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 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
487 * a consistent read of the memory array, so that results, even though
488 * racy, are not corrupted.
490 word = READ_ONCE(bitmap[word_bitidx]);
491 return (word >> bitidx) & mask;
495 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
496 * @page: The page within the block of interest
497 * @pfn: The target page frame number
498 * @mask: mask of bits that the caller is interested in
500 * Return: pageblock_bits flags
502 unsigned long get_pfnblock_flags_mask(const struct page *page,
503 unsigned long pfn, unsigned long mask)
505 return __get_pfnblock_flags_mask(page, pfn, mask);
508 static __always_inline int get_pfnblock_migratetype(const struct page *page,
511 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
515 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
516 * @page: The page within the block of interest
517 * @flags: The flags to set
518 * @pfn: The target page frame number
519 * @mask: mask of bits that the caller is interested in
521 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
525 unsigned long *bitmap;
526 unsigned long bitidx, word_bitidx;
527 unsigned long old_word, word;
529 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
530 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
532 bitmap = get_pageblock_bitmap(page, pfn);
533 bitidx = pfn_to_bitidx(page, pfn);
534 word_bitidx = bitidx / BITS_PER_LONG;
535 bitidx &= (BITS_PER_LONG-1);
537 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
542 word = READ_ONCE(bitmap[word_bitidx]);
544 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
545 if (word == old_word)
551 void set_pageblock_migratetype(struct page *page, int migratetype)
553 if (unlikely(page_group_by_mobility_disabled &&
554 migratetype < MIGRATE_PCPTYPES))
555 migratetype = MIGRATE_UNMOVABLE;
557 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
558 page_to_pfn(page), MIGRATETYPE_MASK);
561 #ifdef CONFIG_DEBUG_VM
562 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
566 unsigned long pfn = page_to_pfn(page);
567 unsigned long sp, start_pfn;
570 seq = zone_span_seqbegin(zone);
571 start_pfn = zone->zone_start_pfn;
572 sp = zone->spanned_pages;
573 if (!zone_spans_pfn(zone, pfn))
575 } while (zone_span_seqretry(zone, seq));
578 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
579 pfn, zone_to_nid(zone), zone->name,
580 start_pfn, start_pfn + sp);
585 static int page_is_consistent(struct zone *zone, struct page *page)
587 if (zone != page_zone(page))
593 * Temporary debugging check for pages not lying within a given zone.
595 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
597 if (page_outside_zone_boundaries(zone, page))
599 if (!page_is_consistent(zone, page))
605 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
611 static void bad_page(struct page *page, const char *reason)
613 static unsigned long resume;
614 static unsigned long nr_shown;
615 static unsigned long nr_unshown;
618 * Allow a burst of 60 reports, then keep quiet for that minute;
619 * or allow a steady drip of one report per second.
621 if (nr_shown == 60) {
622 if (time_before(jiffies, resume)) {
628 "BUG: Bad page state: %lu messages suppressed\n",
635 resume = jiffies + 60 * HZ;
637 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
638 current->comm, page_to_pfn(page));
639 dump_page(page, reason);
644 /* Leave bad fields for debug, except PageBuddy could make trouble */
645 page_mapcount_reset(page); /* remove PageBuddy */
646 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
649 static inline unsigned int order_to_pindex(int migratetype, int order)
653 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
654 if (order > PAGE_ALLOC_COSTLY_ORDER) {
655 VM_BUG_ON(order != pageblock_order);
656 base = PAGE_ALLOC_COSTLY_ORDER + 1;
659 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
662 return (MIGRATE_PCPTYPES * base) + migratetype;
665 static inline int pindex_to_order(unsigned int pindex)
667 int order = pindex / MIGRATE_PCPTYPES;
669 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
670 if (order > PAGE_ALLOC_COSTLY_ORDER)
671 order = pageblock_order;
673 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
679 static inline bool pcp_allowed_order(unsigned int order)
681 if (order <= PAGE_ALLOC_COSTLY_ORDER)
683 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
684 if (order == pageblock_order)
690 static inline void free_the_page(struct page *page, unsigned int order)
692 if (pcp_allowed_order(order)) /* Via pcp? */
693 free_unref_page(page, order);
695 __free_pages_ok(page, order, FPI_NONE);
699 * Higher-order pages are called "compound pages". They are structured thusly:
701 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
703 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
704 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
706 * The first tail page's ->compound_dtor holds the offset in array of compound
707 * page destructors. See compound_page_dtors.
709 * The first tail page's ->compound_order holds the order of allocation.
710 * This usage means that zero-order pages may not be compound.
713 void free_compound_page(struct page *page)
715 mem_cgroup_uncharge(page_folio(page));
716 free_the_page(page, compound_order(page));
719 static void prep_compound_head(struct page *page, unsigned int order)
721 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
722 set_compound_order(page, order);
723 atomic_set(compound_mapcount_ptr(page), -1);
724 atomic_set(compound_pincount_ptr(page), 0);
727 static void prep_compound_tail(struct page *head, int tail_idx)
729 struct page *p = head + tail_idx;
731 p->mapping = TAIL_MAPPING;
732 set_compound_head(p, head);
735 void prep_compound_page(struct page *page, unsigned int order)
738 int nr_pages = 1 << order;
741 for (i = 1; i < nr_pages; i++)
742 prep_compound_tail(page, i);
744 prep_compound_head(page, order);
747 #ifdef CONFIG_DEBUG_PAGEALLOC
748 unsigned int _debug_guardpage_minorder;
750 bool _debug_pagealloc_enabled_early __read_mostly
751 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
752 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
753 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
754 EXPORT_SYMBOL(_debug_pagealloc_enabled);
756 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
758 static int __init early_debug_pagealloc(char *buf)
760 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
762 early_param("debug_pagealloc", early_debug_pagealloc);
764 static int __init debug_guardpage_minorder_setup(char *buf)
768 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
769 pr_err("Bad debug_guardpage_minorder value\n");
772 _debug_guardpage_minorder = res;
773 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
776 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
778 static inline bool set_page_guard(struct zone *zone, struct page *page,
779 unsigned int order, int migratetype)
781 if (!debug_guardpage_enabled())
784 if (order >= debug_guardpage_minorder())
787 __SetPageGuard(page);
788 INIT_LIST_HEAD(&page->lru);
789 set_page_private(page, order);
790 /* Guard pages are not available for any usage */
791 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
796 static inline void clear_page_guard(struct zone *zone, struct page *page,
797 unsigned int order, int migratetype)
799 if (!debug_guardpage_enabled())
802 __ClearPageGuard(page);
804 set_page_private(page, 0);
805 if (!is_migrate_isolate(migratetype))
806 __mod_zone_freepage_state(zone, (1 << order), migratetype);
809 static inline bool set_page_guard(struct zone *zone, struct page *page,
810 unsigned int order, int migratetype) { return false; }
811 static inline void clear_page_guard(struct zone *zone, struct page *page,
812 unsigned int order, int migratetype) {}
816 * Enable static keys related to various memory debugging and hardening options.
817 * Some override others, and depend on early params that are evaluated in the
818 * order of appearance. So we need to first gather the full picture of what was
819 * enabled, and then make decisions.
821 void init_mem_debugging_and_hardening(void)
823 bool page_poisoning_requested = false;
825 #ifdef CONFIG_PAGE_POISONING
827 * Page poisoning is debug page alloc for some arches. If
828 * either of those options are enabled, enable poisoning.
830 if (page_poisoning_enabled() ||
831 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
832 debug_pagealloc_enabled())) {
833 static_branch_enable(&_page_poisoning_enabled);
834 page_poisoning_requested = true;
838 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
839 page_poisoning_requested) {
840 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
841 "will take precedence over init_on_alloc and init_on_free\n");
842 _init_on_alloc_enabled_early = false;
843 _init_on_free_enabled_early = false;
846 if (_init_on_alloc_enabled_early)
847 static_branch_enable(&init_on_alloc);
849 static_branch_disable(&init_on_alloc);
851 if (_init_on_free_enabled_early)
852 static_branch_enable(&init_on_free);
854 static_branch_disable(&init_on_free);
856 #ifdef CONFIG_DEBUG_PAGEALLOC
857 if (!debug_pagealloc_enabled())
860 static_branch_enable(&_debug_pagealloc_enabled);
862 if (!debug_guardpage_minorder())
865 static_branch_enable(&_debug_guardpage_enabled);
869 static inline void set_buddy_order(struct page *page, unsigned int order)
871 set_page_private(page, order);
872 __SetPageBuddy(page);
875 #ifdef CONFIG_COMPACTION
876 static inline struct capture_control *task_capc(struct zone *zone)
878 struct capture_control *capc = current->capture_control;
880 return unlikely(capc) &&
881 !(current->flags & PF_KTHREAD) &&
883 capc->cc->zone == zone ? capc : NULL;
887 compaction_capture(struct capture_control *capc, struct page *page,
888 int order, int migratetype)
890 if (!capc || order != capc->cc->order)
893 /* Do not accidentally pollute CMA or isolated regions*/
894 if (is_migrate_cma(migratetype) ||
895 is_migrate_isolate(migratetype))
899 * Do not let lower order allocations pollute a movable pageblock.
900 * This might let an unmovable request use a reclaimable pageblock
901 * and vice-versa but no more than normal fallback logic which can
902 * have trouble finding a high-order free page.
904 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
912 static inline struct capture_control *task_capc(struct zone *zone)
918 compaction_capture(struct capture_control *capc, struct page *page,
919 int order, int migratetype)
923 #endif /* CONFIG_COMPACTION */
925 /* Used for pages not on another list */
926 static inline void add_to_free_list(struct page *page, struct zone *zone,
927 unsigned int order, int migratetype)
929 struct free_area *area = &zone->free_area[order];
931 list_add(&page->lru, &area->free_list[migratetype]);
935 /* Used for pages not on another list */
936 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
937 unsigned int order, int migratetype)
939 struct free_area *area = &zone->free_area[order];
941 list_add_tail(&page->lru, &area->free_list[migratetype]);
946 * Used for pages which are on another list. Move the pages to the tail
947 * of the list - so the moved pages won't immediately be considered for
948 * allocation again (e.g., optimization for memory onlining).
950 static inline void move_to_free_list(struct page *page, struct zone *zone,
951 unsigned int order, int migratetype)
953 struct free_area *area = &zone->free_area[order];
955 list_move_tail(&page->lru, &area->free_list[migratetype]);
958 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
961 /* clear reported state and update reported page count */
962 if (page_reported(page))
963 __ClearPageReported(page);
965 list_del(&page->lru);
966 __ClearPageBuddy(page);
967 set_page_private(page, 0);
968 zone->free_area[order].nr_free--;
972 * If this is not the largest possible page, check if the buddy
973 * of the next-highest order is free. If it is, it's possible
974 * that pages are being freed that will coalesce soon. In case,
975 * that is happening, add the free page to the tail of the list
976 * so it's less likely to be used soon and more likely to be merged
977 * as a higher order page
980 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
981 struct page *page, unsigned int order)
983 unsigned long higher_page_pfn;
984 struct page *higher_page;
986 if (order >= MAX_ORDER - 2)
989 higher_page_pfn = buddy_pfn & pfn;
990 higher_page = page + (higher_page_pfn - pfn);
992 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
997 * Freeing function for a buddy system allocator.
999 * The concept of a buddy system is to maintain direct-mapped table
1000 * (containing bit values) for memory blocks of various "orders".
1001 * The bottom level table contains the map for the smallest allocatable
1002 * units of memory (here, pages), and each level above it describes
1003 * pairs of units from the levels below, hence, "buddies".
1004 * At a high level, all that happens here is marking the table entry
1005 * at the bottom level available, and propagating the changes upward
1006 * as necessary, plus some accounting needed to play nicely with other
1007 * parts of the VM system.
1008 * At each level, we keep a list of pages, which are heads of continuous
1009 * free pages of length of (1 << order) and marked with PageBuddy.
1010 * Page's order is recorded in page_private(page) field.
1011 * So when we are allocating or freeing one, we can derive the state of the
1012 * other. That is, if we allocate a small block, and both were
1013 * free, the remainder of the region must be split into blocks.
1014 * If a block is freed, and its buddy is also free, then this
1015 * triggers coalescing into a block of larger size.
1020 static inline void __free_one_page(struct page *page,
1022 struct zone *zone, unsigned int order,
1023 int migratetype, fpi_t fpi_flags)
1025 struct capture_control *capc = task_capc(zone);
1026 unsigned long buddy_pfn;
1027 unsigned long combined_pfn;
1031 VM_BUG_ON(!zone_is_initialized(zone));
1032 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1034 VM_BUG_ON(migratetype == -1);
1035 if (likely(!is_migrate_isolate(migratetype)))
1036 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1038 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1039 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1041 while (order < MAX_ORDER - 1) {
1042 if (compaction_capture(capc, page, order, migratetype)) {
1043 __mod_zone_freepage_state(zone, -(1 << order),
1048 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
1052 if (unlikely(order >= pageblock_order)) {
1054 * We want to prevent merge between freepages on pageblock
1055 * without fallbacks and normal pageblock. Without this,
1056 * pageblock isolation could cause incorrect freepage or CMA
1057 * accounting or HIGHATOMIC accounting.
1059 int buddy_mt = get_pageblock_migratetype(buddy);
1061 if (migratetype != buddy_mt
1062 && (!migratetype_is_mergeable(migratetype) ||
1063 !migratetype_is_mergeable(buddy_mt)))
1068 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1069 * merge with it and move up one order.
1071 if (page_is_guard(buddy))
1072 clear_page_guard(zone, buddy, order, migratetype);
1074 del_page_from_free_list(buddy, zone, order);
1075 combined_pfn = buddy_pfn & pfn;
1076 page = page + (combined_pfn - pfn);
1082 set_buddy_order(page, order);
1084 if (fpi_flags & FPI_TO_TAIL)
1086 else if (is_shuffle_order(order))
1087 to_tail = shuffle_pick_tail();
1089 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1092 add_to_free_list_tail(page, zone, order, migratetype);
1094 add_to_free_list(page, zone, order, migratetype);
1096 /* Notify page reporting subsystem of freed page */
1097 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1098 page_reporting_notify_free(order);
1102 * split_free_page() -- split a free page at split_pfn_offset
1103 * @free_page: the original free page
1104 * @order: the order of the page
1105 * @split_pfn_offset: split offset within the page
1107 * Return -ENOENT if the free page is changed, otherwise 0
1109 * It is used when the free page crosses two pageblocks with different migratetypes
1110 * at split_pfn_offset within the page. The split free page will be put into
1111 * separate migratetype lists afterwards. Otherwise, the function achieves
1114 int split_free_page(struct page *free_page,
1115 unsigned int order, unsigned long split_pfn_offset)
1117 struct zone *zone = page_zone(free_page);
1118 unsigned long free_page_pfn = page_to_pfn(free_page);
1120 unsigned long flags;
1121 int free_page_order;
1125 if (split_pfn_offset == 0)
1128 spin_lock_irqsave(&zone->lock, flags);
1130 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
1135 mt = get_pageblock_migratetype(free_page);
1136 if (likely(!is_migrate_isolate(mt)))
1137 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1139 del_page_from_free_list(free_page, zone, order);
1140 for (pfn = free_page_pfn;
1141 pfn < free_page_pfn + (1UL << order);) {
1142 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
1144 free_page_order = min_t(unsigned int,
1145 pfn ? __ffs(pfn) : order,
1146 __fls(split_pfn_offset));
1147 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
1149 pfn += 1UL << free_page_order;
1150 split_pfn_offset -= (1UL << free_page_order);
1151 /* we have done the first part, now switch to second part */
1152 if (split_pfn_offset == 0)
1153 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
1156 spin_unlock_irqrestore(&zone->lock, flags);
1160 * A bad page could be due to a number of fields. Instead of multiple branches,
1161 * try and check multiple fields with one check. The caller must do a detailed
1162 * check if necessary.
1164 static inline bool page_expected_state(struct page *page,
1165 unsigned long check_flags)
1167 if (unlikely(atomic_read(&page->_mapcount) != -1))
1170 if (unlikely((unsigned long)page->mapping |
1171 page_ref_count(page) |
1175 (page->flags & check_flags)))
1181 static const char *page_bad_reason(struct page *page, unsigned long flags)
1183 const char *bad_reason = NULL;
1185 if (unlikely(atomic_read(&page->_mapcount) != -1))
1186 bad_reason = "nonzero mapcount";
1187 if (unlikely(page->mapping != NULL))
1188 bad_reason = "non-NULL mapping";
1189 if (unlikely(page_ref_count(page) != 0))
1190 bad_reason = "nonzero _refcount";
1191 if (unlikely(page->flags & flags)) {
1192 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1193 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1195 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1198 if (unlikely(page->memcg_data))
1199 bad_reason = "page still charged to cgroup";
1204 static void check_free_page_bad(struct page *page)
1207 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1210 static inline int check_free_page(struct page *page)
1212 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1215 /* Something has gone sideways, find it */
1216 check_free_page_bad(page);
1220 static int free_tail_pages_check(struct page *head_page, struct page *page)
1225 * We rely page->lru.next never has bit 0 set, unless the page
1226 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1228 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1230 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1234 switch (page - head_page) {
1236 /* the first tail page: ->mapping may be compound_mapcount() */
1237 if (unlikely(compound_mapcount(page))) {
1238 bad_page(page, "nonzero compound_mapcount");
1244 * the second tail page: ->mapping is
1245 * deferred_list.next -- ignore value.
1249 if (page->mapping != TAIL_MAPPING) {
1250 bad_page(page, "corrupted mapping in tail page");
1255 if (unlikely(!PageTail(page))) {
1256 bad_page(page, "PageTail not set");
1259 if (unlikely(compound_head(page) != head_page)) {
1260 bad_page(page, "compound_head not consistent");
1265 page->mapping = NULL;
1266 clear_compound_head(page);
1271 * Skip KASAN memory poisoning when either:
1273 * 1. Deferred memory initialization has not yet completed,
1274 * see the explanation below.
1275 * 2. Skipping poisoning is requested via FPI_SKIP_KASAN_POISON,
1276 * see the comment next to it.
1277 * 3. Skipping poisoning is requested via __GFP_SKIP_KASAN_POISON,
1278 * see the comment next to it.
1280 * Poisoning pages during deferred memory init will greatly lengthen the
1281 * process and cause problem in large memory systems as the deferred pages
1282 * initialization is done with interrupt disabled.
1284 * Assuming that there will be no reference to those newly initialized
1285 * pages before they are ever allocated, this should have no effect on
1286 * KASAN memory tracking as the poison will be properly inserted at page
1287 * allocation time. The only corner case is when pages are allocated by
1288 * on-demand allocation and then freed again before the deferred pages
1289 * initialization is done, but this is not likely to happen.
1291 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1293 return deferred_pages_enabled() ||
1294 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
1295 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
1296 PageSkipKASanPoison(page);
1299 static void kernel_init_pages(struct page *page, int numpages)
1303 /* s390's use of memset() could override KASAN redzones. */
1304 kasan_disable_current();
1305 for (i = 0; i < numpages; i++)
1306 clear_highpage_kasan_tagged(page + i);
1307 kasan_enable_current();
1310 static __always_inline bool free_pages_prepare(struct page *page,
1311 unsigned int order, bool check_free, fpi_t fpi_flags)
1314 bool init = want_init_on_free();
1316 VM_BUG_ON_PAGE(PageTail(page), page);
1318 trace_mm_page_free(page, order);
1320 if (unlikely(PageHWPoison(page)) && !order) {
1322 * Do not let hwpoison pages hit pcplists/buddy
1323 * Untie memcg state and reset page's owner
1325 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1326 __memcg_kmem_uncharge_page(page, order);
1327 reset_page_owner(page, order);
1328 page_table_check_free(page, order);
1333 * Check tail pages before head page information is cleared to
1334 * avoid checking PageCompound for order-0 pages.
1336 if (unlikely(order)) {
1337 bool compound = PageCompound(page);
1340 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1343 ClearPageDoubleMap(page);
1344 ClearPageHasHWPoisoned(page);
1346 for (i = 1; i < (1 << order); i++) {
1348 bad += free_tail_pages_check(page, page + i);
1349 if (unlikely(check_free_page(page + i))) {
1353 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1356 if (PageMappingFlags(page))
1357 page->mapping = NULL;
1358 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1359 __memcg_kmem_uncharge_page(page, order);
1361 bad += check_free_page(page);
1365 page_cpupid_reset_last(page);
1366 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1367 reset_page_owner(page, order);
1368 page_table_check_free(page, order);
1370 if (!PageHighMem(page)) {
1371 debug_check_no_locks_freed(page_address(page),
1372 PAGE_SIZE << order);
1373 debug_check_no_obj_freed(page_address(page),
1374 PAGE_SIZE << order);
1377 kernel_poison_pages(page, 1 << order);
1380 * As memory initialization might be integrated into KASAN,
1381 * KASAN poisoning and memory initialization code must be
1382 * kept together to avoid discrepancies in behavior.
1384 * With hardware tag-based KASAN, memory tags must be set before the
1385 * page becomes unavailable via debug_pagealloc or arch_free_page.
1387 if (!should_skip_kasan_poison(page, fpi_flags)) {
1388 kasan_poison_pages(page, order, init);
1390 /* Memory is already initialized if KASAN did it internally. */
1391 if (kasan_has_integrated_init())
1395 kernel_init_pages(page, 1 << order);
1398 * arch_free_page() can make the page's contents inaccessible. s390
1399 * does this. So nothing which can access the page's contents should
1400 * happen after this.
1402 arch_free_page(page, order);
1404 debug_pagealloc_unmap_pages(page, 1 << order);
1409 #ifdef CONFIG_DEBUG_VM
1411 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1412 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1413 * moved from pcp lists to free lists.
1415 static bool free_pcp_prepare(struct page *page, unsigned int order)
1417 return free_pages_prepare(page, order, true, FPI_NONE);
1420 static bool bulkfree_pcp_prepare(struct page *page)
1422 if (debug_pagealloc_enabled_static())
1423 return check_free_page(page);
1429 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1430 * moving from pcp lists to free list in order to reduce overhead. With
1431 * debug_pagealloc enabled, they are checked also immediately when being freed
1434 static bool free_pcp_prepare(struct page *page, unsigned int order)
1436 if (debug_pagealloc_enabled_static())
1437 return free_pages_prepare(page, order, true, FPI_NONE);
1439 return free_pages_prepare(page, order, false, FPI_NONE);
1442 static bool bulkfree_pcp_prepare(struct page *page)
1444 return check_free_page(page);
1446 #endif /* CONFIG_DEBUG_VM */
1449 * Frees a number of pages from the PCP lists
1450 * Assumes all pages on list are in same zone.
1451 * count is the number of pages to free.
1453 static void free_pcppages_bulk(struct zone *zone, int count,
1454 struct per_cpu_pages *pcp,
1458 int max_pindex = NR_PCP_LISTS - 1;
1460 bool isolated_pageblocks;
1464 * Ensure proper count is passed which otherwise would stuck in the
1465 * below while (list_empty(list)) loop.
1467 count = min(pcp->count, count);
1469 /* Ensure requested pindex is drained first. */
1470 pindex = pindex - 1;
1473 * local_lock_irq held so equivalent to spin_lock_irqsave for
1474 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1476 spin_lock(&zone->lock);
1477 isolated_pageblocks = has_isolate_pageblock(zone);
1480 struct list_head *list;
1483 /* Remove pages from lists in a round-robin fashion. */
1485 if (++pindex > max_pindex)
1486 pindex = min_pindex;
1487 list = &pcp->lists[pindex];
1488 if (!list_empty(list))
1491 if (pindex == max_pindex)
1493 if (pindex == min_pindex)
1497 order = pindex_to_order(pindex);
1498 nr_pages = 1 << order;
1499 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1503 page = list_last_entry(list, struct page, lru);
1504 mt = get_pcppage_migratetype(page);
1506 /* must delete to avoid corrupting pcp list */
1507 list_del(&page->lru);
1509 pcp->count -= nr_pages;
1511 if (bulkfree_pcp_prepare(page))
1514 /* MIGRATE_ISOLATE page should not go to pcplists */
1515 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1516 /* Pageblock could have been isolated meanwhile */
1517 if (unlikely(isolated_pageblocks))
1518 mt = get_pageblock_migratetype(page);
1520 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1521 trace_mm_page_pcpu_drain(page, order, mt);
1522 } while (count > 0 && !list_empty(list));
1525 spin_unlock(&zone->lock);
1528 static void free_one_page(struct zone *zone,
1529 struct page *page, unsigned long pfn,
1531 int migratetype, fpi_t fpi_flags)
1533 unsigned long flags;
1535 spin_lock_irqsave(&zone->lock, flags);
1536 if (unlikely(has_isolate_pageblock(zone) ||
1537 is_migrate_isolate(migratetype))) {
1538 migratetype = get_pfnblock_migratetype(page, pfn);
1540 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1541 spin_unlock_irqrestore(&zone->lock, flags);
1544 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1545 unsigned long zone, int nid)
1547 mm_zero_struct_page(page);
1548 set_page_links(page, zone, nid, pfn);
1549 init_page_count(page);
1550 page_mapcount_reset(page);
1551 page_cpupid_reset_last(page);
1552 page_kasan_tag_reset(page);
1554 INIT_LIST_HEAD(&page->lru);
1555 #ifdef WANT_PAGE_VIRTUAL
1556 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1557 if (!is_highmem_idx(zone))
1558 set_page_address(page, __va(pfn << PAGE_SHIFT));
1562 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1563 static void __meminit init_reserved_page(unsigned long pfn)
1568 if (!early_page_uninitialised(pfn))
1571 nid = early_pfn_to_nid(pfn);
1572 pgdat = NODE_DATA(nid);
1574 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1575 struct zone *zone = &pgdat->node_zones[zid];
1577 if (zone_spans_pfn(zone, pfn))
1580 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1583 static inline void init_reserved_page(unsigned long pfn)
1586 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1589 * Initialised pages do not have PageReserved set. This function is
1590 * called for each range allocated by the bootmem allocator and
1591 * marks the pages PageReserved. The remaining valid pages are later
1592 * sent to the buddy page allocator.
