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.
2370 if (kasan_has_integrated_init()) {
2371 if (gfp_flags & __GFP_SKIP_KASAN_POISON)
2372 SetPageSkipKASanPoison(page);
2377 for (i = 0; i != 1 << order; ++i)
2378 tag_clear_highpage(page + i);
2380 kasan_unpoison_pages(page, order, init);
2383 kasan_unpoison_pages(page, order, init);
2388 for (i = 0; i < 1 << order; i++)
2389 tag_clear_highpage(page + i);
2395 kernel_init_free_pages(page, 1 << order);
2398 set_page_owner(page, order, gfp_flags);
2399 page_table_check_alloc(page, order);
2402 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2403 unsigned int alloc_flags)
2405 post_alloc_hook(page, order, gfp_flags);
2407 if (order && (gfp_flags & __GFP_COMP))
2408 prep_compound_page(page, order);
2411 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2412 * allocate the page. The expectation is that the caller is taking
2413 * steps that will free more memory. The caller should avoid the page
2414 * being used for !PFMEMALLOC purposes.
2416 if (alloc_flags & ALLOC_NO_WATERMARKS)
2417 set_page_pfmemalloc(page);
2419 clear_page_pfmemalloc(page);
2423 * Go through the free lists for the given migratetype and remove
2424 * the smallest available page from the freelists
2426 static __always_inline
2427 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2430 unsigned int current_order;
2431 struct free_area *area;
2434 /* Find a page of the appropriate size in the preferred list */
2435 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2436 area = &(zone->free_area[current_order]);
2437 page = get_page_from_free_area(area, migratetype);
2440 del_page_from_free_list(page, zone, current_order);
2441 expand(zone, page, order, current_order, migratetype);
2442 set_pcppage_migratetype(page, migratetype);
2451 * This array describes the order lists are fallen back to when
2452 * the free lists for the desirable migrate type are depleted
2454 * The other migratetypes do not have fallbacks.
2456 static int fallbacks[MIGRATE_TYPES][3] = {
2457 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2458 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2459 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2463 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2466 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2469 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2470 unsigned int order) { return NULL; }
2474 * Move the free pages in a range to the freelist tail of the requested type.
2475 * Note that start_page and end_pages are not aligned on a pageblock
2476 * boundary. If alignment is required, use move_freepages_block()
2478 static int move_freepages(struct zone *zone,
2479 unsigned long start_pfn, unsigned long end_pfn,
2480 int migratetype, int *num_movable)
2485 int pages_moved = 0;
2487 for (pfn = start_pfn; pfn <= end_pfn;) {
2488 page = pfn_to_page(pfn);
2489 if (!PageBuddy(page)) {
2491 * We assume that pages that could be isolated for
2492 * migration are movable. But we don't actually try
2493 * isolating, as that would be expensive.
2496 (PageLRU(page) || __PageMovable(page)))
2502 /* Make sure we are not inadvertently changing nodes */
2503 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2504 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2506 order = buddy_order(page);
2507 move_to_free_list(page, zone, order, migratetype);
2509 pages_moved += 1 << order;
2515 int move_freepages_block(struct zone *zone, struct page *page,
2516 int migratetype, int *num_movable)
2518 unsigned long start_pfn, end_pfn, pfn;
2523 pfn = page_to_pfn(page);
2524 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2525 end_pfn = start_pfn + pageblock_nr_pages - 1;
2527 /* Do not cross zone boundaries */
2528 if (!zone_spans_pfn(zone, start_pfn))
2530 if (!zone_spans_pfn(zone, end_pfn))
2533 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2537 static void change_pageblock_range(struct page *pageblock_page,
2538 int start_order, int migratetype)
2540 int nr_pageblocks = 1 << (start_order - pageblock_order);
2542 while (nr_pageblocks--) {
2543 set_pageblock_migratetype(pageblock_page, migratetype);
2544 pageblock_page += pageblock_nr_pages;
2549 * When we are falling back to another migratetype during allocation, try to
2550 * steal extra free pages from the same pageblocks to satisfy further
2551 * allocations, instead of polluting multiple pageblocks.
2553 * If we are stealing a relatively large buddy page, it is likely there will
2554 * be more free pages in the pageblock, so try to steal them all. For
2555 * reclaimable and unmovable allocations, we steal regardless of page size,
2556 * as fragmentation caused by those allocations polluting movable pageblocks
2557 * is worse than movable allocations stealing from unmovable and reclaimable
2560 static bool can_steal_fallback(unsigned int order, int start_mt)
2563 * Leaving this order check is intended, although there is
2564 * relaxed order check in next check. The reason is that
2565 * we can actually steal whole pageblock if this condition met,
2566 * but, below check doesn't guarantee it and that is just heuristic
2567 * so could be changed anytime.
2569 if (order >= pageblock_order)
2572 if (order >= pageblock_order / 2 ||
2573 start_mt == MIGRATE_RECLAIMABLE ||
2574 start_mt == MIGRATE_UNMOVABLE ||
2575 page_group_by_mobility_disabled)
2581 static inline bool boost_watermark(struct zone *zone)
2583 unsigned long max_boost;
2585 if (!watermark_boost_factor)
2588 * Don't bother in zones that are unlikely to produce results.
2589 * On small machines, including kdump capture kernels running
2590 * in a small area, boosting the watermark can cause an out of
2591 * memory situation immediately.
2593 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2596 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2597 watermark_boost_factor, 10000);
2600 * high watermark may be uninitialised if fragmentation occurs
2601 * very early in boot so do not boost. We do not fall
2602 * through and boost by pageblock_nr_pages as failing
2603 * allocations that early means that reclaim is not going
2604 * to help and it may even be impossible to reclaim the
2605 * boosted watermark resulting in a hang.
2610 max_boost = max(pageblock_nr_pages, max_boost);
2612 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2619 * This function implements actual steal behaviour. If order is large enough,
2620 * we can steal whole pageblock. If not, we first move freepages in this
2621 * pageblock to our migratetype and determine how many already-allocated pages
2622 * are there in the pageblock with a compatible migratetype. If at least half
2623 * of pages are free or compatible, we can change migratetype of the pageblock
2624 * itself, so pages freed in the future will be put on the correct free list.
2626 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2627 unsigned int alloc_flags, int start_type, bool whole_block)
2629 unsigned int current_order = buddy_order(page);
2630 int free_pages, movable_pages, alike_pages;
2633 old_block_type = get_pageblock_migratetype(page);
2636 * This can happen due to races and we want to prevent broken
2637 * highatomic accounting.
2639 if (is_migrate_highatomic(old_block_type))
2642 /* Take ownership for orders >= pageblock_order */
2643 if (current_order >= pageblock_order) {
2644 change_pageblock_range(page, current_order, start_type);
2649 * Boost watermarks to increase reclaim pressure to reduce the
2650 * likelihood of future fallbacks. Wake kswapd now as the node
2651 * may be balanced overall and kswapd will not wake naturally.
2653 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2654 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2656 /* We are not allowed to try stealing from the whole block */
2660 free_pages = move_freepages_block(zone, page, start_type,
2663 * Determine how many pages are compatible with our allocation.
2664 * For movable allocation, it's the number of movable pages which
2665 * we just obtained. For other types it's a bit more tricky.
2667 if (start_type == MIGRATE_MOVABLE) {
2668 alike_pages = movable_pages;
2671 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2672 * to MOVABLE pageblock, consider all non-movable pages as
2673 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2674 * vice versa, be conservative since we can't distinguish the
2675 * exact migratetype of non-movable pages.
2677 if (old_block_type == MIGRATE_MOVABLE)
2678 alike_pages = pageblock_nr_pages
2679 - (free_pages + movable_pages);
2684 /* moving whole block can fail due to zone boundary conditions */
2689 * If a sufficient number of pages in the block are either free or of
2690 * comparable migratability as our allocation, claim the whole block.
2692 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2693 page_group_by_mobility_disabled)
2694 set_pageblock_migratetype(page, start_type);
2699 move_to_free_list(page, zone, current_order, start_type);
2703 * Check whether there is a suitable fallback freepage with requested order.
2704 * If only_stealable is true, this function returns fallback_mt only if
2705 * we can steal other freepages all together. This would help to reduce
2706 * fragmentation due to mixed migratetype pages in one pageblock.
2708 int find_suitable_fallback(struct free_area *area, unsigned int order,
2709 int migratetype, bool only_stealable, bool *can_steal)
2714 if (area->nr_free == 0)
2719 fallback_mt = fallbacks[migratetype][i];
2720 if (fallback_mt == MIGRATE_TYPES)
2723 if (free_area_empty(area, fallback_mt))
2726 if (can_steal_fallback(order, migratetype))
2729 if (!only_stealable)
2740 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2741 * there are no empty page blocks that contain a page with a suitable order
2743 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2744 unsigned int alloc_order)
2747 unsigned long max_managed, flags;
2750 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2751 * Check is race-prone but harmless.
2753 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2754 if (zone->nr_reserved_highatomic >= max_managed)
2757 spin_lock_irqsave(&zone->lock, flags);
2759 /* Recheck the nr_reserved_highatomic limit under the lock */
2760 if (zone->nr_reserved_highatomic >= max_managed)
2764 mt = get_pageblock_migratetype(page);
2765 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2766 if (migratetype_is_mergeable(mt)) {
2767 zone->nr_reserved_highatomic += pageblock_nr_pages;
2768 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2769 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2773 spin_unlock_irqrestore(&zone->lock, flags);
2777 * Used when an allocation is about to fail under memory pressure. This
2778 * potentially hurts the reliability of high-order allocations when under
2779 * intense memory pressure but failed atomic allocations should be easier
2780 * to recover from than an OOM.
2782 * If @force is true, try to unreserve a pageblock even though highatomic
2783 * pageblock is exhausted.
2785 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2788 struct zonelist *zonelist = ac->zonelist;
2789 unsigned long flags;
2796 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2799 * Preserve at least one pageblock unless memory pressure
2802 if (!force && zone->nr_reserved_highatomic <=
2806 spin_lock_irqsave(&zone->lock, flags);
2807 for (order = 0; order < MAX_ORDER; order++) {
2808 struct free_area *area = &(zone->free_area[order]);
2810 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2815 * In page freeing path, migratetype change is racy so
2816 * we can counter several free pages in a pageblock
2817 * in this loop although we changed the pageblock type
2818 * from highatomic to ac->migratetype. So we should
2819 * adjust the count once.
2821 if (is_migrate_highatomic_page(page)) {
2823 * It should never happen but changes to
2824 * locking could inadvertently allow a per-cpu
2825 * drain to add pages to MIGRATE_HIGHATOMIC
2826 * while unreserving so be safe and watch for
2829 zone->nr_reserved_highatomic -= min(
2831 zone->nr_reserved_highatomic);
2835 * Convert to ac->migratetype and avoid the normal
2836 * pageblock stealing heuristics. Minimally, the caller
2837 * is doing the work and needs the pages. More
2838 * importantly, if the block was always converted to
2839 * MIGRATE_UNMOVABLE or another type then the number
2840 * of pageblocks that cannot be completely freed
2843 set_pageblock_migratetype(page, ac->migratetype);
2844 ret = move_freepages_block(zone, page, ac->migratetype,
2847 spin_unlock_irqrestore(&zone->lock, flags);
2851 spin_unlock_irqrestore(&zone->lock, flags);
2858 * Try finding a free buddy page on the fallback list and put it on the free
2859 * list of requested migratetype, possibly along with other pages from the same
2860 * block, depending on fragmentation avoidance heuristics. Returns true if
2861 * fallback was found so that __rmqueue_smallest() can grab it.
2863 * The use of signed ints for order and current_order is a deliberate
2864 * deviation from the rest of this file, to make the for loop
2865 * condition simpler.
2867 static __always_inline bool
2868 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2869 unsigned int alloc_flags)
2871 struct free_area *area;
2873 int min_order = order;
2879 * Do not steal pages from freelists belonging to other pageblocks
2880 * i.e. orders < pageblock_order. If there are no local zones free,
2881 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2883 if (alloc_flags & ALLOC_NOFRAGMENT)
2884 min_order = pageblock_order;
2887 * Find the largest available free page in the other list. This roughly
2888 * approximates finding the pageblock with the most free pages, which
2889 * would be too costly to do exactly.
2891 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2893 area = &(zone->free_area[current_order]);
2894 fallback_mt = find_suitable_fallback(area, current_order,
2895 start_migratetype, false, &can_steal);
2896 if (fallback_mt == -1)
2900 * We cannot steal all free pages from the pageblock and the
2901 * requested migratetype is movable. In that case it's better to
2902 * steal and split the smallest available page instead of the
2903 * largest available page, because even if the next movable
2904 * allocation falls back into a different pageblock than this
2905 * one, it won't cause permanent fragmentation.
2907 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2908 && current_order > order)
2917 for (current_order = order; current_order < MAX_ORDER;
2919 area = &(zone->free_area[current_order]);
2920 fallback_mt = find_suitable_fallback(area, current_order,
2921 start_migratetype, false, &can_steal);
2922 if (fallback_mt != -1)
2927 * This should not happen - we already found a suitable fallback
2928 * when looking for the largest page.
2930 VM_BUG_ON(current_order == MAX_ORDER);
2933 page = get_page_from_free_area(area, fallback_mt);
2935 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2938 trace_mm_page_alloc_extfrag(page, order, current_order,
2939 start_migratetype, fallback_mt);
2946 * Do the hard work of removing an element from the buddy allocator.
2947 * Call me with the zone->lock already held.
2949 static __always_inline struct page *
2950 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2951 unsigned int alloc_flags)
2955 if (IS_ENABLED(CONFIG_CMA)) {
2957 * Balance movable allocations between regular and CMA areas by
2958 * allocating from CMA when over half of the zone's free memory
2959 * is in the CMA area.
2961 if (alloc_flags & ALLOC_CMA &&
2962 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2963 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2964 page = __rmqueue_cma_fallback(zone, order);
2970 page = __rmqueue_smallest(zone, order, migratetype);
2971 if (unlikely(!page)) {
2972 if (alloc_flags & ALLOC_CMA)
2973 page = __rmqueue_cma_fallback(zone, order);
2975 if (!page && __rmqueue_fallback(zone, order, migratetype,
2981 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2986 * Obtain a specified number of elements from the buddy allocator, all under
2987 * a single hold of the lock, for efficiency. Add them to the supplied list.
2988 * Returns the number of new pages which were placed at *list.
2990 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2991 unsigned long count, struct list_head *list,
2992 int migratetype, unsigned int alloc_flags)
2994 int i, allocated = 0;
2997 * local_lock_irq held so equivalent to spin_lock_irqsave for
2998 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3000 spin_lock(&zone->lock);
3001 for (i = 0; i < count; ++i) {
3002 struct page *page = __rmqueue(zone, order, migratetype,
3004 if (unlikely(page == NULL))
3007 if (unlikely(check_pcp_refill(page, order)))
3011 * Split buddy pages returned by expand() are received here in
3012 * physical page order. The page is added to the tail of
3013 * caller's list. From the callers perspective, the linked list
3014 * is ordered by page number under some conditions. This is
3015 * useful for IO devices that can forward direction from the
3016 * head, thus also in the physical page order. This is useful
3017 * for IO devices that can merge IO requests if the physical
3018 * pages are ordered properly.
3020 list_add_tail(&page->lru, list);
3022 if (is_migrate_cma(get_pcppage_migratetype(page)))
3023 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3028 * i pages were removed from the buddy list even if some leak due
3029 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3030 * on i. Do not confuse with 'allocated' which is the number of
3031 * pages added to the pcp list.
3033 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3034 spin_unlock(&zone->lock);
3040 * Called from the vmstat counter updater to drain pagesets of this
3041 * currently executing processor on remote nodes after they have
3044 * Note that this function must be called with the thread pinned to
3045 * a single processor.
3047 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3049 unsigned long flags;
3050 int to_drain, batch;
3052 local_lock_irqsave(&pagesets.lock, flags);
3053 batch = READ_ONCE(pcp->batch);
3054 to_drain = min(pcp->count, batch);
3056 free_pcppages_bulk(zone, to_drain, pcp, 0);
3057 local_unlock_irqrestore(&pagesets.lock, flags);
3062 * Drain pcplists of the indicated processor and zone.
3064 * The processor must either be the current processor and the
3065 * thread pinned to the current processor or a processor that
3068 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3070 unsigned long flags;
3071 struct per_cpu_pages *pcp;
3073 local_lock_irqsave(&pagesets.lock, flags);
3075 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3077 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3079 local_unlock_irqrestore(&pagesets.lock, flags);
3083 * Drain pcplists of all zones on the indicated processor.
3085 * The processor must either be the current processor and the
3086 * thread pinned to the current processor or a processor that
3089 static void drain_pages(unsigned int cpu)
3093 for_each_populated_zone(zone) {
3094 drain_pages_zone(cpu, zone);
3099 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3101 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3102 * the single zone's pages.
3104 void drain_local_pages(struct zone *zone)
3106 int cpu = smp_processor_id();
3109 drain_pages_zone(cpu, zone);
3114 static void drain_local_pages_wq(struct work_struct *work)
3116 struct pcpu_drain *drain;
3118 drain = container_of(work, struct pcpu_drain, work);
3121 * drain_all_pages doesn't use proper cpu hotplug protection so
3122 * we can race with cpu offline when the WQ can move this from
3123 * a cpu pinned worker to an unbound one. We can operate on a different
3124 * cpu which is alright but we also have to make sure to not move to
3128 drain_local_pages(drain->zone);
3133 * The implementation of drain_all_pages(), exposing an extra parameter to
3134 * drain on all cpus.
3136 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3137 * not empty. The check for non-emptiness can however race with a free to
3138 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3139 * that need the guarantee that every CPU has drained can disable the
3140 * optimizing racy check.
3142 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3147 * Allocate in the BSS so we won't require allocation in
3148 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3150 static cpumask_t cpus_with_pcps;
3153 * Make sure nobody triggers this path before mm_percpu_wq is fully
3156 if (WARN_ON_ONCE(!mm_percpu_wq))
3160 * Do not drain if one is already in progress unless it's specific to
3161 * a zone. Such callers are primarily CMA and memory hotplug and need
3162 * the drain to be complete when the call returns.
3164 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3167 mutex_lock(&pcpu_drain_mutex);
3171 * We don't care about racing with CPU hotplug event
3172 * as offline notification will cause the notified
3173 * cpu to drain that CPU pcps and on_each_cpu_mask
3174 * disables preemption as part of its processing
3176 for_each_online_cpu(cpu) {
3177 struct per_cpu_pages *pcp;
3179 bool has_pcps = false;
3181 if (force_all_cpus) {
3183 * The pcp.count check is racy, some callers need a
3184 * guarantee that no cpu is missed.
3188 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3192 for_each_populated_zone(z) {
3193 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3202 cpumask_set_cpu(cpu, &cpus_with_pcps);
3204 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3207 for_each_cpu(cpu, &cpus_with_pcps) {
3208 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3211 INIT_WORK(&drain->work, drain_local_pages_wq);
3212 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3214 for_each_cpu(cpu, &cpus_with_pcps)
3215 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3217 mutex_unlock(&pcpu_drain_mutex);
3221 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3223 * When zone parameter is non-NULL, spill just the single zone's pages.
3225 * Note that this can be extremely slow as the draining happens in a workqueue.
3227 void drain_all_pages(struct zone *zone)
3229 __drain_all_pages(zone, false);
3232 #ifdef CONFIG_HIBERNATION
3235 * Touch the watchdog for every WD_PAGE_COUNT pages.
3237 #define WD_PAGE_COUNT (128*1024)
3239 void mark_free_pages(struct zone *zone)
3241 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3242 unsigned long flags;
3243 unsigned int order, t;
3246 if (zone_is_empty(zone))
3249 spin_lock_irqsave(&zone->lock, flags);
3251 max_zone_pfn = zone_end_pfn(zone);
3252 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3253 if (pfn_valid(pfn)) {
3254 page = pfn_to_page(pfn);
3256 if (!--page_count) {
3257 touch_nmi_watchdog();
3258 page_count = WD_PAGE_COUNT;
3261 if (page_zone(page) != zone)
3264 if (!swsusp_page_is_forbidden(page))
3265 swsusp_unset_page_free(page);
3268 for_each_migratetype_order(order, t) {
3269 list_for_each_entry(page,
3270 &zone->free_area[order].free_list[t], lru) {
3273 pfn = page_to_pfn(page);
3274 for (i = 0; i < (1UL << order); i++) {
3275 if (!--page_count) {
3276 touch_nmi_watchdog();
3277 page_count = WD_PAGE_COUNT;
3279 swsusp_set_page_free(pfn_to_page(pfn + i));
3283 spin_unlock_irqrestore(&zone->lock, flags);
3285 #endif /* CONFIG_PM */
3287 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3292 if (!free_pcp_prepare(page, order))
3295 migratetype = get_pfnblock_migratetype(page, pfn);
3296 set_pcppage_migratetype(page, migratetype);
3300 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3303 int min_nr_free, max_nr_free;
3305 /* Free everything if batch freeing high-order pages. */
3306 if (unlikely(free_high))
3309 /* Check for PCP disabled or boot pageset */
3310 if (unlikely(high < batch))
3313 /* Leave at least pcp->batch pages on the list */
3314 min_nr_free = batch;
3315 max_nr_free = high - batch;
3318 * Double the number of pages freed each time there is subsequent
3319 * freeing of pages without any allocation.
