1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/swapops.h>
23 #include <linux/interrupt.h>
24 #include <linux/pagemap.h>
25 #include <linux/jiffies.h>
26 #include <linux/memblock.h>
27 #include <linux/compiler.h>
28 #include <linux/kernel.h>
29 #include <linux/kasan.h>
30 #include <linux/module.h>
31 #include <linux/suspend.h>
32 #include <linux/pagevec.h>
33 #include <linux/blkdev.h>
34 #include <linux/slab.h>
35 #include <linux/ratelimit.h>
36 #include <linux/oom.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/random.h>
49 #include <linux/sort.h>
50 #include <linux/pfn.h>
51 #include <linux/backing-dev.h>
52 #include <linux/fault-inject.h>
53 #include <linux/page-isolation.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/migrate.h>
63 #include <linux/hugetlb.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/mm.h>
66 #include <linux/page_owner.h>
67 #include <linux/page_table_check.h>
68 #include <linux/kthread.h>
69 #include <linux/memcontrol.h>
70 #include <linux/ftrace.h>
71 #include <linux/lockdep.h>
72 #include <linux/nmi.h>
73 #include <linux/psi.h>
74 #include <linux/padata.h>
75 #include <linux/khugepaged.h>
76 #include <linux/buffer_head.h>
77 #include <linux/delayacct.h>
78 #include <asm/sections.h>
79 #include <asm/tlbflush.h>
80 #include <asm/div64.h>
83 #include "page_reporting.h"
85 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
86 typedef int __bitwise fpi_t;
88 /* No special request */
89 #define FPI_NONE ((__force fpi_t)0)
92 * Skip free page reporting notification for the (possibly merged) page.
93 * This does not hinder free page reporting from grabbing the page,
94 * reporting it and marking it "reported" - it only skips notifying
95 * the free page reporting infrastructure about a newly freed page. For
96 * example, used when temporarily pulling a page from a freelist and
97 * putting it back unmodified.
99 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
102 * Place the (possibly merged) page to the tail of the freelist. Will ignore
103 * page shuffling (relevant code - e.g., memory onlining - is expected to
104 * shuffle the whole zone).
106 * Note: No code should rely on this flag for correctness - it's purely
107 * to allow for optimizations when handing back either fresh pages
108 * (memory onlining) or untouched pages (page isolation, free page
111 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
114 * Don't poison memory with KASAN (only for the tag-based modes).
115 * During boot, all non-reserved memblock memory is exposed to page_alloc.
116 * Poisoning all that memory lengthens boot time, especially on systems with
117 * large amount of RAM. This flag is used to skip that poisoning.
118 * This is only done for the tag-based KASAN modes, as those are able to
119 * detect memory corruptions with the memory tags assigned by default.
120 * All memory allocated normally after boot gets poisoned as usual.
122 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
124 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
125 static DEFINE_MUTEX(pcp_batch_high_lock);
126 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
131 static DEFINE_PER_CPU(struct pagesets, pagesets) __maybe_unused = {
132 .lock = INIT_LOCAL_LOCK(lock),
135 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
136 DEFINE_PER_CPU(int, numa_node);
137 EXPORT_PER_CPU_SYMBOL(numa_node);
140 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
142 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
144 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
145 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
146 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
147 * defined in <linux/topology.h>.
149 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
150 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
153 /* work_structs for global per-cpu drains */
156 struct work_struct work;
158 static DEFINE_MUTEX(pcpu_drain_mutex);
159 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
161 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
162 volatile unsigned long latent_entropy __latent_entropy;
163 EXPORT_SYMBOL(latent_entropy);
167 * Array of node states.
169 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
170 [N_POSSIBLE] = NODE_MASK_ALL,
171 [N_ONLINE] = { { [0] = 1UL } },
173 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
174 #ifdef CONFIG_HIGHMEM
175 [N_HIGH_MEMORY] = { { [0] = 1UL } },
177 [N_MEMORY] = { { [0] = 1UL } },
178 [N_CPU] = { { [0] = 1UL } },
181 EXPORT_SYMBOL(node_states);
183 atomic_long_t _totalram_pages __read_mostly;
184 EXPORT_SYMBOL(_totalram_pages);
185 unsigned long totalreserve_pages __read_mostly;
186 unsigned long totalcma_pages __read_mostly;
188 int percpu_pagelist_high_fraction;
189 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
190 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
191 EXPORT_SYMBOL(init_on_alloc);
193 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
194 EXPORT_SYMBOL(init_on_free);
196 static bool _init_on_alloc_enabled_early __read_mostly
197 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
198 static int __init early_init_on_alloc(char *buf)
201 return kstrtobool(buf, &_init_on_alloc_enabled_early);
203 early_param("init_on_alloc", early_init_on_alloc);
205 static bool _init_on_free_enabled_early __read_mostly
206 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
207 static int __init early_init_on_free(char *buf)
209 return kstrtobool(buf, &_init_on_free_enabled_early);
211 early_param("init_on_free", early_init_on_free);
214 * A cached value of the page's pageblock's migratetype, used when the page is
215 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
216 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
217 * Also the migratetype set in the page does not necessarily match the pcplist
218 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
219 * other index - this ensures that it will be put on the correct CMA freelist.
221 static inline int get_pcppage_migratetype(struct page *page)
226 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
228 page->index = migratetype;
231 #ifdef CONFIG_PM_SLEEP
233 * The following functions are used by the suspend/hibernate code to temporarily
234 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
235 * while devices are suspended. To avoid races with the suspend/hibernate code,
236 * they should always be called with system_transition_mutex held
237 * (gfp_allowed_mask also should only be modified with system_transition_mutex
238 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
239 * with that modification).
242 static gfp_t saved_gfp_mask;
244 void pm_restore_gfp_mask(void)
246 WARN_ON(!mutex_is_locked(&system_transition_mutex));
247 if (saved_gfp_mask) {
248 gfp_allowed_mask = saved_gfp_mask;
253 void pm_restrict_gfp_mask(void)
255 WARN_ON(!mutex_is_locked(&system_transition_mutex));
256 WARN_ON(saved_gfp_mask);
257 saved_gfp_mask = gfp_allowed_mask;
258 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
261 bool pm_suspended_storage(void)
263 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
267 #endif /* CONFIG_PM_SLEEP */
269 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
270 unsigned int pageblock_order __read_mostly;
273 static void __free_pages_ok(struct page *page, unsigned int order,
277 * results with 256, 32 in the lowmem_reserve sysctl:
278 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
279 * 1G machine -> (16M dma, 784M normal, 224M high)
280 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
281 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
282 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
284 * TBD: should special case ZONE_DMA32 machines here - in those we normally
285 * don't need any ZONE_NORMAL reservation
287 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
288 #ifdef CONFIG_ZONE_DMA
291 #ifdef CONFIG_ZONE_DMA32
295 #ifdef CONFIG_HIGHMEM
301 static char * const zone_names[MAX_NR_ZONES] = {
302 #ifdef CONFIG_ZONE_DMA
305 #ifdef CONFIG_ZONE_DMA32
309 #ifdef CONFIG_HIGHMEM
313 #ifdef CONFIG_ZONE_DEVICE
318 const char * const migratetype_names[MIGRATE_TYPES] = {
326 #ifdef CONFIG_MEMORY_ISOLATION
331 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
332 [NULL_COMPOUND_DTOR] = NULL,
333 [COMPOUND_PAGE_DTOR] = free_compound_page,
334 #ifdef CONFIG_HUGETLB_PAGE
335 [HUGETLB_PAGE_DTOR] = free_huge_page,
337 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
338 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
342 int min_free_kbytes = 1024;
343 int user_min_free_kbytes = -1;
344 int watermark_boost_factor __read_mostly = 15000;
345 int watermark_scale_factor = 10;
347 static unsigned long nr_kernel_pages __initdata;
348 static unsigned long nr_all_pages __initdata;
349 static unsigned long dma_reserve __initdata;
351 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
352 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
353 static unsigned long required_kernelcore __initdata;
354 static unsigned long required_kernelcore_percent __initdata;
355 static unsigned long required_movablecore __initdata;
356 static unsigned long required_movablecore_percent __initdata;
357 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
358 static bool mirrored_kernelcore __meminitdata;
360 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
362 EXPORT_SYMBOL(movable_zone);
365 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
366 unsigned int nr_online_nodes __read_mostly = 1;
367 EXPORT_SYMBOL(nr_node_ids);
368 EXPORT_SYMBOL(nr_online_nodes);
371 int page_group_by_mobility_disabled __read_mostly;
373 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
375 * During boot we initialize deferred pages on-demand, as needed, but once
376 * page_alloc_init_late() has finished, the deferred pages are all initialized,
377 * and we can permanently disable that path.
379 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
381 static inline bool deferred_pages_enabled(void)
383 return static_branch_unlikely(&deferred_pages);
386 /* Returns true if the struct page for the pfn is uninitialised */
387 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
389 int nid = early_pfn_to_nid(pfn);
391 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
398 * Returns true when the remaining initialisation should be deferred until
399 * later in the boot cycle when it can be parallelised.
401 static bool __meminit
402 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
404 static unsigned long prev_end_pfn, nr_initialised;
407 * prev_end_pfn static that contains the end of previous zone
408 * No need to protect because called very early in boot before smp_init.
410 if (prev_end_pfn != end_pfn) {
411 prev_end_pfn = end_pfn;
415 /* Always populate low zones for address-constrained allocations */
416 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
419 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
422 * We start only with one section of pages, more pages are added as
423 * needed until the rest of deferred pages are initialized.
426 if ((nr_initialised > PAGES_PER_SECTION) &&
427 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
428 NODE_DATA(nid)->first_deferred_pfn = pfn;
434 static inline bool deferred_pages_enabled(void)
439 static inline bool early_page_uninitialised(unsigned long pfn)
444 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
450 /* Return a pointer to the bitmap storing bits affecting a block of pages */
451 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
454 #ifdef CONFIG_SPARSEMEM
455 return section_to_usemap(__pfn_to_section(pfn));
457 return page_zone(page)->pageblock_flags;
458 #endif /* CONFIG_SPARSEMEM */
461 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
463 #ifdef CONFIG_SPARSEMEM
464 pfn &= (PAGES_PER_SECTION-1);
466 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
467 #endif /* CONFIG_SPARSEMEM */
468 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
471 static __always_inline
472 unsigned long __get_pfnblock_flags_mask(const struct page *page,
476 unsigned long *bitmap;
477 unsigned long bitidx, word_bitidx;
480 bitmap = get_pageblock_bitmap(page, pfn);
481 bitidx = pfn_to_bitidx(page, pfn);
482 word_bitidx = bitidx / BITS_PER_LONG;
483 bitidx &= (BITS_PER_LONG-1);
485 word = bitmap[word_bitidx];
486 return (word >> bitidx) & mask;
490 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
491 * @page: The page within the block of interest
492 * @pfn: The target page frame number
493 * @mask: mask of bits that the caller is interested in
495 * Return: pageblock_bits flags
497 unsigned long get_pfnblock_flags_mask(const struct page *page,
498 unsigned long pfn, unsigned long mask)
500 return __get_pfnblock_flags_mask(page, pfn, mask);
503 static __always_inline int get_pfnblock_migratetype(const struct page *page,
506 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
510 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
511 * @page: The page within the block of interest
512 * @flags: The flags to set
513 * @pfn: The target page frame number
514 * @mask: mask of bits that the caller is interested in
516 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
520 unsigned long *bitmap;
521 unsigned long bitidx, word_bitidx;
522 unsigned long old_word, word;
524 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
525 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
527 bitmap = get_pageblock_bitmap(page, pfn);
528 bitidx = pfn_to_bitidx(page, pfn);
529 word_bitidx = bitidx / BITS_PER_LONG;
530 bitidx &= (BITS_PER_LONG-1);
532 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
537 word = READ_ONCE(bitmap[word_bitidx]);
539 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
540 if (word == old_word)
546 void set_pageblock_migratetype(struct page *page, int migratetype)
548 if (unlikely(page_group_by_mobility_disabled &&
549 migratetype < MIGRATE_PCPTYPES))
550 migratetype = MIGRATE_UNMOVABLE;
552 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
553 page_to_pfn(page), MIGRATETYPE_MASK);
556 #ifdef CONFIG_DEBUG_VM
557 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
561 unsigned long pfn = page_to_pfn(page);
562 unsigned long sp, start_pfn;
565 seq = zone_span_seqbegin(zone);
566 start_pfn = zone->zone_start_pfn;
567 sp = zone->spanned_pages;
568 if (!zone_spans_pfn(zone, pfn))
570 } while (zone_span_seqretry(zone, seq));
573 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
574 pfn, zone_to_nid(zone), zone->name,
575 start_pfn, start_pfn + sp);
580 static int page_is_consistent(struct zone *zone, struct page *page)
582 if (zone != page_zone(page))
588 * Temporary debugging check for pages not lying within a given zone.
590 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
592 if (page_outside_zone_boundaries(zone, page))
594 if (!page_is_consistent(zone, page))
600 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
606 static void bad_page(struct page *page, const char *reason)
608 static unsigned long resume;
609 static unsigned long nr_shown;
610 static unsigned long nr_unshown;
613 * Allow a burst of 60 reports, then keep quiet for that minute;
614 * or allow a steady drip of one report per second.
616 if (nr_shown == 60) {
617 if (time_before(jiffies, resume)) {
623 "BUG: Bad page state: %lu messages suppressed\n",
630 resume = jiffies + 60 * HZ;
632 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
633 current->comm, page_to_pfn(page));
634 dump_page(page, reason);
639 /* Leave bad fields for debug, except PageBuddy could make trouble */
640 page_mapcount_reset(page); /* remove PageBuddy */
641 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
644 static inline unsigned int order_to_pindex(int migratetype, int order)
648 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
649 if (order > PAGE_ALLOC_COSTLY_ORDER) {
650 VM_BUG_ON(order != pageblock_order);
651 base = PAGE_ALLOC_COSTLY_ORDER + 1;
654 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
657 return (MIGRATE_PCPTYPES * base) + migratetype;
660 static inline int pindex_to_order(unsigned int pindex)
662 int order = pindex / MIGRATE_PCPTYPES;
664 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
665 if (order > PAGE_ALLOC_COSTLY_ORDER)
666 order = pageblock_order;
668 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
674 static inline bool pcp_allowed_order(unsigned int order)
676 if (order <= PAGE_ALLOC_COSTLY_ORDER)
678 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
679 if (order == pageblock_order)
685 static inline void free_the_page(struct page *page, unsigned int order)
687 if (pcp_allowed_order(order)) /* Via pcp? */
688 free_unref_page(page, order);
690 __free_pages_ok(page, order, FPI_NONE);
694 * Higher-order pages are called "compound pages". They are structured thusly:
696 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
698 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
699 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
701 * The first tail page's ->compound_dtor holds the offset in array of compound
702 * page destructors. See compound_page_dtors.
704 * The first tail page's ->compound_order holds the order of allocation.
705 * This usage means that zero-order pages may not be compound.
708 void free_compound_page(struct page *page)
710 mem_cgroup_uncharge(page_folio(page));
711 free_the_page(page, compound_order(page));
714 static void prep_compound_head(struct page *page, unsigned int order)
716 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
717 set_compound_order(page, order);
718 atomic_set(compound_mapcount_ptr(page), -1);
719 atomic_set(compound_pincount_ptr(page), 0);
722 static void prep_compound_tail(struct page *head, int tail_idx)
724 struct page *p = head + tail_idx;
726 p->mapping = TAIL_MAPPING;
727 set_compound_head(p, head);
730 void prep_compound_page(struct page *page, unsigned int order)
733 int nr_pages = 1 << order;
736 for (i = 1; i < nr_pages; i++)
737 prep_compound_tail(page, i);
739 prep_compound_head(page, order);
742 #ifdef CONFIG_DEBUG_PAGEALLOC
743 unsigned int _debug_guardpage_minorder;
745 bool _debug_pagealloc_enabled_early __read_mostly
746 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
747 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
748 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
749 EXPORT_SYMBOL(_debug_pagealloc_enabled);
751 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
753 static int __init early_debug_pagealloc(char *buf)
755 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
757 early_param("debug_pagealloc", early_debug_pagealloc);
759 static int __init debug_guardpage_minorder_setup(char *buf)
763 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
764 pr_err("Bad debug_guardpage_minorder value\n");
767 _debug_guardpage_minorder = res;
768 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
771 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
773 static inline bool set_page_guard(struct zone *zone, struct page *page,
774 unsigned int order, int migratetype)
776 if (!debug_guardpage_enabled())
779 if (order >= debug_guardpage_minorder())
782 __SetPageGuard(page);
783 INIT_LIST_HEAD(&page->lru);
784 set_page_private(page, order);
785 /* Guard pages are not available for any usage */
786 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
791 static inline void clear_page_guard(struct zone *zone, struct page *page,
792 unsigned int order, int migratetype)
794 if (!debug_guardpage_enabled())
797 __ClearPageGuard(page);
799 set_page_private(page, 0);
800 if (!is_migrate_isolate(migratetype))
801 __mod_zone_freepage_state(zone, (1 << order), migratetype);
804 static inline bool set_page_guard(struct zone *zone, struct page *page,
805 unsigned int order, int migratetype) { return false; }
806 static inline void clear_page_guard(struct zone *zone, struct page *page,
807 unsigned int order, int migratetype) {}
811 * Enable static keys related to various memory debugging and hardening options.
812 * Some override others, and depend on early params that are evaluated in the
813 * order of appearance. So we need to first gather the full picture of what was
814 * enabled, and then make decisions.
816 void init_mem_debugging_and_hardening(void)
818 bool page_poisoning_requested = false;
820 #ifdef CONFIG_PAGE_POISONING
822 * Page poisoning is debug page alloc for some arches. If
823 * either of those options are enabled, enable poisoning.
825 if (page_poisoning_enabled() ||
826 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
827 debug_pagealloc_enabled())) {
828 static_branch_enable(&_page_poisoning_enabled);
829 page_poisoning_requested = true;
833 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
834 page_poisoning_requested) {
835 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
836 "will take precedence over init_on_alloc and init_on_free\n");
837 _init_on_alloc_enabled_early = false;
838 _init_on_free_enabled_early = false;
841 if (_init_on_alloc_enabled_early)
842 static_branch_enable(&init_on_alloc);
844 static_branch_disable(&init_on_alloc);
846 if (_init_on_free_enabled_early)
847 static_branch_enable(&init_on_free);
849 static_branch_disable(&init_on_free);
851 #ifdef CONFIG_DEBUG_PAGEALLOC
852 if (!debug_pagealloc_enabled())
855 static_branch_enable(&_debug_pagealloc_enabled);
857 if (!debug_guardpage_minorder())
860 static_branch_enable(&_debug_guardpage_enabled);
864 static inline void set_buddy_order(struct page *page, unsigned int order)
866 set_page_private(page, order);
867 __SetPageBuddy(page);
871 * This function checks whether a page is free && is the buddy
872 * we can coalesce a page and its buddy if
873 * (a) the buddy is not in a hole (check before calling!) &&
874 * (b) the buddy is in the buddy system &&
875 * (c) a page and its buddy have the same order &&
876 * (d) a page and its buddy are in the same zone.
878 * For recording whether a page is in the buddy system, we set PageBuddy.
879 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
881 * For recording page's order, we use page_private(page).
883 static inline bool page_is_buddy(struct page *page, struct page *buddy,
886 if (!page_is_guard(buddy) && !PageBuddy(buddy))
889 if (buddy_order(buddy) != order)
893 * zone check is done late to avoid uselessly calculating
894 * zone/node ids for pages that could never merge.
896 if (page_zone_id(page) != page_zone_id(buddy))
899 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
904 #ifdef CONFIG_COMPACTION
905 static inline struct capture_control *task_capc(struct zone *zone)
907 struct capture_control *capc = current->capture_control;
909 return unlikely(capc) &&
910 !(current->flags & PF_KTHREAD) &&
912 capc->cc->zone == zone ? capc : NULL;
916 compaction_capture(struct capture_control *capc, struct page *page,
917 int order, int migratetype)
919 if (!capc || order != capc->cc->order)
922 /* Do not accidentally pollute CMA or isolated regions*/
923 if (is_migrate_cma(migratetype) ||
924 is_migrate_isolate(migratetype))
928 * Do not let lower order allocations pollute a movable pageblock.
929 * This might let an unmovable request use a reclaimable pageblock
930 * and vice-versa but no more than normal fallback logic which can
931 * have trouble finding a high-order free page.
933 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
941 static inline struct capture_control *task_capc(struct zone *zone)
947 compaction_capture(struct capture_control *capc, struct page *page,
948 int order, int migratetype)
952 #endif /* CONFIG_COMPACTION */
954 /* Used for pages not on another list */
955 static inline void add_to_free_list(struct page *page, struct zone *zone,
956 unsigned int order, int migratetype)
958 struct free_area *area = &zone->free_area[order];
960 list_add(&page->lru, &area->free_list[migratetype]);
964 /* Used for pages not on another list */
965 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
966 unsigned int order, int migratetype)
968 struct free_area *area = &zone->free_area[order];
970 list_add_tail(&page->lru, &area->free_list[migratetype]);
975 * Used for pages which are on another list. Move the pages to the tail
976 * of the list - so the moved pages won't immediately be considered for
977 * allocation again (e.g., optimization for memory onlining).
979 static inline void move_to_free_list(struct page *page, struct zone *zone,
980 unsigned int order, int migratetype)
982 struct free_area *area = &zone->free_area[order];
984 list_move_tail(&page->lru, &area->free_list[migratetype]);
987 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
990 /* clear reported state and update reported page count */
991 if (page_reported(page))
992 __ClearPageReported(page);
994 list_del(&page->lru);
995 __ClearPageBuddy(page);
996 set_page_private(page, 0);
997 zone->free_area[order].nr_free--;
1001 * If this is not the largest possible page, check if the buddy
1002 * of the next-highest order is free. If it is, it's possible
1003 * that pages are being freed that will coalesce soon. In case,
1004 * that is happening, add the free page to the tail of the list
1005 * so it's less likely to be used soon and more likely to be merged
1006 * as a higher order page
1009 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1010 struct page *page, unsigned int order)
1012 struct page *higher_page, *higher_buddy;
1013 unsigned long combined_pfn;
1015 if (order >= MAX_ORDER - 2)
1018 combined_pfn = buddy_pfn & pfn;
1019 higher_page = page + (combined_pfn - pfn);
1020 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1021 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1023 return page_is_buddy(higher_page, higher_buddy, order + 1);
1027 * Freeing function for a buddy system allocator.
1029 * The concept of a buddy system is to maintain direct-mapped table
1030 * (containing bit values) for memory blocks of various "orders".
1031 * The bottom level table contains the map for the smallest allocatable
1032 * units of memory (here, pages), and each level above it describes
1033 * pairs of units from the levels below, hence, "buddies".
1034 * At a high level, all that happens here is marking the table entry
1035 * at the bottom level available, and propagating the changes upward
1036 * as necessary, plus some accounting needed to play nicely with other
1037 * parts of the VM system.
1038 * At each level, we keep a list of pages, which are heads of continuous
1039 * free pages of length of (1 << order) and marked with PageBuddy.
1040 * Page's order is recorded in page_private(page) field.
1041 * So when we are allocating or freeing one, we can derive the state of the
1042 * other. That is, if we allocate a small block, and both were
1043 * free, the remainder of the region must be split into blocks.
1044 * If a block is freed, and its buddy is also free, then this
1045 * triggers coalescing into a block of larger size.
1050 static inline void __free_one_page(struct page *page,
1052 struct zone *zone, unsigned int order,
1053 int migratetype, fpi_t fpi_flags)
1055 struct capture_control *capc = task_capc(zone);
1056 unsigned int max_order = pageblock_order;
1057 unsigned long buddy_pfn;
1058 unsigned long combined_pfn;
1062 VM_BUG_ON(!zone_is_initialized(zone));
1063 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1065 VM_BUG_ON(migratetype == -1);
1066 if (likely(!is_migrate_isolate(migratetype)))
1067 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1069 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1070 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1073 while (order < max_order) {
1074 if (compaction_capture(capc, page, order, migratetype)) {
1075 __mod_zone_freepage_state(zone, -(1 << order),
1079 buddy_pfn = __find_buddy_pfn(pfn, order);
1080 buddy = page + (buddy_pfn - pfn);
1082 if (!page_is_buddy(page, buddy, order))
1085 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1086 * merge with it and move up one order.
1088 if (page_is_guard(buddy))
1089 clear_page_guard(zone, buddy, order, migratetype);
1091 del_page_from_free_list(buddy, zone, order);
1092 combined_pfn = buddy_pfn & pfn;
1093 page = page + (combined_pfn - pfn);
1097 if (order < MAX_ORDER - 1) {
1098 /* If we are here, it means order is >= pageblock_order.
1099 * We want to prevent merge between freepages on pageblock
1100 * without fallbacks and normal pageblock. Without this,
1101 * pageblock isolation could cause incorrect freepage or CMA
1102 * accounting or HIGHATOMIC accounting.
1104 * We don't want to hit this code for the more frequent
1105 * low-order merging.
1109 buddy_pfn = __find_buddy_pfn(pfn, order);
1110 buddy = page + (buddy_pfn - pfn);
1111 buddy_mt = get_pageblock_migratetype(buddy);
1113 if (migratetype != buddy_mt
1114 && (!migratetype_is_mergeable(migratetype) ||
1115 !migratetype_is_mergeable(buddy_mt)))
1117 max_order = order + 1;
1118 goto continue_merging;
1122 set_buddy_order(page, order);
1124 if (fpi_flags & FPI_TO_TAIL)
1126 else if (is_shuffle_order(order))
1127 to_tail = shuffle_pick_tail();
1129 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1132 add_to_free_list_tail(page, zone, order, migratetype);
1134 add_to_free_list(page, zone, order, migratetype);
1136 /* Notify page reporting subsystem of freed page */
1137 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1138 page_reporting_notify_free(order);
1142 * A bad page could be due to a number of fields. Instead of multiple branches,
1143 * try and check multiple fields with one check. The caller must do a detailed
1144 * check if necessary.
1146 static inline bool page_expected_state(struct page *page,
1147 unsigned long check_flags)
1149 if (unlikely(atomic_read(&page->_mapcount) != -1))
1152 if (unlikely((unsigned long)page->mapping |
1153 page_ref_count(page) |
1157 (page->flags & check_flags)))
1163 static const char *page_bad_reason(struct page *page, unsigned long flags)
1165 const char *bad_reason = NULL;
1167 if (unlikely(atomic_read(&page->_mapcount) != -1))
1168 bad_reason = "nonzero mapcount";
1169 if (unlikely(page->mapping != NULL))
1170 bad_reason = "non-NULL mapping";
1171 if (unlikely(page_ref_count(page) != 0))
1172 bad_reason = "nonzero _refcount";
1173 if (unlikely(page->flags & flags)) {
1174 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1175 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1177 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1180 if (unlikely(page->memcg_data))
1181 bad_reason = "page still charged to cgroup";
1186 static void check_free_page_bad(struct page *page)
1189 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1192 static inline int check_free_page(struct page *page)
1194 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1197 /* Something has gone sideways, find it */
1198 check_free_page_bad(page);
1202 static int free_tail_pages_check(struct page *head_page, struct page *page)
1207 * We rely page->lru.next never has bit 0 set, unless the page
1208 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1210 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1212 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1216 switch (page - head_page) {
1218 /* the first tail page: ->mapping may be compound_mapcount() */
1219 if (unlikely(compound_mapcount(page))) {
1220 bad_page(page, "nonzero compound_mapcount");
1226 * the second tail page: ->mapping is
1227 * deferred_list.next -- ignore value.
1231 if (page->mapping != TAIL_MAPPING) {
1232 bad_page(page, "corrupted mapping in tail page");
1237 if (unlikely(!PageTail(page))) {
1238 bad_page(page, "PageTail not set");
1241 if (unlikely(compound_head(page) != head_page)) {
1242 bad_page(page, "compound_head not consistent");
1247 page->mapping = NULL;
1248 clear_compound_head(page);
1253 * Skip KASAN memory poisoning when either:
1255 * 1. Deferred memory initialization has not yet completed,
1256 * see the explanation below.