1594 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1596 unsigned long start_pfn = PFN_DOWN(start);
1597 unsigned long end_pfn = PFN_UP(end);
1599 for (; start_pfn < end_pfn; start_pfn++) {
1600 if (pfn_valid(start_pfn)) {
1601 struct page *page = pfn_to_page(start_pfn);
1603 init_reserved_page(start_pfn);
1605 /* Avoid false-positive PageTail() */
1606 INIT_LIST_HEAD(&page->lru);
1609 * no need for atomic set_bit because the struct
1610 * page is not visible yet so nobody should
1613 __SetPageReserved(page);
1618 static void __free_pages_ok(struct page *page, unsigned int order,
1621 unsigned long flags;
1623 unsigned long pfn = page_to_pfn(page);
1624 struct zone *zone = page_zone(page);
1626 if (!free_pages_prepare(page, order, true, fpi_flags))
1629 migratetype = get_pfnblock_migratetype(page, pfn);
1631 spin_lock_irqsave(&zone->lock, flags);
1632 if (unlikely(has_isolate_pageblock(zone) ||
1633 is_migrate_isolate(migratetype))) {
1634 migratetype = get_pfnblock_migratetype(page, pfn);
1636 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1637 spin_unlock_irqrestore(&zone->lock, flags);
1639 __count_vm_events(PGFREE, 1 << order);
1642 void __free_pages_core(struct page *page, unsigned int order)
1644 unsigned int nr_pages = 1 << order;
1645 struct page *p = page;
1649 * When initializing the memmap, __init_single_page() sets the refcount
1650 * of all pages to 1 ("allocated"/"not free"). We have to set the
1651 * refcount of all involved pages to 0.
1654 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1656 __ClearPageReserved(p);
1657 set_page_count(p, 0);
1659 __ClearPageReserved(p);
1660 set_page_count(p, 0);
1662 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1665 * Bypass PCP and place fresh pages right to the tail, primarily
1666 * relevant for memory onlining.
1668 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1674 * During memory init memblocks map pfns to nids. The search is expensive and
1675 * this caches recent lookups. The implementation of __early_pfn_to_nid
1676 * treats start/end as pfns.
1678 struct mminit_pfnnid_cache {
1679 unsigned long last_start;
1680 unsigned long last_end;
1684 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1687 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1689 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1690 struct mminit_pfnnid_cache *state)
1692 unsigned long start_pfn, end_pfn;
1695 if (state->last_start <= pfn && pfn < state->last_end)
1696 return state->last_nid;
1698 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1699 if (nid != NUMA_NO_NODE) {
1700 state->last_start = start_pfn;
1701 state->last_end = end_pfn;
1702 state->last_nid = nid;
1708 int __meminit early_pfn_to_nid(unsigned long pfn)
1710 static DEFINE_SPINLOCK(early_pfn_lock);
1713 spin_lock(&early_pfn_lock);
1714 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1716 nid = first_online_node;
1717 spin_unlock(&early_pfn_lock);
1721 #endif /* CONFIG_NUMA */
1723 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1726 if (early_page_uninitialised(pfn))
1728 __free_pages_core(page, order);
1732 * Check that the whole (or subset of) a pageblock given by the interval of
1733 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1734 * with the migration of free compaction scanner.
1736 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1738 * It's possible on some configurations to have a setup like node0 node1 node0
1739 * i.e. it's possible that all pages within a zones range of pages do not
1740 * belong to a single zone. We assume that a border between node0 and node1
1741 * can occur within a single pageblock, but not a node0 node1 node0
1742 * interleaving within a single pageblock. It is therefore sufficient to check
1743 * the first and last page of a pageblock and avoid checking each individual
1744 * page in a pageblock.
1746 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1747 unsigned long end_pfn, struct zone *zone)
1749 struct page *start_page;
1750 struct page *end_page;
1752 /* end_pfn is one past the range we are checking */
1755 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1758 start_page = pfn_to_online_page(start_pfn);
1762 if (page_zone(start_page) != zone)
1765 end_page = pfn_to_page(end_pfn);
1767 /* This gives a shorter code than deriving page_zone(end_page) */
1768 if (page_zone_id(start_page) != page_zone_id(end_page))
1774 void set_zone_contiguous(struct zone *zone)
1776 unsigned long block_start_pfn = zone->zone_start_pfn;
1777 unsigned long block_end_pfn;
1779 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1780 for (; block_start_pfn < zone_end_pfn(zone);
1781 block_start_pfn = block_end_pfn,
1782 block_end_pfn += pageblock_nr_pages) {
1784 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1786 if (!__pageblock_pfn_to_page(block_start_pfn,
1787 block_end_pfn, zone))
1792 /* We confirm that there is no hole */
1793 zone->contiguous = true;
1796 void clear_zone_contiguous(struct zone *zone)
1798 zone->contiguous = false;
1801 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1802 static void __init deferred_free_range(unsigned long pfn,
1803 unsigned long nr_pages)
1811 page = pfn_to_page(pfn);
1813 /* Free a large naturally-aligned chunk if possible */
1814 if (nr_pages == pageblock_nr_pages &&
1815 (pfn & (pageblock_nr_pages - 1)) == 0) {
1816 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1817 __free_pages_core(page, pageblock_order);
1821 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1822 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1823 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1824 __free_pages_core(page, 0);
1828 /* Completion tracking for deferred_init_memmap() threads */
1829 static atomic_t pgdat_init_n_undone __initdata;
1830 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1832 static inline void __init pgdat_init_report_one_done(void)
1834 if (atomic_dec_and_test(&pgdat_init_n_undone))
1835 complete(&pgdat_init_all_done_comp);
1839 * Returns true if page needs to be initialized or freed to buddy allocator.
1841 * First we check if pfn is valid on architectures where it is possible to have
1842 * holes within pageblock_nr_pages. On systems where it is not possible, this
1843 * function is optimized out.
1845 * Then, we check if a current large page is valid by only checking the validity
1848 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1850 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1856 * Free pages to buddy allocator. Try to free aligned pages in
1857 * pageblock_nr_pages sizes.
1859 static void __init deferred_free_pages(unsigned long pfn,
1860 unsigned long end_pfn)
1862 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1863 unsigned long nr_free = 0;
1865 for (; pfn < end_pfn; pfn++) {
1866 if (!deferred_pfn_valid(pfn)) {
1867 deferred_free_range(pfn - nr_free, nr_free);
1869 } else if (!(pfn & nr_pgmask)) {
1870 deferred_free_range(pfn - nr_free, nr_free);
1876 /* Free the last block of pages to allocator */
1877 deferred_free_range(pfn - nr_free, nr_free);
1881 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1882 * by performing it only once every pageblock_nr_pages.
1883 * Return number of pages initialized.
1885 static unsigned long __init deferred_init_pages(struct zone *zone,
1887 unsigned long end_pfn)
1889 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1890 int nid = zone_to_nid(zone);
1891 unsigned long nr_pages = 0;
1892 int zid = zone_idx(zone);
1893 struct page *page = NULL;
1895 for (; pfn < end_pfn; pfn++) {
1896 if (!deferred_pfn_valid(pfn)) {
1899 } else if (!page || !(pfn & nr_pgmask)) {
1900 page = pfn_to_page(pfn);
1904 __init_single_page(page, pfn, zid, nid);
1911 * This function is meant to pre-load the iterator for the zone init.
1912 * Specifically it walks through the ranges until we are caught up to the
1913 * first_init_pfn value and exits there. If we never encounter the value we
1914 * return false indicating there are no valid ranges left.
1917 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1918 unsigned long *spfn, unsigned long *epfn,
1919 unsigned long first_init_pfn)
1924 * Start out by walking through the ranges in this zone that have
1925 * already been initialized. We don't need to do anything with them
1926 * so we just need to flush them out of the system.
1928 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1929 if (*epfn <= first_init_pfn)
1931 if (*spfn < first_init_pfn)
1932 *spfn = first_init_pfn;
1941 * Initialize and free pages. We do it in two loops: first we initialize
1942 * struct page, then free to buddy allocator, because while we are
1943 * freeing pages we can access pages that are ahead (computing buddy
1944 * page in __free_one_page()).
1946 * In order to try and keep some memory in the cache we have the loop
1947 * broken along max page order boundaries. This way we will not cause
1948 * any issues with the buddy page computation.
1950 static unsigned long __init
1951 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1952 unsigned long *end_pfn)
1954 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1955 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1956 unsigned long nr_pages = 0;
1959 /* First we loop through and initialize the page values */
1960 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1963 if (mo_pfn <= *start_pfn)
1966 t = min(mo_pfn, *end_pfn);
1967 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1969 if (mo_pfn < *end_pfn) {
1970 *start_pfn = mo_pfn;
1975 /* Reset values and now loop through freeing pages as needed */
1978 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1984 t = min(mo_pfn, epfn);
1985 deferred_free_pages(spfn, t);
1995 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1998 unsigned long spfn, epfn;
1999 struct zone *zone = arg;
2002 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2005 * Initialize and free pages in MAX_ORDER sized increments so that we
2006 * can avoid introducing any issues with the buddy allocator.
2008 while (spfn < end_pfn) {
2009 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2014 /* An arch may override for more concurrency. */
2016 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2021 /* Initialise remaining memory on a node */
2022 static int __init deferred_init_memmap(void *data)
2024 pg_data_t *pgdat = data;
2025 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2026 unsigned long spfn = 0, epfn = 0;
2027 unsigned long first_init_pfn, flags;
2028 unsigned long start = jiffies;
2030 int zid, max_threads;
2033 /* Bind memory initialisation thread to a local node if possible */
2034 if (!cpumask_empty(cpumask))
2035 set_cpus_allowed_ptr(current, cpumask);
2037 pgdat_resize_lock(pgdat, &flags);
2038 first_init_pfn = pgdat->first_deferred_pfn;
2039 if (first_init_pfn == ULONG_MAX) {
2040 pgdat_resize_unlock(pgdat, &flags);
2041 pgdat_init_report_one_done();
2045 /* Sanity check boundaries */
2046 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2047 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2048 pgdat->first_deferred_pfn = ULONG_MAX;
2051 * Once we unlock here, the zone cannot be grown anymore, thus if an
2052 * interrupt thread must allocate this early in boot, zone must be
2053 * pre-grown prior to start of deferred page initialization.
2055 pgdat_resize_unlock(pgdat, &flags);
2057 /* Only the highest zone is deferred so find it */
2058 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2059 zone = pgdat->node_zones + zid;
2060 if (first_init_pfn < zone_end_pfn(zone))
2064 /* If the zone is empty somebody else may have cleared out the zone */
2065 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2069 max_threads = deferred_page_init_max_threads(cpumask);
2071 while (spfn < epfn) {
2072 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2073 struct padata_mt_job job = {
2074 .thread_fn = deferred_init_memmap_chunk,
2077 .size = epfn_align - spfn,
2078 .align = PAGES_PER_SECTION,
2079 .min_chunk = PAGES_PER_SECTION,
2080 .max_threads = max_threads,
2083 padata_do_multithreaded(&job);
2084 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2088 /* Sanity check that the next zone really is unpopulated */
2089 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2091 pr_info("node %d deferred pages initialised in %ums\n",
2092 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2094 pgdat_init_report_one_done();
2099 * If this zone has deferred pages, try to grow it by initializing enough
2100 * deferred pages to satisfy the allocation specified by order, rounded up to
2101 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2102 * of SECTION_SIZE bytes by initializing struct pages in increments of
2103 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2105 * Return true when zone was grown, otherwise return false. We return true even
2106 * when we grow less than requested, to let the caller decide if there are
2107 * enough pages to satisfy the allocation.
2109 * Note: We use noinline because this function is needed only during boot, and
2110 * it is called from a __ref function _deferred_grow_zone. This way we are
2111 * making sure that it is not inlined into permanent text section.
2113 static noinline bool __init
2114 deferred_grow_zone(struct zone *zone, unsigned int order)
2116 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2117 pg_data_t *pgdat = zone->zone_pgdat;
2118 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2119 unsigned long spfn, epfn, flags;
2120 unsigned long nr_pages = 0;
2123 /* Only the last zone may have deferred pages */
2124 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2127 pgdat_resize_lock(pgdat, &flags);
2130 * If someone grew this zone while we were waiting for spinlock, return
2131 * true, as there might be enough pages already.
2133 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2134 pgdat_resize_unlock(pgdat, &flags);
2138 /* If the zone is empty somebody else may have cleared out the zone */
2139 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2140 first_deferred_pfn)) {
2141 pgdat->first_deferred_pfn = ULONG_MAX;
2142 pgdat_resize_unlock(pgdat, &flags);
2143 /* Retry only once. */
2144 return first_deferred_pfn != ULONG_MAX;
2148 * Initialize and free pages in MAX_ORDER sized increments so
2149 * that we can avoid introducing any issues with the buddy
2152 while (spfn < epfn) {
2153 /* update our first deferred PFN for this section */
2154 first_deferred_pfn = spfn;
2156 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2157 touch_nmi_watchdog();
2159 /* We should only stop along section boundaries */
2160 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2163 /* If our quota has been met we can stop here */
2164 if (nr_pages >= nr_pages_needed)
2168 pgdat->first_deferred_pfn = spfn;
2169 pgdat_resize_unlock(pgdat, &flags);
2171 return nr_pages > 0;
2175 * deferred_grow_zone() is __init, but it is called from
2176 * get_page_from_freelist() during early boot until deferred_pages permanently
2177 * disables this call. This is why we have refdata wrapper to avoid warning,
2178 * and to ensure that the function body gets unloaded.
2181 _deferred_grow_zone(struct zone *zone, unsigned int order)
2183 return deferred_grow_zone(zone, order);
2186 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2188 void __init page_alloc_init_late(void)
2193 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2195 /* There will be num_node_state(N_MEMORY) threads */
2196 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2197 for_each_node_state(nid, N_MEMORY) {
2198 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2201 /* Block until all are initialised */
2202 wait_for_completion(&pgdat_init_all_done_comp);
2205 * We initialized the rest of the deferred pages. Permanently disable
2206 * on-demand struct page initialization.
2208 static_branch_disable(&deferred_pages);
2210 /* Reinit limits that are based on free pages after the kernel is up */
2211 files_maxfiles_init();
2216 /* Discard memblock private memory */
2219 for_each_node_state(nid, N_MEMORY)
2220 shuffle_free_memory(NODE_DATA(nid));
2222 for_each_populated_zone(zone)
2223 set_zone_contiguous(zone);
2227 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2228 void __init init_cma_reserved_pageblock(struct page *page)
2230 unsigned i = pageblock_nr_pages;
2231 struct page *p = page;
2234 __ClearPageReserved(p);
2235 set_page_count(p, 0);
2238 set_pageblock_migratetype(page, MIGRATE_CMA);
2239 set_page_refcounted(page);
2240 __free_pages(page, pageblock_order);
2242 adjust_managed_page_count(page, pageblock_nr_pages);
2243 page_zone(page)->cma_pages += pageblock_nr_pages;
2248 * The order of subdivision here is critical for the IO subsystem.
2249 * Please do not alter this order without good reasons and regression
2250 * testing. Specifically, as large blocks of memory are subdivided,
2251 * the order in which smaller blocks are delivered depends on the order
2252 * they're subdivided in this function. This is the primary factor
2253 * influencing the order in which pages are delivered to the IO
2254 * subsystem according to empirical testing, and this is also justified
2255 * by considering the behavior of a buddy system containing a single
2256 * large block of memory acted on by a series of small allocations.
2257 * This behavior is a critical factor in sglist merging's success.
2261 static inline void expand(struct zone *zone, struct page *page,
2262 int low, int high, int migratetype)
2264 unsigned long size = 1 << high;
2266 while (high > low) {
2269 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2272 * Mark as guard pages (or page), that will allow to
2273 * merge back to allocator when buddy will be freed.
2274 * Corresponding page table entries will not be touched,
2275 * pages will stay not present in virtual address space
2277 if (set_page_guard(zone, &page[size], high, migratetype))
2280 add_to_free_list(&page[size], zone, high, migratetype);
2281 set_buddy_order(&page[size], high);
2285 static void check_new_page_bad(struct page *page)
2287 if (unlikely(page->flags & __PG_HWPOISON)) {
2288 /* Don't complain about hwpoisoned pages */
2289 page_mapcount_reset(page); /* remove PageBuddy */
2294 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2298 * This page is about to be returned from the page allocator
2300 static inline int check_new_page(struct page *page)
2302 if (likely(page_expected_state(page,
2303 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2306 check_new_page_bad(page);
2310 static bool check_new_pages(struct page *page, unsigned int order)
2313 for (i = 0; i < (1 << order); i++) {
2314 struct page *p = page + i;
2316 if (unlikely(check_new_page(p)))
2323 #ifdef CONFIG_DEBUG_VM
2325 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2326 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2327 * also checked when pcp lists are refilled from the free lists.
2329 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2331 if (debug_pagealloc_enabled_static())
2332 return check_new_pages(page, order);
2337 static inline bool check_new_pcp(struct page *page, unsigned int order)
2339 return check_new_pages(page, order);
2343 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2344 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2345 * enabled, they are also checked when being allocated from the pcp lists.
2347 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2349 return check_new_pages(page, order);
2351 static inline bool check_new_pcp(struct page *page, unsigned int order)
2353 if (debug_pagealloc_enabled_static())
2354 return check_new_pages(page, order);
2358 #endif /* CONFIG_DEBUG_VM */
2360 static inline bool should_skip_kasan_unpoison(gfp_t flags, bool init_tags)
2362 /* Don't skip if a software KASAN mode is enabled. */
2363 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
2364 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
2367 /* Skip, if hardware tag-based KASAN is not enabled. */
2368 if (!kasan_hw_tags_enabled())
2372 * With hardware tag-based KASAN enabled, skip if either:
2374 * 1. Memory tags have already been cleared via tag_clear_highpage().
2375 * 2. Skipping has been requested via __GFP_SKIP_KASAN_UNPOISON.
2377 return init_tags || (flags & __GFP_SKIP_KASAN_UNPOISON);
2380 static inline bool should_skip_init(gfp_t flags)
2382 /* Don't skip, if hardware tag-based KASAN is not enabled. */
2383 if (!kasan_hw_tags_enabled())
2386 /* For hardware tag-based KASAN, skip if requested. */
2387 return (flags & __GFP_SKIP_ZERO);
2390 inline void post_alloc_hook(struct page *page, unsigned int order,
2393 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
2394 !should_skip_init(gfp_flags);
2395 bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
2397 set_page_private(page, 0);
2398 set_page_refcounted(page);
2400 arch_alloc_page(page, order);
2401 debug_pagealloc_map_pages(page, 1 << order);
2404 * Page unpoisoning must happen before memory initialization.
2405 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2406 * allocations and the page unpoisoning code will complain.
2408 kernel_unpoison_pages(page, 1 << order);
2411 * As memory initialization might be integrated into KASAN,
2412 * KASAN unpoisoning and memory initializion code must be
2413 * kept together to avoid discrepancies in behavior.
2417 * If memory tags should be zeroed (which happens only when memory
2418 * should be initialized as well).
2423 /* Initialize both memory and tags. */
2424 for (i = 0; i != 1 << order; ++i)
2425 tag_clear_highpage(page + i);
2427 /* Note that memory is already initialized by the loop above. */
2430 if (!should_skip_kasan_unpoison(gfp_flags, init_tags)) {
2431 /* Unpoison shadow memory or set memory tags. */
2432 kasan_unpoison_pages(page, order, init);
2434 /* Note that memory is already initialized by KASAN. */
2435 if (kasan_has_integrated_init())
2438 /* If memory is still not initialized, do it now. */
2440 kernel_init_pages(page, 1 << order);
2441 /* Propagate __GFP_SKIP_KASAN_POISON to page flags. */
2442 if (kasan_hw_tags_enabled() && (gfp_flags & __GFP_SKIP_KASAN_POISON))
2443 SetPageSkipKASanPoison(page);
2445 set_page_owner(page, order, gfp_flags);
2446 page_table_check_alloc(page, order);
2449 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2450 unsigned int alloc_flags)
2452 post_alloc_hook(page, order, gfp_flags);
2454 if (order && (gfp_flags & __GFP_COMP))
2455 prep_compound_page(page, order);
2458 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2459 * allocate the page. The expectation is that the caller is taking
2460 * steps that will free more memory. The caller should avoid the page
2461 * being used for !PFMEMALLOC purposes.
2463 if (alloc_flags & ALLOC_NO_WATERMARKS)
2464 set_page_pfmemalloc(page);
2466 clear_page_pfmemalloc(page);
2470 * Go through the free lists for the given migratetype and remove
2471 * the smallest available page from the freelists
2473 static __always_inline
2474 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2477 unsigned int current_order;
2478 struct free_area *area;
2481 /* Find a page of the appropriate size in the preferred list */
2482 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2483 area = &(zone->free_area[current_order]);
2484 page = get_page_from_free_area(area, migratetype);
2487 del_page_from_free_list(page, zone, current_order);
2488 expand(zone, page, order, current_order, migratetype);
2489 set_pcppage_migratetype(page, migratetype);
2490 trace_mm_page_alloc_zone_locked(page, order, migratetype,
2491 pcp_allowed_order(order) &&
2492 migratetype < MIGRATE_PCPTYPES);
2501 * This array describes the order lists are fallen back to when
2502 * the free lists for the desirable migrate type are depleted
2504 * The other migratetypes do not have fallbacks.
2506 static int fallbacks[MIGRATE_TYPES][3] = {
2507 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2508 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2509 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2513 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2516 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2519 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2520 unsigned int order) { return NULL; }
2524 * Move the free pages in a range to the freelist tail of the requested type.
2525 * Note that start_page and end_pages are not aligned on a pageblock
2526 * boundary. If alignment is required, use move_freepages_block()
2528 static int move_freepages(struct zone *zone,
2529 unsigned long start_pfn, unsigned long end_pfn,
2530 int migratetype, int *num_movable)
2535 int pages_moved = 0;
2537 for (pfn = start_pfn; pfn <= end_pfn;) {
2538 page = pfn_to_page(pfn);
2539 if (!PageBuddy(page)) {
2541 * We assume that pages that could be isolated for
2542 * migration are movable. But we don't actually try
2543 * isolating, as that would be expensive.
2546 (PageLRU(page) || __PageMovable(page)))
2552 /* Make sure we are not inadvertently changing nodes */
2553 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2554 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2556 order = buddy_order(page);
2557 move_to_free_list(page, zone, order, migratetype);
2559 pages_moved += 1 << order;
2565 int move_freepages_block(struct zone *zone, struct page *page,
2566 int migratetype, int *num_movable)
2568 unsigned long start_pfn, end_pfn, pfn;
2573 pfn = page_to_pfn(page);
2574 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2575 end_pfn = start_pfn + pageblock_nr_pages - 1;
2577 /* Do not cross zone boundaries */
2578 if (!zone_spans_pfn(zone, start_pfn))
2580 if (!zone_spans_pfn(zone, end_pfn))
2583 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2587 static void change_pageblock_range(struct page *pageblock_page,
2588 int start_order, int migratetype)
2590 int nr_pageblocks = 1 << (start_order - pageblock_order);
2592 while (nr_pageblocks--) {
2593 set_pageblock_migratetype(pageblock_page, migratetype);
2594 pageblock_page += pageblock_nr_pages;
2599 * When we are falling back to another migratetype during allocation, try to
2600 * steal extra free pages from the same pageblocks to satisfy further
2601 * allocations, instead of polluting multiple pageblocks.
2603 * If we are stealing a relatively large buddy page, it is likely there will
2604 * be more free pages in the pageblock, so try to steal them all. For
2605 * reclaimable and unmovable allocations, we steal regardless of page size,
2606 * as fragmentation caused by those allocations polluting movable pageblocks
2607 * is worse than movable allocations stealing from unmovable and reclaimable
2610 static bool can_steal_fallback(unsigned int order, int start_mt)
2613 * Leaving this order check is intended, although there is
2614 * relaxed order check in next check. The reason is that
2615 * we can actually steal whole pageblock if this condition met,
2616 * but, below check doesn't guarantee it and that is just heuristic
2617 * so could be changed anytime.
2619 if (order >= pageblock_order)
2622 if (order >= pageblock_order / 2 ||
2623 start_mt == MIGRATE_RECLAIMABLE ||
2624 start_mt == MIGRATE_UNMOVABLE ||
2625 page_group_by_mobility_disabled)
2631 static inline bool boost_watermark(struct zone *zone)
2633 unsigned long max_boost;
2635 if (!watermark_boost_factor)
2638 * Don't bother in zones that are unlikely to produce results.
2639 * On small machines, including kdump capture kernels running
2640 * in a small area, boosting the watermark can cause an out of
2641 * memory situation immediately.
2643 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2646 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2647 watermark_boost_factor, 10000);
2650 * high watermark may be uninitialised if fragmentation occurs
2651 * very early in boot so do not boost. We do not fall
2652 * through and boost by pageblock_nr_pages as failing
2653 * allocations that early means that reclaim is not going
2654 * to help and it may even be impossible to reclaim the
2655 * boosted watermark resulting in a hang.
