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"
86 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
87 typedef int __bitwise fpi_t;
89 /* No special request */
90 #define FPI_NONE ((__force fpi_t)0)
93 * Skip free page reporting notification for the (possibly merged) page.
94 * This does not hinder free page reporting from grabbing the page,
95 * reporting it and marking it "reported" - it only skips notifying
96 * the free page reporting infrastructure about a newly freed page. For
97 * example, used when temporarily pulling a page from a freelist and
98 * putting it back unmodified.
100 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
103 * Place the (possibly merged) page to the tail of the freelist. Will ignore
104 * page shuffling (relevant code - e.g., memory onlining - is expected to
105 * shuffle the whole zone).
107 * Note: No code should rely on this flag for correctness - it's purely
108 * to allow for optimizations when handing back either fresh pages
109 * (memory onlining) or untouched pages (page isolation, free page
112 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
115 * Don't poison memory with KASAN (only for the tag-based modes).
116 * During boot, all non-reserved memblock memory is exposed to page_alloc.
117 * Poisoning all that memory lengthens boot time, especially on systems with
118 * large amount of RAM. This flag is used to skip that poisoning.
119 * This is only done for the tag-based KASAN modes, as those are able to
120 * detect memory corruptions with the memory tags assigned by default.
121 * All memory allocated normally after boot gets poisoned as usual.
123 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
125 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
126 static DEFINE_MUTEX(pcp_batch_high_lock);
127 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
132 static DEFINE_PER_CPU(struct pagesets, pagesets) = {
133 .lock = INIT_LOCAL_LOCK(lock),
136 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
137 DEFINE_PER_CPU(int, numa_node);
138 EXPORT_PER_CPU_SYMBOL(numa_node);
141 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
143 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
145 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
146 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
147 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
148 * defined in <linux/topology.h>.
150 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
151 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
154 /* work_structs for global per-cpu drains */
157 struct work_struct work;
159 static DEFINE_MUTEX(pcpu_drain_mutex);
160 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
162 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
163 volatile unsigned long latent_entropy __latent_entropy;
164 EXPORT_SYMBOL(latent_entropy);
168 * Array of node states.
170 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
171 [N_POSSIBLE] = NODE_MASK_ALL,
172 [N_ONLINE] = { { [0] = 1UL } },
174 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
175 #ifdef CONFIG_HIGHMEM
176 [N_HIGH_MEMORY] = { { [0] = 1UL } },
178 [N_MEMORY] = { { [0] = 1UL } },
179 [N_CPU] = { { [0] = 1UL } },
182 EXPORT_SYMBOL(node_states);
184 atomic_long_t _totalram_pages __read_mostly;
185 EXPORT_SYMBOL(_totalram_pages);
186 unsigned long totalreserve_pages __read_mostly;
187 unsigned long totalcma_pages __read_mostly;
189 int percpu_pagelist_high_fraction;
190 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
191 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
192 EXPORT_SYMBOL(init_on_alloc);
194 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
195 EXPORT_SYMBOL(init_on_free);
197 static bool _init_on_alloc_enabled_early __read_mostly
198 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
199 static int __init early_init_on_alloc(char *buf)
202 return kstrtobool(buf, &_init_on_alloc_enabled_early);
204 early_param("init_on_alloc", early_init_on_alloc);
206 static bool _init_on_free_enabled_early __read_mostly
207 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
208 static int __init early_init_on_free(char *buf)
210 return kstrtobool(buf, &_init_on_free_enabled_early);
212 early_param("init_on_free", early_init_on_free);
215 * A cached value of the page's pageblock's migratetype, used when the page is
216 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
217 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
218 * Also the migratetype set in the page does not necessarily match the pcplist
219 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
220 * other index - this ensures that it will be put on the correct CMA freelist.
222 static inline int get_pcppage_migratetype(struct page *page)
227 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
229 page->index = migratetype;
232 #ifdef CONFIG_PM_SLEEP
234 * The following functions are used by the suspend/hibernate code to temporarily
235 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
236 * while devices are suspended. To avoid races with the suspend/hibernate code,
237 * they should always be called with system_transition_mutex held
238 * (gfp_allowed_mask also should only be modified with system_transition_mutex
239 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
240 * with that modification).
243 static gfp_t saved_gfp_mask;
245 void pm_restore_gfp_mask(void)
247 WARN_ON(!mutex_is_locked(&system_transition_mutex));
248 if (saved_gfp_mask) {
249 gfp_allowed_mask = saved_gfp_mask;
254 void pm_restrict_gfp_mask(void)
256 WARN_ON(!mutex_is_locked(&system_transition_mutex));
257 WARN_ON(saved_gfp_mask);
258 saved_gfp_mask = gfp_allowed_mask;
259 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
262 bool pm_suspended_storage(void)
264 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
268 #endif /* CONFIG_PM_SLEEP */
270 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
271 unsigned int pageblock_order __read_mostly;
274 static void __free_pages_ok(struct page *page, unsigned int order,
278 * results with 256, 32 in the lowmem_reserve sysctl:
279 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
280 * 1G machine -> (16M dma, 784M normal, 224M high)
281 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
282 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
283 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
285 * TBD: should special case ZONE_DMA32 machines here - in those we normally
286 * don't need any ZONE_NORMAL reservation
288 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
289 #ifdef CONFIG_ZONE_DMA
292 #ifdef CONFIG_ZONE_DMA32
296 #ifdef CONFIG_HIGHMEM
302 static char * const zone_names[MAX_NR_ZONES] = {
303 #ifdef CONFIG_ZONE_DMA
306 #ifdef CONFIG_ZONE_DMA32
310 #ifdef CONFIG_HIGHMEM
314 #ifdef CONFIG_ZONE_DEVICE
319 const char * const migratetype_names[MIGRATE_TYPES] = {
327 #ifdef CONFIG_MEMORY_ISOLATION
332 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
333 [NULL_COMPOUND_DTOR] = NULL,
334 [COMPOUND_PAGE_DTOR] = free_compound_page,
335 #ifdef CONFIG_HUGETLB_PAGE
336 [HUGETLB_PAGE_DTOR] = free_huge_page,
338 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
339 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
343 int min_free_kbytes = 1024;
344 int user_min_free_kbytes = -1;
345 int watermark_boost_factor __read_mostly = 15000;
346 int watermark_scale_factor = 10;
348 static unsigned long nr_kernel_pages __initdata;
349 static unsigned long nr_all_pages __initdata;
350 static unsigned long dma_reserve __initdata;
352 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
353 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
354 static unsigned long required_kernelcore __initdata;
355 static unsigned long required_kernelcore_percent __initdata;
356 static unsigned long required_movablecore __initdata;
357 static unsigned long required_movablecore_percent __initdata;
358 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
359 static bool mirrored_kernelcore __meminitdata;
361 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
363 EXPORT_SYMBOL(movable_zone);
366 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
367 unsigned int nr_online_nodes __read_mostly = 1;
368 EXPORT_SYMBOL(nr_node_ids);
369 EXPORT_SYMBOL(nr_online_nodes);
372 int page_group_by_mobility_disabled __read_mostly;
374 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
376 * During boot we initialize deferred pages on-demand, as needed, but once
377 * page_alloc_init_late() has finished, the deferred pages are all initialized,
378 * and we can permanently disable that path.
380 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
382 static inline bool deferred_pages_enabled(void)
384 return static_branch_unlikely(&deferred_pages);
387 /* Returns true if the struct page for the pfn is uninitialised */
388 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
390 int nid = early_pfn_to_nid(pfn);
392 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
399 * Returns true when the remaining initialisation should be deferred until
400 * later in the boot cycle when it can be parallelised.
402 static bool __meminit
403 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
405 static unsigned long prev_end_pfn, nr_initialised;
408 * prev_end_pfn static that contains the end of previous zone
409 * No need to protect because called very early in boot before smp_init.
411 if (prev_end_pfn != end_pfn) {
412 prev_end_pfn = end_pfn;
416 /* Always populate low zones for address-constrained allocations */
417 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
420 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
423 * We start only with one section of pages, more pages are added as
424 * needed until the rest of deferred pages are initialized.
427 if ((nr_initialised > PAGES_PER_SECTION) &&
428 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
429 NODE_DATA(nid)->first_deferred_pfn = pfn;
435 static inline bool deferred_pages_enabled(void)
440 static inline bool early_page_uninitialised(unsigned long pfn)
445 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
451 /* Return a pointer to the bitmap storing bits affecting a block of pages */
452 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
455 #ifdef CONFIG_SPARSEMEM
456 return section_to_usemap(__pfn_to_section(pfn));
458 return page_zone(page)->pageblock_flags;
459 #endif /* CONFIG_SPARSEMEM */
462 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
464 #ifdef CONFIG_SPARSEMEM
465 pfn &= (PAGES_PER_SECTION-1);
467 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
468 #endif /* CONFIG_SPARSEMEM */
469 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
472 static __always_inline
473 unsigned long __get_pfnblock_flags_mask(const struct page *page,
477 unsigned long *bitmap;
478 unsigned long bitidx, word_bitidx;
481 bitmap = get_pageblock_bitmap(page, pfn);
482 bitidx = pfn_to_bitidx(page, pfn);
483 word_bitidx = bitidx / BITS_PER_LONG;
484 bitidx &= (BITS_PER_LONG-1);
486 word = bitmap[word_bitidx];
487 return (word >> bitidx) & mask;
491 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
492 * @page: The page within the block of interest
493 * @pfn: The target page frame number
494 * @mask: mask of bits that the caller is interested in
496 * Return: pageblock_bits flags
498 unsigned long get_pfnblock_flags_mask(const struct page *page,
499 unsigned long pfn, unsigned long mask)
501 return __get_pfnblock_flags_mask(page, pfn, mask);
504 static __always_inline int get_pfnblock_migratetype(const struct page *page,
507 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
511 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
512 * @page: The page within the block of interest
513 * @flags: The flags to set
514 * @pfn: The target page frame number
515 * @mask: mask of bits that the caller is interested in
517 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
521 unsigned long *bitmap;
522 unsigned long bitidx, word_bitidx;
523 unsigned long old_word, word;
525 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
526 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
528 bitmap = get_pageblock_bitmap(page, pfn);
529 bitidx = pfn_to_bitidx(page, pfn);
530 word_bitidx = bitidx / BITS_PER_LONG;
531 bitidx &= (BITS_PER_LONG-1);
533 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
538 word = READ_ONCE(bitmap[word_bitidx]);
540 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
541 if (word == old_word)
547 void set_pageblock_migratetype(struct page *page, int migratetype)
549 if (unlikely(page_group_by_mobility_disabled &&
550 migratetype < MIGRATE_PCPTYPES))
551 migratetype = MIGRATE_UNMOVABLE;
553 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
554 page_to_pfn(page), MIGRATETYPE_MASK);
557 #ifdef CONFIG_DEBUG_VM
558 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
562 unsigned long pfn = page_to_pfn(page);
563 unsigned long sp, start_pfn;
566 seq = zone_span_seqbegin(zone);
567 start_pfn = zone->zone_start_pfn;
568 sp = zone->spanned_pages;
569 if (!zone_spans_pfn(zone, pfn))
571 } while (zone_span_seqretry(zone, seq));
574 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
575 pfn, zone_to_nid(zone), zone->name,
576 start_pfn, start_pfn + sp);
581 static int page_is_consistent(struct zone *zone, struct page *page)
583 if (zone != page_zone(page))
589 * Temporary debugging check for pages not lying within a given zone.
591 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
593 if (page_outside_zone_boundaries(zone, page))
595 if (!page_is_consistent(zone, page))
601 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
607 static void bad_page(struct page *page, const char *reason)
609 static unsigned long resume;
610 static unsigned long nr_shown;
611 static unsigned long nr_unshown;
614 * Allow a burst of 60 reports, then keep quiet for that minute;
615 * or allow a steady drip of one report per second.
617 if (nr_shown == 60) {
618 if (time_before(jiffies, resume)) {
624 "BUG: Bad page state: %lu messages suppressed\n",
631 resume = jiffies + 60 * HZ;
633 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
634 current->comm, page_to_pfn(page));
635 dump_page(page, reason);
640 /* Leave bad fields for debug, except PageBuddy could make trouble */
641 page_mapcount_reset(page); /* remove PageBuddy */
642 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
645 static inline unsigned int order_to_pindex(int migratetype, int order)
649 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
650 if (order > PAGE_ALLOC_COSTLY_ORDER) {
651 VM_BUG_ON(order != pageblock_order);
652 base = PAGE_ALLOC_COSTLY_ORDER + 1;
655 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
658 return (MIGRATE_PCPTYPES * base) + migratetype;
661 static inline int pindex_to_order(unsigned int pindex)
663 int order = pindex / MIGRATE_PCPTYPES;
665 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
666 if (order > PAGE_ALLOC_COSTLY_ORDER)
667 order = pageblock_order;
669 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
675 static inline bool pcp_allowed_order(unsigned int order)
677 if (order <= PAGE_ALLOC_COSTLY_ORDER)
679 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
680 if (order == pageblock_order)
686 static inline void free_the_page(struct page *page, unsigned int order)
688 if (pcp_allowed_order(order)) /* Via pcp? */
689 free_unref_page(page, order);
691 __free_pages_ok(page, order, FPI_NONE);
695 * Higher-order pages are called "compound pages". They are structured thusly:
697 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
699 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
700 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
702 * The first tail page's ->compound_dtor holds the offset in array of compound
703 * page destructors. See compound_page_dtors.
705 * The first tail page's ->compound_order holds the order of allocation.
706 * This usage means that zero-order pages may not be compound.
709 void free_compound_page(struct page *page)
711 mem_cgroup_uncharge(page_folio(page));
712 free_the_page(page, compound_order(page));
715 static void prep_compound_head(struct page *page, unsigned int order)
717 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
718 set_compound_order(page, order);
719 atomic_set(compound_mapcount_ptr(page), -1);
720 atomic_set(compound_pincount_ptr(page), 0);
723 static void prep_compound_tail(struct page *head, int tail_idx)
725 struct page *p = head + tail_idx;
727 p->mapping = TAIL_MAPPING;
728 set_compound_head(p, head);
731 void prep_compound_page(struct page *page, unsigned int order)
734 int nr_pages = 1 << order;
737 for (i = 1; i < nr_pages; i++)
738 prep_compound_tail(page, i);
740 prep_compound_head(page, order);
743 #ifdef CONFIG_DEBUG_PAGEALLOC
744 unsigned int _debug_guardpage_minorder;
746 bool _debug_pagealloc_enabled_early __read_mostly
747 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
748 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
749 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
750 EXPORT_SYMBOL(_debug_pagealloc_enabled);
752 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
754 static int __init early_debug_pagealloc(char *buf)
756 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
758 early_param("debug_pagealloc", early_debug_pagealloc);
760 static int __init debug_guardpage_minorder_setup(char *buf)
764 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
765 pr_err("Bad debug_guardpage_minorder value\n");
768 _debug_guardpage_minorder = res;
769 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
772 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
774 static inline bool set_page_guard(struct zone *zone, struct page *page,
775 unsigned int order, int migratetype)
777 if (!debug_guardpage_enabled())
780 if (order >= debug_guardpage_minorder())
783 __SetPageGuard(page);
784 INIT_LIST_HEAD(&page->lru);
785 set_page_private(page, order);
786 /* Guard pages are not available for any usage */
787 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
792 static inline void clear_page_guard(struct zone *zone, struct page *page,
793 unsigned int order, int migratetype)
795 if (!debug_guardpage_enabled())
798 __ClearPageGuard(page);
800 set_page_private(page, 0);
801 if (!is_migrate_isolate(migratetype))
802 __mod_zone_freepage_state(zone, (1 << order), migratetype);
805 static inline bool set_page_guard(struct zone *zone, struct page *page,
806 unsigned int order, int migratetype) { return false; }
807 static inline void clear_page_guard(struct zone *zone, struct page *page,
808 unsigned int order, int migratetype) {}
812 * Enable static keys related to various memory debugging and hardening options.
813 * Some override others, and depend on early params that are evaluated in the
814 * order of appearance. So we need to first gather the full picture of what was
815 * enabled, and then make decisions.
817 void init_mem_debugging_and_hardening(void)
819 bool page_poisoning_requested = false;
821 #ifdef CONFIG_PAGE_POISONING
823 * Page poisoning is debug page alloc for some arches. If
824 * either of those options are enabled, enable poisoning.
826 if (page_poisoning_enabled() ||
827 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
828 debug_pagealloc_enabled())) {
829 static_branch_enable(&_page_poisoning_enabled);
830 page_poisoning_requested = true;
834 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
835 page_poisoning_requested) {
836 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
837 "will take precedence over init_on_alloc and init_on_free\n");
838 _init_on_alloc_enabled_early = false;
839 _init_on_free_enabled_early = false;
842 if (_init_on_alloc_enabled_early)
843 static_branch_enable(&init_on_alloc);
845 static_branch_disable(&init_on_alloc);
847 if (_init_on_free_enabled_early)
848 static_branch_enable(&init_on_free);
850 static_branch_disable(&init_on_free);
852 #ifdef CONFIG_DEBUG_PAGEALLOC
853 if (!debug_pagealloc_enabled())
856 static_branch_enable(&_debug_pagealloc_enabled);
858 if (!debug_guardpage_minorder())
861 static_branch_enable(&_debug_guardpage_enabled);
865 static inline void set_buddy_order(struct page *page, unsigned int order)
867 set_page_private(page, order);
868 __SetPageBuddy(page);
871 #ifdef CONFIG_COMPACTION
872 static inline struct capture_control *task_capc(struct zone *zone)
874 struct capture_control *capc = current->capture_control;
876 return unlikely(capc) &&
877 !(current->flags & PF_KTHREAD) &&
879 capc->cc->zone == zone ? capc : NULL;
883 compaction_capture(struct capture_control *capc, struct page *page,
884 int order, int migratetype)
886 if (!capc || order != capc->cc->order)
889 /* Do not accidentally pollute CMA or isolated regions*/
890 if (is_migrate_cma(migratetype) ||
891 is_migrate_isolate(migratetype))
895 * Do not let lower order allocations pollute a movable pageblock.
896 * This might let an unmovable request use a reclaimable pageblock
897 * and vice-versa but no more than normal fallback logic which can
898 * have trouble finding a high-order free page.
900 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
908 static inline struct capture_control *task_capc(struct zone *zone)
914 compaction_capture(struct capture_control *capc, struct page *page,
915 int order, int migratetype)
919 #endif /* CONFIG_COMPACTION */
921 /* Used for pages not on another list */
922 static inline void add_to_free_list(struct page *page, struct zone *zone,
923 unsigned int order, int migratetype)
925 struct free_area *area = &zone->free_area[order];
927 list_add(&page->lru, &area->free_list[migratetype]);
931 /* Used for pages not on another list */
932 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
933 unsigned int order, int migratetype)
935 struct free_area *area = &zone->free_area[order];
937 list_add_tail(&page->lru, &area->free_list[migratetype]);
942 * Used for pages which are on another list. Move the pages to the tail
943 * of the list - so the moved pages won't immediately be considered for
944 * allocation again (e.g., optimization for memory onlining).
946 static inline void move_to_free_list(struct page *page, struct zone *zone,
947 unsigned int order, int migratetype)
949 struct free_area *area = &zone->free_area[order];
951 list_move_tail(&page->lru, &area->free_list[migratetype]);
954 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
957 /* clear reported state and update reported page count */
958 if (page_reported(page))
959 __ClearPageReported(page);
961 list_del(&page->lru);
962 __ClearPageBuddy(page);
963 set_page_private(page, 0);
964 zone->free_area[order].nr_free--;
968 * If this is not the largest possible page, check if the buddy
969 * of the next-highest order is free. If it is, it's possible
970 * that pages are being freed that will coalesce soon. In case,
971 * that is happening, add the free page to the tail of the list
972 * so it's less likely to be used soon and more likely to be merged
973 * as a higher order page
976 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
977 struct page *page, unsigned int order)
979 unsigned long higher_page_pfn;
980 struct page *higher_page;
982 if (order >= MAX_ORDER - 2)
985 higher_page_pfn = buddy_pfn & pfn;
986 higher_page = page + (higher_page_pfn - pfn);
988 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
993 * Freeing function for a buddy system allocator.
995 * The concept of a buddy system is to maintain direct-mapped table
996 * (containing bit values) for memory blocks of various "orders".
997 * The bottom level table contains the map for the smallest allocatable
998 * units of memory (here, pages), and each level above it describes
999 * pairs of units from the levels below, hence, "buddies".
1000 * At a high level, all that happens here is marking the table entry
1001 * at the bottom level available, and propagating the changes upward
1002 * as necessary, plus some accounting needed to play nicely with other
1003 * parts of the VM system.
1004 * At each level, we keep a list of pages, which are heads of continuous
1005 * free pages of length of (1 << order) and marked with PageBuddy.
1006 * Page's order is recorded in page_private(page) field.
1007 * So when we are allocating or freeing one, we can derive the state of the
1008 * other. That is, if we allocate a small block, and both were
1009 * free, the remainder of the region must be split into blocks.
1010 * If a block is freed, and its buddy is also free, then this
1011 * triggers coalescing into a block of larger size.
1016 static inline void __free_one_page(struct page *page,
1018 struct zone *zone, unsigned int order,
1019 int migratetype, fpi_t fpi_flags)
1021 struct capture_control *capc = task_capc(zone);
1022 unsigned long buddy_pfn;
1023 unsigned long combined_pfn;
1027 VM_BUG_ON(!zone_is_initialized(zone));
1028 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1030 VM_BUG_ON(migratetype == -1);
1031 if (likely(!is_migrate_isolate(migratetype)))
1032 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1034 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1035 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1037 while (order < MAX_ORDER - 1) {
1038 if (compaction_capture(capc, page, order, migratetype)) {
1039 __mod_zone_freepage_state(zone, -(1 << order),
1044 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
1048 if (unlikely(order >= pageblock_order)) {
1050 * We want to prevent merge between freepages on pageblock
1051 * without fallbacks and normal pageblock. Without this,
1052 * pageblock isolation could cause incorrect freepage or CMA
1053 * accounting or HIGHATOMIC accounting.
1055 int buddy_mt = get_pageblock_migratetype(buddy);
1057 if (migratetype != buddy_mt
1058 && (!migratetype_is_mergeable(migratetype) ||
1059 !migratetype_is_mergeable(buddy_mt)))
1064 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1065 * merge with it and move up one order.
1067 if (page_is_guard(buddy))
1068 clear_page_guard(zone, buddy, order, migratetype);
1070 del_page_from_free_list(buddy, zone, order);
1071 combined_pfn = buddy_pfn & pfn;
1072 page = page + (combined_pfn - pfn);
1078 set_buddy_order(page, order);
1080 if (fpi_flags & FPI_TO_TAIL)
1082 else if (is_shuffle_order(order))
1083 to_tail = shuffle_pick_tail();
1085 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1088 add_to_free_list_tail(page, zone, order, migratetype);
1090 add_to_free_list(page, zone, order, migratetype);
1092 /* Notify page reporting subsystem of freed page */
1093 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1094 page_reporting_notify_free(order);
1098 * split_free_page() -- split a free page at split_pfn_offset
1099 * @free_page: the original free page
1100 * @order: the order of the page
1101 * @split_pfn_offset: split offset within the page
1103 * It is used when the free page crosses two pageblocks with different migratetypes
1104 * at split_pfn_offset within the page. The split free page will be put into
1105 * separate migratetype lists afterwards. Otherwise, the function achieves
1108 void split_free_page(struct page *free_page,
1109 int order, unsigned long split_pfn_offset)
1111 struct zone *zone = page_zone(free_page);
1112 unsigned long free_page_pfn = page_to_pfn(free_page);
1114 unsigned long flags;
1115 int free_page_order;
1117 if (split_pfn_offset == 0)
1120 spin_lock_irqsave(&zone->lock, flags);
1121 del_page_from_free_list(free_page, zone, order);
1122 for (pfn = free_page_pfn;
1123 pfn < free_page_pfn + (1UL << order);) {
1124 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
1126 free_page_order = min_t(int,
1127 pfn ? __ffs(pfn) : order,
1128 __fls(split_pfn_offset));
1129 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
1131 pfn += 1UL << free_page_order;
1132 split_pfn_offset -= (1UL << free_page_order);
1133 /* we have done the first part, now switch to second part */
1134 if (split_pfn_offset == 0)
1135 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
1137 spin_unlock_irqrestore(&zone->lock, flags);
1140 * A bad page could be due to a number of fields. Instead of multiple branches,
1141 * try and check multiple fields with one check. The caller must do a detailed
1142 * check if necessary.
1144 static inline bool page_expected_state(struct page *page,
1145 unsigned long check_flags)
1147 if (unlikely(atomic_read(&page->_mapcount) != -1))
1150 if (unlikely((unsigned long)page->mapping |
1151 page_ref_count(page) |
1155 (page->flags & check_flags)))
1161 static const char *page_bad_reason(struct page *page, unsigned long flags)
1163 const char *bad_reason = NULL;
1165 if (unlikely(atomic_read(&page->_mapcount) != -1))
1166 bad_reason = "nonzero mapcount";
1167 if (unlikely(page->mapping != NULL))
1168 bad_reason = "non-NULL mapping";
1169 if (unlikely(page_ref_count(page) != 0))
1170 bad_reason = "nonzero _refcount";
1171 if (unlikely(page->flags & flags)) {
1172 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1173 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1175 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1178 if (unlikely(page->memcg_data))
1179 bad_reason = "page still charged to cgroup";
1184 static void check_free_page_bad(struct page *page)
1187 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1190 static inline int check_free_page(struct page *page)
1192 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1195 /* Something has gone sideways, find it */
1196 check_free_page_bad(page);
1200 static int free_tail_pages_check(struct page *head_page, struct page *page)
1205 * We rely page->lru.next never has bit 0 set, unless the page
1206 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1208 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1210 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1214 switch (page - head_page) {
1216 /* the first tail page: ->mapping may be compound_mapcount() */
1217 if (unlikely(compound_mapcount(page))) {
1218 bad_page(page, "nonzero compound_mapcount");
1224 * the second tail page: ->mapping is
1225 * deferred_list.next -- ignore value.
1229 if (page->mapping != TAIL_MAPPING) {
1230 bad_page(page, "corrupted mapping in tail page");
1235 if (unlikely(!PageTail(page))) {
1236 bad_page(page, "PageTail not set");
1239 if (unlikely(compound_head(page) != head_page)) {
1240 bad_page(page, "compound_head not consistent");
1245 page->mapping = NULL;
1246 clear_compound_head(page);
1251 * Skip KASAN memory poisoning when either:
1253 * 1. Deferred memory initialization has not yet completed,
1254 * see the explanation below.
1255 * 2. Skipping poisoning is requested via FPI_SKIP_KASAN_POISON,
1256 * see the comment next to it.
1257 * 3. Skipping poisoning is requested via __GFP_SKIP_KASAN_POISON,
1258 * see the comment next to it.
1260 * Poisoning pages during deferred memory init will greatly lengthen the
1261 * process and cause problem in large memory systems as the deferred pages
1262 * initialization is done with interrupt disabled.
1264 * Assuming that there will be no reference to those newly initialized
1265 * pages before they are ever allocated, this should have no effect on
1266 * KASAN memory tracking as the poison will be properly inserted at page
1267 * allocation time. The only corner case is when pages are allocated by
1268 * on-demand allocation and then freed again before the deferred pages
1269 * initialization is done, but this is not likely to happen.