1257 * 2. Skipping poisoning is requested via FPI_SKIP_KASAN_POISON,
1258 * see the comment next to it.
1259 * 3. Skipping poisoning is requested via __GFP_SKIP_KASAN_POISON,
1260 * see the comment next to it.
1262 * Poisoning pages during deferred memory init will greatly lengthen the
1263 * process and cause problem in large memory systems as the deferred pages
1264 * initialization is done with interrupt disabled.
1266 * Assuming that there will be no reference to those newly initialized
1267 * pages before they are ever allocated, this should have no effect on
1268 * KASAN memory tracking as the poison will be properly inserted at page
1269 * allocation time. The only corner case is when pages are allocated by
1270 * on-demand allocation and then freed again before the deferred pages
1271 * initialization is done, but this is not likely to happen.
1273 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1275 return deferred_pages_enabled() ||
1276 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
1277 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
1278 PageSkipKASanPoison(page);
1281 static void kernel_init_free_pages(struct page *page, int numpages)
1285 /* s390's use of memset() could override KASAN redzones. */
1286 kasan_disable_current();
1287 for (i = 0; i < numpages; i++) {
1288 u8 tag = page_kasan_tag(page + i);
1289 page_kasan_tag_reset(page + i);
1290 clear_highpage(page + i);
1291 page_kasan_tag_set(page + i, tag);
1293 kasan_enable_current();
1296 static __always_inline bool free_pages_prepare(struct page *page,
1297 unsigned int order, bool check_free, fpi_t fpi_flags)
1300 bool init = want_init_on_free();
1302 VM_BUG_ON_PAGE(PageTail(page), page);
1304 trace_mm_page_free(page, order);
1306 if (unlikely(PageHWPoison(page)) && !order) {
1308 * Do not let hwpoison pages hit pcplists/buddy
1309 * Untie memcg state and reset page's owner
1311 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1312 __memcg_kmem_uncharge_page(page, order);
1313 reset_page_owner(page, order);
1314 page_table_check_free(page, order);
1319 * Check tail pages before head page information is cleared to
1320 * avoid checking PageCompound for order-0 pages.
1322 if (unlikely(order)) {
1323 bool compound = PageCompound(page);
1326 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1329 ClearPageDoubleMap(page);
1330 ClearPageHasHWPoisoned(page);
1332 for (i = 1; i < (1 << order); i++) {
1334 bad += free_tail_pages_check(page, page + i);
1335 if (unlikely(check_free_page(page + i))) {
1339 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1342 if (PageMappingFlags(page))
1343 page->mapping = NULL;
1344 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1345 __memcg_kmem_uncharge_page(page, order);
1347 bad += check_free_page(page);
1351 page_cpupid_reset_last(page);
1352 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1353 reset_page_owner(page, order);
1354 page_table_check_free(page, order);
1356 if (!PageHighMem(page)) {
1357 debug_check_no_locks_freed(page_address(page),
1358 PAGE_SIZE << order);
1359 debug_check_no_obj_freed(page_address(page),
1360 PAGE_SIZE << order);
1363 kernel_poison_pages(page, 1 << order);
1366 * As memory initialization might be integrated into KASAN,
1367 * KASAN poisoning and memory initialization code must be
1368 * kept together to avoid discrepancies in behavior.
1370 * With hardware tag-based KASAN, memory tags must be set before the
1371 * page becomes unavailable via debug_pagealloc or arch_free_page.
1373 if (!should_skip_kasan_poison(page, fpi_flags)) {
1374 kasan_poison_pages(page, order, init);
1376 /* Memory is already initialized if KASAN did it internally. */
1377 if (kasan_has_integrated_init())
1381 kernel_init_free_pages(page, 1 << order);
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 */
1435 * Frees a number of pages from the PCP lists
1436 * Assumes all pages on list are in same zone.
1437 * count is the number of pages to free.
1439 static void free_pcppages_bulk(struct zone *zone, int count,
1440 struct per_cpu_pages *pcp,
1444 int max_pindex = NR_PCP_LISTS - 1;
1446 bool isolated_pageblocks;
1450 * Ensure proper count is passed which otherwise would stuck in the
1451 * below while (list_empty(list)) loop.
1453 count = min(pcp->count, count);
1455 /* Ensure requested pindex is drained first. */
1456 pindex = pindex - 1;
1459 * local_lock_irq held so equivalent to spin_lock_irqsave for
1460 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1462 spin_lock(&zone->lock);
1463 isolated_pageblocks = has_isolate_pageblock(zone);
1466 struct list_head *list;
1469 /* Remove pages from lists in a round-robin fashion. */
1471 if (++pindex > max_pindex)
1472 pindex = min_pindex;
1473 list = &pcp->lists[pindex];
1474 if (!list_empty(list))
1477 if (pindex == max_pindex)
1479 if (pindex == min_pindex)
1483 order = pindex_to_order(pindex);
1484 nr_pages = 1 << order;
1485 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1489 page = list_last_entry(list, struct page, lru);
1490 mt = get_pcppage_migratetype(page);
1492 /* must delete to avoid corrupting pcp list */
1493 list_del(&page->lru);
1495 pcp->count -= nr_pages;
1497 if (bulkfree_pcp_prepare(page))
1500 /* MIGRATE_ISOLATE page should not go to pcplists */
1501 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1502 /* Pageblock could have been isolated meanwhile */
1503 if (unlikely(isolated_pageblocks))
1504 mt = get_pageblock_migratetype(page);
1506 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1507 trace_mm_page_pcpu_drain(page, order, mt);
1508 } while (count > 0 && !list_empty(list));
1511 spin_unlock(&zone->lock);
1514 static void free_one_page(struct zone *zone,
1515 struct page *page, unsigned long pfn,
1517 int migratetype, fpi_t fpi_flags)
1519 unsigned long flags;
1521 spin_lock_irqsave(&zone->lock, flags);
1522 if (unlikely(has_isolate_pageblock(zone) ||
1523 is_migrate_isolate(migratetype))) {
1524 migratetype = get_pfnblock_migratetype(page, pfn);
1526 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1527 spin_unlock_irqrestore(&zone->lock, flags);
1530 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1531 unsigned long zone, int nid)
1533 mm_zero_struct_page(page);
1534 set_page_links(page, zone, nid, pfn);
1535 init_page_count(page);
1536 page_mapcount_reset(page);
1537 page_cpupid_reset_last(page);
1538 page_kasan_tag_reset(page);
1540 INIT_LIST_HEAD(&page->lru);
1541 #ifdef WANT_PAGE_VIRTUAL
1542 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1543 if (!is_highmem_idx(zone))
1544 set_page_address(page, __va(pfn << PAGE_SHIFT));
1548 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1549 static void __meminit init_reserved_page(unsigned long pfn)
1554 if (!early_page_uninitialised(pfn))
1557 nid = early_pfn_to_nid(pfn);
1558 pgdat = NODE_DATA(nid);
1560 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1561 struct zone *zone = &pgdat->node_zones[zid];
1563 if (zone_spans_pfn(zone, pfn))
1566 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1569 static inline void init_reserved_page(unsigned long pfn)
1572 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1575 * Initialised pages do not have PageReserved set. This function is
1576 * called for each range allocated by the bootmem allocator and
1577 * marks the pages PageReserved. The remaining valid pages are later
1578 * sent to the buddy page allocator.
1580 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1582 unsigned long start_pfn = PFN_DOWN(start);
1583 unsigned long end_pfn = PFN_UP(end);
1585 for (; start_pfn < end_pfn; start_pfn++) {
1586 if (pfn_valid(start_pfn)) {
1587 struct page *page = pfn_to_page(start_pfn);
1589 init_reserved_page(start_pfn);
1591 /* Avoid false-positive PageTail() */
1592 INIT_LIST_HEAD(&page->lru);
1595 * no need for atomic set_bit because the struct
1596 * page is not visible yet so nobody should
1599 __SetPageReserved(page);
1604 static void __free_pages_ok(struct page *page, unsigned int order,
1607 unsigned long flags;
1609 unsigned long pfn = page_to_pfn(page);
1610 struct zone *zone = page_zone(page);
1612 if (!free_pages_prepare(page, order, true, fpi_flags))
1615 migratetype = get_pfnblock_migratetype(page, pfn);
1617 spin_lock_irqsave(&zone->lock, flags);
1618 if (unlikely(has_isolate_pageblock(zone) ||
1619 is_migrate_isolate(migratetype))) {
1620 migratetype = get_pfnblock_migratetype(page, pfn);
1622 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1623 spin_unlock_irqrestore(&zone->lock, flags);
1625 __count_vm_events(PGFREE, 1 << order);
1628 void __free_pages_core(struct page *page, unsigned int order)
1630 unsigned int nr_pages = 1 << order;
1631 struct page *p = page;
1635 * When initializing the memmap, __init_single_page() sets the refcount
1636 * of all pages to 1 ("allocated"/"not free"). We have to set the
1637 * refcount of all involved pages to 0.
1640 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1642 __ClearPageReserved(p);
1643 set_page_count(p, 0);
1645 __ClearPageReserved(p);
1646 set_page_count(p, 0);
1648 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1651 * Bypass PCP and place fresh pages right to the tail, primarily
1652 * relevant for memory onlining.
1654 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1660 * During memory init memblocks map pfns to nids. The search is expensive and
1661 * this caches recent lookups. The implementation of __early_pfn_to_nid
1662 * treats start/end as pfns.
1664 struct mminit_pfnnid_cache {
1665 unsigned long last_start;
1666 unsigned long last_end;
1670 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1673 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1675 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1676 struct mminit_pfnnid_cache *state)
1678 unsigned long start_pfn, end_pfn;
1681 if (state->last_start <= pfn && pfn < state->last_end)
1682 return state->last_nid;
1684 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1685 if (nid != NUMA_NO_NODE) {
1686 state->last_start = start_pfn;
1687 state->last_end = end_pfn;
1688 state->last_nid = nid;
1694 int __meminit early_pfn_to_nid(unsigned long pfn)
1696 static DEFINE_SPINLOCK(early_pfn_lock);
1699 spin_lock(&early_pfn_lock);
1700 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1702 nid = first_online_node;
1703 spin_unlock(&early_pfn_lock);
1707 #endif /* CONFIG_NUMA */
1709 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1712 if (early_page_uninitialised(pfn))
1714 __free_pages_core(page, order);
1718 * Check that the whole (or subset of) a pageblock given by the interval of
1719 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1720 * with the migration of free compaction scanner.
1722 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1724 * It's possible on some configurations to have a setup like node0 node1 node0
1725 * i.e. it's possible that all pages within a zones range of pages do not
1726 * belong to a single zone. We assume that a border between node0 and node1
1727 * can occur within a single pageblock, but not a node0 node1 node0
1728 * interleaving within a single pageblock. It is therefore sufficient to check
1729 * the first and last page of a pageblock and avoid checking each individual
1730 * page in a pageblock.
1732 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1733 unsigned long end_pfn, struct zone *zone)
1735 struct page *start_page;
1736 struct page *end_page;
1738 /* end_pfn is one past the range we are checking */
1741 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1744 start_page = pfn_to_online_page(start_pfn);
1748 if (page_zone(start_page) != zone)
1751 end_page = pfn_to_page(end_pfn);
1753 /* This gives a shorter code than deriving page_zone(end_page) */
1754 if (page_zone_id(start_page) != page_zone_id(end_page))
1760 void set_zone_contiguous(struct zone *zone)
1762 unsigned long block_start_pfn = zone->zone_start_pfn;
1763 unsigned long block_end_pfn;
1765 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1766 for (; block_start_pfn < zone_end_pfn(zone);
1767 block_start_pfn = block_end_pfn,
1768 block_end_pfn += pageblock_nr_pages) {
1770 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1772 if (!__pageblock_pfn_to_page(block_start_pfn,
1773 block_end_pfn, zone))
1778 /* We confirm that there is no hole */
1779 zone->contiguous = true;
1782 void clear_zone_contiguous(struct zone *zone)
1784 zone->contiguous = false;
1787 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1788 static void __init deferred_free_range(unsigned long pfn,
1789 unsigned long nr_pages)
1797 page = pfn_to_page(pfn);
1799 /* Free a large naturally-aligned chunk if possible */
1800 if (nr_pages == pageblock_nr_pages &&
1801 (pfn & (pageblock_nr_pages - 1)) == 0) {
1802 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1803 __free_pages_core(page, pageblock_order);
1807 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1808 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1809 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1810 __free_pages_core(page, 0);
1814 /* Completion tracking for deferred_init_memmap() threads */
1815 static atomic_t pgdat_init_n_undone __initdata;
1816 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1818 static inline void __init pgdat_init_report_one_done(void)
1820 if (atomic_dec_and_test(&pgdat_init_n_undone))
1821 complete(&pgdat_init_all_done_comp);
1825 * Returns true if page needs to be initialized or freed to buddy allocator.
1827 * First we check if pfn is valid on architectures where it is possible to have
1828 * holes within pageblock_nr_pages. On systems where it is not possible, this
1829 * function is optimized out.
1831 * Then, we check if a current large page is valid by only checking the validity
1834 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1836 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1842 * Free pages to buddy allocator. Try to free aligned pages in
1843 * pageblock_nr_pages sizes.
1845 static void __init deferred_free_pages(unsigned long pfn,
1846 unsigned long end_pfn)
1848 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1849 unsigned long nr_free = 0;
1851 for (; pfn < end_pfn; pfn++) {
1852 if (!deferred_pfn_valid(pfn)) {
1853 deferred_free_range(pfn - nr_free, nr_free);
1855 } else if (!(pfn & nr_pgmask)) {
1856 deferred_free_range(pfn - nr_free, nr_free);
1862 /* Free the last block of pages to allocator */
1863 deferred_free_range(pfn - nr_free, nr_free);
1867 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1868 * by performing it only once every pageblock_nr_pages.
1869 * Return number of pages initialized.
1871 static unsigned long __init deferred_init_pages(struct zone *zone,
1873 unsigned long end_pfn)
1875 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1876 int nid = zone_to_nid(zone);
1877 unsigned long nr_pages = 0;
1878 int zid = zone_idx(zone);
1879 struct page *page = NULL;
1881 for (; pfn < end_pfn; pfn++) {
1882 if (!deferred_pfn_valid(pfn)) {
1885 } else if (!page || !(pfn & nr_pgmask)) {
1886 page = pfn_to_page(pfn);
1890 __init_single_page(page, pfn, zid, nid);
1897 * This function is meant to pre-load the iterator for the zone init.
1898 * Specifically it walks through the ranges until we are caught up to the
1899 * first_init_pfn value and exits there. If we never encounter the value we
1900 * return false indicating there are no valid ranges left.
1903 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1904 unsigned long *spfn, unsigned long *epfn,
1905 unsigned long first_init_pfn)
1910 * Start out by walking through the ranges in this zone that have
1911 * already been initialized. We don't need to do anything with them
1912 * so we just need to flush them out of the system.
1914 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1915 if (*epfn <= first_init_pfn)
1917 if (*spfn < first_init_pfn)
1918 *spfn = first_init_pfn;
1927 * Initialize and free pages. We do it in two loops: first we initialize
1928 * struct page, then free to buddy allocator, because while we are
1929 * freeing pages we can access pages that are ahead (computing buddy
1930 * page in __free_one_page()).
1932 * In order to try and keep some memory in the cache we have the loop
1933 * broken along max page order boundaries. This way we will not cause
1934 * any issues with the buddy page computation.
1936 static unsigned long __init
1937 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1938 unsigned long *end_pfn)
1940 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1941 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1942 unsigned long nr_pages = 0;
1945 /* First we loop through and initialize the page values */
1946 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1949 if (mo_pfn <= *start_pfn)
1952 t = min(mo_pfn, *end_pfn);
1953 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1955 if (mo_pfn < *end_pfn) {
1956 *start_pfn = mo_pfn;
1961 /* Reset values and now loop through freeing pages as needed */
1964 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1970 t = min(mo_pfn, epfn);
1971 deferred_free_pages(spfn, t);
1981 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1984 unsigned long spfn, epfn;
1985 struct zone *zone = arg;
1988 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1991 * Initialize and free pages in MAX_ORDER sized increments so that we
1992 * can avoid introducing any issues with the buddy allocator.
1994 while (spfn < end_pfn) {
1995 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2000 /* An arch may override for more concurrency. */
2002 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2007 /* Initialise remaining memory on a node */
2008 static int __init deferred_init_memmap(void *data)
2010 pg_data_t *pgdat = data;
2011 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2012 unsigned long spfn = 0, epfn = 0;
2013 unsigned long first_init_pfn, flags;
2014 unsigned long start = jiffies;
2016 int zid, max_threads;
2019 /* Bind memory initialisation thread to a local node if possible */
2020 if (!cpumask_empty(cpumask))
2021 set_cpus_allowed_ptr(current, cpumask);
2023 pgdat_resize_lock(pgdat, &flags);
2024 first_init_pfn = pgdat->first_deferred_pfn;
2025 if (first_init_pfn == ULONG_MAX) {
2026 pgdat_resize_unlock(pgdat, &flags);
2027 pgdat_init_report_one_done();
2031 /* Sanity check boundaries */
2032 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2033 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2034 pgdat->first_deferred_pfn = ULONG_MAX;
2037 * Once we unlock here, the zone cannot be grown anymore, thus if an
2038 * interrupt thread must allocate this early in boot, zone must be
2039 * pre-grown prior to start of deferred page initialization.
2041 pgdat_resize_unlock(pgdat, &flags);
2043 /* Only the highest zone is deferred so find it */
2044 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2045 zone = pgdat->node_zones + zid;
2046 if (first_init_pfn < zone_end_pfn(zone))
2050 /* If the zone is empty somebody else may have cleared out the zone */
2051 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2055 max_threads = deferred_page_init_max_threads(cpumask);
2057 while (spfn < epfn) {
2058 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2059 struct padata_mt_job job = {
2060 .thread_fn = deferred_init_memmap_chunk,
2063 .size = epfn_align - spfn,
2064 .align = PAGES_PER_SECTION,
2065 .min_chunk = PAGES_PER_SECTION,
2066 .max_threads = max_threads,
2069 padata_do_multithreaded(&job);
2070 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2074 /* Sanity check that the next zone really is unpopulated */
2075 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2077 pr_info("node %d deferred pages initialised in %ums\n",
2078 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2080 pgdat_init_report_one_done();
2085 * If this zone has deferred pages, try to grow it by initializing enough
2086 * deferred pages to satisfy the allocation specified by order, rounded up to
2087 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2088 * of SECTION_SIZE bytes by initializing struct pages in increments of
2089 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2091 * Return true when zone was grown, otherwise return false. We return true even
2092 * when we grow less than requested, to let the caller decide if there are
2093 * enough pages to satisfy the allocation.
2095 * Note: We use noinline because this function is needed only during boot, and
2096 * it is called from a __ref function _deferred_grow_zone. This way we are
2097 * making sure that it is not inlined into permanent text section.
2099 static noinline bool __init
2100 deferred_grow_zone(struct zone *zone, unsigned int order)
2102 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2103 pg_data_t *pgdat = zone->zone_pgdat;
2104 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2105 unsigned long spfn, epfn, flags;
2106 unsigned long nr_pages = 0;
2109 /* Only the last zone may have deferred pages */
2110 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2113 pgdat_resize_lock(pgdat, &flags);
2116 * If someone grew this zone while we were waiting for spinlock, return
2117 * true, as there might be enough pages already.
2119 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2120 pgdat_resize_unlock(pgdat, &flags);
2124 /* If the zone is empty somebody else may have cleared out the zone */
2125 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2126 first_deferred_pfn)) {
2127 pgdat->first_deferred_pfn = ULONG_MAX;
2128 pgdat_resize_unlock(pgdat, &flags);
2129 /* Retry only once. */
2130 return first_deferred_pfn != ULONG_MAX;
2134 * Initialize and free pages in MAX_ORDER sized increments so
2135 * that we can avoid introducing any issues with the buddy
2138 while (spfn < epfn) {
2139 /* update our first deferred PFN for this section */
2140 first_deferred_pfn = spfn;
2142 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2143 touch_nmi_watchdog();
2145 /* We should only stop along section boundaries */
2146 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2149 /* If our quota has been met we can stop here */
2150 if (nr_pages >= nr_pages_needed)
2154 pgdat->first_deferred_pfn = spfn;
2155 pgdat_resize_unlock(pgdat, &flags);
2157 return nr_pages > 0;
2161 * deferred_grow_zone() is __init, but it is called from
2162 * get_page_from_freelist() during early boot until deferred_pages permanently
2163 * disables this call. This is why we have refdata wrapper to avoid warning,
2164 * and to ensure that the function body gets unloaded.
2167 _deferred_grow_zone(struct zone *zone, unsigned int order)
2169 return deferred_grow_zone(zone, order);
2172 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2174 void __init page_alloc_init_late(void)
2179 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2181 /* There will be num_node_state(N_MEMORY) threads */
2182 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2183 for_each_node_state(nid, N_MEMORY) {
2184 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2187 /* Block until all are initialised */
2188 wait_for_completion(&pgdat_init_all_done_comp);
2191 * We initialized the rest of the deferred pages. Permanently disable
2192 * on-demand struct page initialization.
2194 static_branch_disable(&deferred_pages);
2196 /* Reinit limits that are based on free pages after the kernel is up */
2197 files_maxfiles_init();
2202 /* Discard memblock private memory */
2205 for_each_node_state(nid, N_MEMORY)
2206 shuffle_free_memory(NODE_DATA(nid));
2208 for_each_populated_zone(zone)
2209 set_zone_contiguous(zone);
2213 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2214 void __init init_cma_reserved_pageblock(struct page *page)
2216 unsigned i = pageblock_nr_pages;
2217 struct page *p = page;
2220 __ClearPageReserved(p);
2221 set_page_count(p, 0);
2224 set_pageblock_migratetype(page, MIGRATE_CMA);
2225 set_page_refcounted(page);
2226 __free_pages(page, pageblock_order);
2228 adjust_managed_page_count(page, pageblock_nr_pages);
2229 page_zone(page)->cma_pages += pageblock_nr_pages;
2234 * The order of subdivision here is critical for the IO subsystem.
2235 * Please do not alter this order without good reasons and regression
2236 * testing. Specifically, as large blocks of memory are subdivided,
2237 * the order in which smaller blocks are delivered depends on the order
2238 * they're subdivided in this function. This is the primary factor
2239 * influencing the order in which pages are delivered to the IO
2240 * subsystem according to empirical testing, and this is also justified
2241 * by considering the behavior of a buddy system containing a single
2242 * large block of memory acted on by a series of small allocations.
2243 * This behavior is a critical factor in sglist merging's success.
2247 static inline void expand(struct zone *zone, struct page *page,
2248 int low, int high, int migratetype)
2250 unsigned long size = 1 << high;
2252 while (high > low) {
2255 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2258 * Mark as guard pages (or page), that will allow to
2259 * merge back to allocator when buddy will be freed.
2260 * Corresponding page table entries will not be touched,
2261 * pages will stay not present in virtual address space
2263 if (set_page_guard(zone, &page[size], high, migratetype))
2266 add_to_free_list(&page[size], zone, high, migratetype);
2267 set_buddy_order(&page[size], high);
2271 static void check_new_page_bad(struct page *page)
2273 if (unlikely(page->flags & __PG_HWPOISON)) {
2274 /* Don't complain about hwpoisoned pages */
2275 page_mapcount_reset(page); /* remove PageBuddy */
2280 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2284 * This page is about to be returned from the page allocator
2286 static inline int check_new_page(struct page *page)
2288 if (likely(page_expected_state(page,
2289 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2292 check_new_page_bad(page);
2296 static bool check_new_pages(struct page *page, unsigned int order)
2299 for (i = 0; i < (1 << order); i++) {
2300 struct page *p = page + i;
2302 if (unlikely(check_new_page(p)))
2309 #ifdef CONFIG_DEBUG_VM
2311 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2312 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2313 * also checked when pcp lists are refilled from the free lists.
2315 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2317 if (debug_pagealloc_enabled_static())
2318 return check_new_pages(page, order);
2323 static inline bool check_new_pcp(struct page *page, unsigned int order)
2325 return check_new_pages(page, order);
2329 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2330 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2331 * enabled, they are also checked when being allocated from the pcp lists.
2333 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2335 return check_new_pages(page, order);
2337 static inline bool check_new_pcp(struct page *page, unsigned int order)
2339 if (debug_pagealloc_enabled_static())
2340 return check_new_pages(page, order);
2344 #endif /* CONFIG_DEBUG_VM */
2346 inline void post_alloc_hook(struct page *page, unsigned int order,
2349 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2350 bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
2352 set_page_private(page, 0);
2353 set_page_refcounted(page);
2355 arch_alloc_page(page, order);
2356 debug_pagealloc_map_pages(page, 1 << order);
2359 * Page unpoisoning must happen before memory initialization.
2360 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2361 * allocations and the page unpoisoning code will complain.
2363 kernel_unpoison_pages(page, 1 << order);
2366 * As memory initialization might be integrated into KASAN,
2367 * KASAN unpoisoning and memory initializion code must be
2368 * kept together to avoid discrepancies in behavior.
2372 * If memory tags should be zeroed (which happens only when memory
2373 * should be initialized as well).
2378 /* Initialize both memory and tags. */
2379 for (i = 0; i != 1 << order; ++i)
2380 tag_clear_highpage(page + i);
2382 /* Note that memory is already initialized by the loop above. */
2386 * If either a software KASAN mode is enabled, or,
2387 * in the case of hardware tag-based KASAN,
2388 * if memory tags have not been cleared via tag_clear_highpage().
2390 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
2391 IS_ENABLED(CONFIG_KASAN_SW_TAGS) ||
2392 kasan_hw_tags_enabled() && !init_tags) {
2393 /* Mark shadow memory or set memory tags. */
2394 kasan_unpoison_pages(page, order, init);
2396 /* Note that memory is already initialized by KASAN. */
2397 if (kasan_has_integrated_init())
2400 /* If memory is still not initialized, do it now. */
2402 kernel_init_free_pages(page, 1 << order);
2403 /* Propagate __GFP_SKIP_KASAN_POISON to page flags. */
2404 if (kasan_hw_tags_enabled() && (gfp_flags & __GFP_SKIP_KASAN_POISON))
2405 SetPageSkipKASanPoison(page);
2407 set_page_owner(page, order, gfp_flags);
2408 page_table_check_alloc(page, order);
2411 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2412 unsigned int alloc_flags)
2414 post_alloc_hook(page, order, gfp_flags);
2416 if (order && (gfp_flags & __GFP_COMP))
2417 prep_compound_page(page, order);
2420 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2421 * allocate the page. The expectation is that the caller is taking
2422 * steps that will free more memory. The caller should avoid the page
2423 * being used for !PFMEMALLOC purposes.
2425 if (alloc_flags & ALLOC_NO_WATERMARKS)
2426 set_page_pfmemalloc(page);
2428 clear_page_pfmemalloc(page);
2432 * Go through the free lists for the given migratetype and remove
2433 * the smallest available page from the freelists
2435 static __always_inline
2436 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2439 unsigned int current_order;
2440 struct free_area *area;
2443 /* Find a page of the appropriate size in the preferred list */
2444 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2445 area = &(zone->free_area[current_order]);
2446 page = get_page_from_free_area(area, migratetype);
2449 del_page_from_free_list(page, zone, current_order);
2450 expand(zone, page, order, current_order, migratetype);
2451 set_pcppage_migratetype(page, migratetype);
2460 * This array describes the order lists are fallen back to when
2461 * the free lists for the desirable migrate type are depleted
2463 * The other migratetypes do not have fallbacks.
2465 static int fallbacks[MIGRATE_TYPES][3] = {
2466 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2467 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2468 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2472 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2475 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2478 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2479 unsigned int order) { return NULL; }
2483 * Move the free pages in a range to the freelist tail of the requested type.