2660 max_boost = max(pageblock_nr_pages, max_boost);
2662 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2669 * This function implements actual steal behaviour. If order is large enough,
2670 * we can steal whole pageblock. If not, we first move freepages in this
2671 * pageblock to our migratetype and determine how many already-allocated pages
2672 * are there in the pageblock with a compatible migratetype. If at least half
2673 * of pages are free or compatible, we can change migratetype of the pageblock
2674 * itself, so pages freed in the future will be put on the correct free list.
2676 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2677 unsigned int alloc_flags, int start_type, bool whole_block)
2679 unsigned int current_order = buddy_order(page);
2680 int free_pages, movable_pages, alike_pages;
2683 old_block_type = get_pageblock_migratetype(page);
2686 * This can happen due to races and we want to prevent broken
2687 * highatomic accounting.
2689 if (is_migrate_highatomic(old_block_type))
2692 /* Take ownership for orders >= pageblock_order */
2693 if (current_order >= pageblock_order) {
2694 change_pageblock_range(page, current_order, start_type);
2699 * Boost watermarks to increase reclaim pressure to reduce the
2700 * likelihood of future fallbacks. Wake kswapd now as the node
2701 * may be balanced overall and kswapd will not wake naturally.
2703 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2704 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2706 /* We are not allowed to try stealing from the whole block */
2710 free_pages = move_freepages_block(zone, page, start_type,
2713 * Determine how many pages are compatible with our allocation.
2714 * For movable allocation, it's the number of movable pages which
2715 * we just obtained. For other types it's a bit more tricky.
2717 if (start_type == MIGRATE_MOVABLE) {
2718 alike_pages = movable_pages;
2721 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2722 * to MOVABLE pageblock, consider all non-movable pages as
2723 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2724 * vice versa, be conservative since we can't distinguish the
2725 * exact migratetype of non-movable pages.
2727 if (old_block_type == MIGRATE_MOVABLE)
2728 alike_pages = pageblock_nr_pages
2729 - (free_pages + movable_pages);
2734 /* moving whole block can fail due to zone boundary conditions */
2739 * If a sufficient number of pages in the block are either free or of
2740 * comparable migratability as our allocation, claim the whole block.
2742 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2743 page_group_by_mobility_disabled)
2744 set_pageblock_migratetype(page, start_type);
2749 move_to_free_list(page, zone, current_order, start_type);
2753 * Check whether there is a suitable fallback freepage with requested order.
2754 * If only_stealable is true, this function returns fallback_mt only if
2755 * we can steal other freepages all together. This would help to reduce
2756 * fragmentation due to mixed migratetype pages in one pageblock.
2758 int find_suitable_fallback(struct free_area *area, unsigned int order,
2759 int migratetype, bool only_stealable, bool *can_steal)
2764 if (area->nr_free == 0)
2769 fallback_mt = fallbacks[migratetype][i];
2770 if (fallback_mt == MIGRATE_TYPES)
2773 if (free_area_empty(area, fallback_mt))
2776 if (can_steal_fallback(order, migratetype))
2779 if (!only_stealable)
2790 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2791 * there are no empty page blocks that contain a page with a suitable order
2793 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2794 unsigned int alloc_order)
2797 unsigned long max_managed, flags;
2800 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2801 * Check is race-prone but harmless.
2803 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2804 if (zone->nr_reserved_highatomic >= max_managed)
2807 spin_lock_irqsave(&zone->lock, flags);
2809 /* Recheck the nr_reserved_highatomic limit under the lock */
2810 if (zone->nr_reserved_highatomic >= max_managed)
2814 mt = get_pageblock_migratetype(page);
2815 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2816 if (migratetype_is_mergeable(mt)) {
2817 zone->nr_reserved_highatomic += pageblock_nr_pages;
2818 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2819 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2823 spin_unlock_irqrestore(&zone->lock, flags);
2827 * Used when an allocation is about to fail under memory pressure. This
2828 * potentially hurts the reliability of high-order allocations when under
2829 * intense memory pressure but failed atomic allocations should be easier
2830 * to recover from than an OOM.
2832 * If @force is true, try to unreserve a pageblock even though highatomic
2833 * pageblock is exhausted.
2835 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2838 struct zonelist *zonelist = ac->zonelist;
2839 unsigned long flags;
2846 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2849 * Preserve at least one pageblock unless memory pressure
2852 if (!force && zone->nr_reserved_highatomic <=
2856 spin_lock_irqsave(&zone->lock, flags);
2857 for (order = 0; order < MAX_ORDER; order++) {
2858 struct free_area *area = &(zone->free_area[order]);
2860 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2865 * In page freeing path, migratetype change is racy so
2866 * we can counter several free pages in a pageblock
2867 * in this loop although we changed the pageblock type
2868 * from highatomic to ac->migratetype. So we should
2869 * adjust the count once.
2871 if (is_migrate_highatomic_page(page)) {
2873 * It should never happen but changes to
2874 * locking could inadvertently allow a per-cpu
2875 * drain to add pages to MIGRATE_HIGHATOMIC
2876 * while unreserving so be safe and watch for
2879 zone->nr_reserved_highatomic -= min(
2881 zone->nr_reserved_highatomic);
2885 * Convert to ac->migratetype and avoid the normal
2886 * pageblock stealing heuristics. Minimally, the caller
2887 * is doing the work and needs the pages. More
2888 * importantly, if the block was always converted to
2889 * MIGRATE_UNMOVABLE or another type then the number
2890 * of pageblocks that cannot be completely freed
2893 set_pageblock_migratetype(page, ac->migratetype);
2894 ret = move_freepages_block(zone, page, ac->migratetype,
2897 spin_unlock_irqrestore(&zone->lock, flags);
2901 spin_unlock_irqrestore(&zone->lock, flags);
2908 * Try finding a free buddy page on the fallback list and put it on the free
2909 * list of requested migratetype, possibly along with other pages from the same
2910 * block, depending on fragmentation avoidance heuristics. Returns true if
2911 * fallback was found so that __rmqueue_smallest() can grab it.
2913 * The use of signed ints for order and current_order is a deliberate
2914 * deviation from the rest of this file, to make the for loop
2915 * condition simpler.
2917 static __always_inline bool
2918 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2919 unsigned int alloc_flags)
2921 struct free_area *area;
2923 int min_order = order;
2929 * Do not steal pages from freelists belonging to other pageblocks
2930 * i.e. orders < pageblock_order. If there are no local zones free,
2931 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2933 if (alloc_flags & ALLOC_NOFRAGMENT)
2934 min_order = pageblock_order;
2937 * Find the largest available free page in the other list. This roughly
2938 * approximates finding the pageblock with the most free pages, which
2939 * would be too costly to do exactly.
2941 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2943 area = &(zone->free_area[current_order]);
2944 fallback_mt = find_suitable_fallback(area, current_order,
2945 start_migratetype, false, &can_steal);
2946 if (fallback_mt == -1)
2950 * We cannot steal all free pages from the pageblock and the
2951 * requested migratetype is movable. In that case it's better to
2952 * steal and split the smallest available page instead of the
2953 * largest available page, because even if the next movable
2954 * allocation falls back into a different pageblock than this
2955 * one, it won't cause permanent fragmentation.
2957 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2958 && current_order > order)
2967 for (current_order = order; current_order < MAX_ORDER;
2969 area = &(zone->free_area[current_order]);
2970 fallback_mt = find_suitable_fallback(area, current_order,
2971 start_migratetype, false, &can_steal);
2972 if (fallback_mt != -1)
2977 * This should not happen - we already found a suitable fallback
2978 * when looking for the largest page.
2980 VM_BUG_ON(current_order == MAX_ORDER);
2983 page = get_page_from_free_area(area, fallback_mt);
2985 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2988 trace_mm_page_alloc_extfrag(page, order, current_order,
2989 start_migratetype, fallback_mt);
2996 * Do the hard work of removing an element from the buddy allocator.
2997 * Call me with the zone->lock already held.
2999 static __always_inline struct page *
3000 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
3001 unsigned int alloc_flags)
3005 if (IS_ENABLED(CONFIG_CMA)) {
3007 * Balance movable allocations between regular and CMA areas by
3008 * allocating from CMA when over half of the zone's free memory
3009 * is in the CMA area.
3011 if (alloc_flags & ALLOC_CMA &&
3012 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3013 zone_page_state(zone, NR_FREE_PAGES) / 2) {
3014 page = __rmqueue_cma_fallback(zone, order);
3020 page = __rmqueue_smallest(zone, order, migratetype);
3021 if (unlikely(!page)) {
3022 if (alloc_flags & ALLOC_CMA)
3023 page = __rmqueue_cma_fallback(zone, order);
3025 if (!page && __rmqueue_fallback(zone, order, migratetype,
3033 * Obtain a specified number of elements from the buddy allocator, all under
3034 * a single hold of the lock, for efficiency. Add them to the supplied list.
3035 * Returns the number of new pages which were placed at *list.
3037 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3038 unsigned long count, struct list_head *list,
3039 int migratetype, unsigned int alloc_flags)
3041 int i, allocated = 0;
3044 * local_lock_irq held so equivalent to spin_lock_irqsave for
3045 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3047 spin_lock(&zone->lock);
3048 for (i = 0; i < count; ++i) {
3049 struct page *page = __rmqueue(zone, order, migratetype,
3051 if (unlikely(page == NULL))
3054 if (unlikely(check_pcp_refill(page, order)))
3058 * Split buddy pages returned by expand() are received here in
3059 * physical page order. The page is added to the tail of
3060 * caller's list. From the callers perspective, the linked list
3061 * is ordered by page number under some conditions. This is
3062 * useful for IO devices that can forward direction from the
3063 * head, thus also in the physical page order. This is useful
3064 * for IO devices that can merge IO requests if the physical
3065 * pages are ordered properly.
3067 list_add_tail(&page->lru, list);
3069 if (is_migrate_cma(get_pcppage_migratetype(page)))
3070 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3075 * i pages were removed from the buddy list even if some leak due
3076 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3077 * on i. Do not confuse with 'allocated' which is the number of
3078 * pages added to the pcp list.
3080 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3081 spin_unlock(&zone->lock);
3087 * Called from the vmstat counter updater to drain pagesets of this
3088 * currently executing processor on remote nodes after they have
3091 * Note that this function must be called with the thread pinned to
3092 * a single processor.
3094 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3096 unsigned long flags;
3097 int to_drain, batch;
3099 local_lock_irqsave(&pagesets.lock, flags);
3100 batch = READ_ONCE(pcp->batch);
3101 to_drain = min(pcp->count, batch);
3103 free_pcppages_bulk(zone, to_drain, pcp, 0);
3104 local_unlock_irqrestore(&pagesets.lock, flags);
3109 * Drain pcplists of the indicated processor and zone.
3111 * The processor must either be the current processor and the
3112 * thread pinned to the current processor or a processor that
3115 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3117 unsigned long flags;
3118 struct per_cpu_pages *pcp;
3120 local_lock_irqsave(&pagesets.lock, flags);
3122 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3124 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3126 local_unlock_irqrestore(&pagesets.lock, flags);
3130 * Drain pcplists of all zones on the indicated processor.
3132 * The processor must either be the current processor and the
3133 * thread pinned to the current processor or a processor that
3136 static void drain_pages(unsigned int cpu)
3140 for_each_populated_zone(zone) {
3141 drain_pages_zone(cpu, zone);
3146 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3148 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3149 * the single zone's pages.
3151 void drain_local_pages(struct zone *zone)
3153 int cpu = smp_processor_id();
3156 drain_pages_zone(cpu, zone);
3161 static void drain_local_pages_wq(struct work_struct *work)
3163 struct pcpu_drain *drain;
3165 drain = container_of(work, struct pcpu_drain, work);
3168 * drain_all_pages doesn't use proper cpu hotplug protection so
3169 * we can race with cpu offline when the WQ can move this from
3170 * a cpu pinned worker to an unbound one. We can operate on a different
3171 * cpu which is alright but we also have to make sure to not move to
3175 drain_local_pages(drain->zone);
3180 * The implementation of drain_all_pages(), exposing an extra parameter to
3181 * drain on all cpus.
3183 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3184 * not empty. The check for non-emptiness can however race with a free to
3185 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3186 * that need the guarantee that every CPU has drained can disable the
3187 * optimizing racy check.
3189 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3194 * Allocate in the BSS so we won't require allocation in
3195 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3197 static cpumask_t cpus_with_pcps;
3200 * Make sure nobody triggers this path before mm_percpu_wq is fully
3203 if (WARN_ON_ONCE(!mm_percpu_wq))
3207 * Do not drain if one is already in progress unless it's specific to
3208 * a zone. Such callers are primarily CMA and memory hotplug and need
3209 * the drain to be complete when the call returns.
3211 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3214 mutex_lock(&pcpu_drain_mutex);
3218 * We don't care about racing with CPU hotplug event
3219 * as offline notification will cause the notified
3220 * cpu to drain that CPU pcps and on_each_cpu_mask
3221 * disables preemption as part of its processing
3223 for_each_online_cpu(cpu) {
3224 struct per_cpu_pages *pcp;
3226 bool has_pcps = false;
3228 if (force_all_cpus) {
3230 * The pcp.count check is racy, some callers need a
3231 * guarantee that no cpu is missed.
3235 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3239 for_each_populated_zone(z) {
3240 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3249 cpumask_set_cpu(cpu, &cpus_with_pcps);
3251 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3254 for_each_cpu(cpu, &cpus_with_pcps) {
3255 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3258 INIT_WORK(&drain->work, drain_local_pages_wq);
3259 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3261 for_each_cpu(cpu, &cpus_with_pcps)
3262 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3264 mutex_unlock(&pcpu_drain_mutex);
3268 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3270 * When zone parameter is non-NULL, spill just the single zone's pages.
3272 * Note that this can be extremely slow as the draining happens in a workqueue.
3274 void drain_all_pages(struct zone *zone)
3276 __drain_all_pages(zone, false);
3279 #ifdef CONFIG_HIBERNATION
3282 * Touch the watchdog for every WD_PAGE_COUNT pages.
3284 #define WD_PAGE_COUNT (128*1024)
3286 void mark_free_pages(struct zone *zone)
3288 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3289 unsigned long flags;
3290 unsigned int order, t;
3293 if (zone_is_empty(zone))
3296 spin_lock_irqsave(&zone->lock, flags);
3298 max_zone_pfn = zone_end_pfn(zone);
3299 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3300 if (pfn_valid(pfn)) {
3301 page = pfn_to_page(pfn);
3303 if (!--page_count) {
3304 touch_nmi_watchdog();
3305 page_count = WD_PAGE_COUNT;
3308 if (page_zone(page) != zone)
3311 if (!swsusp_page_is_forbidden(page))
3312 swsusp_unset_page_free(page);
3315 for_each_migratetype_order(order, t) {
3316 list_for_each_entry(page,
3317 &zone->free_area[order].free_list[t], lru) {
3320 pfn = page_to_pfn(page);
3321 for (i = 0; i < (1UL << order); i++) {
3322 if (!--page_count) {
3323 touch_nmi_watchdog();
3324 page_count = WD_PAGE_COUNT;
3326 swsusp_set_page_free(pfn_to_page(pfn + i));
3330 spin_unlock_irqrestore(&zone->lock, flags);
3332 #endif /* CONFIG_PM */
3334 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3339 if (!free_pcp_prepare(page, order))
3342 migratetype = get_pfnblock_migratetype(page, pfn);
3343 set_pcppage_migratetype(page, migratetype);
3347 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3350 int min_nr_free, max_nr_free;
3352 /* Free everything if batch freeing high-order pages. */
3353 if (unlikely(free_high))
3356 /* Check for PCP disabled or boot pageset */
3357 if (unlikely(high < batch))
3360 /* Leave at least pcp->batch pages on the list */
3361 min_nr_free = batch;
3362 max_nr_free = high - batch;
3365 * Double the number of pages freed each time there is subsequent
3366 * freeing of pages without any allocation.
3368 batch <<= pcp->free_factor;
3369 if (batch < max_nr_free)
3371 batch = clamp(batch, min_nr_free, max_nr_free);
3376 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3379 int high = READ_ONCE(pcp->high);
3381 if (unlikely(!high || free_high))
3384 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3388 * If reclaim is active, limit the number of pages that can be
3389 * stored on pcp lists
3391 return min(READ_ONCE(pcp->batch) << 2, high);
3394 static void free_unref_page_commit(struct page *page, int migratetype,
3397 struct zone *zone = page_zone(page);
3398 struct per_cpu_pages *pcp;
3403 __count_vm_event(PGFREE);
3404 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3405 pindex = order_to_pindex(migratetype, order);
3406 list_add(&page->lru, &pcp->lists[pindex]);
3407 pcp->count += 1 << order;
3410 * As high-order pages other than THP's stored on PCP can contribute
3411 * to fragmentation, limit the number stored when PCP is heavily
3412 * freeing without allocation. The remainder after bulk freeing
3413 * stops will be drained from vmstat refresh context.
3415 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3417 high = nr_pcp_high(pcp, zone, free_high);
3418 if (pcp->count >= high) {
3419 int batch = READ_ONCE(pcp->batch);
3421 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3428 void free_unref_page(struct page *page, unsigned int order)
3430 unsigned long flags;
3431 unsigned long pfn = page_to_pfn(page);
3434 if (!free_unref_page_prepare(page, pfn, order))
3438 * We only track unmovable, reclaimable and movable on pcp lists.
3439 * Place ISOLATE pages on the isolated list because they are being
3440 * offlined but treat HIGHATOMIC as movable pages so we can get those
3441 * areas back if necessary. Otherwise, we may have to free
3442 * excessively into the page allocator
3444 migratetype = get_pcppage_migratetype(page);
3445 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3446 if (unlikely(is_migrate_isolate(migratetype))) {
3447 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3450 migratetype = MIGRATE_MOVABLE;
3453 local_lock_irqsave(&pagesets.lock, flags);
3454 free_unref_page_commit(page, migratetype, order);
3455 local_unlock_irqrestore(&pagesets.lock, flags);
3459 * Free a list of 0-order pages
3461 void free_unref_page_list(struct list_head *list)
3463 struct page *page, *next;
3464 unsigned long flags;
3465 int batch_count = 0;
3468 /* Prepare pages for freeing */
3469 list_for_each_entry_safe(page, next, list, lru) {
3470 unsigned long pfn = page_to_pfn(page);
3471 if (!free_unref_page_prepare(page, pfn, 0)) {
3472 list_del(&page->lru);
3477 * Free isolated pages directly to the allocator, see
3478 * comment in free_unref_page.
3480 migratetype = get_pcppage_migratetype(page);
3481 if (unlikely(is_migrate_isolate(migratetype))) {
3482 list_del(&page->lru);
3483 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3488 local_lock_irqsave(&pagesets.lock, flags);
3489 list_for_each_entry_safe(page, next, list, lru) {
3491 * Non-isolated types over MIGRATE_PCPTYPES get added
3492 * to the MIGRATE_MOVABLE pcp list.
3494 migratetype = get_pcppage_migratetype(page);
3495 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3496 migratetype = MIGRATE_MOVABLE;
3498 trace_mm_page_free_batched(page);
3499 free_unref_page_commit(page, migratetype, 0);
3502 * Guard against excessive IRQ disabled times when we get
3503 * a large list of pages to free.
3505 if (++batch_count == SWAP_CLUSTER_MAX) {
3506 local_unlock_irqrestore(&pagesets.lock, flags);
3508 local_lock_irqsave(&pagesets.lock, flags);
3511 local_unlock_irqrestore(&pagesets.lock, flags);
3515 * split_page takes a non-compound higher-order page, and splits it into
3516 * n (1<<order) sub-pages: page[0..n]
3517 * Each sub-page must be freed individually.
3519 * Note: this is probably too low level an operation for use in drivers.
3520 * Please consult with lkml before using this in your driver.
3522 void split_page(struct page *page, unsigned int order)
3526 VM_BUG_ON_PAGE(PageCompound(page), page);
3527 VM_BUG_ON_PAGE(!page_count(page), page);
3529 for (i = 1; i < (1 << order); i++)
3530 set_page_refcounted(page + i);
3531 split_page_owner(page, 1 << order);
3532 split_page_memcg(page, 1 << order);
3534 EXPORT_SYMBOL_GPL(split_page);
3536 int __isolate_free_page(struct page *page, unsigned int order)
3538 unsigned long watermark;
3542 BUG_ON(!PageBuddy(page));
3544 zone = page_zone(page);
3545 mt = get_pageblock_migratetype(page);
3547 if (!is_migrate_isolate(mt)) {
3549 * Obey watermarks as if the page was being allocated. We can
3550 * emulate a high-order watermark check with a raised order-0
3551 * watermark, because we already know our high-order page
3554 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3555 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3558 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3561 /* Remove page from free list */
3563 del_page_from_free_list(page, zone, order);
3566 * Set the pageblock if the isolated page is at least half of a
3569 if (order >= pageblock_order - 1) {
3570 struct page *endpage = page + (1 << order) - 1;
3571 for (; page < endpage; page += pageblock_nr_pages) {
3572 int mt = get_pageblock_migratetype(page);
3574 * Only change normal pageblocks (i.e., they can merge
3577 if (migratetype_is_mergeable(mt))
3578 set_pageblock_migratetype(page,
3584 return 1UL << order;
3588 * __putback_isolated_page - Return a now-isolated page back where we got it
3589 * @page: Page that was isolated
3590 * @order: Order of the isolated page
3591 * @mt: The page's pageblock's migratetype
3593 * This function is meant to return a page pulled from the free lists via
3594 * __isolate_free_page back to the free lists they were pulled from.
3596 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3598 struct zone *zone = page_zone(page);
3600 /* zone lock should be held when this function is called */
3601 lockdep_assert_held(&zone->lock);
3603 /* Return isolated page to tail of freelist. */
3604 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3605 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3609 * Update NUMA hit/miss statistics
3611 * Must be called with interrupts disabled.
3613 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3617 enum numa_stat_item local_stat = NUMA_LOCAL;
3619 /* skip numa counters update if numa stats is disabled */
3620 if (!static_branch_likely(&vm_numa_stat_key))
3623 if (zone_to_nid(z) != numa_node_id())
3624 local_stat = NUMA_OTHER;
3626 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3627 __count_numa_events(z, NUMA_HIT, nr_account);
3629 __count_numa_events(z, NUMA_MISS, nr_account);
3630 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3632 __count_numa_events(z, local_stat, nr_account);
3636 /* Remove page from the per-cpu list, caller must protect the list */
3638 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3640 unsigned int alloc_flags,
3641 struct per_cpu_pages *pcp,
3642 struct list_head *list)
3647 if (list_empty(list)) {
3648 int batch = READ_ONCE(pcp->batch);
3652 * Scale batch relative to order if batch implies
3653 * free pages can be stored on the PCP. Batch can
3654 * be 1 for small zones or for boot pagesets which
3655 * should never store free pages as the pages may
3656 * belong to arbitrary zones.
3659 batch = max(batch >> order, 2);
3660 alloced = rmqueue_bulk(zone, order,
3662 migratetype, alloc_flags);
3664 pcp->count += alloced << order;
3665 if (unlikely(list_empty(list)))
3669 page = list_first_entry(list, struct page, lru);
3670 list_del(&page->lru);
3671 pcp->count -= 1 << order;
3672 } while (check_new_pcp(page, order));
3677 /* Lock and remove page from the per-cpu list */
3678 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3679 struct zone *zone, unsigned int order,
3680 gfp_t gfp_flags, int migratetype,
3681 unsigned int alloc_flags)
3683 struct per_cpu_pages *pcp;
3684 struct list_head *list;
3686 unsigned long flags;
3688 local_lock_irqsave(&pagesets.lock, flags);
3691 * On allocation, reduce the number of pages that are batch freed.
3692 * See nr_pcp_free() where free_factor is increased for subsequent
3695 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3696 pcp->free_factor >>= 1;
3697 list = &pcp->lists[order_to_pindex(migratetype, order)];
3698 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3699 local_unlock_irqrestore(&pagesets.lock, flags);
3701 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3702 zone_statistics(preferred_zone, zone, 1);
3708 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3711 struct page *rmqueue(struct zone *preferred_zone,
3712 struct zone *zone, unsigned int order,
3713 gfp_t gfp_flags, unsigned int alloc_flags,
3716 unsigned long flags;
3719 if (likely(pcp_allowed_order(order))) {
3721 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3722 * we need to skip it when CMA area isn't allowed.
3724 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3725 migratetype != MIGRATE_MOVABLE) {
3726 page = rmqueue_pcplist(preferred_zone, zone, order,
3727 gfp_flags, migratetype, alloc_flags);
3733 * We most definitely don't want callers attempting to
3734 * allocate greater than order-1 page units with __GFP_NOFAIL.
3736 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3740 spin_lock_irqsave(&zone->lock, flags);
3742 * order-0 request can reach here when the pcplist is skipped
3743 * due to non-CMA allocation context. HIGHATOMIC area is
3744 * reserved for high-order atomic allocation, so order-0
3745 * request should skip it.