1271 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1273 return deferred_pages_enabled() ||
1274 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
1275 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
1276 PageSkipKASanPoison(page);
1279 static void kernel_init_free_pages(struct page *page, int numpages)
1283 /* s390's use of memset() could override KASAN redzones. */
1284 kasan_disable_current();
1285 for (i = 0; i < numpages; i++) {
1286 u8 tag = page_kasan_tag(page + i);
1287 page_kasan_tag_reset(page + i);
1288 clear_highpage(page + i);
1289 page_kasan_tag_set(page + i, tag);
1291 kasan_enable_current();
1294 static __always_inline bool free_pages_prepare(struct page *page,
1295 unsigned int order, bool check_free, fpi_t fpi_flags)
1298 bool init = want_init_on_free();
1300 VM_BUG_ON_PAGE(PageTail(page), page);
1302 trace_mm_page_free(page, order);
1304 if (unlikely(PageHWPoison(page)) && !order) {
1306 * Do not let hwpoison pages hit pcplists/buddy
1307 * Untie memcg state and reset page's owner
1309 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1310 __memcg_kmem_uncharge_page(page, order);
1311 reset_page_owner(page, order);
1312 page_table_check_free(page, order);
1317 * Check tail pages before head page information is cleared to
1318 * avoid checking PageCompound for order-0 pages.
1320 if (unlikely(order)) {
1321 bool compound = PageCompound(page);
1324 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1327 ClearPageDoubleMap(page);
1328 ClearPageHasHWPoisoned(page);
1330 for (i = 1; i < (1 << order); i++) {
1332 bad += free_tail_pages_check(page, page + i);
1333 if (unlikely(check_free_page(page + i))) {
1337 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1340 if (PageMappingFlags(page))
1341 page->mapping = NULL;
1342 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1343 __memcg_kmem_uncharge_page(page, order);
1345 bad += check_free_page(page);
1349 page_cpupid_reset_last(page);
1350 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1351 reset_page_owner(page, order);
1352 page_table_check_free(page, order);
1354 if (!PageHighMem(page)) {
1355 debug_check_no_locks_freed(page_address(page),
1356 PAGE_SIZE << order);
1357 debug_check_no_obj_freed(page_address(page),
1358 PAGE_SIZE << order);
1361 kernel_poison_pages(page, 1 << order);
1364 * As memory initialization might be integrated into KASAN,
1365 * KASAN poisoning and memory initialization code must be
1366 * kept together to avoid discrepancies in behavior.
1368 * With hardware tag-based KASAN, memory tags must be set before the
1369 * page becomes unavailable via debug_pagealloc or arch_free_page.
1371 if (!should_skip_kasan_poison(page, fpi_flags)) {
1372 kasan_poison_pages(page, order, init);
1374 /* Memory is already initialized if KASAN did it internally. */
1375 if (kasan_has_integrated_init())
1379 kernel_init_free_pages(page, 1 << order);
1382 * arch_free_page() can make the page's contents inaccessible. s390
1383 * does this. So nothing which can access the page's contents should
1384 * happen after this.
1386 arch_free_page(page, order);
1388 debug_pagealloc_unmap_pages(page, 1 << order);
1393 #ifdef CONFIG_DEBUG_VM
1395 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1396 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1397 * moved from pcp lists to free lists.
1399 static bool free_pcp_prepare(struct page *page, unsigned int order)
1401 return free_pages_prepare(page, order, true, FPI_NONE);
1404 static bool bulkfree_pcp_prepare(struct page *page)
1406 if (debug_pagealloc_enabled_static())
1407 return check_free_page(page);
1413 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1414 * moving from pcp lists to free list in order to reduce overhead. With
1415 * debug_pagealloc enabled, they are checked also immediately when being freed
1418 static bool free_pcp_prepare(struct page *page, unsigned int order)
1420 if (debug_pagealloc_enabled_static())
1421 return free_pages_prepare(page, order, true, FPI_NONE);
1423 return free_pages_prepare(page, order, false, FPI_NONE);
1426 static bool bulkfree_pcp_prepare(struct page *page)
1428 return check_free_page(page);
1430 #endif /* CONFIG_DEBUG_VM */
1433 * Frees a number of pages from the PCP lists
1434 * Assumes all pages on list are in same zone.
1435 * count is the number of pages to free.
1437 static void free_pcppages_bulk(struct zone *zone, int count,
1438 struct per_cpu_pages *pcp,
1442 int max_pindex = NR_PCP_LISTS - 1;
1444 bool isolated_pageblocks;
1448 * Ensure proper count is passed which otherwise would stuck in the
1449 * below while (list_empty(list)) loop.
1451 count = min(pcp->count, count);
1453 /* Ensure requested pindex is drained first. */
1454 pindex = pindex - 1;
1457 * local_lock_irq held so equivalent to spin_lock_irqsave for
1458 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1460 spin_lock(&zone->lock);
1461 isolated_pageblocks = has_isolate_pageblock(zone);
1464 struct list_head *list;
1467 /* Remove pages from lists in a round-robin fashion. */
1469 if (++pindex > max_pindex)
1470 pindex = min_pindex;
1471 list = &pcp->lists[pindex];
1472 if (!list_empty(list))
1475 if (pindex == max_pindex)
1477 if (pindex == min_pindex)
1481 order = pindex_to_order(pindex);
1482 nr_pages = 1 << order;
1483 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1487 page = list_last_entry(list, struct page, lru);
1488 mt = get_pcppage_migratetype(page);
1490 /* must delete to avoid corrupting pcp list */
1491 list_del(&page->lru);
1493 pcp->count -= nr_pages;
1495 if (bulkfree_pcp_prepare(page))
1498 /* MIGRATE_ISOLATE page should not go to pcplists */
1499 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1500 /* Pageblock could have been isolated meanwhile */
1501 if (unlikely(isolated_pageblocks))
1502 mt = get_pageblock_migratetype(page);
1504 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1505 trace_mm_page_pcpu_drain(page, order, mt);
1506 } while (count > 0 && !list_empty(list));
1509 spin_unlock(&zone->lock);
1512 static void free_one_page(struct zone *zone,
1513 struct page *page, unsigned long pfn,
1515 int migratetype, fpi_t fpi_flags)
1517 unsigned long flags;
1519 spin_lock_irqsave(&zone->lock, flags);
1520 if (unlikely(has_isolate_pageblock(zone) ||
1521 is_migrate_isolate(migratetype))) {
1522 migratetype = get_pfnblock_migratetype(page, pfn);
1524 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1525 spin_unlock_irqrestore(&zone->lock, flags);
1528 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1529 unsigned long zone, int nid)
1531 mm_zero_struct_page(page);
1532 set_page_links(page, zone, nid, pfn);
1533 init_page_count(page);
1534 page_mapcount_reset(page);
1535 page_cpupid_reset_last(page);
1536 page_kasan_tag_reset(page);
1538 INIT_LIST_HEAD(&page->lru);
1539 #ifdef WANT_PAGE_VIRTUAL
1540 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1541 if (!is_highmem_idx(zone))
1542 set_page_address(page, __va(pfn << PAGE_SHIFT));
1546 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1547 static void __meminit init_reserved_page(unsigned long pfn)
1552 if (!early_page_uninitialised(pfn))
1555 nid = early_pfn_to_nid(pfn);
1556 pgdat = NODE_DATA(nid);
1558 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1559 struct zone *zone = &pgdat->node_zones[zid];
1561 if (zone_spans_pfn(zone, pfn))
1564 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1567 static inline void init_reserved_page(unsigned long pfn)
1570 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1573 * Initialised pages do not have PageReserved set. This function is
1574 * called for each range allocated by the bootmem allocator and
1575 * marks the pages PageReserved. The remaining valid pages are later
1576 * sent to the buddy page allocator.
1578 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1580 unsigned long start_pfn = PFN_DOWN(start);
1581 unsigned long end_pfn = PFN_UP(end);
1583 for (; start_pfn < end_pfn; start_pfn++) {
1584 if (pfn_valid(start_pfn)) {
1585 struct page *page = pfn_to_page(start_pfn);
1587 init_reserved_page(start_pfn);
1589 /* Avoid false-positive PageTail() */
1590 INIT_LIST_HEAD(&page->lru);
1593 * no need for atomic set_bit because the struct
1594 * page is not visible yet so nobody should
1597 __SetPageReserved(page);
1602 static void __free_pages_ok(struct page *page, unsigned int order,
1605 unsigned long flags;
1607 unsigned long pfn = page_to_pfn(page);
1608 struct zone *zone = page_zone(page);
1610 if (!free_pages_prepare(page, order, true, fpi_flags))
1613 migratetype = get_pfnblock_migratetype(page, pfn);
1615 spin_lock_irqsave(&zone->lock, flags);
1616 if (unlikely(has_isolate_pageblock(zone) ||
1617 is_migrate_isolate(migratetype))) {
1618 migratetype = get_pfnblock_migratetype(page, pfn);
1620 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1621 spin_unlock_irqrestore(&zone->lock, flags);
1623 __count_vm_events(PGFREE, 1 << order);
1626 void __free_pages_core(struct page *page, unsigned int order)
1628 unsigned int nr_pages = 1 << order;
1629 struct page *p = page;
1633 * When initializing the memmap, __init_single_page() sets the refcount
1634 * of all pages to 1 ("allocated"/"not free"). We have to set the
1635 * refcount of all involved pages to 0.
1638 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1640 __ClearPageReserved(p);
1641 set_page_count(p, 0);
1643 __ClearPageReserved(p);
1644 set_page_count(p, 0);
1646 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1649 * Bypass PCP and place fresh pages right to the tail, primarily
1650 * relevant for memory onlining.
1652 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1658 * During memory init memblocks map pfns to nids. The search is expensive and
1659 * this caches recent lookups. The implementation of __early_pfn_to_nid
1660 * treats start/end as pfns.
1662 struct mminit_pfnnid_cache {
1663 unsigned long last_start;
1664 unsigned long last_end;
1668 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1671 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1673 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1674 struct mminit_pfnnid_cache *state)
1676 unsigned long start_pfn, end_pfn;
1679 if (state->last_start <= pfn && pfn < state->last_end)
1680 return state->last_nid;
1682 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1683 if (nid != NUMA_NO_NODE) {
1684 state->last_start = start_pfn;
1685 state->last_end = end_pfn;
1686 state->last_nid = nid;
1692 int __meminit early_pfn_to_nid(unsigned long pfn)
1694 static DEFINE_SPINLOCK(early_pfn_lock);
1697 spin_lock(&early_pfn_lock);
1698 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1700 nid = first_online_node;
1701 spin_unlock(&early_pfn_lock);
1705 #endif /* CONFIG_NUMA */
1707 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1710 if (early_page_uninitialised(pfn))
1712 __free_pages_core(page, order);
1716 * Check that the whole (or subset of) a pageblock given by the interval of
1717 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1718 * with the migration of free compaction scanner.
1720 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1722 * It's possible on some configurations to have a setup like node0 node1 node0
1723 * i.e. it's possible that all pages within a zones range of pages do not
1724 * belong to a single zone. We assume that a border between node0 and node1
1725 * can occur within a single pageblock, but not a node0 node1 node0
1726 * interleaving within a single pageblock. It is therefore sufficient to check
1727 * the first and last page of a pageblock and avoid checking each individual
1728 * page in a pageblock.
1730 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1731 unsigned long end_pfn, struct zone *zone)
1733 struct page *start_page;
1734 struct page *end_page;
1736 /* end_pfn is one past the range we are checking */
1739 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1742 start_page = pfn_to_online_page(start_pfn);
1746 if (page_zone(start_page) != zone)
1749 end_page = pfn_to_page(end_pfn);
1751 /* This gives a shorter code than deriving page_zone(end_page) */
1752 if (page_zone_id(start_page) != page_zone_id(end_page))
1758 void set_zone_contiguous(struct zone *zone)
1760 unsigned long block_start_pfn = zone->zone_start_pfn;
1761 unsigned long block_end_pfn;
1763 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1764 for (; block_start_pfn < zone_end_pfn(zone);
1765 block_start_pfn = block_end_pfn,
1766 block_end_pfn += pageblock_nr_pages) {
1768 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1770 if (!__pageblock_pfn_to_page(block_start_pfn,
1771 block_end_pfn, zone))
1776 /* We confirm that there is no hole */
1777 zone->contiguous = true;
1780 void clear_zone_contiguous(struct zone *zone)
1782 zone->contiguous = false;
1785 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1786 static void __init deferred_free_range(unsigned long pfn,
1787 unsigned long nr_pages)
1795 page = pfn_to_page(pfn);
1797 /* Free a large naturally-aligned chunk if possible */
1798 if (nr_pages == pageblock_nr_pages &&
1799 (pfn & (pageblock_nr_pages - 1)) == 0) {
1800 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1801 __free_pages_core(page, pageblock_order);
1805 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1806 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1807 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1808 __free_pages_core(page, 0);
1812 /* Completion tracking for deferred_init_memmap() threads */
1813 static atomic_t pgdat_init_n_undone __initdata;
1814 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1816 static inline void __init pgdat_init_report_one_done(void)
1818 if (atomic_dec_and_test(&pgdat_init_n_undone))
1819 complete(&pgdat_init_all_done_comp);
1823 * Returns true if page needs to be initialized or freed to buddy allocator.
1825 * First we check if pfn is valid on architectures where it is possible to have
1826 * holes within pageblock_nr_pages. On systems where it is not possible, this
1827 * function is optimized out.
1829 * Then, we check if a current large page is valid by only checking the validity
1832 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1834 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1840 * Free pages to buddy allocator. Try to free aligned pages in
1841 * pageblock_nr_pages sizes.
1843 static void __init deferred_free_pages(unsigned long pfn,
1844 unsigned long end_pfn)
1846 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1847 unsigned long nr_free = 0;
1849 for (; pfn < end_pfn; pfn++) {
1850 if (!deferred_pfn_valid(pfn)) {
1851 deferred_free_range(pfn - nr_free, nr_free);
1853 } else if (!(pfn & nr_pgmask)) {
1854 deferred_free_range(pfn - nr_free, nr_free);
1860 /* Free the last block of pages to allocator */
1861 deferred_free_range(pfn - nr_free, nr_free);
1865 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1866 * by performing it only once every pageblock_nr_pages.
1867 * Return number of pages initialized.
1869 static unsigned long __init deferred_init_pages(struct zone *zone,
1871 unsigned long end_pfn)
1873 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1874 int nid = zone_to_nid(zone);
1875 unsigned long nr_pages = 0;
1876 int zid = zone_idx(zone);
1877 struct page *page = NULL;
1879 for (; pfn < end_pfn; pfn++) {
1880 if (!deferred_pfn_valid(pfn)) {
1883 } else if (!page || !(pfn & nr_pgmask)) {
1884 page = pfn_to_page(pfn);
1888 __init_single_page(page, pfn, zid, nid);
1895 * This function is meant to pre-load the iterator for the zone init.
1896 * Specifically it walks through the ranges until we are caught up to the
1897 * first_init_pfn value and exits there. If we never encounter the value we
1898 * return false indicating there are no valid ranges left.
1901 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1902 unsigned long *spfn, unsigned long *epfn,
1903 unsigned long first_init_pfn)
1908 * Start out by walking through the ranges in this zone that have
1909 * already been initialized. We don't need to do anything with them
1910 * so we just need to flush them out of the system.
1912 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1913 if (*epfn <= first_init_pfn)
1915 if (*spfn < first_init_pfn)
1916 *spfn = first_init_pfn;
1925 * Initialize and free pages. We do it in two loops: first we initialize
1926 * struct page, then free to buddy allocator, because while we are
1927 * freeing pages we can access pages that are ahead (computing buddy
1928 * page in __free_one_page()).
1930 * In order to try and keep some memory in the cache we have the loop
1931 * broken along max page order boundaries. This way we will not cause
1932 * any issues with the buddy page computation.
1934 static unsigned long __init
1935 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1936 unsigned long *end_pfn)
1938 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1939 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1940 unsigned long nr_pages = 0;
1943 /* First we loop through and initialize the page values */
1944 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1947 if (mo_pfn <= *start_pfn)
1950 t = min(mo_pfn, *end_pfn);
1951 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1953 if (mo_pfn < *end_pfn) {
1954 *start_pfn = mo_pfn;
1959 /* Reset values and now loop through freeing pages as needed */
1962 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1968 t = min(mo_pfn, epfn);
1969 deferred_free_pages(spfn, t);
1979 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1982 unsigned long spfn, epfn;
1983 struct zone *zone = arg;
1986 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1989 * Initialize and free pages in MAX_ORDER sized increments so that we
1990 * can avoid introducing any issues with the buddy allocator.
1992 while (spfn < end_pfn) {
1993 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1998 /* An arch may override for more concurrency. */
2000 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2005 /* Initialise remaining memory on a node */
2006 static int __init deferred_init_memmap(void *data)
2008 pg_data_t *pgdat = data;
2009 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2010 unsigned long spfn = 0, epfn = 0;
2011 unsigned long first_init_pfn, flags;
2012 unsigned long start = jiffies;
2014 int zid, max_threads;
2017 /* Bind memory initialisation thread to a local node if possible */
2018 if (!cpumask_empty(cpumask))
2019 set_cpus_allowed_ptr(current, cpumask);
2021 pgdat_resize_lock(pgdat, &flags);
2022 first_init_pfn = pgdat->first_deferred_pfn;
2023 if (first_init_pfn == ULONG_MAX) {
2024 pgdat_resize_unlock(pgdat, &flags);
2025 pgdat_init_report_one_done();
2029 /* Sanity check boundaries */
2030 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2031 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2032 pgdat->first_deferred_pfn = ULONG_MAX;
2035 * Once we unlock here, the zone cannot be grown anymore, thus if an
2036 * interrupt thread must allocate this early in boot, zone must be
2037 * pre-grown prior to start of deferred page initialization.
2039 pgdat_resize_unlock(pgdat, &flags);
2041 /* Only the highest zone is deferred so find it */
2042 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2043 zone = pgdat->node_zones + zid;
2044 if (first_init_pfn < zone_end_pfn(zone))
2048 /* If the zone is empty somebody else may have cleared out the zone */
2049 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2053 max_threads = deferred_page_init_max_threads(cpumask);
2055 while (spfn < epfn) {
2056 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2057 struct padata_mt_job job = {
2058 .thread_fn = deferred_init_memmap_chunk,
2061 .size = epfn_align - spfn,
2062 .align = PAGES_PER_SECTION,
2063 .min_chunk = PAGES_PER_SECTION,
2064 .max_threads = max_threads,
2067 padata_do_multithreaded(&job);
2068 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2072 /* Sanity check that the next zone really is unpopulated */
2073 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2075 pr_info("node %d deferred pages initialised in %ums\n",
2076 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2078 pgdat_init_report_one_done();
2083 * If this zone has deferred pages, try to grow it by initializing enough
2084 * deferred pages to satisfy the allocation specified by order, rounded up to
2085 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2086 * of SECTION_SIZE bytes by initializing struct pages in increments of
2087 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2089 * Return true when zone was grown, otherwise return false. We return true even
2090 * when we grow less than requested, to let the caller decide if there are
2091 * enough pages to satisfy the allocation.
2093 * Note: We use noinline because this function is needed only during boot, and
2094 * it is called from a __ref function _deferred_grow_zone. This way we are
2095 * making sure that it is not inlined into permanent text section.
2097 static noinline bool __init
2098 deferred_grow_zone(struct zone *zone, unsigned int order)
2100 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2101 pg_data_t *pgdat = zone->zone_pgdat;
2102 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2103 unsigned long spfn, epfn, flags;
2104 unsigned long nr_pages = 0;
2107 /* Only the last zone may have deferred pages */
2108 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2111 pgdat_resize_lock(pgdat, &flags);
2114 * If someone grew this zone while we were waiting for spinlock, return
2115 * true, as there might be enough pages already.
2117 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2118 pgdat_resize_unlock(pgdat, &flags);
2122 /* If the zone is empty somebody else may have cleared out the zone */
2123 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2124 first_deferred_pfn)) {
2125 pgdat->first_deferred_pfn = ULONG_MAX;
2126 pgdat_resize_unlock(pgdat, &flags);
2127 /* Retry only once. */
2128 return first_deferred_pfn != ULONG_MAX;
2132 * Initialize and free pages in MAX_ORDER sized increments so
2133 * that we can avoid introducing any issues with the buddy
2136 while (spfn < epfn) {
2137 /* update our first deferred PFN for this section */
2138 first_deferred_pfn = spfn;
2140 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2141 touch_nmi_watchdog();
2143 /* We should only stop along section boundaries */
2144 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2147 /* If our quota has been met we can stop here */
2148 if (nr_pages >= nr_pages_needed)
2152 pgdat->first_deferred_pfn = spfn;
2153 pgdat_resize_unlock(pgdat, &flags);
2155 return nr_pages > 0;
2159 * deferred_grow_zone() is __init, but it is called from
2160 * get_page_from_freelist() during early boot until deferred_pages permanently
2161 * disables this call. This is why we have refdata wrapper to avoid warning,
2162 * and to ensure that the function body gets unloaded.
2165 _deferred_grow_zone(struct zone *zone, unsigned int order)
2167 return deferred_grow_zone(zone, order);
2170 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2172 void __init page_alloc_init_late(void)
2177 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2179 /* There will be num_node_state(N_MEMORY) threads */
2180 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2181 for_each_node_state(nid, N_MEMORY) {
2182 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2185 /* Block until all are initialised */
2186 wait_for_completion(&pgdat_init_all_done_comp);
2189 * We initialized the rest of the deferred pages. Permanently disable
2190 * on-demand struct page initialization.
2192 static_branch_disable(&deferred_pages);
2194 /* Reinit limits that are based on free pages after the kernel is up */
2195 files_maxfiles_init();
2200 /* Discard memblock private memory */
2203 for_each_node_state(nid, N_MEMORY)
2204 shuffle_free_memory(NODE_DATA(nid));
2206 for_each_populated_zone(zone)
2207 set_zone_contiguous(zone);
2211 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2212 void __init init_cma_reserved_pageblock(struct page *page)
2214 unsigned i = pageblock_nr_pages;
2215 struct page *p = page;
2218 __ClearPageReserved(p);
2219 set_page_count(p, 0);
2222 set_pageblock_migratetype(page, MIGRATE_CMA);
2223 set_page_refcounted(page);
2224 __free_pages(page, pageblock_order);
2226 adjust_managed_page_count(page, pageblock_nr_pages);
2227 page_zone(page)->cma_pages += pageblock_nr_pages;
2232 * The order of subdivision here is critical for the IO subsystem.
2233 * Please do not alter this order without good reasons and regression
2234 * testing. Specifically, as large blocks of memory are subdivided,
2235 * the order in which smaller blocks are delivered depends on the order
2236 * they're subdivided in this function. This is the primary factor
2237 * influencing the order in which pages are delivered to the IO
2238 * subsystem according to empirical testing, and this is also justified
2239 * by considering the behavior of a buddy system containing a single
2240 * large block of memory acted on by a series of small allocations.
2241 * This behavior is a critical factor in sglist merging's success.
2245 static inline void expand(struct zone *zone, struct page *page,
2246 int low, int high, int migratetype)
2248 unsigned long size = 1 << high;
2250 while (high > low) {
2253 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2256 * Mark as guard pages (or page), that will allow to
2257 * merge back to allocator when buddy will be freed.
2258 * Corresponding page table entries will not be touched,
2259 * pages will stay not present in virtual address space
2261 if (set_page_guard(zone, &page[size], high, migratetype))
2264 add_to_free_list(&page[size], zone, high, migratetype);
2265 set_buddy_order(&page[size], high);
2269 static void check_new_page_bad(struct page *page)
2271 if (unlikely(page->flags & __PG_HWPOISON)) {
2272 /* Don't complain about hwpoisoned pages */
2273 page_mapcount_reset(page); /* remove PageBuddy */
2278 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2282 * This page is about to be returned from the page allocator
2284 static inline int check_new_page(struct page *page)
2286 if (likely(page_expected_state(page,
2287 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2290 check_new_page_bad(page);
2294 static bool check_new_pages(struct page *page, unsigned int order)
2297 for (i = 0; i < (1 << order); i++) {
2298 struct page *p = page + i;
2300 if (unlikely(check_new_page(p)))
2307 #ifdef CONFIG_DEBUG_VM
2309 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2310 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2311 * also checked when pcp lists are refilled from the free lists.
2313 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2315 if (debug_pagealloc_enabled_static())
2316 return check_new_pages(page, order);
2321 static inline bool check_new_pcp(struct page *page, unsigned int order)
2323 return check_new_pages(page, order);
2327 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2328 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2329 * enabled, they are also checked when being allocated from the pcp lists.
2331 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2333 return check_new_pages(page, order);
2335 static inline bool check_new_pcp(struct page *page, unsigned int order)
2337 if (debug_pagealloc_enabled_static())
2338 return check_new_pages(page, order);
2342 #endif /* CONFIG_DEBUG_VM */
2344 static inline bool should_skip_kasan_unpoison(gfp_t flags, bool init_tags)
2346 /* Don't skip if a software KASAN mode is enabled. */
2347 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
2348 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
2351 /* Skip, if hardware tag-based KASAN is not enabled. */
2352 if (!kasan_hw_tags_enabled())
2356 * With hardware tag-based KASAN enabled, skip if either:
2358 * 1. Memory tags have already been cleared via tag_clear_highpage().
2359 * 2. Skipping has been requested via __GFP_SKIP_KASAN_UNPOISON.
2361 return init_tags || (flags & __GFP_SKIP_KASAN_UNPOISON);
2364 static inline bool should_skip_init(gfp_t flags)
2366 /* Don't skip, if hardware tag-based KASAN is not enabled. */
2367 if (!kasan_hw_tags_enabled())
2370 /* For hardware tag-based KASAN, skip if requested. */
2371 return (flags & __GFP_SKIP_ZERO);
2374 inline void post_alloc_hook(struct page *page, unsigned int order,
2377 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
2378 !should_skip_init(gfp_flags);
2379 bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
2381 set_page_private(page, 0);
2382 set_page_refcounted(page);
2384 arch_alloc_page(page, order);
2385 debug_pagealloc_map_pages(page, 1 << order);
2388 * Page unpoisoning must happen before memory initialization.
2389 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2390 * allocations and the page unpoisoning code will complain.
2392 kernel_unpoison_pages(page, 1 << order);
2395 * As memory initialization might be integrated into KASAN,
2396 * KASAN unpoisoning and memory initializion code must be
2397 * kept together to avoid discrepancies in behavior.
2401 * If memory tags should be zeroed (which happens only when memory
2402 * should be initialized as well).
2407 /* Initialize both memory and tags. */
2408 for (i = 0; i != 1 << order; ++i)
2409 tag_clear_highpage(page + i);
2411 /* Note that memory is already initialized by the loop above. */
2414 if (!should_skip_kasan_unpoison(gfp_flags, init_tags)) {
2415 /* Unpoison shadow memory or set memory tags. */
2416 kasan_unpoison_pages(page, order, init);
2418 /* Note that memory is already initialized by KASAN. */
2419 if (kasan_has_integrated_init())
2422 /* If memory is still not initialized, do it now. */
2424 kernel_init_free_pages(page, 1 << order);
2425 /* Propagate __GFP_SKIP_KASAN_POISON to page flags. */
2426 if (kasan_hw_tags_enabled() && (gfp_flags & __GFP_SKIP_KASAN_POISON))
2427 SetPageSkipKASanPoison(page);
2429 set_page_owner(page, order, gfp_flags);
2430 page_table_check_alloc(page, order);
2433 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2434 unsigned int alloc_flags)
2436 post_alloc_hook(page, order, gfp_flags);
2438 if (order && (gfp_flags & __GFP_COMP))
2439 prep_compound_page(page, order);
2442 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2443 * allocate the page. The expectation is that the caller is taking
2444 * steps that will free more memory. The caller should avoid the page
2445 * being used for !PFMEMALLOC purposes.