2484 * Note that start_page and end_pages are not aligned on a pageblock
2485 * boundary. If alignment is required, use move_freepages_block()
2487 static int move_freepages(struct zone *zone,
2488 unsigned long start_pfn, unsigned long end_pfn,
2489 int migratetype, int *num_movable)
2494 int pages_moved = 0;
2496 for (pfn = start_pfn; pfn <= end_pfn;) {
2497 page = pfn_to_page(pfn);
2498 if (!PageBuddy(page)) {
2500 * We assume that pages that could be isolated for
2501 * migration are movable. But we don't actually try
2502 * isolating, as that would be expensive.
2505 (PageLRU(page) || __PageMovable(page)))
2511 /* Make sure we are not inadvertently changing nodes */
2512 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2513 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2515 order = buddy_order(page);
2516 move_to_free_list(page, zone, order, migratetype);
2518 pages_moved += 1 << order;
2524 int move_freepages_block(struct zone *zone, struct page *page,
2525 int migratetype, int *num_movable)
2527 unsigned long start_pfn, end_pfn, pfn;
2532 pfn = page_to_pfn(page);
2533 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2534 end_pfn = start_pfn + pageblock_nr_pages - 1;
2536 /* Do not cross zone boundaries */
2537 if (!zone_spans_pfn(zone, start_pfn))
2539 if (!zone_spans_pfn(zone, end_pfn))
2542 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2546 static void change_pageblock_range(struct page *pageblock_page,
2547 int start_order, int migratetype)
2549 int nr_pageblocks = 1 << (start_order - pageblock_order);
2551 while (nr_pageblocks--) {
2552 set_pageblock_migratetype(pageblock_page, migratetype);
2553 pageblock_page += pageblock_nr_pages;
2558 * When we are falling back to another migratetype during allocation, try to
2559 * steal extra free pages from the same pageblocks to satisfy further
2560 * allocations, instead of polluting multiple pageblocks.
2562 * If we are stealing a relatively large buddy page, it is likely there will
2563 * be more free pages in the pageblock, so try to steal them all. For
2564 * reclaimable and unmovable allocations, we steal regardless of page size,
2565 * as fragmentation caused by those allocations polluting movable pageblocks
2566 * is worse than movable allocations stealing from unmovable and reclaimable
2569 static bool can_steal_fallback(unsigned int order, int start_mt)
2572 * Leaving this order check is intended, although there is
2573 * relaxed order check in next check. The reason is that
2574 * we can actually steal whole pageblock if this condition met,
2575 * but, below check doesn't guarantee it and that is just heuristic
2576 * so could be changed anytime.
2578 if (order >= pageblock_order)
2581 if (order >= pageblock_order / 2 ||
2582 start_mt == MIGRATE_RECLAIMABLE ||
2583 start_mt == MIGRATE_UNMOVABLE ||
2584 page_group_by_mobility_disabled)
2590 static inline bool boost_watermark(struct zone *zone)
2592 unsigned long max_boost;
2594 if (!watermark_boost_factor)
2597 * Don't bother in zones that are unlikely to produce results.
2598 * On small machines, including kdump capture kernels running
2599 * in a small area, boosting the watermark can cause an out of
2600 * memory situation immediately.
2602 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2605 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2606 watermark_boost_factor, 10000);
2609 * high watermark may be uninitialised if fragmentation occurs
2610 * very early in boot so do not boost. We do not fall
2611 * through and boost by pageblock_nr_pages as failing
2612 * allocations that early means that reclaim is not going
2613 * to help and it may even be impossible to reclaim the
2614 * boosted watermark resulting in a hang.
2619 max_boost = max(pageblock_nr_pages, max_boost);
2621 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2628 * This function implements actual steal behaviour. If order is large enough,
2629 * we can steal whole pageblock. If not, we first move freepages in this
2630 * pageblock to our migratetype and determine how many already-allocated pages
2631 * are there in the pageblock with a compatible migratetype. If at least half
2632 * of pages are free or compatible, we can change migratetype of the pageblock
2633 * itself, so pages freed in the future will be put on the correct free list.
2635 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2636 unsigned int alloc_flags, int start_type, bool whole_block)
2638 unsigned int current_order = buddy_order(page);
2639 int free_pages, movable_pages, alike_pages;
2642 old_block_type = get_pageblock_migratetype(page);
2645 * This can happen due to races and we want to prevent broken
2646 * highatomic accounting.
2648 if (is_migrate_highatomic(old_block_type))
2651 /* Take ownership for orders >= pageblock_order */
2652 if (current_order >= pageblock_order) {
2653 change_pageblock_range(page, current_order, start_type);
2658 * Boost watermarks to increase reclaim pressure to reduce the
2659 * likelihood of future fallbacks. Wake kswapd now as the node
2660 * may be balanced overall and kswapd will not wake naturally.
2662 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2663 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2665 /* We are not allowed to try stealing from the whole block */
2669 free_pages = move_freepages_block(zone, page, start_type,
2672 * Determine how many pages are compatible with our allocation.
2673 * For movable allocation, it's the number of movable pages which
2674 * we just obtained. For other types it's a bit more tricky.
2676 if (start_type == MIGRATE_MOVABLE) {
2677 alike_pages = movable_pages;
2680 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2681 * to MOVABLE pageblock, consider all non-movable pages as
2682 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2683 * vice versa, be conservative since we can't distinguish the
2684 * exact migratetype of non-movable pages.
2686 if (old_block_type == MIGRATE_MOVABLE)
2687 alike_pages = pageblock_nr_pages
2688 - (free_pages + movable_pages);
2693 /* moving whole block can fail due to zone boundary conditions */
2698 * If a sufficient number of pages in the block are either free or of
2699 * comparable migratability as our allocation, claim the whole block.
2701 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2702 page_group_by_mobility_disabled)
2703 set_pageblock_migratetype(page, start_type);
2708 move_to_free_list(page, zone, current_order, start_type);
2712 * Check whether there is a suitable fallback freepage with requested order.
2713 * If only_stealable is true, this function returns fallback_mt only if
2714 * we can steal other freepages all together. This would help to reduce
2715 * fragmentation due to mixed migratetype pages in one pageblock.
2717 int find_suitable_fallback(struct free_area *area, unsigned int order,
2718 int migratetype, bool only_stealable, bool *can_steal)
2723 if (area->nr_free == 0)
2728 fallback_mt = fallbacks[migratetype][i];
2729 if (fallback_mt == MIGRATE_TYPES)
2732 if (free_area_empty(area, fallback_mt))
2735 if (can_steal_fallback(order, migratetype))
2738 if (!only_stealable)
2749 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2750 * there are no empty page blocks that contain a page with a suitable order
2752 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2753 unsigned int alloc_order)
2756 unsigned long max_managed, flags;
2759 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2760 * Check is race-prone but harmless.
2762 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2763 if (zone->nr_reserved_highatomic >= max_managed)
2766 spin_lock_irqsave(&zone->lock, flags);
2768 /* Recheck the nr_reserved_highatomic limit under the lock */
2769 if (zone->nr_reserved_highatomic >= max_managed)
2773 mt = get_pageblock_migratetype(page);
2774 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2775 if (migratetype_is_mergeable(mt)) {
2776 zone->nr_reserved_highatomic += pageblock_nr_pages;
2777 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2778 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2782 spin_unlock_irqrestore(&zone->lock, flags);
2786 * Used when an allocation is about to fail under memory pressure. This
2787 * potentially hurts the reliability of high-order allocations when under
2788 * intense memory pressure but failed atomic allocations should be easier
2789 * to recover from than an OOM.
2791 * If @force is true, try to unreserve a pageblock even though highatomic
2792 * pageblock is exhausted.
2794 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2797 struct zonelist *zonelist = ac->zonelist;
2798 unsigned long flags;
2805 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2808 * Preserve at least one pageblock unless memory pressure
2811 if (!force && zone->nr_reserved_highatomic <=
2815 spin_lock_irqsave(&zone->lock, flags);
2816 for (order = 0; order < MAX_ORDER; order++) {
2817 struct free_area *area = &(zone->free_area[order]);
2819 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2824 * In page freeing path, migratetype change is racy so
2825 * we can counter several free pages in a pageblock
2826 * in this loop although we changed the pageblock type
2827 * from highatomic to ac->migratetype. So we should
2828 * adjust the count once.
2830 if (is_migrate_highatomic_page(page)) {
2832 * It should never happen but changes to
2833 * locking could inadvertently allow a per-cpu
2834 * drain to add pages to MIGRATE_HIGHATOMIC
2835 * while unreserving so be safe and watch for
2838 zone->nr_reserved_highatomic -= min(
2840 zone->nr_reserved_highatomic);
2844 * Convert to ac->migratetype and avoid the normal
2845 * pageblock stealing heuristics. Minimally, the caller
2846 * is doing the work and needs the pages. More
2847 * importantly, if the block was always converted to
2848 * MIGRATE_UNMOVABLE or another type then the number
2849 * of pageblocks that cannot be completely freed
2852 set_pageblock_migratetype(page, ac->migratetype);
2853 ret = move_freepages_block(zone, page, ac->migratetype,
2856 spin_unlock_irqrestore(&zone->lock, flags);
2860 spin_unlock_irqrestore(&zone->lock, flags);
2867 * Try finding a free buddy page on the fallback list and put it on the free
2868 * list of requested migratetype, possibly along with other pages from the same
2869 * block, depending on fragmentation avoidance heuristics. Returns true if
2870 * fallback was found so that __rmqueue_smallest() can grab it.
2872 * The use of signed ints for order and current_order is a deliberate
2873 * deviation from the rest of this file, to make the for loop
2874 * condition simpler.
2876 static __always_inline bool
2877 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2878 unsigned int alloc_flags)
2880 struct free_area *area;
2882 int min_order = order;
2888 * Do not steal pages from freelists belonging to other pageblocks
2889 * i.e. orders < pageblock_order. If there are no local zones free,
2890 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2892 if (alloc_flags & ALLOC_NOFRAGMENT)
2893 min_order = pageblock_order;
2896 * Find the largest available free page in the other list. This roughly
2897 * approximates finding the pageblock with the most free pages, which
2898 * would be too costly to do exactly.
2900 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2902 area = &(zone->free_area[current_order]);
2903 fallback_mt = find_suitable_fallback(area, current_order,
2904 start_migratetype, false, &can_steal);
2905 if (fallback_mt == -1)
2909 * We cannot steal all free pages from the pageblock and the
2910 * requested migratetype is movable. In that case it's better to
2911 * steal and split the smallest available page instead of the
2912 * largest available page, because even if the next movable
2913 * allocation falls back into a different pageblock than this
2914 * one, it won't cause permanent fragmentation.
2916 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2917 && current_order > order)
2926 for (current_order = order; current_order < MAX_ORDER;
2928 area = &(zone->free_area[current_order]);
2929 fallback_mt = find_suitable_fallback(area, current_order,
2930 start_migratetype, false, &can_steal);
2931 if (fallback_mt != -1)
2936 * This should not happen - we already found a suitable fallback
2937 * when looking for the largest page.
2939 VM_BUG_ON(current_order == MAX_ORDER);
2942 page = get_page_from_free_area(area, fallback_mt);
2944 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2947 trace_mm_page_alloc_extfrag(page, order, current_order,
2948 start_migratetype, fallback_mt);
2955 * Do the hard work of removing an element from the buddy allocator.
2956 * Call me with the zone->lock already held.
2958 static __always_inline struct page *
2959 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2960 unsigned int alloc_flags)
2964 if (IS_ENABLED(CONFIG_CMA)) {
2966 * Balance movable allocations between regular and CMA areas by
2967 * allocating from CMA when over half of the zone's free memory
2968 * is in the CMA area.
2970 if (alloc_flags & ALLOC_CMA &&
2971 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2972 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2973 page = __rmqueue_cma_fallback(zone, order);
2979 page = __rmqueue_smallest(zone, order, migratetype);
2980 if (unlikely(!page)) {
2981 if (alloc_flags & ALLOC_CMA)
2982 page = __rmqueue_cma_fallback(zone, order);
2984 if (!page && __rmqueue_fallback(zone, order, migratetype,
2990 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2995 * Obtain a specified number of elements from the buddy allocator, all under
2996 * a single hold of the lock, for efficiency. Add them to the supplied list.
2997 * Returns the number of new pages which were placed at *list.
2999 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3000 unsigned long count, struct list_head *list,
3001 int migratetype, unsigned int alloc_flags)
3003 int i, allocated = 0;
3006 * local_lock_irq held so equivalent to spin_lock_irqsave for
3007 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3009 spin_lock(&zone->lock);
3010 for (i = 0; i < count; ++i) {
3011 struct page *page = __rmqueue(zone, order, migratetype,
3013 if (unlikely(page == NULL))
3016 if (unlikely(check_pcp_refill(page, order)))
3020 * Split buddy pages returned by expand() are received here in
3021 * physical page order. The page is added to the tail of
3022 * caller's list. From the callers perspective, the linked list
3023 * is ordered by page number under some conditions. This is
3024 * useful for IO devices that can forward direction from the
3025 * head, thus also in the physical page order. This is useful
3026 * for IO devices that can merge IO requests if the physical
3027 * pages are ordered properly.
3029 list_add_tail(&page->lru, list);
3031 if (is_migrate_cma(get_pcppage_migratetype(page)))
3032 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3037 * i pages were removed from the buddy list even if some leak due
3038 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3039 * on i. Do not confuse with 'allocated' which is the number of
3040 * pages added to the pcp list.
3042 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3043 spin_unlock(&zone->lock);
3049 * Called from the vmstat counter updater to drain pagesets of this
3050 * currently executing processor on remote nodes after they have
3053 * Note that this function must be called with the thread pinned to
3054 * a single processor.
3056 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3058 unsigned long flags;
3059 int to_drain, batch;
3061 local_lock_irqsave(&pagesets.lock, flags);
3062 batch = READ_ONCE(pcp->batch);
3063 to_drain = min(pcp->count, batch);
3065 free_pcppages_bulk(zone, to_drain, pcp, 0);
3066 local_unlock_irqrestore(&pagesets.lock, flags);
3071 * Drain pcplists of the indicated processor and zone.
3073 * The processor must either be the current processor and the
3074 * thread pinned to the current processor or a processor that
3077 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3079 unsigned long flags;
3080 struct per_cpu_pages *pcp;
3082 local_lock_irqsave(&pagesets.lock, flags);
3084 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3086 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3088 local_unlock_irqrestore(&pagesets.lock, flags);
3092 * Drain pcplists of all zones on the indicated processor.
3094 * The processor must either be the current processor and the
3095 * thread pinned to the current processor or a processor that
3098 static void drain_pages(unsigned int cpu)
3102 for_each_populated_zone(zone) {
3103 drain_pages_zone(cpu, zone);
3108 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3110 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3111 * the single zone's pages.
3113 void drain_local_pages(struct zone *zone)
3115 int cpu = smp_processor_id();
3118 drain_pages_zone(cpu, zone);
3123 static void drain_local_pages_wq(struct work_struct *work)
3125 struct pcpu_drain *drain;
3127 drain = container_of(work, struct pcpu_drain, work);
3130 * drain_all_pages doesn't use proper cpu hotplug protection so
3131 * we can race with cpu offline when the WQ can move this from
3132 * a cpu pinned worker to an unbound one. We can operate on a different
3133 * cpu which is alright but we also have to make sure to not move to
3137 drain_local_pages(drain->zone);
3142 * The implementation of drain_all_pages(), exposing an extra parameter to
3143 * drain on all cpus.
3145 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3146 * not empty. The check for non-emptiness can however race with a free to
3147 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3148 * that need the guarantee that every CPU has drained can disable the
3149 * optimizing racy check.
3151 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3156 * Allocate in the BSS so we won't require allocation in
3157 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3159 static cpumask_t cpus_with_pcps;
3162 * Make sure nobody triggers this path before mm_percpu_wq is fully
3165 if (WARN_ON_ONCE(!mm_percpu_wq))
3169 * Do not drain if one is already in progress unless it's specific to
3170 * a zone. Such callers are primarily CMA and memory hotplug and need
3171 * the drain to be complete when the call returns.
3173 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3176 mutex_lock(&pcpu_drain_mutex);
3180 * We don't care about racing with CPU hotplug event
3181 * as offline notification will cause the notified
3182 * cpu to drain that CPU pcps and on_each_cpu_mask
3183 * disables preemption as part of its processing
3185 for_each_online_cpu(cpu) {
3186 struct per_cpu_pages *pcp;
3188 bool has_pcps = false;
3190 if (force_all_cpus) {
3192 * The pcp.count check is racy, some callers need a
3193 * guarantee that no cpu is missed.
3197 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3201 for_each_populated_zone(z) {
3202 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3211 cpumask_set_cpu(cpu, &cpus_with_pcps);
3213 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3216 for_each_cpu(cpu, &cpus_with_pcps) {
3217 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3220 INIT_WORK(&drain->work, drain_local_pages_wq);
3221 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3223 for_each_cpu(cpu, &cpus_with_pcps)
3224 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3226 mutex_unlock(&pcpu_drain_mutex);
3230 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3232 * When zone parameter is non-NULL, spill just the single zone's pages.
3234 * Note that this can be extremely slow as the draining happens in a workqueue.
3236 void drain_all_pages(struct zone *zone)
3238 __drain_all_pages(zone, false);
3241 #ifdef CONFIG_HIBERNATION
3244 * Touch the watchdog for every WD_PAGE_COUNT pages.
3246 #define WD_PAGE_COUNT (128*1024)
3248 void mark_free_pages(struct zone *zone)
3250 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3251 unsigned long flags;
3252 unsigned int order, t;
3255 if (zone_is_empty(zone))
3258 spin_lock_irqsave(&zone->lock, flags);
3260 max_zone_pfn = zone_end_pfn(zone);
3261 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3262 if (pfn_valid(pfn)) {
3263 page = pfn_to_page(pfn);
3265 if (!--page_count) {
3266 touch_nmi_watchdog();
3267 page_count = WD_PAGE_COUNT;
3270 if (page_zone(page) != zone)
3273 if (!swsusp_page_is_forbidden(page))
3274 swsusp_unset_page_free(page);
3277 for_each_migratetype_order(order, t) {
3278 list_for_each_entry(page,
3279 &zone->free_area[order].free_list[t], lru) {
3282 pfn = page_to_pfn(page);
3283 for (i = 0; i < (1UL << order); i++) {
3284 if (!--page_count) {
3285 touch_nmi_watchdog();
3286 page_count = WD_PAGE_COUNT;
3288 swsusp_set_page_free(pfn_to_page(pfn + i));
3292 spin_unlock_irqrestore(&zone->lock, flags);
3294 #endif /* CONFIG_PM */
3296 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3301 if (!free_pcp_prepare(page, order))
3304 migratetype = get_pfnblock_migratetype(page, pfn);
3305 set_pcppage_migratetype(page, migratetype);
3309 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3312 int min_nr_free, max_nr_free;
3314 /* Free everything if batch freeing high-order pages. */
3315 if (unlikely(free_high))
3318 /* Check for PCP disabled or boot pageset */
3319 if (unlikely(high < batch))
3322 /* Leave at least pcp->batch pages on the list */
3323 min_nr_free = batch;
3324 max_nr_free = high - batch;
3327 * Double the number of pages freed each time there is subsequent
3328 * freeing of pages without any allocation.
3330 batch <<= pcp->free_factor;
3331 if (batch < max_nr_free)
3333 batch = clamp(batch, min_nr_free, max_nr_free);
3338 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3341 int high = READ_ONCE(pcp->high);
3343 if (unlikely(!high || free_high))
3346 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3350 * If reclaim is active, limit the number of pages that can be
3351 * stored on pcp lists
3353 return min(READ_ONCE(pcp->batch) << 2, high);
3356 static void free_unref_page_commit(struct page *page, int migratetype,
3359 struct zone *zone = page_zone(page);
3360 struct per_cpu_pages *pcp;
3365 __count_vm_event(PGFREE);
3366 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3367 pindex = order_to_pindex(migratetype, order);
3368 list_add(&page->lru, &pcp->lists[pindex]);
3369 pcp->count += 1 << order;
3372 * As high-order pages other than THP's stored on PCP can contribute
3373 * to fragmentation, limit the number stored when PCP is heavily
3374 * freeing without allocation. The remainder after bulk freeing
3375 * stops will be drained from vmstat refresh context.
3377 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3379 high = nr_pcp_high(pcp, zone, free_high);
3380 if (pcp->count >= high) {
3381 int batch = READ_ONCE(pcp->batch);
3383 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3390 void free_unref_page(struct page *page, unsigned int order)
3392 unsigned long flags;
3393 unsigned long pfn = page_to_pfn(page);
3396 if (!free_unref_page_prepare(page, pfn, order))
3400 * We only track unmovable, reclaimable and movable on pcp lists.
3401 * Place ISOLATE pages on the isolated list because they are being
3402 * offlined but treat HIGHATOMIC as movable pages so we can get those
3403 * areas back if necessary. Otherwise, we may have to free
3404 * excessively into the page allocator
3406 migratetype = get_pcppage_migratetype(page);
3407 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3408 if (unlikely(is_migrate_isolate(migratetype))) {
3409 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3412 migratetype = MIGRATE_MOVABLE;
3415 local_lock_irqsave(&pagesets.lock, flags);
3416 free_unref_page_commit(page, migratetype, order);
3417 local_unlock_irqrestore(&pagesets.lock, flags);
3421 * Free a list of 0-order pages
3423 void free_unref_page_list(struct list_head *list)
3425 struct page *page, *next;
3426 unsigned long flags;
3427 int batch_count = 0;
3430 /* Prepare pages for freeing */
3431 list_for_each_entry_safe(page, next, list, lru) {
3432 unsigned long pfn = page_to_pfn(page);
3433 if (!free_unref_page_prepare(page, pfn, 0)) {
3434 list_del(&page->lru);
3439 * Free isolated pages directly to the allocator, see
3440 * comment in free_unref_page.
3442 migratetype = get_pcppage_migratetype(page);
3443 if (unlikely(is_migrate_isolate(migratetype))) {
3444 list_del(&page->lru);
3445 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3450 local_lock_irqsave(&pagesets.lock, flags);
3451 list_for_each_entry_safe(page, next, list, lru) {
3453 * Non-isolated types over MIGRATE_PCPTYPES get added
3454 * to the MIGRATE_MOVABLE pcp list.
3456 migratetype = get_pcppage_migratetype(page);
3457 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3458 migratetype = MIGRATE_MOVABLE;
3460 trace_mm_page_free_batched(page);
3461 free_unref_page_commit(page, migratetype, 0);
3464 * Guard against excessive IRQ disabled times when we get
3465 * a large list of pages to free.
3467 if (++batch_count == SWAP_CLUSTER_MAX) {
3468 local_unlock_irqrestore(&pagesets.lock, flags);
3470 local_lock_irqsave(&pagesets.lock, flags);
3473 local_unlock_irqrestore(&pagesets.lock, flags);
3477 * split_page takes a non-compound higher-order page, and splits it into
3478 * n (1<<order) sub-pages: page[0..n]
3479 * Each sub-page must be freed individually.
3481 * Note: this is probably too low level an operation for use in drivers.
3482 * Please consult with lkml before using this in your driver.
3484 void split_page(struct page *page, unsigned int order)
3488 VM_BUG_ON_PAGE(PageCompound(page), page);
3489 VM_BUG_ON_PAGE(!page_count(page), page);
3491 for (i = 1; i < (1 << order); i++)
3492 set_page_refcounted(page + i);
3493 split_page_owner(page, 1 << order);
3494 split_page_memcg(page, 1 << order);
3496 EXPORT_SYMBOL_GPL(split_page);
3498 int __isolate_free_page(struct page *page, unsigned int order)
3500 unsigned long watermark;
3504 BUG_ON(!PageBuddy(page));
3506 zone = page_zone(page);
3507 mt = get_pageblock_migratetype(page);
3509 if (!is_migrate_isolate(mt)) {
3511 * Obey watermarks as if the page was being allocated. We can
3512 * emulate a high-order watermark check with a raised order-0
3513 * watermark, because we already know our high-order page
3516 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3517 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3520 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3523 /* Remove page from free list */
3525 del_page_from_free_list(page, zone, order);
3528 * Set the pageblock if the isolated page is at least half of a
3531 if (order >= pageblock_order - 1) {
3532 struct page *endpage = page + (1 << order) - 1;
3533 for (; page < endpage; page += pageblock_nr_pages) {
3534 int mt = get_pageblock_migratetype(page);
3536 * Only change normal pageblocks (i.e., they can merge
3539 if (migratetype_is_mergeable(mt))
3540 set_pageblock_migratetype(page,
3546 return 1UL << order;
3550 * __putback_isolated_page - Return a now-isolated page back where we got it
3551 * @page: Page that was isolated
3552 * @order: Order of the isolated page
3553 * @mt: The page's pageblock's migratetype
3555 * This function is meant to return a page pulled from the free lists via
3556 * __isolate_free_page back to the free lists they were pulled from.
3558 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3560 struct zone *zone = page_zone(page);
3562 /* zone lock should be held when this function is called */
3563 lockdep_assert_held(&zone->lock);
3565 /* Return isolated page to tail of freelist. */
3566 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3567 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3571 * Update NUMA hit/miss statistics
3573 * Must be called with interrupts disabled.
3575 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3579 enum numa_stat_item local_stat = NUMA_LOCAL;
3581 /* skip numa counters update if numa stats is disabled */
3582 if (!static_branch_likely(&vm_numa_stat_key))
3585 if (zone_to_nid(z) != numa_node_id())
3586 local_stat = NUMA_OTHER;
3588 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3589 __count_numa_events(z, NUMA_HIT, nr_account);
3591 __count_numa_events(z, NUMA_MISS, nr_account);
3592 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3594 __count_numa_events(z, local_stat, nr_account);
3598 /* Remove page from the per-cpu list, caller must protect the list */
3600 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3602 unsigned int alloc_flags,
3603 struct per_cpu_pages *pcp,
3604 struct list_head *list)
3609 if (list_empty(list)) {
3610 int batch = READ_ONCE(pcp->batch);
3614 * Scale batch relative to order if batch implies
3615 * free pages can be stored on the PCP. Batch can
3616 * be 1 for small zones or for boot pagesets which
3617 * should never store free pages as the pages may
3618 * belong to arbitrary zones.
3621 batch = max(batch >> order, 2);
3622 alloced = rmqueue_bulk(zone, order,
3624 migratetype, alloc_flags);
3626 pcp->count += alloced << order;
3627 if (unlikely(list_empty(list)))
3631 page = list_first_entry(list, struct page, lru);
3632 list_del(&page->lru);
3633 pcp->count -= 1 << order;
3634 } while (check_new_pcp(page, order));
3639 /* Lock and remove page from the per-cpu list */
3640 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3641 struct zone *zone, unsigned int order,
3642 gfp_t gfp_flags, int migratetype,
3643 unsigned int alloc_flags)
3645 struct per_cpu_pages *pcp;
3646 struct list_head *list;
3648 unsigned long flags;
3650 local_lock_irqsave(&pagesets.lock, flags);
3653 * On allocation, reduce the number of pages that are batch freed.
3654 * See nr_pcp_free() where free_factor is increased for subsequent
3657 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3658 pcp->free_factor >>= 1;
3659 list = &pcp->lists[order_to_pindex(migratetype, order)];
3660 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3661 local_unlock_irqrestore(&pagesets.lock, flags);
3663 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3664 zone_statistics(preferred_zone, zone, 1);
3670 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3673 struct page *rmqueue(struct zone *preferred_zone,
3674 struct zone *zone, unsigned int order,
3675 gfp_t gfp_flags, unsigned int alloc_flags,
3678 unsigned long flags;
3681 if (likely(pcp_allowed_order(order))) {
3683 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3684 * we need to skip it when CMA area isn't allowed.
3686 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3687 migratetype != MIGRATE_MOVABLE) {
3688 page = rmqueue_pcplist(preferred_zone, zone, order,
3689 gfp_flags, migratetype, alloc_flags);
3695 * We most definitely don't want callers attempting to
3696 * allocate greater than order-1 page units with __GFP_NOFAIL.