3747 if (order > 0 && alloc_flags & ALLOC_HARDER)
3748 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3750 page = __rmqueue(zone, order, migratetype, alloc_flags);
3754 __mod_zone_freepage_state(zone, -(1 << order),
3755 get_pcppage_migratetype(page));
3756 spin_unlock_irqrestore(&zone->lock, flags);
3757 } while (check_new_pages(page, order));
3759 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3760 zone_statistics(preferred_zone, zone, 1);
3763 /* Separate test+clear to avoid unnecessary atomics */
3764 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3765 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3766 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3769 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3773 spin_unlock_irqrestore(&zone->lock, flags);
3777 #ifdef CONFIG_FAIL_PAGE_ALLOC
3780 struct fault_attr attr;
3782 bool ignore_gfp_highmem;
3783 bool ignore_gfp_reclaim;
3785 } fail_page_alloc = {
3786 .attr = FAULT_ATTR_INITIALIZER,
3787 .ignore_gfp_reclaim = true,
3788 .ignore_gfp_highmem = true,
3792 static int __init setup_fail_page_alloc(char *str)
3794 return setup_fault_attr(&fail_page_alloc.attr, str);
3796 __setup("fail_page_alloc=", setup_fail_page_alloc);
3798 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3800 if (order < fail_page_alloc.min_order)
3802 if (gfp_mask & __GFP_NOFAIL)
3804 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3806 if (fail_page_alloc.ignore_gfp_reclaim &&
3807 (gfp_mask & __GFP_DIRECT_RECLAIM))
3810 if (gfp_mask & __GFP_NOWARN)
3811 fail_page_alloc.attr.no_warn = true;
3813 return should_fail(&fail_page_alloc.attr, 1 << order);
3816 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3818 static int __init fail_page_alloc_debugfs(void)
3820 umode_t mode = S_IFREG | 0600;
3823 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3824 &fail_page_alloc.attr);
3826 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3827 &fail_page_alloc.ignore_gfp_reclaim);
3828 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3829 &fail_page_alloc.ignore_gfp_highmem);
3830 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3835 late_initcall(fail_page_alloc_debugfs);
3837 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3839 #else /* CONFIG_FAIL_PAGE_ALLOC */
3841 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3846 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3848 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3850 return __should_fail_alloc_page(gfp_mask, order);
3852 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3854 static inline long __zone_watermark_unusable_free(struct zone *z,
3855 unsigned int order, unsigned int alloc_flags)
3857 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3858 long unusable_free = (1 << order) - 1;
3861 * If the caller does not have rights to ALLOC_HARDER then subtract
3862 * the high-atomic reserves. This will over-estimate the size of the
3863 * atomic reserve but it avoids a search.
3865 if (likely(!alloc_harder))
3866 unusable_free += z->nr_reserved_highatomic;
3869 /* If allocation can't use CMA areas don't use free CMA pages */
3870 if (!(alloc_flags & ALLOC_CMA))
3871 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3874 return unusable_free;
3878 * Return true if free base pages are above 'mark'. For high-order checks it
3879 * will return true of the order-0 watermark is reached and there is at least
3880 * one free page of a suitable size. Checking now avoids taking the zone lock
3881 * to check in the allocation paths if no pages are free.
3883 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3884 int highest_zoneidx, unsigned int alloc_flags,
3889 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3891 /* free_pages may go negative - that's OK */
3892 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3894 if (alloc_flags & ALLOC_HIGH)
3897 if (unlikely(alloc_harder)) {
3899 * OOM victims can try even harder than normal ALLOC_HARDER
3900 * users on the grounds that it's definitely going to be in
3901 * the exit path shortly and free memory. Any allocation it
3902 * makes during the free path will be small and short-lived.
3904 if (alloc_flags & ALLOC_OOM)
3911 * Check watermarks for an order-0 allocation request. If these
3912 * are not met, then a high-order request also cannot go ahead
3913 * even if a suitable page happened to be free.
3915 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3918 /* If this is an order-0 request then the watermark is fine */
3922 /* For a high-order request, check at least one suitable page is free */
3923 for (o = order; o < MAX_ORDER; o++) {
3924 struct free_area *area = &z->free_area[o];
3930 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3931 if (!free_area_empty(area, mt))
3936 if ((alloc_flags & ALLOC_CMA) &&
3937 !free_area_empty(area, MIGRATE_CMA)) {
3941 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3947 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3948 int highest_zoneidx, unsigned int alloc_flags)
3950 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3951 zone_page_state(z, NR_FREE_PAGES));
3954 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3955 unsigned long mark, int highest_zoneidx,
3956 unsigned int alloc_flags, gfp_t gfp_mask)
3960 free_pages = zone_page_state(z, NR_FREE_PAGES);
3963 * Fast check for order-0 only. If this fails then the reserves
3964 * need to be calculated.
3969 fast_free = free_pages;
3970 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3971 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3975 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3979 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3980 * when checking the min watermark. The min watermark is the
3981 * point where boosting is ignored so that kswapd is woken up
3982 * when below the low watermark.
3984 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3985 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3986 mark = z->_watermark[WMARK_MIN];
3987 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3988 alloc_flags, free_pages);
3994 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3995 unsigned long mark, int highest_zoneidx)
3997 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3999 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
4000 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
4002 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
4007 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
4009 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4011 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
4012 node_reclaim_distance;
4014 #else /* CONFIG_NUMA */
4015 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4019 #endif /* CONFIG_NUMA */
4022 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4023 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4024 * premature use of a lower zone may cause lowmem pressure problems that
4025 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4026 * probably too small. It only makes sense to spread allocations to avoid
4027 * fragmentation between the Normal and DMA32 zones.
4029 static inline unsigned int
4030 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4032 unsigned int alloc_flags;
4035 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4038 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4040 #ifdef CONFIG_ZONE_DMA32
4044 if (zone_idx(zone) != ZONE_NORMAL)
4048 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4049 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4050 * on UMA that if Normal is populated then so is DMA32.
4052 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4053 if (nr_online_nodes > 1 && !populated_zone(--zone))
4056 alloc_flags |= ALLOC_NOFRAGMENT;
4057 #endif /* CONFIG_ZONE_DMA32 */
4061 /* Must be called after current_gfp_context() which can change gfp_mask */
4062 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4063 unsigned int alloc_flags)
4066 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4067 alloc_flags |= ALLOC_CMA;
4073 * get_page_from_freelist goes through the zonelist trying to allocate
4076 static struct page *
4077 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4078 const struct alloc_context *ac)
4082 struct pglist_data *last_pgdat = NULL;
4083 bool last_pgdat_dirty_ok = false;
4088 * Scan zonelist, looking for a zone with enough free.
4089 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4091 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4092 z = ac->preferred_zoneref;
4093 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4098 if (cpusets_enabled() &&
4099 (alloc_flags & ALLOC_CPUSET) &&
4100 !__cpuset_zone_allowed(zone, gfp_mask))
4103 * When allocating a page cache page for writing, we
4104 * want to get it from a node that is within its dirty
4105 * limit, such that no single node holds more than its
4106 * proportional share of globally allowed dirty pages.
4107 * The dirty limits take into account the node's
4108 * lowmem reserves and high watermark so that kswapd
4109 * should be able to balance it without having to
4110 * write pages from its LRU list.
4112 * XXX: For now, allow allocations to potentially
4113 * exceed the per-node dirty limit in the slowpath
4114 * (spread_dirty_pages unset) before going into reclaim,
4115 * which is important when on a NUMA setup the allowed
4116 * nodes are together not big enough to reach the
4117 * global limit. The proper fix for these situations
4118 * will require awareness of nodes in the
4119 * dirty-throttling and the flusher threads.
4121 if (ac->spread_dirty_pages) {
4122 if (last_pgdat != zone->zone_pgdat) {
4123 last_pgdat = zone->zone_pgdat;
4124 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
4127 if (!last_pgdat_dirty_ok)
4131 if (no_fallback && nr_online_nodes > 1 &&
4132 zone != ac->preferred_zoneref->zone) {
4136 * If moving to a remote node, retry but allow
4137 * fragmenting fallbacks. Locality is more important
4138 * than fragmentation avoidance.
4140 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4141 if (zone_to_nid(zone) != local_nid) {
4142 alloc_flags &= ~ALLOC_NOFRAGMENT;
4147 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4148 if (!zone_watermark_fast(zone, order, mark,
4149 ac->highest_zoneidx, alloc_flags,
4153 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4155 * Watermark failed for this zone, but see if we can
4156 * grow this zone if it contains deferred pages.
4158 if (static_branch_unlikely(&deferred_pages)) {
4159 if (_deferred_grow_zone(zone, order))
4163 /* Checked here to keep the fast path fast */
4164 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4165 if (alloc_flags & ALLOC_NO_WATERMARKS)
4168 if (!node_reclaim_enabled() ||
4169 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4172 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4174 case NODE_RECLAIM_NOSCAN:
4177 case NODE_RECLAIM_FULL:
4178 /* scanned but unreclaimable */
4181 /* did we reclaim enough */
4182 if (zone_watermark_ok(zone, order, mark,
4183 ac->highest_zoneidx, alloc_flags))
4191 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4192 gfp_mask, alloc_flags, ac->migratetype);
4194 prep_new_page(page, order, gfp_mask, alloc_flags);
4197 * If this is a high-order atomic allocation then check
4198 * if the pageblock should be reserved for the future
4200 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4201 reserve_highatomic_pageblock(page, zone, order);
4205 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4206 /* Try again if zone has deferred pages */
4207 if (static_branch_unlikely(&deferred_pages)) {
4208 if (_deferred_grow_zone(zone, order))
4216 * It's possible on a UMA machine to get through all zones that are
4217 * fragmented. If avoiding fragmentation, reset and try again.
4220 alloc_flags &= ~ALLOC_NOFRAGMENT;
4227 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4229 unsigned int filter = SHOW_MEM_FILTER_NODES;
4232 * This documents exceptions given to allocations in certain
4233 * contexts that are allowed to allocate outside current's set
4236 if (!(gfp_mask & __GFP_NOMEMALLOC))
4237 if (tsk_is_oom_victim(current) ||
4238 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4239 filter &= ~SHOW_MEM_FILTER_NODES;
4240 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4241 filter &= ~SHOW_MEM_FILTER_NODES;
4243 show_mem(filter, nodemask);
4246 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4248 struct va_format vaf;
4250 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4252 if ((gfp_mask & __GFP_NOWARN) ||
4253 !__ratelimit(&nopage_rs) ||
4254 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4257 va_start(args, fmt);
4260 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4261 current->comm, &vaf, gfp_mask, &gfp_mask,
4262 nodemask_pr_args(nodemask));
4265 cpuset_print_current_mems_allowed();
4268 warn_alloc_show_mem(gfp_mask, nodemask);
4271 static inline struct page *
4272 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4273 unsigned int alloc_flags,
4274 const struct alloc_context *ac)
4278 page = get_page_from_freelist(gfp_mask, order,
4279 alloc_flags|ALLOC_CPUSET, ac);
4281 * fallback to ignore cpuset restriction if our nodes
4285 page = get_page_from_freelist(gfp_mask, order,
4291 static inline struct page *
4292 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4293 const struct alloc_context *ac, unsigned long *did_some_progress)
4295 struct oom_control oc = {
4296 .zonelist = ac->zonelist,
4297 .nodemask = ac->nodemask,
4299 .gfp_mask = gfp_mask,
4304 *did_some_progress = 0;
4307 * Acquire the oom lock. If that fails, somebody else is
4308 * making progress for us.
4310 if (!mutex_trylock(&oom_lock)) {
4311 *did_some_progress = 1;
4312 schedule_timeout_uninterruptible(1);
4317 * Go through the zonelist yet one more time, keep very high watermark
4318 * here, this is only to catch a parallel oom killing, we must fail if
4319 * we're still under heavy pressure. But make sure that this reclaim
4320 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4321 * allocation which will never fail due to oom_lock already held.
4323 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4324 ~__GFP_DIRECT_RECLAIM, order,
4325 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4329 /* Coredumps can quickly deplete all memory reserves */
4330 if (current->flags & PF_DUMPCORE)
4332 /* The OOM killer will not help higher order allocs */
4333 if (order > PAGE_ALLOC_COSTLY_ORDER)
4336 * We have already exhausted all our reclaim opportunities without any
4337 * success so it is time to admit defeat. We will skip the OOM killer
4338 * because it is very likely that the caller has a more reasonable
4339 * fallback than shooting a random task.
4341 * The OOM killer may not free memory on a specific node.
4343 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4345 /* The OOM killer does not needlessly kill tasks for lowmem */
4346 if (ac->highest_zoneidx < ZONE_NORMAL)
4348 if (pm_suspended_storage())
4351 * XXX: GFP_NOFS allocations should rather fail than rely on
4352 * other request to make a forward progress.
4353 * We are in an unfortunate situation where out_of_memory cannot
4354 * do much for this context but let's try it to at least get
4355 * access to memory reserved if the current task is killed (see
4356 * out_of_memory). Once filesystems are ready to handle allocation
4357 * failures more gracefully we should just bail out here.
4360 /* Exhausted what can be done so it's blame time */
4361 if (out_of_memory(&oc) ||
4362 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4363 *did_some_progress = 1;
4366 * Help non-failing allocations by giving them access to memory
4369 if (gfp_mask & __GFP_NOFAIL)
4370 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4371 ALLOC_NO_WATERMARKS, ac);
4374 mutex_unlock(&oom_lock);
4379 * Maximum number of compaction retries with a progress before OOM
4380 * killer is consider as the only way to move forward.
4382 #define MAX_COMPACT_RETRIES 16
4384 #ifdef CONFIG_COMPACTION
4385 /* Try memory compaction for high-order allocations before reclaim */
4386 static struct page *
4387 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4388 unsigned int alloc_flags, const struct alloc_context *ac,
4389 enum compact_priority prio, enum compact_result *compact_result)
4391 struct page *page = NULL;
4392 unsigned long pflags;
4393 unsigned int noreclaim_flag;
4398 psi_memstall_enter(&pflags);
4399 delayacct_compact_start();
4400 noreclaim_flag = memalloc_noreclaim_save();
4402 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4405 memalloc_noreclaim_restore(noreclaim_flag);
4406 psi_memstall_leave(&pflags);
4407 delayacct_compact_end();
4409 if (*compact_result == COMPACT_SKIPPED)
4412 * At least in one zone compaction wasn't deferred or skipped, so let's
4413 * count a compaction stall
4415 count_vm_event(COMPACTSTALL);
4417 /* Prep a captured page if available */
4419 prep_new_page(page, order, gfp_mask, alloc_flags);
4421 /* Try get a page from the freelist if available */
4423 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4426 struct zone *zone = page_zone(page);
4428 zone->compact_blockskip_flush = false;
4429 compaction_defer_reset(zone, order, true);
4430 count_vm_event(COMPACTSUCCESS);
4435 * It's bad if compaction run occurs and fails. The most likely reason
4436 * is that pages exist, but not enough to satisfy watermarks.
4438 count_vm_event(COMPACTFAIL);
4446 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4447 enum compact_result compact_result,
4448 enum compact_priority *compact_priority,
4449 int *compaction_retries)
4451 int max_retries = MAX_COMPACT_RETRIES;
4454 int retries = *compaction_retries;
4455 enum compact_priority priority = *compact_priority;
4460 if (fatal_signal_pending(current))
4463 if (compaction_made_progress(compact_result))
4464 (*compaction_retries)++;
4467 * compaction considers all the zone as desperately out of memory
4468 * so it doesn't really make much sense to retry except when the
4469 * failure could be caused by insufficient priority
4471 if (compaction_failed(compact_result))
4472 goto check_priority;
4475 * compaction was skipped because there are not enough order-0 pages
4476 * to work with, so we retry only if it looks like reclaim can help.
4478 if (compaction_needs_reclaim(compact_result)) {
4479 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4484 * make sure the compaction wasn't deferred or didn't bail out early
4485 * due to locks contention before we declare that we should give up.
4486 * But the next retry should use a higher priority if allowed, so
4487 * we don't just keep bailing out endlessly.
4489 if (compaction_withdrawn(compact_result)) {
4490 goto check_priority;
4494 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4495 * costly ones because they are de facto nofail and invoke OOM
4496 * killer to move on while costly can fail and users are ready
4497 * to cope with that. 1/4 retries is rather arbitrary but we
4498 * would need much more detailed feedback from compaction to
4499 * make a better decision.
4501 if (order > PAGE_ALLOC_COSTLY_ORDER)
4503 if (*compaction_retries <= max_retries) {
4509 * Make sure there are attempts at the highest priority if we exhausted
4510 * all retries or failed at the lower priorities.
4513 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4514 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4516 if (*compact_priority > min_priority) {
4517 (*compact_priority)--;
4518 *compaction_retries = 0;
4522 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4526 static inline struct page *
4527 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4528 unsigned int alloc_flags, const struct alloc_context *ac,
4529 enum compact_priority prio, enum compact_result *compact_result)
4531 *compact_result = COMPACT_SKIPPED;
4536 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4537 enum compact_result compact_result,
4538 enum compact_priority *compact_priority,
4539 int *compaction_retries)
4544 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4548 * There are setups with compaction disabled which would prefer to loop
4549 * inside the allocator rather than hit the oom killer prematurely.
4550 * Let's give them a good hope and keep retrying while the order-0
4551 * watermarks are OK.
4553 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4554 ac->highest_zoneidx, ac->nodemask) {
4555 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4556 ac->highest_zoneidx, alloc_flags))
4561 #endif /* CONFIG_COMPACTION */
4563 #ifdef CONFIG_LOCKDEP
4564 static struct lockdep_map __fs_reclaim_map =
4565 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4567 static bool __need_reclaim(gfp_t gfp_mask)
4569 /* no reclaim without waiting on it */
4570 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4573 /* this guy won't enter reclaim */
4574 if (current->flags & PF_MEMALLOC)
4577 if (gfp_mask & __GFP_NOLOCKDEP)
4583 void __fs_reclaim_acquire(unsigned long ip)
4585 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4588 void __fs_reclaim_release(unsigned long ip)
4590 lock_release(&__fs_reclaim_map, ip);
4593 void fs_reclaim_acquire(gfp_t gfp_mask)
4595 gfp_mask = current_gfp_context(gfp_mask);
4597 if (__need_reclaim(gfp_mask)) {
4598 if (gfp_mask & __GFP_FS)
4599 __fs_reclaim_acquire(_RET_IP_);
4601 #ifdef CONFIG_MMU_NOTIFIER
4602 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4603 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4608 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4610 void fs_reclaim_release(gfp_t gfp_mask)
4612 gfp_mask = current_gfp_context(gfp_mask);
4614 if (__need_reclaim(gfp_mask)) {
4615 if (gfp_mask & __GFP_FS)
4616 __fs_reclaim_release(_RET_IP_);
4619 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4622 /* Perform direct synchronous page reclaim */
4623 static unsigned long
4624 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4625 const struct alloc_context *ac)
4627 unsigned int noreclaim_flag;
4628 unsigned long progress;
4632 /* We now go into synchronous reclaim */
4633 cpuset_memory_pressure_bump();
4634 fs_reclaim_acquire(gfp_mask);
4635 noreclaim_flag = memalloc_noreclaim_save();
4637 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4640 memalloc_noreclaim_restore(noreclaim_flag);
4641 fs_reclaim_release(gfp_mask);
4648 /* The really slow allocator path where we enter direct reclaim */
4649 static inline struct page *
4650 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4651 unsigned int alloc_flags, const struct alloc_context *ac,
4652 unsigned long *did_some_progress)
4654 struct page *page = NULL;
4655 unsigned long pflags;
4656 bool drained = false;
4658 psi_memstall_enter(&pflags);
4659 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4660 if (unlikely(!(*did_some_progress)))
4664 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4667 * If an allocation failed after direct reclaim, it could be because
4668 * pages are pinned on the per-cpu lists or in high alloc reserves.
4669 * Shrink them and try again
4671 if (!page && !drained) {
4672 unreserve_highatomic_pageblock(ac, false);
4673 drain_all_pages(NULL);
4678 psi_memstall_leave(&pflags);
4683 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4684 const struct alloc_context *ac)
4688 pg_data_t *last_pgdat = NULL;
4689 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4691 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4693 if (!managed_zone(zone))
4695 if (last_pgdat != zone->zone_pgdat) {
4696 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4697 last_pgdat = zone->zone_pgdat;
4702 static inline unsigned int
4703 gfp_to_alloc_flags(gfp_t gfp_mask)
4705 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4708 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4709 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4710 * to save two branches.
4712 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4713 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4716 * The caller may dip into page reserves a bit more if the caller
4717 * cannot run direct reclaim, or if the caller has realtime scheduling
4718 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4719 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4721 alloc_flags |= (__force int)
4722 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4724 if (gfp_mask & __GFP_ATOMIC) {
4726 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4727 * if it can't schedule.
4729 if (!(gfp_mask & __GFP_NOMEMALLOC))
4730 alloc_flags |= ALLOC_HARDER;
4732 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4733 * comment for __cpuset_node_allowed().
4735 alloc_flags &= ~ALLOC_CPUSET;
4736 } else if (unlikely(rt_task(current)) && in_task())
4737 alloc_flags |= ALLOC_HARDER;
4739 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4744 static bool oom_reserves_allowed(struct task_struct *tsk)
4746 if (!tsk_is_oom_victim(tsk))
4750 * !MMU doesn't have oom reaper so give access to memory reserves
4751 * only to the thread with TIF_MEMDIE set
4753 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4760 * Distinguish requests which really need access to full memory
4761 * reserves from oom victims which can live with a portion of it
4763 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4765 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4767 if (gfp_mask & __GFP_MEMALLOC)
4768 return ALLOC_NO_WATERMARKS;
4769 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4770 return ALLOC_NO_WATERMARKS;
4771 if (!in_interrupt()) {
4772 if (current->flags & PF_MEMALLOC)
4773 return ALLOC_NO_WATERMARKS;
4774 else if (oom_reserves_allowed(current))
4781 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4783 return !!__gfp_pfmemalloc_flags(gfp_mask);
4787 * Checks whether it makes sense to retry the reclaim to make a forward progress
4788 * for the given allocation request.
4790 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4791 * without success, or when we couldn't even meet the watermark if we
4792 * reclaimed all remaining pages on the LRU lists.
4794 * Returns true if a retry is viable or false to enter the oom path.
4797 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4798 struct alloc_context *ac, int alloc_flags,
4799 bool did_some_progress, int *no_progress_loops)
4806 * Costly allocations might have made a progress but this doesn't mean
4807 * their order will become available due to high fragmentation so
4808 * always increment the no progress counter for them
4810 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4811 *no_progress_loops = 0;
4813 (*no_progress_loops)++;
4816 * Make sure we converge to OOM if we cannot make any progress
4817 * several times in the row.
4819 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4820 /* Before OOM, exhaust highatomic_reserve */
4821 return unreserve_highatomic_pageblock(ac, true);
4825 * Keep reclaiming pages while there is a chance this will lead
4826 * somewhere. If none of the target zones can satisfy our allocation
4827 * request even if all reclaimable pages are considered then we are
4828 * screwed and have to go OOM.
4830 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4831 ac->highest_zoneidx, ac->nodemask) {
4832 unsigned long available;
4833 unsigned long reclaimable;
4834 unsigned long min_wmark = min_wmark_pages(zone);
4837 available = reclaimable = zone_reclaimable_pages(zone);
4838 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4841 * Would the allocation succeed if we reclaimed all
4842 * reclaimable pages?
4844 wmark = __zone_watermark_ok(zone, order, min_wmark,
4845 ac->highest_zoneidx, alloc_flags, available);
4846 trace_reclaim_retry_zone(z, order, reclaimable,
4847 available, min_wmark, *no_progress_loops, wmark);
4855 * Memory allocation/reclaim might be called from a WQ context and the
4856 * current implementation of the WQ concurrency control doesn't
4857 * recognize that a particular WQ is congested if the worker thread is
4858 * looping without ever sleeping. Therefore we have to do a short sleep
4859 * here rather than calling cond_resched().
4861 if (current->flags & PF_WQ_WORKER)
4862 schedule_timeout_uninterruptible(1);
4869 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4872 * It's possible that cpuset's mems_allowed and the nodemask from
4873 * mempolicy don't intersect. This should be normally dealt with by
4874 * policy_nodemask(), but it's possible to race with cpuset update in
4875 * such a way the check therein was true, and then it became false
4876 * before we got our cpuset_mems_cookie here.
4877 * This assumes that for all allocations, ac->nodemask can come only
4878 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4879 * when it does not intersect with the cpuset restrictions) or the
4880 * caller can deal with a violated nodemask.
4882 if (cpusets_enabled() && ac->nodemask &&
4883 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4884 ac->nodemask = NULL;
4889 * When updating a task's mems_allowed or mempolicy nodemask, it is
4890 * possible to race with parallel threads in such a way that our
4891 * allocation can fail while the mask is being updated. If we are about
4892 * to fail, check if the cpuset changed during allocation and if so,
4895 if (read_mems_allowed_retry(cpuset_mems_cookie))
4901 static inline struct page *
4902 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4903 struct alloc_context *ac)
4905 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4906 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4907 struct page *page = NULL;
4908 unsigned int alloc_flags;
4909 unsigned long did_some_progress;
4910 enum compact_priority compact_priority;
4911 enum compact_result compact_result;
4912 int compaction_retries;
4913 int no_progress_loops;
4914 unsigned int cpuset_mems_cookie;
4918 * We also sanity check to catch abuse of atomic reserves being used by
4919 * callers that are not in atomic context.
4921 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4922 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4923 gfp_mask &= ~__GFP_ATOMIC;
4926 compaction_retries = 0;
4927 no_progress_loops = 0;
4928 compact_priority = DEF_COMPACT_PRIORITY;
4929 cpuset_mems_cookie = read_mems_allowed_begin();
4932 * The fast path uses conservative alloc_flags to succeed only until
4933 * kswapd needs to be woken up, and to avoid the cost of setting up
4934 * alloc_flags precisely. So we do that now.