3321 batch <<= pcp->free_factor;
3322 if (batch < max_nr_free)
3324 batch = clamp(batch, min_nr_free, max_nr_free);
3329 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3332 int high = READ_ONCE(pcp->high);
3334 if (unlikely(!high || free_high))
3337 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3341 * If reclaim is active, limit the number of pages that can be
3342 * stored on pcp lists
3344 return min(READ_ONCE(pcp->batch) << 2, high);
3347 static void free_unref_page_commit(struct page *page, int migratetype,
3350 struct zone *zone = page_zone(page);
3351 struct per_cpu_pages *pcp;
3356 __count_vm_event(PGFREE);
3357 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3358 pindex = order_to_pindex(migratetype, order);
3359 list_add(&page->lru, &pcp->lists[pindex]);
3360 pcp->count += 1 << order;
3363 * As high-order pages other than THP's stored on PCP can contribute
3364 * to fragmentation, limit the number stored when PCP is heavily
3365 * freeing without allocation. The remainder after bulk freeing
3366 * stops will be drained from vmstat refresh context.
3368 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3370 high = nr_pcp_high(pcp, zone, free_high);
3371 if (pcp->count >= high) {
3372 int batch = READ_ONCE(pcp->batch);
3374 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3381 void free_unref_page(struct page *page, unsigned int order)
3383 unsigned long flags;
3384 unsigned long pfn = page_to_pfn(page);
3387 if (!free_unref_page_prepare(page, pfn, order))
3391 * We only track unmovable, reclaimable and movable on pcp lists.
3392 * Place ISOLATE pages on the isolated list because they are being
3393 * offlined but treat HIGHATOMIC as movable pages so we can get those
3394 * areas back if necessary. Otherwise, we may have to free
3395 * excessively into the page allocator
3397 migratetype = get_pcppage_migratetype(page);
3398 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3399 if (unlikely(is_migrate_isolate(migratetype))) {
3400 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3403 migratetype = MIGRATE_MOVABLE;
3406 local_lock_irqsave(&pagesets.lock, flags);
3407 free_unref_page_commit(page, migratetype, order);
3408 local_unlock_irqrestore(&pagesets.lock, flags);
3412 * Free a list of 0-order pages
3414 void free_unref_page_list(struct list_head *list)
3416 struct page *page, *next;
3417 unsigned long flags;
3418 int batch_count = 0;
3421 /* Prepare pages for freeing */
3422 list_for_each_entry_safe(page, next, list, lru) {
3423 unsigned long pfn = page_to_pfn(page);
3424 if (!free_unref_page_prepare(page, pfn, 0)) {
3425 list_del(&page->lru);
3430 * Free isolated pages directly to the allocator, see
3431 * comment in free_unref_page.
3433 migratetype = get_pcppage_migratetype(page);
3434 if (unlikely(is_migrate_isolate(migratetype))) {
3435 list_del(&page->lru);
3436 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3441 local_lock_irqsave(&pagesets.lock, flags);
3442 list_for_each_entry_safe(page, next, list, lru) {
3444 * Non-isolated types over MIGRATE_PCPTYPES get added
3445 * to the MIGRATE_MOVABLE pcp list.
3447 migratetype = get_pcppage_migratetype(page);
3448 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3449 migratetype = MIGRATE_MOVABLE;
3451 trace_mm_page_free_batched(page);
3452 free_unref_page_commit(page, migratetype, 0);
3455 * Guard against excessive IRQ disabled times when we get
3456 * a large list of pages to free.
3458 if (++batch_count == SWAP_CLUSTER_MAX) {
3459 local_unlock_irqrestore(&pagesets.lock, flags);
3461 local_lock_irqsave(&pagesets.lock, flags);
3464 local_unlock_irqrestore(&pagesets.lock, flags);
3468 * split_page takes a non-compound higher-order page, and splits it into
3469 * n (1<<order) sub-pages: page[0..n]
3470 * Each sub-page must be freed individually.
3472 * Note: this is probably too low level an operation for use in drivers.
3473 * Please consult with lkml before using this in your driver.
3475 void split_page(struct page *page, unsigned int order)
3479 VM_BUG_ON_PAGE(PageCompound(page), page);
3480 VM_BUG_ON_PAGE(!page_count(page), page);
3482 for (i = 1; i < (1 << order); i++)
3483 set_page_refcounted(page + i);
3484 split_page_owner(page, 1 << order);
3485 split_page_memcg(page, 1 << order);
3487 EXPORT_SYMBOL_GPL(split_page);
3489 int __isolate_free_page(struct page *page, unsigned int order)
3491 unsigned long watermark;
3495 BUG_ON(!PageBuddy(page));
3497 zone = page_zone(page);
3498 mt = get_pageblock_migratetype(page);
3500 if (!is_migrate_isolate(mt)) {
3502 * Obey watermarks as if the page was being allocated. We can
3503 * emulate a high-order watermark check with a raised order-0
3504 * watermark, because we already know our high-order page
3507 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3508 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3511 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3514 /* Remove page from free list */
3516 del_page_from_free_list(page, zone, order);
3519 * Set the pageblock if the isolated page is at least half of a
3522 if (order >= pageblock_order - 1) {
3523 struct page *endpage = page + (1 << order) - 1;
3524 for (; page < endpage; page += pageblock_nr_pages) {
3525 int mt = get_pageblock_migratetype(page);
3527 * Only change normal pageblocks (i.e., they can merge
3530 if (migratetype_is_mergeable(mt))
3531 set_pageblock_migratetype(page,
3537 return 1UL << order;
3541 * __putback_isolated_page - Return a now-isolated page back where we got it
3542 * @page: Page that was isolated
3543 * @order: Order of the isolated page
3544 * @mt: The page's pageblock's migratetype
3546 * This function is meant to return a page pulled from the free lists via
3547 * __isolate_free_page back to the free lists they were pulled from.
3549 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3551 struct zone *zone = page_zone(page);
3553 /* zone lock should be held when this function is called */
3554 lockdep_assert_held(&zone->lock);
3556 /* Return isolated page to tail of freelist. */
3557 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3558 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3562 * Update NUMA hit/miss statistics
3564 * Must be called with interrupts disabled.
3566 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3570 enum numa_stat_item local_stat = NUMA_LOCAL;
3572 /* skip numa counters update if numa stats is disabled */
3573 if (!static_branch_likely(&vm_numa_stat_key))
3576 if (zone_to_nid(z) != numa_node_id())
3577 local_stat = NUMA_OTHER;
3579 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3580 __count_numa_events(z, NUMA_HIT, nr_account);
3582 __count_numa_events(z, NUMA_MISS, nr_account);
3583 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3585 __count_numa_events(z, local_stat, nr_account);
3589 /* Remove page from the per-cpu list, caller must protect the list */
3591 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3593 unsigned int alloc_flags,
3594 struct per_cpu_pages *pcp,
3595 struct list_head *list)
3600 if (list_empty(list)) {
3601 int batch = READ_ONCE(pcp->batch);
3605 * Scale batch relative to order if batch implies
3606 * free pages can be stored on the PCP. Batch can
3607 * be 1 for small zones or for boot pagesets which
3608 * should never store free pages as the pages may
3609 * belong to arbitrary zones.
3612 batch = max(batch >> order, 2);
3613 alloced = rmqueue_bulk(zone, order,
3615 migratetype, alloc_flags);
3617 pcp->count += alloced << order;
3618 if (unlikely(list_empty(list)))
3622 page = list_first_entry(list, struct page, lru);
3623 list_del(&page->lru);
3624 pcp->count -= 1 << order;
3625 } while (check_new_pcp(page, order));
3630 /* Lock and remove page from the per-cpu list */
3631 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3632 struct zone *zone, unsigned int order,
3633 gfp_t gfp_flags, int migratetype,
3634 unsigned int alloc_flags)
3636 struct per_cpu_pages *pcp;
3637 struct list_head *list;
3639 unsigned long flags;
3641 local_lock_irqsave(&pagesets.lock, flags);
3644 * On allocation, reduce the number of pages that are batch freed.
3645 * See nr_pcp_free() where free_factor is increased for subsequent
3648 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3649 pcp->free_factor >>= 1;
3650 list = &pcp->lists[order_to_pindex(migratetype, order)];
3651 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3652 local_unlock_irqrestore(&pagesets.lock, flags);
3654 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3655 zone_statistics(preferred_zone, zone, 1);
3661 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3664 struct page *rmqueue(struct zone *preferred_zone,
3665 struct zone *zone, unsigned int order,
3666 gfp_t gfp_flags, unsigned int alloc_flags,
3669 unsigned long flags;
3672 if (likely(pcp_allowed_order(order))) {
3674 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3675 * we need to skip it when CMA area isn't allowed.
3677 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3678 migratetype != MIGRATE_MOVABLE) {
3679 page = rmqueue_pcplist(preferred_zone, zone, order,
3680 gfp_flags, migratetype, alloc_flags);
3686 * We most definitely don't want callers attempting to
3687 * allocate greater than order-1 page units with __GFP_NOFAIL.
3689 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3693 spin_lock_irqsave(&zone->lock, flags);
3695 * order-0 request can reach here when the pcplist is skipped
3696 * due to non-CMA allocation context. HIGHATOMIC area is
3697 * reserved for high-order atomic allocation, so order-0
3698 * request should skip it.
3700 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3701 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3703 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3706 page = __rmqueue(zone, order, migratetype, alloc_flags);
3710 __mod_zone_freepage_state(zone, -(1 << order),
3711 get_pcppage_migratetype(page));
3712 spin_unlock_irqrestore(&zone->lock, flags);
3713 } while (check_new_pages(page, order));
3715 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3716 zone_statistics(preferred_zone, zone, 1);
3719 /* Separate test+clear to avoid unnecessary atomics */
3720 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3721 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3722 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3725 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3729 spin_unlock_irqrestore(&zone->lock, flags);
3733 #ifdef CONFIG_FAIL_PAGE_ALLOC
3736 struct fault_attr attr;
3738 bool ignore_gfp_highmem;
3739 bool ignore_gfp_reclaim;
3741 } fail_page_alloc = {
3742 .attr = FAULT_ATTR_INITIALIZER,
3743 .ignore_gfp_reclaim = true,
3744 .ignore_gfp_highmem = true,
3748 static int __init setup_fail_page_alloc(char *str)
3750 return setup_fault_attr(&fail_page_alloc.attr, str);
3752 __setup("fail_page_alloc=", setup_fail_page_alloc);
3754 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3756 if (order < fail_page_alloc.min_order)
3758 if (gfp_mask & __GFP_NOFAIL)
3760 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3762 if (fail_page_alloc.ignore_gfp_reclaim &&
3763 (gfp_mask & __GFP_DIRECT_RECLAIM))
3766 return should_fail(&fail_page_alloc.attr, 1 << order);
3769 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3771 static int __init fail_page_alloc_debugfs(void)
3773 umode_t mode = S_IFREG | 0600;
3776 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3777 &fail_page_alloc.attr);
3779 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3780 &fail_page_alloc.ignore_gfp_reclaim);
3781 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3782 &fail_page_alloc.ignore_gfp_highmem);
3783 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3788 late_initcall(fail_page_alloc_debugfs);
3790 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3792 #else /* CONFIG_FAIL_PAGE_ALLOC */
3794 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3799 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3801 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3803 return __should_fail_alloc_page(gfp_mask, order);
3805 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3807 static inline long __zone_watermark_unusable_free(struct zone *z,
3808 unsigned int order, unsigned int alloc_flags)
3810 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3811 long unusable_free = (1 << order) - 1;
3814 * If the caller does not have rights to ALLOC_HARDER then subtract
3815 * the high-atomic reserves. This will over-estimate the size of the
3816 * atomic reserve but it avoids a search.
3818 if (likely(!alloc_harder))
3819 unusable_free += z->nr_reserved_highatomic;
3822 /* If allocation can't use CMA areas don't use free CMA pages */
3823 if (!(alloc_flags & ALLOC_CMA))
3824 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3827 return unusable_free;
3831 * Return true if free base pages are above 'mark'. For high-order checks it
3832 * will return true of the order-0 watermark is reached and there is at least
3833 * one free page of a suitable size. Checking now avoids taking the zone lock
3834 * to check in the allocation paths if no pages are free.
3836 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3837 int highest_zoneidx, unsigned int alloc_flags,
3842 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3844 /* free_pages may go negative - that's OK */
3845 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3847 if (alloc_flags & ALLOC_HIGH)
3850 if (unlikely(alloc_harder)) {
3852 * OOM victims can try even harder than normal ALLOC_HARDER
3853 * users on the grounds that it's definitely going to be in
3854 * the exit path shortly and free memory. Any allocation it
3855 * makes during the free path will be small and short-lived.
3857 if (alloc_flags & ALLOC_OOM)
3864 * Check watermarks for an order-0 allocation request. If these
3865 * are not met, then a high-order request also cannot go ahead
3866 * even if a suitable page happened to be free.
3868 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3871 /* If this is an order-0 request then the watermark is fine */
3875 /* For a high-order request, check at least one suitable page is free */
3876 for (o = order; o < MAX_ORDER; o++) {
3877 struct free_area *area = &z->free_area[o];
3883 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3884 if (!free_area_empty(area, mt))
3889 if ((alloc_flags & ALLOC_CMA) &&
3890 !free_area_empty(area, MIGRATE_CMA)) {
3894 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3900 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3901 int highest_zoneidx, unsigned int alloc_flags)
3903 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3904 zone_page_state(z, NR_FREE_PAGES));
3907 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3908 unsigned long mark, int highest_zoneidx,
3909 unsigned int alloc_flags, gfp_t gfp_mask)
3913 free_pages = zone_page_state(z, NR_FREE_PAGES);
3916 * Fast check for order-0 only. If this fails then the reserves
3917 * need to be calculated.
3922 fast_free = free_pages;
3923 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3924 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3928 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3932 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3933 * when checking the min watermark. The min watermark is the
3934 * point where boosting is ignored so that kswapd is woken up
3935 * when below the low watermark.
3937 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3938 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3939 mark = z->_watermark[WMARK_MIN];
3940 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3941 alloc_flags, free_pages);
3947 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3948 unsigned long mark, int highest_zoneidx)
3950 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3952 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3953 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3955 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3960 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3962 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3964 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3965 node_reclaim_distance;
3967 #else /* CONFIG_NUMA */
3968 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3972 #endif /* CONFIG_NUMA */
3975 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3976 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3977 * premature use of a lower zone may cause lowmem pressure problems that
3978 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3979 * probably too small. It only makes sense to spread allocations to avoid
3980 * fragmentation between the Normal and DMA32 zones.
3982 static inline unsigned int
3983 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3985 unsigned int alloc_flags;
3988 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3991 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3993 #ifdef CONFIG_ZONE_DMA32
3997 if (zone_idx(zone) != ZONE_NORMAL)
4001 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4002 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4003 * on UMA that if Normal is populated then so is DMA32.
4005 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4006 if (nr_online_nodes > 1 && !populated_zone(--zone))
4009 alloc_flags |= ALLOC_NOFRAGMENT;
4010 #endif /* CONFIG_ZONE_DMA32 */
4014 /* Must be called after current_gfp_context() which can change gfp_mask */
4015 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4016 unsigned int alloc_flags)
4019 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4020 alloc_flags |= ALLOC_CMA;
4026 * get_page_from_freelist goes through the zonelist trying to allocate
4029 static struct page *
4030 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4031 const struct alloc_context *ac)
4035 struct pglist_data *last_pgdat_dirty_limit = NULL;
4040 * Scan zonelist, looking for a zone with enough free.
4041 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4043 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4044 z = ac->preferred_zoneref;
4045 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4050 if (cpusets_enabled() &&
4051 (alloc_flags & ALLOC_CPUSET) &&
4052 !__cpuset_zone_allowed(zone, gfp_mask))
4055 * When allocating a page cache page for writing, we
4056 * want to get it from a node that is within its dirty
4057 * limit, such that no single node holds more than its
4058 * proportional share of globally allowed dirty pages.
4059 * The dirty limits take into account the node's
4060 * lowmem reserves and high watermark so that kswapd
4061 * should be able to balance it without having to
4062 * write pages from its LRU list.
4064 * XXX: For now, allow allocations to potentially
4065 * exceed the per-node dirty limit in the slowpath
4066 * (spread_dirty_pages unset) before going into reclaim,
4067 * which is important when on a NUMA setup the allowed
4068 * nodes are together not big enough to reach the
4069 * global limit. The proper fix for these situations
4070 * will require awareness of nodes in the
4071 * dirty-throttling and the flusher threads.
4073 if (ac->spread_dirty_pages) {
4074 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4077 if (!node_dirty_ok(zone->zone_pgdat)) {
4078 last_pgdat_dirty_limit = zone->zone_pgdat;
4083 if (no_fallback && nr_online_nodes > 1 &&
4084 zone != ac->preferred_zoneref->zone) {
4088 * If moving to a remote node, retry but allow
4089 * fragmenting fallbacks. Locality is more important
4090 * than fragmentation avoidance.
4092 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4093 if (zone_to_nid(zone) != local_nid) {
4094 alloc_flags &= ~ALLOC_NOFRAGMENT;
4099 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4100 if (!zone_watermark_fast(zone, order, mark,
4101 ac->highest_zoneidx, alloc_flags,
4105 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4107 * Watermark failed for this zone, but see if we can
4108 * grow this zone if it contains deferred pages.
4110 if (static_branch_unlikely(&deferred_pages)) {
4111 if (_deferred_grow_zone(zone, order))
4115 /* Checked here to keep the fast path fast */
4116 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4117 if (alloc_flags & ALLOC_NO_WATERMARKS)
4120 if (!node_reclaim_enabled() ||
4121 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4124 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4126 case NODE_RECLAIM_NOSCAN:
4129 case NODE_RECLAIM_FULL:
4130 /* scanned but unreclaimable */
4133 /* did we reclaim enough */
4134 if (zone_watermark_ok(zone, order, mark,
4135 ac->highest_zoneidx, alloc_flags))
4143 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4144 gfp_mask, alloc_flags, ac->migratetype);
4146 prep_new_page(page, order, gfp_mask, alloc_flags);
4149 * If this is a high-order atomic allocation then check
4150 * if the pageblock should be reserved for the future
4152 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4153 reserve_highatomic_pageblock(page, zone, order);
4157 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4158 /* Try again if zone has deferred pages */
4159 if (static_branch_unlikely(&deferred_pages)) {
4160 if (_deferred_grow_zone(zone, order))
4168 * It's possible on a UMA machine to get through all zones that are
4169 * fragmented. If avoiding fragmentation, reset and try again.
4172 alloc_flags &= ~ALLOC_NOFRAGMENT;
4179 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4181 unsigned int filter = SHOW_MEM_FILTER_NODES;
4184 * This documents exceptions given to allocations in certain
4185 * contexts that are allowed to allocate outside current's set
4188 if (!(gfp_mask & __GFP_NOMEMALLOC))
4189 if (tsk_is_oom_victim(current) ||
4190 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4191 filter &= ~SHOW_MEM_FILTER_NODES;
4192 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4193 filter &= ~SHOW_MEM_FILTER_NODES;
4195 show_mem(filter, nodemask);
4198 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4200 struct va_format vaf;
4202 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4204 if ((gfp_mask & __GFP_NOWARN) ||
4205 !__ratelimit(&nopage_rs) ||
4206 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4209 va_start(args, fmt);
4212 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4213 current->comm, &vaf, gfp_mask, &gfp_mask,
4214 nodemask_pr_args(nodemask));
4217 cpuset_print_current_mems_allowed();
4220 warn_alloc_show_mem(gfp_mask, nodemask);
4223 static inline struct page *
4224 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4225 unsigned int alloc_flags,
4226 const struct alloc_context *ac)
4230 page = get_page_from_freelist(gfp_mask, order,
4231 alloc_flags|ALLOC_CPUSET, ac);
4233 * fallback to ignore cpuset restriction if our nodes
4237 page = get_page_from_freelist(gfp_mask, order,
4243 static inline struct page *
4244 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4245 const struct alloc_context *ac, unsigned long *did_some_progress)
4247 struct oom_control oc = {
4248 .zonelist = ac->zonelist,
4249 .nodemask = ac->nodemask,
4251 .gfp_mask = gfp_mask,
4256 *did_some_progress = 0;
4259 * Acquire the oom lock. If that fails, somebody else is
4260 * making progress for us.
4262 if (!mutex_trylock(&oom_lock)) {
4263 *did_some_progress = 1;
4264 schedule_timeout_uninterruptible(1);
4269 * Go through the zonelist yet one more time, keep very high watermark
4270 * here, this is only to catch a parallel oom killing, we must fail if
4271 * we're still under heavy pressure. But make sure that this reclaim
4272 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4273 * allocation which will never fail due to oom_lock already held.
4275 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4276 ~__GFP_DIRECT_RECLAIM, order,
4277 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4281 /* Coredumps can quickly deplete all memory reserves */
4282 if (current->flags & PF_DUMPCORE)
4284 /* The OOM killer will not help higher order allocs */
4285 if (order > PAGE_ALLOC_COSTLY_ORDER)
4288 * We have already exhausted all our reclaim opportunities without any
4289 * success so it is time to admit defeat. We will skip the OOM killer
4290 * because it is very likely that the caller has a more reasonable
4291 * fallback than shooting a random task.