3698 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3702 spin_lock_irqsave(&zone->lock, flags);
3704 * order-0 request can reach here when the pcplist is skipped
3705 * due to non-CMA allocation context. HIGHATOMIC area is
3706 * reserved for high-order atomic allocation, so order-0
3707 * request should skip it.
3709 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3710 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3712 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3715 page = __rmqueue(zone, order, migratetype, alloc_flags);
3719 __mod_zone_freepage_state(zone, -(1 << order),
3720 get_pcppage_migratetype(page));
3721 spin_unlock_irqrestore(&zone->lock, flags);
3722 } while (check_new_pages(page, order));
3724 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3725 zone_statistics(preferred_zone, zone, 1);
3728 /* Separate test+clear to avoid unnecessary atomics */
3729 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3730 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3731 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3734 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3738 spin_unlock_irqrestore(&zone->lock, flags);
3742 #ifdef CONFIG_FAIL_PAGE_ALLOC
3745 struct fault_attr attr;
3747 bool ignore_gfp_highmem;
3748 bool ignore_gfp_reclaim;
3750 } fail_page_alloc = {
3751 .attr = FAULT_ATTR_INITIALIZER,
3752 .ignore_gfp_reclaim = true,
3753 .ignore_gfp_highmem = true,
3757 static int __init setup_fail_page_alloc(char *str)
3759 return setup_fault_attr(&fail_page_alloc.attr, str);
3761 __setup("fail_page_alloc=", setup_fail_page_alloc);
3763 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3765 if (order < fail_page_alloc.min_order)
3767 if (gfp_mask & __GFP_NOFAIL)
3769 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3771 if (fail_page_alloc.ignore_gfp_reclaim &&
3772 (gfp_mask & __GFP_DIRECT_RECLAIM))
3775 return should_fail(&fail_page_alloc.attr, 1 << order);
3778 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3780 static int __init fail_page_alloc_debugfs(void)
3782 umode_t mode = S_IFREG | 0600;
3785 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3786 &fail_page_alloc.attr);
3788 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3789 &fail_page_alloc.ignore_gfp_reclaim);
3790 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3791 &fail_page_alloc.ignore_gfp_highmem);
3792 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3797 late_initcall(fail_page_alloc_debugfs);
3799 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3801 #else /* CONFIG_FAIL_PAGE_ALLOC */
3803 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3808 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3810 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3812 return __should_fail_alloc_page(gfp_mask, order);
3814 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3816 static inline long __zone_watermark_unusable_free(struct zone *z,
3817 unsigned int order, unsigned int alloc_flags)
3819 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3820 long unusable_free = (1 << order) - 1;
3823 * If the caller does not have rights to ALLOC_HARDER then subtract
3824 * the high-atomic reserves. This will over-estimate the size of the
3825 * atomic reserve but it avoids a search.
3827 if (likely(!alloc_harder))
3828 unusable_free += z->nr_reserved_highatomic;
3831 /* If allocation can't use CMA areas don't use free CMA pages */
3832 if (!(alloc_flags & ALLOC_CMA))
3833 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3836 return unusable_free;
3840 * Return true if free base pages are above 'mark'. For high-order checks it
3841 * will return true of the order-0 watermark is reached and there is at least
3842 * one free page of a suitable size. Checking now avoids taking the zone lock
3843 * to check in the allocation paths if no pages are free.
3845 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3846 int highest_zoneidx, unsigned int alloc_flags,
3851 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3853 /* free_pages may go negative - that's OK */
3854 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3856 if (alloc_flags & ALLOC_HIGH)
3859 if (unlikely(alloc_harder)) {
3861 * OOM victims can try even harder than normal ALLOC_HARDER
3862 * users on the grounds that it's definitely going to be in
3863 * the exit path shortly and free memory. Any allocation it
3864 * makes during the free path will be small and short-lived.
3866 if (alloc_flags & ALLOC_OOM)
3873 * Check watermarks for an order-0 allocation request. If these
3874 * are not met, then a high-order request also cannot go ahead
3875 * even if a suitable page happened to be free.
3877 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3880 /* If this is an order-0 request then the watermark is fine */
3884 /* For a high-order request, check at least one suitable page is free */
3885 for (o = order; o < MAX_ORDER; o++) {
3886 struct free_area *area = &z->free_area[o];
3892 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3893 if (!free_area_empty(area, mt))
3898 if ((alloc_flags & ALLOC_CMA) &&
3899 !free_area_empty(area, MIGRATE_CMA)) {
3903 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3909 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3910 int highest_zoneidx, unsigned int alloc_flags)
3912 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3913 zone_page_state(z, NR_FREE_PAGES));
3916 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3917 unsigned long mark, int highest_zoneidx,
3918 unsigned int alloc_flags, gfp_t gfp_mask)
3922 free_pages = zone_page_state(z, NR_FREE_PAGES);
3925 * Fast check for order-0 only. If this fails then the reserves
3926 * need to be calculated.
3931 fast_free = free_pages;
3932 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3933 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3937 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3941 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3942 * when checking the min watermark. The min watermark is the
3943 * point where boosting is ignored so that kswapd is woken up
3944 * when below the low watermark.
3946 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3947 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3948 mark = z->_watermark[WMARK_MIN];
3949 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3950 alloc_flags, free_pages);
3956 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3957 unsigned long mark, int highest_zoneidx)
3959 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3961 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3962 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3964 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3969 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3971 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3973 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3974 node_reclaim_distance;
3976 #else /* CONFIG_NUMA */
3977 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3981 #endif /* CONFIG_NUMA */
3984 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3985 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3986 * premature use of a lower zone may cause lowmem pressure problems that
3987 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3988 * probably too small. It only makes sense to spread allocations to avoid
3989 * fragmentation between the Normal and DMA32 zones.
3991 static inline unsigned int
3992 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3994 unsigned int alloc_flags;
3997 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4000 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4002 #ifdef CONFIG_ZONE_DMA32
4006 if (zone_idx(zone) != ZONE_NORMAL)
4010 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4011 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4012 * on UMA that if Normal is populated then so is DMA32.
4014 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4015 if (nr_online_nodes > 1 && !populated_zone(--zone))
4018 alloc_flags |= ALLOC_NOFRAGMENT;
4019 #endif /* CONFIG_ZONE_DMA32 */
4023 /* Must be called after current_gfp_context() which can change gfp_mask */
4024 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4025 unsigned int alloc_flags)
4028 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4029 alloc_flags |= ALLOC_CMA;
4035 * get_page_from_freelist goes through the zonelist trying to allocate
4038 static struct page *
4039 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4040 const struct alloc_context *ac)
4044 struct pglist_data *last_pgdat_dirty_limit = NULL;
4049 * Scan zonelist, looking for a zone with enough free.
4050 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4052 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4053 z = ac->preferred_zoneref;
4054 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4059 if (cpusets_enabled() &&
4060 (alloc_flags & ALLOC_CPUSET) &&
4061 !__cpuset_zone_allowed(zone, gfp_mask))
4064 * When allocating a page cache page for writing, we
4065 * want to get it from a node that is within its dirty
4066 * limit, such that no single node holds more than its
4067 * proportional share of globally allowed dirty pages.
4068 * The dirty limits take into account the node's
4069 * lowmem reserves and high watermark so that kswapd
4070 * should be able to balance it without having to
4071 * write pages from its LRU list.
4073 * XXX: For now, allow allocations to potentially
4074 * exceed the per-node dirty limit in the slowpath
4075 * (spread_dirty_pages unset) before going into reclaim,
4076 * which is important when on a NUMA setup the allowed
4077 * nodes are together not big enough to reach the
4078 * global limit. The proper fix for these situations
4079 * will require awareness of nodes in the
4080 * dirty-throttling and the flusher threads.
4082 if (ac->spread_dirty_pages) {
4083 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4086 if (!node_dirty_ok(zone->zone_pgdat)) {
4087 last_pgdat_dirty_limit = zone->zone_pgdat;
4092 if (no_fallback && nr_online_nodes > 1 &&
4093 zone != ac->preferred_zoneref->zone) {
4097 * If moving to a remote node, retry but allow
4098 * fragmenting fallbacks. Locality is more important
4099 * than fragmentation avoidance.
4101 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4102 if (zone_to_nid(zone) != local_nid) {
4103 alloc_flags &= ~ALLOC_NOFRAGMENT;
4108 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4109 if (!zone_watermark_fast(zone, order, mark,
4110 ac->highest_zoneidx, alloc_flags,
4114 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4116 * Watermark failed for this zone, but see if we can
4117 * grow this zone if it contains deferred pages.
4119 if (static_branch_unlikely(&deferred_pages)) {
4120 if (_deferred_grow_zone(zone, order))
4124 /* Checked here to keep the fast path fast */
4125 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4126 if (alloc_flags & ALLOC_NO_WATERMARKS)
4129 if (!node_reclaim_enabled() ||
4130 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4133 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4135 case NODE_RECLAIM_NOSCAN:
4138 case NODE_RECLAIM_FULL:
4139 /* scanned but unreclaimable */
4142 /* did we reclaim enough */
4143 if (zone_watermark_ok(zone, order, mark,
4144 ac->highest_zoneidx, alloc_flags))
4152 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4153 gfp_mask, alloc_flags, ac->migratetype);
4155 prep_new_page(page, order, gfp_mask, alloc_flags);
4158 * If this is a high-order atomic allocation then check
4159 * if the pageblock should be reserved for the future
4161 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4162 reserve_highatomic_pageblock(page, zone, order);
4166 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4167 /* Try again if zone has deferred pages */
4168 if (static_branch_unlikely(&deferred_pages)) {
4169 if (_deferred_grow_zone(zone, order))
4177 * It's possible on a UMA machine to get through all zones that are
4178 * fragmented. If avoiding fragmentation, reset and try again.
4181 alloc_flags &= ~ALLOC_NOFRAGMENT;
4188 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4190 unsigned int filter = SHOW_MEM_FILTER_NODES;
4193 * This documents exceptions given to allocations in certain
4194 * contexts that are allowed to allocate outside current's set
4197 if (!(gfp_mask & __GFP_NOMEMALLOC))
4198 if (tsk_is_oom_victim(current) ||
4199 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4200 filter &= ~SHOW_MEM_FILTER_NODES;
4201 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4202 filter &= ~SHOW_MEM_FILTER_NODES;
4204 show_mem(filter, nodemask);
4207 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4209 struct va_format vaf;
4211 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4213 if ((gfp_mask & __GFP_NOWARN) ||
4214 !__ratelimit(&nopage_rs) ||
4215 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4218 va_start(args, fmt);
4221 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4222 current->comm, &vaf, gfp_mask, &gfp_mask,
4223 nodemask_pr_args(nodemask));
4226 cpuset_print_current_mems_allowed();
4229 warn_alloc_show_mem(gfp_mask, nodemask);
4232 static inline struct page *
4233 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4234 unsigned int alloc_flags,
4235 const struct alloc_context *ac)
4239 page = get_page_from_freelist(gfp_mask, order,
4240 alloc_flags|ALLOC_CPUSET, ac);
4242 * fallback to ignore cpuset restriction if our nodes
4246 page = get_page_from_freelist(gfp_mask, order,
4252 static inline struct page *
4253 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4254 const struct alloc_context *ac, unsigned long *did_some_progress)
4256 struct oom_control oc = {
4257 .zonelist = ac->zonelist,
4258 .nodemask = ac->nodemask,
4260 .gfp_mask = gfp_mask,
4265 *did_some_progress = 0;
4268 * Acquire the oom lock. If that fails, somebody else is
4269 * making progress for us.
4271 if (!mutex_trylock(&oom_lock)) {
4272 *did_some_progress = 1;
4273 schedule_timeout_uninterruptible(1);
4278 * Go through the zonelist yet one more time, keep very high watermark
4279 * here, this is only to catch a parallel oom killing, we must fail if
4280 * we're still under heavy pressure. But make sure that this reclaim
4281 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4282 * allocation which will never fail due to oom_lock already held.
4284 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4285 ~__GFP_DIRECT_RECLAIM, order,
4286 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4290 /* Coredumps can quickly deplete all memory reserves */
4291 if (current->flags & PF_DUMPCORE)
4293 /* The OOM killer will not help higher order allocs */
4294 if (order > PAGE_ALLOC_COSTLY_ORDER)
4297 * We have already exhausted all our reclaim opportunities without any
4298 * success so it is time to admit defeat. We will skip the OOM killer
4299 * because it is very likely that the caller has a more reasonable
4300 * fallback than shooting a random task.
4302 * The OOM killer may not free memory on a specific node.
4304 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4306 /* The OOM killer does not needlessly kill tasks for lowmem */
4307 if (ac->highest_zoneidx < ZONE_NORMAL)
4309 if (pm_suspended_storage())
4312 * XXX: GFP_NOFS allocations should rather fail than rely on
4313 * other request to make a forward progress.
4314 * We are in an unfortunate situation where out_of_memory cannot
4315 * do much for this context but let's try it to at least get
4316 * access to memory reserved if the current task is killed (see
4317 * out_of_memory). Once filesystems are ready to handle allocation
4318 * failures more gracefully we should just bail out here.
4321 /* Exhausted what can be done so it's blame time */
4322 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4323 *did_some_progress = 1;
4326 * Help non-failing allocations by giving them access to memory
4329 if (gfp_mask & __GFP_NOFAIL)
4330 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4331 ALLOC_NO_WATERMARKS, ac);
4334 mutex_unlock(&oom_lock);
4339 * Maximum number of compaction retries with a progress before OOM
4340 * killer is consider as the only way to move forward.
4342 #define MAX_COMPACT_RETRIES 16
4344 #ifdef CONFIG_COMPACTION
4345 /* Try memory compaction for high-order allocations before reclaim */
4346 static struct page *
4347 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4348 unsigned int alloc_flags, const struct alloc_context *ac,
4349 enum compact_priority prio, enum compact_result *compact_result)
4351 struct page *page = NULL;
4352 unsigned long pflags;
4353 unsigned int noreclaim_flag;
4358 psi_memstall_enter(&pflags);
4359 delayacct_compact_start();
4360 noreclaim_flag = memalloc_noreclaim_save();
4362 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4365 memalloc_noreclaim_restore(noreclaim_flag);
4366 psi_memstall_leave(&pflags);
4367 delayacct_compact_end();
4369 if (*compact_result == COMPACT_SKIPPED)
4372 * At least in one zone compaction wasn't deferred or skipped, so let's
4373 * count a compaction stall
4375 count_vm_event(COMPACTSTALL);
4377 /* Prep a captured page if available */
4379 prep_new_page(page, order, gfp_mask, alloc_flags);
4381 /* Try get a page from the freelist if available */
4383 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4386 struct zone *zone = page_zone(page);
4388 zone->compact_blockskip_flush = false;
4389 compaction_defer_reset(zone, order, true);
4390 count_vm_event(COMPACTSUCCESS);
4395 * It's bad if compaction run occurs and fails. The most likely reason
4396 * is that pages exist, but not enough to satisfy watermarks.
4398 count_vm_event(COMPACTFAIL);
4406 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4407 enum compact_result compact_result,
4408 enum compact_priority *compact_priority,
4409 int *compaction_retries)
4411 int max_retries = MAX_COMPACT_RETRIES;
4414 int retries = *compaction_retries;
4415 enum compact_priority priority = *compact_priority;
4420 if (fatal_signal_pending(current))
4423 if (compaction_made_progress(compact_result))
4424 (*compaction_retries)++;
4427 * compaction considers all the zone as desperately out of memory
4428 * so it doesn't really make much sense to retry except when the
4429 * failure could be caused by insufficient priority
4431 if (compaction_failed(compact_result))
4432 goto check_priority;
4435 * compaction was skipped because there are not enough order-0 pages
4436 * to work with, so we retry only if it looks like reclaim can help.
4438 if (compaction_needs_reclaim(compact_result)) {
4439 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4444 * make sure the compaction wasn't deferred or didn't bail out early
4445 * due to locks contention before we declare that we should give up.
4446 * But the next retry should use a higher priority if allowed, so
4447 * we don't just keep bailing out endlessly.
4449 if (compaction_withdrawn(compact_result)) {
4450 goto check_priority;
4454 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4455 * costly ones because they are de facto nofail and invoke OOM
4456 * killer to move on while costly can fail and users are ready
4457 * to cope with that. 1/4 retries is rather arbitrary but we
4458 * would need much more detailed feedback from compaction to
4459 * make a better decision.
4461 if (order > PAGE_ALLOC_COSTLY_ORDER)
4463 if (*compaction_retries <= max_retries) {
4469 * Make sure there are attempts at the highest priority if we exhausted
4470 * all retries or failed at the lower priorities.
4473 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4474 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4476 if (*compact_priority > min_priority) {
4477 (*compact_priority)--;
4478 *compaction_retries = 0;
4482 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4486 static inline struct page *
4487 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4488 unsigned int alloc_flags, const struct alloc_context *ac,
4489 enum compact_priority prio, enum compact_result *compact_result)
4491 *compact_result = COMPACT_SKIPPED;
4496 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4497 enum compact_result compact_result,
4498 enum compact_priority *compact_priority,
4499 int *compaction_retries)
4504 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4508 * There are setups with compaction disabled which would prefer to loop
4509 * inside the allocator rather than hit the oom killer prematurely.
4510 * Let's give them a good hope and keep retrying while the order-0
4511 * watermarks are OK.
4513 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4514 ac->highest_zoneidx, ac->nodemask) {
4515 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4516 ac->highest_zoneidx, alloc_flags))
4521 #endif /* CONFIG_COMPACTION */
4523 #ifdef CONFIG_LOCKDEP
4524 static struct lockdep_map __fs_reclaim_map =
4525 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4527 static bool __need_reclaim(gfp_t gfp_mask)
4529 /* no reclaim without waiting on it */
4530 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4533 /* this guy won't enter reclaim */
4534 if (current->flags & PF_MEMALLOC)
4537 if (gfp_mask & __GFP_NOLOCKDEP)
4543 void __fs_reclaim_acquire(unsigned long ip)
4545 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4548 void __fs_reclaim_release(unsigned long ip)
4550 lock_release(&__fs_reclaim_map, ip);
4553 void fs_reclaim_acquire(gfp_t gfp_mask)
4555 gfp_mask = current_gfp_context(gfp_mask);
4557 if (__need_reclaim(gfp_mask)) {
4558 if (gfp_mask & __GFP_FS)
4559 __fs_reclaim_acquire(_RET_IP_);
4561 #ifdef CONFIG_MMU_NOTIFIER
4562 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4563 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4568 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4570 void fs_reclaim_release(gfp_t gfp_mask)
4572 gfp_mask = current_gfp_context(gfp_mask);
4574 if (__need_reclaim(gfp_mask)) {
4575 if (gfp_mask & __GFP_FS)
4576 __fs_reclaim_release(_RET_IP_);
4579 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4582 /* Perform direct synchronous page reclaim */
4583 static unsigned long
4584 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4585 const struct alloc_context *ac)
4587 unsigned int noreclaim_flag;
4588 unsigned long progress;
4592 /* We now go into synchronous reclaim */
4593 cpuset_memory_pressure_bump();
4594 fs_reclaim_acquire(gfp_mask);
4595 noreclaim_flag = memalloc_noreclaim_save();
4597 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4600 memalloc_noreclaim_restore(noreclaim_flag);
4601 fs_reclaim_release(gfp_mask);
4608 /* The really slow allocator path where we enter direct reclaim */
4609 static inline struct page *
4610 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4611 unsigned int alloc_flags, const struct alloc_context *ac,
4612 unsigned long *did_some_progress)
4614 struct page *page = NULL;
4615 unsigned long pflags;
4616 bool drained = false;
4618 psi_memstall_enter(&pflags);
4619 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4620 if (unlikely(!(*did_some_progress)))
4624 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4627 * If an allocation failed after direct reclaim, it could be because
4628 * pages are pinned on the per-cpu lists or in high alloc reserves.
4629 * Shrink them and try again
4631 if (!page && !drained) {
4632 unreserve_highatomic_pageblock(ac, false);
4633 drain_all_pages(NULL);
4638 psi_memstall_leave(&pflags);
4643 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4644 const struct alloc_context *ac)
4648 pg_data_t *last_pgdat = NULL;
4649 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4651 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4653 if (last_pgdat != zone->zone_pgdat)
4654 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4655 last_pgdat = zone->zone_pgdat;
4659 static inline unsigned int
4660 gfp_to_alloc_flags(gfp_t gfp_mask)
4662 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4665 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4666 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4667 * to save two branches.
4669 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4670 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4673 * The caller may dip into page reserves a bit more if the caller
4674 * cannot run direct reclaim, or if the caller has realtime scheduling
4675 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4676 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4678 alloc_flags |= (__force int)
4679 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4681 if (gfp_mask & __GFP_ATOMIC) {
4683 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4684 * if it can't schedule.
4686 if (!(gfp_mask & __GFP_NOMEMALLOC))
4687 alloc_flags |= ALLOC_HARDER;
4689 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4690 * comment for __cpuset_node_allowed().
4692 alloc_flags &= ~ALLOC_CPUSET;
4693 } else if (unlikely(rt_task(current)) && in_task())
4694 alloc_flags |= ALLOC_HARDER;
4696 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4701 static bool oom_reserves_allowed(struct task_struct *tsk)
4703 if (!tsk_is_oom_victim(tsk))
4707 * !MMU doesn't have oom reaper so give access to memory reserves
4708 * only to the thread with TIF_MEMDIE set
4710 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4717 * Distinguish requests which really need access to full memory
4718 * reserves from oom victims which can live with a portion of it
4720 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4722 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4724 if (gfp_mask & __GFP_MEMALLOC)
4725 return ALLOC_NO_WATERMARKS;
4726 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4727 return ALLOC_NO_WATERMARKS;
4728 if (!in_interrupt()) {
4729 if (current->flags & PF_MEMALLOC)
4730 return ALLOC_NO_WATERMARKS;
4731 else if (oom_reserves_allowed(current))
4738 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4740 return !!__gfp_pfmemalloc_flags(gfp_mask);
4744 * Checks whether it makes sense to retry the reclaim to make a forward progress
4745 * for the given allocation request.
4747 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4748 * without success, or when we couldn't even meet the watermark if we
4749 * reclaimed all remaining pages on the LRU lists.
4751 * Returns true if a retry is viable or false to enter the oom path.
4754 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4755 struct alloc_context *ac, int alloc_flags,
4756 bool did_some_progress, int *no_progress_loops)
4763 * Costly allocations might have made a progress but this doesn't mean
4764 * their order will become available due to high fragmentation so
4765 * always increment the no progress counter for them
4767 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4768 *no_progress_loops = 0;
4770 (*no_progress_loops)++;
4773 * Make sure we converge to OOM if we cannot make any progress
4774 * several times in the row.
4776 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4777 /* Before OOM, exhaust highatomic_reserve */
4778 return unreserve_highatomic_pageblock(ac, true);
4782 * Keep reclaiming pages while there is a chance this will lead
4783 * somewhere. If none of the target zones can satisfy our allocation
4784 * request even if all reclaimable pages are considered then we are
4785 * screwed and have to go OOM.
4787 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4788 ac->highest_zoneidx, ac->nodemask) {
4789 unsigned long available;
4790 unsigned long reclaimable;
4791 unsigned long min_wmark = min_wmark_pages(zone);
4794 available = reclaimable = zone_reclaimable_pages(zone);
4795 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4798 * Would the allocation succeed if we reclaimed all
4799 * reclaimable pages?
4801 wmark = __zone_watermark_ok(zone, order, min_wmark,
4802 ac->highest_zoneidx, alloc_flags, available);
4803 trace_reclaim_retry_zone(z, order, reclaimable,
4804 available, min_wmark, *no_progress_loops, wmark);
4812 * Memory allocation/reclaim might be called from a WQ context and the
4813 * current implementation of the WQ concurrency control doesn't
4814 * recognize that a particular WQ is congested if the worker thread is
4815 * looping without ever sleeping. Therefore we have to do a short sleep
4816 * here rather than calling cond_resched().
4818 if (current->flags & PF_WQ_WORKER)
4819 schedule_timeout_uninterruptible(1);
4826 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4829 * It's possible that cpuset's mems_allowed and the nodemask from
4830 * mempolicy don't intersect. This should be normally dealt with by
4831 * policy_nodemask(), but it's possible to race with cpuset update in
4832 * such a way the check therein was true, and then it became false
4833 * before we got our cpuset_mems_cookie here.
4834 * This assumes that for all allocations, ac->nodemask can come only
4835 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4836 * when it does not intersect with the cpuset restrictions) or the
4837 * caller can deal with a violated nodemask.
4839 if (cpusets_enabled() && ac->nodemask &&
4840 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4841 ac->nodemask = NULL;
4846 * When updating a task's mems_allowed or mempolicy nodemask, it is
4847 * possible to race with parallel threads in such a way that our
4848 * allocation can fail while the mask is being updated. If we are about
4849 * to fail, check if the cpuset changed during allocation and if so,
4852 if (read_mems_allowed_retry(cpuset_mems_cookie))
4858 static inline struct page *
4859 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4860 struct alloc_context *ac)
4862 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4863 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4864 struct page *page = NULL;
4865 unsigned int alloc_flags;
4866 unsigned long did_some_progress;
4867 enum compact_priority compact_priority;
4868 enum compact_result compact_result;
4869 int compaction_retries;
4870 int no_progress_loops;
4871 unsigned int cpuset_mems_cookie;
4875 * We also sanity check to catch abuse of atomic reserves being used by
4876 * callers that are not in atomic context.
4878 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4879 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4880 gfp_mask &= ~__GFP_ATOMIC;
4883 compaction_retries = 0;
4884 no_progress_loops = 0;
4885 compact_priority = DEF_COMPACT_PRIORITY;
4886 cpuset_mems_cookie = read_mems_allowed_begin();
4889 * The fast path uses conservative alloc_flags to succeed only until
4890 * kswapd needs to be woken up, and to avoid the cost of setting up
4891 * alloc_flags precisely. So we do that now.
4893 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4896 * We need to recalculate the starting point for the zonelist iterator
4897 * because we might have used different nodemask in the fast path, or
4898 * there was a cpuset modification and we are retrying - otherwise we
4899 * could end up iterating over non-eligible zones endlessly.
4901 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4902 ac->highest_zoneidx, ac->nodemask);
4903 if (!ac->preferred_zoneref->zone)
4907 * Check for insane configurations where the cpuset doesn't contain
4908 * any suitable zone to satisfy the request - e.g. non-movable
4909 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4911 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4912 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4913 ac->highest_zoneidx,
4914 &cpuset_current_mems_allowed);
4919 if (alloc_flags & ALLOC_KSWAPD)
4920 wake_all_kswapds(order, gfp_mask, ac);
4923 * The adjusted alloc_flags might result in immediate success, so try
4926 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4931 * For costly allocations, try direct compaction first, as it's likely
4932 * that we have enough base pages and don't need to reclaim. For non-
4933 * movable high-order allocations, do that as well, as compaction will
4934 * try prevent permanent fragmentation by migrating from blocks of the
4936 * Don't try this for allocations that are allowed to ignore
4937 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4939 if (can_direct_reclaim &&
4941 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4942 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4943 page = __alloc_pages_direct_compact(gfp_mask, order,
4945 INIT_COMPACT_PRIORITY,
4951 * Checks for costly allocations with __GFP_NORETRY, which
4952 * includes some THP page fault allocations
4954 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4956 * If allocating entire pageblock(s) and compaction
4957 * failed because all zones are below low watermarks
4958 * or is prohibited because it recently failed at this
4959 * order, fail immediately unless the allocator has
4960 * requested compaction and reclaim retry.
4963 * - potentially very expensive because zones are far
4964 * below their low watermarks or this is part of very
4965 * bursty high order allocations,
4966 * - not guaranteed to help because isolate_freepages()
4967 * may not iterate over freed pages as part of its
4969 * - unlikely to make entire pageblocks free on its
4972 if (compact_result == COMPACT_SKIPPED ||
4973 compact_result == COMPACT_DEFERRED)
4977 * Looks like reclaim/compaction is worth trying, but
4978 * sync compaction could be very expensive, so keep
4979 * using async compaction.