4936 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4939 * We need to recalculate the starting point for the zonelist iterator
4940 * because we might have used different nodemask in the fast path, or
4941 * there was a cpuset modification and we are retrying - otherwise we
4942 * could end up iterating over non-eligible zones endlessly.
4944 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4945 ac->highest_zoneidx, ac->nodemask);
4946 if (!ac->preferred_zoneref->zone)
4950 * Check for insane configurations where the cpuset doesn't contain
4951 * any suitable zone to satisfy the request - e.g. non-movable
4952 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4954 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4955 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4956 ac->highest_zoneidx,
4957 &cpuset_current_mems_allowed);
4962 if (alloc_flags & ALLOC_KSWAPD)
4963 wake_all_kswapds(order, gfp_mask, ac);
4966 * The adjusted alloc_flags might result in immediate success, so try
4969 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4974 * For costly allocations, try direct compaction first, as it's likely
4975 * that we have enough base pages and don't need to reclaim. For non-
4976 * movable high-order allocations, do that as well, as compaction will
4977 * try prevent permanent fragmentation by migrating from blocks of the
4979 * Don't try this for allocations that are allowed to ignore
4980 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4982 if (can_direct_reclaim &&
4984 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4985 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4986 page = __alloc_pages_direct_compact(gfp_mask, order,
4988 INIT_COMPACT_PRIORITY,
4994 * Checks for costly allocations with __GFP_NORETRY, which
4995 * includes some THP page fault allocations
4997 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4999 * If allocating entire pageblock(s) and compaction
5000 * failed because all zones are below low watermarks
5001 * or is prohibited because it recently failed at this
5002 * order, fail immediately unless the allocator has
5003 * requested compaction and reclaim retry.
5006 * - potentially very expensive because zones are far
5007 * below their low watermarks or this is part of very
5008 * bursty high order allocations,
5009 * - not guaranteed to help because isolate_freepages()
5010 * may not iterate over freed pages as part of its
5012 * - unlikely to make entire pageblocks free on its
5015 if (compact_result == COMPACT_SKIPPED ||
5016 compact_result == COMPACT_DEFERRED)
5020 * Looks like reclaim/compaction is worth trying, but
5021 * sync compaction could be very expensive, so keep
5022 * using async compaction.
5024 compact_priority = INIT_COMPACT_PRIORITY;
5029 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5030 if (alloc_flags & ALLOC_KSWAPD)
5031 wake_all_kswapds(order, gfp_mask, ac);
5033 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5035 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
5038 * Reset the nodemask and zonelist iterators if memory policies can be
5039 * ignored. These allocations are high priority and system rather than
5042 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5043 ac->nodemask = NULL;
5044 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5045 ac->highest_zoneidx, ac->nodemask);
5048 /* Attempt with potentially adjusted zonelist and alloc_flags */
5049 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5053 /* Caller is not willing to reclaim, we can't balance anything */
5054 if (!can_direct_reclaim)
5057 /* Avoid recursion of direct reclaim */
5058 if (current->flags & PF_MEMALLOC)
5061 /* Try direct reclaim and then allocating */
5062 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5063 &did_some_progress);
5067 /* Try direct compaction and then allocating */
5068 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5069 compact_priority, &compact_result);
5073 /* Do not loop if specifically requested */
5074 if (gfp_mask & __GFP_NORETRY)
5078 * Do not retry costly high order allocations unless they are
5079 * __GFP_RETRY_MAYFAIL
5081 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5084 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5085 did_some_progress > 0, &no_progress_loops))
5089 * It doesn't make any sense to retry for the compaction if the order-0
5090 * reclaim is not able to make any progress because the current
5091 * implementation of the compaction depends on the sufficient amount
5092 * of free memory (see __compaction_suitable)
5094 if (did_some_progress > 0 &&
5095 should_compact_retry(ac, order, alloc_flags,
5096 compact_result, &compact_priority,
5097 &compaction_retries))
5101 /* Deal with possible cpuset update races before we start OOM killing */
5102 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5105 /* Reclaim has failed us, start killing things */
5106 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5110 /* Avoid allocations with no watermarks from looping endlessly */
5111 if (tsk_is_oom_victim(current) &&
5112 (alloc_flags & ALLOC_OOM ||
5113 (gfp_mask & __GFP_NOMEMALLOC)))
5116 /* Retry as long as the OOM killer is making progress */
5117 if (did_some_progress) {
5118 no_progress_loops = 0;
5123 /* Deal with possible cpuset update races before we fail */
5124 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5128 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5131 if (gfp_mask & __GFP_NOFAIL) {
5133 * All existing users of the __GFP_NOFAIL are blockable, so warn
5134 * of any new users that actually require GFP_NOWAIT
5136 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
5140 * PF_MEMALLOC request from this context is rather bizarre
5141 * because we cannot reclaim anything and only can loop waiting
5142 * for somebody to do a work for us
5144 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
5147 * non failing costly orders are a hard requirement which we
5148 * are not prepared for much so let's warn about these users
5149 * so that we can identify them and convert them to something
5152 WARN_ON_ONCE_GFP(order > PAGE_ALLOC_COSTLY_ORDER, gfp_mask);
5155 * Help non-failing allocations by giving them access to memory
5156 * reserves but do not use ALLOC_NO_WATERMARKS because this
5157 * could deplete whole memory reserves which would just make
5158 * the situation worse
5160 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5168 warn_alloc(gfp_mask, ac->nodemask,
5169 "page allocation failure: order:%u", order);
5174 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5175 int preferred_nid, nodemask_t *nodemask,
5176 struct alloc_context *ac, gfp_t *alloc_gfp,
5177 unsigned int *alloc_flags)
5179 ac->highest_zoneidx = gfp_zone(gfp_mask);
5180 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5181 ac->nodemask = nodemask;
5182 ac->migratetype = gfp_migratetype(gfp_mask);
5184 if (cpusets_enabled()) {
5185 *alloc_gfp |= __GFP_HARDWALL;
5187 * When we are in the interrupt context, it is irrelevant
5188 * to the current task context. It means that any node ok.
5190 if (in_task() && !ac->nodemask)
5191 ac->nodemask = &cpuset_current_mems_allowed;
5193 *alloc_flags |= ALLOC_CPUSET;
5196 might_alloc(gfp_mask);
5198 if (should_fail_alloc_page(gfp_mask, order))
5201 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5203 /* Dirty zone balancing only done in the fast path */
5204 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5207 * The preferred zone is used for statistics but crucially it is
5208 * also used as the starting point for the zonelist iterator. It
5209 * may get reset for allocations that ignore memory policies.
5211 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5212 ac->highest_zoneidx, ac->nodemask);
5218 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5219 * @gfp: GFP flags for the allocation
5220 * @preferred_nid: The preferred NUMA node ID to allocate from
5221 * @nodemask: Set of nodes to allocate from, may be NULL
5222 * @nr_pages: The number of pages desired on the list or array
5223 * @page_list: Optional list to store the allocated pages
5224 * @page_array: Optional array to store the pages
5226 * This is a batched version of the page allocator that attempts to
5227 * allocate nr_pages quickly. Pages are added to page_list if page_list
5228 * is not NULL, otherwise it is assumed that the page_array is valid.
5230 * For lists, nr_pages is the number of pages that should be allocated.
5232 * For arrays, only NULL elements are populated with pages and nr_pages
5233 * is the maximum number of pages that will be stored in the array.
5235 * Returns the number of pages on the list or array.
5237 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5238 nodemask_t *nodemask, int nr_pages,
5239 struct list_head *page_list,
5240 struct page **page_array)
5243 unsigned long flags;
5246 struct per_cpu_pages *pcp;
5247 struct list_head *pcp_list;
5248 struct alloc_context ac;
5250 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5251 int nr_populated = 0, nr_account = 0;
5254 * Skip populated array elements to determine if any pages need
5255 * to be allocated before disabling IRQs.
5257 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5260 /* No pages requested? */
5261 if (unlikely(nr_pages <= 0))
5264 /* Already populated array? */
5265 if (unlikely(page_array && nr_pages - nr_populated == 0))
5268 /* Bulk allocator does not support memcg accounting. */
5269 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5272 /* Use the single page allocator for one page. */
5273 if (nr_pages - nr_populated == 1)
5276 #ifdef CONFIG_PAGE_OWNER
5278 * PAGE_OWNER may recurse into the allocator to allocate space to
5279 * save the stack with pagesets.lock held. Releasing/reacquiring
5280 * removes much of the performance benefit of bulk allocation so
5281 * force the caller to allocate one page at a time as it'll have
5282 * similar performance to added complexity to the bulk allocator.
5284 if (static_branch_unlikely(&page_owner_inited))
5288 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5289 gfp &= gfp_allowed_mask;
5291 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5295 /* Find an allowed local zone that meets the low watermark. */
5296 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5299 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5300 !__cpuset_zone_allowed(zone, gfp)) {
5304 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5305 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5309 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5310 if (zone_watermark_fast(zone, 0, mark,
5311 zonelist_zone_idx(ac.preferred_zoneref),
5312 alloc_flags, gfp)) {
5318 * If there are no allowed local zones that meets the watermarks then
5319 * try to allocate a single page and reclaim if necessary.
5321 if (unlikely(!zone))
5324 /* Attempt the batch allocation */
5325 local_lock_irqsave(&pagesets.lock, flags);
5326 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5327 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5329 while (nr_populated < nr_pages) {
5331 /* Skip existing pages */
5332 if (page_array && page_array[nr_populated]) {
5337 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5339 if (unlikely(!page)) {
5340 /* Try and allocate at least one page */
5347 prep_new_page(page, 0, gfp, 0);
5349 list_add(&page->lru, page_list);
5351 page_array[nr_populated] = page;
5355 local_unlock_irqrestore(&pagesets.lock, flags);
5357 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5358 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5361 return nr_populated;
5364 local_unlock_irqrestore(&pagesets.lock, flags);
5367 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5370 list_add(&page->lru, page_list);
5372 page_array[nr_populated] = page;
5378 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5381 * This is the 'heart' of the zoned buddy allocator.
5383 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5384 nodemask_t *nodemask)
5387 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5388 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5389 struct alloc_context ac = { };
5392 * There are several places where we assume that the order value is sane
5393 * so bail out early if the request is out of bound.
5395 if (WARN_ON_ONCE_GFP(order >= MAX_ORDER, gfp))
5398 gfp &= gfp_allowed_mask;
5400 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5401 * resp. GFP_NOIO which has to be inherited for all allocation requests
5402 * from a particular context which has been marked by
5403 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5404 * movable zones are not used during allocation.
5406 gfp = current_gfp_context(gfp);
5408 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5409 &alloc_gfp, &alloc_flags))
5413 * Forbid the first pass from falling back to types that fragment
5414 * memory until all local zones are considered.
5416 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5418 /* First allocation attempt */
5419 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5424 ac.spread_dirty_pages = false;
5427 * Restore the original nodemask if it was potentially replaced with
5428 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5430 ac.nodemask = nodemask;
5432 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5435 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5436 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5437 __free_pages(page, order);
5441 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5445 EXPORT_SYMBOL(__alloc_pages);
5447 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5448 nodemask_t *nodemask)
5450 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5451 preferred_nid, nodemask);
5453 if (page && order > 1)
5454 prep_transhuge_page(page);
5455 return (struct folio *)page;
5457 EXPORT_SYMBOL(__folio_alloc);
5460 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5461 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5462 * you need to access high mem.
5464 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5468 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5471 return (unsigned long) page_address(page);
5473 EXPORT_SYMBOL(__get_free_pages);
5475 unsigned long get_zeroed_page(gfp_t gfp_mask)
5477 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5479 EXPORT_SYMBOL(get_zeroed_page);
5482 * __free_pages - Free pages allocated with alloc_pages().
5483 * @page: The page pointer returned from alloc_pages().
5484 * @order: The order of the allocation.
5486 * This function can free multi-page allocations that are not compound
5487 * pages. It does not check that the @order passed in matches that of
5488 * the allocation, so it is easy to leak memory. Freeing more memory
5489 * than was allocated will probably emit a warning.
5491 * If the last reference to this page is speculative, it will be released
5492 * by put_page() which only frees the first page of a non-compound
5493 * allocation. To prevent the remaining pages from being leaked, we free
5494 * the subsequent pages here. If you want to use the page's reference
5495 * count to decide when to free the allocation, you should allocate a
5496 * compound page, and use put_page() instead of __free_pages().
5498 * Context: May be called in interrupt context or while holding a normal
5499 * spinlock, but not in NMI context or while holding a raw spinlock.
5501 void __free_pages(struct page *page, unsigned int order)
5503 if (put_page_testzero(page))
5504 free_the_page(page, order);
5505 else if (!PageHead(page))
5507 free_the_page(page + (1 << order), order);
5509 EXPORT_SYMBOL(__free_pages);
5511 void free_pages(unsigned long addr, unsigned int order)
5514 VM_BUG_ON(!virt_addr_valid((void *)addr));
5515 __free_pages(virt_to_page((void *)addr), order);
5519 EXPORT_SYMBOL(free_pages);
5523 * An arbitrary-length arbitrary-offset area of memory which resides
5524 * within a 0 or higher order page. Multiple fragments within that page
5525 * are individually refcounted, in the page's reference counter.
5527 * The page_frag functions below provide a simple allocation framework for
5528 * page fragments. This is used by the network stack and network device
5529 * drivers to provide a backing region of memory for use as either an
5530 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5532 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5535 struct page *page = NULL;
5536 gfp_t gfp = gfp_mask;
5538 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5539 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5541 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5542 PAGE_FRAG_CACHE_MAX_ORDER);
5543 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5545 if (unlikely(!page))
5546 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5548 nc->va = page ? page_address(page) : NULL;
5553 void __page_frag_cache_drain(struct page *page, unsigned int count)
5555 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5557 if (page_ref_sub_and_test(page, count))
5558 free_the_page(page, compound_order(page));
5560 EXPORT_SYMBOL(__page_frag_cache_drain);
5562 void *page_frag_alloc_align(struct page_frag_cache *nc,
5563 unsigned int fragsz, gfp_t gfp_mask,
5564 unsigned int align_mask)
5566 unsigned int size = PAGE_SIZE;
5570 if (unlikely(!nc->va)) {
5572 page = __page_frag_cache_refill(nc, gfp_mask);
5576 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5577 /* if size can vary use size else just use PAGE_SIZE */
5580 /* Even if we own the page, we do not use atomic_set().
5581 * This would break get_page_unless_zero() users.
5583 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5585 /* reset page count bias and offset to start of new frag */
5586 nc->pfmemalloc = page_is_pfmemalloc(page);
5587 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5591 offset = nc->offset - fragsz;
5592 if (unlikely(offset < 0)) {
5593 page = virt_to_page(nc->va);
5595 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5598 if (unlikely(nc->pfmemalloc)) {
5599 free_the_page(page, compound_order(page));
5603 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5604 /* if size can vary use size else just use PAGE_SIZE */
5607 /* OK, page count is 0, we can safely set it */
5608 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5610 /* reset page count bias and offset to start of new frag */
5611 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5612 offset = size - fragsz;
5616 offset &= align_mask;
5617 nc->offset = offset;
5619 return nc->va + offset;
5621 EXPORT_SYMBOL(page_frag_alloc_align);
5624 * Frees a page fragment allocated out of either a compound or order 0 page.
5626 void page_frag_free(void *addr)
5628 struct page *page = virt_to_head_page(addr);
5630 if (unlikely(put_page_testzero(page)))
5631 free_the_page(page, compound_order(page));
5633 EXPORT_SYMBOL(page_frag_free);
5635 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5639 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5640 unsigned long used = addr + PAGE_ALIGN(size);
5642 split_page(virt_to_page((void *)addr), order);
5643 while (used < alloc_end) {
5648 return (void *)addr;
5652 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5653 * @size: the number of bytes to allocate
5654 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5656 * This function is similar to alloc_pages(), except that it allocates the
5657 * minimum number of pages to satisfy the request. alloc_pages() can only
5658 * allocate memory in power-of-two pages.
5660 * This function is also limited by MAX_ORDER.
5662 * Memory allocated by this function must be released by free_pages_exact().
5664 * Return: pointer to the allocated area or %NULL in case of error.
5666 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5668 unsigned int order = get_order(size);
5671 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5672 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5674 addr = __get_free_pages(gfp_mask, order);
5675 return make_alloc_exact(addr, order, size);
5677 EXPORT_SYMBOL(alloc_pages_exact);
5680 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5682 * @nid: the preferred node ID where memory should be allocated
5683 * @size: the number of bytes to allocate
5684 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5686 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5689 * Return: pointer to the allocated area or %NULL in case of error.
5691 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5693 unsigned int order = get_order(size);
5696 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5697 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5699 p = alloc_pages_node(nid, gfp_mask, order);
5702 return make_alloc_exact((unsigned long)page_address(p), order, size);
5706 * free_pages_exact - release memory allocated via alloc_pages_exact()
5707 * @virt: the value returned by alloc_pages_exact.
5708 * @size: size of allocation, same value as passed to alloc_pages_exact().
5710 * Release the memory allocated by a previous call to alloc_pages_exact.
5712 void free_pages_exact(void *virt, size_t size)
5714 unsigned long addr = (unsigned long)virt;
5715 unsigned long end = addr + PAGE_ALIGN(size);
5717 while (addr < end) {
5722 EXPORT_SYMBOL(free_pages_exact);
5725 * nr_free_zone_pages - count number of pages beyond high watermark
5726 * @offset: The zone index of the highest zone
5728 * nr_free_zone_pages() counts the number of pages which are beyond the
5729 * high watermark within all zones at or below a given zone index. For each
5730 * zone, the number of pages is calculated as:
5732 * nr_free_zone_pages = managed_pages - high_pages
5734 * Return: number of pages beyond high watermark.
5736 static unsigned long nr_free_zone_pages(int offset)
5741 /* Just pick one node, since fallback list is circular */
5742 unsigned long sum = 0;
5744 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5746 for_each_zone_zonelist(zone, z, zonelist, offset) {
5747 unsigned long size = zone_managed_pages(zone);
5748 unsigned long high = high_wmark_pages(zone);
5757 * nr_free_buffer_pages - count number of pages beyond high watermark
5759 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5760 * watermark within ZONE_DMA and ZONE_NORMAL.
5762 * Return: number of pages beyond high watermark within ZONE_DMA and
5765 unsigned long nr_free_buffer_pages(void)
5767 return nr_free_zone_pages(gfp_zone(GFP_USER));
5769 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5771 static inline void show_node(struct zone *zone)
5773 if (IS_ENABLED(CONFIG_NUMA))
5774 printk("Node %d ", zone_to_nid(zone));
5777 long si_mem_available(void)
5780 unsigned long pagecache;
5781 unsigned long wmark_low = 0;
5782 unsigned long pages[NR_LRU_LISTS];
5783 unsigned long reclaimable;
5787 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5788 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5791 wmark_low += low_wmark_pages(zone);
5794 * Estimate the amount of memory available for userspace allocations,
5795 * without causing swapping.
5797 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5800 * Not all the page cache can be freed, otherwise the system will
5801 * start swapping. Assume at least half of the page cache, or the
5802 * low watermark worth of cache, needs to stay.
5804 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5805 pagecache -= min(pagecache / 2, wmark_low);
5806 available += pagecache;
5809 * Part of the reclaimable slab and other kernel memory consists of
5810 * items that are in use, and cannot be freed. Cap this estimate at the
5813 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5814 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5815 available += reclaimable - min(reclaimable / 2, wmark_low);
5821 EXPORT_SYMBOL_GPL(si_mem_available);
5823 void si_meminfo(struct sysinfo *val)
5825 val->totalram = totalram_pages();
5826 val->sharedram = global_node_page_state(NR_SHMEM);
5827 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5828 val->bufferram = nr_blockdev_pages();
5829 val->totalhigh = totalhigh_pages();
5830 val->freehigh = nr_free_highpages();
5831 val->mem_unit = PAGE_SIZE;
5834 EXPORT_SYMBOL(si_meminfo);
5837 void si_meminfo_node(struct sysinfo *val, int nid)
5839 int zone_type; /* needs to be signed */
5840 unsigned long managed_pages = 0;
5841 unsigned long managed_highpages = 0;
5842 unsigned long free_highpages = 0;
5843 pg_data_t *pgdat = NODE_DATA(nid);
5845 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5846 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5847 val->totalram = managed_pages;
5848 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5849 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5850 #ifdef CONFIG_HIGHMEM
5851 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5852 struct zone *zone = &pgdat->node_zones[zone_type];
5854 if (is_highmem(zone)) {
5855 managed_highpages += zone_managed_pages(zone);
5856 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5859 val->totalhigh = managed_highpages;
5860 val->freehigh = free_highpages;
5862 val->totalhigh = managed_highpages;
5863 val->freehigh = free_highpages;
5865 val->mem_unit = PAGE_SIZE;
5870 * Determine whether the node should be displayed or not, depending on whether
5871 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5873 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5875 if (!(flags & SHOW_MEM_FILTER_NODES))
5879 * no node mask - aka implicit memory numa policy. Do not bother with
5880 * the synchronization - read_mems_allowed_begin - because we do not
5881 * have to be precise here.
5884 nodemask = &cpuset_current_mems_allowed;
5886 return !node_isset(nid, *nodemask);
5889 #define K(x) ((x) << (PAGE_SHIFT-10))
5891 static void show_migration_types(unsigned char type)
5893 static const char types[MIGRATE_TYPES] = {
5894 [MIGRATE_UNMOVABLE] = 'U',
5895 [MIGRATE_MOVABLE] = 'M',
5896 [MIGRATE_RECLAIMABLE] = 'E',
5897 [MIGRATE_HIGHATOMIC] = 'H',
5899 [MIGRATE_CMA] = 'C',
5901 #ifdef CONFIG_MEMORY_ISOLATION
5902 [MIGRATE_ISOLATE] = 'I',
5905 char tmp[MIGRATE_TYPES + 1];
5909 for (i = 0; i < MIGRATE_TYPES; i++) {
5910 if (type & (1 << i))
5915 printk(KERN_CONT "(%s) ", tmp);
5919 * Show free area list (used inside shift_scroll-lock stuff)
5920 * We also calculate the percentage fragmentation. We do this by counting the
5921 * memory on each free list with the exception of the first item on the list.
5924 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5927 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5929 unsigned long free_pcp = 0;
5934 for_each_populated_zone(zone) {
5935 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5938 for_each_online_cpu(cpu)
5939 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5942 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5943 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5944 " unevictable:%lu dirty:%lu writeback:%lu\n"
5945 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5946 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5947 " kernel_misc_reclaimable:%lu\n"
5948 " free:%lu free_pcp:%lu free_cma:%lu\n",
5949 global_node_page_state(NR_ACTIVE_ANON),
5950 global_node_page_state(NR_INACTIVE_ANON),
5951 global_node_page_state(NR_ISOLATED_ANON),
5952 global_node_page_state(NR_ACTIVE_FILE),
5953 global_node_page_state(NR_INACTIVE_FILE),
5954 global_node_page_state(NR_ISOLATED_FILE),
5955 global_node_page_state(NR_UNEVICTABLE),
5956 global_node_page_state(NR_FILE_DIRTY),
5957 global_node_page_state(NR_WRITEBACK),
5958 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5959 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5960 global_node_page_state(NR_FILE_MAPPED),
5961 global_node_page_state(NR_SHMEM),
5962 global_node_page_state(NR_PAGETABLE),
5963 global_zone_page_state(NR_BOUNCE),
5964 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5965 global_zone_page_state(NR_FREE_PAGES),
5967 global_zone_page_state(NR_FREE_CMA_PAGES));
5969 for_each_online_pgdat(pgdat) {
5970 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5974 " active_anon:%lukB"
5975 " inactive_anon:%lukB"
5976 " active_file:%lukB"
5977 " inactive_file:%lukB"
5978 " unevictable:%lukB"
5979 " isolated(anon):%lukB"
5980 " isolated(file):%lukB"
5985 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5987 " shmem_pmdmapped: %lukB"
5990 " writeback_tmp:%lukB"
5991 " kernel_stack:%lukB"
5992 #ifdef CONFIG_SHADOW_CALL_STACK
5993 " shadow_call_stack:%lukB"
5996 " all_unreclaimable? %s"
5999 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
6000 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
6001 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
6002 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
6003 K(node_page_state(pgdat, NR_UNEVICTABLE)),
6004 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
6005 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
6006 K(node_page_state(pgdat, NR_FILE_MAPPED)),
6007 K(node_page_state(pgdat, NR_FILE_DIRTY)),
6008 K(node_page_state(pgdat, NR_WRITEBACK)),
6009 K(node_page_state(pgdat, NR_SHMEM)),
6010 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6011 K(node_page_state(pgdat, NR_SHMEM_THPS)),
6012 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6013 K(node_page_state(pgdat, NR_ANON_THPS)),
6015 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6016 node_page_state(pgdat, NR_KERNEL_STACK_KB),
6017 #ifdef CONFIG_SHADOW_CALL_STACK
6018 node_page_state(pgdat, NR_KERNEL_SCS_KB),
6020 K(node_page_state(pgdat, NR_PAGETABLE)),
6021 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6025 for_each_populated_zone(zone) {
6028 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6032 for_each_online_cpu(cpu)
6033 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6043 " reserved_highatomic:%luKB"
6044 " active_anon:%lukB"
6045 " inactive_anon:%lukB"
6046 " active_file:%lukB"
6047 " inactive_file:%lukB"
6048 " unevictable:%lukB"
6049 " writepending:%lukB"
6059 K(zone_page_state(zone, NR_FREE_PAGES)),
6060 K(zone->watermark_boost),
6061 K(min_wmark_pages(zone)),
6062 K(low_wmark_pages(zone)),
6063 K(high_wmark_pages(zone)),
6064 K(zone->nr_reserved_highatomic),
6065 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6066 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6067 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6068 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6069 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6070 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6071 K(zone->present_pages),
6072 K(zone_managed_pages(zone)),
6073 K(zone_page_state(zone, NR_MLOCK)),
6074 K(zone_page_state(zone, NR_BOUNCE)),
6076 K(this_cpu_read(zone->per_cpu_pageset->count)),
6077 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6078 printk("lowmem_reserve[]:");
6079 for (i = 0; i < MAX_NR_ZONES; i++)
6080 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6081 printk(KERN_CONT "\n");
6084 for_each_populated_zone(zone) {
6086 unsigned long nr[MAX_ORDER], flags, total = 0;
6087 unsigned char types[MAX_ORDER];
6089 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6092 printk(KERN_CONT "%s: ", zone->name);
6094 spin_lock_irqsave(&zone->lock, flags);
6095 for (order = 0; order < MAX_ORDER; order++) {
6096 struct free_area *area = &zone->free_area[order];
6099 nr[order] = area->nr_free;
6100 total += nr[order] << order;
6103 for (type = 0; type < MIGRATE_TYPES; type++) {
6104 if (!free_area_empty(area, type))
6105 types[order] |= 1 << type;
6108 spin_unlock_irqrestore(&zone->lock, flags);
6109 for (order = 0; order < MAX_ORDER; order++) {
6110 printk(KERN_CONT "%lu*%lukB ",
6111 nr[order], K(1UL) << order);
6113 show_migration_types(types[order]);
6115 printk(KERN_CONT "= %lukB\n", K(total));
6118 hugetlb_show_meminfo();
6120 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6122 show_swap_cache_info();
6125 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6127 zoneref->zone = zone;
6128 zoneref->zone_idx = zone_idx(zone);
6132 * Builds allocation fallback zone lists.