4293 * The OOM killer may not free memory on a specific node.
4295 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4297 /* The OOM killer does not needlessly kill tasks for lowmem */
4298 if (ac->highest_zoneidx < ZONE_NORMAL)
4300 if (pm_suspended_storage())
4303 * XXX: GFP_NOFS allocations should rather fail than rely on
4304 * other request to make a forward progress.
4305 * We are in an unfortunate situation where out_of_memory cannot
4306 * do much for this context but let's try it to at least get
4307 * access to memory reserved if the current task is killed (see
4308 * out_of_memory). Once filesystems are ready to handle allocation
4309 * failures more gracefully we should just bail out here.
4312 /* Exhausted what can be done so it's blame time */
4313 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4314 *did_some_progress = 1;
4317 * Help non-failing allocations by giving them access to memory
4320 if (gfp_mask & __GFP_NOFAIL)
4321 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4322 ALLOC_NO_WATERMARKS, ac);
4325 mutex_unlock(&oom_lock);
4330 * Maximum number of compaction retries with a progress before OOM
4331 * killer is consider as the only way to move forward.
4333 #define MAX_COMPACT_RETRIES 16
4335 #ifdef CONFIG_COMPACTION
4336 /* Try memory compaction for high-order allocations before reclaim */
4337 static struct page *
4338 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4339 unsigned int alloc_flags, const struct alloc_context *ac,
4340 enum compact_priority prio, enum compact_result *compact_result)
4342 struct page *page = NULL;
4343 unsigned long pflags;
4344 unsigned int noreclaim_flag;
4349 psi_memstall_enter(&pflags);
4350 delayacct_compact_start();
4351 noreclaim_flag = memalloc_noreclaim_save();
4353 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4356 memalloc_noreclaim_restore(noreclaim_flag);
4357 psi_memstall_leave(&pflags);
4358 delayacct_compact_end();
4360 if (*compact_result == COMPACT_SKIPPED)
4363 * At least in one zone compaction wasn't deferred or skipped, so let's
4364 * count a compaction stall
4366 count_vm_event(COMPACTSTALL);
4368 /* Prep a captured page if available */
4370 prep_new_page(page, order, gfp_mask, alloc_flags);
4372 /* Try get a page from the freelist if available */
4374 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4377 struct zone *zone = page_zone(page);
4379 zone->compact_blockskip_flush = false;
4380 compaction_defer_reset(zone, order, true);
4381 count_vm_event(COMPACTSUCCESS);
4386 * It's bad if compaction run occurs and fails. The most likely reason
4387 * is that pages exist, but not enough to satisfy watermarks.
4389 count_vm_event(COMPACTFAIL);
4397 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4398 enum compact_result compact_result,
4399 enum compact_priority *compact_priority,
4400 int *compaction_retries)
4402 int max_retries = MAX_COMPACT_RETRIES;
4405 int retries = *compaction_retries;
4406 enum compact_priority priority = *compact_priority;
4411 if (fatal_signal_pending(current))
4414 if (compaction_made_progress(compact_result))
4415 (*compaction_retries)++;
4418 * compaction considers all the zone as desperately out of memory
4419 * so it doesn't really make much sense to retry except when the
4420 * failure could be caused by insufficient priority
4422 if (compaction_failed(compact_result))
4423 goto check_priority;
4426 * compaction was skipped because there are not enough order-0 pages
4427 * to work with, so we retry only if it looks like reclaim can help.
4429 if (compaction_needs_reclaim(compact_result)) {
4430 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4435 * make sure the compaction wasn't deferred or didn't bail out early
4436 * due to locks contention before we declare that we should give up.
4437 * But the next retry should use a higher priority if allowed, so
4438 * we don't just keep bailing out endlessly.
4440 if (compaction_withdrawn(compact_result)) {
4441 goto check_priority;
4445 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4446 * costly ones because they are de facto nofail and invoke OOM
4447 * killer to move on while costly can fail and users are ready
4448 * to cope with that. 1/4 retries is rather arbitrary but we
4449 * would need much more detailed feedback from compaction to
4450 * make a better decision.
4452 if (order > PAGE_ALLOC_COSTLY_ORDER)
4454 if (*compaction_retries <= max_retries) {
4460 * Make sure there are attempts at the highest priority if we exhausted
4461 * all retries or failed at the lower priorities.
4464 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4465 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4467 if (*compact_priority > min_priority) {
4468 (*compact_priority)--;
4469 *compaction_retries = 0;
4473 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4477 static inline struct page *
4478 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4479 unsigned int alloc_flags, const struct alloc_context *ac,
4480 enum compact_priority prio, enum compact_result *compact_result)
4482 *compact_result = COMPACT_SKIPPED;
4487 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4488 enum compact_result compact_result,
4489 enum compact_priority *compact_priority,
4490 int *compaction_retries)
4495 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4499 * There are setups with compaction disabled which would prefer to loop
4500 * inside the allocator rather than hit the oom killer prematurely.
4501 * Let's give them a good hope and keep retrying while the order-0
4502 * watermarks are OK.
4504 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4505 ac->highest_zoneidx, ac->nodemask) {
4506 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4507 ac->highest_zoneidx, alloc_flags))
4512 #endif /* CONFIG_COMPACTION */
4514 #ifdef CONFIG_LOCKDEP
4515 static struct lockdep_map __fs_reclaim_map =
4516 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4518 static bool __need_reclaim(gfp_t gfp_mask)
4520 /* no reclaim without waiting on it */
4521 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4524 /* this guy won't enter reclaim */
4525 if (current->flags & PF_MEMALLOC)
4528 if (gfp_mask & __GFP_NOLOCKDEP)
4534 void __fs_reclaim_acquire(unsigned long ip)
4536 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4539 void __fs_reclaim_release(unsigned long ip)
4541 lock_release(&__fs_reclaim_map, ip);
4544 void fs_reclaim_acquire(gfp_t gfp_mask)
4546 gfp_mask = current_gfp_context(gfp_mask);
4548 if (__need_reclaim(gfp_mask)) {
4549 if (gfp_mask & __GFP_FS)
4550 __fs_reclaim_acquire(_RET_IP_);
4552 #ifdef CONFIG_MMU_NOTIFIER
4553 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4554 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4559 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4561 void fs_reclaim_release(gfp_t gfp_mask)
4563 gfp_mask = current_gfp_context(gfp_mask);
4565 if (__need_reclaim(gfp_mask)) {
4566 if (gfp_mask & __GFP_FS)
4567 __fs_reclaim_release(_RET_IP_);
4570 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4573 /* Perform direct synchronous page reclaim */
4574 static unsigned long
4575 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4576 const struct alloc_context *ac)
4578 unsigned int noreclaim_flag;
4579 unsigned long progress;
4583 /* We now go into synchronous reclaim */
4584 cpuset_memory_pressure_bump();
4585 fs_reclaim_acquire(gfp_mask);
4586 noreclaim_flag = memalloc_noreclaim_save();
4588 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4591 memalloc_noreclaim_restore(noreclaim_flag);
4592 fs_reclaim_release(gfp_mask);
4599 /* The really slow allocator path where we enter direct reclaim */
4600 static inline struct page *
4601 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4602 unsigned int alloc_flags, const struct alloc_context *ac,
4603 unsigned long *did_some_progress)
4605 struct page *page = NULL;
4606 unsigned long pflags;
4607 bool drained = false;
4609 psi_memstall_enter(&pflags);
4610 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4611 if (unlikely(!(*did_some_progress)))
4615 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4618 * If an allocation failed after direct reclaim, it could be because
4619 * pages are pinned on the per-cpu lists or in high alloc reserves.
4620 * Shrink them and try again
4622 if (!page && !drained) {
4623 unreserve_highatomic_pageblock(ac, false);
4624 drain_all_pages(NULL);
4629 psi_memstall_leave(&pflags);
4634 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4635 const struct alloc_context *ac)
4639 pg_data_t *last_pgdat = NULL;
4640 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4642 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4644 if (last_pgdat != zone->zone_pgdat)
4645 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4646 last_pgdat = zone->zone_pgdat;
4650 static inline unsigned int
4651 gfp_to_alloc_flags(gfp_t gfp_mask)
4653 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4656 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4657 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4658 * to save two branches.
4660 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4661 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4664 * The caller may dip into page reserves a bit more if the caller
4665 * cannot run direct reclaim, or if the caller has realtime scheduling
4666 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4667 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4669 alloc_flags |= (__force int)
4670 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4672 if (gfp_mask & __GFP_ATOMIC) {
4674 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4675 * if it can't schedule.
4677 if (!(gfp_mask & __GFP_NOMEMALLOC))
4678 alloc_flags |= ALLOC_HARDER;
4680 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4681 * comment for __cpuset_node_allowed().
4683 alloc_flags &= ~ALLOC_CPUSET;
4684 } else if (unlikely(rt_task(current)) && in_task())
4685 alloc_flags |= ALLOC_HARDER;
4687 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4692 static bool oom_reserves_allowed(struct task_struct *tsk)
4694 if (!tsk_is_oom_victim(tsk))
4698 * !MMU doesn't have oom reaper so give access to memory reserves
4699 * only to the thread with TIF_MEMDIE set
4701 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4708 * Distinguish requests which really need access to full memory
4709 * reserves from oom victims which can live with a portion of it
4711 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4713 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4715 if (gfp_mask & __GFP_MEMALLOC)
4716 return ALLOC_NO_WATERMARKS;
4717 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4718 return ALLOC_NO_WATERMARKS;
4719 if (!in_interrupt()) {
4720 if (current->flags & PF_MEMALLOC)
4721 return ALLOC_NO_WATERMARKS;
4722 else if (oom_reserves_allowed(current))
4729 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4731 return !!__gfp_pfmemalloc_flags(gfp_mask);
4735 * Checks whether it makes sense to retry the reclaim to make a forward progress
4736 * for the given allocation request.
4738 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4739 * without success, or when we couldn't even meet the watermark if we
4740 * reclaimed all remaining pages on the LRU lists.
4742 * Returns true if a retry is viable or false to enter the oom path.
4745 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4746 struct alloc_context *ac, int alloc_flags,
4747 bool did_some_progress, int *no_progress_loops)
4754 * Costly allocations might have made a progress but this doesn't mean
4755 * their order will become available due to high fragmentation so
4756 * always increment the no progress counter for them
4758 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4759 *no_progress_loops = 0;
4761 (*no_progress_loops)++;
4764 * Make sure we converge to OOM if we cannot make any progress
4765 * several times in the row.
4767 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4768 /* Before OOM, exhaust highatomic_reserve */
4769 return unreserve_highatomic_pageblock(ac, true);
4773 * Keep reclaiming pages while there is a chance this will lead
4774 * somewhere. If none of the target zones can satisfy our allocation
4775 * request even if all reclaimable pages are considered then we are
4776 * screwed and have to go OOM.
4778 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4779 ac->highest_zoneidx, ac->nodemask) {
4780 unsigned long available;
4781 unsigned long reclaimable;
4782 unsigned long min_wmark = min_wmark_pages(zone);
4785 available = reclaimable = zone_reclaimable_pages(zone);
4786 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4789 * Would the allocation succeed if we reclaimed all
4790 * reclaimable pages?
4792 wmark = __zone_watermark_ok(zone, order, min_wmark,
4793 ac->highest_zoneidx, alloc_flags, available);
4794 trace_reclaim_retry_zone(z, order, reclaimable,
4795 available, min_wmark, *no_progress_loops, wmark);
4803 * Memory allocation/reclaim might be called from a WQ context and the
4804 * current implementation of the WQ concurrency control doesn't
4805 * recognize that a particular WQ is congested if the worker thread is
4806 * looping without ever sleeping. Therefore we have to do a short sleep
4807 * here rather than calling cond_resched().
4809 if (current->flags & PF_WQ_WORKER)
4810 schedule_timeout_uninterruptible(1);
4817 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4820 * It's possible that cpuset's mems_allowed and the nodemask from
4821 * mempolicy don't intersect. This should be normally dealt with by
4822 * policy_nodemask(), but it's possible to race with cpuset update in
4823 * such a way the check therein was true, and then it became false
4824 * before we got our cpuset_mems_cookie here.
4825 * This assumes that for all allocations, ac->nodemask can come only
4826 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4827 * when it does not intersect with the cpuset restrictions) or the
4828 * caller can deal with a violated nodemask.
4830 if (cpusets_enabled() && ac->nodemask &&
4831 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4832 ac->nodemask = NULL;
4837 * When updating a task's mems_allowed or mempolicy nodemask, it is
4838 * possible to race with parallel threads in such a way that our
4839 * allocation can fail while the mask is being updated. If we are about
4840 * to fail, check if the cpuset changed during allocation and if so,
4843 if (read_mems_allowed_retry(cpuset_mems_cookie))
4849 static inline struct page *
4850 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4851 struct alloc_context *ac)
4853 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4854 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4855 struct page *page = NULL;
4856 unsigned int alloc_flags;
4857 unsigned long did_some_progress;
4858 enum compact_priority compact_priority;
4859 enum compact_result compact_result;
4860 int compaction_retries;
4861 int no_progress_loops;
4862 unsigned int cpuset_mems_cookie;
4866 * We also sanity check to catch abuse of atomic reserves being used by
4867 * callers that are not in atomic context.
4869 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4870 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4871 gfp_mask &= ~__GFP_ATOMIC;
4874 compaction_retries = 0;
4875 no_progress_loops = 0;
4876 compact_priority = DEF_COMPACT_PRIORITY;
4877 cpuset_mems_cookie = read_mems_allowed_begin();
4880 * The fast path uses conservative alloc_flags to succeed only until
4881 * kswapd needs to be woken up, and to avoid the cost of setting up
4882 * alloc_flags precisely. So we do that now.
4884 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4887 * We need to recalculate the starting point for the zonelist iterator
4888 * because we might have used different nodemask in the fast path, or
4889 * there was a cpuset modification and we are retrying - otherwise we
4890 * could end up iterating over non-eligible zones endlessly.
4892 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4893 ac->highest_zoneidx, ac->nodemask);
4894 if (!ac->preferred_zoneref->zone)
4898 * Check for insane configurations where the cpuset doesn't contain
4899 * any suitable zone to satisfy the request - e.g. non-movable
4900 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4902 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4903 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4904 ac->highest_zoneidx,
4905 &cpuset_current_mems_allowed);
4910 if (alloc_flags & ALLOC_KSWAPD)
4911 wake_all_kswapds(order, gfp_mask, ac);
4914 * The adjusted alloc_flags might result in immediate success, so try
4917 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4922 * For costly allocations, try direct compaction first, as it's likely
4923 * that we have enough base pages and don't need to reclaim. For non-
4924 * movable high-order allocations, do that as well, as compaction will
4925 * try prevent permanent fragmentation by migrating from blocks of the
4927 * Don't try this for allocations that are allowed to ignore
4928 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4930 if (can_direct_reclaim &&
4932 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4933 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4934 page = __alloc_pages_direct_compact(gfp_mask, order,
4936 INIT_COMPACT_PRIORITY,
4942 * Checks for costly allocations with __GFP_NORETRY, which
4943 * includes some THP page fault allocations
4945 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4947 * If allocating entire pageblock(s) and compaction
4948 * failed because all zones are below low watermarks
4949 * or is prohibited because it recently failed at this
4950 * order, fail immediately unless the allocator has
4951 * requested compaction and reclaim retry.
4954 * - potentially very expensive because zones are far
4955 * below their low watermarks or this is part of very
4956 * bursty high order allocations,
4957 * - not guaranteed to help because isolate_freepages()
4958 * may not iterate over freed pages as part of its
4960 * - unlikely to make entire pageblocks free on its
4963 if (compact_result == COMPACT_SKIPPED ||
4964 compact_result == COMPACT_DEFERRED)
4968 * Looks like reclaim/compaction is worth trying, but
4969 * sync compaction could be very expensive, so keep
4970 * using async compaction.
4972 compact_priority = INIT_COMPACT_PRIORITY;
4977 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4978 if (alloc_flags & ALLOC_KSWAPD)
4979 wake_all_kswapds(order, gfp_mask, ac);
4981 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4983 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
4986 * Reset the nodemask and zonelist iterators if memory policies can be
4987 * ignored. These allocations are high priority and system rather than
4990 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4991 ac->nodemask = NULL;
4992 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4993 ac->highest_zoneidx, ac->nodemask);
4996 /* Attempt with potentially adjusted zonelist and alloc_flags */
4997 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5001 /* Caller is not willing to reclaim, we can't balance anything */
5002 if (!can_direct_reclaim)
5005 /* Avoid recursion of direct reclaim */
5006 if (current->flags & PF_MEMALLOC)
5009 /* Try direct reclaim and then allocating */
5010 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5011 &did_some_progress);
5015 /* Try direct compaction and then allocating */
5016 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5017 compact_priority, &compact_result);
5021 /* Do not loop if specifically requested */
5022 if (gfp_mask & __GFP_NORETRY)
5026 * Do not retry costly high order allocations unless they are
5027 * __GFP_RETRY_MAYFAIL
5029 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5032 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5033 did_some_progress > 0, &no_progress_loops))
5037 * It doesn't make any sense to retry for the compaction if the order-0
5038 * reclaim is not able to make any progress because the current
5039 * implementation of the compaction depends on the sufficient amount
5040 * of free memory (see __compaction_suitable)
5042 if (did_some_progress > 0 &&
5043 should_compact_retry(ac, order, alloc_flags,
5044 compact_result, &compact_priority,
5045 &compaction_retries))
5049 /* Deal with possible cpuset update races before we start OOM killing */
5050 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5053 /* Reclaim has failed us, start killing things */
5054 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5058 /* Avoid allocations with no watermarks from looping endlessly */
5059 if (tsk_is_oom_victim(current) &&
5060 (alloc_flags & ALLOC_OOM ||
5061 (gfp_mask & __GFP_NOMEMALLOC)))
5064 /* Retry as long as the OOM killer is making progress */
5065 if (did_some_progress) {
5066 no_progress_loops = 0;
5071 /* Deal with possible cpuset update races before we fail */
5072 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5076 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5079 if (gfp_mask & __GFP_NOFAIL) {
5081 * All existing users of the __GFP_NOFAIL are blockable, so warn
5082 * of any new users that actually require GFP_NOWAIT
5084 if (WARN_ON_ONCE(!can_direct_reclaim))
5088 * PF_MEMALLOC request from this context is rather bizarre
5089 * because we cannot reclaim anything and only can loop waiting
5090 * for somebody to do a work for us
5092 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5095 * non failing costly orders are a hard requirement which we
5096 * are not prepared for much so let's warn about these users
5097 * so that we can identify them and convert them to something
5100 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5103 * Help non-failing allocations by giving them access to memory
5104 * reserves but do not use ALLOC_NO_WATERMARKS because this
5105 * could deplete whole memory reserves which would just make
5106 * the situation worse
5108 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5116 warn_alloc(gfp_mask, ac->nodemask,
5117 "page allocation failure: order:%u", order);
5122 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5123 int preferred_nid, nodemask_t *nodemask,
5124 struct alloc_context *ac, gfp_t *alloc_gfp,
5125 unsigned int *alloc_flags)
5127 ac->highest_zoneidx = gfp_zone(gfp_mask);
5128 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5129 ac->nodemask = nodemask;
5130 ac->migratetype = gfp_migratetype(gfp_mask);
5132 if (cpusets_enabled()) {
5133 *alloc_gfp |= __GFP_HARDWALL;
5135 * When we are in the interrupt context, it is irrelevant
5136 * to the current task context. It means that any node ok.
5138 if (in_task() && !ac->nodemask)
5139 ac->nodemask = &cpuset_current_mems_allowed;
5141 *alloc_flags |= ALLOC_CPUSET;
5144 fs_reclaim_acquire(gfp_mask);
5145 fs_reclaim_release(gfp_mask);
5147 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5149 if (should_fail_alloc_page(gfp_mask, order))
5152 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5154 /* Dirty zone balancing only done in the fast path */
5155 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5158 * The preferred zone is used for statistics but crucially it is
5159 * also used as the starting point for the zonelist iterator. It
5160 * may get reset for allocations that ignore memory policies.
5162 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5163 ac->highest_zoneidx, ac->nodemask);
5169 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5170 * @gfp: GFP flags for the allocation
5171 * @preferred_nid: The preferred NUMA node ID to allocate from
5172 * @nodemask: Set of nodes to allocate from, may be NULL
5173 * @nr_pages: The number of pages desired on the list or array
5174 * @page_list: Optional list to store the allocated pages
5175 * @page_array: Optional array to store the pages
5177 * This is a batched version of the page allocator that attempts to
5178 * allocate nr_pages quickly. Pages are added to page_list if page_list
5179 * is not NULL, otherwise it is assumed that the page_array is valid.
5181 * For lists, nr_pages is the number of pages that should be allocated.
5183 * For arrays, only NULL elements are populated with pages and nr_pages
5184 * is the maximum number of pages that will be stored in the array.
5186 * Returns the number of pages on the list or array.
5188 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5189 nodemask_t *nodemask, int nr_pages,
5190 struct list_head *page_list,
5191 struct page **page_array)
5194 unsigned long flags;
5197 struct per_cpu_pages *pcp;
5198 struct list_head *pcp_list;
5199 struct alloc_context ac;
5201 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5202 int nr_populated = 0, nr_account = 0;
5205 * Skip populated array elements to determine if any pages need
5206 * to be allocated before disabling IRQs.