4981 compact_priority = INIT_COMPACT_PRIORITY;
4986 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4987 if (alloc_flags & ALLOC_KSWAPD)
4988 wake_all_kswapds(order, gfp_mask, ac);
4990 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4992 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
4995 * Reset the nodemask and zonelist iterators if memory policies can be
4996 * ignored. These allocations are high priority and system rather than
4999 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5000 ac->nodemask = NULL;
5001 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5002 ac->highest_zoneidx, ac->nodemask);
5005 /* Attempt with potentially adjusted zonelist and alloc_flags */
5006 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5010 /* Caller is not willing to reclaim, we can't balance anything */
5011 if (!can_direct_reclaim)
5014 /* Avoid recursion of direct reclaim */
5015 if (current->flags & PF_MEMALLOC)
5018 /* Try direct reclaim and then allocating */
5019 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5020 &did_some_progress);
5024 /* Try direct compaction and then allocating */
5025 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5026 compact_priority, &compact_result);
5030 /* Do not loop if specifically requested */
5031 if (gfp_mask & __GFP_NORETRY)
5035 * Do not retry costly high order allocations unless they are
5036 * __GFP_RETRY_MAYFAIL
5038 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5041 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5042 did_some_progress > 0, &no_progress_loops))
5046 * It doesn't make any sense to retry for the compaction if the order-0
5047 * reclaim is not able to make any progress because the current
5048 * implementation of the compaction depends on the sufficient amount
5049 * of free memory (see __compaction_suitable)
5051 if (did_some_progress > 0 &&
5052 should_compact_retry(ac, order, alloc_flags,
5053 compact_result, &compact_priority,
5054 &compaction_retries))
5058 /* Deal with possible cpuset update races before we start OOM killing */
5059 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5062 /* Reclaim has failed us, start killing things */
5063 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5067 /* Avoid allocations with no watermarks from looping endlessly */
5068 if (tsk_is_oom_victim(current) &&
5069 (alloc_flags & ALLOC_OOM ||
5070 (gfp_mask & __GFP_NOMEMALLOC)))
5073 /* Retry as long as the OOM killer is making progress */
5074 if (did_some_progress) {
5075 no_progress_loops = 0;
5080 /* Deal with possible cpuset update races before we fail */
5081 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5085 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5088 if (gfp_mask & __GFP_NOFAIL) {
5090 * All existing users of the __GFP_NOFAIL are blockable, so warn
5091 * of any new users that actually require GFP_NOWAIT
5093 if (WARN_ON_ONCE(!can_direct_reclaim))
5097 * PF_MEMALLOC request from this context is rather bizarre
5098 * because we cannot reclaim anything and only can loop waiting
5099 * for somebody to do a work for us
5101 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5104 * non failing costly orders are a hard requirement which we
5105 * are not prepared for much so let's warn about these users
5106 * so that we can identify them and convert them to something
5109 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5112 * Help non-failing allocations by giving them access to memory
5113 * reserves but do not use ALLOC_NO_WATERMARKS because this
5114 * could deplete whole memory reserves which would just make
5115 * the situation worse
5117 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5125 warn_alloc(gfp_mask, ac->nodemask,
5126 "page allocation failure: order:%u", order);
5131 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5132 int preferred_nid, nodemask_t *nodemask,
5133 struct alloc_context *ac, gfp_t *alloc_gfp,
5134 unsigned int *alloc_flags)
5136 ac->highest_zoneidx = gfp_zone(gfp_mask);
5137 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5138 ac->nodemask = nodemask;
5139 ac->migratetype = gfp_migratetype(gfp_mask);
5141 if (cpusets_enabled()) {
5142 *alloc_gfp |= __GFP_HARDWALL;
5144 * When we are in the interrupt context, it is irrelevant
5145 * to the current task context. It means that any node ok.
5147 if (in_task() && !ac->nodemask)
5148 ac->nodemask = &cpuset_current_mems_allowed;
5150 *alloc_flags |= ALLOC_CPUSET;
5153 fs_reclaim_acquire(gfp_mask);
5154 fs_reclaim_release(gfp_mask);
5156 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5158 if (should_fail_alloc_page(gfp_mask, order))
5161 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5163 /* Dirty zone balancing only done in the fast path */
5164 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5167 * The preferred zone is used for statistics but crucially it is
5168 * also used as the starting point for the zonelist iterator. It
5169 * may get reset for allocations that ignore memory policies.
5171 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5172 ac->highest_zoneidx, ac->nodemask);
5178 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5179 * @gfp: GFP flags for the allocation
5180 * @preferred_nid: The preferred NUMA node ID to allocate from
5181 * @nodemask: Set of nodes to allocate from, may be NULL
5182 * @nr_pages: The number of pages desired on the list or array
5183 * @page_list: Optional list to store the allocated pages
5184 * @page_array: Optional array to store the pages
5186 * This is a batched version of the page allocator that attempts to
5187 * allocate nr_pages quickly. Pages are added to page_list if page_list
5188 * is not NULL, otherwise it is assumed that the page_array is valid.
5190 * For lists, nr_pages is the number of pages that should be allocated.
5192 * For arrays, only NULL elements are populated with pages and nr_pages
5193 * is the maximum number of pages that will be stored in the array.
5195 * Returns the number of pages on the list or array.
5197 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5198 nodemask_t *nodemask, int nr_pages,
5199 struct list_head *page_list,
5200 struct page **page_array)
5203 unsigned long flags;
5206 struct per_cpu_pages *pcp;
5207 struct list_head *pcp_list;
5208 struct alloc_context ac;
5210 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5211 int nr_populated = 0, nr_account = 0;
5214 * Skip populated array elements to determine if any pages need
5215 * to be allocated before disabling IRQs.
5217 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5220 /* No pages requested? */
5221 if (unlikely(nr_pages <= 0))
5224 /* Already populated array? */
5225 if (unlikely(page_array && nr_pages - nr_populated == 0))
5228 /* Bulk allocator does not support memcg accounting. */
5229 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5232 /* Use the single page allocator for one page. */
5233 if (nr_pages - nr_populated == 1)
5236 #ifdef CONFIG_PAGE_OWNER
5238 * PAGE_OWNER may recurse into the allocator to allocate space to
5239 * save the stack with pagesets.lock held. Releasing/reacquiring
5240 * removes much of the performance benefit of bulk allocation so
5241 * force the caller to allocate one page at a time as it'll have
5242 * similar performance to added complexity to the bulk allocator.
5244 if (static_branch_unlikely(&page_owner_inited))
5248 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5249 gfp &= gfp_allowed_mask;
5251 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5255 /* Find an allowed local zone that meets the low watermark. */
5256 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5259 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5260 !__cpuset_zone_allowed(zone, gfp)) {
5264 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5265 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5269 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5270 if (zone_watermark_fast(zone, 0, mark,
5271 zonelist_zone_idx(ac.preferred_zoneref),
5272 alloc_flags, gfp)) {
5278 * If there are no allowed local zones that meets the watermarks then
5279 * try to allocate a single page and reclaim if necessary.
5281 if (unlikely(!zone))
5284 /* Attempt the batch allocation */
5285 local_lock_irqsave(&pagesets.lock, flags);
5286 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5287 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5289 while (nr_populated < nr_pages) {
5291 /* Skip existing pages */
5292 if (page_array && page_array[nr_populated]) {
5297 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5299 if (unlikely(!page)) {
5300 /* Try and get at least one page */
5307 prep_new_page(page, 0, gfp, 0);
5309 list_add(&page->lru, page_list);
5311 page_array[nr_populated] = page;
5315 local_unlock_irqrestore(&pagesets.lock, flags);
5317 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5318 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5321 return nr_populated;
5324 local_unlock_irqrestore(&pagesets.lock, flags);
5327 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5330 list_add(&page->lru, page_list);
5332 page_array[nr_populated] = page;
5338 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5341 * This is the 'heart' of the zoned buddy allocator.
5343 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5344 nodemask_t *nodemask)
5347 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5348 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5349 struct alloc_context ac = { };
5352 * There are several places where we assume that the order value is sane
5353 * so bail out early if the request is out of bound.
5355 if (unlikely(order >= MAX_ORDER)) {
5356 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5360 gfp &= gfp_allowed_mask;
5362 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5363 * resp. GFP_NOIO which has to be inherited for all allocation requests
5364 * from a particular context which has been marked by
5365 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5366 * movable zones are not used during allocation.
5368 gfp = current_gfp_context(gfp);
5370 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5371 &alloc_gfp, &alloc_flags))
5375 * Forbid the first pass from falling back to types that fragment
5376 * memory until all local zones are considered.
5378 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5380 /* First allocation attempt */
5381 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5386 ac.spread_dirty_pages = false;
5389 * Restore the original nodemask if it was potentially replaced with
5390 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5392 ac.nodemask = nodemask;
5394 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5397 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5398 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5399 __free_pages(page, order);
5403 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5407 EXPORT_SYMBOL(__alloc_pages);
5409 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5410 nodemask_t *nodemask)
5412 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5413 preferred_nid, nodemask);
5415 if (page && order > 1)
5416 prep_transhuge_page(page);
5417 return (struct folio *)page;
5419 EXPORT_SYMBOL(__folio_alloc);
5422 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5423 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5424 * you need to access high mem.
5426 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5430 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5433 return (unsigned long) page_address(page);
5435 EXPORT_SYMBOL(__get_free_pages);
5437 unsigned long get_zeroed_page(gfp_t gfp_mask)
5439 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5441 EXPORT_SYMBOL(get_zeroed_page);
5444 * __free_pages - Free pages allocated with alloc_pages().
5445 * @page: The page pointer returned from alloc_pages().
5446 * @order: The order of the allocation.
5448 * This function can free multi-page allocations that are not compound
5449 * pages. It does not check that the @order passed in matches that of
5450 * the allocation, so it is easy to leak memory. Freeing more memory
5451 * than was allocated will probably emit a warning.
5453 * If the last reference to this page is speculative, it will be released
5454 * by put_page() which only frees the first page of a non-compound
5455 * allocation. To prevent the remaining pages from being leaked, we free
5456 * the subsequent pages here. If you want to use the page's reference
5457 * count to decide when to free the allocation, you should allocate a
5458 * compound page, and use put_page() instead of __free_pages().
5460 * Context: May be called in interrupt context or while holding a normal
5461 * spinlock, but not in NMI context or while holding a raw spinlock.
5463 void __free_pages(struct page *page, unsigned int order)
5465 if (put_page_testzero(page))
5466 free_the_page(page, order);
5467 else if (!PageHead(page))
5469 free_the_page(page + (1 << order), order);
5471 EXPORT_SYMBOL(__free_pages);
5473 void free_pages(unsigned long addr, unsigned int order)
5476 VM_BUG_ON(!virt_addr_valid((void *)addr));
5477 __free_pages(virt_to_page((void *)addr), order);
5481 EXPORT_SYMBOL(free_pages);
5485 * An arbitrary-length arbitrary-offset area of memory which resides
5486 * within a 0 or higher order page. Multiple fragments within that page
5487 * are individually refcounted, in the page's reference counter.
5489 * The page_frag functions below provide a simple allocation framework for
5490 * page fragments. This is used by the network stack and network device
5491 * drivers to provide a backing region of memory for use as either an
5492 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5494 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5497 struct page *page = NULL;
5498 gfp_t gfp = gfp_mask;
5500 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5501 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5503 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5504 PAGE_FRAG_CACHE_MAX_ORDER);
5505 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5507 if (unlikely(!page))
5508 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5510 nc->va = page ? page_address(page) : NULL;
5515 void __page_frag_cache_drain(struct page *page, unsigned int count)
5517 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5519 if (page_ref_sub_and_test(page, count))
5520 free_the_page(page, compound_order(page));
5522 EXPORT_SYMBOL(__page_frag_cache_drain);
5524 void *page_frag_alloc_align(struct page_frag_cache *nc,
5525 unsigned int fragsz, gfp_t gfp_mask,
5526 unsigned int align_mask)
5528 unsigned int size = PAGE_SIZE;
5532 if (unlikely(!nc->va)) {
5534 page = __page_frag_cache_refill(nc, gfp_mask);
5538 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5539 /* if size can vary use size else just use PAGE_SIZE */
5542 /* Even if we own the page, we do not use atomic_set().
5543 * This would break get_page_unless_zero() users.
5545 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5547 /* reset page count bias and offset to start of new frag */
5548 nc->pfmemalloc = page_is_pfmemalloc(page);
5549 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5553 offset = nc->offset - fragsz;
5554 if (unlikely(offset < 0)) {
5555 page = virt_to_page(nc->va);
5557 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5560 if (unlikely(nc->pfmemalloc)) {
5561 free_the_page(page, compound_order(page));
5565 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5566 /* if size can vary use size else just use PAGE_SIZE */
5569 /* OK, page count is 0, we can safely set it */
5570 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5572 /* reset page count bias and offset to start of new frag */
5573 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5574 offset = size - fragsz;
5578 offset &= align_mask;
5579 nc->offset = offset;
5581 return nc->va + offset;
5583 EXPORT_SYMBOL(page_frag_alloc_align);
5586 * Frees a page fragment allocated out of either a compound or order 0 page.
5588 void page_frag_free(void *addr)
5590 struct page *page = virt_to_head_page(addr);
5592 if (unlikely(put_page_testzero(page)))
5593 free_the_page(page, compound_order(page));
5595 EXPORT_SYMBOL(page_frag_free);
5597 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5601 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5602 unsigned long used = addr + PAGE_ALIGN(size);
5604 split_page(virt_to_page((void *)addr), order);
5605 while (used < alloc_end) {
5610 return (void *)addr;
5614 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5615 * @size: the number of bytes to allocate
5616 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5618 * This function is similar to alloc_pages(), except that it allocates the
5619 * minimum number of pages to satisfy the request. alloc_pages() can only
5620 * allocate memory in power-of-two pages.
5622 * This function is also limited by MAX_ORDER.
5624 * Memory allocated by this function must be released by free_pages_exact().
5626 * Return: pointer to the allocated area or %NULL in case of error.
5628 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5630 unsigned int order = get_order(size);
5633 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5634 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5636 addr = __get_free_pages(gfp_mask, order);
5637 return make_alloc_exact(addr, order, size);
5639 EXPORT_SYMBOL(alloc_pages_exact);
5642 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5644 * @nid: the preferred node ID where memory should be allocated
5645 * @size: the number of bytes to allocate
5646 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5648 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5651 * Return: pointer to the allocated area or %NULL in case of error.
5653 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5655 unsigned int order = get_order(size);
5658 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5659 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5661 p = alloc_pages_node(nid, gfp_mask, order);
5664 return make_alloc_exact((unsigned long)page_address(p), order, size);
5668 * free_pages_exact - release memory allocated via alloc_pages_exact()
5669 * @virt: the value returned by alloc_pages_exact.
5670 * @size: size of allocation, same value as passed to alloc_pages_exact().
5672 * Release the memory allocated by a previous call to alloc_pages_exact.
5674 void free_pages_exact(void *virt, size_t size)
5676 unsigned long addr = (unsigned long)virt;
5677 unsigned long end = addr + PAGE_ALIGN(size);
5679 while (addr < end) {
5684 EXPORT_SYMBOL(free_pages_exact);
5687 * nr_free_zone_pages - count number of pages beyond high watermark
5688 * @offset: The zone index of the highest zone
5690 * nr_free_zone_pages() counts the number of pages which are beyond the
5691 * high watermark within all zones at or below a given zone index. For each
5692 * zone, the number of pages is calculated as:
5694 * nr_free_zone_pages = managed_pages - high_pages
5696 * Return: number of pages beyond high watermark.
5698 static unsigned long nr_free_zone_pages(int offset)
5703 /* Just pick one node, since fallback list is circular */
5704 unsigned long sum = 0;
5706 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5708 for_each_zone_zonelist(zone, z, zonelist, offset) {
5709 unsigned long size = zone_managed_pages(zone);
5710 unsigned long high = high_wmark_pages(zone);
5719 * nr_free_buffer_pages - count number of pages beyond high watermark
5721 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5722 * watermark within ZONE_DMA and ZONE_NORMAL.
5724 * Return: number of pages beyond high watermark within ZONE_DMA and
5727 unsigned long nr_free_buffer_pages(void)
5729 return nr_free_zone_pages(gfp_zone(GFP_USER));
5731 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5733 static inline void show_node(struct zone *zone)
5735 if (IS_ENABLED(CONFIG_NUMA))
5736 printk("Node %d ", zone_to_nid(zone));
5739 long si_mem_available(void)
5742 unsigned long pagecache;
5743 unsigned long wmark_low = 0;
5744 unsigned long pages[NR_LRU_LISTS];
5745 unsigned long reclaimable;
5749 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5750 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5753 wmark_low += low_wmark_pages(zone);
5756 * Estimate the amount of memory available for userspace allocations,
5757 * without causing swapping.
5759 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5762 * Not all the page cache can be freed, otherwise the system will
5763 * start swapping. Assume at least half of the page cache, or the
5764 * low watermark worth of cache, needs to stay.
5766 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5767 pagecache -= min(pagecache / 2, wmark_low);
5768 available += pagecache;
5771 * Part of the reclaimable slab and other kernel memory consists of
5772 * items that are in use, and cannot be freed. Cap this estimate at the
5775 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5776 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5777 available += reclaimable - min(reclaimable / 2, wmark_low);
5783 EXPORT_SYMBOL_GPL(si_mem_available);
5785 void si_meminfo(struct sysinfo *val)
5787 val->totalram = totalram_pages();
5788 val->sharedram = global_node_page_state(NR_SHMEM);
5789 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5790 val->bufferram = nr_blockdev_pages();
5791 val->totalhigh = totalhigh_pages();
5792 val->freehigh = nr_free_highpages();
5793 val->mem_unit = PAGE_SIZE;
5796 EXPORT_SYMBOL(si_meminfo);
5799 void si_meminfo_node(struct sysinfo *val, int nid)
5801 int zone_type; /* needs to be signed */
5802 unsigned long managed_pages = 0;
5803 unsigned long managed_highpages = 0;
5804 unsigned long free_highpages = 0;
5805 pg_data_t *pgdat = NODE_DATA(nid);
5807 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5808 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5809 val->totalram = managed_pages;
5810 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5811 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5812 #ifdef CONFIG_HIGHMEM
5813 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5814 struct zone *zone = &pgdat->node_zones[zone_type];
5816 if (is_highmem(zone)) {
5817 managed_highpages += zone_managed_pages(zone);
5818 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5821 val->totalhigh = managed_highpages;
5822 val->freehigh = free_highpages;
5824 val->totalhigh = managed_highpages;
5825 val->freehigh = free_highpages;
5827 val->mem_unit = PAGE_SIZE;
5832 * Determine whether the node should be displayed or not, depending on whether
5833 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5835 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5837 if (!(flags & SHOW_MEM_FILTER_NODES))
5841 * no node mask - aka implicit memory numa policy. Do not bother with
5842 * the synchronization - read_mems_allowed_begin - because we do not
5843 * have to be precise here.
5846 nodemask = &cpuset_current_mems_allowed;
5848 return !node_isset(nid, *nodemask);
5851 #define K(x) ((x) << (PAGE_SHIFT-10))
5853 static void show_migration_types(unsigned char type)
5855 static const char types[MIGRATE_TYPES] = {
5856 [MIGRATE_UNMOVABLE] = 'U',
5857 [MIGRATE_MOVABLE] = 'M',
5858 [MIGRATE_RECLAIMABLE] = 'E',
5859 [MIGRATE_HIGHATOMIC] = 'H',
5861 [MIGRATE_CMA] = 'C',
5863 #ifdef CONFIG_MEMORY_ISOLATION
5864 [MIGRATE_ISOLATE] = 'I',
5867 char tmp[MIGRATE_TYPES + 1];
5871 for (i = 0; i < MIGRATE_TYPES; i++) {
5872 if (type & (1 << i))
5877 printk(KERN_CONT "(%s) ", tmp);
5881 * Show free area list (used inside shift_scroll-lock stuff)
5882 * We also calculate the percentage fragmentation. We do this by counting the
5883 * memory on each free list with the exception of the first item on the list.
5886 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5889 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5891 unsigned long free_pcp = 0;
5896 for_each_populated_zone(zone) {
5897 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5900 for_each_online_cpu(cpu)
5901 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5904 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5905 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5906 " unevictable:%lu dirty:%lu writeback:%lu\n"
5907 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5908 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5909 " kernel_misc_reclaimable:%lu\n"
5910 " free:%lu free_pcp:%lu free_cma:%lu\n",
5911 global_node_page_state(NR_ACTIVE_ANON),
5912 global_node_page_state(NR_INACTIVE_ANON),
5913 global_node_page_state(NR_ISOLATED_ANON),
5914 global_node_page_state(NR_ACTIVE_FILE),
5915 global_node_page_state(NR_INACTIVE_FILE),
5916 global_node_page_state(NR_ISOLATED_FILE),
5917 global_node_page_state(NR_UNEVICTABLE),
5918 global_node_page_state(NR_FILE_DIRTY),
5919 global_node_page_state(NR_WRITEBACK),
5920 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5921 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5922 global_node_page_state(NR_FILE_MAPPED),
5923 global_node_page_state(NR_SHMEM),
5924 global_node_page_state(NR_PAGETABLE),
5925 global_zone_page_state(NR_BOUNCE),
5926 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5927 global_zone_page_state(NR_FREE_PAGES),
5929 global_zone_page_state(NR_FREE_CMA_PAGES));
5931 for_each_online_pgdat(pgdat) {
5932 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5936 " active_anon:%lukB"
5937 " inactive_anon:%lukB"
5938 " active_file:%lukB"
5939 " inactive_file:%lukB"
5940 " unevictable:%lukB"
5941 " isolated(anon):%lukB"
5942 " isolated(file):%lukB"
5947 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5949 " shmem_pmdmapped: %lukB"
5952 " writeback_tmp:%lukB"
5953 " kernel_stack:%lukB"
5954 #ifdef CONFIG_SHADOW_CALL_STACK
5955 " shadow_call_stack:%lukB"
5958 " all_unreclaimable? %s"
5961 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5962 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5963 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5964 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5965 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5966 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5967 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5968 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5969 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5970 K(node_page_state(pgdat, NR_WRITEBACK)),
5971 K(node_page_state(pgdat, NR_SHMEM)),
5972 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5973 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5974 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5975 K(node_page_state(pgdat, NR_ANON_THPS)),
5977 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5978 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5979 #ifdef CONFIG_SHADOW_CALL_STACK
5980 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5982 K(node_page_state(pgdat, NR_PAGETABLE)),
5983 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5987 for_each_populated_zone(zone) {
5990 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5994 for_each_online_cpu(cpu)
5995 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6005 " reserved_highatomic:%luKB"
6006 " active_anon:%lukB"
6007 " inactive_anon:%lukB"
6008 " active_file:%lukB"
6009 " inactive_file:%lukB"
6010 " unevictable:%lukB"
6011 " writepending:%lukB"
6021 K(zone_page_state(zone, NR_FREE_PAGES)),
6022 K(zone->watermark_boost),
6023 K(min_wmark_pages(zone)),
6024 K(low_wmark_pages(zone)),
6025 K(high_wmark_pages(zone)),
6026 K(zone->nr_reserved_highatomic),
6027 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6028 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6029 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6030 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6031 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6032 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6033 K(zone->present_pages),
6034 K(zone_managed_pages(zone)),
6035 K(zone_page_state(zone, NR_MLOCK)),
6036 K(zone_page_state(zone, NR_BOUNCE)),
6038 K(this_cpu_read(zone->per_cpu_pageset->count)),
6039 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6040 printk("lowmem_reserve[]:");
6041 for (i = 0; i < MAX_NR_ZONES; i++)
6042 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6043 printk(KERN_CONT "\n");
6046 for_each_populated_zone(zone) {
6048 unsigned long nr[MAX_ORDER], flags, total = 0;
6049 unsigned char types[MAX_ORDER];
6051 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6054 printk(KERN_CONT "%s: ", zone->name);
6056 spin_lock_irqsave(&zone->lock, flags);
6057 for (order = 0; order < MAX_ORDER; order++) {
6058 struct free_area *area = &zone->free_area[order];
6061 nr[order] = area->nr_free;
6062 total += nr[order] << order;
6065 for (type = 0; type < MIGRATE_TYPES; type++) {
6066 if (!free_area_empty(area, type))
6067 types[order] |= 1 << type;
6070 spin_unlock_irqrestore(&zone->lock, flags);
6071 for (order = 0; order < MAX_ORDER; order++) {
6072 printk(KERN_CONT "%lu*%lukB ",
6073 nr[order], K(1UL) << order);
6075 show_migration_types(types[order]);
6077 printk(KERN_CONT "= %lukB\n", K(total));
6080 hugetlb_show_meminfo();
6082 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6084 show_swap_cache_info();
6087 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6089 zoneref->zone = zone;
6090 zoneref->zone_idx = zone_idx(zone);
6094 * Builds allocation fallback zone lists.
6096 * Add all populated zones of a node to the zonelist.
6098 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6101 enum zone_type zone_type = MAX_NR_ZONES;
6106 zone = pgdat->node_zones + zone_type;
6107 if (managed_zone(zone)) {
6108 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6109 check_highest_zone(zone_type);
6111 } while (zone_type);
6118 static int __parse_numa_zonelist_order(char *s)
6121 * We used to support different zonelists modes but they turned
6122 * out to be just not useful. Let's keep the warning in place
6123 * if somebody still use the cmd line parameter so that we do
6124 * not fail it silently
6126 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6127 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6133 char numa_zonelist_order[] = "Node";
6136 * sysctl handler for numa_zonelist_order
6138 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6139 void *buffer, size_t *length, loff_t *ppos)
6142 return __parse_numa_zonelist_order(buffer);
6143 return proc_dostring(table, write, buffer, length, ppos);
6147 #define MAX_NODE_LOAD (nr_online_nodes)
6148 static int node_load[MAX_NUMNODES];
6151 * find_next_best_node - find the next node that should appear in a given node's fallback list
6152 * @node: node whose fallback list we're appending
6153 * @used_node_mask: nodemask_t of already used nodes
6155 * We use a number of factors to determine which is the next node that should
6156 * appear on a given node's fallback list. The node should not have appeared
6157 * already in @node's fallback list, and it should be the next closest node
6158 * according to the distance array (which contains arbitrary distance values
6159 * from each node to each node in the system), and should also prefer nodes
6160 * with no CPUs, since presumably they'll have very little allocation pressure
6161 * on them otherwise.
6163 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6165 int find_next_best_node(int node, nodemask_t *used_node_mask)
6168 int min_val = INT_MAX;
6169 int best_node = NUMA_NO_NODE;
6171 /* Use the local node if we haven't already */
6172 if (!node_isset(node, *used_node_mask)) {
6173 node_set(node, *used_node_mask);
6177 for_each_node_state(n, N_MEMORY) {
6179 /* Don't want a node to appear more than once */
6180 if (node_isset(n, *used_node_mask))
6183 /* Use the distance array to find the distance */
6184 val = node_distance(node, n);
6186 /* Penalize nodes under us ("prefer the next node") */
6189 /* Give preference to headless and unused nodes */
6190 if (!cpumask_empty(cpumask_of_node(n)))
6191 val += PENALTY_FOR_NODE_WITH_CPUS;
6193 /* Slight preference for less loaded node */
6194 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6195 val += node_load[n];
6197 if (val < min_val) {
6204 node_set(best_node, *used_node_mask);
6211 * Build zonelists ordered by node and zones within node.
6212 * This results in maximum locality--normal zone overflows into local
6213 * DMA zone, if any--but risks exhausting DMA zone.