2447 if (alloc_flags & ALLOC_NO_WATERMARKS)
2448 set_page_pfmemalloc(page);
2450 clear_page_pfmemalloc(page);
2454 * Go through the free lists for the given migratetype and remove
2455 * the smallest available page from the freelists
2457 static __always_inline
2458 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2461 unsigned int current_order;
2462 struct free_area *area;
2465 /* Find a page of the appropriate size in the preferred list */
2466 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2467 area = &(zone->free_area[current_order]);
2468 page = get_page_from_free_area(area, migratetype);
2471 del_page_from_free_list(page, zone, current_order);
2472 expand(zone, page, order, current_order, migratetype);
2473 set_pcppage_migratetype(page, migratetype);
2474 trace_mm_page_alloc_zone_locked(page, order, migratetype,
2475 pcp_allowed_order(order) &&
2476 migratetype < MIGRATE_PCPTYPES);
2485 * This array describes the order lists are fallen back to when
2486 * the free lists for the desirable migrate type are depleted
2488 * The other migratetypes do not have fallbacks.
2490 static int fallbacks[MIGRATE_TYPES][3] = {
2491 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2492 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2493 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2497 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2500 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2503 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2504 unsigned int order) { return NULL; }
2508 * Move the free pages in a range to the freelist tail of the requested type.
2509 * Note that start_page and end_pages are not aligned on a pageblock
2510 * boundary. If alignment is required, use move_freepages_block()
2512 static int move_freepages(struct zone *zone,
2513 unsigned long start_pfn, unsigned long end_pfn,
2514 int migratetype, int *num_movable)
2519 int pages_moved = 0;
2521 for (pfn = start_pfn; pfn <= end_pfn;) {
2522 page = pfn_to_page(pfn);
2523 if (!PageBuddy(page)) {
2525 * We assume that pages that could be isolated for
2526 * migration are movable. But we don't actually try
2527 * isolating, as that would be expensive.
2530 (PageLRU(page) || __PageMovable(page)))
2536 /* Make sure we are not inadvertently changing nodes */
2537 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2538 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2540 order = buddy_order(page);
2541 move_to_free_list(page, zone, order, migratetype);
2543 pages_moved += 1 << order;
2549 int move_freepages_block(struct zone *zone, struct page *page,
2550 int migratetype, int *num_movable)
2552 unsigned long start_pfn, end_pfn, pfn;
2557 pfn = page_to_pfn(page);
2558 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2559 end_pfn = start_pfn + pageblock_nr_pages - 1;
2561 /* Do not cross zone boundaries */
2562 if (!zone_spans_pfn(zone, start_pfn))
2564 if (!zone_spans_pfn(zone, end_pfn))
2567 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2571 static void change_pageblock_range(struct page *pageblock_page,
2572 int start_order, int migratetype)
2574 int nr_pageblocks = 1 << (start_order - pageblock_order);
2576 while (nr_pageblocks--) {
2577 set_pageblock_migratetype(pageblock_page, migratetype);
2578 pageblock_page += pageblock_nr_pages;
2583 * When we are falling back to another migratetype during allocation, try to
2584 * steal extra free pages from the same pageblocks to satisfy further
2585 * allocations, instead of polluting multiple pageblocks.
2587 * If we are stealing a relatively large buddy page, it is likely there will
2588 * be more free pages in the pageblock, so try to steal them all. For
2589 * reclaimable and unmovable allocations, we steal regardless of page size,
2590 * as fragmentation caused by those allocations polluting movable pageblocks
2591 * is worse than movable allocations stealing from unmovable and reclaimable
2594 static bool can_steal_fallback(unsigned int order, int start_mt)
2597 * Leaving this order check is intended, although there is
2598 * relaxed order check in next check. The reason is that
2599 * we can actually steal whole pageblock if this condition met,
2600 * but, below check doesn't guarantee it and that is just heuristic
2601 * so could be changed anytime.
2603 if (order >= pageblock_order)
2606 if (order >= pageblock_order / 2 ||
2607 start_mt == MIGRATE_RECLAIMABLE ||
2608 start_mt == MIGRATE_UNMOVABLE ||
2609 page_group_by_mobility_disabled)
2615 static inline bool boost_watermark(struct zone *zone)
2617 unsigned long max_boost;
2619 if (!watermark_boost_factor)
2622 * Don't bother in zones that are unlikely to produce results.
2623 * On small machines, including kdump capture kernels running
2624 * in a small area, boosting the watermark can cause an out of
2625 * memory situation immediately.
2627 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2630 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2631 watermark_boost_factor, 10000);
2634 * high watermark may be uninitialised if fragmentation occurs
2635 * very early in boot so do not boost. We do not fall
2636 * through and boost by pageblock_nr_pages as failing
2637 * allocations that early means that reclaim is not going
2638 * to help and it may even be impossible to reclaim the
2639 * boosted watermark resulting in a hang.
2644 max_boost = max(pageblock_nr_pages, max_boost);
2646 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2653 * This function implements actual steal behaviour. If order is large enough,
2654 * we can steal whole pageblock. If not, we first move freepages in this
2655 * pageblock to our migratetype and determine how many already-allocated pages
2656 * are there in the pageblock with a compatible migratetype. If at least half
2657 * of pages are free or compatible, we can change migratetype of the pageblock
2658 * itself, so pages freed in the future will be put on the correct free list.
2660 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2661 unsigned int alloc_flags, int start_type, bool whole_block)
2663 unsigned int current_order = buddy_order(page);
2664 int free_pages, movable_pages, alike_pages;
2667 old_block_type = get_pageblock_migratetype(page);
2670 * This can happen due to races and we want to prevent broken
2671 * highatomic accounting.
2673 if (is_migrate_highatomic(old_block_type))
2676 /* Take ownership for orders >= pageblock_order */
2677 if (current_order >= pageblock_order) {
2678 change_pageblock_range(page, current_order, start_type);
2683 * Boost watermarks to increase reclaim pressure to reduce the
2684 * likelihood of future fallbacks. Wake kswapd now as the node
2685 * may be balanced overall and kswapd will not wake naturally.
2687 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2688 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2690 /* We are not allowed to try stealing from the whole block */
2694 free_pages = move_freepages_block(zone, page, start_type,
2697 * Determine how many pages are compatible with our allocation.
2698 * For movable allocation, it's the number of movable pages which
2699 * we just obtained. For other types it's a bit more tricky.
2701 if (start_type == MIGRATE_MOVABLE) {
2702 alike_pages = movable_pages;
2705 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2706 * to MOVABLE pageblock, consider all non-movable pages as
2707 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2708 * vice versa, be conservative since we can't distinguish the
2709 * exact migratetype of non-movable pages.
2711 if (old_block_type == MIGRATE_MOVABLE)
2712 alike_pages = pageblock_nr_pages
2713 - (free_pages + movable_pages);
2718 /* moving whole block can fail due to zone boundary conditions */
2723 * If a sufficient number of pages in the block are either free or of
2724 * comparable migratability as our allocation, claim the whole block.
2726 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2727 page_group_by_mobility_disabled)
2728 set_pageblock_migratetype(page, start_type);
2733 move_to_free_list(page, zone, current_order, start_type);
2737 * Check whether there is a suitable fallback freepage with requested order.
2738 * If only_stealable is true, this function returns fallback_mt only if
2739 * we can steal other freepages all together. This would help to reduce
2740 * fragmentation due to mixed migratetype pages in one pageblock.
2742 int find_suitable_fallback(struct free_area *area, unsigned int order,
2743 int migratetype, bool only_stealable, bool *can_steal)
2748 if (area->nr_free == 0)
2753 fallback_mt = fallbacks[migratetype][i];
2754 if (fallback_mt == MIGRATE_TYPES)
2757 if (free_area_empty(area, fallback_mt))
2760 if (can_steal_fallback(order, migratetype))
2763 if (!only_stealable)
2774 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2775 * there are no empty page blocks that contain a page with a suitable order
2777 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2778 unsigned int alloc_order)
2781 unsigned long max_managed, flags;
2784 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2785 * Check is race-prone but harmless.
2787 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2788 if (zone->nr_reserved_highatomic >= max_managed)
2791 spin_lock_irqsave(&zone->lock, flags);
2793 /* Recheck the nr_reserved_highatomic limit under the lock */
2794 if (zone->nr_reserved_highatomic >= max_managed)
2798 mt = get_pageblock_migratetype(page);
2799 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2800 if (migratetype_is_mergeable(mt)) {
2801 zone->nr_reserved_highatomic += pageblock_nr_pages;
2802 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2803 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2807 spin_unlock_irqrestore(&zone->lock, flags);
2811 * Used when an allocation is about to fail under memory pressure. This
2812 * potentially hurts the reliability of high-order allocations when under
2813 * intense memory pressure but failed atomic allocations should be easier
2814 * to recover from than an OOM.
2816 * If @force is true, try to unreserve a pageblock even though highatomic
2817 * pageblock is exhausted.
2819 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2822 struct zonelist *zonelist = ac->zonelist;
2823 unsigned long flags;
2830 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2833 * Preserve at least one pageblock unless memory pressure
2836 if (!force && zone->nr_reserved_highatomic <=
2840 spin_lock_irqsave(&zone->lock, flags);
2841 for (order = 0; order < MAX_ORDER; order++) {
2842 struct free_area *area = &(zone->free_area[order]);
2844 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2849 * In page freeing path, migratetype change is racy so
2850 * we can counter several free pages in a pageblock
2851 * in this loop although we changed the pageblock type
2852 * from highatomic to ac->migratetype. So we should
2853 * adjust the count once.
2855 if (is_migrate_highatomic_page(page)) {
2857 * It should never happen but changes to
2858 * locking could inadvertently allow a per-cpu
2859 * drain to add pages to MIGRATE_HIGHATOMIC
2860 * while unreserving so be safe and watch for
2863 zone->nr_reserved_highatomic -= min(
2865 zone->nr_reserved_highatomic);
2869 * Convert to ac->migratetype and avoid the normal
2870 * pageblock stealing heuristics. Minimally, the caller
2871 * is doing the work and needs the pages. More
2872 * importantly, if the block was always converted to
2873 * MIGRATE_UNMOVABLE or another type then the number
2874 * of pageblocks that cannot be completely freed
2877 set_pageblock_migratetype(page, ac->migratetype);
2878 ret = move_freepages_block(zone, page, ac->migratetype,
2881 spin_unlock_irqrestore(&zone->lock, flags);
2885 spin_unlock_irqrestore(&zone->lock, flags);
2892 * Try finding a free buddy page on the fallback list and put it on the free
2893 * list of requested migratetype, possibly along with other pages from the same
2894 * block, depending on fragmentation avoidance heuristics. Returns true if
2895 * fallback was found so that __rmqueue_smallest() can grab it.
2897 * The use of signed ints for order and current_order is a deliberate
2898 * deviation from the rest of this file, to make the for loop
2899 * condition simpler.
2901 static __always_inline bool
2902 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2903 unsigned int alloc_flags)
2905 struct free_area *area;
2907 int min_order = order;
2913 * Do not steal pages from freelists belonging to other pageblocks
2914 * i.e. orders < pageblock_order. If there are no local zones free,
2915 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2917 if (alloc_flags & ALLOC_NOFRAGMENT)
2918 min_order = pageblock_order;
2921 * Find the largest available free page in the other list. This roughly
2922 * approximates finding the pageblock with the most free pages, which
2923 * would be too costly to do exactly.
2925 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2927 area = &(zone->free_area[current_order]);
2928 fallback_mt = find_suitable_fallback(area, current_order,
2929 start_migratetype, false, &can_steal);
2930 if (fallback_mt == -1)
2934 * We cannot steal all free pages from the pageblock and the
2935 * requested migratetype is movable. In that case it's better to
2936 * steal and split the smallest available page instead of the
2937 * largest available page, because even if the next movable
2938 * allocation falls back into a different pageblock than this
2939 * one, it won't cause permanent fragmentation.
2941 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2942 && current_order > order)
2951 for (current_order = order; current_order < MAX_ORDER;
2953 area = &(zone->free_area[current_order]);
2954 fallback_mt = find_suitable_fallback(area, current_order,
2955 start_migratetype, false, &can_steal);
2956 if (fallback_mt != -1)
2961 * This should not happen - we already found a suitable fallback
2962 * when looking for the largest page.
2964 VM_BUG_ON(current_order == MAX_ORDER);
2967 page = get_page_from_free_area(area, fallback_mt);
2969 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2972 trace_mm_page_alloc_extfrag(page, order, current_order,
2973 start_migratetype, fallback_mt);
2980 * Do the hard work of removing an element from the buddy allocator.
2981 * Call me with the zone->lock already held.
2983 static __always_inline struct page *
2984 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2985 unsigned int alloc_flags)
2989 if (IS_ENABLED(CONFIG_CMA)) {
2991 * Balance movable allocations between regular and CMA areas by
2992 * allocating from CMA when over half of the zone's free memory
2993 * is in the CMA area.
2995 if (alloc_flags & ALLOC_CMA &&
2996 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2997 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2998 page = __rmqueue_cma_fallback(zone, order);
3004 page = __rmqueue_smallest(zone, order, migratetype);
3005 if (unlikely(!page)) {
3006 if (alloc_flags & ALLOC_CMA)
3007 page = __rmqueue_cma_fallback(zone, order);
3009 if (!page && __rmqueue_fallback(zone, order, migratetype,
3017 * Obtain a specified number of elements from the buddy allocator, all under
3018 * a single hold of the lock, for efficiency. Add them to the supplied list.
3019 * Returns the number of new pages which were placed at *list.
3021 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3022 unsigned long count, struct list_head *list,
3023 int migratetype, unsigned int alloc_flags)
3025 int i, allocated = 0;
3028 * local_lock_irq held so equivalent to spin_lock_irqsave for
3029 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3031 spin_lock(&zone->lock);
3032 for (i = 0; i < count; ++i) {
3033 struct page *page = __rmqueue(zone, order, migratetype,
3035 if (unlikely(page == NULL))
3038 if (unlikely(check_pcp_refill(page, order)))
3042 * Split buddy pages returned by expand() are received here in
3043 * physical page order. The page is added to the tail of
3044 * caller's list. From the callers perspective, the linked list
3045 * is ordered by page number under some conditions. This is
3046 * useful for IO devices that can forward direction from the
3047 * head, thus also in the physical page order. This is useful
3048 * for IO devices that can merge IO requests if the physical
3049 * pages are ordered properly.
3051 list_add_tail(&page->lru, list);
3053 if (is_migrate_cma(get_pcppage_migratetype(page)))
3054 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3059 * i pages were removed from the buddy list even if some leak due
3060 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3061 * on i. Do not confuse with 'allocated' which is the number of
3062 * pages added to the pcp list.
3064 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3065 spin_unlock(&zone->lock);
3071 * Called from the vmstat counter updater to drain pagesets of this
3072 * currently executing processor on remote nodes after they have
3075 * Note that this function must be called with the thread pinned to
3076 * a single processor.
3078 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3080 unsigned long flags;
3081 int to_drain, batch;
3083 local_lock_irqsave(&pagesets.lock, flags);
3084 batch = READ_ONCE(pcp->batch);
3085 to_drain = min(pcp->count, batch);
3087 free_pcppages_bulk(zone, to_drain, pcp, 0);
3088 local_unlock_irqrestore(&pagesets.lock, flags);
3093 * Drain pcplists of the indicated processor and zone.
3095 * The processor must either be the current processor and the
3096 * thread pinned to the current processor or a processor that
3099 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3101 unsigned long flags;
3102 struct per_cpu_pages *pcp;
3104 local_lock_irqsave(&pagesets.lock, flags);
3106 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3108 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3110 local_unlock_irqrestore(&pagesets.lock, flags);
3114 * Drain pcplists of all zones on the indicated processor.
3116 * The processor must either be the current processor and the
3117 * thread pinned to the current processor or a processor that
3120 static void drain_pages(unsigned int cpu)
3124 for_each_populated_zone(zone) {
3125 drain_pages_zone(cpu, zone);
3130 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3132 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3133 * the single zone's pages.
3135 void drain_local_pages(struct zone *zone)
3137 int cpu = smp_processor_id();
3140 drain_pages_zone(cpu, zone);
3145 static void drain_local_pages_wq(struct work_struct *work)
3147 struct pcpu_drain *drain;
3149 drain = container_of(work, struct pcpu_drain, work);
3152 * drain_all_pages doesn't use proper cpu hotplug protection so
3153 * we can race with cpu offline when the WQ can move this from
3154 * a cpu pinned worker to an unbound one. We can operate on a different
3155 * cpu which is alright but we also have to make sure to not move to
3159 drain_local_pages(drain->zone);
3164 * The implementation of drain_all_pages(), exposing an extra parameter to
3165 * drain on all cpus.
3167 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3168 * not empty. The check for non-emptiness can however race with a free to
3169 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3170 * that need the guarantee that every CPU has drained can disable the
3171 * optimizing racy check.
3173 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3178 * Allocate in the BSS so we won't require allocation in
3179 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3181 static cpumask_t cpus_with_pcps;
3184 * Make sure nobody triggers this path before mm_percpu_wq is fully
3187 if (WARN_ON_ONCE(!mm_percpu_wq))
3191 * Do not drain if one is already in progress unless it's specific to
3192 * a zone. Such callers are primarily CMA and memory hotplug and need
3193 * the drain to be complete when the call returns.
3195 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3198 mutex_lock(&pcpu_drain_mutex);
3202 * We don't care about racing with CPU hotplug event
3203 * as offline notification will cause the notified
3204 * cpu to drain that CPU pcps and on_each_cpu_mask
3205 * disables preemption as part of its processing
3207 for_each_online_cpu(cpu) {
3208 struct per_cpu_pages *pcp;
3210 bool has_pcps = false;
3212 if (force_all_cpus) {
3214 * The pcp.count check is racy, some callers need a
3215 * guarantee that no cpu is missed.
3219 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3223 for_each_populated_zone(z) {
3224 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3233 cpumask_set_cpu(cpu, &cpus_with_pcps);
3235 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3238 for_each_cpu(cpu, &cpus_with_pcps) {
3239 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3242 INIT_WORK(&drain->work, drain_local_pages_wq);
3243 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3245 for_each_cpu(cpu, &cpus_with_pcps)
3246 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3248 mutex_unlock(&pcpu_drain_mutex);
3252 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3254 * When zone parameter is non-NULL, spill just the single zone's pages.
3256 * Note that this can be extremely slow as the draining happens in a workqueue.
3258 void drain_all_pages(struct zone *zone)
3260 __drain_all_pages(zone, false);
3263 #ifdef CONFIG_HIBERNATION
3266 * Touch the watchdog for every WD_PAGE_COUNT pages.
3268 #define WD_PAGE_COUNT (128*1024)
3270 void mark_free_pages(struct zone *zone)
3272 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3273 unsigned long flags;
3274 unsigned int order, t;
3277 if (zone_is_empty(zone))
3280 spin_lock_irqsave(&zone->lock, flags);
3282 max_zone_pfn = zone_end_pfn(zone);
3283 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3284 if (pfn_valid(pfn)) {
3285 page = pfn_to_page(pfn);
3287 if (!--page_count) {
3288 touch_nmi_watchdog();
3289 page_count = WD_PAGE_COUNT;
3292 if (page_zone(page) != zone)
3295 if (!swsusp_page_is_forbidden(page))
3296 swsusp_unset_page_free(page);
3299 for_each_migratetype_order(order, t) {
3300 list_for_each_entry(page,
3301 &zone->free_area[order].free_list[t], lru) {
3304 pfn = page_to_pfn(page);
3305 for (i = 0; i < (1UL << order); i++) {
3306 if (!--page_count) {
3307 touch_nmi_watchdog();
3308 page_count = WD_PAGE_COUNT;
3310 swsusp_set_page_free(pfn_to_page(pfn + i));
3314 spin_unlock_irqrestore(&zone->lock, flags);
3316 #endif /* CONFIG_PM */
3318 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3323 if (!free_pcp_prepare(page, order))
3326 migratetype = get_pfnblock_migratetype(page, pfn);
3327 set_pcppage_migratetype(page, migratetype);
3331 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3334 int min_nr_free, max_nr_free;
3336 /* Free everything if batch freeing high-order pages. */
3337 if (unlikely(free_high))
3340 /* Check for PCP disabled or boot pageset */
3341 if (unlikely(high < batch))
3344 /* Leave at least pcp->batch pages on the list */
3345 min_nr_free = batch;
3346 max_nr_free = high - batch;
3349 * Double the number of pages freed each time there is subsequent
3350 * freeing of pages without any allocation.
3352 batch <<= pcp->free_factor;
3353 if (batch < max_nr_free)
3355 batch = clamp(batch, min_nr_free, max_nr_free);
3360 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3363 int high = READ_ONCE(pcp->high);
3365 if (unlikely(!high || free_high))
3368 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3372 * If reclaim is active, limit the number of pages that can be
3373 * stored on pcp lists
3375 return min(READ_ONCE(pcp->batch) << 2, high);
3378 static void free_unref_page_commit(struct page *page, int migratetype,
3381 struct zone *zone = page_zone(page);
3382 struct per_cpu_pages *pcp;
3387 __count_vm_event(PGFREE);
3388 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3389 pindex = order_to_pindex(migratetype, order);
3390 list_add(&page->lru, &pcp->lists[pindex]);
3391 pcp->count += 1 << order;
3394 * As high-order pages other than THP's stored on PCP can contribute
3395 * to fragmentation, limit the number stored when PCP is heavily
3396 * freeing without allocation. The remainder after bulk freeing
3397 * stops will be drained from vmstat refresh context.
3399 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3401 high = nr_pcp_high(pcp, zone, free_high);
3402 if (pcp->count >= high) {
3403 int batch = READ_ONCE(pcp->batch);
3405 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3412 void free_unref_page(struct page *page, unsigned int order)
3414 unsigned long flags;
3415 unsigned long pfn = page_to_pfn(page);
3418 if (!free_unref_page_prepare(page, pfn, order))
3422 * We only track unmovable, reclaimable and movable on pcp lists.
3423 * Place ISOLATE pages on the isolated list because they are being
3424 * offlined but treat HIGHATOMIC as movable pages so we can get those
3425 * areas back if necessary. Otherwise, we may have to free
3426 * excessively into the page allocator
3428 migratetype = get_pcppage_migratetype(page);
3429 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3430 if (unlikely(is_migrate_isolate(migratetype))) {
3431 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3434 migratetype = MIGRATE_MOVABLE;
3437 local_lock_irqsave(&pagesets.lock, flags);
3438 free_unref_page_commit(page, migratetype, order);
3439 local_unlock_irqrestore(&pagesets.lock, flags);
3443 * Free a list of 0-order pages
3445 void free_unref_page_list(struct list_head *list)
3447 struct page *page, *next;
3448 unsigned long flags;
3449 int batch_count = 0;
3452 /* Prepare pages for freeing */
3453 list_for_each_entry_safe(page, next, list, lru) {
3454 unsigned long pfn = page_to_pfn(page);
3455 if (!free_unref_page_prepare(page, pfn, 0)) {
3456 list_del(&page->lru);
3461 * Free isolated pages directly to the allocator, see
3462 * comment in free_unref_page.
3464 migratetype = get_pcppage_migratetype(page);
3465 if (unlikely(is_migrate_isolate(migratetype))) {
3466 list_del(&page->lru);
3467 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3472 local_lock_irqsave(&pagesets.lock, flags);
3473 list_for_each_entry_safe(page, next, list, lru) {
3475 * Non-isolated types over MIGRATE_PCPTYPES get added
3476 * to the MIGRATE_MOVABLE pcp list.
3478 migratetype = get_pcppage_migratetype(page);
3479 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3480 migratetype = MIGRATE_MOVABLE;
3482 trace_mm_page_free_batched(page);
3483 free_unref_page_commit(page, migratetype, 0);
3486 * Guard against excessive IRQ disabled times when we get
3487 * a large list of pages to free.
3489 if (++batch_count == SWAP_CLUSTER_MAX) {
3490 local_unlock_irqrestore(&pagesets.lock, flags);
3492 local_lock_irqsave(&pagesets.lock, flags);
3495 local_unlock_irqrestore(&pagesets.lock, flags);
3499 * split_page takes a non-compound higher-order page, and splits it into
3500 * n (1<<order) sub-pages: page[0..n]
3501 * Each sub-page must be freed individually.
3503 * Note: this is probably too low level an operation for use in drivers.
3504 * Please consult with lkml before using this in your driver.
3506 void split_page(struct page *page, unsigned int order)
3510 VM_BUG_ON_PAGE(PageCompound(page), page);
3511 VM_BUG_ON_PAGE(!page_count(page), page);
3513 for (i = 1; i < (1 << order); i++)
3514 set_page_refcounted(page + i);
3515 split_page_owner(page, 1 << order);
3516 split_page_memcg(page, 1 << order);
3518 EXPORT_SYMBOL_GPL(split_page);
3520 int __isolate_free_page(struct page *page, unsigned int order)
3522 unsigned long watermark;
3526 BUG_ON(!PageBuddy(page));
3528 zone = page_zone(page);
3529 mt = get_pageblock_migratetype(page);
3531 if (!is_migrate_isolate(mt)) {
3533 * Obey watermarks as if the page was being allocated. We can
3534 * emulate a high-order watermark check with a raised order-0
3535 * watermark, because we already know our high-order page
3538 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3539 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3542 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3545 /* Remove page from free list */
3547 del_page_from_free_list(page, zone, order);
3550 * Set the pageblock if the isolated page is at least half of a
3553 if (order >= pageblock_order - 1) {
3554 struct page *endpage = page + (1 << order) - 1;
3555 for (; page < endpage; page += pageblock_nr_pages) {
3556 int mt = get_pageblock_migratetype(page);
3558 * Only change normal pageblocks (i.e., they can merge
3561 if (migratetype_is_mergeable(mt))
3562 set_pageblock_migratetype(page,
3568 return 1UL << order;
3572 * __putback_isolated_page - Return a now-isolated page back where we got it
3573 * @page: Page that was isolated
3574 * @order: Order of the isolated page
3575 * @mt: The page's pageblock's migratetype
3577 * This function is meant to return a page pulled from the free lists via
3578 * __isolate_free_page back to the free lists they were pulled from.
3580 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3582 struct zone *zone = page_zone(page);
3584 /* zone lock should be held when this function is called */
3585 lockdep_assert_held(&zone->lock);
3587 /* Return isolated page to tail of freelist. */
3588 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3589 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3593 * Update NUMA hit/miss statistics
3595 * Must be called with interrupts disabled.
3597 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3601 enum numa_stat_item local_stat = NUMA_LOCAL;
3603 /* skip numa counters update if numa stats is disabled */
3604 if (!static_branch_likely(&vm_numa_stat_key))
3607 if (zone_to_nid(z) != numa_node_id())
3608 local_stat = NUMA_OTHER;
3610 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3611 __count_numa_events(z, NUMA_HIT, nr_account);
3613 __count_numa_events(z, NUMA_MISS, nr_account);
3614 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3616 __count_numa_events(z, local_stat, nr_account);
3620 /* Remove page from the per-cpu list, caller must protect the list */
3622 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3624 unsigned int alloc_flags,
3625 struct per_cpu_pages *pcp,
3626 struct list_head *list)
3631 if (list_empty(list)) {
3632 int batch = READ_ONCE(pcp->batch);
3636 * Scale batch relative to order if batch implies
3637 * free pages can be stored on the PCP. Batch can
3638 * be 1 for small zones or for boot pagesets which
3639 * should never store free pages as the pages may
3640 * belong to arbitrary zones.
