]s handed to this
6554 * function. They do not exist on hotplugged memory.
6556 if (context == MEMINIT_EARLY) {
6557 if (overlap_memmap_init(zone, &pfn))
6559 if (defer_init(nid, pfn, zone_end_pfn))
6563 page = pfn_to_page(pfn);
6564 __init_single_page(page, pfn, zone, nid);
6565 if (context == MEMINIT_HOTPLUG)
6566 __SetPageReserved(page);
6569 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6570 * such that unmovable allocations won't be scattered all
6571 * over the place during system boot.
6573 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6574 set_pageblock_migratetype(page, migratetype);
6581 #ifdef CONFIG_ZONE_DEVICE
6582 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6583 unsigned long zone_idx, int nid,
6584 struct dev_pagemap *pgmap)
6587 __init_single_page(page, pfn, zone_idx, nid);
6590 * Mark page reserved as it will need to wait for onlining
6591 * phase for it to be fully associated with a zone.
6593 * We can use the non-atomic __set_bit operation for setting
6594 * the flag as we are still initializing the pages.
6596 __SetPageReserved(page);
6599 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6600 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6601 * ever freed or placed on a driver-private list.
6603 page->pgmap = pgmap;
6604 page->zone_device_data = NULL;
6607 * Mark the block movable so that blocks are reserved for
6608 * movable at startup. This will force kernel allocations
6609 * to reserve their blocks rather than leaking throughout
6610 * the address space during boot when many long-lived
6611 * kernel allocations are made.
6613 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6614 * because this is done early in section_activate()
6616 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6617 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6622 static void __ref memmap_init_compound(struct page *head,
6623 unsigned long head_pfn,
6624 unsigned long zone_idx, int nid,
6625 struct dev_pagemap *pgmap,
6626 unsigned long nr_pages)
6628 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6629 unsigned int order = pgmap->vmemmap_shift;
6631 __SetPageHead(head);
6632 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6633 struct page *page = pfn_to_page(pfn);
6635 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6636 prep_compound_tail(head, pfn - head_pfn);
6637 set_page_count(page, 0);
6640 * The first tail page stores compound_mapcount_ptr() and
6641 * compound_order() and the second tail page stores
6642 * compound_pincount_ptr(). Call prep_compound_head() after
6643 * the first and second tail pages have been initialized to
6644 * not have the data overwritten.
6646 if (pfn == head_pfn + 2)
6647 prep_compound_head(head, order);
6651 void __ref memmap_init_zone_device(struct zone *zone,
6652 unsigned long start_pfn,
6653 unsigned long nr_pages,
6654 struct dev_pagemap *pgmap)
6656 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6657 struct pglist_data *pgdat = zone->zone_pgdat;
6658 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6659 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6660 unsigned long zone_idx = zone_idx(zone);
6661 unsigned long start = jiffies;
6662 int nid = pgdat->node_id;
6664 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6668 * The call to memmap_init should have already taken care
6669 * of the pages reserved for the memmap, so we can just jump to
6670 * the end of that region and start processing the device pages.
6673 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6674 nr_pages = end_pfn - start_pfn;
6677 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6678 struct page *page = pfn_to_page(pfn);
6680 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6682 if (pfns_per_compound == 1)
6685 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6689 pr_info("%s initialised %lu pages in %ums\n", __func__,
6690 nr_pages, jiffies_to_msecs(jiffies - start));
6694 static void __meminit zone_init_free_lists(struct zone *zone)
6696 unsigned int order, t;
6697 for_each_migratetype_order(order, t) {
6698 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6699 zone->free_area[order].nr_free = 0;
6704 * Only struct pages that correspond to ranges defined by memblock.memory
6705 * are zeroed and initialized by going through __init_single_page() during
6706 * memmap_init_zone_range().
6708 * But, there could be struct pages that correspond to holes in
6709 * memblock.memory. This can happen because of the following reasons:
6710 * - physical memory bank size is not necessarily the exact multiple of the
6711 * arbitrary section size
6712 * - early reserved memory may not be listed in memblock.memory
6713 * - memory layouts defined with memmap= kernel parameter may not align
6714 * nicely with memmap sections
6716 * Explicitly initialize those struct pages so that:
6717 * - PG_Reserved is set
6718 * - zone and node links point to zone and node that span the page if the
6719 * hole is in the middle of a zone
6720 * - zone and node links point to adjacent zone/node if the hole falls on
6721 * the zone boundary; the pages in such holes will be prepended to the
6722 * zone/node above the hole except for the trailing pages in the last
6723 * section that will be appended to the zone/node below.
6725 static void __init init_unavailable_range(unsigned long spfn,
6732 for (pfn = spfn; pfn < epfn; pfn++) {
6733 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6734 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6735 + pageblock_nr_pages - 1;
6738 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6739 __SetPageReserved(pfn_to_page(pfn));
6744 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6745 node, zone_names[zone], pgcnt);
6748 static void __init memmap_init_zone_range(struct zone *zone,
6749 unsigned long start_pfn,
6750 unsigned long end_pfn,
6751 unsigned long *hole_pfn)
6753 unsigned long zone_start_pfn = zone->zone_start_pfn;
6754 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6755 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6757 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6758 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6760 if (start_pfn >= end_pfn)
6763 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6764 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6766 if (*hole_pfn < start_pfn)
6767 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6769 *hole_pfn = end_pfn;
6772 static void __init memmap_init(void)
6774 unsigned long start_pfn, end_pfn;
6775 unsigned long hole_pfn = 0;
6776 int i, j, zone_id = 0, nid;
6778 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6779 struct pglist_data *node = NODE_DATA(nid);
6781 for (j = 0; j < MAX_NR_ZONES; j++) {
6782 struct zone *zone = node->node_zones + j;
6784 if (!populated_zone(zone))
6787 memmap_init_zone_range(zone, start_pfn, end_pfn,
6793 #ifdef CONFIG_SPARSEMEM
6795 * Initialize the memory map for hole in the range [memory_end,
6797 * Append the pages in this hole to the highest zone in the last
6799 * The call to init_unavailable_range() is outside the ifdef to
6800 * silence the compiler warining about zone_id set but not used;
6801 * for FLATMEM it is a nop anyway
6803 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6804 if (hole_pfn < end_pfn)
6806 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6809 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6810 phys_addr_t min_addr, int nid, bool exact_nid)
6815 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6816 MEMBLOCK_ALLOC_ACCESSIBLE,
6819 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6820 MEMBLOCK_ALLOC_ACCESSIBLE,
6823 if (ptr && size > 0)
6824 page_init_poison(ptr, size);
6829 static int zone_batchsize(struct zone *zone)
6835 * The number of pages to batch allocate is either ~0.1%
6836 * of the zone or 1MB, whichever is smaller. The batch
6837 * size is striking a balance between allocation latency
6838 * and zone lock contention.
6840 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6841 batch /= 4; /* We effectively *= 4 below */
6846 * Clamp the batch to a 2^n - 1 value. Having a power
6847 * of 2 value was found to be more likely to have
6848 * suboptimal cache aliasing properties in some cases.
6850 * For example if 2 tasks are alternately allocating
6851 * batches of pages, one task can end up with a lot
6852 * of pages of one half of the possible page colors
6853 * and the other with pages of the other colors.
6855 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6860 /* The deferral and batching of frees should be suppressed under NOMMU
6863 * The problem is that NOMMU needs to be able to allocate large chunks
6864 * of contiguous memory as there's no hardware page translation to
6865 * assemble apparent contiguous memory from discontiguous pages.
6867 * Queueing large contiguous runs of pages for batching, however,
6868 * causes the pages to actually be freed in smaller chunks. As there
6869 * can be a significant delay between the individual batches being
6870 * recycled, this leads to the once large chunks of space being
6871 * fragmented and becoming unavailable for high-order allocations.
6877 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6882 unsigned long total_pages;
6884 if (!percpu_pagelist_high_fraction) {
6886 * By default, the high value of the pcp is based on the zone
6887 * low watermark so that if they are full then background
6888 * reclaim will not be started prematurely.
6890 total_pages = low_wmark_pages(zone);
6893 * If percpu_pagelist_high_fraction is configured, the high
6894 * value is based on a fraction of the managed pages in the
6897 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6901 * Split the high value across all online CPUs local to the zone. Note
6902 * that early in boot that CPUs may not be online yet and that during
6903 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6904 * onlined. For memory nodes that have no CPUs, split pcp->high across
6905 * all online CPUs to mitigate the risk that reclaim is triggered
6906 * prematurely due to pages stored on pcp lists.
6908 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6910 nr_split_cpus = num_online_cpus();
6911 high = total_pages / nr_split_cpus;
6914 * Ensure high is at least batch*4. The multiple is based on the
6915 * historical relationship between high and batch.
6917 high = max(high, batch << 2);
6926 * pcp->high and pcp->batch values are related and generally batch is lower
6927 * than high. They are also related to pcp->count such that count is lower
6928 * than high, and as soon as it reaches high, the pcplist is flushed.
6930 * However, guaranteeing these relations at all times would require e.g. write
6931 * barriers here but also careful usage of read barriers at the read side, and
6932 * thus be prone to error and bad for performance. Thus the update only prevents
6933 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6934 * can cope with those fields changing asynchronously, and fully trust only the
6935 * pcp->count field on the local CPU with interrupts disabled.