5208 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5211 /* No pages requested? */
5212 if (unlikely(nr_pages <= 0))
5215 /* Already populated array? */
5216 if (unlikely(page_array && nr_pages - nr_populated == 0))
5219 /* Bulk allocator does not support memcg accounting. */
5220 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5223 /* Use the single page allocator for one page. */
5224 if (nr_pages - nr_populated == 1)
5227 #ifdef CONFIG_PAGE_OWNER
5229 * PAGE_OWNER may recurse into the allocator to allocate space to
5230 * save the stack with pagesets.lock held. Releasing/reacquiring
5231 * removes much of the performance benefit of bulk allocation so
5232 * force the caller to allocate one page at a time as it'll have
5233 * similar performance to added complexity to the bulk allocator.
5235 if (static_branch_unlikely(&page_owner_inited))
5239 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5240 gfp &= gfp_allowed_mask;
5242 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5246 /* Find an allowed local zone that meets the low watermark. */
5247 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5250 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5251 !__cpuset_zone_allowed(zone, gfp)) {
5255 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5256 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5260 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5261 if (zone_watermark_fast(zone, 0, mark,
5262 zonelist_zone_idx(ac.preferred_zoneref),
5263 alloc_flags, gfp)) {
5269 * If there are no allowed local zones that meets the watermarks then
5270 * try to allocate a single page and reclaim if necessary.
5272 if (unlikely(!zone))
5275 /* Attempt the batch allocation */
5276 local_lock_irqsave(&pagesets.lock, flags);
5277 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5278 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5280 while (nr_populated < nr_pages) {
5282 /* Skip existing pages */
5283 if (page_array && page_array[nr_populated]) {
5288 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5290 if (unlikely(!page)) {
5291 /* Try and get at least one page */
5298 prep_new_page(page, 0, gfp, 0);
5300 list_add(&page->lru, page_list);
5302 page_array[nr_populated] = page;
5306 local_unlock_irqrestore(&pagesets.lock, flags);
5308 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5309 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5312 return nr_populated;
5315 local_unlock_irqrestore(&pagesets.lock, flags);
5318 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5321 list_add(&page->lru, page_list);
5323 page_array[nr_populated] = page;
5329 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5332 * This is the 'heart' of the zoned buddy allocator.
5334 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5335 nodemask_t *nodemask)
5338 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5339 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5340 struct alloc_context ac = { };
5343 * There are several places where we assume that the order value is sane
5344 * so bail out early if the request is out of bound.
5346 if (unlikely(order >= MAX_ORDER)) {
5347 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5351 gfp &= gfp_allowed_mask;
5353 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5354 * resp. GFP_NOIO which has to be inherited for all allocation requests
5355 * from a particular context which has been marked by
5356 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5357 * movable zones are not used during allocation.
5359 gfp = current_gfp_context(gfp);
5361 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5362 &alloc_gfp, &alloc_flags))
5366 * Forbid the first pass from falling back to types that fragment
5367 * memory until all local zones are considered.
5369 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5371 /* First allocation attempt */
5372 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5377 ac.spread_dirty_pages = false;
5380 * Restore the original nodemask if it was potentially replaced with
5381 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5383 ac.nodemask = nodemask;
5385 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5388 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5389 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5390 __free_pages(page, order);
5394 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5398 EXPORT_SYMBOL(__alloc_pages);
5400 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5401 nodemask_t *nodemask)
5403 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5404 preferred_nid, nodemask);
5406 if (page && order > 1)
5407 prep_transhuge_page(page);
5408 return (struct folio *)page;
5410 EXPORT_SYMBOL(__folio_alloc);
5413 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5414 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5415 * you need to access high mem.
5417 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5421 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5424 return (unsigned long) page_address(page);
5426 EXPORT_SYMBOL(__get_free_pages);
5428 unsigned long get_zeroed_page(gfp_t gfp_mask)
5430 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5432 EXPORT_SYMBOL(get_zeroed_page);
5435 * __free_pages - Free pages allocated with alloc_pages().
5436 * @page: The page pointer returned from alloc_pages().
5437 * @order: The order of the allocation.
5439 * This function can free multi-page allocations that are not compound
5440 * pages. It does not check that the @order passed in matches that of
5441 * the allocation, so it is easy to leak memory. Freeing more memory
5442 * than was allocated will probably emit a warning.
5444 * If the last reference to this page is speculative, it will be released
5445 * by put_page() which only frees the first page of a non-compound
5446 * allocation. To prevent the remaining pages from being leaked, we free
5447 * the subsequent pages here. If you want to use the page's reference
5448 * count to decide when to free the allocation, you should allocate a
5449 * compound page, and use put_page() instead of __free_pages().
5451 * Context: May be called in interrupt context or while holding a normal
5452 * spinlock, but not in NMI context or while holding a raw spinlock.
5454 void __free_pages(struct page *page, unsigned int order)
5456 if (put_page_testzero(page))
5457 free_the_page(page, order);
5458 else if (!PageHead(page))
5460 free_the_page(page + (1 << order), order);
5462 EXPORT_SYMBOL(__free_pages);
5464 void free_pages(unsigned long addr, unsigned int order)
5467 VM_BUG_ON(!virt_addr_valid((void *)addr));
5468 __free_pages(virt_to_page((void *)addr), order);
5472 EXPORT_SYMBOL(free_pages);
5476 * An arbitrary-length arbitrary-offset area of memory which resides
5477 * within a 0 or higher order page. Multiple fragments within that page
5478 * are individually refcounted, in the page's reference counter.
5480 * The page_frag functions below provide a simple allocation framework for
5481 * page fragments. This is used by the network stack and network device
5482 * drivers to provide a backing region of memory for use as either an
5483 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5485 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5488 struct page *page = NULL;
5489 gfp_t gfp = gfp_mask;
5491 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5492 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5494 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5495 PAGE_FRAG_CACHE_MAX_ORDER);
5496 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5498 if (unlikely(!page))
5499 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5501 nc->va = page ? page_address(page) : NULL;
5506 void __page_frag_cache_drain(struct page *page, unsigned int count)
5508 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5510 if (page_ref_sub_and_test(page, count))
5511 free_the_page(page, compound_order(page));
5513 EXPORT_SYMBOL(__page_frag_cache_drain);
5515 void *page_frag_alloc_align(struct page_frag_cache *nc,
5516 unsigned int fragsz, gfp_t gfp_mask,
5517 unsigned int align_mask)
5519 unsigned int size = PAGE_SIZE;
5523 if (unlikely(!nc->va)) {
5525 page = __page_frag_cache_refill(nc, gfp_mask);
5529 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5530 /* if size can vary use size else just use PAGE_SIZE */
5533 /* Even if we own the page, we do not use atomic_set().
5534 * This would break get_page_unless_zero() users.
5536 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5538 /* reset page count bias and offset to start of new frag */
5539 nc->pfmemalloc = page_is_pfmemalloc(page);
5540 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5544 offset = nc->offset - fragsz;
5545 if (unlikely(offset < 0)) {
5546 page = virt_to_page(nc->va);
5548 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5551 if (unlikely(nc->pfmemalloc)) {
5552 free_the_page(page, compound_order(page));
5556 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5557 /* if size can vary use size else just use PAGE_SIZE */
5560 /* OK, page count is 0, we can safely set it */
5561 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5563 /* reset page count bias and offset to start of new frag */
5564 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5565 offset = size - fragsz;
5569 offset &= align_mask;
5570 nc->offset = offset;
5572 return nc->va + offset;
5574 EXPORT_SYMBOL(page_frag_alloc_align);
5577 * Frees a page fragment allocated out of either a compound or order 0 page.
5579 void page_frag_free(void *addr)
5581 struct page *page = virt_to_head_page(addr);
5583 if (unlikely(put_page_testzero(page)))
5584 free_the_page(page, compound_order(page));
5586 EXPORT_SYMBOL(page_frag_free);
5588 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5592 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5593 unsigned long used = addr + PAGE_ALIGN(size);
5595 split_page(virt_to_page((void *)addr), order);
5596 while (used < alloc_end) {
5601 return (void *)addr;
5605 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5606 * @size: the number of bytes to allocate
5607 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5609 * This function is similar to alloc_pages(), except that it allocates the
5610 * minimum number of pages to satisfy the request. alloc_pages() can only
5611 * allocate memory in power-of-two pages.
5613 * This function is also limited by MAX_ORDER.
5615 * Memory allocated by this function must be released by free_pages_exact().
5617 * Return: pointer to the allocated area or %NULL in case of error.
5619 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5621 unsigned int order = get_order(size);
5624 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5625 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5627 addr = __get_free_pages(gfp_mask, order);
5628 return make_alloc_exact(addr, order, size);
5630 EXPORT_SYMBOL(alloc_pages_exact);
5633 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5635 * @nid: the preferred node ID where memory should be allocated
5636 * @size: the number of bytes to allocate
5637 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5639 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5642 * Return: pointer to the allocated area or %NULL in case of error.
5644 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5646 unsigned int order = get_order(size);
5649 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5650 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5652 p = alloc_pages_node(nid, gfp_mask, order);
5655 return make_alloc_exact((unsigned long)page_address(p), order, size);
5659 * free_pages_exact - release memory allocated via alloc_pages_exact()
5660 * @virt: the value returned by alloc_pages_exact.
5661 * @size: size of allocation, same value as passed to alloc_pages_exact().
5663 * Release the memory allocated by a previous call to alloc_pages_exact.
5665 void free_pages_exact(void *virt, size_t size)
5667 unsigned long addr = (unsigned long)virt;
5668 unsigned long end = addr + PAGE_ALIGN(size);
5670 while (addr < end) {
5675 EXPORT_SYMBOL(free_pages_exact);
5678 * nr_free_zone_pages - count number of pages beyond high watermark
5679 * @offset: The zone index of the highest zone
5681 * nr_free_zone_pages() counts the number of pages which are beyond the
5682 * high watermark within all zones at or below a given zone index. For each
5683 * zone, the number of pages is calculated as:
5685 * nr_free_zone_pages = managed_pages - high_pages
5687 * Return: number of pages beyond high watermark.
5689 static unsigned long nr_free_zone_pages(int offset)
5694 /* Just pick one node, since fallback list is circular */
5695 unsigned long sum = 0;
5697 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5699 for_each_zone_zonelist(zone, z, zonelist, offset) {
5700 unsigned long size = zone_managed_pages(zone);
5701 unsigned long high = high_wmark_pages(zone);
5710 * nr_free_buffer_pages - count number of pages beyond high watermark
5712 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5713 * watermark within ZONE_DMA and ZONE_NORMAL.
5715 * Return: number of pages beyond high watermark within ZONE_DMA and
5718 unsigned long nr_free_buffer_pages(void)
5720 return nr_free_zone_pages(gfp_zone(GFP_USER));
5722 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5724 static inline void show_node(struct zone *zone)
5726 if (IS_ENABLED(CONFIG_NUMA))
5727 printk("Node %d ", zone_to_nid(zone));
5730 long si_mem_available(void)
5733 unsigned long pagecache;
5734 unsigned long wmark_low = 0;
5735 unsigned long pages[NR_LRU_LISTS];
5736 unsigned long reclaimable;
5740 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5741 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5744 wmark_low += low_wmark_pages(zone);
5747 * Estimate the amount of memory available for userspace allocations,
5748 * without causing swapping.
5750 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5753 * Not all the page cache can be freed, otherwise the system will
5754 * start swapping. Assume at least half of the page cache, or the
5755 * low watermark worth of cache, needs to stay.
5757 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5758 pagecache -= min(pagecache / 2, wmark_low);
5759 available += pagecache;
5762 * Part of the reclaimable slab and other kernel memory consists of
5763 * items that are in use, and cannot be freed. Cap this estimate at the
5766 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5767 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5768 available += reclaimable - min(reclaimable / 2, wmark_low);
5774 EXPORT_SYMBOL_GPL(si_mem_available);
5776 void si_meminfo(struct sysinfo *val)
5778 val->totalram = totalram_pages();
5779 val->sharedram = global_node_page_state(NR_SHMEM);
5780 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5781 val->bufferram = nr_blockdev_pages();
5782 val->totalhigh = totalhigh_pages();
5783 val->freehigh = nr_free_highpages();
5784 val->mem_unit = PAGE_SIZE;
5787 EXPORT_SYMBOL(si_meminfo);
5790 void si_meminfo_node(struct sysinfo *val, int nid)
5792 int zone_type; /* needs to be signed */
5793 unsigned long managed_pages = 0;
5794 unsigned long managed_highpages = 0;
5795 unsigned long free_highpages = 0;
5796 pg_data_t *pgdat = NODE_DATA(nid);
5798 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5799 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5800 val->totalram = managed_pages;
5801 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5802 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5803 #ifdef CONFIG_HIGHMEM
5804 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5805 struct zone *zone = &pgdat->node_zones[zone_type];
5807 if (is_highmem(zone)) {
5808 managed_highpages += zone_managed_pages(zone);
5809 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5812 val->totalhigh = managed_highpages;
5813 val->freehigh = free_highpages;
5815 val->totalhigh = managed_highpages;
5816 val->freehigh = free_highpages;
5818 val->mem_unit = PAGE_SIZE;
5823 * Determine whether the node should be displayed or not, depending on whether
5824 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5826 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5828 if (!(flags & SHOW_MEM_FILTER_NODES))
5832 * no node mask - aka implicit memory numa policy. Do not bother with
5833 * the synchronization - read_mems_allowed_begin - because we do not
5834 * have to be precise here.
5837 nodemask = &cpuset_current_mems_allowed;
5839 return !node_isset(nid, *nodemask);
5842 #define K(x) ((x) << (PAGE_SHIFT-10))
5844 static void show_migration_types(unsigned char type)
5846 static const char types[MIGRATE_TYPES] = {
5847 [MIGRATE_UNMOVABLE] = 'U',
5848 [MIGRATE_MOVABLE] = 'M',
5849 [MIGRATE_RECLAIMABLE] = 'E',
5850 [MIGRATE_HIGHATOMIC] = 'H',
5852 [MIGRATE_CMA] = 'C',
5854 #ifdef CONFIG_MEMORY_ISOLATION
5855 [MIGRATE_ISOLATE] = 'I',
5858 char tmp[MIGRATE_TYPES + 1];
5862 for (i = 0; i < MIGRATE_TYPES; i++) {
5863 if (type & (1 << i))
5868 printk(KERN_CONT "(%s) ", tmp);
5872 * Show free area list (used inside shift_scroll-lock stuff)
5873 * We also calculate the percentage fragmentation. We do this by counting the
5874 * memory on each free list with the exception of the first item on the list.
5877 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5880 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5882 unsigned long free_pcp = 0;
5887 for_each_populated_zone(zone) {
5888 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5891 for_each_online_cpu(cpu)
5892 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5895 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5896 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5897 " unevictable:%lu dirty:%lu writeback:%lu\n"
5898 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5899 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5900 " kernel_misc_reclaimable:%lu\n"
5901 " free:%lu free_pcp:%lu free_cma:%lu\n",
5902 global_node_page_state(NR_ACTIVE_ANON),
5903 global_node_page_state(NR_INACTIVE_ANON),
5904 global_node_page_state(NR_ISOLATED_ANON),
5905 global_node_page_state(NR_ACTIVE_FILE),
5906 global_node_page_state(NR_INACTIVE_FILE),
5907 global_node_page_state(NR_ISOLATED_FILE),
5908 global_node_page_state(NR_UNEVICTABLE),
5909 global_node_page_state(NR_FILE_DIRTY),
5910 global_node_page_state(NR_WRITEBACK),
5911 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5912 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5913 global_node_page_state(NR_FILE_MAPPED),
5914 global_node_page_state(NR_SHMEM),
5915 global_node_page_state(NR_PAGETABLE),
5916 global_zone_page_state(NR_BOUNCE),
5917 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5918 global_zone_page_state(NR_FREE_PAGES),
5920 global_zone_page_state(NR_FREE_CMA_PAGES));
5922 for_each_online_pgdat(pgdat) {
5923 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5927 " active_anon:%lukB"
5928 " inactive_anon:%lukB"
5929 " active_file:%lukB"
5930 " inactive_file:%lukB"
5931 " unevictable:%lukB"
5932 " isolated(anon):%lukB"
5933 " isolated(file):%lukB"
5938 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5940 " shmem_pmdmapped: %lukB"
5943 " writeback_tmp:%lukB"
5944 " kernel_stack:%lukB"
5945 #ifdef CONFIG_SHADOW_CALL_STACK
5946 " shadow_call_stack:%lukB"
5949 " all_unreclaimable? %s"
5952 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5953 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5954 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5955 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5956 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5957 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5958 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5959 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5960 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5961 K(node_page_state(pgdat, NR_WRITEBACK)),
5962 K(node_page_state(pgdat, NR_SHMEM)),
5963 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5964 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5965 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5966 K(node_page_state(pgdat, NR_ANON_THPS)),
5968 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5969 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5970 #ifdef CONFIG_SHADOW_CALL_STACK
5971 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5973 K(node_page_state(pgdat, NR_PAGETABLE)),
5974 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5978 for_each_populated_zone(zone) {
5981 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5985 for_each_online_cpu(cpu)
5986 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5996 " reserved_highatomic:%luKB"
5997 " active_anon:%lukB"
5998 " inactive_anon:%lukB"
5999 " active_file:%lukB"
6000 " inactive_file:%lukB"
6001 " unevictable:%lukB"
6002 " writepending:%lukB"
6012 K(zone_page_state(zone, NR_FREE_PAGES)),
6013 K(zone->watermark_boost),
6014 K(min_wmark_pages(zone)),
6015 K(low_wmark_pages(zone)),
6016 K(high_wmark_pages(zone)),
6017 K(zone->nr_reserved_highatomic),
6018 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6019 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6020 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6021 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6022 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6023 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6024 K(zone->present_pages),
6025 K(zone_managed_pages(zone)),
6026 K(zone_page_state(zone, NR_MLOCK)),
6027 K(zone_page_state(zone, NR_BOUNCE)),
6029 K(this_cpu_read(zone->per_cpu_pageset->count)),
6030 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6031 printk("lowmem_reserve[]:");
6032 for (i = 0; i < MAX_NR_ZONES; i++)
6033 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6034 printk(KERN_CONT "\n");
6037 for_each_populated_zone(zone) {
6039 unsigned long nr[MAX_ORDER], flags, total = 0;
6040 unsigned char types[MAX_ORDER];
6042 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6045 printk(KERN_CONT "%s: ", zone->name);
6047 spin_lock_irqsave(&zone->lock, flags);
6048 for (order = 0; order < MAX_ORDER; order++) {
6049 struct free_area *area = &zone->free_area[order];
6052 nr[order] = area->nr_free;
6053 total += nr[order] << order;
6056 for (type = 0; type < MIGRATE_TYPES; type++) {
6057 if (!free_area_empty(area, type))
6058 types[order] |= 1 << type;
6061 spin_unlock_irqrestore(&zone->lock, flags);
6062 for (order = 0; order < MAX_ORDER; order++) {
6063 printk(KERN_CONT "%lu*%lukB ",
6064 nr[order], K(1UL) << order);
6066 show_migration_types(types[order]);
6068 printk(KERN_CONT "= %lukB\n", K(total));
6071 hugetlb_show_meminfo();
6073 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6075 show_swap_cache_info();
6078 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6080 zoneref->zone = zone;
6081 zoneref->zone_idx = zone_idx(zone);
6085 * Builds allocation fallback zone lists.
6087 * Add all populated zones of a node to the zonelist.
6089 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6092 enum zone_type zone_type = MAX_NR_ZONES;
6097 zone = pgdat->node_zones + zone_type;
6098 if (managed_zone(zone)) {
6099 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6100 check_highest_zone(zone_type);
6102 } while (zone_type);
6109 static int __parse_numa_zonelist_order(char *s)
6112 * We used to support different zonelists modes but they turned
6113 * out to be just not useful. Let's keep the warning in place
6114 * if somebody still use the cmd line parameter so that we do
6115 * not fail it silently
6117 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6118 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6124 char numa_zonelist_order[] = "Node";
6127 * sysctl handler for numa_zonelist_order
6129 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6130 void *buffer, size_t *length, loff_t *ppos)
6133 return __parse_numa_zonelist_order(buffer);
6134 return proc_dostring(table, write, buffer, length, ppos);
6138 #define MAX_NODE_LOAD (nr_online_nodes)
6139 static int node_load[MAX_NUMNODES];
6142 * find_next_best_node - find the next node that should appear in a given node's fallback list
6143 * @node: node whose fallback list we're appending
6144 * @used_node_mask: nodemask_t of already used nodes
6146 * We use a number of factors to determine which is the next node that should
6147 * appear on a given node's fallback list. The node should not have appeared
6148 * already in @node's fallback list, and it should be the next closest node
6149 * according to the distance array (which contains arbitrary distance values
6150 * from each node to each node in the system), and should also prefer nodes
6151 * with no CPUs, since presumably they'll have very little allocation pressure
6152 * on them otherwise.