3643 batch = max(batch >> order, 2);
3644 alloced = rmqueue_bulk(zone, order,
3646 migratetype, alloc_flags);
3648 pcp->count += alloced << order;
3649 if (unlikely(list_empty(list)))
3653 page = list_first_entry(list, struct page, lru);
3654 list_del(&page->lru);
3655 pcp->count -= 1 << order;
3656 } while (check_new_pcp(page, order));
3661 /* Lock and remove page from the per-cpu list */
3662 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3663 struct zone *zone, unsigned int order,
3664 gfp_t gfp_flags, int migratetype,
3665 unsigned int alloc_flags)
3667 struct per_cpu_pages *pcp;
3668 struct list_head *list;
3670 unsigned long flags;
3672 local_lock_irqsave(&pagesets.lock, flags);
3675 * On allocation, reduce the number of pages that are batch freed.
3676 * See nr_pcp_free() where free_factor is increased for subsequent
3679 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3680 pcp->free_factor >>= 1;
3681 list = &pcp->lists[order_to_pindex(migratetype, order)];
3682 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3683 local_unlock_irqrestore(&pagesets.lock, flags);
3685 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3686 zone_statistics(preferred_zone, zone, 1);
3692 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3695 struct page *rmqueue(struct zone *preferred_zone,
3696 struct zone *zone, unsigned int order,
3697 gfp_t gfp_flags, unsigned int alloc_flags,
3700 unsigned long flags;
3703 if (likely(pcp_allowed_order(order))) {
3705 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3706 * we need to skip it when CMA area isn't allowed.
3708 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3709 migratetype != MIGRATE_MOVABLE) {
3710 page = rmqueue_pcplist(preferred_zone, zone, order,
3711 gfp_flags, migratetype, alloc_flags);
3717 * We most definitely don't want callers attempting to
3718 * allocate greater than order-1 page units with __GFP_NOFAIL.
3720 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3724 spin_lock_irqsave(&zone->lock, flags);
3726 * order-0 request can reach here when the pcplist is skipped
3727 * due to non-CMA allocation context. HIGHATOMIC area is
3728 * reserved for high-order atomic allocation, so order-0
3729 * request should skip it.
3731 if (order > 0 && alloc_flags & ALLOC_HARDER)
3732 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3734 page = __rmqueue(zone, order, migratetype, alloc_flags);
3738 __mod_zone_freepage_state(zone, -(1 << order),
3739 get_pcppage_migratetype(page));
3740 spin_unlock_irqrestore(&zone->lock, flags);
3741 } while (check_new_pages(page, order));
3743 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3744 zone_statistics(preferred_zone, zone, 1);
3747 /* Separate test+clear to avoid unnecessary atomics */
3748 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3749 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3750 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3753 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3757 spin_unlock_irqrestore(&zone->lock, flags);
3761 #ifdef CONFIG_FAIL_PAGE_ALLOC
3764 struct fault_attr attr;
3766 bool ignore_gfp_highmem;
3767 bool ignore_gfp_reclaim;
3769 } fail_page_alloc = {
3770 .attr = FAULT_ATTR_INITIALIZER,
3771 .ignore_gfp_reclaim = true,
3772 .ignore_gfp_highmem = true,
3776 static int __init setup_fail_page_alloc(char *str)
3778 return setup_fault_attr(&fail_page_alloc.attr, str);
3780 __setup("fail_page_alloc=", setup_fail_page_alloc);
3782 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3784 if (order < fail_page_alloc.min_order)
3786 if (gfp_mask & __GFP_NOFAIL)
3788 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3790 if (fail_page_alloc.ignore_gfp_reclaim &&
3791 (gfp_mask & __GFP_DIRECT_RECLAIM))
3794 if (gfp_mask & __GFP_NOWARN)
3795 fail_page_alloc.attr.no_warn = true;
3797 return should_fail(&fail_page_alloc.attr, 1 << order);
3800 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3802 static int __init fail_page_alloc_debugfs(void)
3804 umode_t mode = S_IFREG | 0600;
3807 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3808 &fail_page_alloc.attr);
3810 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3811 &fail_page_alloc.ignore_gfp_reclaim);
3812 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3813 &fail_page_alloc.ignore_gfp_highmem);
3814 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3819 late_initcall(fail_page_alloc_debugfs);
3821 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3823 #else /* CONFIG_FAIL_PAGE_ALLOC */
3825 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3830 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3832 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3834 return __should_fail_alloc_page(gfp_mask, order);
3836 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3838 static inline long __zone_watermark_unusable_free(struct zone *z,
3839 unsigned int order, unsigned int alloc_flags)
3841 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3842 long unusable_free = (1 << order) - 1;
3845 * If the caller does not have rights to ALLOC_HARDER then subtract
3846 * the high-atomic reserves. This will over-estimate the size of the
3847 * atomic reserve but it avoids a search.
3849 if (likely(!alloc_harder))
3850 unusable_free += z->nr_reserved_highatomic;
3853 /* If allocation can't use CMA areas don't use free CMA pages */
3854 if (!(alloc_flags & ALLOC_CMA))
3855 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3858 return unusable_free;
3862 * Return true if free base pages are above 'mark'. For high-order checks it
3863 * will return true of the order-0 watermark is reached and there is at least
3864 * one free page of a suitable size. Checking now avoids taking the zone lock
3865 * to check in the allocation paths if no pages are free.
3867 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3868 int highest_zoneidx, unsigned int alloc_flags,
3873 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3875 /* free_pages may go negative - that's OK */
3876 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3878 if (alloc_flags & ALLOC_HIGH)
3881 if (unlikely(alloc_harder)) {
3883 * OOM victims can try even harder than normal ALLOC_HARDER
3884 * users on the grounds that it's definitely going to be in
3885 * the exit path shortly and free memory. Any allocation it
3886 * makes during the free path will be small and short-lived.
3888 if (alloc_flags & ALLOC_OOM)
3895 * Check watermarks for an order-0 allocation request. If these
3896 * are not met, then a high-order request also cannot go ahead
3897 * even if a suitable page happened to be free.
3899 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3902 /* If this is an order-0 request then the watermark is fine */
3906 /* For a high-order request, check at least one suitable page is free */
3907 for (o = order; o < MAX_ORDER; o++) {
3908 struct free_area *area = &z->free_area[o];
3914 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3915 if (!free_area_empty(area, mt))
3920 if ((alloc_flags & ALLOC_CMA) &&
3921 !free_area_empty(area, MIGRATE_CMA)) {
3925 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3931 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3932 int highest_zoneidx, unsigned int alloc_flags)
3934 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3935 zone_page_state(z, NR_FREE_PAGES));
3938 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3939 unsigned long mark, int highest_zoneidx,
3940 unsigned int alloc_flags, gfp_t gfp_mask)
3944 free_pages = zone_page_state(z, NR_FREE_PAGES);
3947 * Fast check for order-0 only. If this fails then the reserves
3948 * need to be calculated.
3953 fast_free = free_pages;
3954 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3955 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3959 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3963 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3964 * when checking the min watermark. The min watermark is the
3965 * point where boosting is ignored so that kswapd is woken up
3966 * when below the low watermark.
3968 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3969 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3970 mark = z->_watermark[WMARK_MIN];
3971 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3972 alloc_flags, free_pages);
3978 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3979 unsigned long mark, int highest_zoneidx)
3981 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3983 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3984 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3986 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3991 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3993 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3995 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3996 node_reclaim_distance;
3998 #else /* CONFIG_NUMA */
3999 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4003 #endif /* CONFIG_NUMA */
4006 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4007 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4008 * premature use of a lower zone may cause lowmem pressure problems that
4009 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4010 * probably too small. It only makes sense to spread allocations to avoid
4011 * fragmentation between the Normal and DMA32 zones.
4013 static inline unsigned int
4014 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4016 unsigned int alloc_flags;
4019 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4022 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4024 #ifdef CONFIG_ZONE_DMA32
4028 if (zone_idx(zone) != ZONE_NORMAL)
4032 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4033 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4034 * on UMA that if Normal is populated then so is DMA32.
4036 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4037 if (nr_online_nodes > 1 && !populated_zone(--zone))
4040 alloc_flags |= ALLOC_NOFRAGMENT;
4041 #endif /* CONFIG_ZONE_DMA32 */
4045 /* Must be called after current_gfp_context() which can change gfp_mask */
4046 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4047 unsigned int alloc_flags)
4050 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4051 alloc_flags |= ALLOC_CMA;
4057 * get_page_from_freelist goes through the zonelist trying to allocate
4060 static struct page *
4061 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4062 const struct alloc_context *ac)
4066 struct pglist_data *last_pgdat = NULL;
4067 bool last_pgdat_dirty_ok = false;
4072 * Scan zonelist, looking for a zone with enough free.
4073 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4075 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4076 z = ac->preferred_zoneref;
4077 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4082 if (cpusets_enabled() &&
4083 (alloc_flags & ALLOC_CPUSET) &&
4084 !__cpuset_zone_allowed(zone, gfp_mask))
4087 * When allocating a page cache page for writing, we
4088 * want to get it from a node that is within its dirty
4089 * limit, such that no single node holds more than its
4090 * proportional share of globally allowed dirty pages.
4091 * The dirty limits take into account the node's
4092 * lowmem reserves and high watermark so that kswapd
4093 * should be able to balance it without having to
4094 * write pages from its LRU list.
4096 * XXX: For now, allow allocations to potentially
4097 * exceed the per-node dirty limit in the slowpath
4098 * (spread_dirty_pages unset) before going into reclaim,
4099 * which is important when on a NUMA setup the allowed
4100 * nodes are together not big enough to reach the
4101 * global limit. The proper fix for these situations
4102 * will require awareness of nodes in the
4103 * dirty-throttling and the flusher threads.
4105 if (ac->spread_dirty_pages) {
4106 if (last_pgdat != zone->zone_pgdat) {
4107 last_pgdat = zone->zone_pgdat;
4108 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
4111 if (!last_pgdat_dirty_ok)
4115 if (no_fallback && nr_online_nodes > 1 &&
4116 zone != ac->preferred_zoneref->zone) {
4120 * If moving to a remote node, retry but allow
4121 * fragmenting fallbacks. Locality is more important
4122 * than fragmentation avoidance.
4124 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4125 if (zone_to_nid(zone) != local_nid) {
4126 alloc_flags &= ~ALLOC_NOFRAGMENT;
4131 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4132 if (!zone_watermark_fast(zone, order, mark,
4133 ac->highest_zoneidx, alloc_flags,
4137 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4139 * Watermark failed for this zone, but see if we can
4140 * grow this zone if it contains deferred pages.
4142 if (static_branch_unlikely(&deferred_pages)) {
4143 if (_deferred_grow_zone(zone, order))
4147 /* Checked here to keep the fast path fast */
4148 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4149 if (alloc_flags & ALLOC_NO_WATERMARKS)
4152 if (!node_reclaim_enabled() ||
4153 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4156 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4158 case NODE_RECLAIM_NOSCAN:
4161 case NODE_RECLAIM_FULL:
4162 /* scanned but unreclaimable */
4165 /* did we reclaim enough */
4166 if (zone_watermark_ok(zone, order, mark,
4167 ac->highest_zoneidx, alloc_flags))
4175 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4176 gfp_mask, alloc_flags, ac->migratetype);
4178 prep_new_page(page, order, gfp_mask, alloc_flags);
4181 * If this is a high-order atomic allocation then check
4182 * if the pageblock should be reserved for the future
4184 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4185 reserve_highatomic_pageblock(page, zone, order);
4189 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4190 /* Try again if zone has deferred pages */
4191 if (static_branch_unlikely(&deferred_pages)) {
4192 if (_deferred_grow_zone(zone, order))
4200 * It's possible on a UMA machine to get through all zones that are
4201 * fragmented. If avoiding fragmentation, reset and try again.
4204 alloc_flags &= ~ALLOC_NOFRAGMENT;
4211 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4213 unsigned int filter = SHOW_MEM_FILTER_NODES;
4216 * This documents exceptions given to allocations in certain
4217 * contexts that are allowed to allocate outside current's set
4220 if (!(gfp_mask & __GFP_NOMEMALLOC))
4221 if (tsk_is_oom_victim(current) ||
4222 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4223 filter &= ~SHOW_MEM_FILTER_NODES;
4224 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4225 filter &= ~SHOW_MEM_FILTER_NODES;
4227 show_mem(filter, nodemask);
4230 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4232 struct va_format vaf;
4234 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4236 if ((gfp_mask & __GFP_NOWARN) ||
4237 !__ratelimit(&nopage_rs) ||
4238 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4241 va_start(args, fmt);
4244 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4245 current->comm, &vaf, gfp_mask, &gfp_mask,
4246 nodemask_pr_args(nodemask));
4249 cpuset_print_current_mems_allowed();
4252 warn_alloc_show_mem(gfp_mask, nodemask);
4255 static inline struct page *
4256 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4257 unsigned int alloc_flags,
4258 const struct alloc_context *ac)
4262 page = get_page_from_freelist(gfp_mask, order,
4263 alloc_flags|ALLOC_CPUSET, ac);
4265 * fallback to ignore cpuset restriction if our nodes
4269 page = get_page_from_freelist(gfp_mask, order,
4275 static inline struct page *
4276 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4277 const struct alloc_context *ac, unsigned long *did_some_progress)
4279 struct oom_control oc = {
4280 .zonelist = ac->zonelist,
4281 .nodemask = ac->nodemask,
4283 .gfp_mask = gfp_mask,
4288 *did_some_progress = 0;
4291 * Acquire the oom lock. If that fails, somebody else is
4292 * making progress for us.
4294 if (!mutex_trylock(&oom_lock)) {
4295 *did_some_progress = 1;
4296 schedule_timeout_uninterruptible(1);
4301 * Go through the zonelist yet one more time, keep very high watermark
4302 * here, this is only to catch a parallel oom killing, we must fail if
4303 * we're still under heavy pressure. But make sure that this reclaim
4304 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4305 * allocation which will never fail due to oom_lock already held.
4307 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4308 ~__GFP_DIRECT_RECLAIM, order,
4309 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4313 /* Coredumps can quickly deplete all memory reserves */
4314 if (current->flags & PF_DUMPCORE)
4316 /* The OOM killer will not help higher order allocs */
4317 if (order > PAGE_ALLOC_COSTLY_ORDER)
4320 * We have already exhausted all our reclaim opportunities without any
4321 * success so it is time to admit defeat. We will skip the OOM killer
4322 * because it is very likely that the caller has a more reasonable
4323 * fallback than shooting a random task.
4325 * The OOM killer may not free memory on a specific node.
4327 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4329 /* The OOM killer does not needlessly kill tasks for lowmem */
4330 if (ac->highest_zoneidx < ZONE_NORMAL)
4332 if (pm_suspended_storage())
4335 * XXX: GFP_NOFS allocations should rather fail than rely on
4336 * other request to make a forward progress.
4337 * We are in an unfortunate situation where out_of_memory cannot
4338 * do much for this context but let's try it to at least get
4339 * access to memory reserved if the current task is killed (see
4340 * out_of_memory). Once filesystems are ready to handle allocation
4341 * failures more gracefully we should just bail out here.
4344 /* Exhausted what can be done so it's blame time */
4345 if (out_of_memory(&oc) ||
4346 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4347 *did_some_progress = 1;
4350 * Help non-failing allocations by giving them access to memory
4353 if (gfp_mask & __GFP_NOFAIL)
4354 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4355 ALLOC_NO_WATERMARKS, ac);
4358 mutex_unlock(&oom_lock);
4363 * Maximum number of compaction retries with a progress before OOM
4364 * killer is consider as the only way to move forward.
4366 #define MAX_COMPACT_RETRIES 16
4368 #ifdef CONFIG_COMPACTION
4369 /* Try memory compaction for high-order allocations before reclaim */
4370 static struct page *
4371 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4372 unsigned int alloc_flags, const struct alloc_context *ac,
4373 enum compact_priority prio, enum compact_result *compact_result)
4375 struct page *page = NULL;
4376 unsigned long pflags;
4377 unsigned int noreclaim_flag;
4382 psi_memstall_enter(&pflags);
4383 delayacct_compact_start();
4384 noreclaim_flag = memalloc_noreclaim_save();
4386 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4389 memalloc_noreclaim_restore(noreclaim_flag);
4390 psi_memstall_leave(&pflags);
4391 delayacct_compact_end();
4393 if (*compact_result == COMPACT_SKIPPED)
4396 * At least in one zone compaction wasn't deferred or skipped, so let's
4397 * count a compaction stall
4399 count_vm_event(COMPACTSTALL);
4401 /* Prep a captured page if available */
4403 prep_new_page(page, order, gfp_mask, alloc_flags);
4405 /* Try get a page from the freelist if available */
4407 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4410 struct zone *zone = page_zone(page);
4412 zone->compact_blockskip_flush = false;
4413 compaction_defer_reset(zone, order, true);
4414 count_vm_event(COMPACTSUCCESS);
4419 * It's bad if compaction run occurs and fails. The most likely reason
4420 * is that pages exist, but not enough to satisfy watermarks.
4422 count_vm_event(COMPACTFAIL);
4430 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4431 enum compact_result compact_result,
4432 enum compact_priority *compact_priority,
4433 int *compaction_retries)
4435 int max_retries = MAX_COMPACT_RETRIES;
4438 int retries = *compaction_retries;
4439 enum compact_priority priority = *compact_priority;
4444 if (fatal_signal_pending(current))
4447 if (compaction_made_progress(compact_result))
4448 (*compaction_retries)++;
4451 * compaction considers all the zone as desperately out of memory
4452 * so it doesn't really make much sense to retry except when the
4453 * failure could be caused by insufficient priority
4455 if (compaction_failed(compact_result))
4456 goto check_priority;
4459 * compaction was skipped because there are not enough order-0 pages
4460 * to work with, so we retry only if it looks like reclaim can help.
4462 if (compaction_needs_reclaim(compact_result)) {
4463 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4468 * make sure the compaction wasn't deferred or didn't bail out early
4469 * due to locks contention before we declare that we should give up.
4470 * But the next retry should use a higher priority if allowed, so
4471 * we don't just keep bailing out endlessly.
4473 if (compaction_withdrawn(compact_result)) {
4474 goto check_priority;
4478 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4479 * costly ones because they are de facto nofail and invoke OOM
4480 * killer to move on while costly can fail and users are ready
4481 * to cope with that. 1/4 retries is rather arbitrary but we
4482 * would need much more detailed feedback from compaction to
4483 * make a better decision.
4485 if (order > PAGE_ALLOC_COSTLY_ORDER)
4487 if (*compaction_retries <= max_retries) {
4493 * Make sure there are attempts at the highest priority if we exhausted
4494 * all retries or failed at the lower priorities.
4497 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4498 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4500 if (*compact_priority > min_priority) {
4501 (*compact_priority)--;
4502 *compaction_retries = 0;
4506 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4510 static inline struct page *
4511 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4512 unsigned int alloc_flags, const struct alloc_context *ac,
4513 enum compact_priority prio, enum compact_result *compact_result)
4515 *compact_result = COMPACT_SKIPPED;
4520 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4521 enum compact_result compact_result,
4522 enum compact_priority *compact_priority,
4523 int *compaction_retries)
4528 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4532 * There are setups with compaction disabled which would prefer to loop
4533 * inside the allocator rather than hit the oom killer prematurely.
4534 * Let's give them a good hope and keep retrying while the order-0
4535 * watermarks are OK.
4537 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4538 ac->highest_zoneidx, ac->nodemask) {
4539 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4540 ac->highest_zoneidx, alloc_flags))
4545 #endif /* CONFIG_COMPACTION */
4547 #ifdef CONFIG_LOCKDEP
4548 static struct lockdep_map __fs_reclaim_map =
4549 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4551 static bool __need_reclaim(gfp_t gfp_mask)
4553 /* no reclaim without waiting on it */
4554 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4557 /* this guy won't enter reclaim */
4558 if (current->flags & PF_MEMALLOC)
4561 if (gfp_mask & __GFP_NOLOCKDEP)
4567 void __fs_reclaim_acquire(unsigned long ip)
4569 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4572 void __fs_reclaim_release(unsigned long ip)
4574 lock_release(&__fs_reclaim_map, ip);
4577 void fs_reclaim_acquire(gfp_t gfp_mask)
4579 gfp_mask = current_gfp_context(gfp_mask);
4581 if (__need_reclaim(gfp_mask)) {
4582 if (gfp_mask & __GFP_FS)
4583 __fs_reclaim_acquire(_RET_IP_);
4585 #ifdef CONFIG_MMU_NOTIFIER
4586 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4587 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4592 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4594 void fs_reclaim_release(gfp_t gfp_mask)
4596 gfp_mask = current_gfp_context(gfp_mask);
4598 if (__need_reclaim(gfp_mask)) {
4599 if (gfp_mask & __GFP_FS)
4600 __fs_reclaim_release(_RET_IP_);
4603 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4606 /* Perform direct synchronous page reclaim */
4607 static unsigned long
4608 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4609 const struct alloc_context *ac)
4611 unsigned int noreclaim_flag;
4612 unsigned long progress;
4616 /* We now go into synchronous reclaim */
4617 cpuset_memory_pressure_bump();
4618 fs_reclaim_acquire(gfp_mask);
4619 noreclaim_flag = memalloc_noreclaim_save();
4621 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4624 memalloc_noreclaim_restore(noreclaim_flag);
4625 fs_reclaim_release(gfp_mask);
4632 /* The really slow allocator path where we enter direct reclaim */
4633 static inline struct page *
4634 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4635 unsigned int alloc_flags, const struct alloc_context *ac,
4636 unsigned long *did_some_progress)
4638 struct page *page = NULL;
4639 unsigned long pflags;
4640 bool drained = false;
4642 psi_memstall_enter(&pflags);
4643 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4644 if (unlikely(!(*did_some_progress)))
4648 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4651 * If an allocation failed after direct reclaim, it could be because
4652 * pages are pinned on the per-cpu lists or in high alloc reserves.
4653 * Shrink them and try again
4655 if (!page && !drained) {
4656 unreserve_highatomic_pageblock(ac, false);
4657 drain_all_pages(NULL);
4662 psi_memstall_leave(&pflags);
4667 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4668 const struct alloc_context *ac)
4672 pg_data_t *last_pgdat = NULL;
4673 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4675 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4677 if (!managed_zone(zone))
4679 if (last_pgdat != zone->zone_pgdat) {
4680 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4681 last_pgdat = zone->zone_pgdat;
4686 static inline unsigned int
4687 gfp_to_alloc_flags(gfp_t gfp_mask)
4689 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4692 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4693 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4694 * to save two branches.
4696 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4697 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4700 * The caller may dip into page reserves a bit more if the caller
4701 * cannot run direct reclaim, or if the caller has realtime scheduling
4702 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4703 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4705 alloc_flags |= (__force int)
4706 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4708 if (gfp_mask & __GFP_ATOMIC) {
4710 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4711 * if it can't schedule.
4713 if (!(gfp_mask & __GFP_NOMEMALLOC))
4714 alloc_flags |= ALLOC_HARDER;
4716 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4717 * comment for __cpuset_node_allowed().
4719 alloc_flags &= ~ALLOC_CPUSET;
4720 } else if (unlikely(rt_task(current)) && in_task())
4721 alloc_flags |= ALLOC_HARDER;
4723 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4728 static bool oom_reserves_allowed(struct task_struct *tsk)
4730 if (!tsk_is_oom_victim(tsk))
4734 * !MMU doesn't have oom reaper so give access to memory reserves
4735 * only to the thread with TIF_MEMDIE set
4737 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4744 * Distinguish requests which really need access to full memory
4745 * reserves from oom victims which can live with a portion of it
4747 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4749 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4751 if (gfp_mask & __GFP_MEMALLOC)
4752 return ALLOC_NO_WATERMARKS;
4753 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4754 return ALLOC_NO_WATERMARKS;
4755 if (!in_interrupt()) {
4756 if (current->flags & PF_MEMALLOC)
4757 return ALLOC_NO_WATERMARKS;
4758 else if (oom_reserves_allowed(current))
4765 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4767 return !!__gfp_pfmemalloc_flags(gfp_mask);
4771 * Checks whether it makes sense to retry the reclaim to make a forward progress
4772 * for the given allocation request.
4774 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4775 * without success, or when we couldn't even meet the watermark if we
4776 * reclaimed all remaining pages on the LRU lists.
4778 * Returns true if a retry is viable or false to enter the oom path.
4781 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4782 struct alloc_context *ac, int alloc_flags,
4783 bool did_some_progress, int *no_progress_loops)
4790 * Costly allocations might have made a progress but this doesn't mean
4791 * their order will become available due to high fragmentation so
4792 * always increment the no progress counter for them
4794 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4795 *no_progress_loops = 0;
4797 (*no_progress_loops)++;
4800 * Make sure we converge to OOM if we cannot make any progress
4801 * several times in the row.
4803 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4804 /* Before OOM, exhaust highatomic_reserve */
4805 return unreserve_highatomic_pageblock(ac, true);
4809 * Keep reclaiming pages while there is a chance this will lead
4810 * somewhere. If none of the target zones can satisfy our allocation
4811 * request even if all reclaimable pages are considered then we are
4812 * screwed and have to go OOM.
4814 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4815 ac->highest_zoneidx, ac->nodemask) {
4816 unsigned long available;
4817 unsigned long reclaimable;
4818 unsigned long min_wmark = min_wmark_pages(zone);
4821 available = reclaimable = zone_reclaimable_pages(zone);
4822 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4825 * Would the allocation succeed if we reclaimed all
4826 * reclaimable pages?
4828 wmark = __zone_watermark_ok(zone, order, min_wmark,
4829 ac->highest_zoneidx, alloc_flags, available);
4830 trace_reclaim_retry_zone(z, order, reclaimable,
4831 available, min_wmark, *no_progress_loops, wmark);
4839 * Memory allocation/reclaim might be called from a WQ context and the
4840 * current implementation of the WQ concurrency control doesn't
4841 * recognize that a particular WQ is congested if the worker thread is
4842 * looping without ever sleeping. Therefore we have to do a short sleep
4843 * here rather than calling cond_resched().
4845 if (current->flags & PF_WQ_WORKER)
4846 schedule_timeout_uninterruptible(1);
4853 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4856 * It's possible that cpuset's mems_allowed and the nodemask from
4857 * mempolicy don't intersect. This should be normally dealt with by
4858 * policy_nodemask(), but it's possible to race with cpuset update in
4859 * such a way the check therein was true, and then it became false
4860 * before we got our cpuset_mems_cookie here.
4861 * This assumes that for all allocations, ac->nodemask can come only
4862 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4863 * when it does not intersect with the cpuset restrictions) or the
4864 * caller can deal with a violated nodemask.
4866 if (cpusets_enabled() && ac->nodemask &&
4867 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4868 ac->nodemask = NULL;
4873 * When updating a task's mems_allowed or mempolicy nodemask, it is
4874 * possible to race with parallel threads in such a way that our
4875 * allocation can fail while the mask is being updated. If we are about
4876 * to fail, check if the cpuset changed during allocation and if so,
4879 if (read_mems_allowed_retry(cpuset_mems_cookie))
4885 static inline struct page *
4886 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4887 struct alloc_context *ac)
4889 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4890 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4891 struct page *page = NULL;
4892 unsigned int alloc_flags;
4893 unsigned long did_some_progress;
4894 enum compact_priority compact_priority;
4895 enum compact_result compact_result;
4896 int compaction_retries;
4897 int no_progress_loops;
4898 unsigned int cpuset_mems_cookie;
4902 * We also sanity check to catch abuse of atomic reserves being used by
4903 * callers that are not in atomic context.
4905 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4906 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4907 gfp_mask &= ~__GFP_ATOMIC;
4910 compaction_retries = 0;
4911 no_progress_loops = 0;
4912 compact_priority = DEF_COMPACT_PRIORITY;
4913 cpuset_mems_cookie = read_mems_allowed_begin();
4916 * The fast path uses conservative alloc_flags to succeed only until
4917 * kswapd needs to be woken up, and to avoid the cost of setting up
4918 * alloc_flags precisely. So we do that now.