6937 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6938 * outside of boot time (or some other assurance that no concurrent updaters
6941 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6942 unsigned long batch)
6944 WRITE_ONCE(pcp->batch, batch);
6945 WRITE_ONCE(pcp->high, high);
6948 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6952 memset(pcp, 0, sizeof(*pcp));
6953 memset(pzstats, 0, sizeof(*pzstats));
6955 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6956 INIT_LIST_HEAD(&pcp->lists[pindex]);
6959 * Set batch and high values safe for a boot pageset. A true percpu
6960 * pageset's initialization will update them subsequently. Here we don't
6961 * need to be as careful as pageset_update() as nobody can access the
6964 pcp->high = BOOT_PAGESET_HIGH;
6965 pcp->batch = BOOT_PAGESET_BATCH;
6966 pcp->free_factor = 0;
6969 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6970 unsigned long batch)
6972 struct per_cpu_pages *pcp;
6975 for_each_possible_cpu(cpu) {
6976 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6977 pageset_update(pcp, high, batch);
6982 * Calculate and set new high and batch values for all per-cpu pagesets of a
6983 * zone based on the zone's size.
6985 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6987 int new_high, new_batch;
6989 new_batch = max(1, zone_batchsize(zone));
6990 new_high = zone_highsize(zone, new_batch, cpu_online);
6992 if (zone->pageset_high == new_high &&
6993 zone->pageset_batch == new_batch)
6996 zone->pageset_high = new_high;
6997 zone->pageset_batch = new_batch;
6999 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7002 void __meminit setup_zone_pageset(struct zone *zone)
7006 /* Size may be 0 on !SMP && !NUMA */
7007 if (sizeof(struct per_cpu_zonestat) > 0)
7008 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7010 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7011 for_each_possible_cpu(cpu) {
7012 struct per_cpu_pages *pcp;
7013 struct per_cpu_zonestat *pzstats;
7015 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7016 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7017 per_cpu_pages_init(pcp, pzstats);
7020 zone_set_pageset_high_and_batch(zone, 0);
7024 * Allocate per cpu pagesets and initialize them.
7025 * Before this call only boot pagesets were available.
7027 void __init setup_per_cpu_pageset(void)
7029 struct pglist_data *pgdat;
7031 int __maybe_unused cpu;
7033 for_each_populated_zone(zone)
7034 setup_zone_pageset(zone);
7038 * Unpopulated zones continue using the boot pagesets.
7039 * The numa stats for these pagesets need to be reset.
7040 * Otherwise, they will end up skewing the stats of
7041 * the nodes these zones are associated with.
7043 for_each_possible_cpu(cpu) {
7044 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7045 memset(pzstats->vm_numa_event, 0,
7046 sizeof(pzstats->vm_numa_event));
7050 for_each_online_pgdat(pgdat)
7051 pgdat->per_cpu_nodestats =
7052 alloc_percpu(struct per_cpu_nodestat);
7055 static __meminit void zone_pcp_init(struct zone *zone)
7058 * per cpu subsystem is not up at this point. The following code
7059 * relies on the ability of the linker to provide the
7060 * offset of a (static) per cpu variable into the per cpu area.
7062 zone->per_cpu_pageset = &boot_pageset;
7063 zone->per_cpu_zonestats = &boot_zonestats;
7064 zone->pageset_high = BOOT_PAGESET_HIGH;
7065 zone->pageset_batch = BOOT_PAGESET_BATCH;
7067 if (populated_zone(zone))
7068 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7069 zone->present_pages, zone_batchsize(zone));
7072 void __meminit init_currently_empty_zone(struct zone *zone,
7073 unsigned long zone_start_pfn,
7076 struct pglist_data *pgdat = zone->zone_pgdat;
7077 int zone_idx = zone_idx(zone) + 1;
7079 if (zone_idx > pgdat->nr_zones)
7080 pgdat->nr_zones = zone_idx;
7082 zone->zone_start_pfn = zone_start_pfn;
7084 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7085 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7087 (unsigned long)zone_idx(zone),
7088 zone_start_pfn, (zone_start_pfn + size));
7090 zone_init_free_lists(zone);
7091 zone->initialized = 1;
7095 * get_pfn_range_for_nid - Return the start and end page frames for a node
7096 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7097 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7098 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7100 * It returns the start and end page frame of a node based on information
7101 * provided by memblock_set_node(). If called for a node
7102 * with no available memory, a warning is printed and the start and end
7105 void __init get_pfn_range_for_nid(unsigned int nid,
7106 unsigned long *start_pfn, unsigned long *end_pfn)
7108 unsigned long this_start_pfn, this_end_pfn;
7114 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7115 *start_pfn = min(*start_pfn, this_start_pfn);
7116 *end_pfn = max(*end_pfn, this_end_pfn);
7119 if (*start_pfn == -1UL)
7124 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7125 * assumption is made that zones within a node are ordered in monotonic
7126 * increasing memory addresses so that the "highest" populated zone is used
7128 static void __init find_usable_zone_for_movable(void)
7131 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7132 if (zone_index == ZONE_MOVABLE)
7135 if (arch_zone_highest_possible_pfn[zone_index] >
7136 arch_zone_lowest_possible_pfn[zone_index])
7140 VM_BUG_ON(zone_index == -1);
7141 movable_zone = zone_index;
7145 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7146 * because it is sized independent of architecture. Unlike the other zones,
7147 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7148 * in each node depending on the size of each node and how evenly kernelcore
7149 * is distributed. This helper function adjusts the zone ranges
7150 * provided by the architecture for a given node by using the end of the
7151 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7152 * zones within a node are in order of monotonic increases memory addresses
7154 static void __init adjust_zone_range_for_zone_movable(int nid,
7155 unsigned long zone_type,
7156 unsigned long node_start_pfn,
7157 unsigned long node_end_pfn,
7158 unsigned long *zone_start_pfn,
7159 unsigned long *zone_end_pfn)
7161 /* Only adjust if ZONE_MOVABLE is on this node */
7162 if (zone_movable_pfn[nid]) {
7163 /* Size ZONE_MOVABLE */
7164 if (zone_type == ZONE_MOVABLE) {
7165 *zone_start_pfn = zone_movable_pfn[nid];
7166 *zone_end_pfn = min(node_end_pfn,
7167 arch_zone_highest_possible_pfn[movable_zone]);
7169 /* Adjust for ZONE_MOVABLE starting within this range */
7170 } else if (!mirrored_kernelcore &&
7171 *zone_start_pfn < zone_movable_pfn[nid] &&
7172 *zone_end_pfn > zone_movable_pfn[nid]) {
7173 *zone_end_pfn = zone_movable_pfn[nid];
7175 /* Check if this whole range is within ZONE_MOVABLE */
7176 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7177 *zone_start_pfn = *zone_end_pfn;
7182 * Return the number of pages a zone spans in a node, including holes
7183 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7185 static unsigned long __init zone_spanned_pages_in_node(int nid,
7186 unsigned long zone_type,
7187 unsigned long node_start_pfn,
7188 unsigned long node_end_pfn,
7189 unsigned long *zone_start_pfn,
7190 unsigned long *zone_end_pfn)
7192 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7193 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7194 /* When hotadd a new node from cpu_up(), the node should be empty */
7195 if (!node_start_pfn && !node_end_pfn)
7198 /* Get the start and end of the zone */
7199 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7200 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7201 adjust_zone_range_for_zone_movable(nid, zone_type,
7202 node_start_pfn, node_end_pfn,
7203 zone_start_pfn, zone_end_pfn);
7205 /* Check that this node has pages within the zone's required range */
7206 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7209 /* Move the zone boundaries inside the node if necessary */
7210 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7211 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7213 /* Return the spanned pages */
7214 return *zone_end_pfn - *zone_start_pfn;
7218 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7219 * then all holes in the requested range will be accounted for.
7221 unsigned long __init __absent_pages_in_range(int nid,
7222 unsigned long range_start_pfn,
7223 unsigned long range_end_pfn)
7225 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7226 unsigned long start_pfn, end_pfn;
7229 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7230 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7231 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7232 nr_absent -= end_pfn - start_pfn;
7238 * absent_pages_in_range - Return number of page frames in holes within a range
7239 * @start_pfn: The start PFN to start searching for holes
7240 * @end_pfn: The end PFN to stop searching for holes
7242 * Return: the number of pages frames in memory holes within a range.
7244 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7245 unsigned long end_pfn)
7247 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7250 /* Return the number of page frames in holes in a zone on a node */
7251 static unsigned long __init zone_absent_pages_in_node(int nid,
7252 unsigned long zone_type,
7253 unsigned long node_start_pfn,
7254 unsigned long node_end_pfn)
7256 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7257 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7258 unsigned long zone_start_pfn, zone_end_pfn;
7259 unsigned long nr_absent;
7261 /* When hotadd a new node from cpu_up(), the node should be empty */
7262 if (!node_start_pfn && !node_end_pfn)
7265 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7266 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7268 adjust_zone_range_for_zone_movable(nid, zone_type,
7269 node_start_pfn, node_end_pfn,
7270 &zone_start_pfn, &zone_end_pfn);
7271 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7274 * ZONE_MOVABLE handling.