6134 * Add all populated zones of a node to the zonelist.
6136 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6139 enum zone_type zone_type = MAX_NR_ZONES;
6144 zone = pgdat->node_zones + zone_type;
6145 if (populated_zone(zone)) {
6146 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6147 check_highest_zone(zone_type);
6149 } while (zone_type);
6156 static int __parse_numa_zonelist_order(char *s)
6159 * We used to support different zonelists modes but they turned
6160 * out to be just not useful. Let's keep the warning in place
6161 * if somebody still use the cmd line parameter so that we do
6162 * not fail it silently
6164 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6165 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6171 char numa_zonelist_order[] = "Node";
6174 * sysctl handler for numa_zonelist_order
6176 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6177 void *buffer, size_t *length, loff_t *ppos)
6180 return __parse_numa_zonelist_order(buffer);
6181 return proc_dostring(table, write, buffer, length, ppos);
6185 static int node_load[MAX_NUMNODES];
6188 * find_next_best_node - find the next node that should appear in a given node's fallback list
6189 * @node: node whose fallback list we're appending
6190 * @used_node_mask: nodemask_t of already used nodes
6192 * We use a number of factors to determine which is the next node that should
6193 * appear on a given node's fallback list. The node should not have appeared
6194 * already in @node's fallback list, and it should be the next closest node
6195 * according to the distance array (which contains arbitrary distance values
6196 * from each node to each node in the system), and should also prefer nodes
6197 * with no CPUs, since presumably they'll have very little allocation pressure
6198 * on them otherwise.
6200 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6202 int find_next_best_node(int node, nodemask_t *used_node_mask)
6205 int min_val = INT_MAX;
6206 int best_node = NUMA_NO_NODE;
6208 /* Use the local node if we haven't already */
6209 if (!node_isset(node, *used_node_mask)) {
6210 node_set(node, *used_node_mask);
6214 for_each_node_state(n, N_MEMORY) {
6216 /* Don't want a node to appear more than once */
6217 if (node_isset(n, *used_node_mask))
6220 /* Use the distance array to find the distance */
6221 val = node_distance(node, n);
6223 /* Penalize nodes under us ("prefer the next node") */
6226 /* Give preference to headless and unused nodes */
6227 if (!cpumask_empty(cpumask_of_node(n)))
6228 val += PENALTY_FOR_NODE_WITH_CPUS;
6230 /* Slight preference for less loaded node */
6231 val *= MAX_NUMNODES;
6232 val += node_load[n];
6234 if (val < min_val) {
6241 node_set(best_node, *used_node_mask);
6248 * Build zonelists ordered by node and zones within node.
6249 * This results in maximum locality--normal zone overflows into local
6250 * DMA zone, if any--but risks exhausting DMA zone.
6252 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6255 struct zoneref *zonerefs;
6258 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6260 for (i = 0; i < nr_nodes; i++) {
6263 pg_data_t *node = NODE_DATA(node_order[i]);
6265 nr_zones = build_zonerefs_node(node, zonerefs);
6266 zonerefs += nr_zones;
6268 zonerefs->zone = NULL;
6269 zonerefs->zone_idx = 0;
6273 * Build gfp_thisnode zonelists
6275 static void build_thisnode_zonelists(pg_data_t *pgdat)
6277 struct zoneref *zonerefs;
6280 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6281 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6282 zonerefs += nr_zones;
6283 zonerefs->zone = NULL;
6284 zonerefs->zone_idx = 0;
6288 * Build zonelists ordered by zone and nodes within zones.
6289 * This results in conserving DMA zone[s] until all Normal memory is
6290 * exhausted, but results in overflowing to remote node while memory
6291 * may still exist in local DMA zone.
6294 static void build_zonelists(pg_data_t *pgdat)
6296 static int node_order[MAX_NUMNODES];
6297 int node, nr_nodes = 0;
6298 nodemask_t used_mask = NODE_MASK_NONE;
6299 int local_node, prev_node;
6301 /* NUMA-aware ordering of nodes */
6302 local_node = pgdat->node_id;
6303 prev_node = local_node;
6305 memset(node_order, 0, sizeof(node_order));
6306 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6308 * We don't want to pressure a particular node.
6309 * So adding penalty to the first node in same
6310 * distance group to make it round-robin.
6312 if (node_distance(local_node, node) !=
6313 node_distance(local_node, prev_node))
6314 node_load[node] += 1;
6316 node_order[nr_nodes++] = node;
6320 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6321 build_thisnode_zonelists(pgdat);
6322 pr_info("Fallback order for Node %d: ", local_node);
6323 for (node = 0; node < nr_nodes; node++)
6324 pr_cont("%d ", node_order[node]);
6328 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6330 * Return node id of node used for "local" allocations.
6331 * I.e., first node id of first zone in arg node's generic zonelist.
6332 * Used for initializing percpu 'numa_mem', which is used primarily
6333 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6335 int local_memory_node(int node)
6339 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6340 gfp_zone(GFP_KERNEL),
6342 return zone_to_nid(z->zone);
6346 static void setup_min_unmapped_ratio(void);
6347 static void setup_min_slab_ratio(void);
6348 #else /* CONFIG_NUMA */
6350 static void build_zonelists(pg_data_t *pgdat)
6352 int node, local_node;
6353 struct zoneref *zonerefs;
6356 local_node = pgdat->node_id;
6358 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6359 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6360 zonerefs += nr_zones;
6363 * Now we build the zonelist so that it contains the zones
6364 * of all the other nodes.
6365 * We don't want to pressure a particular node, so when
6366 * building the zones for node N, we make sure that the
6367 * zones coming right after the local ones are those from
6368 * node N+1 (modulo N)
6370 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6371 if (!node_online(node))
6373 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6374 zonerefs += nr_zones;
6376 for (node = 0; node < local_node; node++) {
6377 if (!node_online(node))
6379 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6380 zonerefs += nr_zones;
6383 zonerefs->zone = NULL;
6384 zonerefs->zone_idx = 0;
6387 #endif /* CONFIG_NUMA */
6390 * Boot pageset table. One per cpu which is going to be used for all
6391 * zones and all nodes. The parameters will be set in such a way
6392 * that an item put on a list will immediately be handed over to
6393 * the buddy list. This is safe since pageset manipulation is done
6394 * with interrupts disabled.
6396 * The boot_pagesets must be kept even after bootup is complete for
6397 * unused processors and/or zones. They do play a role for bootstrapping
6398 * hotplugged processors.
6400 * zoneinfo_show() and maybe other functions do
6401 * not check if the processor is online before following the pageset pointer.
6402 * Other parts of the kernel may not check if the zone is available.
6404 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6405 /* These effectively disable the pcplists in the boot pageset completely */
6406 #define BOOT_PAGESET_HIGH 0
6407 #define BOOT_PAGESET_BATCH 1
6408 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6409 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6410 DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6412 static void __build_all_zonelists(void *data)
6415 int __maybe_unused cpu;
6416 pg_data_t *self = data;
6417 static DEFINE_SPINLOCK(lock);
6422 memset(node_load, 0, sizeof(node_load));
6426 * This node is hotadded and no memory is yet present. So just
6427 * building zonelists is fine - no need to touch other nodes.
6429 if (self && !node_online(self->node_id)) {
6430 build_zonelists(self);
6433 * All possible nodes have pgdat preallocated
6436 for_each_node(nid) {
6437 pg_data_t *pgdat = NODE_DATA(nid);
6439 build_zonelists(pgdat);
6442 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6444 * We now know the "local memory node" for each node--
6445 * i.e., the node of the first zone in the generic zonelist.
6446 * Set up numa_mem percpu variable for on-line cpus. During
6447 * boot, only the boot cpu should be on-line; we'll init the
6448 * secondary cpus' numa_mem as they come on-line. During
6449 * node/memory hotplug, we'll fixup all on-line cpus.
6451 for_each_online_cpu(cpu)
6452 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6459 static noinline void __init
6460 build_all_zonelists_init(void)
6464 __build_all_zonelists(NULL);
6467 * Initialize the boot_pagesets that are going to be used
6468 * for bootstrapping processors. The real pagesets for
6469 * each zone will be allocated later when the per cpu
6470 * allocator is available.
6472 * boot_pagesets are used also for bootstrapping offline
6473 * cpus if the system is already booted because the pagesets
6474 * are needed to initialize allocators on a specific cpu too.
6475 * F.e. the percpu allocator needs the page allocator which
6476 * needs the percpu allocator in order to allocate its pagesets
6477 * (a chicken-egg dilemma).
6479 for_each_possible_cpu(cpu)
6480 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6482 mminit_verify_zonelist();
6483 cpuset_init_current_mems_allowed();
6487 * unless system_state == SYSTEM_BOOTING.
6489 * __ref due to call of __init annotated helper build_all_zonelists_init
6490 * [protected by SYSTEM_BOOTING].
6492 void __ref build_all_zonelists(pg_data_t *pgdat)
6494 unsigned long vm_total_pages;
6496 if (system_state == SYSTEM_BOOTING) {
6497 build_all_zonelists_init();
6499 __build_all_zonelists(pgdat);
6500 /* cpuset refresh routine should be here */
6502 /* Get the number of free pages beyond high watermark in all zones. */
6503 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6505 * Disable grouping by mobility if the number of pages in the
6506 * system is too low to allow the mechanism to work. It would be
6507 * more accurate, but expensive to check per-zone. This check is
6508 * made on memory-hotadd so a system can start with mobility
6509 * disabled and enable it later
6511 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6512 page_group_by_mobility_disabled = 1;
6514 page_group_by_mobility_disabled = 0;
6516 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6518 page_group_by_mobility_disabled ? "off" : "on",
6521 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6525 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6526 static bool __meminit
6527 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6529 static struct memblock_region *r;
6531 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6532 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6533 for_each_mem_region(r) {
6534 if (*pfn < memblock_region_memory_end_pfn(r))
6538 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6539 memblock_is_mirror(r)) {
6540 *pfn = memblock_region_memory_end_pfn(r);
6548 * Initially all pages are reserved - free ones are freed
6549 * up by memblock_free_all() once the early boot process is
6550 * done. Non-atomic initialization, single-pass.
6552 * All aligned pageblocks are initialized to the specified migratetype
6553 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6554 * zone stats (e.g., nr_isolate_pageblock) are touched.
6556 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6557 unsigned long start_pfn, unsigned long zone_end_pfn,
6558 enum meminit_context context,
6559 struct vmem_altmap *altmap, int migratetype)
6561 unsigned long pfn, end_pfn = start_pfn + size;
6564 if (highest_memmap_pfn < end_pfn - 1)
6565 highest_memmap_pfn = end_pfn - 1;
6567 #ifdef CONFIG_ZONE_DEVICE
6569 * Honor reservation requested by the driver for this ZONE_DEVICE
6570 * memory. We limit the total number of pages to initialize to just
6571 * those that might contain the memory mapping. We will defer the
6572 * ZONE_DEVICE page initialization until after we have released
6575 if (zone == ZONE_DEVICE) {
6579 if (start_pfn == altmap->base_pfn)
6580 start_pfn += altmap->reserve;
6581 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6585 for (pfn = start_pfn; pfn < end_pfn; ) {
6587 * There can be holes in boot-time mem_map[]s handed to this
6588 * function. They do not exist on hotplugged memory.
6590 if (context == MEMINIT_EARLY) {
6591 if (overlap_memmap_init(zone, &pfn))
6593 if (defer_init(nid, pfn, zone_end_pfn))
6597 page = pfn_to_page(pfn);
6598 __init_single_page(page, pfn, zone, nid);
6599 if (context == MEMINIT_HOTPLUG)
6600 __SetPageReserved(page);
6603 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6604 * such that unmovable allocations won't be scattered all
6605 * over the place during system boot.
6607 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6608 set_pageblock_migratetype(page, migratetype);
6615 #ifdef CONFIG_ZONE_DEVICE
6616 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6617 unsigned long zone_idx, int nid,
6618 struct dev_pagemap *pgmap)
6621 __init_single_page(page, pfn, zone_idx, nid);
6624 * Mark page reserved as it will need to wait for onlining
6625 * phase for it to be fully associated with a zone.
6627 * We can use the non-atomic __set_bit operation for setting
6628 * the flag as we are still initializing the pages.
6630 __SetPageReserved(page);
6633 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6634 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6635 * ever freed or placed on a driver-private list.
6637 page->pgmap = pgmap;
6638 page->zone_device_data = NULL;
6641 * Mark the block movable so that blocks are reserved for
6642 * movable at startup. This will force kernel allocations
6643 * to reserve their blocks rather than leaking throughout
6644 * the address space during boot when many long-lived
6645 * kernel allocations are made.
6647 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6648 * because this is done early in section_activate()
6650 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6651 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6657 * With compound page geometry and when struct pages are stored in ram most
6658 * tail pages are reused. Consequently, the amount of unique struct pages to
6659 * initialize is a lot smaller that the total amount of struct pages being
6660 * mapped. This is a paired / mild layering violation with explicit knowledge
6661 * of how the sparse_vmemmap internals handle compound pages in the lack
6662 * of an altmap. See vmemmap_populate_compound_pages().
6664 static inline unsigned long compound_nr_pages(struct vmem_altmap *altmap,
6665 unsigned long nr_pages)
6667 return is_power_of_2(sizeof(struct page)) &&
6668 !altmap ? 2 * (PAGE_SIZE / sizeof(struct page)) : nr_pages;
6671 static void __ref memmap_init_compound(struct page *head,
6672 unsigned long head_pfn,
6673 unsigned long zone_idx, int nid,
6674 struct dev_pagemap *pgmap,
6675 unsigned long nr_pages)
6677 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6678 unsigned int order = pgmap->vmemmap_shift;
6680 __SetPageHead(head);
6681 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6682 struct page *page = pfn_to_page(pfn);
6684 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6685 prep_compound_tail(head, pfn - head_pfn);
6686 set_page_count(page, 0);
6689 * The first tail page stores compound_mapcount_ptr() and
6690 * compound_order() and the second tail page stores
6691 * compound_pincount_ptr(). Call prep_compound_head() after
6692 * the first and second tail pages have been initialized to
6693 * not have the data overwritten.
6695 if (pfn == head_pfn + 2)
6696 prep_compound_head(head, order);
6700 void __ref memmap_init_zone_device(struct zone *zone,
6701 unsigned long start_pfn,
6702 unsigned long nr_pages,
6703 struct dev_pagemap *pgmap)
6705 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6706 struct pglist_data *pgdat = zone->zone_pgdat;
6707 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6708 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6709 unsigned long zone_idx = zone_idx(zone);
6710 unsigned long start = jiffies;
6711 int nid = pgdat->node_id;
6713 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6717 * The call to memmap_init should have already taken care
6718 * of the pages reserved for the memmap, so we can just jump to
6719 * the end of that region and start processing the device pages.
6722 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6723 nr_pages = end_pfn - start_pfn;
6726 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6727 struct page *page = pfn_to_page(pfn);
6729 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6731 if (pfns_per_compound == 1)
6734 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6735 compound_nr_pages(altmap, pfns_per_compound));
6738 pr_info("%s initialised %lu pages in %ums\n", __func__,
6739 nr_pages, jiffies_to_msecs(jiffies - start));
6743 static void __meminit zone_init_free_lists(struct zone *zone)
6745 unsigned int order, t;
6746 for_each_migratetype_order(order, t) {
6747 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6748 zone->free_area[order].nr_free = 0;
6753 * Only struct pages that correspond to ranges defined by memblock.memory
6754 * are zeroed and initialized by going through __init_single_page() during
6755 * memmap_init_zone_range().
6757 * But, there could be struct pages that correspond to holes in
6758 * memblock.memory. This can happen because of the following reasons:
6759 * - physical memory bank size is not necessarily the exact multiple of the
6760 * arbitrary section size
6761 * - early reserved memory may not be listed in memblock.memory
6762 * - memory layouts defined with memmap= kernel parameter may not align
6763 * nicely with memmap sections
6765 * Explicitly initialize those struct pages so that:
6766 * - PG_Reserved is set
6767 * - zone and node links point to zone and node that span the page if the
6768 * hole is in the middle of a zone
6769 * - zone and node links point to adjacent zone/node if the hole falls on
6770 * the zone boundary; the pages in such holes will be prepended to the
6771 * zone/node above the hole except for the trailing pages in the last
6772 * section that will be appended to the zone/node below.
6774 static void __init init_unavailable_range(unsigned long spfn,
6781 for (pfn = spfn; pfn < epfn; pfn++) {
6782 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6783 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6784 + pageblock_nr_pages - 1;
6787 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6788 __SetPageReserved(pfn_to_page(pfn));
6793 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6794 node, zone_names[zone], pgcnt);
6797 static void __init memmap_init_zone_range(struct zone *zone,
6798 unsigned long start_pfn,
6799 unsigned long end_pfn,
6800 unsigned long *hole_pfn)
6802 unsigned long zone_start_pfn = zone->zone_start_pfn;
6803 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6804 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6806 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6807 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6809 if (start_pfn >= end_pfn)
6812 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6813 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6815 if (*hole_pfn < start_pfn)
6816 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6818 *hole_pfn = end_pfn;
6821 static void __init memmap_init(void)
6823 unsigned long start_pfn, end_pfn;
6824 unsigned long hole_pfn = 0;
6825 int i, j, zone_id = 0, nid;
6827 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6828 struct pglist_data *node = NODE_DATA(nid);
6830 for (j = 0; j < MAX_NR_ZONES; j++) {
6831 struct zone *zone = node->node_zones + j;
6833 if (!populated_zone(zone))
6836 memmap_init_zone_range(zone, start_pfn, end_pfn,
6842 #ifdef CONFIG_SPARSEMEM
6844 * Initialize the memory map for hole in the range [memory_end,
6846 * Append the pages in this hole to the highest zone in the last
6848 * The call to init_unavailable_range() is outside the ifdef to
6849 * silence the compiler warining about zone_id set but not used;
6850 * for FLATMEM it is a nop anyway
6852 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6853 if (hole_pfn < end_pfn)
6855 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6858 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6859 phys_addr_t min_addr, int nid, bool exact_nid)
6864 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6865 MEMBLOCK_ALLOC_ACCESSIBLE,
6868 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6869 MEMBLOCK_ALLOC_ACCESSIBLE,
6872 if (ptr && size > 0)
6873 page_init_poison(ptr, size);
6878 static int zone_batchsize(struct zone *zone)
6884 * The number of pages to batch allocate is either ~0.1%
6885 * of the zone or 1MB, whichever is smaller. The batch
6886 * size is striking a balance between allocation latency
6887 * and zone lock contention.
6889 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6890 batch /= 4; /* We effectively *= 4 below */
6895 * Clamp the batch to a 2^n - 1 value. Having a power
6896 * of 2 value was found to be more likely to have
6897 * suboptimal cache aliasing properties in some cases.
6899 * For example if 2 tasks are alternately allocating
6900 * batches of pages, one task can end up with a lot
6901 * of pages of one half of the possible page colors
6902 * and the other with pages of the other colors.
6904 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6909 /* The deferral and batching of frees should be suppressed under NOMMU
6912 * The problem is that NOMMU needs to be able to allocate large chunks
6913 * of contiguous memory as there's no hardware page translation to
6914 * assemble apparent contiguous memory from discontiguous pages.
6916 * Queueing large contiguous runs of pages for batching, however,
6917 * causes the pages to actually be freed in smaller chunks. As there
6918 * can be a significant delay between the individual batches being
6919 * recycled, this leads to the once large chunks of space being
6920 * fragmented and becoming unavailable for high-order allocations.
6926 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6931 unsigned long total_pages;
6933 if (!percpu_pagelist_high_fraction) {
6935 * By default, the high value of the pcp is based on the zone
6936 * low watermark so that if they are full then background
6937 * reclaim will not be started prematurely.
6939 total_pages = low_wmark_pages(zone);
6942 * If percpu_pagelist_high_fraction is configured, the high
6943 * value is based on a fraction of the managed pages in the
6946 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6950 * Split the high value across all online CPUs local to the zone. Note
6951 * that early in boot that CPUs may not be online yet and that during
6952 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6953 * onlined. For memory nodes that have no CPUs, split pcp->high across
6954 * all online CPUs to mitigate the risk that reclaim is triggered
6955 * prematurely due to pages stored on pcp lists.
6957 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6959 nr_split_cpus = num_online_cpus();
6960 high = total_pages / nr_split_cpus;
6963 * Ensure high is at least batch*4. The multiple is based on the
6964 * historical relationship between high and batch.
6966 high = max(high, batch << 2);
6975 * pcp->high and pcp->batch values are related and generally batch is lower
6976 * than high. They are also related to pcp->count such that count is lower
6977 * than high, and as soon as it reaches high, the pcplist is flushed.
6979 * However, guaranteeing these relations at all times would require e.g. write
6980 * barriers here but also careful usage of read barriers at the read side, and
6981 * thus be prone to error and bad for performance. Thus the update only prevents
6982 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6983 * can cope with those fields changing asynchronously, and fully trust only the
6984 * pcp->count field on the local CPU with interrupts disabled.
6986 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6987 * outside of boot time (or some other assurance that no concurrent updaters
6990 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6991 unsigned long batch)
6993 WRITE_ONCE(pcp->batch, batch);
6994 WRITE_ONCE(pcp->high, high);
6997 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
7001 memset(pcp, 0, sizeof(*pcp));
7002 memset(pzstats, 0, sizeof(*pzstats));
7004 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
7005 INIT_LIST_HEAD(&pcp->lists[pindex]);
7008 * Set batch and high values safe for a boot pageset. A true percpu
7009 * pageset's initialization will update them subsequently. Here we don't
7010 * need to be as careful as pageset_update() as nobody can access the
7013 pcp->high = BOOT_PAGESET_HIGH;
7014 pcp->batch = BOOT_PAGESET_BATCH;
7015 pcp->free_factor = 0;
7018 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
7019 unsigned long batch)
7021 struct per_cpu_pages *pcp;
7024 for_each_possible_cpu(cpu) {
7025 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7026 pageset_update(pcp, high, batch);
7031 * Calculate and set new high and batch values for all per-cpu pagesets of a
7032 * zone based on the zone's size.
7034 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
7036 int new_high, new_batch;
7038 new_batch = max(1, zone_batchsize(zone));
7039 new_high = zone_highsize(zone, new_batch, cpu_online);
7041 if (zone->pageset_high == new_high &&
7042 zone->pageset_batch == new_batch)
7045 zone->pageset_high = new_high;
7046 zone->pageset_batch = new_batch;
7048 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7051 void __meminit setup_zone_pageset(struct zone *zone)
7055 /* Size may be 0 on !SMP && !NUMA */
7056 if (sizeof(struct per_cpu_zonestat) > 0)
7057 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7059 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7060 for_each_possible_cpu(cpu) {
7061 struct per_cpu_pages *pcp;
7062 struct per_cpu_zonestat *pzstats;
7064 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7065 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7066 per_cpu_pages_init(pcp, pzstats);
7069 zone_set_pageset_high_and_batch(zone, 0);
7073 * Allocate per cpu pagesets and initialize them.
7074 * Before this call only boot pagesets were available.
7076 void __init setup_per_cpu_pageset(void)
7078 struct pglist_data *pgdat;
7080 int __maybe_unused cpu;
7082 for_each_populated_zone(zone)
7083 setup_zone_pageset(zone);
7087 * Unpopulated zones continue using the boot pagesets.
7088 * The numa stats for these pagesets need to be reset.