4920 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4923 * We need to recalculate the starting point for the zonelist iterator
4924 * because we might have used different nodemask in the fast path, or
4925 * there was a cpuset modification and we are retrying - otherwise we
4926 * could end up iterating over non-eligible zones endlessly.
4928 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4929 ac->highest_zoneidx, ac->nodemask);
4930 if (!ac->preferred_zoneref->zone)
4934 * Check for insane configurations where the cpuset doesn't contain
4935 * any suitable zone to satisfy the request - e.g. non-movable
4936 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4938 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4939 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4940 ac->highest_zoneidx,
4941 &cpuset_current_mems_allowed);
4946 if (alloc_flags & ALLOC_KSWAPD)
4947 wake_all_kswapds(order, gfp_mask, ac);
4950 * The adjusted alloc_flags might result in immediate success, so try
4953 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4958 * For costly allocations, try direct compaction first, as it's likely
4959 * that we have enough base pages and don't need to reclaim. For non-
4960 * movable high-order allocations, do that as well, as compaction will
4961 * try prevent permanent fragmentation by migrating from blocks of the
4963 * Don't try this for allocations that are allowed to ignore
4964 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4966 if (can_direct_reclaim &&
4968 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4969 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4970 page = __alloc_pages_direct_compact(gfp_mask, order,
4972 INIT_COMPACT_PRIORITY,
4978 * Checks for costly allocations with __GFP_NORETRY, which
4979 * includes some THP page fault allocations
4981 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4983 * If allocating entire pageblock(s) and compaction
4984 * failed because all zones are below low watermarks
4985 * or is prohibited because it recently failed at this
4986 * order, fail immediately unless the allocator has
4987 * requested compaction and reclaim retry.
4990 * - potentially very expensive because zones are far
4991 * below their low watermarks or this is part of very
4992 * bursty high order allocations,
4993 * - not guaranteed to help because isolate_freepages()
4994 * may not iterate over freed pages as part of its
4996 * - unlikely to make entire pageblocks free on its
4999 if (compact_result == COMPACT_SKIPPED ||
5000 compact_result == COMPACT_DEFERRED)
5004 * Looks like reclaim/compaction is worth trying, but
5005 * sync compaction could be very expensive, so keep
5006 * using async compaction.
5008 compact_priority = INIT_COMPACT_PRIORITY;
5013 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5014 if (alloc_flags & ALLOC_KSWAPD)
5015 wake_all_kswapds(order, gfp_mask, ac);
5017 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5019 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
5022 * Reset the nodemask and zonelist iterators if memory policies can be
5023 * ignored. These allocations are high priority and system rather than
5026 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5027 ac->nodemask = NULL;
5028 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5029 ac->highest_zoneidx, ac->nodemask);
5032 /* Attempt with potentially adjusted zonelist and alloc_flags */
5033 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5037 /* Caller is not willing to reclaim, we can't balance anything */
5038 if (!can_direct_reclaim)
5041 /* Avoid recursion of direct reclaim */
5042 if (current->flags & PF_MEMALLOC)
5045 /* Try direct reclaim and then allocating */
5046 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5047 &did_some_progress);
5051 /* Try direct compaction and then allocating */
5052 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5053 compact_priority, &compact_result);
5057 /* Do not loop if specifically requested */
5058 if (gfp_mask & __GFP_NORETRY)
5062 * Do not retry costly high order allocations unless they are
5063 * __GFP_RETRY_MAYFAIL
5065 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5068 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5069 did_some_progress > 0, &no_progress_loops))
5073 * It doesn't make any sense to retry for the compaction if the order-0
5074 * reclaim is not able to make any progress because the current
5075 * implementation of the compaction depends on the sufficient amount
5076 * of free memory (see __compaction_suitable)
5078 if (did_some_progress > 0 &&
5079 should_compact_retry(ac, order, alloc_flags,
5080 compact_result, &compact_priority,
5081 &compaction_retries))
5085 /* Deal with possible cpuset update races before we start OOM killing */
5086 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5089 /* Reclaim has failed us, start killing things */
5090 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5094 /* Avoid allocations with no watermarks from looping endlessly */
5095 if (tsk_is_oom_victim(current) &&
5096 (alloc_flags & ALLOC_OOM ||
5097 (gfp_mask & __GFP_NOMEMALLOC)))
5100 /* Retry as long as the OOM killer is making progress */
5101 if (did_some_progress) {
5102 no_progress_loops = 0;
5107 /* Deal with possible cpuset update races before we fail */
5108 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5112 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5115 if (gfp_mask & __GFP_NOFAIL) {
5117 * All existing users of the __GFP_NOFAIL are blockable, so warn
5118 * of any new users that actually require GFP_NOWAIT
5120 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
5124 * PF_MEMALLOC request from this context is rather bizarre
5125 * because we cannot reclaim anything and only can loop waiting
5126 * for somebody to do a work for us
5128 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
5131 * non failing costly orders are a hard requirement which we
5132 * are not prepared for much so let's warn about these users
5133 * so that we can identify them and convert them to something
5136 WARN_ON_ONCE_GFP(order > PAGE_ALLOC_COSTLY_ORDER, gfp_mask);
5139 * Help non-failing allocations by giving them access to memory
5140 * reserves but do not use ALLOC_NO_WATERMARKS because this
5141 * could deplete whole memory reserves which would just make
5142 * the situation worse
5144 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5152 warn_alloc(gfp_mask, ac->nodemask,
5153 "page allocation failure: order:%u", order);
5158 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5159 int preferred_nid, nodemask_t *nodemask,
5160 struct alloc_context *ac, gfp_t *alloc_gfp,
5161 unsigned int *alloc_flags)
5163 ac->highest_zoneidx = gfp_zone(gfp_mask);
5164 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5165 ac->nodemask = nodemask;
5166 ac->migratetype = gfp_migratetype(gfp_mask);
5168 if (cpusets_enabled()) {
5169 *alloc_gfp |= __GFP_HARDWALL;
5171 * When we are in the interrupt context, it is irrelevant
5172 * to the current task context. It means that any node ok.
5174 if (in_task() && !ac->nodemask)
5175 ac->nodemask = &cpuset_current_mems_allowed;
5177 *alloc_flags |= ALLOC_CPUSET;
5180 fs_reclaim_acquire(gfp_mask);
5181 fs_reclaim_release(gfp_mask);
5183 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5185 if (should_fail_alloc_page(gfp_mask, order))
5188 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5190 /* Dirty zone balancing only done in the fast path */
5191 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5194 * The preferred zone is used for statistics but crucially it is
5195 * also used as the starting point for the zonelist iterator. It
5196 * may get reset for allocations that ignore memory policies.
5198 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5199 ac->highest_zoneidx, ac->nodemask);
5205 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5206 * @gfp: GFP flags for the allocation
5207 * @preferred_nid: The preferred NUMA node ID to allocate from
5208 * @nodemask: Set of nodes to allocate from, may be NULL
5209 * @nr_pages: The number of pages desired on the list or array
5210 * @page_list: Optional list to store the allocated pages
5211 * @page_array: Optional array to store the pages
5213 * This is a batched version of the page allocator that attempts to
5214 * allocate nr_pages quickly. Pages are added to page_list if page_list
5215 * is not NULL, otherwise it is assumed that the page_array is valid.
5217 * For lists, nr_pages is the number of pages that should be allocated.
5219 * For arrays, only NULL elements are populated with pages and nr_pages
5220 * is the maximum number of pages that will be stored in the array.
5222 * Returns the number of pages on the list or array.
5224 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5225 nodemask_t *nodemask, int nr_pages,
5226 struct list_head *page_list,
5227 struct page **page_array)
5230 unsigned long flags;
5233 struct per_cpu_pages *pcp;
5234 struct list_head *pcp_list;
5235 struct alloc_context ac;
5237 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5238 int nr_populated = 0, nr_account = 0;
5241 * Skip populated array elements to determine if any pages need
5242 * to be allocated before disabling IRQs.
5244 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5247 /* No pages requested? */
5248 if (unlikely(nr_pages <= 0))
5251 /* Already populated array? */
5252 if (unlikely(page_array && nr_pages - nr_populated == 0))
5255 /* Bulk allocator does not support memcg accounting. */
5256 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5259 /* Use the single page allocator for one page. */
5260 if (nr_pages - nr_populated == 1)
5263 #ifdef CONFIG_PAGE_OWNER
5265 * PAGE_OWNER may recurse into the allocator to allocate space to
5266 * save the stack with pagesets.lock held. Releasing/reacquiring
5267 * removes much of the performance benefit of bulk allocation so
5268 * force the caller to allocate one page at a time as it'll have
5269 * similar performance to added complexity to the bulk allocator.
5271 if (static_branch_unlikely(&page_owner_inited))
5275 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5276 gfp &= gfp_allowed_mask;
5278 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5282 /* Find an allowed local zone that meets the low watermark. */
5283 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5286 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5287 !__cpuset_zone_allowed(zone, gfp)) {
5291 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5292 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5296 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5297 if (zone_watermark_fast(zone, 0, mark,
5298 zonelist_zone_idx(ac.preferred_zoneref),
5299 alloc_flags, gfp)) {
5305 * If there are no allowed local zones that meets the watermarks then
5306 * try to allocate a single page and reclaim if necessary.
5308 if (unlikely(!zone))
5311 /* Attempt the batch allocation */
5312 local_lock_irqsave(&pagesets.lock, flags);
5313 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5314 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5316 while (nr_populated < nr_pages) {
5318 /* Skip existing pages */
5319 if (page_array && page_array[nr_populated]) {
5324 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5326 if (unlikely(!page)) {
5327 /* Try and get at least one page */
5334 prep_new_page(page, 0, gfp, 0);
5336 list_add(&page->lru, page_list);
5338 page_array[nr_populated] = page;
5342 local_unlock_irqrestore(&pagesets.lock, flags);
5344 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5345 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5348 return nr_populated;
5351 local_unlock_irqrestore(&pagesets.lock, flags);
5354 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5357 list_add(&page->lru, page_list);
5359 page_array[nr_populated] = page;
5365 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5368 * This is the 'heart' of the zoned buddy allocator.
5370 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5371 nodemask_t *nodemask)
5374 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5375 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5376 struct alloc_context ac = { };
5379 * There are several places where we assume that the order value is sane
5380 * so bail out early if the request is out of bound.
5382 if (WARN_ON_ONCE_GFP(order >= MAX_ORDER, gfp))
5385 gfp &= gfp_allowed_mask;
5387 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5388 * resp. GFP_NOIO which has to be inherited for all allocation requests
5389 * from a particular context which has been marked by
5390 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5391 * movable zones are not used during allocation.
5393 gfp = current_gfp_context(gfp);
5395 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5396 &alloc_gfp, &alloc_flags))
5400 * Forbid the first pass from falling back to types that fragment
5401 * memory until all local zones are considered.
5403 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5405 /* First allocation attempt */
5406 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5411 ac.spread_dirty_pages = false;
5414 * Restore the original nodemask if it was potentially replaced with
5415 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5417 ac.nodemask = nodemask;
5419 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5422 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5423 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5424 __free_pages(page, order);
5428 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5432 EXPORT_SYMBOL(__alloc_pages);
5434 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5435 nodemask_t *nodemask)
5437 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5438 preferred_nid, nodemask);
5440 if (page && order > 1)
5441 prep_transhuge_page(page);
5442 return (struct folio *)page;
5444 EXPORT_SYMBOL(__folio_alloc);
5447 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5448 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5449 * you need to access high mem.
5451 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5455 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5458 return (unsigned long) page_address(page);
5460 EXPORT_SYMBOL(__get_free_pages);
5462 unsigned long get_zeroed_page(gfp_t gfp_mask)
5464 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5466 EXPORT_SYMBOL(get_zeroed_page);
5469 * __free_pages - Free pages allocated with alloc_pages().
5470 * @page: The page pointer returned from alloc_pages().
5471 * @order: The order of the allocation.
5473 * This function can free multi-page allocations that are not compound
5474 * pages. It does not check that the @order passed in matches that of
5475 * the allocation, so it is easy to leak memory. Freeing more memory
5476 * than was allocated will probably emit a warning.
5478 * If the last reference to this page is speculative, it will be released
5479 * by put_page() which only frees the first page of a non-compound
5480 * allocation. To prevent the remaining pages from being leaked, we free
5481 * the subsequent pages here. If you want to use the page's reference
5482 * count to decide when to free the allocation, you should allocate a
5483 * compound page, and use put_page() instead of __free_pages().
5485 * Context: May be called in interrupt context or while holding a normal
5486 * spinlock, but not in NMI context or while holding a raw spinlock.
5488 void __free_pages(struct page *page, unsigned int order)
5490 if (put_page_testzero(page))
5491 free_the_page(page, order);
5492 else if (!PageHead(page))
5494 free_the_page(page + (1 << order), order);
5496 EXPORT_SYMBOL(__free_pages);
5498 void free_pages(unsigned long addr, unsigned int order)
5501 VM_BUG_ON(!virt_addr_valid((void *)addr));
5502 __free_pages(virt_to_page((void *)addr), order);
5506 EXPORT_SYMBOL(free_pages);
5510 * An arbitrary-length arbitrary-offset area of memory which resides
5511 * within a 0 or higher order page. Multiple fragments within that page
5512 * are individually refcounted, in the page's reference counter.
5514 * The page_frag functions below provide a simple allocation framework for
5515 * page fragments. This is used by the network stack and network device
5516 * drivers to provide a backing region of memory for use as either an
5517 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5519 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5522 struct page *page = NULL;
5523 gfp_t gfp = gfp_mask;
5525 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5526 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5528 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5529 PAGE_FRAG_CACHE_MAX_ORDER);
5530 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5532 if (unlikely(!page))
5533 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5535 nc->va = page ? page_address(page) : NULL;
5540 void __page_frag_cache_drain(struct page *page, unsigned int count)
5542 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5544 if (page_ref_sub_and_test(page, count))
5545 free_the_page(page, compound_order(page));
5547 EXPORT_SYMBOL(__page_frag_cache_drain);
5549 void *page_frag_alloc_align(struct page_frag_cache *nc,
5550 unsigned int fragsz, gfp_t gfp_mask,
5551 unsigned int align_mask)
5553 unsigned int size = PAGE_SIZE;
5557 if (unlikely(!nc->va)) {
5559 page = __page_frag_cache_refill(nc, gfp_mask);
5563 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5564 /* if size can vary use size else just use PAGE_SIZE */
5567 /* Even if we own the page, we do not use atomic_set().
5568 * This would break get_page_unless_zero() users.
5570 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5572 /* reset page count bias and offset to start of new frag */
5573 nc->pfmemalloc = page_is_pfmemalloc(page);
5574 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5578 offset = nc->offset - fragsz;
5579 if (unlikely(offset < 0)) {
5580 page = virt_to_page(nc->va);
5582 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5585 if (unlikely(nc->pfmemalloc)) {
5586 free_the_page(page, compound_order(page));
5590 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5591 /* if size can vary use size else just use PAGE_SIZE */
5594 /* OK, page count is 0, we can safely set it */
5595 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5597 /* reset page count bias and offset to start of new frag */
5598 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5599 offset = size - fragsz;
5603 offset &= align_mask;
5604 nc->offset = offset;
5606 return nc->va + offset;
5608 EXPORT_SYMBOL(page_frag_alloc_align);
5611 * Frees a page fragment allocated out of either a compound or order 0 page.
5613 void page_frag_free(void *addr)
5615 struct page *page = virt_to_head_page(addr);
5617 if (unlikely(put_page_testzero(page)))
5618 free_the_page(page, compound_order(page));
5620 EXPORT_SYMBOL(page_frag_free);
5622 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5626 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5627 unsigned long used = addr + PAGE_ALIGN(size);
5629 split_page(virt_to_page((void *)addr), order);
5630 while (used < alloc_end) {
5635 return (void *)addr;
5639 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5640 * @size: the number of bytes to allocate
5641 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5643 * This function is similar to alloc_pages(), except that it allocates the
5644 * minimum number of pages to satisfy the request. alloc_pages() can only
5645 * allocate memory in power-of-two pages.
5647 * This function is also limited by MAX_ORDER.
5649 * Memory allocated by this function must be released by free_pages_exact().
5651 * Return: pointer to the allocated area or %NULL in case of error.
5653 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5655 unsigned int order = get_order(size);
5658 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5659 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5661 addr = __get_free_pages(gfp_mask, order);
5662 return make_alloc_exact(addr, order, size);
5664 EXPORT_SYMBOL(alloc_pages_exact);
5667 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5669 * @nid: the preferred node ID where memory should be allocated
5670 * @size: the number of bytes to allocate
5671 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5673 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5676 * Return: pointer to the allocated area or %NULL in case of error.
5678 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5680 unsigned int order = get_order(size);
5683 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5684 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5686 p = alloc_pages_node(nid, gfp_mask, order);
5689 return make_alloc_exact((unsigned long)page_address(p), order, size);
5693 * free_pages_exact - release memory allocated via alloc_pages_exact()
5694 * @virt: the value returned by alloc_pages_exact.
5695 * @size: size of allocation, same value as passed to alloc_pages_exact().
5697 * Release the memory allocated by a previous call to alloc_pages_exact.
5699 void free_pages_exact(void *virt, size_t size)
5701 unsigned long addr = (unsigned long)virt;
5702 unsigned long end = addr + PAGE_ALIGN(size);
5704 while (addr < end) {
5709 EXPORT_SYMBOL(free_pages_exact);
5712 * nr_free_zone_pages - count number of pages beyond high watermark
5713 * @offset: The zone index of the highest zone
5715 * nr_free_zone_pages() counts the number of pages which are beyond the
5716 * high watermark within all zones at or below a given zone index. For each
5717 * zone, the number of pages is calculated as:
5719 * nr_free_zone_pages = managed_pages - high_pages
5721 * Return: number of pages beyond high watermark.
5723 static unsigned long nr_free_zone_pages(int offset)
5728 /* Just pick one node, since fallback list is circular */
5729 unsigned long sum = 0;
5731 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5733 for_each_zone_zonelist(zone, z, zonelist, offset) {
5734 unsigned long size = zone_managed_pages(zone);
5735 unsigned long high = high_wmark_pages(zone);
5744 * nr_free_buffer_pages - count number of pages beyond high watermark
5746 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5747 * watermark within ZONE_DMA and ZONE_NORMAL.
5749 * Return: number of pages beyond high watermark within ZONE_DMA and
5752 unsigned long nr_free_buffer_pages(void)
5754 return nr_free_zone_pages(gfp_zone(GFP_USER));
5756 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5758 static inline void show_node(struct zone *zone)
5760 if (IS_ENABLED(CONFIG_NUMA))
5761 printk("Node %d ", zone_to_nid(zone));
5764 long si_mem_available(void)
5767 unsigned long pagecache;
5768 unsigned long wmark_low = 0;
5769 unsigned long pages[NR_LRU_LISTS];
5770 unsigned long reclaimable;
5774 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5775 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5778 wmark_low += low_wmark_pages(zone);
5781 * Estimate the amount of memory available for userspace allocations,
5782 * without causing swapping.
5784 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5787 * Not all the page cache can be freed, otherwise the system will
5788 * start swapping. Assume at least half of the page cache, or the
5789 * low watermark worth of cache, needs to stay.
5791 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5792 pagecache -= min(pagecache / 2, wmark_low);
5793 available += pagecache;
5796 * Part of the reclaimable slab and other kernel memory consists of
5797 * items that are in use, and cannot be freed. Cap this estimate at the
5800 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5801 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5802 available += reclaimable - min(reclaimable / 2, wmark_low);
5808 EXPORT_SYMBOL_GPL(si_mem_available);
5810 void si_meminfo(struct sysinfo *val)
5812 val->totalram = totalram_pages();
5813 val->sharedram = global_node_page_state(NR_SHMEM);
5814 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5815 val->bufferram = nr_blockdev_pages();
5816 val->totalhigh = totalhigh_pages();
5817 val->freehigh = nr_free_highpages();
5818 val->mem_unit = PAGE_SIZE;
5821 EXPORT_SYMBOL(si_meminfo);
5824 void si_meminfo_node(struct sysinfo *val, int nid)
5826 int zone_type; /* needs to be signed */
5827 unsigned long managed_pages = 0;
5828 unsigned long managed_highpages = 0;
5829 unsigned long free_highpages = 0;
5830 pg_data_t *pgdat = NODE_DATA(nid);
5832 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5833 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5834 val->totalram = managed_pages;
5835 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5836 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5837 #ifdef CONFIG_HIGHMEM
5838 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5839 struct zone *zone = &pgdat->node_zones[zone_type];
5841 if (is_highmem(zone)) {
5842 managed_highpages += zone_managed_pages(zone);
5843 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5846 val->totalhigh = managed_highpages;
5847 val->freehigh = free_highpages;
5849 val->totalhigh = managed_highpages;
5850 val->freehigh = free_highpages;
5852 val->mem_unit = PAGE_SIZE;
5857 * Determine whether the node should be displayed or not, depending on whether
5858 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5860 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5862 if (!(flags & SHOW_MEM_FILTER_NODES))
5866 * no node mask - aka implicit memory numa policy. Do not bother with
5867 * the synchronization - read_mems_allowed_begin - because we do not
5868 * have to be precise here.
5871 nodemask = &cpuset_current_mems_allowed;
5873 return !node_isset(nid, *nodemask);
5876 #define K(x) ((x) << (PAGE_SHIFT-10))
5878 static void show_migration_types(unsigned char type)
5880 static const char types[MIGRATE_TYPES] = {
5881 [MIGRATE_UNMOVABLE] = 'U',
5882 [MIGRATE_MOVABLE] = 'M',
5883 [MIGRATE_RECLAIMABLE] = 'E',
5884 [MIGRATE_HIGHATOMIC] = 'H',
5886 [MIGRATE_CMA] = 'C',
5888 #ifdef CONFIG_MEMORY_ISOLATION
5889 [MIGRATE_ISOLATE] = 'I',
5892 char tmp[MIGRATE_TYPES + 1];
5896 for (i = 0; i < MIGRATE_TYPES; i++) {
5897 if (type & (1 << i))
5902 printk(KERN_CONT "(%s) ", tmp);
5906 * Show free area list (used inside shift_scroll-lock stuff)
5907 * We also calculate the percentage fragmentation. We do this by counting the
5908 * memory on each free list with the exception of the first item on the list.
5911 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5914 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5916 unsigned long free_pcp = 0;
5921 for_each_populated_zone(zone) {
5922 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5925 for_each_online_cpu(cpu)
5926 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5929 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5930 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5931 " unevictable:%lu dirty:%lu writeback:%lu\n"
5932 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5933 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5934 " kernel_misc_reclaimable:%lu\n"
5935 " free:%lu free_pcp:%lu free_cma:%lu\n",
5936 global_node_page_state(NR_ACTIVE_ANON),
5937 global_node_page_state(NR_INACTIVE_ANON),
5938 global_node_page_state(NR_ISOLATED_ANON),
5939 global_node_page_state(NR_ACTIVE_FILE),
5940 global_node_page_state(NR_INACTIVE_FILE),
5941 global_node_page_state(NR_ISOLATED_FILE),
5942 global_node_page_state(NR_UNEVICTABLE),
5943 global_node_page_state(NR_FILE_DIRTY),
5944 global_node_page_state(NR_WRITEBACK),
5945 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5946 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5947 global_node_page_state(NR_FILE_MAPPED),
5948 global_node_page_state(NR_SHMEM),
5949 global_node_page_state(NR_PAGETABLE),
5950 global_zone_page_state(NR_BOUNCE),
5951 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5952 global_zone_page_state(NR_FREE_PAGES),
5954 global_zone_page_state(NR_FREE_CMA_PAGES));
5956 for_each_online_pgdat(pgdat) {
5957 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5961 " active_anon:%lukB"
5962 " inactive_anon:%lukB"
5963 " active_file:%lukB"
5964 " inactive_file:%lukB"
5965 " unevictable:%lukB"
5966 " isolated(anon):%lukB"
5967 " isolated(file):%lukB"
5972 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5974 " shmem_pmdmapped: %lukB"
5977 " writeback_tmp:%lukB"
5978 " kernel_stack:%lukB"
5979 #ifdef CONFIG_SHADOW_CALL_STACK
5980 " shadow_call_stack:%lukB"
5983 " all_unreclaimable? %s"
5986 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5987 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5988 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5989 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5990 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5991 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5992 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5993 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5994 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5995 K(node_page_state(pgdat, NR_WRITEBACK)),
5996 K(node_page_state(pgdat, NR_SHMEM)),
5997 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5998 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5999 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6000 K(node_page_state(pgdat, NR_ANON_THPS)),
6002 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6003 node_page_state(pgdat, NR_KERNEL_STACK_KB),
6004 #ifdef CONFIG_SHADOW_CALL_STACK
6005 node_page_state(pgdat, NR_KERNEL_SCS_KB),
6007 K(node_page_state(pgdat, NR_PAGETABLE)),
6008 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6012 for_each_populated_zone(zone) {
6015 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6019 for_each_online_cpu(cpu)
6020 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6030 " reserved_highatomic:%luKB"
6031 " active_anon:%lukB"
6032 " inactive_anon:%lukB"
6033 " active_file:%lukB"
6034 " inactive_file:%lukB"
6035 " unevictable:%lukB"
6036 " writepending:%lukB"
6046 K(zone_page_state(zone, NR_FREE_PAGES)),
6047 K(zone->watermark_boost),
6048 K(min_wmark_pages(zone)),
6049 K(low_wmark_pages(zone)),
6050 K(high_wmark_pages(zone)),
6051 K(zone->nr_reserved_highatomic),
6052 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6053 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6054 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6055 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6056 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6057 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6058 K(zone->present_pages),
6059 K(zone_managed_pages(zone)),
6060 K(zone_page_state(zone, NR_MLOCK)),
6061 K(zone_page_state(zone, NR_BOUNCE)),
6063 K(this_cpu_read(zone->per_cpu_pageset->count)),
6064 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6065 printk("lowmem_reserve[]:");
6066 for (i = 0; i < MAX_NR_ZONES; i++)
6067 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6068 printk(KERN_CONT "\n");
6071 for_each_populated_zone(zone) {
6073 unsigned long nr[MAX_ORDER], flags, total = 0;
6074 unsigned char types[MAX_ORDER];
6076 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6079 printk(KERN_CONT "%s: ", zone->name);
6081 spin_lock_irqsave(&zone->lock, flags);
6082 for (order = 0; order < MAX_ORDER; order++) {
6083 struct free_area *area = &zone->free_area[order];
6086 nr[order] = area->nr_free;
6087 total += nr[order] << order;
6090 for (type = 0; type < MIGRATE_TYPES; type++) {
6091 if (!free_area_empty(area, type))
6092 types[order] |= 1 << type;
6095 spin_unlock_irqrestore(&zone->lock, flags);
6096 for (order = 0; order < MAX_ORDER; order++) {
6097 printk(KERN_CONT "%lu*%lukB ",
6098 nr[order], K(1UL) << order);
6100 show_migration_types(types[order]);
6102 printk(KERN_CONT "= %lukB\n", K(total));
6105 hugetlb_show_meminfo();
6107 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6109 show_swap_cache_info();
6112 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6114 zoneref->zone = zone;
6115 zoneref->zone_idx = zone_idx(zone);
6119 * Builds allocation fallback zone lists.
6121 * Add all populated zones of a node to the zonelist.