7275 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7278 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7279 unsigned long start_pfn, end_pfn;
7280 struct memblock_region *r;
7282 for_each_mem_region(r) {
7283 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7284 zone_start_pfn, zone_end_pfn);
7285 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7286 zone_start_pfn, zone_end_pfn);
7288 if (zone_type == ZONE_MOVABLE &&
7289 memblock_is_mirror(r))
7290 nr_absent += end_pfn - start_pfn;
7292 if (zone_type == ZONE_NORMAL &&
7293 !memblock_is_mirror(r))
7294 nr_absent += end_pfn - start_pfn;
7301 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7302 unsigned long node_start_pfn,
7303 unsigned long node_end_pfn)
7305 unsigned long realtotalpages = 0, totalpages = 0;
7308 for (i = 0; i < MAX_NR_ZONES; i++) {
7309 struct zone *zone = pgdat->node_zones + i;
7310 unsigned long zone_start_pfn, zone_end_pfn;
7311 unsigned long spanned, absent;
7312 unsigned long size, real_size;
7314 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7319 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7324 real_size = size - absent;
7327 zone->zone_start_pfn = zone_start_pfn;
7329 zone->zone_start_pfn = 0;
7330 zone->spanned_pages = size;
7331 zone->present_pages = real_size;
7332 #if defined(CONFIG_MEMORY_HOTPLUG)
7333 zone->present_early_pages = real_size;
7337 realtotalpages += real_size;
7340 pgdat->node_spanned_pages = totalpages;
7341 pgdat->node_present_pages = realtotalpages;
7342 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7345 #ifndef CONFIG_SPARSEMEM
7347 * Calculate the size of the zone->blockflags rounded to an unsigned long
7348 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7349 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7350 * round what is now in bits to nearest long in bits, then return it in
7353 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7355 unsigned long usemapsize;
7357 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7358 usemapsize = roundup(zonesize, pageblock_nr_pages);
7359 usemapsize = usemapsize >> pageblock_order;
7360 usemapsize *= NR_PAGEBLOCK_BITS;
7361 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7363 return usemapsize / 8;
7366 static void __ref setup_usemap(struct zone *zone)
7368 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7369 zone->spanned_pages);
7370 zone->pageblock_flags = NULL;
7372 zone->pageblock_flags =
7373 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7375 if (!zone->pageblock_flags)
7376 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7377 usemapsize, zone->name, zone_to_nid(zone));
7381 static inline void setup_usemap(struct zone *zone) {}
7382 #endif /* CONFIG_SPARSEMEM */
7384 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7386 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7387 void __init set_pageblock_order(void)
7391 /* Check that pageblock_nr_pages has not already been setup */
7392 if (pageblock_order)
7395 if (HPAGE_SHIFT > PAGE_SHIFT)
7396 order = HUGETLB_PAGE_ORDER;
7398 order = MAX_ORDER - 1;
7401 * Assume the largest contiguous order of interest is a huge page.
7402 * This value may be variable depending on boot parameters on IA64 and
7405 pageblock_order = order;
7407 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7410 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7411 * is unused as pageblock_order is set at compile-time. See
7412 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7415 void __init set_pageblock_order(void)
7419 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7421 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7422 unsigned long present_pages)
7424 unsigned long pages = spanned_pages;
7427 * Provide a more accurate estimation if there are holes within
7428 * the zone and SPARSEMEM is in use. If there are holes within the
7429 * zone, each populated memory region may cost us one or two extra
7430 * memmap pages due to alignment because memmap pages for each
7431 * populated regions may not be naturally aligned on page boundary.
7432 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7434 if (spanned_pages > present_pages + (present_pages >> 4) &&
7435 IS_ENABLED(CONFIG_SPARSEMEM))
7436 pages = present_pages;
7438 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7441 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7442 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7444 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7446 spin_lock_init(&ds_queue->split_queue_lock);
7447 INIT_LIST_HEAD(&ds_queue->split_queue);
7448 ds_queue->split_queue_len = 0;
7451 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7454 #ifdef CONFIG_COMPACTION
7455 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7457 init_waitqueue_head(&pgdat->kcompactd_wait);
7460 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7463 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7467 pgdat_resize_init(pgdat);
7469 pgdat_init_split_queue(pgdat);
7470 pgdat_init_kcompactd(pgdat);
7472 init_waitqueue_head(&pgdat->kswapd_wait);
7473 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7475 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7476 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7478 pgdat_page_ext_init(pgdat);
7479 lruvec_init(&pgdat->__lruvec);
7482 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7483 unsigned long remaining_pages)
7485 atomic_long_set(&zone->managed_pages, remaining_pages);
7486 zone_set_nid(zone, nid);
7487 zone->name = zone_names[idx];
7488 zone->zone_pgdat = NODE_DATA(nid);
7489 spin_lock_init(&zone->lock);
7490 zone_seqlock_init(zone);
7491 zone_pcp_init(zone);
7495 * Set up the zone data structures
7496 * - init pgdat internals
7497 * - init all zones belonging to this node
7499 * NOTE: this function is only called during memory hotplug
7501 #ifdef CONFIG_MEMORY_HOTPLUG
7502 void __ref free_area_init_core_hotplug(int nid)
7505 pg_data_t *pgdat = NODE_DATA(nid);
7507 pgdat_init_internals(pgdat);
7508 for (z = 0; z < MAX_NR_ZONES; z++)
7509 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7514 * Set up the zone data structures:
7515 * - mark all pages reserved
7516 * - mark all memory queues empty
7517 * - clear the memory bitmaps
7519 * NOTE: pgdat should get zeroed by caller.
7520 * NOTE: this function is only called during early init.
7522 static void __init free_area_init_core(struct pglist_data *pgdat)
7525 int nid = pgdat->node_id;
7527 pgdat_init_internals(pgdat);
7528 pgdat->per_cpu_nodestats = &boot_nodestats;
7530 for (j = 0; j < MAX_NR_ZONES; j++) {
7531 struct zone *zone = pgdat->node_zones + j;
7532 unsigned long size, freesize, memmap_pages;
7534 size = zone->spanned_pages;
7535 freesize = zone->present_pages;
7538 * Adjust freesize so that it accounts for how much memory
7539 * is used by this zone for memmap. This affects the watermark
7540 * and per-cpu initialisations
7542 memmap_pages = calc_memmap_size(size, freesize);
7543 if (!is_highmem_idx(j)) {
7544 if (freesize >= memmap_pages) {
7545 freesize -= memmap_pages;
7547 pr_debug(" %s zone: %lu pages used for memmap\n",
7548 zone_names[j], memmap_pages);
7550 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7551 zone_names[j], memmap_pages, freesize);
7554 /* Account for reserved pages */
7555 if (j == 0 && freesize > dma_reserve) {
7556 freesize -= dma_reserve;
7557 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7560 if (!is_highmem_idx(j))
7561 nr_kernel_pages += freesize;
7562 /* Charge for highmem memmap if there are enough kernel pages */
7563 else if (nr_kernel_pages > memmap_pages * 2)
7564 nr_kernel_pages -= memmap_pages;
7565 nr_all_pages += freesize;
7568 * Set an approximate value for lowmem here, it will be adjusted
7569 * when the bootmem allocator frees pages into the buddy system.
7570 * And all highmem pages will be managed by the buddy system.
7572 zone_init_internals(zone, j, nid, freesize);
7577 set_pageblock_order();
7579 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7583 #ifdef CONFIG_FLATMEM
7584 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7586 unsigned long __maybe_unused start = 0;
7587 unsigned long __maybe_unused offset = 0;
7589 /* Skip empty nodes */
7590 if (!pgdat->node_spanned_pages)
7593 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7594 offset = pgdat->node_start_pfn - start;
7595 /* ia64 gets its own node_mem_map, before this, without bootmem */
7596 if (!pgdat->node_mem_map) {
7597 unsigned long size, end;
7601 * The zone's endpoints aren't required to be MAX_ORDER
7602 * aligned but the node_mem_map endpoints must be in order
7603 * for the buddy allocator to function correctly.
7605 end = pgdat_end_pfn(pgdat);
7606 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7607 size = (end - start) * sizeof(struct page);
7608 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7609 pgdat->node_id, false);
7611 panic("Failed to allocate %ld bytes for node %d memory map\n",
7612 size, pgdat->node_id);
7613 pgdat->node_mem_map = map + offset;
7615 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7616 __func__, pgdat->node_id, (unsigned long)pgdat,
7617 (unsigned long)pgdat->node_mem_map);
7620 * With no DISCONTIG, the global mem_map is just set as node 0's
7622 if (pgdat == NODE_DATA(0)) {
7623 mem_map = NODE_DATA(0)->node_mem_map;
7624 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7630 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7631 #endif /* CONFIG_FLATMEM */
7633 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7634 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7636 pgdat->first_deferred_pfn = ULONG_MAX;
7639 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7642 static void __init free_area_init_node(int nid)
7644 pg_data_t *pgdat = NODE_DATA(nid);
7645 unsigned long start_pfn = 0;
7646 unsigned long end_pfn = 0;
7648 /* pg_data_t should be reset to zero when it's allocated */
7649 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7651 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7653 pgdat->node_id = nid;
7654 pgdat->node_start_pfn = start_pfn;
7655 pgdat->per_cpu_nodestats = NULL;
7657 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7658 (u64)start_pfn << PAGE_SHIFT,
7659 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7660 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7662 alloc_node_mem_map(pgdat);
7663 pgdat_set_deferred_range(pgdat);
7665 free_area_init_core(pgdat);
7668 void __init free_area_init_memoryless_node(int nid)
7670 free_area_init_node(nid);
7673 #if MAX_NUMNODES > 1
7675 * Figure out the number of possible node ids.