6154 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6156 int find_next_best_node(int node, nodemask_t *used_node_mask)
6159 int min_val = INT_MAX;
6160 int best_node = NUMA_NO_NODE;
6162 /* Use the local node if we haven't already */
6163 if (!node_isset(node, *used_node_mask)) {
6164 node_set(node, *used_node_mask);
6168 for_each_node_state(n, N_MEMORY) {
6170 /* Don't want a node to appear more than once */
6171 if (node_isset(n, *used_node_mask))
6174 /* Use the distance array to find the distance */
6175 val = node_distance(node, n);
6177 /* Penalize nodes under us ("prefer the next node") */
6180 /* Give preference to headless and unused nodes */
6181 if (!cpumask_empty(cpumask_of_node(n)))
6182 val += PENALTY_FOR_NODE_WITH_CPUS;
6184 /* Slight preference for less loaded node */
6185 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6186 val += node_load[n];
6188 if (val < min_val) {
6195 node_set(best_node, *used_node_mask);
6202 * Build zonelists ordered by node and zones within node.
6203 * This results in maximum locality--normal zone overflows into local
6204 * DMA zone, if any--but risks exhausting DMA zone.
6206 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6209 struct zoneref *zonerefs;
6212 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6214 for (i = 0; i < nr_nodes; i++) {
6217 pg_data_t *node = NODE_DATA(node_order[i]);
6219 nr_zones = build_zonerefs_node(node, zonerefs);
6220 zonerefs += nr_zones;
6222 zonerefs->zone = NULL;
6223 zonerefs->zone_idx = 0;
6227 * Build gfp_thisnode zonelists
6229 static void build_thisnode_zonelists(pg_data_t *pgdat)
6231 struct zoneref *zonerefs;
6234 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6235 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6236 zonerefs += nr_zones;
6237 zonerefs->zone = NULL;
6238 zonerefs->zone_idx = 0;
6242 * Build zonelists ordered by zone and nodes within zones.
6243 * This results in conserving DMA zone[s] until all Normal memory is
6244 * exhausted, but results in overflowing to remote node while memory
6245 * may still exist in local DMA zone.
6248 static void build_zonelists(pg_data_t *pgdat)
6250 static int node_order[MAX_NUMNODES];
6251 int node, load, nr_nodes = 0;
6252 nodemask_t used_mask = NODE_MASK_NONE;
6253 int local_node, prev_node;
6255 /* NUMA-aware ordering of nodes */
6256 local_node = pgdat->node_id;
6257 load = nr_online_nodes;
6258 prev_node = local_node;
6260 memset(node_order, 0, sizeof(node_order));
6261 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6263 * We don't want to pressure a particular node.
6264 * So adding penalty to the first node in same
6265 * distance group to make it round-robin.
6267 if (node_distance(local_node, node) !=
6268 node_distance(local_node, prev_node))
6269 node_load[node] += load;
6271 node_order[nr_nodes++] = node;
6276 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6277 build_thisnode_zonelists(pgdat);
6278 pr_info("Fallback order for Node %d: ", local_node);
6279 for (node = 0; node < nr_nodes; node++)
6280 pr_cont("%d ", node_order[node]);
6284 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6286 * Return node id of node used for "local" allocations.
6287 * I.e., first node id of first zone in arg node's generic zonelist.
6288 * Used for initializing percpu 'numa_mem', which is used primarily
6289 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6291 int local_memory_node(int node)
6295 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6296 gfp_zone(GFP_KERNEL),
6298 return zone_to_nid(z->zone);
6302 static void setup_min_unmapped_ratio(void);
6303 static void setup_min_slab_ratio(void);
6304 #else /* CONFIG_NUMA */
6306 static void build_zonelists(pg_data_t *pgdat)
6308 int node, local_node;
6309 struct zoneref *zonerefs;
6312 local_node = pgdat->node_id;
6314 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6315 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6316 zonerefs += nr_zones;
6319 * Now we build the zonelist so that it contains the zones
6320 * of all the other nodes.
6321 * We don't want to pressure a particular node, so when
6322 * building the zones for node N, we make sure that the
6323 * zones coming right after the local ones are those from
6324 * node N+1 (modulo N)
6326 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6327 if (!node_online(node))
6329 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6330 zonerefs += nr_zones;
6332 for (node = 0; node < local_node; node++) {
6333 if (!node_online(node))
6335 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6336 zonerefs += nr_zones;
6339 zonerefs->zone = NULL;
6340 zonerefs->zone_idx = 0;
6343 #endif /* CONFIG_NUMA */
6346 * Boot pageset table. One per cpu which is going to be used for all
6347 * zones and all nodes. The parameters will be set in such a way
6348 * that an item put on a list will immediately be handed over to
6349 * the buddy list. This is safe since pageset manipulation is done
6350 * with interrupts disabled.
6352 * The boot_pagesets must be kept even after bootup is complete for
6353 * unused processors and/or zones. They do play a role for bootstrapping
6354 * hotplugged processors.
6356 * zoneinfo_show() and maybe other functions do
6357 * not check if the processor is online before following the pageset pointer.
6358 * Other parts of the kernel may not check if the zone is available.
6360 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6361 /* These effectively disable the pcplists in the boot pageset completely */
6362 #define BOOT_PAGESET_HIGH 0
6363 #define BOOT_PAGESET_BATCH 1
6364 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6365 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6366 DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6368 static void __build_all_zonelists(void *data)
6371 int __maybe_unused cpu;
6372 pg_data_t *self = data;
6373 static DEFINE_SPINLOCK(lock);
6378 memset(node_load, 0, sizeof(node_load));
6382 * This node is hotadded and no memory is yet present. So just
6383 * building zonelists is fine - no need to touch other nodes.
6385 if (self && !node_online(self->node_id)) {
6386 build_zonelists(self);
6389 * All possible nodes have pgdat preallocated
6392 for_each_node(nid) {
6393 pg_data_t *pgdat = NODE_DATA(nid);
6395 build_zonelists(pgdat);
6398 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6400 * We now know the "local memory node" for each node--
6401 * i.e., the node of the first zone in the generic zonelist.
6402 * Set up numa_mem percpu variable for on-line cpus. During
6403 * boot, only the boot cpu should be on-line; we'll init the
6404 * secondary cpus' numa_mem as they come on-line. During
6405 * node/memory hotplug, we'll fixup all on-line cpus.
6407 for_each_online_cpu(cpu)
6408 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6415 static noinline void __init
6416 build_all_zonelists_init(void)
6420 __build_all_zonelists(NULL);
6423 * Initialize the boot_pagesets that are going to be used
6424 * for bootstrapping processors. The real pagesets for
6425 * each zone will be allocated later when the per cpu
6426 * allocator is available.
6428 * boot_pagesets are used also for bootstrapping offline
6429 * cpus if the system is already booted because the pagesets
6430 * are needed to initialize allocators on a specific cpu too.
6431 * F.e. the percpu allocator needs the page allocator which
6432 * needs the percpu allocator in order to allocate its pagesets
6433 * (a chicken-egg dilemma).
6435 for_each_possible_cpu(cpu)
6436 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6438 mminit_verify_zonelist();
6439 cpuset_init_current_mems_allowed();
6443 * unless system_state == SYSTEM_BOOTING.
6445 * __ref due to call of __init annotated helper build_all_zonelists_init
6446 * [protected by SYSTEM_BOOTING].
6448 void __ref build_all_zonelists(pg_data_t *pgdat)
6450 unsigned long vm_total_pages;
6452 if (system_state == SYSTEM_BOOTING) {
6453 build_all_zonelists_init();
6455 __build_all_zonelists(pgdat);
6456 /* cpuset refresh routine should be here */
6458 /* Get the number of free pages beyond high watermark in all zones. */
6459 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6461 * Disable grouping by mobility if the number of pages in the
6462 * system is too low to allow the mechanism to work. It would be
6463 * more accurate, but expensive to check per-zone. This check is
6464 * made on memory-hotadd so a system can start with mobility
6465 * disabled and enable it later
6467 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6468 page_group_by_mobility_disabled = 1;
6470 page_group_by_mobility_disabled = 0;
6472 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6474 page_group_by_mobility_disabled ? "off" : "on",
6477 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6481 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6482 static bool __meminit
6483 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6485 static struct memblock_region *r;
6487 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6488 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6489 for_each_mem_region(r) {
6490 if (*pfn < memblock_region_memory_end_pfn(r))
6494 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6495 memblock_is_mirror(r)) {
6496 *pfn = memblock_region_memory_end_pfn(r);
6504 * Initially all pages are reserved - free ones are freed
6505 * up by memblock_free_all() once the early boot process is
6506 * done. Non-atomic initialization, single-pass.
6508 * All aligned pageblocks are initialized to the specified migratetype
6509 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6510 * zone stats (e.g., nr_isolate_pageblock) are touched.
6512 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6513 unsigned long start_pfn, unsigned long zone_end_pfn,
6514 enum meminit_context context,
6515 struct vmem_altmap *altmap, int migratetype)
6517 unsigned long pfn, end_pfn = start_pfn + size;
6520 if (highest_memmap_pfn < end_pfn - 1)
6521 highest_memmap_pfn = end_pfn - 1;
6523 #ifdef CONFIG_ZONE_DEVICE
6525 * Honor reservation requested by the driver for this ZONE_DEVICE
6526 * memory. We limit the total number of pages to initialize to just
6527 * those that might contain the memory mapping. We will defer the
6528 * ZONE_DEVICE page initialization until after we have released
6531 if (zone == ZONE_DEVICE) {
6535 if (start_pfn == altmap->base_pfn)
6536 start_pfn += altmap->reserve;
6537 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6541 for (pfn = start_pfn; pfn < end_pfn; ) {
6543 * There can be holes in boot-time mem_map[]s handed to this
6544 * function. They do not exist on hotplugged memory.
6546 if (context == MEMINIT_EARLY) {
6547 if (overlap_memmap_init(zone, &pfn))
6549 if (defer_init(nid, pfn, zone_end_pfn))
6553 page = pfn_to_page(pfn);
6554 __init_single_page(page, pfn, zone, nid);
6555 if (context == MEMINIT_HOTPLUG)
6556 __SetPageReserved(page);
6559 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6560 * such that unmovable allocations won't be scattered all
6561 * over the place during system boot.
6563 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6564 set_pageblock_migratetype(page, migratetype);
6571 #ifdef CONFIG_ZONE_DEVICE
6572 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6573 unsigned long zone_idx, int nid,
6574 struct dev_pagemap *pgmap)
6577 __init_single_page(page, pfn, zone_idx, nid);
6580 * Mark page reserved as it will need to wait for onlining
6581 * phase for it to be fully associated with a zone.
6583 * We can use the non-atomic __set_bit operation for setting
6584 * the flag as we are still initializing the pages.
6586 __SetPageReserved(page);
6589 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6590 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6591 * ever freed or placed on a driver-private list.
6593 page->pgmap = pgmap;
6594 page->zone_device_data = NULL;
6597 * Mark the block movable so that blocks are reserved for
6598 * movable at startup. This will force kernel allocations
6599 * to reserve their blocks rather than leaking throughout
6600 * the address space during boot when many long-lived
6601 * kernel allocations are made.
6603 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6604 * because this is done early in section_activate()
6606 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6607 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6612 static void __ref memmap_init_compound(struct page *head,
6613 unsigned long head_pfn,
6614 unsigned long zone_idx, int nid,
6615 struct dev_pagemap *pgmap,
6616 unsigned long nr_pages)
6618 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6619 unsigned int order = pgmap->vmemmap_shift;
6621 __SetPageHead(head);
6622 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6623 struct page *page = pfn_to_page(pfn);
6625 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6626 prep_compound_tail(head, pfn - head_pfn);
6627 set_page_count(page, 0);
6630 * The first tail page stores compound_mapcount_ptr() and
6631 * compound_order() and the second tail page stores
6632 * compound_pincount_ptr(). Call prep_compound_head() after
6633 * the first and second tail pages have been initialized to
6634 * not have the data overwritten.
6636 if (pfn == head_pfn + 2)
6637 prep_compound_head(head, order);
6641 void __ref memmap_init_zone_device(struct zone *zone,
6642 unsigned long start_pfn,
6643 unsigned long nr_pages,
6644 struct dev_pagemap *pgmap)
6646 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6647 struct pglist_data *pgdat = zone->zone_pgdat;
6648 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6649 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6650 unsigned long zone_idx = zone_idx(zone);
6651 unsigned long start = jiffies;
6652 int nid = pgdat->node_id;
6654 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6658 * The call to memmap_init should have already taken care
6659 * of the pages reserved for the memmap, so we can just jump to
6660 * the end of that region and start processing the device pages.
6663 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6664 nr_pages = end_pfn - start_pfn;
6667 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6668 struct page *page = pfn_to_page(pfn);
6670 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6672 if (pfns_per_compound == 1)
6675 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6679 pr_info("%s initialised %lu pages in %ums\n", __func__,
6680 nr_pages, jiffies_to_msecs(jiffies - start));
6684 static void __meminit zone_init_free_lists(struct zone *zone)
6686 unsigned int order, t;
6687 for_each_migratetype_order(order, t) {
6688 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6689 zone->free_area[order].nr_free = 0;
6694 * Only struct pages that correspond to ranges defined by memblock.memory
6695 * are zeroed and initialized by going through __init_single_page() during
6696 * memmap_init_zone_range().
6698 * But, there could be struct pages that correspond to holes in
6699 * memblock.memory. This can happen because of the following reasons:
6700 * - physical memory bank size is not necessarily the exact multiple of the
6701 * arbitrary section size
6702 * - early reserved memory may not be listed in memblock.memory
6703 * - memory layouts defined with memmap= kernel parameter may not align
6704 * nicely with memmap sections
6706 * Explicitly initialize those struct pages so that:
6707 * - PG_Reserved is set
6708 * - zone and node links point to zone and node that span the page if the
6709 * hole is in the middle of a zone
6710 * - zone and node links point to adjacent zone/node if the hole falls on
6711 * the zone boundary; the pages in such holes will be prepended to the
6712 * zone/node above the hole except for the trailing pages in the last
6713 * section that will be appended to the zone/node below.
6715 static void __init init_unavailable_range(unsigned long spfn,
6722 for (pfn = spfn; pfn < epfn; pfn++) {
6723 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6724 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6725 + pageblock_nr_pages - 1;
6728 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6729 __SetPageReserved(pfn_to_page(pfn));
6734 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6735 node, zone_names[zone], pgcnt);
6738 static void __init memmap_init_zone_range(struct zone *zone,
6739 unsigned long start_pfn,
6740 unsigned long end_pfn,
6741 unsigned long *hole_pfn)
6743 unsigned long zone_start_pfn = zone->zone_start_pfn;
6744 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6745 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6747 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6748 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6750 if (start_pfn >= end_pfn)
6753 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6754 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6756 if (*hole_pfn < start_pfn)
6757 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6759 *hole_pfn = end_pfn;
6762 static void __init memmap_init(void)
6764 unsigned long start_pfn, end_pfn;
6765 unsigned long hole_pfn = 0;
6766 int i, j, zone_id = 0, nid;
6768 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6769 struct pglist_data *node = NODE_DATA(nid);
6771 for (j = 0; j < MAX_NR_ZONES; j++) {
6772 struct zone *zone = node->node_zones + j;
6774 if (!populated_zone(zone))
6777 memmap_init_zone_range(zone, start_pfn, end_pfn,
6783 #ifdef CONFIG_SPARSEMEM
6785 * Initialize the memory map for hole in the range [memory_end,
6787 * Append the pages in this hole to the highest zone in the last
6789 * The call to init_unavailable_range() is outside the ifdef to
6790 * silence the compiler warining about zone_id set but not used;
6791 * for FLATMEM it is a nop anyway
6793 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6794 if (hole_pfn < end_pfn)
6796 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6799 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6800 phys_addr_t min_addr, int nid, bool exact_nid)
6805 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6806 MEMBLOCK_ALLOC_ACCESSIBLE,
6809 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6810 MEMBLOCK_ALLOC_ACCESSIBLE,
6813 if (ptr && size > 0)
6814 page_init_poison(ptr, size);
6819 static int zone_batchsize(struct zone *zone)
6825 * The number of pages to batch allocate is either ~0.1%
6826 * of the zone or 1MB, whichever is smaller. The batch
6827 * size is striking a balance between allocation latency
6828 * and zone lock contention.
6830 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6831 batch /= 4; /* We effectively *= 4 below */
6836 * Clamp the batch to a 2^n - 1 value. Having a power
6837 * of 2 value was found to be more likely to have
6838 * suboptimal cache aliasing properties in some cases.
6840 * For example if 2 tasks are alternately allocating
6841 * batches of pages, one task can end up with a lot
6842 * of pages of one half of the possible page colors
6843 * and the other with pages of the other colors.
6845 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6850 /* The deferral and batching of frees should be suppressed under NOMMU
6853 * The problem is that NOMMU needs to be able to allocate large chunks
6854 * of contiguous memory as there's no hardware page translation to
6855 * assemble apparent contiguous memory from discontiguous pages.
6857 * Queueing large contiguous runs of pages for batching, however,
6858 * causes the pages to actually be freed in smaller chunks. As there
6859 * can be a significant delay between the individual batches being
6860 * recycled, this leads to the once large chunks of space being
6861 * fragmented and becoming unavailable for high-order allocations.
6867 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6872 unsigned long total_pages;
6874 if (!percpu_pagelist_high_fraction) {
6876 * By default, the high value of the pcp is based on the zone
6877 * low watermark so that if they are full then background
6878 * reclaim will not be started prematurely.
6880 total_pages = low_wmark_pages(zone);
6883 * If percpu_pagelist_high_fraction is configured, the high
6884 * value is based on a fraction of the managed pages in the
6887 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6891 * Split the high value across all online CPUs local to the zone. Note
6892 * that early in boot that CPUs may not be online yet and that during
6893 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6894 * onlined. For memory nodes that have no CPUs, split pcp->high across
6895 * all online CPUs to mitigate the risk that reclaim is triggered
6896 * prematurely due to pages stored on pcp lists.
6898 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6900 nr_split_cpus = num_online_cpus();
6901 high = total_pages / nr_split_cpus;
6904 * Ensure high is at least batch*4. The multiple is based on the
6905 * historical relationship between high and batch.
6907 high = max(high, batch << 2);
6916 * pcp->high and pcp->batch values are related and generally batch is lower
6917 * than high. They are also related to pcp->count such that count is lower
6918 * than high, and as soon as it reaches high, the pcplist is flushed.
6920 * However, guaranteeing these relations at all times would require e.g. write
6921 * barriers here but also careful usage of read barriers at the read side, and
6922 * thus be prone to error and bad for performance. Thus the update only prevents
6923 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6924 * can cope with those fields changing asynchronously, and fully trust only the
6925 * pcp->count field on the local CPU with interrupts disabled.
6927 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6928 * outside of boot time (or some other assurance that no concurrent updaters
6931 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6932 unsigned long batch)
6934 WRITE_ONCE(pcp->batch, batch);
6935 WRITE_ONCE(pcp->high, high);
6938 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6942 memset(pcp, 0, sizeof(*pcp));
6943 memset(pzstats, 0, sizeof(*pzstats));
6945 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6946 INIT_LIST_HEAD(&pcp->lists[pindex]);
6949 * Set batch and high values safe for a boot pageset. A true percpu
6950 * pageset's initialization will update them subsequently. Here we don't
6951 * need to be as careful as pageset_update() as nobody can access the
6954 pcp->high = BOOT_PAGESET_HIGH;
6955 pcp->batch = BOOT_PAGESET_BATCH;
6956 pcp->free_factor = 0;
6959 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6960 unsigned long batch)
6962 struct per_cpu_pages *pcp;
6965 for_each_possible_cpu(cpu) {
6966 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6967 pageset_update(pcp, high, batch);
6972 * Calculate and set new high and batch values for all per-cpu pagesets of a
6973 * zone based on the zone's size.
6975 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6977 int new_high, new_batch;
6979 new_batch = max(1, zone_batchsize(zone));
6980 new_high = zone_highsize(zone, new_batch, cpu_online);
6982 if (zone->pageset_high == new_high &&
6983 zone->pageset_batch == new_batch)
6986 zone->pageset_high = new_high;
6987 zone->pageset_batch = new_batch;
6989 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6992 void __meminit setup_zone_pageset(struct zone *zone)
6996 /* Size may be 0 on !SMP && !NUMA */
6997 if (sizeof(struct per_cpu_zonestat) > 0)
6998 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7000 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7001 for_each_possible_cpu(cpu) {
7002 struct per_cpu_pages *pcp;
7003 struct per_cpu_zonestat *pzstats;
7005 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7006 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7007 per_cpu_pages_init(pcp, pzstats);
7010 zone_set_pageset_high_and_batch(zone, 0);
7014 * Allocate per cpu pagesets and initialize them.
7015 * Before this call only boot pagesets were available.
7017 void __init setup_per_cpu_pageset(void)
7019 struct pglist_data *pgdat;
7021 int __maybe_unused cpu;
7023 for_each_populated_zone(zone)
7024 setup_zone_pageset(zone);
7028 * Unpopulated zones continue using the boot pagesets.
7029 * The numa stats for these pagesets need to be reset.
7030 * Otherwise, they will end up skewing the stats of
7031 * the nodes these zones are associated with.
7033 for_each_possible_cpu(cpu) {
7034 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7035 memset(pzstats->vm_numa_event, 0,
7036 sizeof(pzstats->vm_numa_event));
7040 for_each_online_pgdat(pgdat)
7041 pgdat->per_cpu_nodestats =
7042 alloc_percpu(struct per_cpu_nodestat);
7045 static __meminit void zone_pcp_init(struct zone *zone)
7048 * per cpu subsystem is not up at this point. The following code
7049 * relies on the ability of the linker to provide the
7050 * offset of a (static) per cpu variable into the per cpu area.