6123 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6126 enum zone_type zone_type = MAX_NR_ZONES;
6131 zone = pgdat->node_zones + zone_type;
6132 if (populated_zone(zone)) {
6133 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6134 check_highest_zone(zone_type);
6136 } while (zone_type);
6143 static int __parse_numa_zonelist_order(char *s)
6146 * We used to support different zonelists modes but they turned
6147 * out to be just not useful. Let's keep the warning in place
6148 * if somebody still use the cmd line parameter so that we do
6149 * not fail it silently
6151 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6152 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6158 char numa_zonelist_order[] = "Node";
6161 * sysctl handler for numa_zonelist_order
6163 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6164 void *buffer, size_t *length, loff_t *ppos)
6167 return __parse_numa_zonelist_order(buffer);
6168 return proc_dostring(table, write, buffer, length, ppos);
6172 static int node_load[MAX_NUMNODES];
6175 * find_next_best_node - find the next node that should appear in a given node's fallback list
6176 * @node: node whose fallback list we're appending
6177 * @used_node_mask: nodemask_t of already used nodes
6179 * We use a number of factors to determine which is the next node that should
6180 * appear on a given node's fallback list. The node should not have appeared
6181 * already in @node's fallback list, and it should be the next closest node
6182 * according to the distance array (which contains arbitrary distance values
6183 * from each node to each node in the system), and should also prefer nodes
6184 * with no CPUs, since presumably they'll have very little allocation pressure
6185 * on them otherwise.
6187 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6189 int find_next_best_node(int node, nodemask_t *used_node_mask)
6192 int min_val = INT_MAX;
6193 int best_node = NUMA_NO_NODE;
6195 /* Use the local node if we haven't already */
6196 if (!node_isset(node, *used_node_mask)) {
6197 node_set(node, *used_node_mask);
6201 for_each_node_state(n, N_MEMORY) {
6203 /* Don't want a node to appear more than once */
6204 if (node_isset(n, *used_node_mask))
6207 /* Use the distance array to find the distance */
6208 val = node_distance(node, n);
6210 /* Penalize nodes under us ("prefer the next node") */
6213 /* Give preference to headless and unused nodes */
6214 if (!cpumask_empty(cpumask_of_node(n)))
6215 val += PENALTY_FOR_NODE_WITH_CPUS;
6217 /* Slight preference for less loaded node */
6218 val *= MAX_NUMNODES;
6219 val += node_load[n];
6221 if (val < min_val) {
6228 node_set(best_node, *used_node_mask);
6235 * Build zonelists ordered by node and zones within node.
6236 * This results in maximum locality--normal zone overflows into local
6237 * DMA zone, if any--but risks exhausting DMA zone.
6239 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6242 struct zoneref *zonerefs;
6245 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6247 for (i = 0; i < nr_nodes; i++) {
6250 pg_data_t *node = NODE_DATA(node_order[i]);
6252 nr_zones = build_zonerefs_node(node, zonerefs);
6253 zonerefs += nr_zones;
6255 zonerefs->zone = NULL;
6256 zonerefs->zone_idx = 0;
6260 * Build gfp_thisnode zonelists
6262 static void build_thisnode_zonelists(pg_data_t *pgdat)
6264 struct zoneref *zonerefs;
6267 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6268 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6269 zonerefs += nr_zones;
6270 zonerefs->zone = NULL;
6271 zonerefs->zone_idx = 0;
6275 * Build zonelists ordered by zone and nodes within zones.
6276 * This results in conserving DMA zone[s] until all Normal memory is
6277 * exhausted, but results in overflowing to remote node while memory
6278 * may still exist in local DMA zone.
6281 static void build_zonelists(pg_data_t *pgdat)
6283 static int node_order[MAX_NUMNODES];
6284 int node, nr_nodes = 0;
6285 nodemask_t used_mask = NODE_MASK_NONE;
6286 int local_node, prev_node;
6288 /* NUMA-aware ordering of nodes */
6289 local_node = pgdat->node_id;
6290 prev_node = local_node;
6292 memset(node_order, 0, sizeof(node_order));
6293 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6295 * We don't want to pressure a particular node.
6296 * So adding penalty to the first node in same
6297 * distance group to make it round-robin.
6299 if (node_distance(local_node, node) !=
6300 node_distance(local_node, prev_node))
6301 node_load[node] += 1;
6303 node_order[nr_nodes++] = node;
6307 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6308 build_thisnode_zonelists(pgdat);
6309 pr_info("Fallback order for Node %d: ", local_node);
6310 for (node = 0; node < nr_nodes; node++)
6311 pr_cont("%d ", node_order[node]);
6315 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6317 * Return node id of node used for "local" allocations.
6318 * I.e., first node id of first zone in arg node's generic zonelist.
6319 * Used for initializing percpu 'numa_mem', which is used primarily
6320 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6322 int local_memory_node(int node)
6326 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6327 gfp_zone(GFP_KERNEL),
6329 return zone_to_nid(z->zone);
6333 static void setup_min_unmapped_ratio(void);
6334 static void setup_min_slab_ratio(void);
6335 #else /* CONFIG_NUMA */
6337 static void build_zonelists(pg_data_t *pgdat)
6339 int node, local_node;
6340 struct zoneref *zonerefs;
6343 local_node = pgdat->node_id;
6345 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6346 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6347 zonerefs += nr_zones;
6350 * Now we build the zonelist so that it contains the zones
6351 * of all the other nodes.
6352 * We don't want to pressure a particular node, so when
6353 * building the zones for node N, we make sure that the
6354 * zones coming right after the local ones are those from
6355 * node N+1 (modulo N)
6357 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6358 if (!node_online(node))
6360 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6361 zonerefs += nr_zones;
6363 for (node = 0; node < local_node; node++) {
6364 if (!node_online(node))
6366 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6367 zonerefs += nr_zones;
6370 zonerefs->zone = NULL;
6371 zonerefs->zone_idx = 0;
6374 #endif /* CONFIG_NUMA */
6377 * Boot pageset table. One per cpu which is going to be used for all
6378 * zones and all nodes. The parameters will be set in such a way
6379 * that an item put on a list will immediately be handed over to
6380 * the buddy list. This is safe since pageset manipulation is done
6381 * with interrupts disabled.
6383 * The boot_pagesets must be kept even after bootup is complete for
6384 * unused processors and/or zones. They do play a role for bootstrapping
6385 * hotplugged processors.
6387 * zoneinfo_show() and maybe other functions do
6388 * not check if the processor is online before following the pageset pointer.
6389 * Other parts of the kernel may not check if the zone is available.
6391 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6392 /* These effectively disable the pcplists in the boot pageset completely */
6393 #define BOOT_PAGESET_HIGH 0
6394 #define BOOT_PAGESET_BATCH 1
6395 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6396 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6397 DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6399 static void __build_all_zonelists(void *data)
6402 int __maybe_unused cpu;
6403 pg_data_t *self = data;
6404 static DEFINE_SPINLOCK(lock);
6409 memset(node_load, 0, sizeof(node_load));
6413 * This node is hotadded and no memory is yet present. So just
6414 * building zonelists is fine - no need to touch other nodes.
6416 if (self && !node_online(self->node_id)) {
6417 build_zonelists(self);
6420 * All possible nodes have pgdat preallocated
6423 for_each_node(nid) {
6424 pg_data_t *pgdat = NODE_DATA(nid);
6426 build_zonelists(pgdat);
6429 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6431 * We now know the "local memory node" for each node--
6432 * i.e., the node of the first zone in the generic zonelist.
6433 * Set up numa_mem percpu variable for on-line cpus. During
6434 * boot, only the boot cpu should be on-line; we'll init the
6435 * secondary cpus' numa_mem as they come on-line. During
6436 * node/memory hotplug, we'll fixup all on-line cpus.
6438 for_each_online_cpu(cpu)
6439 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6446 static noinline void __init
6447 build_all_zonelists_init(void)
6451 __build_all_zonelists(NULL);
6454 * Initialize the boot_pagesets that are going to be used
6455 * for bootstrapping processors. The real pagesets for
6456 * each zone will be allocated later when the per cpu
6457 * allocator is available.
6459 * boot_pagesets are used also for bootstrapping offline
6460 * cpus if the system is already booted because the pagesets
6461 * are needed to initialize allocators on a specific cpu too.
6462 * F.e. the percpu allocator needs the page allocator which
6463 * needs the percpu allocator in order to allocate its pagesets
6464 * (a chicken-egg dilemma).
6466 for_each_possible_cpu(cpu)
6467 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6469 mminit_verify_zonelist();
6470 cpuset_init_current_mems_allowed();
6474 * unless system_state == SYSTEM_BOOTING.
6476 * __ref due to call of __init annotated helper build_all_zonelists_init
6477 * [protected by SYSTEM_BOOTING].
6479 void __ref build_all_zonelists(pg_data_t *pgdat)
6481 unsigned long vm_total_pages;
6483 if (system_state == SYSTEM_BOOTING) {
6484 build_all_zonelists_init();
6486 __build_all_zonelists(pgdat);
6487 /* cpuset refresh routine should be here */
6489 /* Get the number of free pages beyond high watermark in all zones. */
6490 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6492 * Disable grouping by mobility if the number of pages in the
6493 * system is too low to allow the mechanism to work. It would be
6494 * more accurate, but expensive to check per-zone. This check is
6495 * made on memory-hotadd so a system can start with mobility
6496 * disabled and enable it later
6498 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6499 page_group_by_mobility_disabled = 1;
6501 page_group_by_mobility_disabled = 0;
6503 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6505 page_group_by_mobility_disabled ? "off" : "on",
6508 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6512 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6513 static bool __meminit
6514 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6516 static struct memblock_region *r;
6518 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6519 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6520 for_each_mem_region(r) {
6521 if (*pfn < memblock_region_memory_end_pfn(r))
6525 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6526 memblock_is_mirror(r)) {
6527 *pfn = memblock_region_memory_end_pfn(r);
6535 * Initially all pages are reserved - free ones are freed
6536 * up by memblock_free_all() once the early boot process is
6537 * done. Non-atomic initialization, single-pass.
6539 * All aligned pageblocks are initialized to the specified migratetype
6540 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6541 * zone stats (e.g., nr_isolate_pageblock) are touched.
6543 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6544 unsigned long start_pfn, unsigned long zone_end_pfn,
6545 enum meminit_context context,
6546 struct vmem_altmap *altmap, int migratetype)
6548 unsigned long pfn, end_pfn = start_pfn + size;
6551 if (highest_memmap_pfn < end_pfn - 1)
6552 highest_memmap_pfn = end_pfn - 1;
6554 #ifdef CONFIG_ZONE_DEVICE
6556 * Honor reservation requested by the driver for this ZONE_DEVICE
6557 * memory. We limit the total number of pages to initialize to just
6558 * those that might contain the memory mapping. We will defer the
6559 * ZONE_DEVICE page initialization until after we have released
6562 if (zone == ZONE_DEVICE) {
6566 if (start_pfn == altmap->base_pfn)
6567 start_pfn += altmap->reserve;
6568 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6572 for (pfn = start_pfn; pfn < end_pfn; ) {
6574 * There can be holes in boot-time mem_map[]s handed to this
6575 * function. They do not exist on hotplugged memory.
6577 if (context == MEMINIT_EARLY) {
6578 if (overlap_memmap_init(zone, &pfn))
6580 if (defer_init(nid, pfn, zone_end_pfn))
6584 page = pfn_to_page(pfn);
6585 __init_single_page(page, pfn, zone, nid);
6586 if (context == MEMINIT_HOTPLUG)
6587 __SetPageReserved(page);
6590 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6591 * such that unmovable allocations won't be scattered all
6592 * over the place during system boot.
6594 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6595 set_pageblock_migratetype(page, migratetype);
6602 #ifdef CONFIG_ZONE_DEVICE
6603 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6604 unsigned long zone_idx, int nid,
6605 struct dev_pagemap *pgmap)
6608 __init_single_page(page, pfn, zone_idx, nid);
6611 * Mark page reserved as it will need to wait for onlining
6612 * phase for it to be fully associated with a zone.
6614 * We can use the non-atomic __set_bit operation for setting
6615 * the flag as we are still initializing the pages.
6617 __SetPageReserved(page);
6620 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6621 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6622 * ever freed or placed on a driver-private list.
6624 page->pgmap = pgmap;
6625 page->zone_device_data = NULL;
6628 * Mark the block movable so that blocks are reserved for
6629 * movable at startup. This will force kernel allocations
6630 * to reserve their blocks rather than leaking throughout
6631 * the address space during boot when many long-lived
6632 * kernel allocations are made.
6634 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6635 * because this is done early in section_activate()
6637 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6638 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6644 * With compound page geometry and when struct pages are stored in ram most
6645 * tail pages are reused. Consequently, the amount of unique struct pages to
6646 * initialize is a lot smaller that the total amount of struct pages being
6647 * mapped. This is a paired / mild layering violation with explicit knowledge
6648 * of how the sparse_vmemmap internals handle compound pages in the lack
6649 * of an altmap. See vmemmap_populate_compound_pages().
6651 static inline unsigned long compound_nr_pages(struct vmem_altmap *altmap,
6652 unsigned long nr_pages)
6654 return is_power_of_2(sizeof(struct page)) &&
6655 !altmap ? 2 * (PAGE_SIZE / sizeof(struct page)) : nr_pages;
6658 static void __ref memmap_init_compound(struct page *head,
6659 unsigned long head_pfn,
6660 unsigned long zone_idx, int nid,
6661 struct dev_pagemap *pgmap,
6662 unsigned long nr_pages)
6664 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6665 unsigned int order = pgmap->vmemmap_shift;
6667 __SetPageHead(head);
6668 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6669 struct page *page = pfn_to_page(pfn);
6671 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6672 prep_compound_tail(head, pfn - head_pfn);
6673 set_page_count(page, 0);
6676 * The first tail page stores compound_mapcount_ptr() and
6677 * compound_order() and the second tail page stores
6678 * compound_pincount_ptr(). Call prep_compound_head() after
6679 * the first and second tail pages have been initialized to
6680 * not have the data overwritten.
6682 if (pfn == head_pfn + 2)
6683 prep_compound_head(head, order);
6687 void __ref memmap_init_zone_device(struct zone *zone,
6688 unsigned long start_pfn,
6689 unsigned long nr_pages,
6690 struct dev_pagemap *pgmap)
6692 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6693 struct pglist_data *pgdat = zone->zone_pgdat;
6694 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6695 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6696 unsigned long zone_idx = zone_idx(zone);
6697 unsigned long start = jiffies;
6698 int nid = pgdat->node_id;
6700 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6704 * The call to memmap_init should have already taken care
6705 * of the pages reserved for the memmap, so we can just jump to
6706 * the end of that region and start processing the device pages.
6709 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6710 nr_pages = end_pfn - start_pfn;
6713 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6714 struct page *page = pfn_to_page(pfn);
6716 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6718 if (pfns_per_compound == 1)
6721 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6722 compound_nr_pages(altmap, pfns_per_compound));
6725 pr_info("%s initialised %lu pages in %ums\n", __func__,
6726 nr_pages, jiffies_to_msecs(jiffies - start));
6730 static void __meminit zone_init_free_lists(struct zone *zone)
6732 unsigned int order, t;
6733 for_each_migratetype_order(order, t) {
6734 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6735 zone->free_area[order].nr_free = 0;
6740 * Only struct pages that correspond to ranges defined by memblock.memory
6741 * are zeroed and initialized by going through __init_single_page() during
6742 * memmap_init_zone_range().
6744 * But, there could be struct pages that correspond to holes in
6745 * memblock.memory. This can happen because of the following reasons:
6746 * - physical memory bank size is not necessarily the exact multiple of the
6747 * arbitrary section size
6748 * - early reserved memory may not be listed in memblock.memory
6749 * - memory layouts defined with memmap= kernel parameter may not align
6750 * nicely with memmap sections
6752 * Explicitly initialize those struct pages so that:
6753 * - PG_Reserved is set
6754 * - zone and node links point to zone and node that span the page if the
6755 * hole is in the middle of a zone
6756 * - zone and node links point to adjacent zone/node if the hole falls on
6757 * the zone boundary; the pages in such holes will be prepended to the
6758 * zone/node above the hole except for the trailing pages in the last
6759 * section that will be appended to the zone/node below.
6761 static void __init init_unavailable_range(unsigned long spfn,
6768 for (pfn = spfn; pfn < epfn; pfn++) {
6769 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6770 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6771 + pageblock_nr_pages - 1;
6774 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6775 __SetPageReserved(pfn_to_page(pfn));
6780 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6781 node, zone_names[zone], pgcnt);
6784 static void __init memmap_init_zone_range(struct zone *zone,
6785 unsigned long start_pfn,
6786 unsigned long end_pfn,
6787 unsigned long *hole_pfn)
6789 unsigned long zone_start_pfn = zone->zone_start_pfn;
6790 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6791 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6793 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6794 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6796 if (start_pfn >= end_pfn)
6799 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6800 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6802 if (*hole_pfn < start_pfn)
6803 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6805 *hole_pfn = end_pfn;
6808 static void __init memmap_init(void)
6810 unsigned long start_pfn, end_pfn;
6811 unsigned long hole_pfn = 0;
6812 int i, j, zone_id = 0, nid;
6814 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6815 struct pglist_data *node = NODE_DATA(nid);
6817 for (j = 0; j < MAX_NR_ZONES; j++) {
6818 struct zone *zone = node->node_zones + j;
6820 if (!populated_zone(zone))
6823 memmap_init_zone_range(zone, start_pfn, end_pfn,
6829 #ifdef CONFIG_SPARSEMEM
6831 * Initialize the memory map for hole in the range [memory_end,
6833 * Append the pages in this hole to the highest zone in the last
6835 * The call to init_unavailable_range() is outside the ifdef to
6836 * silence the compiler warining about zone_id set but not used;
6837 * for FLATMEM it is a nop anyway
6839 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6840 if (hole_pfn < end_pfn)
6842 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6845 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6846 phys_addr_t min_addr, int nid, bool exact_nid)
6851 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6852 MEMBLOCK_ALLOC_ACCESSIBLE,
6855 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6856 MEMBLOCK_ALLOC_ACCESSIBLE,
6859 if (ptr && size > 0)
6860 page_init_poison(ptr, size);
6865 static int zone_batchsize(struct zone *zone)
6871 * The number of pages to batch allocate is either ~0.1%
6872 * of the zone or 1MB, whichever is smaller. The batch
6873 * size is striking a balance between allocation latency
6874 * and zone lock contention.
6876 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6877 batch /= 4; /* We effectively *= 4 below */
6882 * Clamp the batch to a 2^n - 1 value. Having a power
6883 * of 2 value was found to be more likely to have
6884 * suboptimal cache aliasing properties in some cases.
6886 * For example if 2 tasks are alternately allocating
6887 * batches of pages, one task can end up with a lot
6888 * of pages of one half of the possible page colors
6889 * and the other with pages of the other colors.
6891 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6896 /* The deferral and batching of frees should be suppressed under NOMMU
6899 * The problem is that NOMMU needs to be able to allocate large chunks
6900 * of contiguous memory as there's no hardware page translation to
6901 * assemble apparent contiguous memory from discontiguous pages.
6903 * Queueing large contiguous runs of pages for batching, however,
6904 * causes the pages to actually be freed in smaller chunks. As there
6905 * can be a significant delay between the individual batches being
6906 * recycled, this leads to the once large chunks of space being
6907 * fragmented and becoming unavailable for high-order allocations.
6913 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6918 unsigned long total_pages;
6920 if (!percpu_pagelist_high_fraction) {
6922 * By default, the high value of the pcp is based on the zone
6923 * low watermark so that if they are full then background
6924 * reclaim will not be started prematurely.
6926 total_pages = low_wmark_pages(zone);
6929 * If percpu_pagelist_high_fraction is configured, the high
6930 * value is based on a fraction of the managed pages in the
6933 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6937 * Split the high value across all online CPUs local to the zone. Note
6938 * that early in boot that CPUs may not be online yet and that during
6939 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6940 * onlined. For memory nodes that have no CPUs, split pcp->high across
6941 * all online CPUs to mitigate the risk that reclaim is triggered
6942 * prematurely due to pages stored on pcp lists.
6944 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6946 nr_split_cpus = num_online_cpus();
6947 high = total_pages / nr_split_cpus;
6950 * Ensure high is at least batch*4. The multiple is based on the
6951 * historical relationship between high and batch.
6953 high = max(high, batch << 2);
6962 * pcp->high and pcp->batch values are related and generally batch is lower
6963 * than high. They are also related to pcp->count such that count is lower
6964 * than high, and as soon as it reaches high, the pcplist is flushed.
6966 * However, guaranteeing these relations at all times would require e.g. write
6967 * barriers here but also careful usage of read barriers at the read side, and
6968 * thus be prone to error and bad for performance. Thus the update only prevents
6969 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6970 * can cope with those fields changing asynchronously, and fully trust only the
6971 * pcp->count field on the local CPU with interrupts disabled.
6973 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6974 * outside of boot time (or some other assurance that no concurrent updaters
6977 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6978 unsigned long batch)
6980 WRITE_ONCE(pcp->batch, batch);
6981 WRITE_ONCE(pcp->high, high);
6984 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6988 memset(pcp, 0, sizeof(*pcp));
6989 memset(pzstats, 0, sizeof(*pzstats));
6991 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6992 INIT_LIST_HEAD(&pcp->lists[pindex]);
6995 * Set batch and high values safe for a boot pageset. A true percpu
6996 * pageset's initialization will update them subsequently. Here we don't
6997 * need to be as careful as pageset_update() as nobody can access the
7000 pcp->high = BOOT_PAGESET_HIGH;
7001 pcp->batch = BOOT_PAGESET_BATCH;
7002 pcp->free_factor = 0;
7005 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
7006 unsigned long batch)
7008 struct per_cpu_pages *pcp;
7011 for_each_possible_cpu(cpu) {
7012 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7013 pageset_update(pcp, high, batch);
7018 * Calculate and set new high and batch values for all per-cpu pagesets of a
7019 * zone based on the zone's size.
7021 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
7023 int new_high, new_batch;
7025 new_batch = max(1, zone_batchsize(zone));
7026 new_high = zone_highsize(zone, new_batch, cpu_online);
7028 if (zone->pageset_high == new_high &&
7029 zone->pageset_batch == new_batch)
7032 zone->pageset_high = new_high;
7033 zone->pageset_batch = new_batch;
7035 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7038 void __meminit setup_zone_pageset(struct zone *zone)
7042 /* Size may be 0 on !SMP && !NUMA */
7043 if (sizeof(struct per_cpu_zonestat) > 0)
7044 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7046 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7047 for_each_possible_cpu(cpu) {
7048 struct per_cpu_pages *pcp;
7049 struct per_cpu_zonestat *pzstats;
7051 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7052 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7053 per_cpu_pages_init(pcp, pzstats);
7056 zone_set_pageset_high_and_batch(zone, 0);
7060 * Allocate per cpu pagesets and initialize them.
7061 * Before this call only boot pagesets were available.
7063 void __init setup_per_cpu_pageset(void)
7065 struct pglist_data *pgdat;
7067 int __maybe_unused cpu;
7069 for_each_populated_zone(zone)
7070 setup_zone_pageset(zone);
7074 * Unpopulated zones continue using the boot pagesets.
7075 * The numa stats for these pagesets need to be reset.
7076 * Otherwise, they will end up skewing the stats of
7077 * the nodes these zones are associated with.
7079 for_each_possible_cpu(cpu) {
7080 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7081 memset(pzstats->vm_numa_event, 0,
7082 sizeof(pzstats->vm_numa_event));
7086 for_each_online_pgdat(pgdat)
7087 pgdat->per_cpu_nodestats =
7088 alloc_percpu(struct per_cpu_nodestat);
7091 static __meminit void zone_pcp_init(struct zone *zone)
7094 * per cpu subsystem is not up at this point. The following code
7095 * relies on the ability of the linker to provide the
7096 * offset of a (static) per cpu variable into the per cpu area.
7098 zone->per_cpu_pageset = &boot_pageset;
7099 zone->per_cpu_zonestats = &boot_zonestats;
7100 zone->pageset_high = BOOT_PAGESET_HIGH;
7101 zone->pageset_batch = BOOT_PAGESET_BATCH;
7103 if (populated_zone(zone))
7104 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7105 zone->present_pages, zone_batchsize(zone));
7108 void __meminit init_currently_empty_zone(struct zone *zone,
7109 unsigned long zone_start_pfn,
7112 struct pglist_data *pgdat = zone->zone_pgdat;
7113 int zone_idx = zone_idx(zone) + 1;
7115 if (zone_idx > pgdat->nr_zones)
7116 pgdat->nr_zones = zone_idx;
7118 zone->zone_start_pfn = zone_start_pfn;
7120 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7121 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7123 (unsigned long)zone_idx(zone),
7124 zone_start_pfn, (zone_start_pfn + size));
7126 zone_init_free_lists(zone);
7127 zone->initialized = 1;
7131 * get_pfn_range_for_nid - Return the start and end page frames for a node
7132 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7133 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7134 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7136 * It returns the start and end page frame of a node based on information
7137 * provided by memblock_set_node(). If called for a node
7138 * with no available memory, a warning is printed and the start and end
7141 void __init get_pfn_range_for_nid(unsigned int nid,
7142 unsigned long *start_pfn, unsigned long *end_pfn)
7144 unsigned long this_start_pfn, this_end_pfn;
7150 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7151 *start_pfn = min(*start_pfn, this_start_pfn);
7152 *end_pfn = max(*end_pfn, this_end_pfn);
7155 if (*start_pfn == -1UL)
7160 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7161 * assumption is made that zones within a node are ordered in monotonic
7162 * increasing memory addresses so that the "highest" populated zone is used
7164 static void __init find_usable_zone_for_movable(void)
7167 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7168 if (zone_index == ZONE_MOVABLE)
7171 if (arch_zone_highest_possible_pfn[zone_index] >
7172 arch_zone_lowest_possible_pfn[zone_index])
7176 VM_BUG_ON(zone_index == -1);
7177 movable_zone = zone_index;
7181 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7182 * because it is sized independent of architecture. Unlike the other zones,
7183 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7184 * in each node depending on the size of each node and how evenly kernelcore
7185 * is distributed. This helper function adjusts the zone ranges
7186 * provided by the architecture for a given node by using the end of the
7187 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7188 * zones within a node are in order of monotonic increases memory addresses
7190 static void __init adjust_zone_range_for_zone_movable(int nid,
7191 unsigned long zone_type,
7192 unsigned long node_start_pfn,
7193 unsigned long node_end_pfn,
7194 unsigned long *zone_start_pfn,
7195 unsigned long *zone_end_pfn)
7197 /* Only adjust if ZONE_MOVABLE is on this node */
7198 if (zone_movable_pfn[nid]) {
7199 /* Size ZONE_MOVABLE */
7200 if (zone_type == ZONE_MOVABLE) {
7201 *zone_start_pfn = zone_movable_pfn[nid];
7202 *zone_end_pfn = min(node_end_pfn,
7203 arch_zone_highest_possible_pfn[movable_zone]);
7205 /* Adjust for ZONE_MOVABLE starting within this range */
7206 } else if (!mirrored_kernelcore &&
7207 *zone_start_pfn < zone_movable_pfn[nid] &&
7208 *zone_end_pfn > zone_movable_pfn[nid]) {
7209 *zone_end_pfn = zone_movable_pfn[nid];
7211 /* Check if this whole range is within ZONE_MOVABLE */
7212 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7213 *zone_start_pfn = *zone_end_pfn;
7218 * Return the number of pages a zone spans in a node, including holes
7219 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7221 static unsigned long __init zone_spanned_pages_in_node(int nid,
7222 unsigned long zone_type,
7223 unsigned long node_start_pfn,
7224 unsigned long node_end_pfn,
7225 unsigned long *zone_start_pfn,
7226 unsigned long *zone_end_pfn)
7228 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7229 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7230 /* When hotadd a new node from cpu_up(), the node should be empty */
7231 if (!node_start_pfn && !node_end_pfn)
7234 /* Get the start and end of the zone */
7235 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7236 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7237 adjust_zone_range_for_zone_movable(nid, zone_type,
7238 node_start_pfn, node_end_pfn,
7239 zone_start_pfn, zone_end_pfn);
7241 /* Check that this node has pages within the zone's required range */
7242 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7245 /* Move the zone boundaries inside the node if necessary */
7246 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7247 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7249 /* Return the spanned pages */
7250 return *zone_end_pfn - *zone_start_pfn;
7254 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7255 * then all holes in the requested range will be accounted for.