7677 void __init setup_nr_node_ids(void)
7679 unsigned int highest;
7681 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7682 nr_node_ids = highest + 1;
7687 * node_map_pfn_alignment - determine the maximum internode alignment
7689 * This function should be called after node map is populated and sorted.
7690 * It calculates the maximum power of two alignment which can distinguish
7693 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7694 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7695 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7696 * shifted, 1GiB is enough and this function will indicate so.
7698 * This is used to test whether pfn -> nid mapping of the chosen memory
7699 * model has fine enough granularity to avoid incorrect mapping for the
7700 * populated node map.
7702 * Return: the determined alignment in pfn's. 0 if there is no alignment
7703 * requirement (single node).
7705 unsigned long __init node_map_pfn_alignment(void)
7707 unsigned long accl_mask = 0, last_end = 0;
7708 unsigned long start, end, mask;
7709 int last_nid = NUMA_NO_NODE;
7712 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7713 if (!start || last_nid < 0 || last_nid == nid) {
7720 * Start with a mask granular enough to pin-point to the
7721 * start pfn and tick off bits one-by-one until it becomes
7722 * too coarse to separate the current node from the last.
7724 mask = ~((1 << __ffs(start)) - 1);
7725 while (mask && last_end <= (start & (mask << 1)))
7728 /* accumulate all internode masks */
7732 /* convert mask to number of pages */
7733 return ~accl_mask + 1;
7737 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7739 * Return: the minimum PFN based on information provided via
7740 * memblock_set_node().
7742 unsigned long __init find_min_pfn_with_active_regions(void)
7744 return PHYS_PFN(memblock_start_of_DRAM());
7748 * early_calculate_totalpages()
7749 * Sum pages in active regions for movable zone.
7750 * Populate N_MEMORY for calculating usable_nodes.
7752 static unsigned long __init early_calculate_totalpages(void)
7754 unsigned long totalpages = 0;
7755 unsigned long start_pfn, end_pfn;
7758 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7759 unsigned long pages = end_pfn - start_pfn;
7761 totalpages += pages;
7763 node_set_state(nid, N_MEMORY);
7769 * Find the PFN the Movable zone begins in each node. Kernel memory
7770 * is spread evenly between nodes as long as the nodes have enough
7771 * memory. When they don't, some nodes will have more kernelcore than
7774 static void __init find_zone_movable_pfns_for_nodes(void)
7777 unsigned long usable_startpfn;
7778 unsigned long kernelcore_node, kernelcore_remaining;
7779 /* save the state before borrow the nodemask */
7780 nodemask_t saved_node_state = node_states[N_MEMORY];
7781 unsigned long totalpages = early_calculate_totalpages();
7782 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7783 struct memblock_region *r;
7785 /* Need to find movable_zone earlier when movable_node is specified. */
7786 find_usable_zone_for_movable();
7789 * If movable_node is specified, ignore kernelcore and movablecore
7792 if (movable_node_is_enabled()) {
7793 for_each_mem_region(r) {
7794 if (!memblock_is_hotpluggable(r))
7797 nid = memblock_get_region_node(r);
7799 usable_startpfn = PFN_DOWN(r->base);
7800 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7801 min(usable_startpfn, zone_movable_pfn[nid]) :
7809 * If kernelcore=mirror is specified, ignore movablecore option
7811 if (mirrored_kernelcore) {
7812 bool mem_below_4gb_not_mirrored = false;
7814 for_each_mem_region(r) {
7815 if (memblock_is_mirror(r))
7818 nid = memblock_get_region_node(r);
7820 usable_startpfn = memblock_region_memory_base_pfn(r);
7822 if (usable_startpfn < 0x100000) {
7823 mem_below_4gb_not_mirrored = true;
7827 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7828 min(usable_startpfn, zone_movable_pfn[nid]) :
7832 if (mem_below_4gb_not_mirrored)
7833 pr_warn("This configuration results in unmirrored kernel memory.\n");
7839 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7840 * amount of necessary memory.
7842 if (required_kernelcore_percent)
7843 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7845 if (required_movablecore_percent)
7846 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7850 * If movablecore= was specified, calculate what size of
7851 * kernelcore that corresponds so that memory usable for
7852 * any allocation type is evenly spread. If both kernelcore
7853 * and movablecore are specified, then the value of kernelcore
7854 * will be used for required_kernelcore if it's greater than
7855 * what movablecore would have allowed.
7857 if (required_movablecore) {
7858 unsigned long corepages;
7861 * Round-up so that ZONE_MOVABLE is at least as large as what
7862 * was requested by the user
7864 required_movablecore =
7865 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7866 required_movablecore = min(totalpages, required_movablecore);
7867 corepages = totalpages - required_movablecore;
7869 required_kernelcore = max(required_kernelcore, corepages);
7873 * If kernelcore was not specified or kernelcore size is larger
7874 * than totalpages, there is no ZONE_MOVABLE.
7876 if (!required_kernelcore || required_kernelcore >= totalpages)
7879 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7880 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7883 /* Spread kernelcore memory as evenly as possible throughout nodes */
7884 kernelcore_node = required_kernelcore / usable_nodes;
7885 for_each_node_state(nid, N_MEMORY) {
7886 unsigned long start_pfn, end_pfn;
7889 * Recalculate kernelcore_node if the division per node
7890 * now exceeds what is necessary to satisfy the requested
7891 * amount of memory for the kernel
7893 if (required_kernelcore < kernelcore_node)
7894 kernelcore_node = required_kernelcore / usable_nodes;
7897 * As the map is walked, we track how much memory is usable
7898 * by the kernel using kernelcore_remaining. When it is
7899 * 0, the rest of the node is usable by ZONE_MOVABLE
7901 kernelcore_remaining = kernelcore_node;
7903 /* Go through each range of PFNs within this node */
7904 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7905 unsigned long size_pages;
7907 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7908 if (start_pfn >= end_pfn)
7911 /* Account for what is only usable for kernelcore */
7912 if (start_pfn < usable_startpfn) {
7913 unsigned long kernel_pages;
7914 kernel_pages = min(end_pfn, usable_startpfn)
7917 kernelcore_remaining -= min(kernel_pages,
7918 kernelcore_remaining);
7919 required_kernelcore -= min(kernel_pages,
7920 required_kernelcore);
7922 /* Continue if range is now fully accounted */
7923 if (end_pfn <= usable_startpfn) {
7926 * Push zone_movable_pfn to the end so
7927 * that if we have to rebalance
7928 * kernelcore across nodes, we will
7929 * not double account here
7931 zone_movable_pfn[nid] = end_pfn;
7934 start_pfn = usable_startpfn;
7938 * The usable PFN range for ZONE_MOVABLE is from
7939 * start_pfn->end_pfn. Calculate size_pages as the
7940 * number of pages used as kernelcore
7942 size_pages = end_pfn - start_pfn;
7943 if (size_pages > kernelcore_remaining)
7944 size_pages = kernelcore_remaining;
7945 zone_movable_pfn[nid] = start_pfn + size_pages;
7948 * Some kernelcore has been met, update counts and
7949 * break if the kernelcore for this node has been
7952 required_kernelcore -= min(required_kernelcore,
7954 kernelcore_remaining -= size_pages;
7955 if (!kernelcore_remaining)
7961 * If there is still required_kernelcore, we do another pass with one
7962 * less node in the count. This will push zone_movable_pfn[nid] further
7963 * along on the nodes that still have memory until kernelcore is
7967 if (usable_nodes && required_kernelcore > usable_nodes)
7971 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7972 for (nid = 0; nid < MAX_NUMNODES; nid++)
7973 zone_movable_pfn[nid] =
7974 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7977 /* restore the node_state */
7978 node_states[N_MEMORY] = saved_node_state;
7981 /* Any regular or high memory on that node ? */
7982 static void check_for_memory(pg_data_t *pgdat, int nid)
7984 enum zone_type zone_type;
7986 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7987 struct zone *zone = &pgdat->node_zones[zone_type];
7988 if (populated_zone(zone)) {
7989 if (IS_ENABLED(CONFIG_HIGHMEM))
7990 node_set_state(nid, N_HIGH_MEMORY);
7991 if (zone_type <= ZONE_NORMAL)
7992 node_set_state(nid, N_NORMAL_MEMORY);
7999 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8000 * such cases we allow max_zone_pfn sorted in the descending order
8002 bool __weak arch_has_descending_max_zone_pfns(void)
8008 * free_area_init - Initialise all pg_data_t and zone data
8009 * @max_zone_pfn: an array of max PFNs for each zone
8011 * This will call free_area_init_node() for each active node in the system.
8012 * Using the page ranges provided by memblock_set_node(), the size of each
8013 * zone in each node and their holes is calculated. If the maximum PFN
8014 * between two adjacent zones match, it is assumed that the zone is empty.
8015 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8016 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8017 * starts where the previous one ended. For example, ZONE_DMA32 starts
8018 * at arch_max_dma_pfn.