7089 * Otherwise, they will end up skewing the stats of
7090 * the nodes these zones are associated with.
7092 for_each_possible_cpu(cpu) {
7093 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7094 memset(pzstats->vm_numa_event, 0,
7095 sizeof(pzstats->vm_numa_event));
7099 for_each_online_pgdat(pgdat)
7100 pgdat->per_cpu_nodestats =
7101 alloc_percpu(struct per_cpu_nodestat);
7104 static __meminit void zone_pcp_init(struct zone *zone)
7107 * per cpu subsystem is not up at this point. The following code
7108 * relies on the ability of the linker to provide the
7109 * offset of a (static) per cpu variable into the per cpu area.
7111 zone->per_cpu_pageset = &boot_pageset;
7112 zone->per_cpu_zonestats = &boot_zonestats;
7113 zone->pageset_high = BOOT_PAGESET_HIGH;
7114 zone->pageset_batch = BOOT_PAGESET_BATCH;
7116 if (populated_zone(zone))
7117 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7118 zone->present_pages, zone_batchsize(zone));
7121 void __meminit init_currently_empty_zone(struct zone *zone,
7122 unsigned long zone_start_pfn,
7125 struct pglist_data *pgdat = zone->zone_pgdat;
7126 int zone_idx = zone_idx(zone) + 1;
7128 if (zone_idx > pgdat->nr_zones)
7129 pgdat->nr_zones = zone_idx;
7131 zone->zone_start_pfn = zone_start_pfn;
7133 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7134 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7136 (unsigned long)zone_idx(zone),
7137 zone_start_pfn, (zone_start_pfn + size));
7139 zone_init_free_lists(zone);
7140 zone->initialized = 1;
7144 * get_pfn_range_for_nid - Return the start and end page frames for a node
7145 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7146 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7147 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7149 * It returns the start and end page frame of a node based on information
7150 * provided by memblock_set_node(). If called for a node
7151 * with no available memory, a warning is printed and the start and end
7154 void __init get_pfn_range_for_nid(unsigned int nid,
7155 unsigned long *start_pfn, unsigned long *end_pfn)
7157 unsigned long this_start_pfn, this_end_pfn;
7163 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7164 *start_pfn = min(*start_pfn, this_start_pfn);
7165 *end_pfn = max(*end_pfn, this_end_pfn);
7168 if (*start_pfn == -1UL)
7173 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7174 * assumption is made that zones within a node are ordered in monotonic
7175 * increasing memory addresses so that the "highest" populated zone is used
7177 static void __init find_usable_zone_for_movable(void)
7180 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7181 if (zone_index == ZONE_MOVABLE)
7184 if (arch_zone_highest_possible_pfn[zone_index] >
7185 arch_zone_lowest_possible_pfn[zone_index])
7189 VM_BUG_ON(zone_index == -1);
7190 movable_zone = zone_index;
7194 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7195 * because it is sized independent of architecture. Unlike the other zones,
7196 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7197 * in each node depending on the size of each node and how evenly kernelcore
7198 * is distributed. This helper function adjusts the zone ranges
7199 * provided by the architecture for a given node by using the end of the
7200 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7201 * zones within a node are in order of monotonic increases memory addresses
7203 static void __init adjust_zone_range_for_zone_movable(int nid,
7204 unsigned long zone_type,
7205 unsigned long node_start_pfn,
7206 unsigned long node_end_pfn,
7207 unsigned long *zone_start_pfn,
7208 unsigned long *zone_end_pfn)
7210 /* Only adjust if ZONE_MOVABLE is on this node */
7211 if (zone_movable_pfn[nid]) {
7212 /* Size ZONE_MOVABLE */
7213 if (zone_type == ZONE_MOVABLE) {
7214 *zone_start_pfn = zone_movable_pfn[nid];
7215 *zone_end_pfn = min(node_end_pfn,
7216 arch_zone_highest_possible_pfn[movable_zone]);
7218 /* Adjust for ZONE_MOVABLE starting within this range */
7219 } else if (!mirrored_kernelcore &&
7220 *zone_start_pfn < zone_movable_pfn[nid] &&
7221 *zone_end_pfn > zone_movable_pfn[nid]) {
7222 *zone_end_pfn = zone_movable_pfn[nid];
7224 /* Check if this whole range is within ZONE_MOVABLE */
7225 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7226 *zone_start_pfn = *zone_end_pfn;
7231 * Return the number of pages a zone spans in a node, including holes
7232 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7234 static unsigned long __init zone_spanned_pages_in_node(int nid,
7235 unsigned long zone_type,
7236 unsigned long node_start_pfn,
7237 unsigned long node_end_pfn,
7238 unsigned long *zone_start_pfn,
7239 unsigned long *zone_end_pfn)
7241 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7242 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7243 /* When hotadd a new node from cpu_up(), the node should be empty */
7244 if (!node_start_pfn && !node_end_pfn)
7247 /* Get the start and end of the zone */
7248 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7249 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7250 adjust_zone_range_for_zone_movable(nid, zone_type,
7251 node_start_pfn, node_end_pfn,
7252 zone_start_pfn, zone_end_pfn);
7254 /* Check that this node has pages within the zone's required range */
7255 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7258 /* Move the zone boundaries inside the node if necessary */
7259 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7260 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7262 /* Return the spanned pages */
7263 return *zone_end_pfn - *zone_start_pfn;
7267 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7268 * then all holes in the requested range will be accounted for.
7270 unsigned long __init __absent_pages_in_range(int nid,
7271 unsigned long range_start_pfn,
7272 unsigned long range_end_pfn)
7274 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7275 unsigned long start_pfn, end_pfn;
7278 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7279 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7280 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7281 nr_absent -= end_pfn - start_pfn;
7287 * absent_pages_in_range - Return number of page frames in holes within a range
7288 * @start_pfn: The start PFN to start searching for holes
7289 * @end_pfn: The end PFN to stop searching for holes
7291 * Return: the number of pages frames in memory holes within a range.
7293 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7294 unsigned long end_pfn)
7296 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7299 /* Return the number of page frames in holes in a zone on a node */
7300 static unsigned long __init zone_absent_pages_in_node(int nid,
7301 unsigned long zone_type,
7302 unsigned long node_start_pfn,
7303 unsigned long node_end_pfn)
7305 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7306 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7307 unsigned long zone_start_pfn, zone_end_pfn;
7308 unsigned long nr_absent;
7310 /* When hotadd a new node from cpu_up(), the node should be empty */
7311 if (!node_start_pfn && !node_end_pfn)
7314 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7315 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7317 adjust_zone_range_for_zone_movable(nid, zone_type,
7318 node_start_pfn, node_end_pfn,
7319 &zone_start_pfn, &zone_end_pfn);
7320 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7323 * ZONE_MOVABLE handling.
7324 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7327 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7328 unsigned long start_pfn, end_pfn;
7329 struct memblock_region *r;
7331 for_each_mem_region(r) {
7332 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7333 zone_start_pfn, zone_end_pfn);
7334 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7335 zone_start_pfn, zone_end_pfn);
7337 if (zone_type == ZONE_MOVABLE &&
7338 memblock_is_mirror(r))
7339 nr_absent += end_pfn - start_pfn;
7341 if (zone_type == ZONE_NORMAL &&
7342 !memblock_is_mirror(r))
7343 nr_absent += end_pfn - start_pfn;
7350 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7351 unsigned long node_start_pfn,
7352 unsigned long node_end_pfn)
7354 unsigned long realtotalpages = 0, totalpages = 0;
7357 for (i = 0; i < MAX_NR_ZONES; i++) {
7358 struct zone *zone = pgdat->node_zones + i;
7359 unsigned long zone_start_pfn, zone_end_pfn;
7360 unsigned long spanned, absent;
7361 unsigned long size, real_size;
7363 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7368 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7373 real_size = size - absent;
7376 zone->zone_start_pfn = zone_start_pfn;
7378 zone->zone_start_pfn = 0;
7379 zone->spanned_pages = size;
7380 zone->present_pages = real_size;
7381 #if defined(CONFIG_MEMORY_HOTPLUG)
7382 zone->present_early_pages = real_size;
7386 realtotalpages += real_size;
7389 pgdat->node_spanned_pages = totalpages;
7390 pgdat->node_present_pages = realtotalpages;
7391 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7394 #ifndef CONFIG_SPARSEMEM
7396 * Calculate the size of the zone->blockflags rounded to an unsigned long
7397 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7398 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7399 * round what is now in bits to nearest long in bits, then return it in
7402 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7404 unsigned long usemapsize;
7406 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7407 usemapsize = roundup(zonesize, pageblock_nr_pages);
7408 usemapsize = usemapsize >> pageblock_order;
7409 usemapsize *= NR_PAGEBLOCK_BITS;
7410 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7412 return usemapsize / 8;
7415 static void __ref setup_usemap(struct zone *zone)
7417 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7418 zone->spanned_pages);
7419 zone->pageblock_flags = NULL;
7421 zone->pageblock_flags =
7422 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7424 if (!zone->pageblock_flags)
7425 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7426 usemapsize, zone->name, zone_to_nid(zone));
7430 static inline void setup_usemap(struct zone *zone) {}
7431 #endif /* CONFIG_SPARSEMEM */
7433 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7435 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7436 void __init set_pageblock_order(void)
7438 unsigned int order = MAX_ORDER - 1;
7440 /* Check that pageblock_nr_pages has not already been setup */
7441 if (pageblock_order)
7444 /* Don't let pageblocks exceed the maximum allocation granularity. */
7445 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7446 order = HUGETLB_PAGE_ORDER;
7449 * Assume the largest contiguous order of interest is a huge page.
7450 * This value may be variable depending on boot parameters on IA64 and
7453 pageblock_order = order;
7455 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7458 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7459 * is unused as pageblock_order is set at compile-time. See
7460 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7463 void __init set_pageblock_order(void)
7467 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7469 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7470 unsigned long present_pages)
7472 unsigned long pages = spanned_pages;
7475 * Provide a more accurate estimation if there are holes within
7476 * the zone and SPARSEMEM is in use. If there are holes within the
7477 * zone, each populated memory region may cost us one or two extra
7478 * memmap pages due to alignment because memmap pages for each
7479 * populated regions may not be naturally aligned on page boundary.
7480 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7482 if (spanned_pages > present_pages + (present_pages >> 4) &&
7483 IS_ENABLED(CONFIG_SPARSEMEM))
7484 pages = present_pages;
7486 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7489 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7490 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7492 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7494 spin_lock_init(&ds_queue->split_queue_lock);
7495 INIT_LIST_HEAD(&ds_queue->split_queue);
7496 ds_queue->split_queue_len = 0;
7499 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7502 #ifdef CONFIG_COMPACTION
7503 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7505 init_waitqueue_head(&pgdat->kcompactd_wait);
7508 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7511 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7515 pgdat_resize_init(pgdat);
7517 pgdat_init_split_queue(pgdat);
7518 pgdat_init_kcompactd(pgdat);
7520 init_waitqueue_head(&pgdat->kswapd_wait);
7521 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7523 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7524 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7526 pgdat_page_ext_init(pgdat);
7527 lruvec_init(&pgdat->__lruvec);
7530 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7531 unsigned long remaining_pages)
7533 atomic_long_set(&zone->managed_pages, remaining_pages);
7534 zone_set_nid(zone, nid);
7535 zone->name = zone_names[idx];
7536 zone->zone_pgdat = NODE_DATA(nid);
7537 spin_lock_init(&zone->lock);
7538 zone_seqlock_init(zone);
7539 zone_pcp_init(zone);
7543 * Set up the zone data structures
7544 * - init pgdat internals
7545 * - init all zones belonging to this node
7547 * NOTE: this function is only called during memory hotplug
7549 #ifdef CONFIG_MEMORY_HOTPLUG
7550 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7552 int nid = pgdat->node_id;
7556 pgdat_init_internals(pgdat);
7558 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7559 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7562 * Reset the nr_zones, order and highest_zoneidx before reuse.
7563 * Note that kswapd will init kswapd_highest_zoneidx properly
7564 * when it starts in the near future.
7566 pgdat->nr_zones = 0;
7567 pgdat->kswapd_order = 0;
7568 pgdat->kswapd_highest_zoneidx = 0;
7569 pgdat->node_start_pfn = 0;
7570 for_each_online_cpu(cpu) {
7571 struct per_cpu_nodestat *p;
7573 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7574 memset(p, 0, sizeof(*p));
7577 for (z = 0; z < MAX_NR_ZONES; z++)
7578 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7583 * Set up the zone data structures:
7584 * - mark all pages reserved
7585 * - mark all memory queues empty
7586 * - clear the memory bitmaps
7588 * NOTE: pgdat should get zeroed by caller.
7589 * NOTE: this function is only called during early init.
7591 static void __init free_area_init_core(struct pglist_data *pgdat)
7594 int nid = pgdat->node_id;
7596 pgdat_init_internals(pgdat);
7597 pgdat->per_cpu_nodestats = &boot_nodestats;
7599 for (j = 0; j < MAX_NR_ZONES; j++) {
7600 struct zone *zone = pgdat->node_zones + j;
7601 unsigned long size, freesize, memmap_pages;
7603 size = zone->spanned_pages;
7604 freesize = zone->present_pages;
7607 * Adjust freesize so that it accounts for how much memory
7608 * is used by this zone for memmap. This affects the watermark
7609 * and per-cpu initialisations
7611 memmap_pages = calc_memmap_size(size, freesize);
7612 if (!is_highmem_idx(j)) {
7613 if (freesize >= memmap_pages) {
7614 freesize -= memmap_pages;
7616 pr_debug(" %s zone: %lu pages used for memmap\n",
7617 zone_names[j], memmap_pages);
7619 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7620 zone_names[j], memmap_pages, freesize);
7623 /* Account for reserved pages */
7624 if (j == 0 && freesize > dma_reserve) {
7625 freesize -= dma_reserve;
7626 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7629 if (!is_highmem_idx(j))
7630 nr_kernel_pages += freesize;
7631 /* Charge for highmem memmap if there are enough kernel pages */
7632 else if (nr_kernel_pages > memmap_pages * 2)
7633 nr_kernel_pages -= memmap_pages;
7634 nr_all_pages += freesize;
7637 * Set an approximate value for lowmem here, it will be adjusted
7638 * when the bootmem allocator frees pages into the buddy system.
7639 * And all highmem pages will be managed by the buddy system.
7641 zone_init_internals(zone, j, nid, freesize);
7646 set_pageblock_order();
7648 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7652 #ifdef CONFIG_FLATMEM
7653 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7655 unsigned long __maybe_unused start = 0;
7656 unsigned long __maybe_unused offset = 0;
7658 /* Skip empty nodes */
7659 if (!pgdat->node_spanned_pages)
7662 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7663 offset = pgdat->node_start_pfn - start;
7664 /* ia64 gets its own node_mem_map, before this, without bootmem */
7665 if (!pgdat->node_mem_map) {
7666 unsigned long size, end;
7670 * The zone's endpoints aren't required to be MAX_ORDER
7671 * aligned but the node_mem_map endpoints must be in order
7672 * for the buddy allocator to function correctly.
7674 end = pgdat_end_pfn(pgdat);
7675 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7676 size = (end - start) * sizeof(struct page);
7677 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7678 pgdat->node_id, false);
7680 panic("Failed to allocate %ld bytes for node %d memory map\n",
7681 size, pgdat->node_id);
7682 pgdat->node_mem_map = map + offset;
7684 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7685 __func__, pgdat->node_id, (unsigned long)pgdat,
7686 (unsigned long)pgdat->node_mem_map);
7689 * With no DISCONTIG, the global mem_map is just set as node 0's
7691 if (pgdat == NODE_DATA(0)) {
7692 mem_map = NODE_DATA(0)->node_mem_map;
7693 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7699 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7700 #endif /* CONFIG_FLATMEM */
7702 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7703 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7705 pgdat->first_deferred_pfn = ULONG_MAX;
7708 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7711 static void __init free_area_init_node(int nid)
7713 pg_data_t *pgdat = NODE_DATA(nid);
7714 unsigned long start_pfn = 0;
7715 unsigned long end_pfn = 0;
7717 /* pg_data_t should be reset to zero when it's allocated */
7718 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7720 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7722 pgdat->node_id = nid;
7723 pgdat->node_start_pfn = start_pfn;
7724 pgdat->per_cpu_nodestats = NULL;
7726 if (start_pfn != end_pfn) {
7727 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7728 (u64)start_pfn << PAGE_SHIFT,
7729 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7731 pr_info("Initmem setup node %d as memoryless\n", nid);
7734 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7736 alloc_node_mem_map(pgdat);
7737 pgdat_set_deferred_range(pgdat);
7739 free_area_init_core(pgdat);
7742 static void __init free_area_init_memoryless_node(int nid)
7744 free_area_init_node(nid);
7747 #if MAX_NUMNODES > 1
7749 * Figure out the number of possible node ids.
7751 void __init setup_nr_node_ids(void)
7753 unsigned int highest;
7755 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7756 nr_node_ids = highest + 1;
7761 * node_map_pfn_alignment - determine the maximum internode alignment
7763 * This function should be called after node map is populated and sorted.
7764 * It calculates the maximum power of two alignment which can distinguish
7767 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7768 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7769 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7770 * shifted, 1GiB is enough and this function will indicate so.
7772 * This is used to test whether pfn -> nid mapping of the chosen memory
7773 * model has fine enough granularity to avoid incorrect mapping for the
7774 * populated node map.
7776 * Return: the determined alignment in pfn's. 0 if there is no alignment
7777 * requirement (single node).
7779 unsigned long __init node_map_pfn_alignment(void)
7781 unsigned long accl_mask = 0, last_end = 0;
7782 unsigned long start, end, mask;
7783 int last_nid = NUMA_NO_NODE;
7786 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7787 if (!start || last_nid < 0 || last_nid == nid) {
7794 * Start with a mask granular enough to pin-point to the
7795 * start pfn and tick off bits one-by-one until it becomes
7796 * too coarse to separate the current node from the last.
7798 mask = ~((1 << __ffs(start)) - 1);
7799 while (mask && last_end <= (start & (mask << 1)))
7802 /* accumulate all internode masks */
7806 /* convert mask to number of pages */
7807 return ~accl_mask + 1;
7811 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7813 * Return: the minimum PFN based on information provided via
7814 * memblock_set_node().
7816 unsigned long __init find_min_pfn_with_active_regions(void)
7818 return PHYS_PFN(memblock_start_of_DRAM());
7822 * early_calculate_totalpages()
7823 * Sum pages in active regions for movable zone.
7824 * Populate N_MEMORY for calculating usable_nodes.
7826 static unsigned long __init early_calculate_totalpages(void)
7828 unsigned long totalpages = 0;
7829 unsigned long start_pfn, end_pfn;
7832 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7833 unsigned long pages = end_pfn - start_pfn;
7835 totalpages += pages;
7837 node_set_state(nid, N_MEMORY);
7843 * Find the PFN the Movable zone begins in each node. Kernel memory
7844 * is spread evenly between nodes as long as the nodes have enough
7845 * memory. When they don't, some nodes will have more kernelcore than
7848 static void __init find_zone_movable_pfns_for_nodes(void)
7851 unsigned long usable_startpfn;
7852 unsigned long kernelcore_node, kernelcore_remaining;
7853 /* save the state before borrow the nodemask */
7854 nodemask_t saved_node_state = node_states[N_MEMORY];
7855 unsigned long totalpages = early_calculate_totalpages();
7856 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7857 struct memblock_region *r;
7859 /* Need to find movable_zone earlier when movable_node is specified. */
7860 find_usable_zone_for_movable();
7863 * If movable_node is specified, ignore kernelcore and movablecore
7866 if (movable_node_is_enabled()) {
7867 for_each_mem_region(r) {
7868 if (!memblock_is_hotpluggable(r))
7871 nid = memblock_get_region_node(r);
7873 usable_startpfn = PFN_DOWN(r->base);
7874 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7875 min(usable_startpfn, zone_movable_pfn[nid]) :
7883 * If kernelcore=mirror is specified, ignore movablecore option
7885 if (mirrored_kernelcore) {
7886 bool mem_below_4gb_not_mirrored = false;
7888 for_each_mem_region(r) {
7889 if (memblock_is_mirror(r))
7892 nid = memblock_get_region_node(r);
7894 usable_startpfn = memblock_region_memory_base_pfn(r);
7896 if (usable_startpfn < PHYS_PFN(SZ_4G)) {
7897 mem_below_4gb_not_mirrored = true;
7901 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7902 min(usable_startpfn, zone_movable_pfn[nid]) :
7906 if (mem_below_4gb_not_mirrored)
7907 pr_warn("This configuration results in unmirrored kernel memory.\n");
7913 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7914 * amount of necessary memory.
7916 if (required_kernelcore_percent)
7917 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7919 if (required_movablecore_percent)
7920 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7924 * If movablecore= was specified, calculate what size of
7925 * kernelcore that corresponds so that memory usable for
7926 * any allocation type is evenly spread. If both kernelcore
7927 * and movablecore are specified, then the value of kernelcore
7928 * will be used for required_kernelcore if it's greater than
7929 * what movablecore would have allowed.
7931 if (required_movablecore) {
7932 unsigned long corepages;
7935 * Round-up so that ZONE_MOVABLE is at least as large as what
7936 * was requested by the user
7938 required_movablecore =
7939 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7940 required_movablecore = min(totalpages, required_movablecore);
7941 corepages = totalpages - required_movablecore;
7943 required_kernelcore = max(required_kernelcore, corepages);
7947 * If kernelcore was not specified or kernelcore size is larger
7948 * than totalpages, there is no ZONE_MOVABLE.
7950 if (!required_kernelcore || required_kernelcore >= totalpages)
7953 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7954 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7957 /* Spread kernelcore memory as evenly as possible throughout nodes */
7958 kernelcore_node = required_kernelcore / usable_nodes;
7959 for_each_node_state(nid, N_MEMORY) {
7960 unsigned long start_pfn, end_pfn;
7963 * Recalculate kernelcore_node if the division per node
7964 * now exceeds what is necessary to satisfy the requested
7965 * amount of memory for the kernel
7967 if (required_kernelcore < kernelcore_node)
7968 kernelcore_node = required_kernelcore / usable_nodes;
7971 * As the map is walked, we track how much memory is usable
7972 * by the kernel using kernelcore_remaining. When it is
7973 * 0, the rest of the node is usable by ZONE_MOVABLE
7975 kernelcore_remaining = kernelcore_node;
7977 /* Go through each range of PFNs within this node */
7978 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7979 unsigned long size_pages;
7981 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7982 if (start_pfn >= end_pfn)
7985 /* Account for what is only usable for kernelcore */
7986 if (start_pfn < usable_startpfn) {
7987 unsigned long kernel_pages;
7988 kernel_pages = min(end_pfn, usable_startpfn)
7991 kernelcore_remaining -= min(kernel_pages,
7992 kernelcore_remaining);
7993 required_kernelcore -= min(kernel_pages,
7994 required_kernelcore);
7996 /* Continue if range is now fully accounted */
7997 if (end_pfn <= usable_startpfn) {
8000 * Push zone_movable_pfn to the end so
8001 * that if we have to rebalance
8002 * kernelcore across nodes, we will
8003 * not double account here
8005 zone_movable_pfn[nid] = end_pfn;
8008 start_pfn = usable_startpfn;
8012 * The usable PFN range for ZONE_MOVABLE is from
8013 * start_pfn->end_pfn. Calculate size_pages as the
8014 * number of pages used as kernelcore
8016 size_pages = end_pfn - start_pfn;
8017 if (size_pages > kernelcore_remaining)
8018 size_pages = kernelcore_remaining;
8019 zone_movable_pfn[nid] = start_pfn + size_pages;
8022 * Some kernelcore has been met, update counts and
8023 * break if the kernelcore for this node has been
8026 required_kernelcore -= min(required_kernelcore,
8028 kernelcore_remaining -= size_pages;
8029 if (!kernelcore_remaining)
8035 * If there is still required_kernelcore, we do another pass with one
8036 * less node in the count. This will push zone_movable_pfn[nid] further
8037 * along on the nodes that still have memory until kernelcore is
8041 if (usable_nodes && required_kernelcore > usable_nodes)
8045 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
8046 for (nid = 0; nid < MAX_NUMNODES; nid++) {
8047 unsigned long start_pfn, end_pfn;
8049 zone_movable_pfn[nid] =
8050 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
8052 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
8053 if (zone_movable_pfn[nid] >= end_pfn)
8054 zone_movable_pfn[nid] = 0;
8058 /* restore the node_state */
8059 node_states[N_MEMORY] = saved_node_state;
8062 /* Any regular or high memory on that node ? */
8063 static void check_for_memory(pg_data_t *pgdat, int nid)
8065 enum zone_type zone_type;
8067 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8068 struct zone *zone = &pgdat->node_zones[zone_type];
8069 if (populated_zone(zone)) {
8070 if (IS_ENABLED(CONFIG_HIGHMEM))
8071 node_set_state(nid, N_HIGH_MEMORY);
8072 if (zone_type <= ZONE_NORMAL)
8073 node_set_state(nid, N_NORMAL_MEMORY);
8080 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8081 * such cases we allow max_zone_pfn sorted in the descending order
8083 bool __weak arch_has_descending_max_zone_pfns(void)
8089 * free_area_init - Initialise all pg_data_t and zone data
8090 * @max_zone_pfn: an array of max PFNs for each zone
8092 * This will call free_area_init_node() for each active node in the system.
8093 * Using the page ranges provided by memblock_set_node(), the size of each
8094 * zone in each node and their holes is calculated. If the maximum PFN
8095 * between two adjacent zones match, it is assumed that the zone is empty.