6215 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6218 struct zoneref *zonerefs;
6221 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6223 for (i = 0; i < nr_nodes; i++) {
6226 pg_data_t *node = NODE_DATA(node_order[i]);
6228 nr_zones = build_zonerefs_node(node, zonerefs);
6229 zonerefs += nr_zones;
6231 zonerefs->zone = NULL;
6232 zonerefs->zone_idx = 0;
6236 * Build gfp_thisnode zonelists
6238 static void build_thisnode_zonelists(pg_data_t *pgdat)
6240 struct zoneref *zonerefs;
6243 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6244 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6245 zonerefs += nr_zones;
6246 zonerefs->zone = NULL;
6247 zonerefs->zone_idx = 0;
6251 * Build zonelists ordered by zone and nodes within zones.
6252 * This results in conserving DMA zone[s] until all Normal memory is
6253 * exhausted, but results in overflowing to remote node while memory
6254 * may still exist in local DMA zone.
6257 static void build_zonelists(pg_data_t *pgdat)
6259 static int node_order[MAX_NUMNODES];
6260 int node, load, nr_nodes = 0;
6261 nodemask_t used_mask = NODE_MASK_NONE;
6262 int local_node, prev_node;
6264 /* NUMA-aware ordering of nodes */
6265 local_node = pgdat->node_id;
6266 load = nr_online_nodes;
6267 prev_node = local_node;
6269 memset(node_order, 0, sizeof(node_order));
6270 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6272 * We don't want to pressure a particular node.
6273 * So adding penalty to the first node in same
6274 * distance group to make it round-robin.
6276 if (node_distance(local_node, node) !=
6277 node_distance(local_node, prev_node))
6278 node_load[node] += load;
6280 node_order[nr_nodes++] = node;
6285 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6286 build_thisnode_zonelists(pgdat);
6287 pr_info("Fallback order for Node %d: ", local_node);
6288 for (node = 0; node < nr_nodes; node++)
6289 pr_cont("%d ", node_order[node]);
6293 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6295 * Return node id of node used for "local" allocations.
6296 * I.e., first node id of first zone in arg node's generic zonelist.
6297 * Used for initializing percpu 'numa_mem', which is used primarily
6298 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6300 int local_memory_node(int node)
6304 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6305 gfp_zone(GFP_KERNEL),
6307 return zone_to_nid(z->zone);
6311 static void setup_min_unmapped_ratio(void);
6312 static void setup_min_slab_ratio(void);
6313 #else /* CONFIG_NUMA */
6315 static void build_zonelists(pg_data_t *pgdat)
6317 int node, local_node;
6318 struct zoneref *zonerefs;
6321 local_node = pgdat->node_id;
6323 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6324 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6325 zonerefs += nr_zones;
6328 * Now we build the zonelist so that it contains the zones
6329 * of all the other nodes.
6330 * We don't want to pressure a particular node, so when
6331 * building the zones for node N, we make sure that the
6332 * zones coming right after the local ones are those from
6333 * node N+1 (modulo N)
6335 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6336 if (!node_online(node))
6338 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6339 zonerefs += nr_zones;
6341 for (node = 0; node < local_node; node++) {
6342 if (!node_online(node))
6344 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6345 zonerefs += nr_zones;
6348 zonerefs->zone = NULL;
6349 zonerefs->zone_idx = 0;
6352 #endif /* CONFIG_NUMA */
6355 * Boot pageset table. One per cpu which is going to be used for all
6356 * zones and all nodes. The parameters will be set in such a way
6357 * that an item put on a list will immediately be handed over to
6358 * the buddy list. This is safe since pageset manipulation is done
6359 * with interrupts disabled.
6361 * The boot_pagesets must be kept even after bootup is complete for
6362 * unused processors and/or zones. They do play a role for bootstrapping
6363 * hotplugged processors.
6365 * zoneinfo_show() and maybe other functions do
6366 * not check if the processor is online before following the pageset pointer.
6367 * Other parts of the kernel may not check if the zone is available.
6369 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6370 /* These effectively disable the pcplists in the boot pageset completely */
6371 #define BOOT_PAGESET_HIGH 0
6372 #define BOOT_PAGESET_BATCH 1
6373 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6374 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6375 DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6377 static void __build_all_zonelists(void *data)
6380 int __maybe_unused cpu;
6381 pg_data_t *self = data;
6382 static DEFINE_SPINLOCK(lock);
6387 memset(node_load, 0, sizeof(node_load));
6391 * This node is hotadded and no memory is yet present. So just
6392 * building zonelists is fine - no need to touch other nodes.
6394 if (self && !node_online(self->node_id)) {
6395 build_zonelists(self);
6398 * All possible nodes have pgdat preallocated
6401 for_each_node(nid) {
6402 pg_data_t *pgdat = NODE_DATA(nid);
6404 build_zonelists(pgdat);
6407 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6409 * We now know the "local memory node" for each node--
6410 * i.e., the node of the first zone in the generic zonelist.
6411 * Set up numa_mem percpu variable for on-line cpus. During
6412 * boot, only the boot cpu should be on-line; we'll init the
6413 * secondary cpus' numa_mem as they come on-line. During
6414 * node/memory hotplug, we'll fixup all on-line cpus.
6416 for_each_online_cpu(cpu)
6417 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6424 static noinline void __init
6425 build_all_zonelists_init(void)
6429 __build_all_zonelists(NULL);
6432 * Initialize the boot_pagesets that are going to be used
6433 * for bootstrapping processors. The real pagesets for
6434 * each zone will be allocated later when the per cpu
6435 * allocator is available.
6437 * boot_pagesets are used also for bootstrapping offline
6438 * cpus if the system is already booted because the pagesets
6439 * are needed to initialize allocators on a specific cpu too.
6440 * F.e. the percpu allocator needs the page allocator which
6441 * needs the percpu allocator in order to allocate its pagesets
6442 * (a chicken-egg dilemma).
6444 for_each_possible_cpu(cpu)
6445 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6447 mminit_verify_zonelist();
6448 cpuset_init_current_mems_allowed();
6452 * unless system_state == SYSTEM_BOOTING.
6454 * __ref due to call of __init annotated helper build_all_zonelists_init
6455 * [protected by SYSTEM_BOOTING].
6457 void __ref build_all_zonelists(pg_data_t *pgdat)
6459 unsigned long vm_total_pages;
6461 if (system_state == SYSTEM_BOOTING) {
6462 build_all_zonelists_init();
6464 __build_all_zonelists(pgdat);
6465 /* cpuset refresh routine should be here */
6467 /* Get the number of free pages beyond high watermark in all zones. */
6468 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6470 * Disable grouping by mobility if the number of pages in the
6471 * system is too low to allow the mechanism to work. It would be
6472 * more accurate, but expensive to check per-zone. This check is
6473 * made on memory-hotadd so a system can start with mobility
6474 * disabled and enable it later
6476 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6477 page_group_by_mobility_disabled = 1;
6479 page_group_by_mobility_disabled = 0;
6481 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6483 page_group_by_mobility_disabled ? "off" : "on",
6486 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6490 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6491 static bool __meminit
6492 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6494 static struct memblock_region *r;
6496 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6497 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6498 for_each_mem_region(r) {
6499 if (*pfn < memblock_region_memory_end_pfn(r))
6503 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6504 memblock_is_mirror(r)) {
6505 *pfn = memblock_region_memory_end_pfn(r);
6513 * Initially all pages are reserved - free ones are freed
6514 * up by memblock_free_all() once the early boot process is
6515 * done. Non-atomic initialization, single-pass.
6517 * All aligned pageblocks are initialized to the specified migratetype
6518 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6519 * zone stats (e.g., nr_isolate_pageblock) are touched.
6521 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6522 unsigned long start_pfn, unsigned long zone_end_pfn,
6523 enum meminit_context context,
6524 struct vmem_altmap *altmap, int migratetype)
6526 unsigned long pfn, end_pfn = start_pfn + size;
6529 if (highest_memmap_pfn < end_pfn - 1)
6530 highest_memmap_pfn = end_pfn - 1;
6532 #ifdef CONFIG_ZONE_DEVICE
6534 * Honor reservation requested by the driver for this ZONE_DEVICE
6535 * memory. We limit the total number of pages to initialize to just
6536 * those that might contain the memory mapping. We will defer the
6537 * ZONE_DEVICE page initialization until after we have released
6540 if (zone == ZONE_DEVICE) {
6544 if (start_pfn == altmap->base_pfn)
6545 start_pfn += altmap->reserve;
6546 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6550 for (pfn = start_pfn; pfn < end_pfn; ) {
6552 * There can be holes in boot-time mem_map[]s handed to this
6553 * function. They do not exist on hotplugged memory.
6555 if (context == MEMINIT_EARLY) {
6556 if (overlap_memmap_init(zone, &pfn))
6558 if (defer_init(nid, pfn, zone_end_pfn))
6562 page = pfn_to_page(pfn);
6563 __init_single_page(page, pfn, zone, nid);
6564 if (context == MEMINIT_HOTPLUG)
6565 __SetPageReserved(page);
6568 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6569 * such that unmovable allocations won't be scattered all
6570 * over the place during system boot.
6572 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6573 set_pageblock_migratetype(page, migratetype);
6580 #ifdef CONFIG_ZONE_DEVICE
6581 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6582 unsigned long zone_idx, int nid,
6583 struct dev_pagemap *pgmap)
6586 __init_single_page(page, pfn, zone_idx, nid);
6589 * Mark page reserved as it will need to wait for onlining
6590 * phase for it to be fully associated with a zone.
6592 * We can use the non-atomic __set_bit operation for setting
6593 * the flag as we are still initializing the pages.
6595 __SetPageReserved(page);
6598 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6599 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6600 * ever freed or placed on a driver-private list.
6602 page->pgmap = pgmap;
6603 page->zone_device_data = NULL;
6606 * Mark the block movable so that blocks are reserved for
6607 * movable at startup. This will force kernel allocations
6608 * to reserve their blocks rather than leaking throughout
6609 * the address space during boot when many long-lived
6610 * kernel allocations are made.
6612 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6613 * because this is done early in section_activate()
6615 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6616 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6621 static void __ref memmap_init_compound(struct page *head,
6622 unsigned long head_pfn,
6623 unsigned long zone_idx, int nid,
6624 struct dev_pagemap *pgmap,
6625 unsigned long nr_pages)
6627 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6628 unsigned int order = pgmap->vmemmap_shift;
6630 __SetPageHead(head);
6631 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6632 struct page *page = pfn_to_page(pfn);
6634 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6635 prep_compound_tail(head, pfn - head_pfn);
6636 set_page_count(page, 0);
6639 * The first tail page stores compound_mapcount_ptr() and
6640 * compound_order() and the second tail page stores
6641 * compound_pincount_ptr(). Call prep_compound_head() after
6642 * the first and second tail pages have been initialized to
6643 * not have the data overwritten.
6645 if (pfn == head_pfn + 2)
6646 prep_compound_head(head, order);
6650 void __ref memmap_init_zone_device(struct zone *zone,
6651 unsigned long start_pfn,
6652 unsigned long nr_pages,
6653 struct dev_pagemap *pgmap)
6655 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6656 struct pglist_data *pgdat = zone->zone_pgdat;
6657 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6658 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6659 unsigned long zone_idx = zone_idx(zone);
6660 unsigned long start = jiffies;
6661 int nid = pgdat->node_id;
6663 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6667 * The call to memmap_init should have already taken care
6668 * of the pages reserved for the memmap, so we can just jump to
6669 * the end of that region and start processing the device pages.
6672 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6673 nr_pages = end_pfn - start_pfn;
6676 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6677 struct page *page = pfn_to_page(pfn);
6679 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6681 if (pfns_per_compound == 1)
6684 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6688 pr_info("%s initialised %lu pages in %ums\n", __func__,
6689 nr_pages, jiffies_to_msecs(jiffies - start));
6693 static void __meminit zone_init_free_lists(struct zone *zone)
6695 unsigned int order, t;
6696 for_each_migratetype_order(order, t) {
6697 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6698 zone->free_area[order].nr_free = 0;
6703 * Only struct pages that correspond to ranges defined by memblock.memory
6704 * are zeroed and initialized by going through __init_single_page() during
6705 * memmap_init_zone_range().
6707 * But, there could be struct pages that correspond to holes in
6708 * memblock.memory. This can happen because of the following reasons:
6709 * - physical memory bank size is not necessarily the exact multiple of the
6710 * arbitrary section size
6711 * - early reserved memory may not be listed in memblock.memory
6712 * - memory layouts defined with memmap= kernel parameter may not align
6713 * nicely with memmap sections
6715 * Explicitly initialize those struct pages so that:
6716 * - PG_Reserved is set
6717 * - zone and node links point to zone and node that span the page if the
6718 * hole is in the middle of a zone
6719 * - zone and node links point to adjacent zone/node if the hole falls on
6720 * the zone boundary; the pages in such holes will be prepended to the
6721 * zone/node above the hole except for the trailing pages in the last
6722 * section that will be appended to the zone/node below.
6724 static void __init init_unavailable_range(unsigned long spfn,
6731 for (pfn = spfn; pfn < epfn; pfn++) {
6732 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6733 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6734 + pageblock_nr_pages - 1;
6737 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6738 __SetPageReserved(pfn_to_page(pfn));
6743 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6744 node, zone_names[zone], pgcnt);
6747 static void __init memmap_init_zone_range(struct zone *zone,
6748 unsigned long start_pfn,
6749 unsigned long end_pfn,
6750 unsigned long *hole_pfn)
6752 unsigned long zone_start_pfn = zone->zone_start_pfn;
6753 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6754 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6756 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6757 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6759 if (start_pfn >= end_pfn)
6762 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6763 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6765 if (*hole_pfn < start_pfn)
6766 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6768 *hole_pfn = end_pfn;
6771 static void __init memmap_init(void)
6773 unsigned long start_pfn, end_pfn;
6774 unsigned long hole_pfn = 0;
6775 int i, j, zone_id = 0, nid;
6777 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6778 struct pglist_data *node = NODE_DATA(nid);
6780 for (j = 0; j < MAX_NR_ZONES; j++) {
6781 struct zone *zone = node->node_zones + j;
6783 if (!populated_zone(zone))
6786 memmap_init_zone_range(zone, start_pfn, end_pfn,
6792 #ifdef CONFIG_SPARSEMEM
6794 * Initialize the memory map for hole in the range [memory_end,
6796 * Append the pages in this hole to the highest zone in the last
6798 * The call to init_unavailable_range() is outside the ifdef to
6799 * silence the compiler warining about zone_id set but not used;
6800 * for FLATMEM it is a nop anyway
6802 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6803 if (hole_pfn < end_pfn)
6805 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6808 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6809 phys_addr_t min_addr, int nid, bool exact_nid)
6814 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6815 MEMBLOCK_ALLOC_ACCESSIBLE,
6818 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6819 MEMBLOCK_ALLOC_ACCESSIBLE,
6822 if (ptr && size > 0)
6823 page_init_poison(ptr, size);
6828 static int zone_batchsize(struct zone *zone)
6834 * The number of pages to batch allocate is either ~0.1%
6835 * of the zone or 1MB, whichever is smaller. The batch
6836 * size is striking a balance between allocation latency
6837 * and zone lock contention.
6839 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6840 batch /= 4; /* We effectively *= 4 below */
6845 * Clamp the batch to a 2^n - 1 value. Having a power
6846 * of 2 value was found to be more likely to have
6847 * suboptimal cache aliasing properties in some cases.
6849 * For example if 2 tasks are alternately allocating
6850 * batches of pages, one task can end up with a lot
6851 * of pages of one half of the possible page colors
6852 * and the other with pages of the other colors.
6854 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6859 /* The deferral and batching of frees should be suppressed under NOMMU
6862 * The problem is that NOMMU needs to be able to allocate large chunks
6863 * of contiguous memory as there's no hardware page translation to
6864 * assemble apparent contiguous memory from discontiguous pages.
6866 * Queueing large contiguous runs of pages for batching, however,
6867 * causes the pages to actually be freed in smaller chunks. As there
6868 * can be a significant delay between the individual batches being
6869 * recycled, this leads to the once large chunks of space being
6870 * fragmented and becoming unavailable for high-order allocations.
6876 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6881 unsigned long total_pages;
6883 if (!percpu_pagelist_high_fraction) {
6885 * By default, the high value of the pcp is based on the zone
6886 * low watermark so that if they are full then background
6887 * reclaim will not be started prematurely.
6889 total_pages = low_wmark_pages(zone);
6892 * If percpu_pagelist_high_fraction is configured, the high
6893 * value is based on a fraction of the managed pages in the
6896 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6900 * Split the high value across all online CPUs local to the zone. Note
6901 * that early in boot that CPUs may not be online yet and that during
6902 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6903 * onlined. For memory nodes that have no CPUs, split pcp->high across
6904 * all online CPUs to mitigate the risk that reclaim is triggered
6905 * prematurely due to pages stored on pcp lists.
6907 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6909 nr_split_cpus = num_online_cpus();
6910 high = total_pages / nr_split_cpus;
6913 * Ensure high is at least batch*4. The multiple is based on the
6914 * historical relationship between high and batch.
6916 high = max(high, batch << 2);
6925 * pcp->high and pcp->batch values are related and generally batch is lower
6926 * than high. They are also related to pcp->count such that count is lower
6927 * than high, and as soon as it reaches high, the pcplist is flushed.
6929 * However, guaranteeing these relations at all times would require e.g. write
6930 * barriers here but also careful usage of read barriers at the read side, and
6931 * thus be prone to error and bad for performance. Thus the update only prevents
6932 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6933 * can cope with those fields changing asynchronously, and fully trust only the
6934 * pcp->count field on the local CPU with interrupts disabled.
6936 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6937 * outside of boot time (or some other assurance that no concurrent updaters
6940 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6941 unsigned long batch)
6943 WRITE_ONCE(pcp->batch, batch);
6944 WRITE_ONCE(pcp->high, high);
6947 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6951 memset(pcp, 0, sizeof(*pcp));
6952 memset(pzstats, 0, sizeof(*pzstats));
6954 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6955 INIT_LIST_HEAD(&pcp->lists[pindex]);
6958 * Set batch and high values safe for a boot pageset. A true percpu
6959 * pageset's initialization will update them subsequently. Here we don't
6960 * need to be as careful as pageset_update() as nobody can access the
6963 pcp->high = BOOT_PAGESET_HIGH;
6964 pcp->batch = BOOT_PAGESET_BATCH;
6965 pcp->free_factor = 0;
6968 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6969 unsigned long batch)
6971 struct per_cpu_pages *pcp;
6974 for_each_possible_cpu(cpu) {
6975 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6976 pageset_update(pcp, high, batch);
6981 * Calculate and set new high and batch values for all per-cpu pagesets of a
6982 * zone based on the zone's size.
6984 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6986 int new_high, new_batch;
6988 new_batch = max(1, zone_batchsize(zone));
6989 new_high = zone_highsize(zone, new_batch, cpu_online);
6991 if (zone->pageset_high == new_high &&
6992 zone->pageset_batch == new_batch)
6995 zone->pageset_high = new_high;
6996 zone->pageset_batch = new_batch;
6998 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7001 void __meminit setup_zone_pageset(struct zone *zone)
7005 /* Size may be 0 on !SMP && !NUMA */
7006 if (sizeof(struct per_cpu_zonestat) > 0)
7007 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7009 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7010 for_each_possible_cpu(cpu) {
7011 struct per_cpu_pages *pcp;
7012 struct per_cpu_zonestat *pzstats;
7014 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7015 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7016 per_cpu_pages_init(pcp, pzstats);
7019 zone_set_pageset_high_and_batch(zone, 0);
7023 * Allocate per cpu pagesets and initialize them.
7024 * Before this call only boot pagesets were available.
7026 void __init setup_per_cpu_pageset(void)
7028 struct pglist_data *pgdat;
7030 int __maybe_unused cpu;
7032 for_each_populated_zone(zone)
7033 setup_zone_pageset(zone);
7037 * Unpopulated zones continue using the boot pagesets.
7038 * The numa stats for these pagesets need to be reset.
7039 * Otherwise, they will end up skewing the stats of
7040 * the nodes these zones are associated with.
7042 for_each_possible_cpu(cpu) {
7043 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7044 memset(pzstats->vm_numa_event, 0,
7045 sizeof(pzstats->vm_numa_event));
7049 for_each_online_pgdat(pgdat)
7050 pgdat->per_cpu_nodestats =
7051 alloc_percpu(struct per_cpu_nodestat);
7054 static __meminit void zone_pcp_init(struct zone *zone)
7057 * per cpu subsystem is not up at this point. The following code
7058 * relies on the ability of the linker to provide the
7059 * offset of a (static) per cpu variable into the per cpu area.
7061 zone->per_cpu_pageset = &boot_pageset;
7062 zone->per_cpu_zonestats = &boot_zonestats;
7063 zone->pageset_high = BOOT_PAGESET_HIGH;
7064 zone->pageset_batch = BOOT_PAGESET_BATCH;
7066 if (populated_zone(zone))
7067 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7068 zone->present_pages, zone_batchsize(zone));
7071 void __meminit init_currently_empty_zone(struct zone *zone,
7072 unsigned long zone_start_pfn,
7075 struct pglist_data *pgdat = zone->zone_pgdat;
7076 int zone_idx = zone_idx(zone) + 1;
7078 if (zone_idx > pgdat->nr_zones)
7079 pgdat->nr_zones = zone_idx;
7081 zone->zone_start_pfn = zone_start_pfn;
7083 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7084 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7086 (unsigned long)zone_idx(zone),
7087 zone_start_pfn, (zone_start_pfn + size));
7089 zone_init_free_lists(zone);
7090 zone->initialized = 1;
7094 * get_pfn_range_for_nid - Return the start and end page frames for a node
7095 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7096 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7097 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7099 * It returns the start and end page frame of a node based on information
7100 * provided by memblock_set_node(). If called for a node
7101 * with no available memory, a warning is printed and the start and end
7104 void __init get_pfn_range_for_nid(unsigned int nid,
7105 unsigned long *start_pfn, unsigned long *end_pfn)
7107 unsigned long this_start_pfn, this_end_pfn;
7113 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7114 *start_pfn = min(*start_pfn, this_start_pfn);
7115 *end_pfn = max(*end_pfn, this_end_pfn);
7118 if (*start_pfn == -1UL)
7123 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7124 * assumption is made that zones within a node are ordered in monotonic
7125 * increasing memory addresses so that the "highest" populated zone is used
7127 static void __init find_usable_zone_for_movable(void)
7130 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7131 if (zone_index == ZONE_MOVABLE)
7134 if (arch_zone_highest_possible_pfn[zone_index] >
7135 arch_zone_lowest_possible_pfn[zone_index])
7139 VM_BUG_ON(zone_index == -1);
7140 movable_zone = zone_index;
7144 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7145 * because it is sized independent of architecture. Unlike the other zones,
7146 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7147 * in each node depending on the size of each node and how evenly kernelcore
7148 * is distributed. This helper function adjusts the zone ranges
7149 * provided by the architecture for a given node by using the end of the
7150 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7151 * zones within a node are in order of monotonic increases memory addresses
7153 static void __init adjust_zone_range_for_zone_movable(int nid,
7154 unsigned long zone_type,
7155 unsigned long node_start_pfn,
7156 unsigned long node_end_pfn,
7157 unsigned long *zone_start_pfn,
7158 unsigned long *zone_end_pfn)
7160 /* Only adjust if ZONE_MOVABLE is on this node */
7161 if (zone_movable_pfn[nid]) {
7162 /* Size ZONE_MOVABLE */
7163 if (zone_type == ZONE_MOVABLE) {
7164 *zone_start_pfn = zone_movable_pfn[nid];
7165 *zone_end_pfn = min(node_end_pfn,
7166 arch_zone_highest_possible_pfn[movable_zone]);
7168 /* Adjust for ZONE_MOVABLE starting within this range */
7169 } else if (!mirrored_kernelcore &&
7170 *zone_start_pfn < zone_movable_pfn[nid] &&
7171 *zone_end_pfn > zone_movable_pfn[nid]) {
7172 *zone_end_pfn = zone_movable_pfn[nid];
7174 /* Check if this whole range is within ZONE_MOVABLE */
7175 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7176 *zone_start_pfn = *zone_end_pfn;
7181 * Return the number of pages a zone spans in a node, including holes
7182 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7184 static unsigned long __init zone_spanned_pages_in_node(int nid,
7185 unsigned long zone_type,
7186 unsigned long node_start_pfn,
7187 unsigned long node_end_pfn,
7188 unsigned long *zone_start_pfn,
7189 unsigned long *zone_end_pfn)
7191 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7192 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7193 /* When hotadd a new node from cpu_up(), the node should be empty */
7194 if (!node_start_pfn && !node_end_pfn)
7197 /* Get the start and end of the zone */
7198 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7199 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7200 adjust_zone_range_for_zone_movable(nid, zone_type,
7201 node_start_pfn, node_end_pfn,
7202 zone_start_pfn, zone_end_pfn);
7204 /* Check that this node has pages within the zone's required range */
7205 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7208 /* Move the zone boundaries inside the node if necessary */
7209 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7210 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7212 /* Return the spanned pages */
7213 return *zone_end_pfn - *zone_start_pfn;
7217 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7218 * then all holes in the requested range will be accounted for.
7220 unsigned long __init __absent_pages_in_range(int nid,
7221 unsigned long range_start_pfn,
7222 unsigned long range_end_pfn)
7224 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7225 unsigned long start_pfn, end_pfn;
7228 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7229 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7230 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7231 nr_absent -= end_pfn - start_pfn;
7237 * absent_pages_in_range - Return number of page frames in holes within a range
7238 * @start_pfn: The start PFN to start searching for holes
7239 * @end_pfn: The end PFN to stop searching for holes
7241 * Return: the number of pages frames in memory holes within a range.
7243 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7244 unsigned long end_pfn)
7246 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7249 /* Return the number of page frames in holes in a zone on a node */
7250 static unsigned long __init zone_absent_pages_in_node(int nid,
7251 unsigned long zone_type,
7252 unsigned long node_start_pfn,
7253 unsigned long node_end_pfn)
7255 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7256 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7257 unsigned long zone_start_pfn, zone_end_pfn;
7258 unsigned long nr_absent;
7260 /* When hotadd a new node from cpu_up(), the node should be empty */
7261 if (!node_start_pfn && !node_end_pfn)
7264 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7265 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7267 adjust_zone_range_for_zone_movable(nid, zone_type,
7268 node_start_pfn, node_end_pfn,
7269 &zone_start_pfn, &zone_end_pfn);
7270 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7273 * ZONE_MOVABLE handling.
7274 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7277 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7278 unsigned long start_pfn, end_pfn;
7279 struct memblock_region *r;
7281 for_each_mem_region(r) {
7282 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7283 zone_start_pfn, zone_end_pfn);
7284 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7285 zone_start_pfn, zone_end_pfn);
7287 if (zone_type == ZONE_MOVABLE &&
7288 memblock_is_mirror(r))
7289 nr_absent += end_pfn - start_pfn;
7291 if (zone_type == ZONE_NORMAL &&
7292 !memblock_is_mirror(r))
7293 nr_absent += end_pfn - start_pfn;
7300 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7301 unsigned long node_start_pfn,
7302 unsigned long node_end_pfn)
7304 unsigned long realtotalpages = 0, totalpages = 0;
7307 for (i = 0; i < MAX_NR_ZONES; i++) {
7308 struct zone *zone = pgdat->node_zones + i;
7309 unsigned long zone_start_pfn, zone_end_pfn;
7310 unsigned long spanned, absent;
7311 unsigned long size, real_size;
7313 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7318 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7323 real_size = size - absent;
7326 zone->zone_start_pfn = zone_start_pfn;
7328 zone->zone_start_pfn = 0;
7329 zone->spanned_pages = size;
7330 zone->present_pages = real_size;
7331 #if defined(CONFIG_MEMORY_HOTPLUG)
7332 zone->present_early_pages = real_size;
7336 realtotalpages += real_size;
7339 pgdat->node_spanned_pages = totalpages;
7340 pgdat->node_present_pages = realtotalpages;
7341 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7344 #ifndef CONFIG_SPARSEMEM
7346 * Calculate the size of the zone->blockflags rounded to an unsigned long
7347 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7348 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7349 * round what is now in bits to nearest long in bits, then return it in
7352 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7354 unsigned long usemapsize;
7356 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7357 usemapsize = roundup(zonesize, pageblock_nr_pages);
7358 usemapsize = usemapsize >> pageblock_order;
7359 usemapsize *= NR_PAGEBLOCK_BITS;
7360 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7362 return usemapsize / 8;
7365 static void __ref setup_usemap(struct zone *zone)
7367 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7368 zone->spanned_pages);
7369 zone->pageblock_flags = NULL;
7371 zone->pageblock_flags =
7372 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7374 if (!zone->pageblock_flags)
7375 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7376 usemapsize, zone->name, zone_to_nid(zone));
7380 static inline void setup_usemap(struct zone *zone) {}
7381 #endif /* CONFIG_SPARSEMEM */
7383 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7385 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7386 void __init set_pageblock_order(void)
7388 unsigned int order = MAX_ORDER - 1;
7390 /* Check that pageblock_nr_pages has not already been setup */
7391 if (pageblock_order)
7394 /* Don't let pageblocks exceed the maximum allocation granularity. */
7395 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7396 order = HUGETLB_PAGE_ORDER;
7399 * Assume the largest contiguous order of interest is a huge page.