8020 void __init free_area_init(unsigned long *max_zone_pfn)
8022 unsigned long start_pfn, end_pfn;
8026 /* Record where the zone boundaries are */
8027 memset(arch_zone_lowest_possible_pfn, 0,
8028 sizeof(arch_zone_lowest_possible_pfn));
8029 memset(arch_zone_highest_possible_pfn, 0,
8030 sizeof(arch_zone_highest_possible_pfn));
8032 start_pfn = find_min_pfn_with_active_regions();
8033 descending = arch_has_descending_max_zone_pfns();
8035 for (i = 0; i < MAX_NR_ZONES; i++) {
8037 zone = MAX_NR_ZONES - i - 1;
8041 if (zone == ZONE_MOVABLE)
8044 end_pfn = max(max_zone_pfn[zone], start_pfn);
8045 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8046 arch_zone_highest_possible_pfn[zone] = end_pfn;
8048 start_pfn = end_pfn;
8051 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8052 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8053 find_zone_movable_pfns_for_nodes();
8055 /* Print out the zone ranges */
8056 pr_info("Zone ranges:\n");
8057 for (i = 0; i < MAX_NR_ZONES; i++) {
8058 if (i == ZONE_MOVABLE)
8060 pr_info(" %-8s ", zone_names[i]);
8061 if (arch_zone_lowest_possible_pfn[i] ==
8062 arch_zone_highest_possible_pfn[i])
8065 pr_cont("[mem %#018Lx-%#018Lx]\n",
8066 (u64)arch_zone_lowest_possible_pfn[i]
8068 ((u64)arch_zone_highest_possible_pfn[i]
8069 << PAGE_SHIFT) - 1);
8072 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8073 pr_info("Movable zone start for each node\n");
8074 for (i = 0; i < MAX_NUMNODES; i++) {
8075 if (zone_movable_pfn[i])
8076 pr_info(" Node %d: %#018Lx\n", i,
8077 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8081 * Print out the early node map, and initialize the
8082 * subsection-map relative to active online memory ranges to
8083 * enable future "sub-section" extensions of the memory map.
8085 pr_info("Early memory node ranges\n");
8086 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8087 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8088 (u64)start_pfn << PAGE_SHIFT,
8089 ((u64)end_pfn << PAGE_SHIFT) - 1);
8090 subsection_map_init(start_pfn, end_pfn - start_pfn);
8093 /* Initialise every node */
8094 mminit_verify_pageflags_layout();
8095 setup_nr_node_ids();
8096 for_each_online_node(nid) {
8097 pg_data_t *pgdat = NODE_DATA(nid);
8098 free_area_init_node(nid);
8100 /* Any memory on that node */
8101 if (pgdat->node_present_pages)
8102 node_set_state(nid, N_MEMORY);
8103 check_for_memory(pgdat, nid);
8109 static int __init cmdline_parse_core(char *p, unsigned long *core,
8110 unsigned long *percent)
8112 unsigned long long coremem;
8118 /* Value may be a percentage of total memory, otherwise bytes */
8119 coremem = simple_strtoull(p, &endptr, 0);
8120 if (*endptr == '%') {
8121 /* Paranoid check for percent values greater than 100 */
8122 WARN_ON(coremem > 100);
8126 coremem = memparse(p, &p);
8127 /* Paranoid check that UL is enough for the coremem value */
8128 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8130 *core = coremem >> PAGE_SHIFT;
8137 * kernelcore=size sets the amount of memory for use for allocations that
8138 * cannot be reclaimed or migrated.
8140 static int __init cmdline_parse_kernelcore(char *p)
8142 /* parse kernelcore=mirror */
8143 if (parse_option_str(p, "mirror")) {
8144 mirrored_kernelcore = true;
8148 return cmdline_parse_core(p, &required_kernelcore,
8149 &required_kernelcore_percent);
8153 * movablecore=size sets the amount of memory for use for allocations that
8154 * can be reclaimed or migrated.
8156 static int __init cmdline_parse_movablecore(char *p)
8158 return cmdline_parse_core(p, &required_movablecore,
8159 &required_movablecore_percent);
8162 early_param("kernelcore", cmdline_parse_kernelcore);
8163 early_param("movablecore", cmdline_parse_movablecore);
8165 void adjust_managed_page_count(struct page *page, long count)
8167 atomic_long_add(count, &page_zone(page)->managed_pages);
8168 totalram_pages_add(count);
8169 #ifdef CONFIG_HIGHMEM
8170 if (PageHighMem(page))
8171 totalhigh_pages_add(count);
8174 EXPORT_SYMBOL(adjust_managed_page_count);
8176 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8179 unsigned long pages = 0;
8181 start = (void *)PAGE_ALIGN((unsigned long)start);
8182 end = (void *)((unsigned long)end & PAGE_MASK);
8183 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8184 struct page *page = virt_to_page(pos);
8185 void *direct_map_addr;
8188 * 'direct_map_addr' might be different from 'pos'
8189 * because some architectures' virt_to_page()
8190 * work with aliases. Getting the direct map
8191 * address ensures that we get a _writeable_
8192 * alias for the memset().
8194 direct_map_addr = page_address(page);
8196 * Perform a kasan-unchecked memset() since this memory
8197 * has not been initialized.
8199 direct_map_addr = kasan_reset_tag(direct_map_addr);
8200 if ((unsigned int)poison <= 0xFF)
8201 memset(direct_map_addr, poison, PAGE_SIZE);
8203 free_reserved_page(page);
8207 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8212 void __init mem_init_print_info(void)
8214 unsigned long physpages, codesize, datasize, rosize, bss_size;
8215 unsigned long init_code_size, init_data_size;
8217 physpages = get_num_physpages();
8218 codesize = _etext - _stext;
8219 datasize = _edata - _sdata;
8220 rosize = __end_rodata - __start_rodata;
8221 bss_size = __bss_stop - __bss_start;
8222 init_data_size = __init_end - __init_begin;
8223 init_code_size = _einittext - _sinittext;
8226 * Detect special cases and adjust section sizes accordingly:
8227 * 1) .init.* may be embedded into .data sections
8228 * 2) .init.text.* may be out of [__init_begin, __init_end],
8229 * please refer to arch/tile/kernel/vmlinux.lds.S.
8230 * 3) .rodata.* may be embedded into .text or .data sections.
8232 #define adj_init_size(start, end, size, pos, adj) \
8234 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8238 adj_init_size(__init_begin, __init_end, init_data_size,
8239 _sinittext, init_code_size);
8240 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8241 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8242 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8243 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8245 #undef adj_init_size
8247 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8248 #ifdef CONFIG_HIGHMEM
8252 K(nr_free_pages()), K(physpages),
8253 codesize >> 10, datasize >> 10, rosize >> 10,
8254 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8255 K(physpages - totalram_pages() - totalcma_pages),
8257 #ifdef CONFIG_HIGHMEM
8258 , K(totalhigh_pages())
8264 * set_dma_reserve - set the specified number of pages reserved in the first zone
8265 * @new_dma_reserve: The number of pages to mark reserved
8267 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8268 * In the DMA zone, a significant percentage may be consumed by kernel image
8269 * and other unfreeable allocations which can skew the watermarks badly. This
8270 * function may optionally be used to account for unfreeable pages in the
8271 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8272 * smaller per-cpu batchsize.
8274 void __init set_dma_reserve(unsigned long new_dma_reserve)
8276 dma_reserve = new_dma_reserve;
8279 static int page_alloc_cpu_dead(unsigned int cpu)
8283 lru_add_drain_cpu(cpu);
8287 * Spill the event counters of the dead processor
8288 * into the current processors event counters.
8289 * This artificially elevates the count of the current
8292 vm_events_fold_cpu(cpu);
8295 * Zero the differential counters of the dead processor
8296 * so that the vm statistics are consistent.
8298 * This is only okay since the processor is dead and cannot
8299 * race with what we are doing.
8301 cpu_vm_stats_fold(cpu);
8303 for_each_populated_zone(zone)
8304 zone_pcp_update(zone, 0);
8309 static int page_alloc_cpu_online(unsigned int cpu)
8313 for_each_populated_zone(zone)
8314 zone_pcp_update(zone, 1);
8319 int hashdist = HASHDIST_DEFAULT;
8321 static int __init set_hashdist(char *str)
8325 hashdist = simple_strtoul(str, &str, 0);
8328 __setup("hashdist=", set_hashdist);
8331 void __init page_alloc_init(void)
8336 if (num_node_state(N_MEMORY) == 1)
8340 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8341 "mm/page_alloc:pcp",
8342 page_alloc_cpu_online,
8343 page_alloc_cpu_dead);
8348 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8349 * or min_free_kbytes changes.
8351 static void calculate_totalreserve_pages(void)
8353 struct pglist_data *pgdat;
8354 unsigned long reserve_pages = 0;
8355 enum zone_type i, j;
8357 for_each_online_pgdat(pgdat) {
8359 pgdat->totalreserve_pages = 0;
8361 for (i = 0; i < MAX_NR_ZONES; i++) {
8362 struct zone *zone = pgdat->node_zones + i;
8364 unsigned long managed_pages = zone_managed_pages(zone);
8366 /* Find valid and maximum lowmem_reserve in the zone */
8367 for (j = i; j < MAX_NR_ZONES; j++) {
8368 if (zone->lowmem_reserve[j] > max)
8369 max = zone->lowmem_reserve[j];
8372 /* we treat the high watermark as reserved pages. */
8373 max += high_wmark_pages(zone);
8375 if (max > managed_pages)
8376 max = managed_pages;
8378 pgdat->totalreserve_pages += max;
8380 reserve_pages += max;
8383 totalreserve_pages = reserve_pages;
8387 * setup_per_zone_lowmem_reserve - called whenever
8388 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8389 * has a correct pages reserved value, so an adequate number of
8390 * pages are left in the zone after a successful __alloc_pages().