7052 zone->per_cpu_pageset = &boot_pageset;
7053 zone->per_cpu_zonestats = &boot_zonestats;
7054 zone->pageset_high = BOOT_PAGESET_HIGH;
7055 zone->pageset_batch = BOOT_PAGESET_BATCH;
7057 if (populated_zone(zone))
7058 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7059 zone->present_pages, zone_batchsize(zone));
7062 void __meminit init_currently_empty_zone(struct zone *zone,
7063 unsigned long zone_start_pfn,
7066 struct pglist_data *pgdat = zone->zone_pgdat;
7067 int zone_idx = zone_idx(zone) + 1;
7069 if (zone_idx > pgdat->nr_zones)
7070 pgdat->nr_zones = zone_idx;
7072 zone->zone_start_pfn = zone_start_pfn;
7074 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7075 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7077 (unsigned long)zone_idx(zone),
7078 zone_start_pfn, (zone_start_pfn + size));
7080 zone_init_free_lists(zone);
7081 zone->initialized = 1;
7085 * get_pfn_range_for_nid - Return the start and end page frames for a node
7086 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7087 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7088 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7090 * It returns the start and end page frame of a node based on information
7091 * provided by memblock_set_node(). If called for a node
7092 * with no available memory, a warning is printed and the start and end
7095 void __init get_pfn_range_for_nid(unsigned int nid,
7096 unsigned long *start_pfn, unsigned long *end_pfn)
7098 unsigned long this_start_pfn, this_end_pfn;
7104 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7105 *start_pfn = min(*start_pfn, this_start_pfn);
7106 *end_pfn = max(*end_pfn, this_end_pfn);
7109 if (*start_pfn == -1UL)
7114 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7115 * assumption is made that zones within a node are ordered in monotonic
7116 * increasing memory addresses so that the "highest" populated zone is used
7118 static void __init find_usable_zone_for_movable(void)
7121 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7122 if (zone_index == ZONE_MOVABLE)
7125 if (arch_zone_highest_possible_pfn[zone_index] >
7126 arch_zone_lowest_possible_pfn[zone_index])
7130 VM_BUG_ON(zone_index == -1);
7131 movable_zone = zone_index;
7135 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7136 * because it is sized independent of architecture. Unlike the other zones,
7137 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7138 * in each node depending on the size of each node and how evenly kernelcore
7139 * is distributed. This helper function adjusts the zone ranges
7140 * provided by the architecture for a given node by using the end of the
7141 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7142 * zones within a node are in order of monotonic increases memory addresses
7144 static void __init adjust_zone_range_for_zone_movable(int nid,
7145 unsigned long zone_type,
7146 unsigned long node_start_pfn,
7147 unsigned long node_end_pfn,
7148 unsigned long *zone_start_pfn,
7149 unsigned long *zone_end_pfn)
7151 /* Only adjust if ZONE_MOVABLE is on this node */
7152 if (zone_movable_pfn[nid]) {
7153 /* Size ZONE_MOVABLE */
7154 if (zone_type == ZONE_MOVABLE) {
7155 *zone_start_pfn = zone_movable_pfn[nid];
7156 *zone_end_pfn = min(node_end_pfn,
7157 arch_zone_highest_possible_pfn[movable_zone]);
7159 /* Adjust for ZONE_MOVABLE starting within this range */
7160 } else if (!mirrored_kernelcore &&
7161 *zone_start_pfn < zone_movable_pfn[nid] &&
7162 *zone_end_pfn > zone_movable_pfn[nid]) {
7163 *zone_end_pfn = zone_movable_pfn[nid];
7165 /* Check if this whole range is within ZONE_MOVABLE */
7166 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7167 *zone_start_pfn = *zone_end_pfn;
7172 * Return the number of pages a zone spans in a node, including holes
7173 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7175 static unsigned long __init zone_spanned_pages_in_node(int nid,
7176 unsigned long zone_type,
7177 unsigned long node_start_pfn,
7178 unsigned long node_end_pfn,
7179 unsigned long *zone_start_pfn,
7180 unsigned long *zone_end_pfn)
7182 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7183 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7184 /* When hotadd a new node from cpu_up(), the node should be empty */
7185 if (!node_start_pfn && !node_end_pfn)
7188 /* Get the start and end of the zone */
7189 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7190 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7191 adjust_zone_range_for_zone_movable(nid, zone_type,
7192 node_start_pfn, node_end_pfn,
7193 zone_start_pfn, zone_end_pfn);
7195 /* Check that this node has pages within the zone's required range */
7196 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7199 /* Move the zone boundaries inside the node if necessary */
7200 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7201 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7203 /* Return the spanned pages */
7204 return *zone_end_pfn - *zone_start_pfn;
7208 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7209 * then all holes in the requested range will be accounted for.
7211 unsigned long __init __absent_pages_in_range(int nid,
7212 unsigned long range_start_pfn,
7213 unsigned long range_end_pfn)
7215 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7216 unsigned long start_pfn, end_pfn;
7219 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7220 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7221 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7222 nr_absent -= end_pfn - start_pfn;
7228 * absent_pages_in_range - Return number of page frames in holes within a range
7229 * @start_pfn: The start PFN to start searching for holes
7230 * @end_pfn: The end PFN to stop searching for holes
7232 * Return: the number of pages frames in memory holes within a range.
7234 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7235 unsigned long end_pfn)
7237 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7240 /* Return the number of page frames in holes in a zone on a node */
7241 static unsigned long __init zone_absent_pages_in_node(int nid,
7242 unsigned long zone_type,
7243 unsigned long node_start_pfn,
7244 unsigned long node_end_pfn)
7246 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7247 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7248 unsigned long zone_start_pfn, zone_end_pfn;
7249 unsigned long nr_absent;
7251 /* When hotadd a new node from cpu_up(), the node should be empty */
7252 if (!node_start_pfn && !node_end_pfn)
7255 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7256 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7258 adjust_zone_range_for_zone_movable(nid, zone_type,
7259 node_start_pfn, node_end_pfn,
7260 &zone_start_pfn, &zone_end_pfn);
7261 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7264 * ZONE_MOVABLE handling.
7265 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7268 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7269 unsigned long start_pfn, end_pfn;
7270 struct memblock_region *r;
7272 for_each_mem_region(r) {
7273 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7274 zone_start_pfn, zone_end_pfn);
7275 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7276 zone_start_pfn, zone_end_pfn);
7278 if (zone_type == ZONE_MOVABLE &&
7279 memblock_is_mirror(r))
7280 nr_absent += end_pfn - start_pfn;
7282 if (zone_type == ZONE_NORMAL &&
7283 !memblock_is_mirror(r))
7284 nr_absent += end_pfn - start_pfn;
7291 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7292 unsigned long node_start_pfn,
7293 unsigned long node_end_pfn)
7295 unsigned long realtotalpages = 0, totalpages = 0;
7298 for (i = 0; i < MAX_NR_ZONES; i++) {
7299 struct zone *zone = pgdat->node_zones + i;
7300 unsigned long zone_start_pfn, zone_end_pfn;
7301 unsigned long spanned, absent;
7302 unsigned long size, real_size;
7304 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7309 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7314 real_size = size - absent;
7317 zone->zone_start_pfn = zone_start_pfn;
7319 zone->zone_start_pfn = 0;
7320 zone->spanned_pages = size;
7321 zone->present_pages = real_size;
7322 #if defined(CONFIG_MEMORY_HOTPLUG)
7323 zone->present_early_pages = real_size;
7327 realtotalpages += real_size;
7330 pgdat->node_spanned_pages = totalpages;
7331 pgdat->node_present_pages = realtotalpages;
7332 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7335 #ifndef CONFIG_SPARSEMEM
7337 * Calculate the size of the zone->blockflags rounded to an unsigned long
7338 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7339 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7340 * round what is now in bits to nearest long in bits, then return it in
7343 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7345 unsigned long usemapsize;
7347 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7348 usemapsize = roundup(zonesize, pageblock_nr_pages);
7349 usemapsize = usemapsize >> pageblock_order;
7350 usemapsize *= NR_PAGEBLOCK_BITS;
7351 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7353 return usemapsize / 8;
7356 static void __ref setup_usemap(struct zone *zone)
7358 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7359 zone->spanned_pages);
7360 zone->pageblock_flags = NULL;
7362 zone->pageblock_flags =
7363 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7365 if (!zone->pageblock_flags)
7366 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7367 usemapsize, zone->name, zone_to_nid(zone));
7371 static inline void setup_usemap(struct zone *zone) {}
7372 #endif /* CONFIG_SPARSEMEM */
7374 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7376 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7377 void __init set_pageblock_order(void)
7379 unsigned int order = MAX_ORDER - 1;
7381 /* Check that pageblock_nr_pages has not already been setup */
7382 if (pageblock_order)
7385 /* Don't let pageblocks exceed the maximum allocation granularity. */
7386 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7387 order = HUGETLB_PAGE_ORDER;
7390 * Assume the largest contiguous order of interest is a huge page.
7391 * This value may be variable depending on boot parameters on IA64 and
7394 pageblock_order = order;
7396 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7399 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7400 * is unused as pageblock_order is set at compile-time. See
7401 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7404 void __init set_pageblock_order(void)
7408 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7410 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7411 unsigned long present_pages)
7413 unsigned long pages = spanned_pages;
7416 * Provide a more accurate estimation if there are holes within
7417 * the zone and SPARSEMEM is in use. If there are holes within the
7418 * zone, each populated memory region may cost us one or two extra
7419 * memmap pages due to alignment because memmap pages for each
7420 * populated regions may not be naturally aligned on page boundary.
7421 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7423 if (spanned_pages > present_pages + (present_pages >> 4) &&
7424 IS_ENABLED(CONFIG_SPARSEMEM))
7425 pages = present_pages;
7427 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7430 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7431 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7433 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7435 spin_lock_init(&ds_queue->split_queue_lock);
7436 INIT_LIST_HEAD(&ds_queue->split_queue);
7437 ds_queue->split_queue_len = 0;
7440 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7443 #ifdef CONFIG_COMPACTION
7444 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7446 init_waitqueue_head(&pgdat->kcompactd_wait);
7449 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7452 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7456 pgdat_resize_init(pgdat);
7458 pgdat_init_split_queue(pgdat);
7459 pgdat_init_kcompactd(pgdat);
7461 init_waitqueue_head(&pgdat->kswapd_wait);
7462 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7464 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7465 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7467 pgdat_page_ext_init(pgdat);
7468 lruvec_init(&pgdat->__lruvec);
7471 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7472 unsigned long remaining_pages)
7474 atomic_long_set(&zone->managed_pages, remaining_pages);
7475 zone_set_nid(zone, nid);
7476 zone->name = zone_names[idx];
7477 zone->zone_pgdat = NODE_DATA(nid);
7478 spin_lock_init(&zone->lock);
7479 zone_seqlock_init(zone);
7480 zone_pcp_init(zone);
7484 * Set up the zone data structures
7485 * - init pgdat internals
7486 * - init all zones belonging to this node
7488 * NOTE: this function is only called during memory hotplug
7490 #ifdef CONFIG_MEMORY_HOTPLUG
7491 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7493 int nid = pgdat->node_id;
7497 pgdat_init_internals(pgdat);
7499 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7500 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7503 * Reset the nr_zones, order and highest_zoneidx before reuse.
7504 * Note that kswapd will init kswapd_highest_zoneidx properly
7505 * when it starts in the near future.
7507 pgdat->nr_zones = 0;
7508 pgdat->kswapd_order = 0;
7509 pgdat->kswapd_highest_zoneidx = 0;
7510 pgdat->node_start_pfn = 0;
7511 for_each_online_cpu(cpu) {
7512 struct per_cpu_nodestat *p;
7514 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7515 memset(p, 0, sizeof(*p));
7518 for (z = 0; z < MAX_NR_ZONES; z++)
7519 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7524 * Set up the zone data structures:
7525 * - mark all pages reserved
7526 * - mark all memory queues empty
7527 * - clear the memory bitmaps
7529 * NOTE: pgdat should get zeroed by caller.
7530 * NOTE: this function is only called during early init.
7532 static void __init free_area_init_core(struct pglist_data *pgdat)
7535 int nid = pgdat->node_id;
7537 pgdat_init_internals(pgdat);
7538 pgdat->per_cpu_nodestats = &boot_nodestats;
7540 for (j = 0; j < MAX_NR_ZONES; j++) {
7541 struct zone *zone = pgdat->node_zones + j;
7542 unsigned long size, freesize, memmap_pages;
7544 size = zone->spanned_pages;
7545 freesize = zone->present_pages;
7548 * Adjust freesize so that it accounts for how much memory
7549 * is used by this zone for memmap. This affects the watermark
7550 * and per-cpu initialisations
7552 memmap_pages = calc_memmap_size(size, freesize);
7553 if (!is_highmem_idx(j)) {
7554 if (freesize >= memmap_pages) {
7555 freesize -= memmap_pages;
7557 pr_debug(" %s zone: %lu pages used for memmap\n",
7558 zone_names[j], memmap_pages);
7560 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7561 zone_names[j], memmap_pages, freesize);
7564 /* Account for reserved pages */
7565 if (j == 0 && freesize > dma_reserve) {
7566 freesize -= dma_reserve;
7567 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7570 if (!is_highmem_idx(j))
7571 nr_kernel_pages += freesize;
7572 /* Charge for highmem memmap if there are enough kernel pages */
7573 else if (nr_kernel_pages > memmap_pages * 2)
7574 nr_kernel_pages -= memmap_pages;
7575 nr_all_pages += freesize;
7578 * Set an approximate value for lowmem here, it will be adjusted
7579 * when the bootmem allocator frees pages into the buddy system.
7580 * And all highmem pages will be managed by the buddy system.
7582 zone_init_internals(zone, j, nid, freesize);
7587 set_pageblock_order();
7589 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7593 #ifdef CONFIG_FLATMEM
7594 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7596 unsigned long __maybe_unused start = 0;
7597 unsigned long __maybe_unused offset = 0;
7599 /* Skip empty nodes */
7600 if (!pgdat->node_spanned_pages)
7603 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7604 offset = pgdat->node_start_pfn - start;
7605 /* ia64 gets its own node_mem_map, before this, without bootmem */
7606 if (!pgdat->node_mem_map) {
7607 unsigned long size, end;
7611 * The zone's endpoints aren't required to be MAX_ORDER
7612 * aligned but the node_mem_map endpoints must be in order
7613 * for the buddy allocator to function correctly.
7615 end = pgdat_end_pfn(pgdat);
7616 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7617 size = (end - start) * sizeof(struct page);
7618 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7619 pgdat->node_id, false);
7621 panic("Failed to allocate %ld bytes for node %d memory map\n",
7622 size, pgdat->node_id);
7623 pgdat->node_mem_map = map + offset;
7625 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7626 __func__, pgdat->node_id, (unsigned long)pgdat,
7627 (unsigned long)pgdat->node_mem_map);
7630 * With no DISCONTIG, the global mem_map is just set as node 0's
7632 if (pgdat == NODE_DATA(0)) {
7633 mem_map = NODE_DATA(0)->node_mem_map;
7634 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7640 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7641 #endif /* CONFIG_FLATMEM */
7643 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7644 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7646 pgdat->first_deferred_pfn = ULONG_MAX;
7649 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7652 static void __init free_area_init_node(int nid)
7654 pg_data_t *pgdat = NODE_DATA(nid);
7655 unsigned long start_pfn = 0;
7656 unsigned long end_pfn = 0;
7658 /* pg_data_t should be reset to zero when it's allocated */
7659 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7661 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7663 pgdat->node_id = nid;
7664 pgdat->node_start_pfn = start_pfn;
7665 pgdat->per_cpu_nodestats = NULL;
7667 if (start_pfn != end_pfn) {
7668 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7669 (u64)start_pfn << PAGE_SHIFT,
7670 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7672 pr_info("Initmem setup node %d as memoryless\n", nid);
7675 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7677 alloc_node_mem_map(pgdat);
7678 pgdat_set_deferred_range(pgdat);
7680 free_area_init_core(pgdat);
7683 static void __init free_area_init_memoryless_node(int nid)
7685 free_area_init_node(nid);
7688 #if MAX_NUMNODES > 1
7690 * Figure out the number of possible node ids.
7692 void __init setup_nr_node_ids(void)
7694 unsigned int highest;
7696 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7697 nr_node_ids = highest + 1;
7702 * node_map_pfn_alignment - determine the maximum internode alignment
7704 * This function should be called after node map is populated and sorted.
7705 * It calculates the maximum power of two alignment which can distinguish
7708 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7709 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7710 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7711 * shifted, 1GiB is enough and this function will indicate so.
7713 * This is used to test whether pfn -> nid mapping of the chosen memory
7714 * model has fine enough granularity to avoid incorrect mapping for the
7715 * populated node map.
7717 * Return: the determined alignment in pfn's. 0 if there is no alignment
7718 * requirement (single node).
7720 unsigned long __init node_map_pfn_alignment(void)
7722 unsigned long accl_mask = 0, last_end = 0;
7723 unsigned long start, end, mask;
7724 int last_nid = NUMA_NO_NODE;
7727 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7728 if (!start || last_nid < 0 || last_nid == nid) {
7735 * Start with a mask granular enough to pin-point to the
7736 * start pfn and tick off bits one-by-one until it becomes
7737 * too coarse to separate the current node from the last.
7739 mask = ~((1 << __ffs(start)) - 1);
7740 while (mask && last_end <= (start & (mask << 1)))
7743 /* accumulate all internode masks */
7747 /* convert mask to number of pages */
7748 return ~accl_mask + 1;
7752 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7754 * Return: the minimum PFN based on information provided via
7755 * memblock_set_node().
7757 unsigned long __init find_min_pfn_with_active_regions(void)
7759 return PHYS_PFN(memblock_start_of_DRAM());
7763 * early_calculate_totalpages()
7764 * Sum pages in active regions for movable zone.
7765 * Populate N_MEMORY for calculating usable_nodes.
7767 static unsigned long __init early_calculate_totalpages(void)
7769 unsigned long totalpages = 0;
7770 unsigned long start_pfn, end_pfn;
7773 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7774 unsigned long pages = end_pfn - start_pfn;
7776 totalpages += pages;
7778 node_set_state(nid, N_MEMORY);
7784 * Find the PFN the Movable zone begins in each node. Kernel memory
7785 * is spread evenly between nodes as long as the nodes have enough
7786 * memory. When they don't, some nodes will have more kernelcore than
7789 static void __init find_zone_movable_pfns_for_nodes(void)
7792 unsigned long usable_startpfn;
7793 unsigned long kernelcore_node, kernelcore_remaining;
7794 /* save the state before borrow the nodemask */
7795 nodemask_t saved_node_state = node_states[N_MEMORY];
7796 unsigned long totalpages = early_calculate_totalpages();
7797 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7798 struct memblock_region *r;
7800 /* Need to find movable_zone earlier when movable_node is specified. */
7801 find_usable_zone_for_movable();
7804 * If movable_node is specified, ignore kernelcore and movablecore
7807 if (movable_node_is_enabled()) {
7808 for_each_mem_region(r) {
7809 if (!memblock_is_hotpluggable(r))
7812 nid = memblock_get_region_node(r);
7814 usable_startpfn = PFN_DOWN(r->base);
7815 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7816 min(usable_startpfn, zone_movable_pfn[nid]) :
7824 * If kernelcore=mirror is specified, ignore movablecore option
7826 if (mirrored_kernelcore) {
7827 bool mem_below_4gb_not_mirrored = false;
7829 for_each_mem_region(r) {
7830 if (memblock_is_mirror(r))
7833 nid = memblock_get_region_node(r);
7835 usable_startpfn = memblock_region_memory_base_pfn(r);
7837 if (usable_startpfn < 0x100000) {
7838 mem_below_4gb_not_mirrored = true;
7842 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7843 min(usable_startpfn, zone_movable_pfn[nid]) :
7847 if (mem_below_4gb_not_mirrored)
7848 pr_warn("This configuration results in unmirrored kernel memory.\n");
7854 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7855 * amount of necessary memory.
7857 if (required_kernelcore_percent)
7858 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7860 if (required_movablecore_percent)
7861 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7865 * If movablecore= was specified, calculate what size of
7866 * kernelcore that corresponds so that memory usable for
7867 * any allocation type is evenly spread. If both kernelcore
7868 * and movablecore are specified, then the value of kernelcore
7869 * will be used for required_kernelcore if it's greater than
7870 * what movablecore would have allowed.
7872 if (required_movablecore) {
7873 unsigned long corepages;
7876 * Round-up so that ZONE_MOVABLE is at least as large as what
7877 * was requested by the user
7879 required_movablecore =
7880 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7881 required_movablecore = min(totalpages, required_movablecore);
7882 corepages = totalpages - required_movablecore;
7884 required_kernelcore = max(required_kernelcore, corepages);
7888 * If kernelcore was not specified or kernelcore size is larger
7889 * than totalpages, there is no ZONE_MOVABLE.