7400 * This value may be variable depending on boot parameters on IA64 and
7403 pageblock_order = order;
7405 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7408 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7409 * is unused as pageblock_order is set at compile-time. See
7410 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7413 void __init set_pageblock_order(void)
7417 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7419 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7420 unsigned long present_pages)
7422 unsigned long pages = spanned_pages;
7425 * Provide a more accurate estimation if there are holes within
7426 * the zone and SPARSEMEM is in use. If there are holes within the
7427 * zone, each populated memory region may cost us one or two extra
7428 * memmap pages due to alignment because memmap pages for each
7429 * populated regions may not be naturally aligned on page boundary.
7430 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7432 if (spanned_pages > present_pages + (present_pages >> 4) &&
7433 IS_ENABLED(CONFIG_SPARSEMEM))
7434 pages = present_pages;
7436 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7439 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7440 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7442 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7444 spin_lock_init(&ds_queue->split_queue_lock);
7445 INIT_LIST_HEAD(&ds_queue->split_queue);
7446 ds_queue->split_queue_len = 0;
7449 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7452 #ifdef CONFIG_COMPACTION
7453 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7455 init_waitqueue_head(&pgdat->kcompactd_wait);
7458 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7461 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7465 pgdat_resize_init(pgdat);
7467 pgdat_init_split_queue(pgdat);
7468 pgdat_init_kcompactd(pgdat);
7470 init_waitqueue_head(&pgdat->kswapd_wait);
7471 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7473 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7474 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7476 pgdat_page_ext_init(pgdat);
7477 lruvec_init(&pgdat->__lruvec);
7480 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7481 unsigned long remaining_pages)
7483 atomic_long_set(&zone->managed_pages, remaining_pages);
7484 zone_set_nid(zone, nid);
7485 zone->name = zone_names[idx];
7486 zone->zone_pgdat = NODE_DATA(nid);
7487 spin_lock_init(&zone->lock);
7488 zone_seqlock_init(zone);
7489 zone_pcp_init(zone);
7493 * Set up the zone data structures
7494 * - init pgdat internals
7495 * - init all zones belonging to this node
7497 * NOTE: this function is only called during memory hotplug
7499 #ifdef CONFIG_MEMORY_HOTPLUG
7500 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7502 int nid = pgdat->node_id;
7506 pgdat_init_internals(pgdat);
7508 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7509 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7512 * Reset the nr_zones, order and highest_zoneidx before reuse.
7513 * Note that kswapd will init kswapd_highest_zoneidx properly
7514 * when it starts in the near future.
7516 pgdat->nr_zones = 0;
7517 pgdat->kswapd_order = 0;
7518 pgdat->kswapd_highest_zoneidx = 0;
7519 pgdat->node_start_pfn = 0;
7520 for_each_online_cpu(cpu) {
7521 struct per_cpu_nodestat *p;
7523 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7524 memset(p, 0, sizeof(*p));
7527 for (z = 0; z < MAX_NR_ZONES; z++)
7528 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7533 * Set up the zone data structures:
7534 * - mark all pages reserved
7535 * - mark all memory queues empty
7536 * - clear the memory bitmaps
7538 * NOTE: pgdat should get zeroed by caller.
7539 * NOTE: this function is only called during early init.
7541 static void __init free_area_init_core(struct pglist_data *pgdat)
7544 int nid = pgdat->node_id;
7546 pgdat_init_internals(pgdat);
7547 pgdat->per_cpu_nodestats = &boot_nodestats;
7549 for (j = 0; j < MAX_NR_ZONES; j++) {
7550 struct zone *zone = pgdat->node_zones + j;
7551 unsigned long size, freesize, memmap_pages;
7553 size = zone->spanned_pages;
7554 freesize = zone->present_pages;
7557 * Adjust freesize so that it accounts for how much memory
7558 * is used by this zone for memmap. This affects the watermark
7559 * and per-cpu initialisations
7561 memmap_pages = calc_memmap_size(size, freesize);
7562 if (!is_highmem_idx(j)) {
7563 if (freesize >= memmap_pages) {
7564 freesize -= memmap_pages;
7566 pr_debug(" %s zone: %lu pages used for memmap\n",
7567 zone_names[j], memmap_pages);
7569 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7570 zone_names[j], memmap_pages, freesize);
7573 /* Account for reserved pages */
7574 if (j == 0 && freesize > dma_reserve) {
7575 freesize -= dma_reserve;
7576 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7579 if (!is_highmem_idx(j))
7580 nr_kernel_pages += freesize;
7581 /* Charge for highmem memmap if there are enough kernel pages */
7582 else if (nr_kernel_pages > memmap_pages * 2)
7583 nr_kernel_pages -= memmap_pages;
7584 nr_all_pages += freesize;
7587 * Set an approximate value for lowmem here, it will be adjusted
7588 * when the bootmem allocator frees pages into the buddy system.
7589 * And all highmem pages will be managed by the buddy system.
7591 zone_init_internals(zone, j, nid, freesize);
7596 set_pageblock_order();
7598 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7602 #ifdef CONFIG_FLATMEM
7603 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7605 unsigned long __maybe_unused start = 0;
7606 unsigned long __maybe_unused offset = 0;
7608 /* Skip empty nodes */
7609 if (!pgdat->node_spanned_pages)
7612 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7613 offset = pgdat->node_start_pfn - start;
7614 /* ia64 gets its own node_mem_map, before this, without bootmem */
7615 if (!pgdat->node_mem_map) {
7616 unsigned long size, end;
7620 * The zone's endpoints aren't required to be MAX_ORDER
7621 * aligned but the node_mem_map endpoints must be in order
7622 * for the buddy allocator to function correctly.
7624 end = pgdat_end_pfn(pgdat);
7625 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7626 size = (end - start) * sizeof(struct page);
7627 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7628 pgdat->node_id, false);
7630 panic("Failed to allocate %ld bytes for node %d memory map\n",
7631 size, pgdat->node_id);
7632 pgdat->node_mem_map = map + offset;
7634 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7635 __func__, pgdat->node_id, (unsigned long)pgdat,
7636 (unsigned long)pgdat->node_mem_map);
7639 * With no DISCONTIG, the global mem_map is just set as node 0's
7641 if (pgdat == NODE_DATA(0)) {
7642 mem_map = NODE_DATA(0)->node_mem_map;
7643 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7649 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7650 #endif /* CONFIG_FLATMEM */
7652 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7653 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7655 pgdat->first_deferred_pfn = ULONG_MAX;
7658 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7661 static void __init free_area_init_node(int nid)
7663 pg_data_t *pgdat = NODE_DATA(nid);
7664 unsigned long start_pfn = 0;
7665 unsigned long end_pfn = 0;
7667 /* pg_data_t should be reset to zero when it's allocated */
7668 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7670 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7672 pgdat->node_id = nid;
7673 pgdat->node_start_pfn = start_pfn;
7674 pgdat->per_cpu_nodestats = NULL;
7676 if (start_pfn != end_pfn) {
7677 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7678 (u64)start_pfn << PAGE_SHIFT,
7679 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7681 pr_info("Initmem setup node %d as memoryless\n", nid);
7684 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7686 alloc_node_mem_map(pgdat);
7687 pgdat_set_deferred_range(pgdat);
7689 free_area_init_core(pgdat);
7692 static void __init free_area_init_memoryless_node(int nid)
7694 free_area_init_node(nid);
7697 #if MAX_NUMNODES > 1
7699 * Figure out the number of possible node ids.
7701 void __init setup_nr_node_ids(void)
7703 unsigned int highest;
7705 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7706 nr_node_ids = highest + 1;
7711 * node_map_pfn_alignment - determine the maximum internode alignment
7713 * This function should be called after node map is populated and sorted.
7714 * It calculates the maximum power of two alignment which can distinguish
7717 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7718 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7719 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7720 * shifted, 1GiB is enough and this function will indicate so.
7722 * This is used to test whether pfn -> nid mapping of the chosen memory
7723 * model has fine enough granularity to avoid incorrect mapping for the
7724 * populated node map.
7726 * Return: the determined alignment in pfn's. 0 if there is no alignment
7727 * requirement (single node).
7729 unsigned long __init node_map_pfn_alignment(void)
7731 unsigned long accl_mask = 0, last_end = 0;
7732 unsigned long start, end, mask;
7733 int last_nid = NUMA_NO_NODE;
7736 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7737 if (!start || last_nid < 0 || last_nid == nid) {
7744 * Start with a mask granular enough to pin-point to the
7745 * start pfn and tick off bits one-by-one until it becomes
7746 * too coarse to separate the current node from the last.
7748 mask = ~((1 << __ffs(start)) - 1);
7749 while (mask && last_end <= (start & (mask << 1)))
7752 /* accumulate all internode masks */
7756 /* convert mask to number of pages */
7757 return ~accl_mask + 1;
7761 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7763 * Return: the minimum PFN based on information provided via
7764 * memblock_set_node().
7766 unsigned long __init find_min_pfn_with_active_regions(void)
7768 return PHYS_PFN(memblock_start_of_DRAM());
7772 * early_calculate_totalpages()
7773 * Sum pages in active regions for movable zone.
7774 * Populate N_MEMORY for calculating usable_nodes.
7776 static unsigned long __init early_calculate_totalpages(void)
7778 unsigned long totalpages = 0;
7779 unsigned long start_pfn, end_pfn;
7782 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7783 unsigned long pages = end_pfn - start_pfn;
7785 totalpages += pages;
7787 node_set_state(nid, N_MEMORY);
7793 * Find the PFN the Movable zone begins in each node. Kernel memory
7794 * is spread evenly between nodes as long as the nodes have enough
7795 * memory. When they don't, some nodes will have more kernelcore than
7798 static void __init find_zone_movable_pfns_for_nodes(void)
7801 unsigned long usable_startpfn;
7802 unsigned long kernelcore_node, kernelcore_remaining;
7803 /* save the state before borrow the nodemask */
7804 nodemask_t saved_node_state = node_states[N_MEMORY];
7805 unsigned long totalpages = early_calculate_totalpages();
7806 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7807 struct memblock_region *r;
7809 /* Need to find movable_zone earlier when movable_node is specified. */
7810 find_usable_zone_for_movable();
7813 * If movable_node is specified, ignore kernelcore and movablecore
7816 if (movable_node_is_enabled()) {
7817 for_each_mem_region(r) {
7818 if (!memblock_is_hotpluggable(r))
7821 nid = memblock_get_region_node(r);
7823 usable_startpfn = PFN_DOWN(r->base);
7824 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7825 min(usable_startpfn, zone_movable_pfn[nid]) :
7833 * If kernelcore=mirror is specified, ignore movablecore option
7835 if (mirrored_kernelcore) {
7836 bool mem_below_4gb_not_mirrored = false;
7838 for_each_mem_region(r) {
7839 if (memblock_is_mirror(r))
7842 nid = memblock_get_region_node(r);
7844 usable_startpfn = memblock_region_memory_base_pfn(r);
7846 if (usable_startpfn < 0x100000) {
7847 mem_below_4gb_not_mirrored = true;
7851 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7852 min(usable_startpfn, zone_movable_pfn[nid]) :
7856 if (mem_below_4gb_not_mirrored)
7857 pr_warn("This configuration results in unmirrored kernel memory.\n");
7863 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7864 * amount of necessary memory.
7866 if (required_kernelcore_percent)
7867 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7869 if (required_movablecore_percent)
7870 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7874 * If movablecore= was specified, calculate what size of
7875 * kernelcore that corresponds so that memory usable for
7876 * any allocation type is evenly spread. If both kernelcore
7877 * and movablecore are specified, then the value of kernelcore
7878 * will be used for required_kernelcore if it's greater than
7879 * what movablecore would have allowed.
7881 if (required_movablecore) {
7882 unsigned long corepages;
7885 * Round-up so that ZONE_MOVABLE is at least as large as what
7886 * was requested by the user
7888 required_movablecore =
7889 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7890 required_movablecore = min(totalpages, required_movablecore);
7891 corepages = totalpages - required_movablecore;
7893 required_kernelcore = max(required_kernelcore, corepages);
7897 * If kernelcore was not specified or kernelcore size is larger
7898 * than totalpages, there is no ZONE_MOVABLE.
7900 if (!required_kernelcore || required_kernelcore >= totalpages)
7903 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7904 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7907 /* Spread kernelcore memory as evenly as possible throughout nodes */
7908 kernelcore_node = required_kernelcore / usable_nodes;
7909 for_each_node_state(nid, N_MEMORY) {
7910 unsigned long start_pfn, end_pfn;
7913 * Recalculate kernelcore_node if the division per node
7914 * now exceeds what is necessary to satisfy the requested
7915 * amount of memory for the kernel
7917 if (required_kernelcore < kernelcore_node)
7918 kernelcore_node = required_kernelcore / usable_nodes;
7921 * As the map is walked, we track how much memory is usable
7922 * by the kernel using kernelcore_remaining. When it is
7923 * 0, the rest of the node is usable by ZONE_MOVABLE
7925 kernelcore_remaining = kernelcore_node;
7927 /* Go through each range of PFNs within this node */
7928 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7929 unsigned long size_pages;
7931 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7932 if (start_pfn >= end_pfn)
7935 /* Account for what is only usable for kernelcore */
7936 if (start_pfn < usable_startpfn) {
7937 unsigned long kernel_pages;
7938 kernel_pages = min(end_pfn, usable_startpfn)
7941 kernelcore_remaining -= min(kernel_pages,
7942 kernelcore_remaining);
7943 required_kernelcore -= min(kernel_pages,
7944 required_kernelcore);
7946 /* Continue if range is now fully accounted */
7947 if (end_pfn <= usable_startpfn) {
7950 * Push zone_movable_pfn to the end so
7951 * that if we have to rebalance
7952 * kernelcore across nodes, we will
7953 * not double account here
7955 zone_movable_pfn[nid] = end_pfn;
7958 start_pfn = usable_startpfn;
7962 * The usable PFN range for ZONE_MOVABLE is from
7963 * start_pfn->end_pfn. Calculate size_pages as the
7964 * number of pages used as kernelcore
7966 size_pages = end_pfn - start_pfn;
7967 if (size_pages > kernelcore_remaining)
7968 size_pages = kernelcore_remaining;
7969 zone_movable_pfn[nid] = start_pfn + size_pages;
7972 * Some kernelcore has been met, update counts and
7973 * break if the kernelcore for this node has been
7976 required_kernelcore -= min(required_kernelcore,
7978 kernelcore_remaining -= size_pages;
7979 if (!kernelcore_remaining)
7985 * If there is still required_kernelcore, we do another pass with one
7986 * less node in the count. This will push zone_movable_pfn[nid] further
7987 * along on the nodes that still have memory until kernelcore is
7991 if (usable_nodes && required_kernelcore > usable_nodes)
7995 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7996 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7997 unsigned long start_pfn, end_pfn;
7999 zone_movable_pfn[nid] =
8000 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
8002 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
8003 if (zone_movable_pfn[nid] >= end_pfn)
8004 zone_movable_pfn[nid] = 0;
8008 /* restore the node_state */
8009 node_states[N_MEMORY] = saved_node_state;
8012 /* Any regular or high memory on that node ? */
8013 static void check_for_memory(pg_data_t *pgdat, int nid)
8015 enum zone_type zone_type;
8017 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8018 struct zone *zone = &pgdat->node_zones[zone_type];
8019 if (populated_zone(zone)) {
8020 if (IS_ENABLED(CONFIG_HIGHMEM))
8021 node_set_state(nid, N_HIGH_MEMORY);
8022 if (zone_type <= ZONE_NORMAL)
8023 node_set_state(nid, N_NORMAL_MEMORY);
8030 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8031 * such cases we allow max_zone_pfn sorted in the descending order
8033 bool __weak arch_has_descending_max_zone_pfns(void)
8039 * free_area_init - Initialise all pg_data_t and zone data
8040 * @max_zone_pfn: an array of max PFNs for each zone
8042 * This will call free_area_init_node() for each active node in the system.
8043 * Using the page ranges provided by memblock_set_node(), the size of each
8044 * zone in each node and their holes is calculated. If the maximum PFN
8045 * between two adjacent zones match, it is assumed that the zone is empty.
8046 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8047 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8048 * starts where the previous one ended. For example, ZONE_DMA32 starts
8049 * at arch_max_dma_pfn.
8051 void __init free_area_init(unsigned long *max_zone_pfn)
8053 unsigned long start_pfn, end_pfn;
8057 /* Record where the zone boundaries are */
8058 memset(arch_zone_lowest_possible_pfn, 0,
8059 sizeof(arch_zone_lowest_possible_pfn));
8060 memset(arch_zone_highest_possible_pfn, 0,
8061 sizeof(arch_zone_highest_possible_pfn));
8063 start_pfn = find_min_pfn_with_active_regions();
8064 descending = arch_has_descending_max_zone_pfns();
8066 for (i = 0; i < MAX_NR_ZONES; i++) {
8068 zone = MAX_NR_ZONES - i - 1;
8072 if (zone == ZONE_MOVABLE)
8075 end_pfn = max(max_zone_pfn[zone], start_pfn);
8076 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8077 arch_zone_highest_possible_pfn[zone] = end_pfn;
8079 start_pfn = end_pfn;
8082 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8083 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8084 find_zone_movable_pfns_for_nodes();
8086 /* Print out the zone ranges */
8087 pr_info("Zone ranges:\n");
8088 for (i = 0; i < MAX_NR_ZONES; i++) {
8089 if (i == ZONE_MOVABLE)
8091 pr_info(" %-8s ", zone_names[i]);
8092 if (arch_zone_lowest_possible_pfn[i] ==
8093 arch_zone_highest_possible_pfn[i])
8096 pr_cont("[mem %#018Lx-%#018Lx]\n",
8097 (u64)arch_zone_lowest_possible_pfn[i]
8099 ((u64)arch_zone_highest_possible_pfn[i]
8100 << PAGE_SHIFT) - 1);
8103 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8104 pr_info("Movable zone start for each node\n");
8105 for (i = 0; i < MAX_NUMNODES; i++) {
8106 if (zone_movable_pfn[i])
8107 pr_info(" Node %d: %#018Lx\n", i,
8108 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8112 * Print out the early node map, and initialize the
8113 * subsection-map relative to active online memory ranges to
8114 * enable future "sub-section" extensions of the memory map.
8116 pr_info("Early memory node ranges\n");
8117 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8118 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8119 (u64)start_pfn << PAGE_SHIFT,
8120 ((u64)end_pfn << PAGE_SHIFT) - 1);
8121 subsection_map_init(start_pfn, end_pfn - start_pfn);
8124 /* Initialise every node */
8125 mminit_verify_pageflags_layout();
8126 setup_nr_node_ids();
8127 for_each_node(nid) {
8130 if (!node_online(nid)) {
8131 pr_info("Initializing node %d as memoryless\n", nid);
8133 /* Allocator not initialized yet */
8134 pgdat = arch_alloc_nodedata(nid);
8136 pr_err("Cannot allocate %zuB for node %d.\n",
8137 sizeof(*pgdat), nid);
8140 arch_refresh_nodedata(nid, pgdat);
8141 free_area_init_memoryless_node(nid);
8144 * We do not want to confuse userspace by sysfs
8145 * files/directories for node without any memory
8146 * attached to it, so this node is not marked as
8147 * N_MEMORY and not marked online so that no sysfs
8148 * hierarchy will be created via register_one_node for
8149 * it. The pgdat will get fully initialized by
8150 * hotadd_init_pgdat() when memory is hotplugged into
8156 pgdat = NODE_DATA(nid);
8157 free_area_init_node(nid);
8159 /* Any memory on that node */
8160 if (pgdat->node_present_pages)
8161 node_set_state(nid, N_MEMORY);
8162 check_for_memory(pgdat, nid);
8168 static int __init cmdline_parse_core(char *p, unsigned long *core,
8169 unsigned long *percent)
8171 unsigned long long coremem;
8177 /* Value may be a percentage of total memory, otherwise bytes */
8178 coremem = simple_strtoull(p, &endptr, 0);
8179 if (*endptr == '%') {
8180 /* Paranoid check for percent values greater than 100 */
8181 WARN_ON(coremem > 100);
8185 coremem = memparse(p, &p);
8186 /* Paranoid check that UL is enough for the coremem value */
8187 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8189 *core = coremem >> PAGE_SHIFT;
8196 * kernelcore=size sets the amount of memory for use for allocations that
8197 * cannot be reclaimed or migrated.
8199 static int __init cmdline_parse_kernelcore(char *p)
8201 /* parse kernelcore=mirror */
8202 if (parse_option_str(p, "mirror")) {
8203 mirrored_kernelcore = true;
8207 return cmdline_parse_core(p, &required_kernelcore,
8208 &required_kernelcore_percent);
8212 * movablecore=size sets the amount of memory for use for allocations that
8213 * can be reclaimed or migrated.
8215 static int __init cmdline_parse_movablecore(char *p)
8217 return cmdline_parse_core(p, &required_movablecore,
8218 &required_movablecore_percent);
8221 early_param("kernelcore", cmdline_parse_kernelcore);
8222 early_param("movablecore", cmdline_parse_movablecore);
8224 void adjust_managed_page_count(struct page *page, long count)
8226 atomic_long_add(count, &page_zone(page)->managed_pages);
8227 totalram_pages_add(count);
8228 #ifdef CONFIG_HIGHMEM
8229 if (PageHighMem(page))
8230 totalhigh_pages_add(count);
8233 EXPORT_SYMBOL(adjust_managed_page_count);
8235 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8238 unsigned long pages = 0;
8240 start = (void *)PAGE_ALIGN((unsigned long)start);
8241 end = (void *)((unsigned long)end & PAGE_MASK);
8242 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8243 struct page *page = virt_to_page(pos);
8244 void *direct_map_addr;
8247 * 'direct_map_addr' might be different from 'pos'
8248 * because some architectures' virt_to_page()
8249 * work with aliases. Getting the direct map
8250 * address ensures that we get a _writeable_
8251 * alias for the memset().
8253 direct_map_addr = page_address(page);
8255 * Perform a kasan-unchecked memset() since this memory
8256 * has not been initialized.
8258 direct_map_addr = kasan_reset_tag(direct_map_addr);
8259 if ((unsigned int)poison <= 0xFF)
8260 memset(direct_map_addr, poison, PAGE_SIZE);
8262 free_reserved_page(page);
8266 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8271 void __init mem_init_print_info(void)
8273 unsigned long physpages, codesize, datasize, rosize, bss_size;
8274 unsigned long init_code_size, init_data_size;
8276 physpages = get_num_physpages();
8277 codesize = _etext - _stext;
8278 datasize = _edata - _sdata;
8279 rosize = __end_rodata - __start_rodata;
8280 bss_size = __bss_stop - __bss_start;
8281 init_data_size = __init_end - __init_begin;
8282 init_code_size = _einittext - _sinittext;
8285 * Detect special cases and adjust section sizes accordingly:
8286 * 1) .init.* may be embedded into .data sections
8287 * 2) .init.text.* may be out of [__init_begin, __init_end],
8288 * please refer to arch/tile/kernel/vmlinux.lds.S.
8289 * 3) .rodata.* may be embedded into .text or .data sections.
8291 #define adj_init_size(start, end, size, pos, adj) \
8293 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8297 adj_init_size(__init_begin, __init_end, init_data_size,
8298 _sinittext, init_code_size);
8299 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8300 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8301 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8302 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8304 #undef adj_init_size
8306 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8307 #ifdef CONFIG_HIGHMEM
8311 K(nr_free_pages()), K(physpages),
8312 codesize >> 10, datasize >> 10, rosize >> 10,
8313 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8314 K(physpages - totalram_pages() - totalcma_pages),
8316 #ifdef CONFIG_HIGHMEM
8317 , K(totalhigh_pages())
8323 * set_dma_reserve - set the specified number of pages reserved in the first zone
8324 * @new_dma_reserve: The number of pages to mark reserved
8326 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8327 * In the DMA zone, a significant percentage may be consumed by kernel image
8328 * and other unfreeable allocations which can skew the watermarks badly. This
8329 * function may optionally be used to account for unfreeable pages in the
8330 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8331 * smaller per-cpu batchsize.
8333 void __init set_dma_reserve(unsigned long new_dma_reserve)
8335 dma_reserve = new_dma_reserve;
8338 static int page_alloc_cpu_dead(unsigned int cpu)
8342 lru_add_drain_cpu(cpu);
8346 * Spill the event counters of the dead processor
8347 * into the current processors event counters.
8348 * This artificially elevates the count of the current
8351 vm_events_fold_cpu(cpu);
8354 * Zero the differential counters of the dead processor
8355 * so that the vm statistics are consistent.
8357 * This is only okay since the processor is dead and cannot
8358 * race with what we are doing.
8360 cpu_vm_stats_fold(cpu);
8362 for_each_populated_zone(zone)
8363 zone_pcp_update(zone, 0);
8368 static int page_alloc_cpu_online(unsigned int cpu)
8372 for_each_populated_zone(zone)
8373 zone_pcp_update(zone, 1);
8378 int hashdist = HASHDIST_DEFAULT;
8380 static int __init set_hashdist(char *str)
8384 hashdist = simple_strtoul(str, &str, 0);
8387 __setup("hashdist=", set_hashdist);
8390 void __init page_alloc_init(void)
8395 if (num_node_state(N_MEMORY) == 1)
8399 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8400 "mm/page_alloc:pcp",
8401 page_alloc_cpu_online,
8402 page_alloc_cpu_dead);
8407 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8408 * or min_free_kbytes changes.
8410 static void calculate_totalreserve_pages(void)
8412 struct pglist_data *pgdat;
8413 unsigned long reserve_pages = 0;
8414 enum zone_type i, j;
8416 for_each_online_pgdat(pgdat) {
8418 pgdat->totalreserve_pages = 0;
8420 for (i = 0; i < MAX_NR_ZONES; i++) {
8421 struct zone *zone = pgdat->node_zones + i;
8423 unsigned long managed_pages = zone_managed_pages(zone);
8425 /* Find valid and maximum lowmem_reserve in the zone */
8426 for (j = i; j < MAX_NR_ZONES; j++) {
8427 if (zone->lowmem_reserve[j] > max)
8428 max = zone->lowmem_reserve[j];
8431 /* we treat the high watermark as reserved pages. */
8432 max += high_wmark_pages(zone);
8434 if (max > managed_pages)
8435 max = managed_pages;
8437 pgdat->totalreserve_pages += max;
8439 reserve_pages += max;
8442 totalreserve_pages = reserve_pages;
8446 * setup_per_zone_lowmem_reserve - called whenever
8447 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8448 * has a correct pages reserved value, so an adequate number of
8449 * pages are left in the zone after a successful __alloc_pages().