8392 static void setup_per_zone_lowmem_reserve(void)
8394 struct pglist_data *pgdat;
8395 enum zone_type i, j;
8397 for_each_online_pgdat(pgdat) {
8398 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8399 struct zone *zone = &pgdat->node_zones[i];
8400 int ratio = sysctl_lowmem_reserve_ratio[i];
8401 bool clear = !ratio || !zone_managed_pages(zone);
8402 unsigned long managed_pages = 0;
8404 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8405 struct zone *upper_zone = &pgdat->node_zones[j];
8407 managed_pages += zone_managed_pages(upper_zone);
8410 zone->lowmem_reserve[j] = 0;
8412 zone->lowmem_reserve[j] = managed_pages / ratio;
8417 /* update totalreserve_pages */
8418 calculate_totalreserve_pages();
8421 static void __setup_per_zone_wmarks(void)
8423 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8424 unsigned long lowmem_pages = 0;
8426 unsigned long flags;
8428 /* Calculate total number of !ZONE_HIGHMEM pages */
8429 for_each_zone(zone) {
8430 if (!is_highmem(zone))
8431 lowmem_pages += zone_managed_pages(zone);
8434 for_each_zone(zone) {
8437 spin_lock_irqsave(&zone->lock, flags);
8438 tmp = (u64)pages_min * zone_managed_pages(zone);
8439 do_div(tmp, lowmem_pages);
8440 if (is_highmem(zone)) {
8442 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8443 * need highmem pages, so cap pages_min to a small
8446 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8447 * deltas control async page reclaim, and so should
8448 * not be capped for highmem.
8450 unsigned long min_pages;
8452 min_pages = zone_managed_pages(zone) / 1024;
8453 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8454 zone->_watermark[WMARK_MIN] = min_pages;
8457 * If it's a lowmem zone, reserve a number of pages
8458 * proportionate to the zone's size.
8460 zone->_watermark[WMARK_MIN] = tmp;
8464 * Set the kswapd watermarks distance according to the
8465 * scale factor in proportion to available memory, but
8466 * ensure a minimum size on small systems.
8468 tmp = max_t(u64, tmp >> 2,
8469 mult_frac(zone_managed_pages(zone),
8470 watermark_scale_factor, 10000));
8472 zone->watermark_boost = 0;
8473 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8474 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8476 spin_unlock_irqrestore(&zone->lock, flags);
8479 /* update totalreserve_pages */
8480 calculate_totalreserve_pages();
8484 * setup_per_zone_wmarks - called when min_free_kbytes changes
8485 * or when memory is hot-{added|removed}
8487 * Ensures that the watermark[min,low,high] values for each zone are set
8488 * correctly with respect to min_free_kbytes.
8490 void setup_per_zone_wmarks(void)
8493 static DEFINE_SPINLOCK(lock);
8496 __setup_per_zone_wmarks();
8500 * The watermark size have changed so update the pcpu batch
8501 * and high limits or the limits may be inappropriate.
8504 zone_pcp_update(zone, 0);
8508 * Initialise min_free_kbytes.
8510 * For small machines we want it small (128k min). For large machines
8511 * we want it large (256MB max). But it is not linear, because network
8512 * bandwidth does not increase linearly with machine size. We use
8514 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8515 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8531 void calculate_min_free_kbytes(void)
8533 unsigned long lowmem_kbytes;
8534 int new_min_free_kbytes;
8536 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8537 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8539 if (new_min_free_kbytes > user_min_free_kbytes)
8540 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8542 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8543 new_min_free_kbytes, user_min_free_kbytes);
8547 int __meminit init_per_zone_wmark_min(void)
8549 calculate_min_free_kbytes();
8550 setup_per_zone_wmarks();
8551 refresh_zone_stat_thresholds();
8552 setup_per_zone_lowmem_reserve();
8555 setup_min_unmapped_ratio();
8556 setup_min_slab_ratio();
8559 khugepaged_min_free_kbytes_update();
8563 postcore_initcall(init_per_zone_wmark_min)
8566 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8567 * that we can call two helper functions whenever min_free_kbytes
8570 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8571 void *buffer, size_t *length, loff_t *ppos)
8575 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8580 user_min_free_kbytes = min_free_kbytes;
8581 setup_per_zone_wmarks();
8586 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8587 void *buffer, size_t *length, loff_t *ppos)
8591 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8596 setup_per_zone_wmarks();
8602 static void setup_min_unmapped_ratio(void)
8607 for_each_online_pgdat(pgdat)
8608 pgdat->min_unmapped_pages = 0;
8611 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8612 sysctl_min_unmapped_ratio) / 100;
8616 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8617 void *buffer, size_t *length, loff_t *ppos)
8621 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8625 setup_min_unmapped_ratio();
8630 static void setup_min_slab_ratio(void)
8635 for_each_online_pgdat(pgdat)
8636 pgdat->min_slab_pages = 0;
8639 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8640 sysctl_min_slab_ratio) / 100;
8643 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8644 void *buffer, size_t *length, loff_t *ppos)
8648 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8652 setup_min_slab_ratio();
8659 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8660 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8661 * whenever sysctl_lowmem_reserve_ratio changes.
8663 * The reserve ratio obviously has absolutely no relation with the
8664 * minimum watermarks. The lowmem reserve ratio can only make sense
8665 * if in function of the boot time zone sizes.
8667 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8668 void *buffer, size_t *length, loff_t *ppos)
8672 proc_dointvec_minmax(table, write, buffer, length, ppos);
8674 for (i = 0; i < MAX_NR_ZONES; i++) {
8675 if (sysctl_lowmem_reserve_ratio[i] < 1)
8676 sysctl_lowmem_reserve_ratio[i] = 0;
8679 setup_per_zone_lowmem_reserve();
8684 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8685 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8686 * pagelist can have before it gets flushed back to buddy allocator.
8688 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8689 int write, void *buffer, size_t *length, loff_t *ppos)
8692 int old_percpu_pagelist_high_fraction;
8695 mutex_lock(&pcp_batch_high_lock);
8696 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8698 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8699 if (!write || ret < 0)
8702 /* Sanity checking to avoid pcp imbalance */
8703 if (percpu_pagelist_high_fraction &&
8704 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8705 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8711 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8714 for_each_populated_zone(zone)
8715 zone_set_pageset_high_and_batch(zone, 0);
8717 mutex_unlock(&pcp_batch_high_lock);
8721 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8723 * Returns the number of pages that arch has reserved but
8724 * is not known to alloc_large_system_hash().
8726 static unsigned long __init arch_reserved_kernel_pages(void)
8733 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8734 * machines. As memory size is increased the scale is also increased but at
8735 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8736 * quadruples the scale is increased by one, which means the size of hash table
8737 * only doubles, instead of quadrupling as well.
8738 * Because 32-bit systems cannot have large physical memory, where this scaling
8739 * makes sense, it is disabled on such platforms.
8741 #if __BITS_PER_LONG > 32
8742 #define ADAPT_SCALE_BASE (64ul << 30)
8743 #define ADAPT_SCALE_SHIFT 2
8744 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8748 * allocate a large system hash table from bootmem
8749 * - it is assumed that the hash table must contain an exact power-of-2
8750 * quantity of entries
8751 * - limit is the number of hash buckets, not the total allocation size
8753 void *__init alloc_large_system_hash(const char *tablename,
8754 unsigned long bucketsize,
8755 unsigned long numentries,
8758 unsigned int *_hash_shift,
8759 unsigned int *_hash_mask,
8760 unsigned long low_limit,
8761 unsigned long high_limit)
8763 unsigned long long max = high_limit;
8764 unsigned long log2qty, size;
8770 /* allow the kernel cmdline to have a say */
8772 /* round applicable memory size up to nearest megabyte */
8773 numentries = nr_kernel_pages;
8774 numentries -= arch_reserved_kernel_pages();
8776 /* It isn't necessary when PAGE_SIZE >= 1MB */
8777 if (PAGE_SHIFT < 20)
8778 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8780 #if __BITS_PER_LONG > 32
8782 unsigned long adapt;
8784 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8785 adapt <<= ADAPT_SCALE_SHIFT)
8790 /* limit to 1 bucket per 2^scale bytes of low memory */
8791 if (scale > PAGE_SHIFT)
8792 numentries >>= (scale - PAGE_SHIFT);
8794 numentries <<= (PAGE_SHIFT - scale);
8796 /* Make sure we've got at least a 0-order allocation.. */
8797 if (unlikely(flags & HASH_SMALL)) {
8798 /* Makes no sense without HASH_EARLY */
8799 WARN_ON(!(flags & HASH_EARLY));
8800 if (!(numentries >> *_hash_shift)) {
8801 numentries = 1UL << *_hash_shift;
8802 BUG_ON(!numentries);
8804 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8805 numentries = PAGE_SIZE / bucketsize;
8807 numentries = roundup_pow_of_two(numentries);
8809 /* limit allocation size to 1/16 total memory by default */
8811 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8812 do_div(max, bucketsize);
8814 max = min(max, 0x80000000ULL);
8816 if (numentries < low_limit)
8817 numentries = low_limit;
8818 if (numentries > max)
8821 log2qty = ilog2(numentries);
8823 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8826 size = bucketsize << log2qty;
8827 if (flags & HASH_EARLY) {
8828 if (flags & HASH_ZERO)
8829 table = memblock_alloc(size, SMP_CACHE_BYTES);
8831 table = memblock_alloc_raw(size,
8833 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8834 table = __vmalloc(size, gfp_flags);
8837 huge = is_vm_area_hugepages(table);
8840 * If bucketsize is not a power-of-two, we may free
8841 * some pages at the end of hash table which
8842 * alloc_pages_exact() automatically does
8844 table = alloc_pages_exact(size, gfp_flags);
8845 kmemleak_alloc(table, size, 1, gfp_flags);
8847 } while (!table && size > PAGE_SIZE && --log2qty);
8850 panic("Failed to allocate %s hash table\n", tablename);
8852 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8853 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8854 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8857 *_hash_shift = log2qty;
8859 *_hash_mask = (1 << log2qty) - 1;
8865 * This function checks whether pageblock includes unmovable pages or not.