7891 if (!required_kernelcore || required_kernelcore >= totalpages)
7894 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7895 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7898 /* Spread kernelcore memory as evenly as possible throughout nodes */
7899 kernelcore_node = required_kernelcore / usable_nodes;
7900 for_each_node_state(nid, N_MEMORY) {
7901 unsigned long start_pfn, end_pfn;
7904 * Recalculate kernelcore_node if the division per node
7905 * now exceeds what is necessary to satisfy the requested
7906 * amount of memory for the kernel
7908 if (required_kernelcore < kernelcore_node)
7909 kernelcore_node = required_kernelcore / usable_nodes;
7912 * As the map is walked, we track how much memory is usable
7913 * by the kernel using kernelcore_remaining. When it is
7914 * 0, the rest of the node is usable by ZONE_MOVABLE
7916 kernelcore_remaining = kernelcore_node;
7918 /* Go through each range of PFNs within this node */
7919 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7920 unsigned long size_pages;
7922 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7923 if (start_pfn >= end_pfn)
7926 /* Account for what is only usable for kernelcore */
7927 if (start_pfn < usable_startpfn) {
7928 unsigned long kernel_pages;
7929 kernel_pages = min(end_pfn, usable_startpfn)
7932 kernelcore_remaining -= min(kernel_pages,
7933 kernelcore_remaining);
7934 required_kernelcore -= min(kernel_pages,
7935 required_kernelcore);
7937 /* Continue if range is now fully accounted */
7938 if (end_pfn <= usable_startpfn) {
7941 * Push zone_movable_pfn to the end so
7942 * that if we have to rebalance
7943 * kernelcore across nodes, we will
7944 * not double account here
7946 zone_movable_pfn[nid] = end_pfn;
7949 start_pfn = usable_startpfn;
7953 * The usable PFN range for ZONE_MOVABLE is from
7954 * start_pfn->end_pfn. Calculate size_pages as the
7955 * number of pages used as kernelcore
7957 size_pages = end_pfn - start_pfn;
7958 if (size_pages > kernelcore_remaining)
7959 size_pages = kernelcore_remaining;
7960 zone_movable_pfn[nid] = start_pfn + size_pages;
7963 * Some kernelcore has been met, update counts and
7964 * break if the kernelcore for this node has been
7967 required_kernelcore -= min(required_kernelcore,
7969 kernelcore_remaining -= size_pages;
7970 if (!kernelcore_remaining)
7976 * If there is still required_kernelcore, we do another pass with one
7977 * less node in the count. This will push zone_movable_pfn[nid] further
7978 * along on the nodes that still have memory until kernelcore is
7982 if (usable_nodes && required_kernelcore > usable_nodes)
7986 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7987 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7988 unsigned long start_pfn, end_pfn;
7990 zone_movable_pfn[nid] =
7991 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7993 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7994 if (zone_movable_pfn[nid] >= end_pfn)
7995 zone_movable_pfn[nid] = 0;
7999 /* restore the node_state */
8000 node_states[N_MEMORY] = saved_node_state;
8003 /* Any regular or high memory on that node ? */
8004 static void check_for_memory(pg_data_t *pgdat, int nid)
8006 enum zone_type zone_type;
8008 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8009 struct zone *zone = &pgdat->node_zones[zone_type];
8010 if (populated_zone(zone)) {
8011 if (IS_ENABLED(CONFIG_HIGHMEM))
8012 node_set_state(nid, N_HIGH_MEMORY);
8013 if (zone_type <= ZONE_NORMAL)
8014 node_set_state(nid, N_NORMAL_MEMORY);
8021 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8022 * such cases we allow max_zone_pfn sorted in the descending order
8024 bool __weak arch_has_descending_max_zone_pfns(void)
8030 * free_area_init - Initialise all pg_data_t and zone data
8031 * @max_zone_pfn: an array of max PFNs for each zone
8033 * This will call free_area_init_node() for each active node in the system.
8034 * Using the page ranges provided by memblock_set_node(), the size of each
8035 * zone in each node and their holes is calculated. If the maximum PFN
8036 * between two adjacent zones match, it is assumed that the zone is empty.
8037 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8038 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8039 * starts where the previous one ended. For example, ZONE_DMA32 starts
8040 * at arch_max_dma_pfn.
8042 void __init free_area_init(unsigned long *max_zone_pfn)
8044 unsigned long start_pfn, end_pfn;
8048 /* Record where the zone boundaries are */
8049 memset(arch_zone_lowest_possible_pfn, 0,
8050 sizeof(arch_zone_lowest_possible_pfn));
8051 memset(arch_zone_highest_possible_pfn, 0,
8052 sizeof(arch_zone_highest_possible_pfn));
8054 start_pfn = find_min_pfn_with_active_regions();
8055 descending = arch_has_descending_max_zone_pfns();
8057 for (i = 0; i < MAX_NR_ZONES; i++) {
8059 zone = MAX_NR_ZONES - i - 1;
8063 if (zone == ZONE_MOVABLE)
8066 end_pfn = max(max_zone_pfn[zone], start_pfn);
8067 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8068 arch_zone_highest_possible_pfn[zone] = end_pfn;
8070 start_pfn = end_pfn;
8073 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8074 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8075 find_zone_movable_pfns_for_nodes();
8077 /* Print out the zone ranges */
8078 pr_info("Zone ranges:\n");
8079 for (i = 0; i < MAX_NR_ZONES; i++) {
8080 if (i == ZONE_MOVABLE)
8082 pr_info(" %-8s ", zone_names[i]);
8083 if (arch_zone_lowest_possible_pfn[i] ==
8084 arch_zone_highest_possible_pfn[i])
8087 pr_cont("[mem %#018Lx-%#018Lx]\n",
8088 (u64)arch_zone_lowest_possible_pfn[i]
8090 ((u64)arch_zone_highest_possible_pfn[i]
8091 << PAGE_SHIFT) - 1);
8094 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8095 pr_info("Movable zone start for each node\n");
8096 for (i = 0; i < MAX_NUMNODES; i++) {
8097 if (zone_movable_pfn[i])
8098 pr_info(" Node %d: %#018Lx\n", i,
8099 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8103 * Print out the early node map, and initialize the
8104 * subsection-map relative to active online memory ranges to
8105 * enable future "sub-section" extensions of the memory map.
8107 pr_info("Early memory node ranges\n");
8108 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8109 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8110 (u64)start_pfn << PAGE_SHIFT,
8111 ((u64)end_pfn << PAGE_SHIFT) - 1);
8112 subsection_map_init(start_pfn, end_pfn - start_pfn);
8115 /* Initialise every node */
8116 mminit_verify_pageflags_layout();
8117 setup_nr_node_ids();
8118 for_each_node(nid) {
8121 if (!node_online(nid)) {
8122 pr_info("Initializing node %d as memoryless\n", nid);
8124 /* Allocator not initialized yet */
8125 pgdat = arch_alloc_nodedata(nid);
8127 pr_err("Cannot allocate %zuB for node %d.\n",
8128 sizeof(*pgdat), nid);
8131 arch_refresh_nodedata(nid, pgdat);
8132 free_area_init_memoryless_node(nid);
8135 * We do not want to confuse userspace by sysfs
8136 * files/directories for node without any memory
8137 * attached to it, so this node is not marked as
8138 * N_MEMORY and not marked online so that no sysfs
8139 * hierarchy will be created via register_one_node for
8140 * it. The pgdat will get fully initialized by
8141 * hotadd_init_pgdat() when memory is hotplugged into
8147 pgdat = NODE_DATA(nid);
8148 free_area_init_node(nid);
8150 /* Any memory on that node */
8151 if (pgdat->node_present_pages)
8152 node_set_state(nid, N_MEMORY);
8153 check_for_memory(pgdat, nid);
8159 static int __init cmdline_parse_core(char *p, unsigned long *core,
8160 unsigned long *percent)
8162 unsigned long long coremem;
8168 /* Value may be a percentage of total memory, otherwise bytes */
8169 coremem = simple_strtoull(p, &endptr, 0);
8170 if (*endptr == '%') {
8171 /* Paranoid check for percent values greater than 100 */
8172 WARN_ON(coremem > 100);
8176 coremem = memparse(p, &p);
8177 /* Paranoid check that UL is enough for the coremem value */
8178 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8180 *core = coremem >> PAGE_SHIFT;
8187 * kernelcore=size sets the amount of memory for use for allocations that
8188 * cannot be reclaimed or migrated.
8190 static int __init cmdline_parse_kernelcore(char *p)
8192 /* parse kernelcore=mirror */
8193 if (parse_option_str(p, "mirror")) {
8194 mirrored_kernelcore = true;
8198 return cmdline_parse_core(p, &required_kernelcore,
8199 &required_kernelcore_percent);
8203 * movablecore=size sets the amount of memory for use for allocations that
8204 * can be reclaimed or migrated.
8206 static int __init cmdline_parse_movablecore(char *p)
8208 return cmdline_parse_core(p, &required_movablecore,
8209 &required_movablecore_percent);
8212 early_param("kernelcore", cmdline_parse_kernelcore);
8213 early_param("movablecore", cmdline_parse_movablecore);
8215 void adjust_managed_page_count(struct page *page, long count)
8217 atomic_long_add(count, &page_zone(page)->managed_pages);
8218 totalram_pages_add(count);
8219 #ifdef CONFIG_HIGHMEM
8220 if (PageHighMem(page))
8221 totalhigh_pages_add(count);
8224 EXPORT_SYMBOL(adjust_managed_page_count);
8226 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8229 unsigned long pages = 0;
8231 start = (void *)PAGE_ALIGN((unsigned long)start);
8232 end = (void *)((unsigned long)end & PAGE_MASK);
8233 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8234 struct page *page = virt_to_page(pos);
8235 void *direct_map_addr;
8238 * 'direct_map_addr' might be different from 'pos'
8239 * because some architectures' virt_to_page()
8240 * work with aliases. Getting the direct map
8241 * address ensures that we get a _writeable_
8242 * alias for the memset().
8244 direct_map_addr = page_address(page);
8246 * Perform a kasan-unchecked memset() since this memory
8247 * has not been initialized.
8249 direct_map_addr = kasan_reset_tag(direct_map_addr);
8250 if ((unsigned int)poison <= 0xFF)
8251 memset(direct_map_addr, poison, PAGE_SIZE);
8253 free_reserved_page(page);
8257 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8262 void __init mem_init_print_info(void)
8264 unsigned long physpages, codesize, datasize, rosize, bss_size;
8265 unsigned long init_code_size, init_data_size;
8267 physpages = get_num_physpages();
8268 codesize = _etext - _stext;
8269 datasize = _edata - _sdata;
8270 rosize = __end_rodata - __start_rodata;
8271 bss_size = __bss_stop - __bss_start;
8272 init_data_size = __init_end - __init_begin;
8273 init_code_size = _einittext - _sinittext;
8276 * Detect special cases and adjust section sizes accordingly:
8277 * 1) .init.* may be embedded into .data sections
8278 * 2) .init.text.* may be out of [__init_begin, __init_end],
8279 * please refer to arch/tile/kernel/vmlinux.lds.S.
8280 * 3) .rodata.* may be embedded into .text or .data sections.
8282 #define adj_init_size(start, end, size, pos, adj) \
8284 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8288 adj_init_size(__init_begin, __init_end, init_data_size,
8289 _sinittext, init_code_size);
8290 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8291 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8292 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8293 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8295 #undef adj_init_size
8297 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8298 #ifdef CONFIG_HIGHMEM
8302 K(nr_free_pages()), K(physpages),
8303 codesize >> 10, datasize >> 10, rosize >> 10,
8304 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8305 K(physpages - totalram_pages() - totalcma_pages),
8307 #ifdef CONFIG_HIGHMEM
8308 , K(totalhigh_pages())
8314 * set_dma_reserve - set the specified number of pages reserved in the first zone
8315 * @new_dma_reserve: The number of pages to mark reserved
8317 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8318 * In the DMA zone, a significant percentage may be consumed by kernel image
8319 * and other unfreeable allocations which can skew the watermarks badly. This
8320 * function may optionally be used to account for unfreeable pages in the
8321 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8322 * smaller per-cpu batchsize.
8324 void __init set_dma_reserve(unsigned long new_dma_reserve)
8326 dma_reserve = new_dma_reserve;
8329 static int page_alloc_cpu_dead(unsigned int cpu)
8333 lru_add_drain_cpu(cpu);
8337 * Spill the event counters of the dead processor
8338 * into the current processors event counters.
8339 * This artificially elevates the count of the current
8342 vm_events_fold_cpu(cpu);
8345 * Zero the differential counters of the dead processor
8346 * so that the vm statistics are consistent.
8348 * This is only okay since the processor is dead and cannot
8349 * race with what we are doing.
8351 cpu_vm_stats_fold(cpu);
8353 for_each_populated_zone(zone)
8354 zone_pcp_update(zone, 0);
8359 static int page_alloc_cpu_online(unsigned int cpu)
8363 for_each_populated_zone(zone)
8364 zone_pcp_update(zone, 1);
8369 int hashdist = HASHDIST_DEFAULT;
8371 static int __init set_hashdist(char *str)
8375 hashdist = simple_strtoul(str, &str, 0);
8378 __setup("hashdist=", set_hashdist);
8381 void __init page_alloc_init(void)
8386 if (num_node_state(N_MEMORY) == 1)
8390 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8391 "mm/page_alloc:pcp",
8392 page_alloc_cpu_online,
8393 page_alloc_cpu_dead);
8398 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8399 * or min_free_kbytes changes.
8401 static void calculate_totalreserve_pages(void)
8403 struct pglist_data *pgdat;
8404 unsigned long reserve_pages = 0;
8405 enum zone_type i, j;
8407 for_each_online_pgdat(pgdat) {
8409 pgdat->totalreserve_pages = 0;
8411 for (i = 0; i < MAX_NR_ZONES; i++) {
8412 struct zone *zone = pgdat->node_zones + i;
8414 unsigned long managed_pages = zone_managed_pages(zone);
8416 /* Find valid and maximum lowmem_reserve in the zone */
8417 for (j = i; j < MAX_NR_ZONES; j++) {
8418 if (zone->lowmem_reserve[j] > max)
8419 max = zone->lowmem_reserve[j];
8422 /* we treat the high watermark as reserved pages. */
8423 max += high_wmark_pages(zone);
8425 if (max > managed_pages)
8426 max = managed_pages;
8428 pgdat->totalreserve_pages += max;
8430 reserve_pages += max;
8433 totalreserve_pages = reserve_pages;
8437 * setup_per_zone_lowmem_reserve - called whenever
8438 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8439 * has a correct pages reserved value, so an adequate number of
8440 * pages are left in the zone after a successful __alloc_pages().
8442 static void setup_per_zone_lowmem_reserve(void)
8444 struct pglist_data *pgdat;
8445 enum zone_type i, j;
8447 for_each_online_pgdat(pgdat) {
8448 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8449 struct zone *zone = &pgdat->node_zones[i];
8450 int ratio = sysctl_lowmem_reserve_ratio[i];
8451 bool clear = !ratio || !zone_managed_pages(zone);
8452 unsigned long managed_pages = 0;
8454 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8455 struct zone *upper_zone = &pgdat->node_zones[j];
8457 managed_pages += zone_managed_pages(upper_zone);
8460 zone->lowmem_reserve[j] = 0;
8462 zone->lowmem_reserve[j] = managed_pages / ratio;
8467 /* update totalreserve_pages */
8468 calculate_totalreserve_pages();
8471 static void __setup_per_zone_wmarks(void)
8473 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8474 unsigned long lowmem_pages = 0;
8476 unsigned long flags;
8478 /* Calculate total number of !ZONE_HIGHMEM pages */
8479 for_each_zone(zone) {
8480 if (!is_highmem(zone))
8481 lowmem_pages += zone_managed_pages(zone);
8484 for_each_zone(zone) {
8487 spin_lock_irqsave(&zone->lock, flags);
8488 tmp = (u64)pages_min * zone_managed_pages(zone);
8489 do_div(tmp, lowmem_pages);
8490 if (is_highmem(zone)) {
8492 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8493 * need highmem pages, so cap pages_min to a small
8496 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8497 * deltas control async page reclaim, and so should
8498 * not be capped for highmem.
8500 unsigned long min_pages;
8502 min_pages = zone_managed_pages(zone) / 1024;
8503 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8504 zone->_watermark[WMARK_MIN] = min_pages;
8507 * If it's a lowmem zone, reserve a number of pages
8508 * proportionate to the zone's size.
8510 zone->_watermark[WMARK_MIN] = tmp;
8514 * Set the kswapd watermarks distance according to the
8515 * scale factor in proportion to available memory, but
8516 * ensure a minimum size on small systems.
8518 tmp = max_t(u64, tmp >> 2,
8519 mult_frac(zone_managed_pages(zone),
8520 watermark_scale_factor, 10000));
8522 zone->watermark_boost = 0;
8523 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8524 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8525 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8527 spin_unlock_irqrestore(&zone->lock, flags);
8530 /* update totalreserve_pages */
8531 calculate_totalreserve_pages();
8535 * setup_per_zone_wmarks - called when min_free_kbytes changes
8536 * or when memory is hot-{added|removed}
8538 * Ensures that the watermark[min,low,high] values for each zone are set
8539 * correctly with respect to min_free_kbytes.
8541 void setup_per_zone_wmarks(void)
8544 static DEFINE_SPINLOCK(lock);
8547 __setup_per_zone_wmarks();
8551 * The watermark size have changed so update the pcpu batch
8552 * and high limits or the limits may be inappropriate.
8555 zone_pcp_update(zone, 0);
8559 * Initialise min_free_kbytes.
8561 * For small machines we want it small (128k min). For large machines
8562 * we want it large (256MB max). But it is not linear, because network
8563 * bandwidth does not increase linearly with machine size. We use
8565 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8566 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8582 void calculate_min_free_kbytes(void)
8584 unsigned long lowmem_kbytes;
8585 int new_min_free_kbytes;
8587 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8588 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8590 if (new_min_free_kbytes > user_min_free_kbytes)
8591 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8593 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8594 new_min_free_kbytes, user_min_free_kbytes);
8598 int __meminit init_per_zone_wmark_min(void)
8600 calculate_min_free_kbytes();
8601 setup_per_zone_wmarks();
8602 refresh_zone_stat_thresholds();
8603 setup_per_zone_lowmem_reserve();
8606 setup_min_unmapped_ratio();
8607 setup_min_slab_ratio();
8610 khugepaged_min_free_kbytes_update();
8614 postcore_initcall(init_per_zone_wmark_min)
8617 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8618 * that we can call two helper functions whenever min_free_kbytes
8621 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8622 void *buffer, size_t *length, loff_t *ppos)
8626 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8631 user_min_free_kbytes = min_free_kbytes;
8632 setup_per_zone_wmarks();
8637 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8638 void *buffer, size_t *length, loff_t *ppos)
8642 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8647 setup_per_zone_wmarks();
8653 static void setup_min_unmapped_ratio(void)
8658 for_each_online_pgdat(pgdat)
8659 pgdat->min_unmapped_pages = 0;
8662 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8663 sysctl_min_unmapped_ratio) / 100;
8667 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8668 void *buffer, size_t *length, loff_t *ppos)
8672 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8676 setup_min_unmapped_ratio();
8681 static void setup_min_slab_ratio(void)
8686 for_each_online_pgdat(pgdat)
8687 pgdat->min_slab_pages = 0;
8690 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8691 sysctl_min_slab_ratio) / 100;
8694 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8695 void *buffer, size_t *length, loff_t *ppos)
8699 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8703 setup_min_slab_ratio();
8710 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8711 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8712 * whenever sysctl_lowmem_reserve_ratio changes.
8714 * The reserve ratio obviously has absolutely no relation with the
8715 * minimum watermarks. The lowmem reserve ratio can only make sense
8716 * if in function of the boot time zone sizes.
8718 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8719 void *buffer, size_t *length, loff_t *ppos)
8723 proc_dointvec_minmax(table, write, buffer, length, ppos);
8725 for (i = 0; i < MAX_NR_ZONES; i++) {
8726 if (sysctl_lowmem_reserve_ratio[i] < 1)
8727 sysctl_lowmem_reserve_ratio[i] = 0;
8730 setup_per_zone_lowmem_reserve();
8735 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8736 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8737 * pagelist can have before it gets flushed back to buddy allocator.
8739 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8740 int write, void *buffer, size_t *length, loff_t *ppos)
8743 int old_percpu_pagelist_high_fraction;
8746 mutex_lock(&pcp_batch_high_lock);
8747 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8749 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8750 if (!write || ret < 0)
8753 /* Sanity checking to avoid pcp imbalance */
8754 if (percpu_pagelist_high_fraction &&
8755 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8756 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8762 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8765 for_each_populated_zone(zone)
8766 zone_set_pageset_high_and_batch(zone, 0);
8768 mutex_unlock(&pcp_batch_high_lock);
8772 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8774 * Returns the number of pages that arch has reserved but
8775 * is not known to alloc_large_system_hash().
8777 static unsigned long __init arch_reserved_kernel_pages(void)
8784 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8785 * machines. As memory size is increased the scale is also increased but at
8786 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8787 * quadruples the scale is increased by one, which means the size of hash table
8788 * only doubles, instead of quadrupling as well.
8789 * Because 32-bit systems cannot have large physical memory, where this scaling
8790 * makes sense, it is disabled on such platforms.