8451 static void setup_per_zone_lowmem_reserve(void)
8453 struct pglist_data *pgdat;
8454 enum zone_type i, j;
8456 for_each_online_pgdat(pgdat) {
8457 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8458 struct zone *zone = &pgdat->node_zones[i];
8459 int ratio = sysctl_lowmem_reserve_ratio[i];
8460 bool clear = !ratio || !zone_managed_pages(zone);
8461 unsigned long managed_pages = 0;
8463 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8464 struct zone *upper_zone = &pgdat->node_zones[j];
8466 managed_pages += zone_managed_pages(upper_zone);
8469 zone->lowmem_reserve[j] = 0;
8471 zone->lowmem_reserve[j] = managed_pages / ratio;
8476 /* update totalreserve_pages */
8477 calculate_totalreserve_pages();
8480 static void __setup_per_zone_wmarks(void)
8482 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8483 unsigned long lowmem_pages = 0;
8485 unsigned long flags;
8487 /* Calculate total number of !ZONE_HIGHMEM pages */
8488 for_each_zone(zone) {
8489 if (!is_highmem(zone))
8490 lowmem_pages += zone_managed_pages(zone);
8493 for_each_zone(zone) {
8496 spin_lock_irqsave(&zone->lock, flags);
8497 tmp = (u64)pages_min * zone_managed_pages(zone);
8498 do_div(tmp, lowmem_pages);
8499 if (is_highmem(zone)) {
8501 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8502 * need highmem pages, so cap pages_min to a small
8505 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8506 * deltas control async page reclaim, and so should
8507 * not be capped for highmem.
8509 unsigned long min_pages;
8511 min_pages = zone_managed_pages(zone) / 1024;
8512 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8513 zone->_watermark[WMARK_MIN] = min_pages;
8516 * If it's a lowmem zone, reserve a number of pages
8517 * proportionate to the zone's size.
8519 zone->_watermark[WMARK_MIN] = tmp;
8523 * Set the kswapd watermarks distance according to the
8524 * scale factor in proportion to available memory, but
8525 * ensure a minimum size on small systems.
8527 tmp = max_t(u64, tmp >> 2,
8528 mult_frac(zone_managed_pages(zone),
8529 watermark_scale_factor, 10000));
8531 zone->watermark_boost = 0;
8532 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8533 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8534 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8536 spin_unlock_irqrestore(&zone->lock, flags);
8539 /* update totalreserve_pages */
8540 calculate_totalreserve_pages();
8544 * setup_per_zone_wmarks - called when min_free_kbytes changes
8545 * or when memory is hot-{added|removed}
8547 * Ensures that the watermark[min,low,high] values for each zone are set
8548 * correctly with respect to min_free_kbytes.
8550 void setup_per_zone_wmarks(void)
8553 static DEFINE_SPINLOCK(lock);
8556 __setup_per_zone_wmarks();
8560 * The watermark size have changed so update the pcpu batch
8561 * and high limits or the limits may be inappropriate.
8564 zone_pcp_update(zone, 0);
8568 * Initialise min_free_kbytes.
8570 * For small machines we want it small (128k min). For large machines
8571 * we want it large (256MB max). But it is not linear, because network
8572 * bandwidth does not increase linearly with machine size. We use
8574 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8575 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8591 void calculate_min_free_kbytes(void)
8593 unsigned long lowmem_kbytes;
8594 int new_min_free_kbytes;
8596 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8597 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8599 if (new_min_free_kbytes > user_min_free_kbytes)
8600 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8602 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8603 new_min_free_kbytes, user_min_free_kbytes);
8607 int __meminit init_per_zone_wmark_min(void)
8609 calculate_min_free_kbytes();
8610 setup_per_zone_wmarks();
8611 refresh_zone_stat_thresholds();
8612 setup_per_zone_lowmem_reserve();
8615 setup_min_unmapped_ratio();
8616 setup_min_slab_ratio();
8619 khugepaged_min_free_kbytes_update();
8623 postcore_initcall(init_per_zone_wmark_min)
8626 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8627 * that we can call two helper functions whenever min_free_kbytes
8630 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8631 void *buffer, size_t *length, loff_t *ppos)
8635 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8640 user_min_free_kbytes = min_free_kbytes;
8641 setup_per_zone_wmarks();
8646 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8647 void *buffer, size_t *length, loff_t *ppos)
8651 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8656 setup_per_zone_wmarks();
8662 static void setup_min_unmapped_ratio(void)
8667 for_each_online_pgdat(pgdat)
8668 pgdat->min_unmapped_pages = 0;
8671 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8672 sysctl_min_unmapped_ratio) / 100;
8676 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8677 void *buffer, size_t *length, loff_t *ppos)
8681 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8685 setup_min_unmapped_ratio();
8690 static void setup_min_slab_ratio(void)
8695 for_each_online_pgdat(pgdat)
8696 pgdat->min_slab_pages = 0;
8699 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8700 sysctl_min_slab_ratio) / 100;
8703 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8704 void *buffer, size_t *length, loff_t *ppos)
8708 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8712 setup_min_slab_ratio();
8719 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8720 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8721 * whenever sysctl_lowmem_reserve_ratio changes.
8723 * The reserve ratio obviously has absolutely no relation with the
8724 * minimum watermarks. The lowmem reserve ratio can only make sense
8725 * if in function of the boot time zone sizes.
8727 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8728 void *buffer, size_t *length, loff_t *ppos)
8732 proc_dointvec_minmax(table, write, buffer, length, ppos);
8734 for (i = 0; i < MAX_NR_ZONES; i++) {
8735 if (sysctl_lowmem_reserve_ratio[i] < 1)
8736 sysctl_lowmem_reserve_ratio[i] = 0;
8739 setup_per_zone_lowmem_reserve();
8744 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8745 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8746 * pagelist can have before it gets flushed back to buddy allocator.
8748 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8749 int write, void *buffer, size_t *length, loff_t *ppos)
8752 int old_percpu_pagelist_high_fraction;
8755 mutex_lock(&pcp_batch_high_lock);
8756 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8758 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8759 if (!write || ret < 0)
8762 /* Sanity checking to avoid pcp imbalance */
8763 if (percpu_pagelist_high_fraction &&
8764 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8765 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8771 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8774 for_each_populated_zone(zone)
8775 zone_set_pageset_high_and_batch(zone, 0);
8777 mutex_unlock(&pcp_batch_high_lock);
8781 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8783 * Returns the number of pages that arch has reserved but
8784 * is not known to alloc_large_system_hash().
8786 static unsigned long __init arch_reserved_kernel_pages(void)
8793 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8794 * machines. As memory size is increased the scale is also increased but at
8795 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8796 * quadruples the scale is increased by one, which means the size of hash table
8797 * only doubles, instead of quadrupling as well.
8798 * Because 32-bit systems cannot have large physical memory, where this scaling
8799 * makes sense, it is disabled on such platforms.
8801 #if __BITS_PER_LONG > 32
8802 #define ADAPT_SCALE_BASE (64ul << 30)
8803 #define ADAPT_SCALE_SHIFT 2
8804 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8808 * allocate a large system hash table from bootmem
8809 * - it is assumed that the hash table must contain an exact power-of-2
8810 * quantity of entries
8811 * - limit is the number of hash buckets, not the total allocation size
8813 void *__init alloc_large_system_hash(const char *tablename,
8814 unsigned long bucketsize,
8815 unsigned long numentries,
8818 unsigned int *_hash_shift,
8819 unsigned int *_hash_mask,
8820 unsigned long low_limit,
8821 unsigned long high_limit)
8823 unsigned long long max = high_limit;
8824 unsigned long log2qty, size;
8830 /* allow the kernel cmdline to have a say */
8832 /* round applicable memory size up to nearest megabyte */
8833 numentries = nr_kernel_pages;
8834 numentries -= arch_reserved_kernel_pages();
8836 /* It isn't necessary when PAGE_SIZE >= 1MB */
8837 if (PAGE_SHIFT < 20)
8838 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8840 #if __BITS_PER_LONG > 32
8842 unsigned long adapt;
8844 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8845 adapt <<= ADAPT_SCALE_SHIFT)
8850 /* limit to 1 bucket per 2^scale bytes of low memory */
8851 if (scale > PAGE_SHIFT)
8852 numentries >>= (scale - PAGE_SHIFT);
8854 numentries <<= (PAGE_SHIFT - scale);
8856 /* Make sure we've got at least a 0-order allocation.. */
8857 if (unlikely(flags & HASH_SMALL)) {
8858 /* Makes no sense without HASH_EARLY */
8859 WARN_ON(!(flags & HASH_EARLY));
8860 if (!(numentries >> *_hash_shift)) {
8861 numentries = 1UL << *_hash_shift;
8862 BUG_ON(!numentries);
8864 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8865 numentries = PAGE_SIZE / bucketsize;
8867 numentries = roundup_pow_of_two(numentries);
8869 /* limit allocation size to 1/16 total memory by default */
8871 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8872 do_div(max, bucketsize);
8874 max = min(max, 0x80000000ULL);
8876 if (numentries < low_limit)
8877 numentries = low_limit;
8878 if (numentries > max)
8881 log2qty = ilog2(numentries);
8883 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8886 size = bucketsize << log2qty;
8887 if (flags & HASH_EARLY) {
8888 if (flags & HASH_ZERO)
8889 table = memblock_alloc(size, SMP_CACHE_BYTES);
8891 table = memblock_alloc_raw(size,
8893 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8894 table = __vmalloc(size, gfp_flags);
8897 huge = is_vm_area_hugepages(table);
8900 * If bucketsize is not a power-of-two, we may free
8901 * some pages at the end of hash table which
8902 * alloc_pages_exact() automatically does
8904 table = alloc_pages_exact(size, gfp_flags);
8905 kmemleak_alloc(table, size, 1, gfp_flags);
8907 } while (!table && size > PAGE_SIZE && --log2qty);
8910 panic("Failed to allocate %s hash table\n", tablename);
8912 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8913 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8914 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8917 *_hash_shift = log2qty;
8919 *_hash_mask = (1 << log2qty) - 1;
8925 * This function checks whether pageblock includes unmovable pages or not.
8927 * PageLRU check without isolation or lru_lock could race so that
8928 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8929 * check without lock_page also may miss some movable non-lru pages at
8930 * race condition. So you can't expect this function should be exact.
8932 * Returns a page without holding a reference. If the caller wants to
8933 * dereference that page (e.g., dumping), it has to make sure that it
8934 * cannot get removed (e.g., via memory unplug) concurrently.
8937 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8938 int migratetype, int flags)
8940 unsigned long iter = 0;
8941 unsigned long pfn = page_to_pfn(page);
8942 unsigned long offset = pfn % pageblock_nr_pages;
8944 if (is_migrate_cma_page(page)) {
8946 * CMA allocations (alloc_contig_range) really need to mark
8947 * isolate CMA pageblocks even when they are not movable in fact
8948 * so consider them movable here.
8950 if (is_migrate_cma(migratetype))
8956 for (; iter < pageblock_nr_pages - offset; iter++) {
8957 page = pfn_to_page(pfn + iter);
8960 * Both, bootmem allocations and memory holes are marked
8961 * PG_reserved and are unmovable. We can even have unmovable
8962 * allocations inside ZONE_MOVABLE, for example when
8963 * specifying "movablecore".
8965 if (PageReserved(page))
8969 * If the zone is movable and we have ruled out all reserved
8970 * pages then it should be reasonably safe to assume the rest
8973 if (zone_idx(zone) == ZONE_MOVABLE)
8977 * Hugepages are not in LRU lists, but they're movable.
8978 * THPs are on the LRU, but need to be counted as #small pages.
8979 * We need not scan over tail pages because we don't
8980 * handle each tail page individually in migration.
8982 if (PageHuge(page) || PageTransCompound(page)) {
8983 struct page *head = compound_head(page);
8984 unsigned int skip_pages;
8986 if (PageHuge(page)) {
8987 if (!hugepage_migration_supported(page_hstate(head)))
8989 } else if (!PageLRU(head) && !__PageMovable(head)) {
8993 skip_pages = compound_nr(head) - (page - head);
8994 iter += skip_pages - 1;
8999 * We can't use page_count without pin a page
9000 * because another CPU can free compound page.
9001 * This check already skips compound tails of THP
9002 * because their page->_refcount is zero at all time.
9004 if (!page_ref_count(page)) {
9005 if (PageBuddy(page))
9006 iter += (1 << buddy_order(page)) - 1;
9011 * The HWPoisoned page may be not in buddy system, and
9012 * page_count() is not 0.
9014 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
9018 * We treat all PageOffline() pages as movable when offlining
9019 * to give drivers a chance to decrement their reference count
9020 * in MEM_GOING_OFFLINE in order to indicate that these pages
9021 * can be offlined as there are no direct references anymore.
9022 * For actually unmovable PageOffline() where the driver does
9023 * not support this, we will fail later when trying to actually
9024 * move these pages that still have a reference count > 0.
9025 * (false negatives in this function only)
9027 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
9030 if (__PageMovable(page) || PageLRU(page))
9034 * If there are RECLAIMABLE pages, we need to check
9035 * it. But now, memory offline itself doesn't call
9036 * shrink_node_slabs() and it still to be fixed.
9043 #ifdef CONFIG_CONTIG_ALLOC
9044 static unsigned long pfn_max_align_down(unsigned long pfn)
9046 return ALIGN_DOWN(pfn, MAX_ORDER_NR_PAGES);
9049 static unsigned long pfn_max_align_up(unsigned long pfn)
9051 return ALIGN(pfn, MAX_ORDER_NR_PAGES);
9054 #if defined(CONFIG_DYNAMIC_DEBUG) || \
9055 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
9056 /* Usage: See admin-guide/dynamic-debug-howto.rst */
9057 static void alloc_contig_dump_pages(struct list_head *page_list)
9059 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9061 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9065 list_for_each_entry(page, page_list, lru)
9066 dump_page(page, "migration failure");
9070 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9075 /* [start, end) must belong to a single zone. */
9076 static int __alloc_contig_migrate_range(struct compact_control *cc,
9077 unsigned long start, unsigned long end)
9079 /* This function is based on compact_zone() from compaction.c. */
9080 unsigned int nr_reclaimed;
9081 unsigned long pfn = start;
9082 unsigned int tries = 0;
9084 struct migration_target_control mtc = {
9085 .nid = zone_to_nid(cc->zone),
9086 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9089 lru_cache_disable();
9091 while (pfn < end || !list_empty(&cc->migratepages)) {
9092 if (fatal_signal_pending(current)) {
9097 if (list_empty(&cc->migratepages)) {
9098 cc->nr_migratepages = 0;
9099 ret = isolate_migratepages_range(cc, pfn, end);
9100 if (ret && ret != -EAGAIN)
9102 pfn = cc->migrate_pfn;
9104 } else if (++tries == 5) {
9109 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9111 cc->nr_migratepages -= nr_reclaimed;
9113 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9114 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9117 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9118 * to retry again over this error, so do the same here.
9127 alloc_contig_dump_pages(&cc->migratepages);
9128 putback_movable_pages(&cc->migratepages);
9135 * alloc_contig_range() -- tries to allocate given range of pages
9136 * @start: start PFN to allocate
9137 * @end: one-past-the-last PFN to allocate
9138 * @migratetype: migratetype of the underlying pageblocks (either
9139 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9140 * in range must have the same migratetype and it must
9141 * be either of the two.
9142 * @gfp_mask: GFP mask to use during compaction
9144 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9145 * aligned. The PFN range must belong to a single zone.
9147 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9148 * pageblocks in the range. Once isolated, the pageblocks should not
9149 * be modified by others.
9151 * Return: zero on success or negative error code. On success all
9152 * pages which PFN is in [start, end) are allocated for the caller and
9153 * need to be freed with free_contig_range().
9155 int alloc_contig_range(unsigned long start, unsigned long end,
9156 unsigned migratetype, gfp_t gfp_mask)
9158 unsigned long outer_start, outer_end;
9162 struct compact_control cc = {
9163 .nr_migratepages = 0,
9165 .zone = page_zone(pfn_to_page(start)),
9166 .mode = MIGRATE_SYNC,
9167 .ignore_skip_hint = true,
9168 .no_set_skip_hint = true,
9169 .gfp_mask = current_gfp_context(gfp_mask),
9170 .alloc_contig = true,
9172 INIT_LIST_HEAD(&cc.migratepages);
9175 * What we do here is we mark all pageblocks in range as
9176 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9177 * have different sizes, and due to the way page allocator
9178 * work, we align the range to biggest of the two pages so
9179 * that page allocator won't try to merge buddies from
9180 * different pageblocks and change MIGRATE_ISOLATE to some
9181 * other migration type.
9183 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9184 * migrate the pages from an unaligned range (ie. pages that
9185 * we are interested in). This will put all the pages in
9186 * range back to page allocator as MIGRATE_ISOLATE.
9188 * When this is done, we take the pages in range from page
9189 * allocator removing them from the buddy system. This way
9190 * page allocator will never consider using them.
9192 * This lets us mark the pageblocks back as
9193 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9194 * aligned range but not in the unaligned, original range are
9195 * put back to page allocator so that buddy can use them.
9198 ret = start_isolate_page_range(pfn_max_align_down(start),
9199 pfn_max_align_up(end), migratetype, 0);
9203 drain_all_pages(cc.zone);
9206 * In case of -EBUSY, we'd like to know which page causes problem.
9207 * So, just fall through. test_pages_isolated() has a tracepoint
9208 * which will report the busy page.
9210 * It is possible that busy pages could become available before
9211 * the call to test_pages_isolated, and the range will actually be
9212 * allocated. So, if we fall through be sure to clear ret so that
9213 * -EBUSY is not accidentally used or returned to caller.
9215 ret = __alloc_contig_migrate_range(&cc, start, end);
9216 if (ret && ret != -EBUSY)
9221 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9222 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9223 * more, all pages in [start, end) are free in page allocator.
9224 * What we are going to do is to allocate all pages from
9225 * [start, end) (that is remove them from page allocator).
9227 * The only problem is that pages at the beginning and at the
9228 * end of interesting range may be not aligned with pages that
9229 * page allocator holds, ie. they can be part of higher order
9230 * pages. Because of this, we reserve the bigger range and
9231 * once this is done free the pages we are not interested in.
9233 * We don't have to hold zone->lock here because the pages are
9234 * isolated thus they won't get removed from buddy.
9238 outer_start = start;
9239 while (!PageBuddy(pfn_to_page(outer_start))) {
9240 if (++order >= MAX_ORDER) {
9241 outer_start = start;
9244 outer_start &= ~0UL << order;
9247 if (outer_start != start) {
9248 order = buddy_order(pfn_to_page(outer_start));
9251 * outer_start page could be small order buddy page and
9252 * it doesn't include start page. Adjust outer_start
9253 * in this case to report failed page properly
9254 * on tracepoint in test_pages_isolated()
9256 if (outer_start + (1UL << order) <= start)
9257 outer_start = start;
9260 /* Make sure the range is really isolated. */
9261 if (test_pages_isolated(outer_start, end, 0)) {
9266 /* Grab isolated pages from freelists. */
9267 outer_end = isolate_freepages_range(&cc, outer_start, end);
9273 /* Free head and tail (if any) */
9274 if (start != outer_start)
9275 free_contig_range(outer_start, start - outer_start);
9276 if (end != outer_end)
9277 free_contig_range(end, outer_end - end);
9280 undo_isolate_page_range(pfn_max_align_down(start),
9281 pfn_max_align_up(end), migratetype);
9284 EXPORT_SYMBOL(alloc_contig_range);
9286 static int __alloc_contig_pages(unsigned long start_pfn,
9287 unsigned long nr_pages, gfp_t gfp_mask)
9289 unsigned long end_pfn = start_pfn + nr_pages;
9291 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9295 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9296 unsigned long nr_pages)
9298 unsigned long i, end_pfn = start_pfn + nr_pages;
9301 for (i = start_pfn; i < end_pfn; i++) {
9302 page = pfn_to_online_page(i);
9306 if (page_zone(page) != z)
9309 if (PageReserved(page))
9315 static bool zone_spans_last_pfn(const struct zone *zone,
9316 unsigned long start_pfn, unsigned long nr_pages)
9318 unsigned long last_pfn = start_pfn + nr_pages - 1;
9320 return zone_spans_pfn(zone, last_pfn);
9324 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9325 * @nr_pages: Number of contiguous pages to allocate
9326 * @gfp_mask: GFP mask to limit search and used during compaction
9328 * @nodemask: Mask for other possible nodes
9330 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9331 * on an applicable zonelist to find a contiguous pfn range which can then be
9332 * tried for allocation with alloc_contig_range(). This routine is intended
9333 * for allocation requests which can not be fulfilled with the buddy allocator.
9335 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9336 * power of two, then allocated range is also guaranteed to be aligned to same
9337 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9339 * Allocated pages can be freed with free_contig_range() or by manually calling
9340 * __free_page() on each allocated page.
9342 * Return: pointer to contiguous pages on success, or NULL if not successful.
9344 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9345 int nid, nodemask_t *nodemask)
9347 unsigned long ret, pfn, flags;
9348 struct zonelist *zonelist;
9352 zonelist = node_zonelist(nid, gfp_mask);
9353 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9354 gfp_zone(gfp_mask), nodemask) {
9355 spin_lock_irqsave(&zone->lock, flags);
9357 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9358 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9359 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9361 * We release the zone lock here because
9362 * alloc_contig_range() will also lock the zone
9363 * at some point. If there's an allocation
9364 * spinning on this lock, it may win the race
9365 * and cause alloc_contig_range() to fail...
9367 spin_unlock_irqrestore(&zone->lock, flags);
9368 ret = __alloc_contig_pages(pfn, nr_pages,
9371 return pfn_to_page(pfn);
9372 spin_lock_irqsave(&zone->lock, flags);
9376 spin_unlock_irqrestore(&zone->lock, flags);
9380 #endif /* CONFIG_CONTIG_ALLOC */
9382 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9384 unsigned long count = 0;
9386 for (; nr_pages--; pfn++) {
9387 struct page *page = pfn_to_page(pfn);
9389 count += page_count(page) != 1;
9392 WARN(count != 0, "%lu pages are still in use!\n", count);
9394 EXPORT_SYMBOL(free_contig_range);
9397 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9398 * page high values need to be recalculated.
9400 void zone_pcp_update(struct zone *zone, int cpu_online)
9402 mutex_lock(&pcp_batch_high_lock);
9403 zone_set_pageset_high_and_batch(zone, cpu_online);
9404 mutex_unlock(&pcp_batch_high_lock);
9408 * Effectively disable pcplists for the zone by setting the high limit to 0
9409 * and draining all cpus. A concurrent page freeing on another CPU that's about
9410 * to put the page on pcplist will either finish before the drain and the page
9411 * will be drained, or observe the new high limit and skip the pcplist.
9413 * Must be paired with a call to zone_pcp_enable().
9415 void zone_pcp_disable(struct zone *zone)
9417 mutex_lock(&pcp_batch_high_lock);
9418 __zone_set_pageset_high_and_batch(zone, 0, 1);
9419 __drain_all_pages(zone, true);
9422 void zone_pcp_enable(struct zone *zone)
9424 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9425 mutex_unlock(&pcp_batch_high_lock);
9428 void zone_pcp_reset(struct zone *zone)
9431 struct per_cpu_zonestat *pzstats;
9433 if (zone->per_cpu_pageset != &boot_pageset) {
9434 for_each_online_cpu(cpu) {
9435 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9436 drain_zonestat(zone, pzstats);
9438 free_percpu(zone->per_cpu_pageset);
9439 free_percpu(zone->per_cpu_zonestats);
9440 zone->per_cpu_pageset = &boot_pageset;
9441 zone->per_cpu_zonestats = &boot_zonestats;
9445 #ifdef CONFIG_MEMORY_HOTREMOVE
9447 * All pages in the range must be in a single zone, must not contain holes,
9448 * must span full sections, and must be isolated before calling this function.
9450 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9452 unsigned long pfn = start_pfn;
9456 unsigned long flags;
9458 offline_mem_sections(pfn, end_pfn);
9459 zone = page_zone(pfn_to_page(pfn));
9460 spin_lock_irqsave(&zone->lock, flags);
9461 while (pfn < end_pfn) {
9462 page = pfn_to_page(pfn);
9464 * The HWPoisoned page may be not in buddy system, and
9465 * page_count() is not 0.
9467 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9472 * At this point all remaining PageOffline() pages have a
9473 * reference count of 0 and can simply be skipped.
9475 if (PageOffline(page)) {
9476 BUG_ON(page_count(page));
9477 BUG_ON(PageBuddy(page));
9482 BUG_ON(page_count(page));
9483 BUG_ON(!PageBuddy(page));
9484 order = buddy_order(page);
9485 del_page_from_free_list(page, zone, order);
9486 pfn += (1 << order);
9488 spin_unlock_irqrestore(&zone->lock, flags);
9493 * This function returns a stable result only if called under zone lock.
9495 bool is_free_buddy_page(struct page *page)
9497 unsigned long pfn = page_to_pfn(page);
9500 for (order = 0; order < MAX_ORDER; order++) {
9501 struct page *page_head = page - (pfn & ((1 << order) - 1));
9503 if (PageBuddy(page_head) &&
9504 buddy_order_unsafe(page_head) >= order)
9508 return order < MAX_ORDER;
9510 EXPORT_SYMBOL(is_free_buddy_page);
9512 #ifdef CONFIG_MEMORY_FAILURE
9514 * Break down a higher-order page in sub-pages, and keep our target out of
9517 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9518 struct page *target, int low, int high,
9521 unsigned long size = 1 << high;
9522 struct page *current_buddy, *next_page;
9524 while (high > low) {
9528 if (target >= &page[size]) {
9529 next_page = page + size;
9530 current_buddy = page;
9533 current_buddy = page + size;
9536 if (set_page_guard(zone, current_buddy, high, migratetype))
9539 if (current_buddy != target) {
9540 add_to_free_list(current_buddy, zone, high, migratetype);
9541 set_buddy_order(current_buddy, high);
9548 * Take a page that will be marked as poisoned off the buddy allocator.
9550 bool take_page_off_buddy(struct page *page)
9552 struct zone *zone = page_zone(page);
9553 unsigned long pfn = page_to_pfn(page);
9554 unsigned long flags;
9558 spin_lock_irqsave(&zone->lock, flags);
9559 for (order = 0; order < MAX_ORDER; order++) {
9560 struct page *page_head = page - (pfn & ((1 << order) - 1));
9561 int page_order = buddy_order(page_head);
9563 if (PageBuddy(page_head) && page_order >= order) {
9564 unsigned long pfn_head = page_to_pfn(page_head);
9565 int migratetype = get_pfnblock_migratetype(page_head,
9568 del_page_from_free_list(page_head, zone, page_order);
9569 break_down_buddy_pages(zone, page_head, page, 0,
9570 page_order, migratetype);
9571 SetPageHWPoisonTakenOff(page);
9572 if (!is_migrate_isolate(migratetype))
9573 __mod_zone_freepage_state(zone, -1, migratetype);
9577 if (page_count(page_head) > 0)
9580 spin_unlock_irqrestore(&zone->lock, flags);
9585 * Cancel takeoff done by take_page_off_buddy().
9587 bool put_page_back_buddy(struct page *page)
9589 struct zone *zone = page_zone(page);
9590 unsigned long pfn = page_to_pfn(page);
9591 unsigned long flags;
9592 int migratetype = get_pfnblock_migratetype(page, pfn);
9595 spin_lock_irqsave(&zone->lock, flags);
9596 if (put_page_testzero(page)) {
9597 ClearPageHWPoisonTakenOff(page);
9598 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9599 if (TestClearPageHWPoison(page)) {
9600 num_poisoned_pages_dec();
9604 spin_unlock_irqrestore(&zone->lock, flags);
9610 #ifdef CONFIG_ZONE_DMA
9611 bool has_managed_dma(void)
9613 struct pglist_data *pgdat;
9615 for_each_online_pgdat(pgdat) {
9616 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9618 if (managed_zone(zone))
9623 #endif /* CONFIG_ZONE_DMA */