8867 * PageLRU check without isolation or lru_lock could race so that
8868 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8869 * check without lock_page also may miss some movable non-lru pages at
8870 * race condition. So you can't expect this function should be exact.
8872 * Returns a page without holding a reference. If the caller wants to
8873 * dereference that page (e.g., dumping), it has to make sure that it
8874 * cannot get removed (e.g., via memory unplug) concurrently.
8877 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8878 int migratetype, int flags)
8880 unsigned long iter = 0;
8881 unsigned long pfn = page_to_pfn(page);
8882 unsigned long offset = pfn % pageblock_nr_pages;
8884 if (is_migrate_cma_page(page)) {
8886 * CMA allocations (alloc_contig_range) really need to mark
8887 * isolate CMA pageblocks even when they are not movable in fact
8888 * so consider them movable here.
8890 if (is_migrate_cma(migratetype))
8896 for (; iter < pageblock_nr_pages - offset; iter++) {
8897 page = pfn_to_page(pfn + iter);
8900 * Both, bootmem allocations and memory holes are marked
8901 * PG_reserved and are unmovable. We can even have unmovable
8902 * allocations inside ZONE_MOVABLE, for example when
8903 * specifying "movablecore".
8905 if (PageReserved(page))
8909 * If the zone is movable and we have ruled out all reserved
8910 * pages then it should be reasonably safe to assume the rest
8913 if (zone_idx(zone) == ZONE_MOVABLE)
8917 * Hugepages are not in LRU lists, but they're movable.
8918 * THPs are on the LRU, but need to be counted as #small pages.
8919 * We need not scan over tail pages because we don't
8920 * handle each tail page individually in migration.
8922 if (PageHuge(page) || PageTransCompound(page)) {
8923 struct page *head = compound_head(page);
8924 unsigned int skip_pages;
8926 if (PageHuge(page)) {
8927 if (!hugepage_migration_supported(page_hstate(head)))
8929 } else if (!PageLRU(head) && !__PageMovable(head)) {
8933 skip_pages = compound_nr(head) - (page - head);
8934 iter += skip_pages - 1;
8939 * We can't use page_count without pin a page
8940 * because another CPU can free compound page.
8941 * This check already skips compound tails of THP
8942 * because their page->_refcount is zero at all time.
8944 if (!page_ref_count(page)) {
8945 if (PageBuddy(page))
8946 iter += (1 << buddy_order(page)) - 1;
8951 * The HWPoisoned page may be not in buddy system, and
8952 * page_count() is not 0.
8954 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8958 * We treat all PageOffline() pages as movable when offlining
8959 * to give drivers a chance to decrement their reference count
8960 * in MEM_GOING_OFFLINE in order to indicate that these pages
8961 * can be offlined as there are no direct references anymore.
8962 * For actually unmovable PageOffline() where the driver does
8963 * not support this, we will fail later when trying to actually
8964 * move these pages that still have a reference count > 0.
8965 * (false negatives in this function only)
8967 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8970 if (__PageMovable(page) || PageLRU(page))
8974 * If there are RECLAIMABLE pages, we need to check
8975 * it. But now, memory offline itself doesn't call
8976 * shrink_node_slabs() and it still to be fixed.
8983 #ifdef CONFIG_CONTIG_ALLOC
8984 static unsigned long pfn_max_align_down(unsigned long pfn)
8986 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8987 pageblock_nr_pages) - 1);
8990 static unsigned long pfn_max_align_up(unsigned long pfn)
8992 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8993 pageblock_nr_pages));
8996 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8997 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8998 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8999 static void alloc_contig_dump_pages(struct list_head *page_list)
9001 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9003 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9007 list_for_each_entry(page, page_list, lru)
9008 dump_page(page, "migration failure");
9012 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9017 /* [start, end) must belong to a single zone. */
9018 static int __alloc_contig_migrate_range(struct compact_control *cc,
9019 unsigned long start, unsigned long end)
9021 /* This function is based on compact_zone() from compaction.c. */
9022 unsigned int nr_reclaimed;
9023 unsigned long pfn = start;
9024 unsigned int tries = 0;
9026 struct migration_target_control mtc = {
9027 .nid = zone_to_nid(cc->zone),
9028 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9031 lru_cache_disable();
9033 while (pfn < end || !list_empty(&cc->migratepages)) {
9034 if (fatal_signal_pending(current)) {
9039 if (list_empty(&cc->migratepages)) {
9040 cc->nr_migratepages = 0;
9041 ret = isolate_migratepages_range(cc, pfn, end);
9042 if (ret && ret != -EAGAIN)
9044 pfn = cc->migrate_pfn;
9046 } else if (++tries == 5) {
9051 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9053 cc->nr_migratepages -= nr_reclaimed;
9055 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9056 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9059 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9060 * to retry again over this error, so do the same here.
9069 alloc_contig_dump_pages(&cc->migratepages);
9070 putback_movable_pages(&cc->migratepages);
9077 * alloc_contig_range() -- tries to allocate given range of pages
9078 * @start: start PFN to allocate
9079 * @end: one-past-the-last PFN to allocate
9080 * @migratetype: migratetype of the underlying pageblocks (either
9081 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9082 * in range must have the same migratetype and it must
9083 * be either of the two.
9084 * @gfp_mask: GFP mask to use during compaction
9086 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9087 * aligned. The PFN range must belong to a single zone.
9089 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9090 * pageblocks in the range. Once isolated, the pageblocks should not
9091 * be modified by others.
9093 * Return: zero on success or negative error code. On success all
9094 * pages which PFN is in [start, end) are allocated for the caller and
9095 * need to be freed with free_contig_range().
9097 int alloc_contig_range(unsigned long start, unsigned long end,
9098 unsigned migratetype, gfp_t gfp_mask)
9100 unsigned long outer_start, outer_end;
9104 struct compact_control cc = {
9105 .nr_migratepages = 0,
9107 .zone = page_zone(pfn_to_page(start)),
9108 .mode = MIGRATE_SYNC,
9109 .ignore_skip_hint = true,
9110 .no_set_skip_hint = true,
9111 .gfp_mask = current_gfp_context(gfp_mask),
9112 .alloc_contig = true,
9114 INIT_LIST_HEAD(&cc.migratepages);
9117 * What we do here is we mark all pageblocks in range as
9118 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9119 * have different sizes, and due to the way page allocator
9120 * work, we align the range to biggest of the two pages so
9121 * that page allocator won't try to merge buddies from
9122 * different pageblocks and change MIGRATE_ISOLATE to some
9123 * other migration type.
9125 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9126 * migrate the pages from an unaligned range (ie. pages that
9127 * we are interested in). This will put all the pages in
9128 * range back to page allocator as MIGRATE_ISOLATE.
9130 * When this is done, we take the pages in range from page
9131 * allocator removing them from the buddy system. This way
9132 * page allocator will never consider using them.
9134 * This lets us mark the pageblocks back as
9135 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9136 * aligned range but not in the unaligned, original range are
9137 * put back to page allocator so that buddy can use them.
9140 ret = start_isolate_page_range(pfn_max_align_down(start),
9141 pfn_max_align_up(end), migratetype, 0);
9145 drain_all_pages(cc.zone);
9148 * In case of -EBUSY, we'd like to know which page causes problem.
9149 * So, just fall through. test_pages_isolated() has a tracepoint
9150 * which will report the busy page.
9152 * It is possible that busy pages could become available before
9153 * the call to test_pages_isolated, and the range will actually be
9154 * allocated. So, if we fall through be sure to clear ret so that
9155 * -EBUSY is not accidentally used or returned to caller.
9157 ret = __alloc_contig_migrate_range(&cc, start, end);
9158 if (ret && ret != -EBUSY)
9163 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9164 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9165 * more, all pages in [start, end) are free in page allocator.
9166 * What we are going to do is to allocate all pages from
9167 * [start, end) (that is remove them from page allocator).
9169 * The only problem is that pages at the beginning and at the
9170 * end of interesting range may be not aligned with pages that
9171 * page allocator holds, ie. they can be part of higher order
9172 * pages. Because of this, we reserve the bigger range and
9173 * once this is done free the pages we are not interested in.
9175 * We don't have to hold zone->lock here because the pages are
9176 * isolated thus they won't get removed from buddy.