8792 #if __BITS_PER_LONG > 32
8793 #define ADAPT_SCALE_BASE (64ul << 30)
8794 #define ADAPT_SCALE_SHIFT 2
8795 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8799 * allocate a large system hash table from bootmem
8800 * - it is assumed that the hash table must contain an exact power-of-2
8801 * quantity of entries
8802 * - limit is the number of hash buckets, not the total allocation size
8804 void *__init alloc_large_system_hash(const char *tablename,
8805 unsigned long bucketsize,
8806 unsigned long numentries,
8809 unsigned int *_hash_shift,
8810 unsigned int *_hash_mask,
8811 unsigned long low_limit,
8812 unsigned long high_limit)
8814 unsigned long long max = high_limit;
8815 unsigned long log2qty, size;
8821 /* allow the kernel cmdline to have a say */
8823 /* round applicable memory size up to nearest megabyte */
8824 numentries = nr_kernel_pages;
8825 numentries -= arch_reserved_kernel_pages();
8827 /* It isn't necessary when PAGE_SIZE >= 1MB */
8828 if (PAGE_SHIFT < 20)
8829 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8831 #if __BITS_PER_LONG > 32
8833 unsigned long adapt;
8835 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8836 adapt <<= ADAPT_SCALE_SHIFT)
8841 /* limit to 1 bucket per 2^scale bytes of low memory */
8842 if (scale > PAGE_SHIFT)
8843 numentries >>= (scale - PAGE_SHIFT);
8845 numentries <<= (PAGE_SHIFT - scale);
8847 /* Make sure we've got at least a 0-order allocation.. */
8848 if (unlikely(flags & HASH_SMALL)) {
8849 /* Makes no sense without HASH_EARLY */
8850 WARN_ON(!(flags & HASH_EARLY));
8851 if (!(numentries >> *_hash_shift)) {
8852 numentries = 1UL << *_hash_shift;
8853 BUG_ON(!numentries);
8855 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8856 numentries = PAGE_SIZE / bucketsize;
8858 numentries = roundup_pow_of_two(numentries);
8860 /* limit allocation size to 1/16 total memory by default */
8862 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8863 do_div(max, bucketsize);
8865 max = min(max, 0x80000000ULL);
8867 if (numentries < low_limit)
8868 numentries = low_limit;
8869 if (numentries > max)
8872 log2qty = ilog2(numentries);
8874 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8877 size = bucketsize << log2qty;
8878 if (flags & HASH_EARLY) {
8879 if (flags & HASH_ZERO)
8880 table = memblock_alloc(size, SMP_CACHE_BYTES);
8882 table = memblock_alloc_raw(size,
8884 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8885 table = __vmalloc(size, gfp_flags);
8888 huge = is_vm_area_hugepages(table);
8891 * If bucketsize is not a power-of-two, we may free
8892 * some pages at the end of hash table which
8893 * alloc_pages_exact() automatically does
8895 table = alloc_pages_exact(size, gfp_flags);
8896 kmemleak_alloc(table, size, 1, gfp_flags);
8898 } while (!table && size > PAGE_SIZE && --log2qty);
8901 panic("Failed to allocate %s hash table\n", tablename);
8903 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8904 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8905 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8908 *_hash_shift = log2qty;
8910 *_hash_mask = (1 << log2qty) - 1;
8916 * This function checks whether pageblock includes unmovable pages or not.
8918 * PageLRU check without isolation or lru_lock could race so that
8919 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8920 * check without lock_page also may miss some movable non-lru pages at
8921 * race condition. So you can't expect this function should be exact.
8923 * Returns a page without holding a reference. If the caller wants to
8924 * dereference that page (e.g., dumping), it has to make sure that it
8925 * cannot get removed (e.g., via memory unplug) concurrently.
8928 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8929 int migratetype, int flags)
8931 unsigned long iter = 0;
8932 unsigned long pfn = page_to_pfn(page);
8933 unsigned long offset = pfn % pageblock_nr_pages;
8935 if (is_migrate_cma_page(page)) {
8937 * CMA allocations (alloc_contig_range) really need to mark
8938 * isolate CMA pageblocks even when they are not movable in fact
8939 * so consider them movable here.
8941 if (is_migrate_cma(migratetype))
8947 for (; iter < pageblock_nr_pages - offset; iter++) {
8948 page = pfn_to_page(pfn + iter);
8951 * Both, bootmem allocations and memory holes are marked
8952 * PG_reserved and are unmovable. We can even have unmovable
8953 * allocations inside ZONE_MOVABLE, for example when
8954 * specifying "movablecore".
8956 if (PageReserved(page))
8960 * If the zone is movable and we have ruled out all reserved
8961 * pages then it should be reasonably safe to assume the rest
8964 if (zone_idx(zone) == ZONE_MOVABLE)
8968 * Hugepages are not in LRU lists, but they're movable.
8969 * THPs are on the LRU, but need to be counted as #small pages.
8970 * We need not scan over tail pages because we don't
8971 * handle each tail page individually in migration.
8973 if (PageHuge(page) || PageTransCompound(page)) {
8974 struct page *head = compound_head(page);
8975 unsigned int skip_pages;
8977 if (PageHuge(page)) {
8978 if (!hugepage_migration_supported(page_hstate(head)))
8980 } else if (!PageLRU(head) && !__PageMovable(head)) {
8984 skip_pages = compound_nr(head) - (page - head);
8985 iter += skip_pages - 1;
8990 * We can't use page_count without pin a page
8991 * because another CPU can free compound page.
8992 * This check already skips compound tails of THP
8993 * because their page->_refcount is zero at all time.
8995 if (!page_ref_count(page)) {
8996 if (PageBuddy(page))
8997 iter += (1 << buddy_order(page)) - 1;
9002 * The HWPoisoned page may be not in buddy system, and
9003 * page_count() is not 0.
9005 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
9009 * We treat all PageOffline() pages as movable when offlining
9010 * to give drivers a chance to decrement their reference count
9011 * in MEM_GOING_OFFLINE in order to indicate that these pages
9012 * can be offlined as there are no direct references anymore.
9013 * For actually unmovable PageOffline() where the driver does
9014 * not support this, we will fail later when trying to actually
9015 * move these pages that still have a reference count > 0.
9016 * (false negatives in this function only)
9018 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
9021 if (__PageMovable(page) || PageLRU(page))
9025 * If there are RECLAIMABLE pages, we need to check
9026 * it. But now, memory offline itself doesn't call
9027 * shrink_node_slabs() and it still to be fixed.
9034 #ifdef CONFIG_CONTIG_ALLOC
9035 static unsigned long pfn_max_align_down(unsigned long pfn)
9037 return ALIGN_DOWN(pfn, MAX_ORDER_NR_PAGES);
9040 static unsigned long pfn_max_align_up(unsigned long pfn)
9042 return ALIGN(pfn, MAX_ORDER_NR_PAGES);
9045 #if defined(CONFIG_DYNAMIC_DEBUG) || \
9046 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
9047 /* Usage: See admin-guide/dynamic-debug-howto.rst */
9048 static void alloc_contig_dump_pages(struct list_head *page_list)
9050 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9052 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9056 list_for_each_entry(page, page_list, lru)
9057 dump_page(page, "migration failure");
9061 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9066 /* [start, end) must belong to a single zone. */
9067 static int __alloc_contig_migrate_range(struct compact_control *cc,
9068 unsigned long start, unsigned long end)
9070 /* This function is based on compact_zone() from compaction.c. */
9071 unsigned int nr_reclaimed;
9072 unsigned long pfn = start;
9073 unsigned int tries = 0;
9075 struct migration_target_control mtc = {
9076 .nid = zone_to_nid(cc->zone),
9077 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9080 lru_cache_disable();
9082 while (pfn < end || !list_empty(&cc->migratepages)) {
9083 if (fatal_signal_pending(current)) {
9088 if (list_empty(&cc->migratepages)) {
9089 cc->nr_migratepages = 0;
9090 ret = isolate_migratepages_range(cc, pfn, end);
9091 if (ret && ret != -EAGAIN)
9093 pfn = cc->migrate_pfn;
9095 } else if (++tries == 5) {
9100 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9102 cc->nr_migratepages -= nr_reclaimed;
9104 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9105 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9108 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9109 * to retry again over this error, so do the same here.
9118 alloc_contig_dump_pages(&cc->migratepages);
9119 putback_movable_pages(&cc->migratepages);
9126 * alloc_contig_range() -- tries to allocate given range of pages
9127 * @start: start PFN to allocate
9128 * @end: one-past-the-last PFN to allocate
9129 * @migratetype: migratetype of the underlying pageblocks (either
9130 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9131 * in range must have the same migratetype and it must
9132 * be either of the two.
9133 * @gfp_mask: GFP mask to use during compaction
9135 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9136 * aligned. The PFN range must belong to a single zone.
9138 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9139 * pageblocks in the range. Once isolated, the pageblocks should not
9140 * be modified by others.
9142 * Return: zero on success or negative error code. On success all
9143 * pages which PFN is in [start, end) are allocated for the caller and
9144 * need to be freed with free_contig_range().
9146 int alloc_contig_range(unsigned long start, unsigned long end,
9147 unsigned migratetype, gfp_t gfp_mask)
9149 unsigned long outer_start, outer_end;
9153 struct compact_control cc = {
9154 .nr_migratepages = 0,
9156 .zone = page_zone(pfn_to_page(start)),
9157 .mode = MIGRATE_SYNC,
9158 .ignore_skip_hint = true,
9159 .no_set_skip_hint = true,
9160 .gfp_mask = current_gfp_context(gfp_mask),
9161 .alloc_contig = true,
9163 INIT_LIST_HEAD(&cc.migratepages);
9166 * What we do here is we mark all pageblocks in range as
9167 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9168 * have different sizes, and due to the way page allocator
9169 * work, we align the range to biggest of the two pages so
9170 * that page allocator won't try to merge buddies from
9171 * different pageblocks and change MIGRATE_ISOLATE to some
9172 * other migration type.
9174 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9175 * migrate the pages from an unaligned range (ie. pages that
9176 * we are interested in). This will put all the pages in
9177 * range back to page allocator as MIGRATE_ISOLATE.
9179 * When this is done, we take the pages in range from page
9180 * allocator removing them from the buddy system. This way
9181 * page allocator will never consider using them.
9183 * This lets us mark the pageblocks back as
9184 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9185 * aligned range but not in the unaligned, original range are
9186 * put back to page allocator so that buddy can use them.
9189 ret = start_isolate_page_range(pfn_max_align_down(start),
9190 pfn_max_align_up(end), migratetype, 0);
9194 drain_all_pages(cc.zone);
9197 * In case of -EBUSY, we'd like to know which page causes problem.
9198 * So, just fall through. test_pages_isolated() has a tracepoint
9199 * which will report the busy page.
9201 * It is possible that busy pages could become available before
9202 * the call to test_pages_isolated, and the range will actually be
9203 * allocated. So, if we fall through be sure to clear ret so that
9204 * -EBUSY is not accidentally used or returned to caller.
9206 ret = __alloc_contig_migrate_range(&cc, start, end);
9207 if (ret && ret != -EBUSY)
9212 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9213 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9214 * more, all pages in [start, end) are free in page allocator.
9215 * What we are going to do is to allocate all pages from
9216 * [start, end) (that is remove them from page allocator).
9218 * The only problem is that pages at the beginning and at the
9219 * end of interesting range may be not aligned with pages that
9220 * page allocator holds, ie. they can be part of higher order
9221 * pages. Because of this, we reserve the bigger range and
9222 * once this is done free the pages we are not interested in.
9224 * We don't have to hold zone->lock here because the pages are
9225 * isolated thus they won't get removed from buddy.
9229 outer_start = start;
9230 while (!PageBuddy(pfn_to_page(outer_start))) {
9231 if (++order >= MAX_ORDER) {
9232 outer_start = start;
9235 outer_start &= ~0UL << order;
9238 if (outer_start != start) {
9239 order = buddy_order(pfn_to_page(outer_start));
9242 * outer_start page could be small order buddy page and
9243 * it doesn't include start page. Adjust outer_start
9244 * in this case to report failed page properly
9245 * on tracepoint in test_pages_isolated()
9247 if (outer_start + (1UL << order) <= start)
9248 outer_start = start;
9251 /* Make sure the range is really isolated. */
9252 if (test_pages_isolated(outer_start, end, 0)) {
9257 /* Grab isolated pages from freelists. */
9258 outer_end = isolate_freepages_range(&cc, outer_start, end);
9264 /* Free head and tail (if any) */
9265 if (start != outer_start)
9266 free_contig_range(outer_start, start - outer_start);
9267 if (end != outer_end)
9268 free_contig_range(end, outer_end - end);
9271 undo_isolate_page_range(pfn_max_align_down(start),
9272 pfn_max_align_up(end), migratetype);
9275 EXPORT_SYMBOL(alloc_contig_range);
9277 static int __alloc_contig_pages(unsigned long start_pfn,
9278 unsigned long nr_pages, gfp_t gfp_mask)
9280 unsigned long end_pfn = start_pfn + nr_pages;
9282 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9286 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9287 unsigned long nr_pages)
9289 unsigned long i, end_pfn = start_pfn + nr_pages;
9292 for (i = start_pfn; i < end_pfn; i++) {
9293 page = pfn_to_online_page(i);
9297 if (page_zone(page) != z)
9300 if (PageReserved(page))
9306 static bool zone_spans_last_pfn(const struct zone *zone,
9307 unsigned long start_pfn, unsigned long nr_pages)
9309 unsigned long last_pfn = start_pfn + nr_pages - 1;
9311 return zone_spans_pfn(zone, last_pfn);
9315 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9316 * @nr_pages: Number of contiguous pages to allocate
9317 * @gfp_mask: GFP mask to limit search and used during compaction
9319 * @nodemask: Mask for other possible nodes
9321 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9322 * on an applicable zonelist to find a contiguous pfn range which can then be
9323 * tried for allocation with alloc_contig_range(). This routine is intended
9324 * for allocation requests which can not be fulfilled with the buddy allocator.
9326 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9327 * power of two, then allocated range is also guaranteed to be aligned to same
9328 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9330 * Allocated pages can be freed with free_contig_range() or by manually calling
9331 * __free_page() on each allocated page.
9333 * Return: pointer to contiguous pages on success, or NULL if not successful.
9335 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9336 int nid, nodemask_t *nodemask)
9338 unsigned long ret, pfn, flags;
9339 struct zonelist *zonelist;
9343 zonelist = node_zonelist(nid, gfp_mask);
9344 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9345 gfp_zone(gfp_mask), nodemask) {
9346 spin_lock_irqsave(&zone->lock, flags);
9348 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9349 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9350 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9352 * We release the zone lock here because
9353 * alloc_contig_range() will also lock the zone
9354 * at some point. If there's an allocation
9355 * spinning on this lock, it may win the race
9356 * and cause alloc_contig_range() to fail...
9358 spin_unlock_irqrestore(&zone->lock, flags);
9359 ret = __alloc_contig_pages(pfn, nr_pages,
9362 return pfn_to_page(pfn);
9363 spin_lock_irqsave(&zone->lock, flags);
9367 spin_unlock_irqrestore(&zone->lock, flags);
9371 #endif /* CONFIG_CONTIG_ALLOC */
9373 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9375 unsigned long count = 0;
9377 for (; nr_pages--; pfn++) {
9378 struct page *page = pfn_to_page(pfn);
9380 count += page_count(page) != 1;
9383 WARN(count != 0, "%lu pages are still in use!\n", count);
9385 EXPORT_SYMBOL(free_contig_range);
9388 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9389 * page high values need to be recalculated.
9391 void zone_pcp_update(struct zone *zone, int cpu_online)
9393 mutex_lock(&pcp_batch_high_lock);
9394 zone_set_pageset_high_and_batch(zone, cpu_online);
9395 mutex_unlock(&pcp_batch_high_lock);
9399 * Effectively disable pcplists for the zone by setting the high limit to 0
9400 * and draining all cpus. A concurrent page freeing on another CPU that's about
9401 * to put the page on pcplist will either finish before the drain and the page
9402 * will be drained, or observe the new high limit and skip the pcplist.
9404 * Must be paired with a call to zone_pcp_enable().
9406 void zone_pcp_disable(struct zone *zone)
9408 mutex_lock(&pcp_batch_high_lock);
9409 __zone_set_pageset_high_and_batch(zone, 0, 1);
9410 __drain_all_pages(zone, true);
9413 void zone_pcp_enable(struct zone *zone)
9415 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9416 mutex_unlock(&pcp_batch_high_lock);
9419 void zone_pcp_reset(struct zone *zone)
9422 struct per_cpu_zonestat *pzstats;
9424 if (zone->per_cpu_pageset != &boot_pageset) {
9425 for_each_online_cpu(cpu) {
9426 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9427 drain_zonestat(zone, pzstats);
9429 free_percpu(zone->per_cpu_pageset);
9430 free_percpu(zone->per_cpu_zonestats);
9431 zone->per_cpu_pageset = &boot_pageset;
9432 zone->per_cpu_zonestats = &boot_zonestats;
9436 #ifdef CONFIG_MEMORY_HOTREMOVE
9438 * All pages in the range must be in a single zone, must not contain holes,
9439 * must span full sections, and must be isolated before calling this function.
9441 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9443 unsigned long pfn = start_pfn;
9447 unsigned long flags;
9449 offline_mem_sections(pfn, end_pfn);
9450 zone = page_zone(pfn_to_page(pfn));
9451 spin_lock_irqsave(&zone->lock, flags);
9452 while (pfn < end_pfn) {
9453 page = pfn_to_page(pfn);
9455 * The HWPoisoned page may be not in buddy system, and
9456 * page_count() is not 0.
9458 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9463 * At this point all remaining PageOffline() pages have a
9464 * reference count of 0 and can simply be skipped.
9466 if (PageOffline(page)) {
9467 BUG_ON(page_count(page));
9468 BUG_ON(PageBuddy(page));
9473 BUG_ON(page_count(page));
9474 BUG_ON(!PageBuddy(page));
9475 order = buddy_order(page);
9476 del_page_from_free_list(page, zone, order);
9477 pfn += (1 << order);
9479 spin_unlock_irqrestore(&zone->lock, flags);
9484 * This function returns a stable result only if called under zone lock.
9486 bool is_free_buddy_page(struct page *page)
9488 unsigned long pfn = page_to_pfn(page);
9491 for (order = 0; order < MAX_ORDER; order++) {
9492 struct page *page_head = page - (pfn & ((1 << order) - 1));
9494 if (PageBuddy(page_head) &&
9495 buddy_order_unsafe(page_head) >= order)
9499 return order < MAX_ORDER;
9501 EXPORT_SYMBOL(is_free_buddy_page);
9503 #ifdef CONFIG_MEMORY_FAILURE
9505 * Break down a higher-order page in sub-pages, and keep our target out of
9508 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9509 struct page *target, int low, int high,
9512 unsigned long size = 1 << high;
9513 struct page *current_buddy, *next_page;
9515 while (high > low) {
9519 if (target >= &page[size]) {
9520 next_page = page + size;
9521 current_buddy = page;
9524 current_buddy = page + size;
9527 if (set_page_guard(zone, current_buddy, high, migratetype))
9530 if (current_buddy != target) {
9531 add_to_free_list(current_buddy, zone, high, migratetype);
9532 set_buddy_order(current_buddy, high);
9539 * Take a page that will be marked as poisoned off the buddy allocator.
9541 bool take_page_off_buddy(struct page *page)
9543 struct zone *zone = page_zone(page);
9544 unsigned long pfn = page_to_pfn(page);
9545 unsigned long flags;
9549 spin_lock_irqsave(&zone->lock, flags);
9550 for (order = 0; order < MAX_ORDER; order++) {
9551 struct page *page_head = page - (pfn & ((1 << order) - 1));
9552 int page_order = buddy_order(page_head);
9554 if (PageBuddy(page_head) && page_order >= order) {
9555 unsigned long pfn_head = page_to_pfn(page_head);
9556 int migratetype = get_pfnblock_migratetype(page_head,
9559 del_page_from_free_list(page_head, zone, page_order);
9560 break_down_buddy_pages(zone, page_head, page, 0,
9561 page_order, migratetype);
9562 SetPageHWPoisonTakenOff(page);
9563 if (!is_migrate_isolate(migratetype))
9564 __mod_zone_freepage_state(zone, -1, migratetype);
9568 if (page_count(page_head) > 0)
9571 spin_unlock_irqrestore(&zone->lock, flags);
9576 * Cancel takeoff done by take_page_off_buddy().
9578 bool put_page_back_buddy(struct page *page)
9580 struct zone *zone = page_zone(page);
9581 unsigned long pfn = page_to_pfn(page);
9582 unsigned long flags;
9583 int migratetype = get_pfnblock_migratetype(page, pfn);
9586 spin_lock_irqsave(&zone->lock, flags);
9587 if (put_page_testzero(page)) {
9588 ClearPageHWPoisonTakenOff(page);
9589 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9590 if (TestClearPageHWPoison(page)) {
9591 num_poisoned_pages_dec();
9595 spin_unlock_irqrestore(&zone->lock, flags);
9601 #ifdef CONFIG_ZONE_DMA
9602 bool has_managed_dma(void)
9604 struct pglist_data *pgdat;
9606 for_each_online_pgdat(pgdat) {
9607 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9609 if (managed_zone(zone))
9614 #endif /* CONFIG_ZONE_DMA */