7257 unsigned long __init __absent_pages_in_range(int nid,
7258 unsigned long range_start_pfn,
7259 unsigned long range_end_pfn)
7261 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7262 unsigned long start_pfn, end_pfn;
7265 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7266 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7267 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7268 nr_absent -= end_pfn - start_pfn;
7274 * absent_pages_in_range - Return number of page frames in holes within a range
7275 * @start_pfn: The start PFN to start searching for holes
7276 * @end_pfn: The end PFN to stop searching for holes
7278 * Return: the number of pages frames in memory holes within a range.
7280 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7281 unsigned long end_pfn)
7283 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7286 /* Return the number of page frames in holes in a zone on a node */
7287 static unsigned long __init zone_absent_pages_in_node(int nid,
7288 unsigned long zone_type,
7289 unsigned long node_start_pfn,
7290 unsigned long node_end_pfn)
7292 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7293 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7294 unsigned long zone_start_pfn, zone_end_pfn;
7295 unsigned long nr_absent;
7297 /* When hotadd a new node from cpu_up(), the node should be empty */
7298 if (!node_start_pfn && !node_end_pfn)
7301 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7302 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7304 adjust_zone_range_for_zone_movable(nid, zone_type,
7305 node_start_pfn, node_end_pfn,
7306 &zone_start_pfn, &zone_end_pfn);
7307 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7310 * ZONE_MOVABLE handling.
7311 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7314 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7315 unsigned long start_pfn, end_pfn;
7316 struct memblock_region *r;
7318 for_each_mem_region(r) {
7319 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7320 zone_start_pfn, zone_end_pfn);
7321 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7322 zone_start_pfn, zone_end_pfn);
7324 if (zone_type == ZONE_MOVABLE &&
7325 memblock_is_mirror(r))
7326 nr_absent += end_pfn - start_pfn;
7328 if (zone_type == ZONE_NORMAL &&
7329 !memblock_is_mirror(r))
7330 nr_absent += end_pfn - start_pfn;
7337 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7338 unsigned long node_start_pfn,
7339 unsigned long node_end_pfn)
7341 unsigned long realtotalpages = 0, totalpages = 0;
7344 for (i = 0; i < MAX_NR_ZONES; i++) {
7345 struct zone *zone = pgdat->node_zones + i;
7346 unsigned long zone_start_pfn, zone_end_pfn;
7347 unsigned long spanned, absent;
7348 unsigned long size, real_size;
7350 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7355 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7360 real_size = size - absent;
7363 zone->zone_start_pfn = zone_start_pfn;
7365 zone->zone_start_pfn = 0;
7366 zone->spanned_pages = size;
7367 zone->present_pages = real_size;
7368 #if defined(CONFIG_MEMORY_HOTPLUG)
7369 zone->present_early_pages = real_size;
7373 realtotalpages += real_size;
7376 pgdat->node_spanned_pages = totalpages;
7377 pgdat->node_present_pages = realtotalpages;
7378 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7381 #ifndef CONFIG_SPARSEMEM
7383 * Calculate the size of the zone->blockflags rounded to an unsigned long
7384 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7385 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7386 * round what is now in bits to nearest long in bits, then return it in
7389 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7391 unsigned long usemapsize;
7393 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7394 usemapsize = roundup(zonesize, pageblock_nr_pages);
7395 usemapsize = usemapsize >> pageblock_order;
7396 usemapsize *= NR_PAGEBLOCK_BITS;
7397 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7399 return usemapsize / 8;
7402 static void __ref setup_usemap(struct zone *zone)
7404 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7405 zone->spanned_pages);
7406 zone->pageblock_flags = NULL;
7408 zone->pageblock_flags =
7409 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7411 if (!zone->pageblock_flags)
7412 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7413 usemapsize, zone->name, zone_to_nid(zone));
7417 static inline void setup_usemap(struct zone *zone) {}
7418 #endif /* CONFIG_SPARSEMEM */
7420 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7422 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7423 void __init set_pageblock_order(void)
7425 unsigned int order = MAX_ORDER - 1;
7427 /* Check that pageblock_nr_pages has not already been setup */
7428 if (pageblock_order)
7431 /* Don't let pageblocks exceed the maximum allocation granularity. */
7432 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7433 order = HUGETLB_PAGE_ORDER;
7436 * Assume the largest contiguous order of interest is a huge page.
7437 * This value may be variable depending on boot parameters on IA64 and
7440 pageblock_order = order;
7442 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7445 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7446 * is unused as pageblock_order is set at compile-time. See
7447 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7450 void __init set_pageblock_order(void)
7454 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7456 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7457 unsigned long present_pages)
7459 unsigned long pages = spanned_pages;
7462 * Provide a more accurate estimation if there are holes within
7463 * the zone and SPARSEMEM is in use. If there are holes within the
7464 * zone, each populated memory region may cost us one or two extra
7465 * memmap pages due to alignment because memmap pages for each
7466 * populated regions may not be naturally aligned on page boundary.
7467 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7469 if (spanned_pages > present_pages + (present_pages >> 4) &&
7470 IS_ENABLED(CONFIG_SPARSEMEM))
7471 pages = present_pages;
7473 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7476 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7477 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7479 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7481 spin_lock_init(&ds_queue->split_queue_lock);
7482 INIT_LIST_HEAD(&ds_queue->split_queue);
7483 ds_queue->split_queue_len = 0;
7486 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7489 #ifdef CONFIG_COMPACTION
7490 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7492 init_waitqueue_head(&pgdat->kcompactd_wait);
7495 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7498 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7502 pgdat_resize_init(pgdat);
7504 pgdat_init_split_queue(pgdat);
7505 pgdat_init_kcompactd(pgdat);
7507 init_waitqueue_head(&pgdat->kswapd_wait);
7508 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7510 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7511 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7513 pgdat_page_ext_init(pgdat);
7514 lruvec_init(&pgdat->__lruvec);
7517 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7518 unsigned long remaining_pages)
7520 atomic_long_set(&zone->managed_pages, remaining_pages);
7521 zone_set_nid(zone, nid);
7522 zone->name = zone_names[idx];
7523 zone->zone_pgdat = NODE_DATA(nid);
7524 spin_lock_init(&zone->lock);
7525 zone_seqlock_init(zone);
7526 zone_pcp_init(zone);
7530 * Set up the zone data structures
7531 * - init pgdat internals
7532 * - init all zones belonging to this node
7534 * NOTE: this function is only called during memory hotplug
7536 #ifdef CONFIG_MEMORY_HOTPLUG
7537 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7539 int nid = pgdat->node_id;
7543 pgdat_init_internals(pgdat);
7545 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7546 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7549 * Reset the nr_zones, order and highest_zoneidx before reuse.
7550 * Note that kswapd will init kswapd_highest_zoneidx properly
7551 * when it starts in the near future.
7553 pgdat->nr_zones = 0;
7554 pgdat->kswapd_order = 0;
7555 pgdat->kswapd_highest_zoneidx = 0;
7556 pgdat->node_start_pfn = 0;
7557 for_each_online_cpu(cpu) {
7558 struct per_cpu_nodestat *p;
7560 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7561 memset(p, 0, sizeof(*p));
7564 for (z = 0; z < MAX_NR_ZONES; z++)
7565 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7570 * Set up the zone data structures:
7571 * - mark all pages reserved
7572 * - mark all memory queues empty
7573 * - clear the memory bitmaps
7575 * NOTE: pgdat should get zeroed by caller.
7576 * NOTE: this function is only called during early init.
7578 static void __init free_area_init_core(struct pglist_data *pgdat)
7581 int nid = pgdat->node_id;
7583 pgdat_init_internals(pgdat);
7584 pgdat->per_cpu_nodestats = &boot_nodestats;
7586 for (j = 0; j < MAX_NR_ZONES; j++) {
7587 struct zone *zone = pgdat->node_zones + j;
7588 unsigned long size, freesize, memmap_pages;
7590 size = zone->spanned_pages;
7591 freesize = zone->present_pages;
7594 * Adjust freesize so that it accounts for how much memory
7595 * is used by this zone for memmap. This affects the watermark
7596 * and per-cpu initialisations
7598 memmap_pages = calc_memmap_size(size, freesize);
7599 if (!is_highmem_idx(j)) {
7600 if (freesize >= memmap_pages) {
7601 freesize -= memmap_pages;
7603 pr_debug(" %s zone: %lu pages used for memmap\n",
7604 zone_names[j], memmap_pages);
7606 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7607 zone_names[j], memmap_pages, freesize);
7610 /* Account for reserved pages */
7611 if (j == 0 && freesize > dma_reserve) {
7612 freesize -= dma_reserve;
7613 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7616 if (!is_highmem_idx(j))
7617 nr_kernel_pages += freesize;
7618 /* Charge for highmem memmap if there are enough kernel pages */
7619 else if (nr_kernel_pages > memmap_pages * 2)
7620 nr_kernel_pages -= memmap_pages;
7621 nr_all_pages += freesize;
7624 * Set an approximate value for lowmem here, it will be adjusted
7625 * when the bootmem allocator frees pages into the buddy system.
7626 * And all highmem pages will be managed by the buddy system.
7628 zone_init_internals(zone, j, nid, freesize);
7633 set_pageblock_order();
7635 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7639 #ifdef CONFIG_FLATMEM
7640 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7642 unsigned long __maybe_unused start = 0;
7643 unsigned long __maybe_unused offset = 0;
7645 /* Skip empty nodes */
7646 if (!pgdat->node_spanned_pages)
7649 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7650 offset = pgdat->node_start_pfn - start;
7651 /* ia64 gets its own node_mem_map, before this, without bootmem */
7652 if (!pgdat->node_mem_map) {
7653 unsigned long size, end;
7657 * The zone's endpoints aren't required to be MAX_ORDER
7658 * aligned but the node_mem_map endpoints must be in order
7659 * for the buddy allocator to function correctly.
7661 end = pgdat_end_pfn(pgdat);
7662 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7663 size = (end - start) * sizeof(struct page);
7664 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7665 pgdat->node_id, false);
7667 panic("Failed to allocate %ld bytes for node %d memory map\n",
7668 size, pgdat->node_id);
7669 pgdat->node_mem_map = map + offset;
7671 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7672 __func__, pgdat->node_id, (unsigned long)pgdat,
7673 (unsigned long)pgdat->node_mem_map);
7676 * With no DISCONTIG, the global mem_map is just set as node 0's
7678 if (pgdat == NODE_DATA(0)) {
7679 mem_map = NODE_DATA(0)->node_mem_map;
7680 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7686 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7687 #endif /* CONFIG_FLATMEM */
7689 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7690 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7692 pgdat->first_deferred_pfn = ULONG_MAX;
7695 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7698 static void __init free_area_init_node(int nid)
7700 pg_data_t *pgdat = NODE_DATA(nid);
7701 unsigned long start_pfn = 0;
7702 unsigned long end_pfn = 0;
7704 /* pg_data_t should be reset to zero when it's allocated */
7705 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7707 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7709 pgdat->node_id = nid;
7710 pgdat->node_start_pfn = start_pfn;
7711 pgdat->per_cpu_nodestats = NULL;
7713 if (start_pfn != end_pfn) {
7714 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7715 (u64)start_pfn << PAGE_SHIFT,
7716 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7718 pr_info("Initmem setup node %d as memoryless\n", nid);
7721 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7723 alloc_node_mem_map(pgdat);
7724 pgdat_set_deferred_range(pgdat);
7726 free_area_init_core(pgdat);
7729 static void __init free_area_init_memoryless_node(int nid)
7731 free_area_init_node(nid);
7734 #if MAX_NUMNODES > 1
7736 * Figure out the number of possible node ids.
7738 void __init setup_nr_node_ids(void)
7740 unsigned int highest;
7742 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7743 nr_node_ids = highest + 1;
7748 * node_map_pfn_alignment - determine the maximum internode alignment
7750 * This function should be called after node map is populated and sorted.
7751 * It calculates the maximum power of two alignment which can distinguish
7754 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7755 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7756 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7757 * shifted, 1GiB is enough and this function will indicate so.
7759 * This is used to test whether pfn -> nid mapping of the chosen memory
7760 * model has fine enough granularity to avoid incorrect mapping for the
7761 * populated node map.
7763 * Return: the determined alignment in pfn's. 0 if there is no alignment
7764 * requirement (single node).
7766 unsigned long __init node_map_pfn_alignment(void)
7768 unsigned long accl_mask = 0, last_end = 0;
7769 unsigned long start, end, mask;
7770 int last_nid = NUMA_NO_NODE;
7773 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7774 if (!start || last_nid < 0 || last_nid == nid) {
7781 * Start with a mask granular enough to pin-point to the
7782 * start pfn and tick off bits one-by-one until it becomes
7783 * too coarse to separate the current node from the last.
7785 mask = ~((1 << __ffs(start)) - 1);
7786 while (mask && last_end <= (start & (mask << 1)))
7789 /* accumulate all internode masks */
7793 /* convert mask to number of pages */
7794 return ~accl_mask + 1;
7798 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7800 * Return: the minimum PFN based on information provided via
7801 * memblock_set_node().
7803 unsigned long __init find_min_pfn_with_active_regions(void)
7805 return PHYS_PFN(memblock_start_of_DRAM());
7809 * early_calculate_totalpages()
7810 * Sum pages in active regions for movable zone.
7811 * Populate N_MEMORY for calculating usable_nodes.
7813 static unsigned long __init early_calculate_totalpages(void)
7815 unsigned long totalpages = 0;
7816 unsigned long start_pfn, end_pfn;
7819 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7820 unsigned long pages = end_pfn - start_pfn;
7822 totalpages += pages;
7824 node_set_state(nid, N_MEMORY);
7830 * Find the PFN the Movable zone begins in each node. Kernel memory
7831 * is spread evenly between nodes as long as the nodes have enough
7832 * memory. When they don't, some nodes will have more kernelcore than
7835 static void __init find_zone_movable_pfns_for_nodes(void)
7838 unsigned long usable_startpfn;
7839 unsigned long kernelcore_node, kernelcore_remaining;
7840 /* save the state before borrow the nodemask */
7841 nodemask_t saved_node_state = node_states[N_MEMORY];
7842 unsigned long totalpages = early_calculate_totalpages();
7843 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7844 struct memblock_region *r;
7846 /* Need to find movable_zone earlier when movable_node is specified. */
7847 find_usable_zone_for_movable();
7850 * If movable_node is specified, ignore kernelcore and movablecore
7853 if (movable_node_is_enabled()) {
7854 for_each_mem_region(r) {
7855 if (!memblock_is_hotpluggable(r))
7858 nid = memblock_get_region_node(r);
7860 usable_startpfn = PFN_DOWN(r->base);
7861 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7862 min(usable_startpfn, zone_movable_pfn[nid]) :
7870 * If kernelcore=mirror is specified, ignore movablecore option
7872 if (mirrored_kernelcore) {
7873 bool mem_below_4gb_not_mirrored = false;
7875 for_each_mem_region(r) {
7876 if (memblock_is_mirror(r))
7879 nid = memblock_get_region_node(r);
7881 usable_startpfn = memblock_region_memory_base_pfn(r);
7883 if (usable_startpfn < PHYS_PFN(SZ_4G)) {
7884 mem_below_4gb_not_mirrored = true;
7888 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7889 min(usable_startpfn, zone_movable_pfn[nid]) :
7893 if (mem_below_4gb_not_mirrored)
7894 pr_warn("This configuration results in unmirrored kernel memory.\n");
7900 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7901 * amount of necessary memory.
7903 if (required_kernelcore_percent)
7904 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7906 if (required_movablecore_percent)
7907 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7911 * If movablecore= was specified, calculate what size of
7912 * kernelcore that corresponds so that memory usable for
7913 * any allocation type is evenly spread. If both kernelcore
7914 * and movablecore are specified, then the value of kernelcore
7915 * will be used for required_kernelcore if it's greater than
7916 * what movablecore would have allowed.
7918 if (required_movablecore) {
7919 unsigned long corepages;
7922 * Round-up so that ZONE_MOVABLE is at least as large as what
7923 * was requested by the user
7925 required_movablecore =
7926 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7927 required_movablecore = min(totalpages, required_movablecore);
7928 corepages = totalpages - required_movablecore;
7930 required_kernelcore = max(required_kernelcore, corepages);
7934 * If kernelcore was not specified or kernelcore size is larger
7935 * than totalpages, there is no ZONE_MOVABLE.
7937 if (!required_kernelcore || required_kernelcore >= totalpages)
7940 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7941 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7944 /* Spread kernelcore memory as evenly as possible throughout nodes */
7945 kernelcore_node = required_kernelcore / usable_nodes;
7946 for_each_node_state(nid, N_MEMORY) {
7947 unsigned long start_pfn, end_pfn;
7950 * Recalculate kernelcore_node if the division per node
7951 * now exceeds what is necessary to satisfy the requested
7952 * amount of memory for the kernel
7954 if (required_kernelcore < kernelcore_node)
7955 kernelcore_node = required_kernelcore / usable_nodes;
7958 * As the map is walked, we track how much memory is usable
7959 * by the kernel using kernelcore_remaining. When it is
7960 * 0, the rest of the node is usable by ZONE_MOVABLE
7962 kernelcore_remaining = kernelcore_node;
7964 /* Go through each range of PFNs within this node */
7965 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7966 unsigned long size_pages;
7968 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7969 if (start_pfn >= end_pfn)
7972 /* Account for what is only usable for kernelcore */
7973 if (start_pfn < usable_startpfn) {
7974 unsigned long kernel_pages;
7975 kernel_pages = min(end_pfn, usable_startpfn)
7978 kernelcore_remaining -= min(kernel_pages,
7979 kernelcore_remaining);
7980 required_kernelcore -= min(kernel_pages,
7981 required_kernelcore);
7983 /* Continue if range is now fully accounted */
7984 if (end_pfn <= usable_startpfn) {
7987 * Push zone_movable_pfn to the end so
7988 * that if we have to rebalance
7989 * kernelcore across nodes, we will
7990 * not double account here
7992 zone_movable_pfn[nid] = end_pfn;
7995 start_pfn = usable_startpfn;
7999 * The usable PFN range for ZONE_MOVABLE is from
8000 * start_pfn->end_pfn. Calculate size_pages as the
8001 * number of pages used as kernelcore
8003 size_pages = end_pfn - start_pfn;
8004 if (size_pages > kernelcore_remaining)
8005 size_pages = kernelcore_remaining;
8006 zone_movable_pfn[nid] = start_pfn + size_pages;
8009 * Some kernelcore has been met, update counts and
8010 * break if the kernelcore for this node has been
8013 required_kernelcore -= min(required_kernelcore,
8015 kernelcore_remaining -= size_pages;
8016 if (!kernelcore_remaining)
8022 * If there is still required_kernelcore, we do another pass with one
8023 * less node in the count. This will push zone_movable_pfn[nid] further
8024 * along on the nodes that still have memory until kernelcore is
8028 if (usable_nodes && required_kernelcore > usable_nodes)
8032 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
8033 for (nid = 0; nid < MAX_NUMNODES; nid++) {
8034 unsigned long start_pfn, end_pfn;
8036 zone_movable_pfn[nid] =
8037 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
8039 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
8040 if (zone_movable_pfn[nid] >= end_pfn)
8041 zone_movable_pfn[nid] = 0;
8045 /* restore the node_state */
8046 node_states[N_MEMORY] = saved_node_state;
8049 /* Any regular or high memory on that node ? */
8050 static void check_for_memory(pg_data_t *pgdat, int nid)
8052 enum zone_type zone_type;
8054 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8055 struct zone *zone = &pgdat->node_zones[zone_type];
8056 if (populated_zone(zone)) {
8057 if (IS_ENABLED(CONFIG_HIGHMEM))
8058 node_set_state(nid, N_HIGH_MEMORY);
8059 if (zone_type <= ZONE_NORMAL)
8060 node_set_state(nid, N_NORMAL_MEMORY);
8067 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8068 * such cases we allow max_zone_pfn sorted in the descending order
8070 bool __weak arch_has_descending_max_zone_pfns(void)
8076 * free_area_init - Initialise all pg_data_t and zone data
8077 * @max_zone_pfn: an array of max PFNs for each zone
8079 * This will call free_area_init_node() for each active node in the system.
8080 * Using the page ranges provided by memblock_set_node(), the size of each
8081 * zone in each node and their holes is calculated. If the maximum PFN
8082 * between two adjacent zones match, it is assumed that the zone is empty.
8083 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8084 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8085 * starts where the previous one ended. For example, ZONE_DMA32 starts
8086 * at arch_max_dma_pfn.
8088 void __init free_area_init(unsigned long *max_zone_pfn)
8090 unsigned long start_pfn, end_pfn;
8094 /* Record where the zone boundaries are */
8095 memset(arch_zone_lowest_possible_pfn, 0,
8096 sizeof(arch_zone_lowest_possible_pfn));
8097 memset(arch_zone_highest_possible_pfn, 0,
8098 sizeof(arch_zone_highest_possible_pfn));
8100 start_pfn = find_min_pfn_with_active_regions();
8101 descending = arch_has_descending_max_zone_pfns();
8103 for (i = 0; i < MAX_NR_ZONES; i++) {
8105 zone = MAX_NR_ZONES - i - 1;
8109 if (zone == ZONE_MOVABLE)
8112 end_pfn = max(max_zone_pfn[zone], start_pfn);
8113 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8114 arch_zone_highest_possible_pfn[zone] = end_pfn;
8116 start_pfn = end_pfn;
8119 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8120 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8121 find_zone_movable_pfns_for_nodes();
8123 /* Print out the zone ranges */
8124 pr_info("Zone ranges:\n");
8125 for (i = 0; i < MAX_NR_ZONES; i++) {
8126 if (i == ZONE_MOVABLE)
8128 pr_info(" %-8s ", zone_names[i]);
8129 if (arch_zone_lowest_possible_pfn[i] ==
8130 arch_zone_highest_possible_pfn[i])
8133 pr_cont("[mem %#018Lx-%#018Lx]\n",
8134 (u64)arch_zone_lowest_possible_pfn[i]
8136 ((u64)arch_zone_highest_possible_pfn[i]
8137 << PAGE_SHIFT) - 1);
8140 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8141 pr_info("Movable zone start for each node\n");
8142 for (i = 0; i < MAX_NUMNODES; i++) {
8143 if (zone_movable_pfn[i])
8144 pr_info(" Node %d: %#018Lx\n", i,
8145 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8149 * Print out the early node map, and initialize the
8150 * subsection-map relative to active online memory ranges to
8151 * enable future "sub-section" extensions of the memory map.
8153 pr_info("Early memory node ranges\n");
8154 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8155 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8156 (u64)start_pfn << PAGE_SHIFT,
8157 ((u64)end_pfn << PAGE_SHIFT) - 1);
8158 subsection_map_init(start_pfn, end_pfn - start_pfn);
8161 /* Initialise every node */
8162 mminit_verify_pageflags_layout();
8163 setup_nr_node_ids();
8164 for_each_node(nid) {
8167 if (!node_online(nid)) {
8168 pr_info("Initializing node %d as memoryless\n", nid);
8170 /* Allocator not initialized yet */
8171 pgdat = arch_alloc_nodedata(nid);
8173 pr_err("Cannot allocate %zuB for node %d.\n",
8174 sizeof(*pgdat), nid);
8177 arch_refresh_nodedata(nid, pgdat);
8178 free_area_init_memoryless_node(nid);
8181 * We do not want to confuse userspace by sysfs
8182 * files/directories for node without any memory
8183 * attached to it, so this node is not marked as
8184 * N_MEMORY and not marked online so that no sysfs
8185 * hierarchy will be created via register_one_node for
8186 * it. The pgdat will get fully initialized by
8187 * hotadd_init_pgdat() when memory is hotplugged into
8193 pgdat = NODE_DATA(nid);
8194 free_area_init_node(nid);
8196 /* Any memory on that node */
8197 if (pgdat->node_present_pages)
8198 node_set_state(nid, N_MEMORY);
8199 check_for_memory(pgdat, nid);
8205 static int __init cmdline_parse_core(char *p, unsigned long *core,
8206 unsigned long *percent)
8208 unsigned long long coremem;
8214 /* Value may be a percentage of total memory, otherwise bytes */
8215 coremem = simple_strtoull(p, &endptr, 0);
8216 if (*endptr == '%') {
8217 /* Paranoid check for percent values greater than 100 */
8218 WARN_ON(coremem > 100);
8222 coremem = memparse(p, &p);
8223 /* Paranoid check that UL is enough for the coremem value */
8224 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8226 *core = coremem >> PAGE_SHIFT;
8233 * kernelcore=size sets the amount of memory for use for allocations that
8234 * cannot be reclaimed or migrated.
8236 static int __init cmdline_parse_kernelcore(char *p)
8238 /* parse kernelcore=mirror */
8239 if (parse_option_str(p, "mirror")) {
8240 mirrored_kernelcore = true;
8244 return cmdline_parse_core(p, &required_kernelcore,
8245 &required_kernelcore_percent);
8249 * movablecore=size sets the amount of memory for use for allocations that
8250 * can be reclaimed or migrated.
8252 static int __init cmdline_parse_movablecore(char *p)
8254 return cmdline_parse_core(p, &required_movablecore,
8255 &required_movablecore_percent);
8258 early_param("kernelcore", cmdline_parse_kernelcore);
8259 early_param("movablecore", cmdline_parse_movablecore);
8261 void adjust_managed_page_count(struct page *page, long count)
8263 atomic_long_add(count, &page_zone(page)->managed_pages);
8264 totalram_pages_add(count);
8265 #ifdef CONFIG_HIGHMEM
8266 if (PageHighMem(page))
8267 totalhigh_pages_add(count);
8270 EXPORT_SYMBOL(adjust_managed_page_count);
8272 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8275 unsigned long pages = 0;
8277 start = (void *)PAGE_ALIGN((unsigned long)start);
8278 end = (void *)((unsigned long)end & PAGE_MASK);
8279 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8280 struct page *page = virt_to_page(pos);
8281 void *direct_map_addr;
8284 * 'direct_map_addr' might be different from 'pos'
8285 * because some architectures' virt_to_page()
8286 * work with aliases. Getting the direct map
8287 * address ensures that we get a _writeable_
8288 * alias for the memset().
8290 direct_map_addr = page_address(page);
8292 * Perform a kasan-unchecked memset() since this memory
8293 * has not been initialized.
8295 direct_map_addr = kasan_reset_tag(direct_map_addr);
8296 if ((unsigned int)poison <= 0xFF)
8297 memset(direct_map_addr, poison, PAGE_SIZE);
8299 free_reserved_page(page);
8303 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8308 void __init mem_init_print_info(void)
8310 unsigned long physpages, codesize, datasize, rosize, bss_size;
8311 unsigned long init_code_size, init_data_size;
8313 physpages = get_num_physpages();
8314 codesize = _etext - _stext;
8315 datasize = _edata - _sdata;
8316 rosize = __end_rodata - __start_rodata;
8317 bss_size = __bss_stop - __bss_start;
8318 init_data_size = __init_end - __init_begin;
8319 init_code_size = _einittext - _sinittext;
8322 * Detect special cases and adjust section sizes accordingly:
8323 * 1) .init.* may be embedded into .data sections
8324 * 2) .init.text.* may be out of [__init_begin, __init_end],
8325 * please refer to arch/tile/kernel/vmlinux.lds.S.
8326 * 3) .rodata.* may be embedded into .text or .data sections.