9180 outer_start = start;
9181 while (!PageBuddy(pfn_to_page(outer_start))) {
9182 if (++order >= MAX_ORDER) {
9183 outer_start = start;
9186 outer_start &= ~0UL << order;
9189 if (outer_start != start) {
9190 order = buddy_order(pfn_to_page(outer_start));
9193 * outer_start page could be small order buddy page and
9194 * it doesn't include start page. Adjust outer_start
9195 * in this case to report failed page properly
9196 * on tracepoint in test_pages_isolated()
9198 if (outer_start + (1UL << order) <= start)
9199 outer_start = start;
9202 /* Make sure the range is really isolated. */
9203 if (test_pages_isolated(outer_start, end, 0)) {
9208 /* Grab isolated pages from freelists. */
9209 outer_end = isolate_freepages_range(&cc, outer_start, end);
9215 /* Free head and tail (if any) */
9216 if (start != outer_start)
9217 free_contig_range(outer_start, start - outer_start);
9218 if (end != outer_end)
9219 free_contig_range(end, outer_end - end);
9222 undo_isolate_page_range(pfn_max_align_down(start),
9223 pfn_max_align_up(end), migratetype);
9226 EXPORT_SYMBOL(alloc_contig_range);
9228 static int __alloc_contig_pages(unsigned long start_pfn,
9229 unsigned long nr_pages, gfp_t gfp_mask)
9231 unsigned long end_pfn = start_pfn + nr_pages;
9233 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9237 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9238 unsigned long nr_pages)
9240 unsigned long i, end_pfn = start_pfn + nr_pages;
9243 for (i = start_pfn; i < end_pfn; i++) {
9244 page = pfn_to_online_page(i);
9248 if (page_zone(page) != z)
9251 if (PageReserved(page))
9257 static bool zone_spans_last_pfn(const struct zone *zone,
9258 unsigned long start_pfn, unsigned long nr_pages)
9260 unsigned long last_pfn = start_pfn + nr_pages - 1;
9262 return zone_spans_pfn(zone, last_pfn);
9266 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9267 * @nr_pages: Number of contiguous pages to allocate
9268 * @gfp_mask: GFP mask to limit search and used during compaction
9270 * @nodemask: Mask for other possible nodes
9272 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9273 * on an applicable zonelist to find a contiguous pfn range which can then be
9274 * tried for allocation with alloc_contig_range(). This routine is intended
9275 * for allocation requests which can not be fulfilled with the buddy allocator.
9277 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9278 * power of two, then allocated range is also guaranteed to be aligned to same
9279 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9281 * Allocated pages can be freed with free_contig_range() or by manually calling
9282 * __free_page() on each allocated page.
9284 * Return: pointer to contiguous pages on success, or NULL if not successful.
9286 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9287 int nid, nodemask_t *nodemask)
9289 unsigned long ret, pfn, flags;
9290 struct zonelist *zonelist;
9294 zonelist = node_zonelist(nid, gfp_mask);
9295 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9296 gfp_zone(gfp_mask), nodemask) {
9297 spin_lock_irqsave(&zone->lock, flags);
9299 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9300 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9301 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9303 * We release the zone lock here because
9304 * alloc_contig_range() will also lock the zone
9305 * at some point. If there's an allocation
9306 * spinning on this lock, it may win the race
9307 * and cause alloc_contig_range() to fail...
9309 spin_unlock_irqrestore(&zone->lock, flags);
9310 ret = __alloc_contig_pages(pfn, nr_pages,
9313 return pfn_to_page(pfn);
9314 spin_lock_irqsave(&zone->lock, flags);
9318 spin_unlock_irqrestore(&zone->lock, flags);
9322 #endif /* CONFIG_CONTIG_ALLOC */
9324 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9326 unsigned long count = 0;
9328 for (; nr_pages--; pfn++) {
9329 struct page *page = pfn_to_page(pfn);
9331 count += page_count(page) != 1;
9334 WARN(count != 0, "%lu pages are still in use!\n", count);
9336 EXPORT_SYMBOL(free_contig_range);
9339 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9340 * page high values need to be recalculated.
9342 void zone_pcp_update(struct zone *zone, int cpu_online)
9344 mutex_lock(&pcp_batch_high_lock);
9345 zone_set_pageset_high_and_batch(zone, cpu_online);
9346 mutex_unlock(&pcp_batch_high_lock);
9350 * Effectively disable pcplists for the zone by setting the high limit to 0
9351 * and draining all cpus. A concurrent page freeing on another CPU that's about
9352 * to put the page on pcplist will either finish before the drain and the page
9353 * will be drained, or observe the new high limit and skip the pcplist.
9355 * Must be paired with a call to zone_pcp_enable().
9357 void zone_pcp_disable(struct zone *zone)
9359 mutex_lock(&pcp_batch_high_lock);
9360 __zone_set_pageset_high_and_batch(zone, 0, 1);
9361 __drain_all_pages(zone, true);
9364 void zone_pcp_enable(struct zone *zone)
9366 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9367 mutex_unlock(&pcp_batch_high_lock);
9370 void zone_pcp_reset(struct zone *zone)
9373 struct per_cpu_zonestat *pzstats;
9375 if (zone->per_cpu_pageset != &boot_pageset) {
9376 for_each_online_cpu(cpu) {
9377 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9378 drain_zonestat(zone, pzstats);
9380 free_percpu(zone->per_cpu_pageset);
9381 free_percpu(zone->per_cpu_zonestats);
9382 zone->per_cpu_pageset = &boot_pageset;
9383 zone->per_cpu_zonestats = &boot_zonestats;
9387 #ifdef CONFIG_MEMORY_HOTREMOVE
9389 * All pages in the range must be in a single zone, must not contain holes,
9390 * must span full sections, and must be isolated before calling this function.
9392 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9394 unsigned long pfn = start_pfn;
9398 unsigned long flags;
9400 offline_mem_sections(pfn, end_pfn);
9401 zone = page_zone(pfn_to_page(pfn));
9402 spin_lock_irqsave(&zone->lock, flags);
9403 while (pfn < end_pfn) {
9404 page = pfn_to_page(pfn);
9406 * The HWPoisoned page may be not in buddy system, and
9407 * page_count() is not 0.
9409 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9414 * At this point all remaining PageOffline() pages have a
9415 * reference count of 0 and can simply be skipped.
9417 if (PageOffline(page)) {
9418 BUG_ON(page_count(page));
9419 BUG_ON(PageBuddy(page));
9424 BUG_ON(page_count(page));
9425 BUG_ON(!PageBuddy(page));
9426 order = buddy_order(page);
9427 del_page_from_free_list(page, zone, order);
9428 pfn += (1 << order);
9430 spin_unlock_irqrestore(&zone->lock, flags);
9435 * This function returns a stable result only if called under zone lock.
9437 bool is_free_buddy_page(struct page *page)
9439 unsigned long pfn = page_to_pfn(page);
9442 for (order = 0; order < MAX_ORDER; order++) {
9443 struct page *page_head = page - (pfn & ((1 << order) - 1));
9445 if (PageBuddy(page_head) &&
9446 buddy_order_unsafe(page_head) >= order)
9450 return order < MAX_ORDER;
9453 #ifdef CONFIG_MEMORY_FAILURE
9455 * Break down a higher-order page in sub-pages, and keep our target out of
9458 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9459 struct page *target, int low, int high,
9462 unsigned long size = 1 << high;
9463 struct page *current_buddy, *next_page;
9465 while (high > low) {
9469 if (target >= &page[size]) {
9470 next_page = page + size;
9471 current_buddy = page;
9474 current_buddy = page + size;
9477 if (set_page_guard(zone, current_buddy, high, migratetype))
9480 if (current_buddy != target) {
9481 add_to_free_list(current_buddy, zone, high, migratetype);
9482 set_buddy_order(current_buddy, high);
9489 * Take a page that will be marked as poisoned off the buddy allocator.
9491 bool take_page_off_buddy(struct page *page)
9493 struct zone *zone = page_zone(page);
9494 unsigned long pfn = page_to_pfn(page);
9495 unsigned long flags;
9499 spin_lock_irqsave(&zone->lock, flags);
9500 for (order = 0; order < MAX_ORDER; order++) {
9501 struct page *page_head = page - (pfn & ((1 << order) - 1));
9502 int page_order = buddy_order(page_head);
9504 if (PageBuddy(page_head) && page_order >= order) {
9505 unsigned long pfn_head = page_to_pfn(page_head);
9506 int migratetype = get_pfnblock_migratetype(page_head,
9509 del_page_from_free_list(page_head, zone, page_order);
9510 break_down_buddy_pages(zone, page_head, page, 0,
9511 page_order, migratetype);
9512 SetPageHWPoisonTakenOff(page);
9513 if (!is_migrate_isolate(migratetype))
9514 __mod_zone_freepage_state(zone, -1, migratetype);
9518 if (page_count(page_head) > 0)
9521 spin_unlock_irqrestore(&zone->lock, flags);
9526 * Cancel takeoff done by take_page_off_buddy().
9528 bool put_page_back_buddy(struct page *page)
9530 struct zone *zone = page_zone(page);
9531 unsigned long pfn = page_to_pfn(page);
9532 unsigned long flags;
9533 int migratetype = get_pfnblock_migratetype(page, pfn);
9536 spin_lock_irqsave(&zone->lock, flags);
9537 if (put_page_testzero(page)) {
9538 ClearPageHWPoisonTakenOff(page);
9539 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9540 if (TestClearPageHWPoison(page)) {
9541 num_poisoned_pages_dec();
9545 spin_unlock_irqrestore(&zone->lock, flags);
9551 #ifdef CONFIG_ZONE_DMA
9552 bool has_managed_dma(void)
9554 struct pglist_data *pgdat;
9556 for_each_online_pgdat(pgdat) {
9557 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9559 if (managed_zone(zone))
9564 #endif /* CONFIG_ZONE_DMA */