8328 #define adj_init_size(start, end, size, pos, adj) \
8330 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8334 adj_init_size(__init_begin, __init_end, init_data_size,
8335 _sinittext, init_code_size);
8336 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8337 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8338 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8339 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8341 #undef adj_init_size
8343 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8344 #ifdef CONFIG_HIGHMEM
8348 K(nr_free_pages()), K(physpages),
8349 codesize >> 10, datasize >> 10, rosize >> 10,
8350 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8351 K(physpages - totalram_pages() - totalcma_pages),
8353 #ifdef CONFIG_HIGHMEM
8354 , K(totalhigh_pages())
8360 * set_dma_reserve - set the specified number of pages reserved in the first zone
8361 * @new_dma_reserve: The number of pages to mark reserved
8363 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8364 * In the DMA zone, a significant percentage may be consumed by kernel image
8365 * and other unfreeable allocations which can skew the watermarks badly. This
8366 * function may optionally be used to account for unfreeable pages in the
8367 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8368 * smaller per-cpu batchsize.
8370 void __init set_dma_reserve(unsigned long new_dma_reserve)
8372 dma_reserve = new_dma_reserve;
8375 static int page_alloc_cpu_dead(unsigned int cpu)
8379 lru_add_drain_cpu(cpu);
8380 mlock_page_drain_remote(cpu);
8384 * Spill the event counters of the dead processor
8385 * into the current processors event counters.
8386 * This artificially elevates the count of the current
8389 vm_events_fold_cpu(cpu);
8392 * Zero the differential counters of the dead processor
8393 * so that the vm statistics are consistent.
8395 * This is only okay since the processor is dead and cannot
8396 * race with what we are doing.
8398 cpu_vm_stats_fold(cpu);
8400 for_each_populated_zone(zone)
8401 zone_pcp_update(zone, 0);
8406 static int page_alloc_cpu_online(unsigned int cpu)
8410 for_each_populated_zone(zone)
8411 zone_pcp_update(zone, 1);
8416 int hashdist = HASHDIST_DEFAULT;
8418 static int __init set_hashdist(char *str)
8422 hashdist = simple_strtoul(str, &str, 0);
8425 __setup("hashdist=", set_hashdist);
8428 void __init page_alloc_init(void)
8433 if (num_node_state(N_MEMORY) == 1)
8437 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8438 "mm/page_alloc:pcp",
8439 page_alloc_cpu_online,
8440 page_alloc_cpu_dead);
8445 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8446 * or min_free_kbytes changes.
8448 static void calculate_totalreserve_pages(void)
8450 struct pglist_data *pgdat;
8451 unsigned long reserve_pages = 0;
8452 enum zone_type i, j;
8454 for_each_online_pgdat(pgdat) {
8456 pgdat->totalreserve_pages = 0;
8458 for (i = 0; i < MAX_NR_ZONES; i++) {
8459 struct zone *zone = pgdat->node_zones + i;
8461 unsigned long managed_pages = zone_managed_pages(zone);
8463 /* Find valid and maximum lowmem_reserve in the zone */
8464 for (j = i; j < MAX_NR_ZONES; j++) {
8465 if (zone->lowmem_reserve[j] > max)
8466 max = zone->lowmem_reserve[j];
8469 /* we treat the high watermark as reserved pages. */
8470 max += high_wmark_pages(zone);
8472 if (max > managed_pages)
8473 max = managed_pages;
8475 pgdat->totalreserve_pages += max;
8477 reserve_pages += max;
8480 totalreserve_pages = reserve_pages;
8484 * setup_per_zone_lowmem_reserve - called whenever
8485 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8486 * has a correct pages reserved value, so an adequate number of
8487 * pages are left in the zone after a successful __alloc_pages().
8489 static void setup_per_zone_lowmem_reserve(void)
8491 struct pglist_data *pgdat;
8492 enum zone_type i, j;
8494 for_each_online_pgdat(pgdat) {
8495 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8496 struct zone *zone = &pgdat->node_zones[i];
8497 int ratio = sysctl_lowmem_reserve_ratio[i];
8498 bool clear = !ratio || !zone_managed_pages(zone);
8499 unsigned long managed_pages = 0;
8501 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8502 struct zone *upper_zone = &pgdat->node_zones[j];
8504 managed_pages += zone_managed_pages(upper_zone);
8507 zone->lowmem_reserve[j] = 0;
8509 zone->lowmem_reserve[j] = managed_pages / ratio;
8514 /* update totalreserve_pages */
8515 calculate_totalreserve_pages();
8518 static void __setup_per_zone_wmarks(void)
8520 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8521 unsigned long lowmem_pages = 0;
8523 unsigned long flags;
8525 /* Calculate total number of !ZONE_HIGHMEM pages */
8526 for_each_zone(zone) {
8527 if (!is_highmem(zone))
8528 lowmem_pages += zone_managed_pages(zone);
8531 for_each_zone(zone) {
8534 spin_lock_irqsave(&zone->lock, flags);
8535 tmp = (u64)pages_min * zone_managed_pages(zone);
8536 do_div(tmp, lowmem_pages);
8537 if (is_highmem(zone)) {
8539 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8540 * need highmem pages, so cap pages_min to a small
8543 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8544 * deltas control async page reclaim, and so should
8545 * not be capped for highmem.
8547 unsigned long min_pages;
8549 min_pages = zone_managed_pages(zone) / 1024;
8550 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8551 zone->_watermark[WMARK_MIN] = min_pages;
8554 * If it's a lowmem zone, reserve a number of pages
8555 * proportionate to the zone's size.
8557 zone->_watermark[WMARK_MIN] = tmp;
8561 * Set the kswapd watermarks distance according to the
8562 * scale factor in proportion to available memory, but
8563 * ensure a minimum size on small systems.
8565 tmp = max_t(u64, tmp >> 2,
8566 mult_frac(zone_managed_pages(zone),
8567 watermark_scale_factor, 10000));
8569 zone->watermark_boost = 0;
8570 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8571 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8572 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8574 spin_unlock_irqrestore(&zone->lock, flags);
8577 /* update totalreserve_pages */
8578 calculate_totalreserve_pages();
8582 * setup_per_zone_wmarks - called when min_free_kbytes changes
8583 * or when memory is hot-{added|removed}
8585 * Ensures that the watermark[min,low,high] values for each zone are set
8586 * correctly with respect to min_free_kbytes.
8588 void setup_per_zone_wmarks(void)
8591 static DEFINE_SPINLOCK(lock);
8594 __setup_per_zone_wmarks();
8598 * The watermark size have changed so update the pcpu batch
8599 * and high limits or the limits may be inappropriate.
8602 zone_pcp_update(zone, 0);
8606 * Initialise min_free_kbytes.
8608 * For small machines we want it small (128k min). For large machines
8609 * we want it large (256MB max). But it is not linear, because network
8610 * bandwidth does not increase linearly with machine size. We use
8612 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8613 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8629 void calculate_min_free_kbytes(void)
8631 unsigned long lowmem_kbytes;
8632 int new_min_free_kbytes;
8634 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8635 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8637 if (new_min_free_kbytes > user_min_free_kbytes)
8638 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8640 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8641 new_min_free_kbytes, user_min_free_kbytes);
8645 int __meminit init_per_zone_wmark_min(void)
8647 calculate_min_free_kbytes();
8648 setup_per_zone_wmarks();
8649 refresh_zone_stat_thresholds();
8650 setup_per_zone_lowmem_reserve();
8653 setup_min_unmapped_ratio();
8654 setup_min_slab_ratio();
8657 khugepaged_min_free_kbytes_update();
8661 postcore_initcall(init_per_zone_wmark_min)
8664 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8665 * that we can call two helper functions whenever min_free_kbytes
8668 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8669 void *buffer, size_t *length, loff_t *ppos)
8673 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8678 user_min_free_kbytes = min_free_kbytes;
8679 setup_per_zone_wmarks();
8684 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8685 void *buffer, size_t *length, loff_t *ppos)
8689 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8694 setup_per_zone_wmarks();
8700 static void setup_min_unmapped_ratio(void)
8705 for_each_online_pgdat(pgdat)
8706 pgdat->min_unmapped_pages = 0;
8709 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8710 sysctl_min_unmapped_ratio) / 100;
8714 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8715 void *buffer, size_t *length, loff_t *ppos)
8719 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8723 setup_min_unmapped_ratio();
8728 static void setup_min_slab_ratio(void)
8733 for_each_online_pgdat(pgdat)
8734 pgdat->min_slab_pages = 0;
8737 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8738 sysctl_min_slab_ratio) / 100;
8741 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8742 void *buffer, size_t *length, loff_t *ppos)
8746 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8750 setup_min_slab_ratio();
8757 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8758 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8759 * whenever sysctl_lowmem_reserve_ratio changes.
8761 * The reserve ratio obviously has absolutely no relation with the
8762 * minimum watermarks. The lowmem reserve ratio can only make sense
8763 * if in function of the boot time zone sizes.
8765 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8766 void *buffer, size_t *length, loff_t *ppos)
8770 proc_dointvec_minmax(table, write, buffer, length, ppos);
8772 for (i = 0; i < MAX_NR_ZONES; i++) {
8773 if (sysctl_lowmem_reserve_ratio[i] < 1)
8774 sysctl_lowmem_reserve_ratio[i] = 0;
8777 setup_per_zone_lowmem_reserve();
8782 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8783 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8784 * pagelist can have before it gets flushed back to buddy allocator.
8786 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8787 int write, void *buffer, size_t *length, loff_t *ppos)
8790 int old_percpu_pagelist_high_fraction;
8793 mutex_lock(&pcp_batch_high_lock);
8794 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8796 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8797 if (!write || ret < 0)
8800 /* Sanity checking to avoid pcp imbalance */
8801 if (percpu_pagelist_high_fraction &&
8802 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8803 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8809 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8812 for_each_populated_zone(zone)
8813 zone_set_pageset_high_and_batch(zone, 0);
8815 mutex_unlock(&pcp_batch_high_lock);
8819 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8821 * Returns the number of pages that arch has reserved but
8822 * is not known to alloc_large_system_hash().
8824 static unsigned long __init arch_reserved_kernel_pages(void)
8831 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8832 * machines. As memory size is increased the scale is also increased but at
8833 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8834 * quadruples the scale is increased by one, which means the size of hash table
8835 * only doubles, instead of quadrupling as well.
8836 * Because 32-bit systems cannot have large physical memory, where this scaling
8837 * makes sense, it is disabled on such platforms.
8839 #if __BITS_PER_LONG > 32
8840 #define ADAPT_SCALE_BASE (64ul << 30)
8841 #define ADAPT_SCALE_SHIFT 2
8842 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8846 * allocate a large system hash table from bootmem
8847 * - it is assumed that the hash table must contain an exact power-of-2
8848 * quantity of entries
8849 * - limit is the number of hash buckets, not the total allocation size
8851 void *__init alloc_large_system_hash(const char *tablename,
8852 unsigned long bucketsize,
8853 unsigned long numentries,
8856 unsigned int *_hash_shift,
8857 unsigned int *_hash_mask,
8858 unsigned long low_limit,
8859 unsigned long high_limit)
8861 unsigned long long max = high_limit;
8862 unsigned long log2qty, size;
8868 /* allow the kernel cmdline to have a say */
8870 /* round applicable memory size up to nearest megabyte */
8871 numentries = nr_kernel_pages;
8872 numentries -= arch_reserved_kernel_pages();
8874 /* It isn't necessary when PAGE_SIZE >= 1MB */
8875 if (PAGE_SHIFT < 20)
8876 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8878 #if __BITS_PER_LONG > 32
8880 unsigned long adapt;
8882 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8883 adapt <<= ADAPT_SCALE_SHIFT)
8888 /* limit to 1 bucket per 2^scale bytes of low memory */
8889 if (scale > PAGE_SHIFT)
8890 numentries >>= (scale - PAGE_SHIFT);
8892 numentries <<= (PAGE_SHIFT - scale);
8894 /* Make sure we've got at least a 0-order allocation.. */
8895 if (unlikely(flags & HASH_SMALL)) {
8896 /* Makes no sense without HASH_EARLY */
8897 WARN_ON(!(flags & HASH_EARLY));
8898 if (!(numentries >> *_hash_shift)) {
8899 numentries = 1UL << *_hash_shift;
8900 BUG_ON(!numentries);
8902 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8903 numentries = PAGE_SIZE / bucketsize;
8905 numentries = roundup_pow_of_two(numentries);
8907 /* limit allocation size to 1/16 total memory by default */
8909 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8910 do_div(max, bucketsize);
8912 max = min(max, 0x80000000ULL);
8914 if (numentries < low_limit)
8915 numentries = low_limit;
8916 if (numentries > max)
8919 log2qty = ilog2(numentries);
8921 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8924 size = bucketsize << log2qty;
8925 if (flags & HASH_EARLY) {
8926 if (flags & HASH_ZERO)
8927 table = memblock_alloc(size, SMP_CACHE_BYTES);
8929 table = memblock_alloc_raw(size,
8931 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8932 table = vmalloc_huge(size, gfp_flags);
8935 huge = is_vm_area_hugepages(table);
8938 * If bucketsize is not a power-of-two, we may free
8939 * some pages at the end of hash table which
8940 * alloc_pages_exact() automatically does
8942 table = alloc_pages_exact(size, gfp_flags);
8943 kmemleak_alloc(table, size, 1, gfp_flags);
8945 } while (!table && size > PAGE_SIZE && --log2qty);
8948 panic("Failed to allocate %s hash table\n", tablename);
8950 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8951 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8952 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8955 *_hash_shift = log2qty;
8957 *_hash_mask = (1 << log2qty) - 1;
8962 #ifdef CONFIG_CONTIG_ALLOC
8963 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8964 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8965 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8966 static void alloc_contig_dump_pages(struct list_head *page_list)
8968 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8970 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8974 list_for_each_entry(page, page_list, lru)
8975 dump_page(page, "migration failure");
8979 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8984 /* [start, end) must belong to a single zone. */
8985 int __alloc_contig_migrate_range(struct compact_control *cc,
8986 unsigned long start, unsigned long end)
8988 /* This function is based on compact_zone() from compaction.c. */
8989 unsigned int nr_reclaimed;
8990 unsigned long pfn = start;
8991 unsigned int tries = 0;
8993 struct migration_target_control mtc = {
8994 .nid = zone_to_nid(cc->zone),
8995 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8998 lru_cache_disable();
9000 while (pfn < end || !list_empty(&cc->migratepages)) {
9001 if (fatal_signal_pending(current)) {
9006 if (list_empty(&cc->migratepages)) {
9007 cc->nr_migratepages = 0;
9008 ret = isolate_migratepages_range(cc, pfn, end);
9009 if (ret && ret != -EAGAIN)
9011 pfn = cc->migrate_pfn;
9013 } else if (++tries == 5) {
9018 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9020 cc->nr_migratepages -= nr_reclaimed;
9022 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9023 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9026 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9027 * to retry again over this error, so do the same here.
9035 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
9036 alloc_contig_dump_pages(&cc->migratepages);
9037 putback_movable_pages(&cc->migratepages);
9044 * alloc_contig_range() -- tries to allocate given range of pages
9045 * @start: start PFN to allocate
9046 * @end: one-past-the-last PFN to allocate
9047 * @migratetype: migratetype of the underlying pageblocks (either
9048 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9049 * in range must have the same migratetype and it must
9050 * be either of the two.
9051 * @gfp_mask: GFP mask to use during compaction
9053 * The PFN range does not have to be pageblock aligned. The PFN range must
9054 * belong to a single zone.
9056 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9057 * pageblocks in the range. Once isolated, the pageblocks should not
9058 * be modified by others.
9060 * Return: zero on success or negative error code. On success all
9061 * pages which PFN is in [start, end) are allocated for the caller and
9062 * need to be freed with free_contig_range().
9064 int alloc_contig_range(unsigned long start, unsigned long end,
9065 unsigned migratetype, gfp_t gfp_mask)
9067 unsigned long outer_start, outer_end;
9071 struct compact_control cc = {
9072 .nr_migratepages = 0,
9074 .zone = page_zone(pfn_to_page(start)),
9075 .mode = MIGRATE_SYNC,
9076 .ignore_skip_hint = true,
9077 .no_set_skip_hint = true,
9078 .gfp_mask = current_gfp_context(gfp_mask),
9079 .alloc_contig = true,
9081 INIT_LIST_HEAD(&cc.migratepages);
9084 * What we do here is we mark all pageblocks in range as
9085 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9086 * have different sizes, and due to the way page allocator
9087 * work, start_isolate_page_range() has special handlings for this.
9089 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9090 * migrate the pages from an unaligned range (ie. pages that
9091 * we are interested in). This will put all the pages in
9092 * range back to page allocator as MIGRATE_ISOLATE.
9094 * When this is done, we take the pages in range from page
9095 * allocator removing them from the buddy system. This way
9096 * page allocator will never consider using them.
9098 * This lets us mark the pageblocks back as
9099 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9100 * aligned range but not in the unaligned, original range are
9101 * put back to page allocator so that buddy can use them.
9104 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
9108 drain_all_pages(cc.zone);
9111 * In case of -EBUSY, we'd like to know which page causes problem.
9112 * So, just fall through. test_pages_isolated() has a tracepoint
9113 * which will report the busy page.
9115 * It is possible that busy pages could become available before
9116 * the call to test_pages_isolated, and the range will actually be
9117 * allocated. So, if we fall through be sure to clear ret so that
9118 * -EBUSY is not accidentally used or returned to caller.
9120 ret = __alloc_contig_migrate_range(&cc, start, end);
9121 if (ret && ret != -EBUSY)
9126 * Pages from [start, end) are within a pageblock_nr_pages
9127 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9128 * more, all pages in [start, end) are free in page allocator.
9129 * What we are going to do is to allocate all pages from
9130 * [start, end) (that is remove them from page allocator).
9132 * The only problem is that pages at the beginning and at the
9133 * end of interesting range may be not aligned with pages that
9134 * page allocator holds, ie. they can be part of higher order
9135 * pages. Because of this, we reserve the bigger range and
9136 * once this is done free the pages we are not interested in.
9138 * We don't have to hold zone->lock here because the pages are
9139 * isolated thus they won't get removed from buddy.
9143 outer_start = start;
9144 while (!PageBuddy(pfn_to_page(outer_start))) {
9145 if (++order >= MAX_ORDER) {
9146 outer_start = start;
9149 outer_start &= ~0UL << order;
9152 if (outer_start != start) {
9153 order = buddy_order(pfn_to_page(outer_start));
9156 * outer_start page could be small order buddy page and
9157 * it doesn't include start page. Adjust outer_start
9158 * in this case to report failed page properly
9159 * on tracepoint in test_pages_isolated()
9161 if (outer_start + (1UL << order) <= start)
9162 outer_start = start;
9165 /* Make sure the range is really isolated. */
9166 if (test_pages_isolated(outer_start, end, 0)) {
9171 /* Grab isolated pages from freelists. */
9172 outer_end = isolate_freepages_range(&cc, outer_start, end);
9178 /* Free head and tail (if any) */
9179 if (start != outer_start)
9180 free_contig_range(outer_start, start - outer_start);
9181 if (end != outer_end)
9182 free_contig_range(end, outer_end - end);
9185 undo_isolate_page_range(start, end, migratetype);
9188 EXPORT_SYMBOL(alloc_contig_range);
9190 static int __alloc_contig_pages(unsigned long start_pfn,
9191 unsigned long nr_pages, gfp_t gfp_mask)
9193 unsigned long end_pfn = start_pfn + nr_pages;
9195 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9199 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9200 unsigned long nr_pages)
9202 unsigned long i, end_pfn = start_pfn + nr_pages;
9205 for (i = start_pfn; i < end_pfn; i++) {
9206 page = pfn_to_online_page(i);
9210 if (page_zone(page) != z)
9213 if (PageReserved(page))
9219 static bool zone_spans_last_pfn(const struct zone *zone,
9220 unsigned long start_pfn, unsigned long nr_pages)
9222 unsigned long last_pfn = start_pfn + nr_pages - 1;
9224 return zone_spans_pfn(zone, last_pfn);
9228 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9229 * @nr_pages: Number of contiguous pages to allocate
9230 * @gfp_mask: GFP mask to limit search and used during compaction
9232 * @nodemask: Mask for other possible nodes
9234 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9235 * on an applicable zonelist to find a contiguous pfn range which can then be
9236 * tried for allocation with alloc_contig_range(). This routine is intended
9237 * for allocation requests which can not be fulfilled with the buddy allocator.
9239 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9240 * power of two, then allocated range is also guaranteed to be aligned to same
9241 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9243 * Allocated pages can be freed with free_contig_range() or by manually calling
9244 * __free_page() on each allocated page.
9246 * Return: pointer to contiguous pages on success, or NULL if not successful.
9248 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9249 int nid, nodemask_t *nodemask)
9251 unsigned long ret, pfn, flags;
9252 struct zonelist *zonelist;
9256 zonelist = node_zonelist(nid, gfp_mask);
9257 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9258 gfp_zone(gfp_mask), nodemask) {
9259 spin_lock_irqsave(&zone->lock, flags);
9261 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9262 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9263 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9265 * We release the zone lock here because
9266 * alloc_contig_range() will also lock the zone
9267 * at some point. If there's an allocation
9268 * spinning on this lock, it may win the race
9269 * and cause alloc_contig_range() to fail...
9271 spin_unlock_irqrestore(&zone->lock, flags);
9272 ret = __alloc_contig_pages(pfn, nr_pages,
9275 return pfn_to_page(pfn);
9276 spin_lock_irqsave(&zone->lock, flags);
9280 spin_unlock_irqrestore(&zone->lock, flags);
9284 #endif /* CONFIG_CONTIG_ALLOC */
9286 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9288 unsigned long count = 0;
9290 for (; nr_pages--; pfn++) {
9291 struct page *page = pfn_to_page(pfn);
9293 count += page_count(page) != 1;
9296 WARN(count != 0, "%lu pages are still in use!\n", count);
9298 EXPORT_SYMBOL(free_contig_range);
9301 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9302 * page high values need to be recalculated.
9304 void zone_pcp_update(struct zone *zone, int cpu_online)
9306 mutex_lock(&pcp_batch_high_lock);
9307 zone_set_pageset_high_and_batch(zone, cpu_online);
9308 mutex_unlock(&pcp_batch_high_lock);
9312 * Effectively disable pcplists for the zone by setting the high limit to 0
9313 * and draining all cpus. A concurrent page freeing on another CPU that's about
9314 * to put the page on pcplist will either finish before the drain and the page
9315 * will be drained, or observe the new high limit and skip the pcplist.
9317 * Must be paired with a call to zone_pcp_enable().
9319 void zone_pcp_disable(struct zone *zone)
9321 mutex_lock(&pcp_batch_high_lock);
9322 __zone_set_pageset_high_and_batch(zone, 0, 1);
9323 __drain_all_pages(zone, true);
9326 void zone_pcp_enable(struct zone *zone)
9328 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9329 mutex_unlock(&pcp_batch_high_lock);
9332 void zone_pcp_reset(struct zone *zone)
9335 struct per_cpu_zonestat *pzstats;
9337 if (zone->per_cpu_pageset != &boot_pageset) {
9338 for_each_online_cpu(cpu) {
9339 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9340 drain_zonestat(zone, pzstats);
9342 free_percpu(zone->per_cpu_pageset);
9343 free_percpu(zone->per_cpu_zonestats);
9344 zone->per_cpu_pageset = &boot_pageset;
9345 zone->per_cpu_zonestats = &boot_zonestats;
9349 #ifdef CONFIG_MEMORY_HOTREMOVE
9351 * All pages in the range must be in a single zone, must not contain holes,
9352 * must span full sections, and must be isolated before calling this function.
9354 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9356 unsigned long pfn = start_pfn;
9360 unsigned long flags;
9362 offline_mem_sections(pfn, end_pfn);
9363 zone = page_zone(pfn_to_page(pfn));
9364 spin_lock_irqsave(&zone->lock, flags);
9365 while (pfn < end_pfn) {
9366 page = pfn_to_page(pfn);
9368 * The HWPoisoned page may be not in buddy system, and
9369 * page_count() is not 0.
9371 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9376 * At this point all remaining PageOffline() pages have a
9377 * reference count of 0 and can simply be skipped.
9379 if (PageOffline(page)) {
9380 BUG_ON(page_count(page));
9381 BUG_ON(PageBuddy(page));
9386 BUG_ON(page_count(page));
9387 BUG_ON(!PageBuddy(page));
9388 order = buddy_order(page);
9389 del_page_from_free_list(page, zone, order);
9390 pfn += (1 << order);
9392 spin_unlock_irqrestore(&zone->lock, flags);
9397 * This function returns a stable result only if called under zone lock.
9399 bool is_free_buddy_page(struct page *page)
9401 unsigned long pfn = page_to_pfn(page);
9404 for (order = 0; order < MAX_ORDER; order++) {
9405 struct page *page_head = page - (pfn & ((1 << order) - 1));
9407 if (PageBuddy(page_head) &&
9408 buddy_order_unsafe(page_head) >= order)
9412 return order < MAX_ORDER;
9414 EXPORT_SYMBOL(is_free_buddy_page);
9416 #ifdef CONFIG_MEMORY_FAILURE
9418 * Break down a higher-order page in sub-pages, and keep our target out of
9421 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9422 struct page *target, int low, int high,
9425 unsigned long size = 1 << high;
9426 struct page *current_buddy, *next_page;
9428 while (high > low) {
9432 if (target >= &page[size]) {
9433 next_page = page + size;
9434 current_buddy = page;
9437 current_buddy = page + size;
9440 if (set_page_guard(zone, current_buddy, high, migratetype))
9443 if (current_buddy != target) {
9444 add_to_free_list(current_buddy, zone, high, migratetype);
9445 set_buddy_order(current_buddy, high);
9452 * Take a page that will be marked as poisoned off the buddy allocator.
9454 bool take_page_off_buddy(struct page *page)
9456 struct zone *zone = page_zone(page);
9457 unsigned long pfn = page_to_pfn(page);
9458 unsigned long flags;
9462 spin_lock_irqsave(&zone->lock, flags);
9463 for (order = 0; order < MAX_ORDER; order++) {
9464 struct page *page_head = page - (pfn & ((1 << order) - 1));
9465 int page_order = buddy_order(page_head);
9467 if (PageBuddy(page_head) && page_order >= order) {
9468 unsigned long pfn_head = page_to_pfn(page_head);
9469 int migratetype = get_pfnblock_migratetype(page_head,
9472 del_page_from_free_list(page_head, zone, page_order);
9473 break_down_buddy_pages(zone, page_head, page, 0,
9474 page_order, migratetype);
9475 SetPageHWPoisonTakenOff(page);
9476 if (!is_migrate_isolate(migratetype))
9477 __mod_zone_freepage_state(zone, -1, migratetype);
9481 if (page_count(page_head) > 0)
9484 spin_unlock_irqrestore(&zone->lock, flags);
9489 * Cancel takeoff done by take_page_off_buddy().
9491 bool put_page_back_buddy(struct page *page)
9493 struct zone *zone = page_zone(page);
9494 unsigned long pfn = page_to_pfn(page);
9495 unsigned long flags;
9496 int migratetype = get_pfnblock_migratetype(page, pfn);
9499 spin_lock_irqsave(&zone->lock, flags);
9500 if (put_page_testzero(page)) {
9501 ClearPageHWPoisonTakenOff(page);
9502 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9503 if (TestClearPageHWPoison(page)) {
9507 spin_unlock_irqrestore(&zone->lock, flags);
9513 #ifdef CONFIG_ZONE_DMA
9514 bool has_managed_dma(void)
9516 struct pglist_data *pgdat;
9518 for_each_online_pgdat(pgdat) {
9519 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9521 if (managed_zone(zone))
9526 #endif /* CONFIG_ZONE_DMA */