8096 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8097 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8098 * starts where the previous one ended. For example, ZONE_DMA32 starts
8099 * at arch_max_dma_pfn.
8101 void __init free_area_init(unsigned long *max_zone_pfn)
8103 unsigned long start_pfn, end_pfn;
8107 /* Record where the zone boundaries are */
8108 memset(arch_zone_lowest_possible_pfn, 0,
8109 sizeof(arch_zone_lowest_possible_pfn));
8110 memset(arch_zone_highest_possible_pfn, 0,
8111 sizeof(arch_zone_highest_possible_pfn));
8113 start_pfn = find_min_pfn_with_active_regions();
8114 descending = arch_has_descending_max_zone_pfns();
8116 for (i = 0; i < MAX_NR_ZONES; i++) {
8118 zone = MAX_NR_ZONES - i - 1;
8122 if (zone == ZONE_MOVABLE)
8125 end_pfn = max(max_zone_pfn[zone], start_pfn);
8126 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8127 arch_zone_highest_possible_pfn[zone] = end_pfn;
8129 start_pfn = end_pfn;
8132 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8133 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8134 find_zone_movable_pfns_for_nodes();
8136 /* Print out the zone ranges */
8137 pr_info("Zone ranges:\n");
8138 for (i = 0; i < MAX_NR_ZONES; i++) {
8139 if (i == ZONE_MOVABLE)
8141 pr_info(" %-8s ", zone_names[i]);
8142 if (arch_zone_lowest_possible_pfn[i] ==
8143 arch_zone_highest_possible_pfn[i])
8146 pr_cont("[mem %#018Lx-%#018Lx]\n",
8147 (u64)arch_zone_lowest_possible_pfn[i]
8149 ((u64)arch_zone_highest_possible_pfn[i]
8150 << PAGE_SHIFT) - 1);
8153 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8154 pr_info("Movable zone start for each node\n");
8155 for (i = 0; i < MAX_NUMNODES; i++) {
8156 if (zone_movable_pfn[i])
8157 pr_info(" Node %d: %#018Lx\n", i,
8158 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8162 * Print out the early node map, and initialize the
8163 * subsection-map relative to active online memory ranges to
8164 * enable future "sub-section" extensions of the memory map.
8166 pr_info("Early memory node ranges\n");
8167 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8168 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8169 (u64)start_pfn << PAGE_SHIFT,
8170 ((u64)end_pfn << PAGE_SHIFT) - 1);
8171 subsection_map_init(start_pfn, end_pfn - start_pfn);
8174 /* Initialise every node */
8175 mminit_verify_pageflags_layout();
8176 setup_nr_node_ids();
8177 for_each_node(nid) {
8180 if (!node_online(nid)) {
8181 pr_info("Initializing node %d as memoryless\n", nid);
8183 /* Allocator not initialized yet */
8184 pgdat = arch_alloc_nodedata(nid);
8186 pr_err("Cannot allocate %zuB for node %d.\n",
8187 sizeof(*pgdat), nid);
8190 arch_refresh_nodedata(nid, pgdat);
8191 free_area_init_memoryless_node(nid);
8194 * We do not want to confuse userspace by sysfs
8195 * files/directories for node without any memory
8196 * attached to it, so this node is not marked as
8197 * N_MEMORY and not marked online so that no sysfs
8198 * hierarchy will be created via register_one_node for
8199 * it. The pgdat will get fully initialized by
8200 * hotadd_init_pgdat() when memory is hotplugged into
8206 pgdat = NODE_DATA(nid);
8207 free_area_init_node(nid);
8209 /* Any memory on that node */
8210 if (pgdat->node_present_pages)
8211 node_set_state(nid, N_MEMORY);
8212 check_for_memory(pgdat, nid);
8218 static int __init cmdline_parse_core(char *p, unsigned long *core,
8219 unsigned long *percent)
8221 unsigned long long coremem;
8227 /* Value may be a percentage of total memory, otherwise bytes */
8228 coremem = simple_strtoull(p, &endptr, 0);
8229 if (*endptr == '%') {
8230 /* Paranoid check for percent values greater than 100 */
8231 WARN_ON(coremem > 100);
8235 coremem = memparse(p, &p);
8236 /* Paranoid check that UL is enough for the coremem value */
8237 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8239 *core = coremem >> PAGE_SHIFT;
8246 * kernelcore=size sets the amount of memory for use for allocations that
8247 * cannot be reclaimed or migrated.
8249 static int __init cmdline_parse_kernelcore(char *p)
8251 /* parse kernelcore=mirror */
8252 if (parse_option_str(p, "mirror")) {
8253 mirrored_kernelcore = true;
8257 return cmdline_parse_core(p, &required_kernelcore,
8258 &required_kernelcore_percent);
8262 * movablecore=size sets the amount of memory for use for allocations that
8263 * can be reclaimed or migrated.
8265 static int __init cmdline_parse_movablecore(char *p)
8267 return cmdline_parse_core(p, &required_movablecore,
8268 &required_movablecore_percent);
8271 early_param("kernelcore", cmdline_parse_kernelcore);
8272 early_param("movablecore", cmdline_parse_movablecore);
8274 void adjust_managed_page_count(struct page *page, long count)
8276 atomic_long_add(count, &page_zone(page)->managed_pages);
8277 totalram_pages_add(count);
8278 #ifdef CONFIG_HIGHMEM
8279 if (PageHighMem(page))
8280 totalhigh_pages_add(count);
8283 EXPORT_SYMBOL(adjust_managed_page_count);
8285 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8288 unsigned long pages = 0;
8290 start = (void *)PAGE_ALIGN((unsigned long)start);
8291 end = (void *)((unsigned long)end & PAGE_MASK);
8292 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8293 struct page *page = virt_to_page(pos);
8294 void *direct_map_addr;
8297 * 'direct_map_addr' might be different from 'pos'
8298 * because some architectures' virt_to_page()
8299 * work with aliases. Getting the direct map
8300 * address ensures that we get a _writeable_
8301 * alias for the memset().
8303 direct_map_addr = page_address(page);
8305 * Perform a kasan-unchecked memset() since this memory
8306 * has not been initialized.
8308 direct_map_addr = kasan_reset_tag(direct_map_addr);
8309 if ((unsigned int)poison <= 0xFF)
8310 memset(direct_map_addr, poison, PAGE_SIZE);
8312 free_reserved_page(page);
8316 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8321 void __init mem_init_print_info(void)
8323 unsigned long physpages, codesize, datasize, rosize, bss_size;
8324 unsigned long init_code_size, init_data_size;
8326 physpages = get_num_physpages();
8327 codesize = _etext - _stext;
8328 datasize = _edata - _sdata;
8329 rosize = __end_rodata - __start_rodata;
8330 bss_size = __bss_stop - __bss_start;
8331 init_data_size = __init_end - __init_begin;
8332 init_code_size = _einittext - _sinittext;
8335 * Detect special cases and adjust section sizes accordingly:
8336 * 1) .init.* may be embedded into .data sections
8337 * 2) .init.text.* may be out of [__init_begin, __init_end],
8338 * please refer to arch/tile/kernel/vmlinux.lds.S.
8339 * 3) .rodata.* may be embedded into .text or .data sections.
8341 #define adj_init_size(start, end, size, pos, adj) \
8343 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8347 adj_init_size(__init_begin, __init_end, init_data_size,
8348 _sinittext, init_code_size);
8349 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8350 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8351 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8352 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8354 #undef adj_init_size
8356 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8357 #ifdef CONFIG_HIGHMEM
8361 K(nr_free_pages()), K(physpages),
8362 codesize >> 10, datasize >> 10, rosize >> 10,
8363 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8364 K(physpages - totalram_pages() - totalcma_pages),
8366 #ifdef CONFIG_HIGHMEM
8367 , K(totalhigh_pages())
8373 * set_dma_reserve - set the specified number of pages reserved in the first zone
8374 * @new_dma_reserve: The number of pages to mark reserved
8376 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8377 * In the DMA zone, a significant percentage may be consumed by kernel image
8378 * and other unfreeable allocations which can skew the watermarks badly. This
8379 * function may optionally be used to account for unfreeable pages in the
8380 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8381 * smaller per-cpu batchsize.
8383 void __init set_dma_reserve(unsigned long new_dma_reserve)
8385 dma_reserve = new_dma_reserve;
8388 static int page_alloc_cpu_dead(unsigned int cpu)
8392 lru_add_drain_cpu(cpu);
8393 mlock_page_drain_remote(cpu);
8397 * Spill the event counters of the dead processor
8398 * into the current processors event counters.
8399 * This artificially elevates the count of the current
8402 vm_events_fold_cpu(cpu);
8405 * Zero the differential counters of the dead processor
8406 * so that the vm statistics are consistent.
8408 * This is only okay since the processor is dead and cannot
8409 * race with what we are doing.
8411 cpu_vm_stats_fold(cpu);
8413 for_each_populated_zone(zone)
8414 zone_pcp_update(zone, 0);
8419 static int page_alloc_cpu_online(unsigned int cpu)
8423 for_each_populated_zone(zone)
8424 zone_pcp_update(zone, 1);
8429 int hashdist = HASHDIST_DEFAULT;
8431 static int __init set_hashdist(char *str)
8435 hashdist = simple_strtoul(str, &str, 0);
8438 __setup("hashdist=", set_hashdist);
8441 void __init page_alloc_init(void)
8446 if (num_node_state(N_MEMORY) == 1)
8450 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8451 "mm/page_alloc:pcp",
8452 page_alloc_cpu_online,
8453 page_alloc_cpu_dead);
8458 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8459 * or min_free_kbytes changes.
8461 static void calculate_totalreserve_pages(void)
8463 struct pglist_data *pgdat;
8464 unsigned long reserve_pages = 0;
8465 enum zone_type i, j;
8467 for_each_online_pgdat(pgdat) {
8469 pgdat->totalreserve_pages = 0;
8471 for (i = 0; i < MAX_NR_ZONES; i++) {
8472 struct zone *zone = pgdat->node_zones + i;
8474 unsigned long managed_pages = zone_managed_pages(zone);
8476 /* Find valid and maximum lowmem_reserve in the zone */
8477 for (j = i; j < MAX_NR_ZONES; j++) {
8478 if (zone->lowmem_reserve[j] > max)
8479 max = zone->lowmem_reserve[j];
8482 /* we treat the high watermark as reserved pages. */
8483 max += high_wmark_pages(zone);
8485 if (max > managed_pages)
8486 max = managed_pages;
8488 pgdat->totalreserve_pages += max;
8490 reserve_pages += max;
8493 totalreserve_pages = reserve_pages;
8497 * setup_per_zone_lowmem_reserve - called whenever
8498 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8499 * has a correct pages reserved value, so an adequate number of
8500 * pages are left in the zone after a successful __alloc_pages().
8502 static void setup_per_zone_lowmem_reserve(void)
8504 struct pglist_data *pgdat;
8505 enum zone_type i, j;
8507 for_each_online_pgdat(pgdat) {
8508 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8509 struct zone *zone = &pgdat->node_zones[i];
8510 int ratio = sysctl_lowmem_reserve_ratio[i];
8511 bool clear = !ratio || !zone_managed_pages(zone);
8512 unsigned long managed_pages = 0;
8514 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8515 struct zone *upper_zone = &pgdat->node_zones[j];
8517 managed_pages += zone_managed_pages(upper_zone);
8520 zone->lowmem_reserve[j] = 0;
8522 zone->lowmem_reserve[j] = managed_pages / ratio;
8527 /* update totalreserve_pages */
8528 calculate_totalreserve_pages();
8531 static void __setup_per_zone_wmarks(void)
8533 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8534 unsigned long lowmem_pages = 0;
8536 unsigned long flags;
8538 /* Calculate total number of !ZONE_HIGHMEM pages */
8539 for_each_zone(zone) {
8540 if (!is_highmem(zone))
8541 lowmem_pages += zone_managed_pages(zone);
8544 for_each_zone(zone) {
8547 spin_lock_irqsave(&zone->lock, flags);
8548 tmp = (u64)pages_min * zone_managed_pages(zone);
8549 do_div(tmp, lowmem_pages);
8550 if (is_highmem(zone)) {
8552 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8553 * need highmem pages, so cap pages_min to a small
8556 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8557 * deltas control async page reclaim, and so should
8558 * not be capped for highmem.
8560 unsigned long min_pages;
8562 min_pages = zone_managed_pages(zone) / 1024;
8563 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8564 zone->_watermark[WMARK_MIN] = min_pages;
8567 * If it's a lowmem zone, reserve a number of pages
8568 * proportionate to the zone's size.
8570 zone->_watermark[WMARK_MIN] = tmp;
8574 * Set the kswapd watermarks distance according to the
8575 * scale factor in proportion to available memory, but
8576 * ensure a minimum size on small systems.
8578 tmp = max_t(u64, tmp >> 2,
8579 mult_frac(zone_managed_pages(zone),
8580 watermark_scale_factor, 10000));
8582 zone->watermark_boost = 0;
8583 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8584 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8585 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8587 spin_unlock_irqrestore(&zone->lock, flags);
8590 /* update totalreserve_pages */
8591 calculate_totalreserve_pages();
8595 * setup_per_zone_wmarks - called when min_free_kbytes changes
8596 * or when memory is hot-{added|removed}
8598 * Ensures that the watermark[min,low,high] values for each zone are set
8599 * correctly with respect to min_free_kbytes.
8601 void setup_per_zone_wmarks(void)
8604 static DEFINE_SPINLOCK(lock);
8607 __setup_per_zone_wmarks();
8611 * The watermark size have changed so update the pcpu batch
8612 * and high limits or the limits may be inappropriate.
8615 zone_pcp_update(zone, 0);
8619 * Initialise min_free_kbytes.
8621 * For small machines we want it small (128k min). For large machines
8622 * we want it large (256MB max). But it is not linear, because network
8623 * bandwidth does not increase linearly with machine size. We use
8625 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8626 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8642 void calculate_min_free_kbytes(void)
8644 unsigned long lowmem_kbytes;
8645 int new_min_free_kbytes;
8647 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8648 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8650 if (new_min_free_kbytes > user_min_free_kbytes)
8651 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8653 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8654 new_min_free_kbytes, user_min_free_kbytes);
8658 int __meminit init_per_zone_wmark_min(void)
8660 calculate_min_free_kbytes();
8661 setup_per_zone_wmarks();
8662 refresh_zone_stat_thresholds();
8663 setup_per_zone_lowmem_reserve();
8666 setup_min_unmapped_ratio();
8667 setup_min_slab_ratio();
8670 khugepaged_min_free_kbytes_update();
8674 postcore_initcall(init_per_zone_wmark_min)
8677 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8678 * that we can call two helper functions whenever min_free_kbytes
8681 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8682 void *buffer, size_t *length, loff_t *ppos)
8686 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8691 user_min_free_kbytes = min_free_kbytes;
8692 setup_per_zone_wmarks();
8697 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8698 void *buffer, size_t *length, loff_t *ppos)
8702 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8707 setup_per_zone_wmarks();
8713 static void setup_min_unmapped_ratio(void)
8718 for_each_online_pgdat(pgdat)
8719 pgdat->min_unmapped_pages = 0;
8722 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8723 sysctl_min_unmapped_ratio) / 100;
8727 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8728 void *buffer, size_t *length, loff_t *ppos)
8732 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8736 setup_min_unmapped_ratio();
8741 static void setup_min_slab_ratio(void)
8746 for_each_online_pgdat(pgdat)
8747 pgdat->min_slab_pages = 0;
8750 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8751 sysctl_min_slab_ratio) / 100;
8754 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8755 void *buffer, size_t *length, loff_t *ppos)
8759 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8763 setup_min_slab_ratio();
8770 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8771 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8772 * whenever sysctl_lowmem_reserve_ratio changes.
8774 * The reserve ratio obviously has absolutely no relation with the
8775 * minimum watermarks. The lowmem reserve ratio can only make sense
8776 * if in function of the boot time zone sizes.
8778 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8779 void *buffer, size_t *length, loff_t *ppos)
8783 proc_dointvec_minmax(table, write, buffer, length, ppos);
8785 for (i = 0; i < MAX_NR_ZONES; i++) {
8786 if (sysctl_lowmem_reserve_ratio[i] < 1)
8787 sysctl_lowmem_reserve_ratio[i] = 0;
8790 setup_per_zone_lowmem_reserve();
8795 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8796 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8797 * pagelist can have before it gets flushed back to buddy allocator.
8799 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8800 int write, void *buffer, size_t *length, loff_t *ppos)
8803 int old_percpu_pagelist_high_fraction;
8806 mutex_lock(&pcp_batch_high_lock);
8807 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8809 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8810 if (!write || ret < 0)
8813 /* Sanity checking to avoid pcp imbalance */
8814 if (percpu_pagelist_high_fraction &&
8815 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8816 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8822 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8825 for_each_populated_zone(zone)
8826 zone_set_pageset_high_and_batch(zone, 0);
8828 mutex_unlock(&pcp_batch_high_lock);
8832 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8834 * Returns the number of pages that arch has reserved but
8835 * is not known to alloc_large_system_hash().
8837 static unsigned long __init arch_reserved_kernel_pages(void)
8844 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8845 * machines. As memory size is increased the scale is also increased but at
8846 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8847 * quadruples the scale is increased by one, which means the size of hash table
8848 * only doubles, instead of quadrupling as well.
8849 * Because 32-bit systems cannot have large physical memory, where this scaling
8850 * makes sense, it is disabled on such platforms.
8852 #if __BITS_PER_LONG > 32
8853 #define ADAPT_SCALE_BASE (64ul << 30)
8854 #define ADAPT_SCALE_SHIFT 2
8855 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8859 * allocate a large system hash table from bootmem
8860 * - it is assumed that the hash table must contain an exact power-of-2
8861 * quantity of entries
8862 * - limit is the number of hash buckets, not the total allocation size
8864 void *__init alloc_large_system_hash(const char *tablename,
8865 unsigned long bucketsize,
8866 unsigned long numentries,
8869 unsigned int *_hash_shift,
8870 unsigned int *_hash_mask,
8871 unsigned long low_limit,
8872 unsigned long high_limit)
8874 unsigned long long max = high_limit;
8875 unsigned long log2qty, size;
8881 /* allow the kernel cmdline to have a say */
8883 /* round applicable memory size up to nearest megabyte */
8884 numentries = nr_kernel_pages;
8885 numentries -= arch_reserved_kernel_pages();
8887 /* It isn't necessary when PAGE_SIZE >= 1MB */
8888 if (PAGE_SHIFT < 20)
8889 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8891 #if __BITS_PER_LONG > 32
8893 unsigned long adapt;
8895 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8896 adapt <<= ADAPT_SCALE_SHIFT)
8901 /* limit to 1 bucket per 2^scale bytes of low memory */
8902 if (scale > PAGE_SHIFT)
8903 numentries >>= (scale - PAGE_SHIFT);
8905 numentries <<= (PAGE_SHIFT - scale);
8907 /* Make sure we've got at least a 0-order allocation.. */
8908 if (unlikely(flags & HASH_SMALL)) {
8909 /* Makes no sense without HASH_EARLY */
8910 WARN_ON(!(flags & HASH_EARLY));
8911 if (!(numentries >> *_hash_shift)) {
8912 numentries = 1UL << *_hash_shift;
8913 BUG_ON(!numentries);
8915 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8916 numentries = PAGE_SIZE / bucketsize;
8918 numentries = roundup_pow_of_two(numentries);
8920 /* limit allocation size to 1/16 total memory by default */
8922 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8923 do_div(max, bucketsize);
8925 max = min(max, 0x80000000ULL);
8927 if (numentries < low_limit)
8928 numentries = low_limit;
8929 if (numentries > max)
8932 log2qty = ilog2(numentries);
8934 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8937 size = bucketsize << log2qty;
8938 if (flags & HASH_EARLY) {
8939 if (flags & HASH_ZERO)
8940 table = memblock_alloc(size, SMP_CACHE_BYTES);
8942 table = memblock_alloc_raw(size,
8944 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8945 table = vmalloc_huge(size, gfp_flags);
8948 huge = is_vm_area_hugepages(table);
8951 * If bucketsize is not a power-of-two, we may free
8952 * some pages at the end of hash table which
8953 * alloc_pages_exact() automatically does
8955 table = alloc_pages_exact(size, gfp_flags);
8956 kmemleak_alloc(table, size, 1, gfp_flags);
8958 } while (!table && size > PAGE_SIZE && --log2qty);
8961 panic("Failed to allocate %s hash table\n", tablename);
8963 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8964 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8965 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8968 *_hash_shift = log2qty;
8970 *_hash_mask = (1 << log2qty) - 1;
8975 #ifdef CONFIG_CONTIG_ALLOC
8976 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8977 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8978 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8979 static void alloc_contig_dump_pages(struct list_head *page_list)
8981 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8983 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8987 list_for_each_entry(page, page_list, lru)
8988 dump_page(page, "migration failure");
8992 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8997 /* [start, end) must belong to a single zone. */
8998 int __alloc_contig_migrate_range(struct compact_control *cc,
8999 unsigned long start, unsigned long end)
9001 /* This function is based on compact_zone() from compaction.c. */
9002 unsigned int nr_reclaimed;
9003 unsigned long pfn = start;
9004 unsigned int tries = 0;
9006 struct migration_target_control mtc = {
9007 .nid = zone_to_nid(cc->zone),
9008 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9011 lru_cache_disable();
9013 while (pfn < end || !list_empty(&cc->migratepages)) {
9014 if (fatal_signal_pending(current)) {
9019 if (list_empty(&cc->migratepages)) {
9020 cc->nr_migratepages = 0;
9021 ret = isolate_migratepages_range(cc, pfn, end);
9022 if (ret && ret != -EAGAIN)
9024 pfn = cc->migrate_pfn;
9026 } else if (++tries == 5) {
9031 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9033 cc->nr_migratepages -= nr_reclaimed;
9035 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9036 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9039 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9040 * to retry again over this error, so do the same here.
9048 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
9049 alloc_contig_dump_pages(&cc->migratepages);
9050 putback_movable_pages(&cc->migratepages);
9057 * alloc_contig_range() -- tries to allocate given range of pages
9058 * @start: start PFN to allocate
9059 * @end: one-past-the-last PFN to allocate
9060 * @migratetype: migratetype of the underlying pageblocks (either
9061 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9062 * in range must have the same migratetype and it must
9063 * be either of the two.
9064 * @gfp_mask: GFP mask to use during compaction
9066 * The PFN range does not have to be pageblock aligned. The PFN range must
9067 * belong to a single zone.
9069 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9070 * pageblocks in the range. Once isolated, the pageblocks should not
9071 * be modified by others.
9073 * Return: zero on success or negative error code. On success all
9074 * pages which PFN is in [start, end) are allocated for the caller and
9075 * need to be freed with free_contig_range().
9077 int alloc_contig_range(unsigned long start, unsigned long end,
9078 unsigned migratetype, gfp_t gfp_mask)
9080 unsigned long outer_start, outer_end;
9084 struct compact_control cc = {
9085 .nr_migratepages = 0,
9087 .zone = page_zone(pfn_to_page(start)),
9088 .mode = MIGRATE_SYNC,
9089 .ignore_skip_hint = true,
9090 .no_set_skip_hint = true,
9091 .gfp_mask = current_gfp_context(gfp_mask),
9092 .alloc_contig = true,
9094 INIT_LIST_HEAD(&cc.migratepages);
9097 * What we do here is we mark all pageblocks in range as
9098 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9099 * have different sizes, and due to the way page allocator
9100 * work, start_isolate_page_range() has special handlings for this.
9102 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9103 * migrate the pages from an unaligned range (ie. pages that
9104 * we are interested in). This will put all the pages in
9105 * range back to page allocator as MIGRATE_ISOLATE.
9107 * When this is done, we take the pages in range from page
9108 * allocator removing them from the buddy system. This way
9109 * page allocator will never consider using them.
9111 * This lets us mark the pageblocks back as
9112 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9113 * aligned range but not in the unaligned, original range are
9114 * put back to page allocator so that buddy can use them.
9117 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
9121 drain_all_pages(cc.zone);
9124 * In case of -EBUSY, we'd like to know which page causes problem.
9125 * So, just fall through. test_pages_isolated() has a tracepoint
9126 * which will report the busy page.
9128 * It is possible that busy pages could become available before
9129 * the call to test_pages_isolated, and the range will actually be
9130 * allocated. So, if we fall through be sure to clear ret so that
9131 * -EBUSY is not accidentally used or returned to caller.
9133 ret = __alloc_contig_migrate_range(&cc, start, end);
9134 if (ret && ret != -EBUSY)
9139 * Pages from [start, end) are within a pageblock_nr_pages
9140 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9141 * more, all pages in [start, end) are free in page allocator.
9142 * What we are going to do is to allocate all pages from
9143 * [start, end) (that is remove them from page allocator).
9145 * The only problem is that pages at the beginning and at the
9146 * end of interesting range may be not aligned with pages that
9147 * page allocator holds, ie. they can be part of higher order
9148 * pages. Because of this, we reserve the bigger range and
9149 * once this is done free the pages we are not interested in.
9151 * We don't have to hold zone->lock here because the pages are
9152 * isolated thus they won't get removed from buddy.
9156 outer_start = start;
9157 while (!PageBuddy(pfn_to_page(outer_start))) {
9158 if (++order >= MAX_ORDER) {
9159 outer_start = start;
9162 outer_start &= ~0UL << order;
9165 if (outer_start != start) {
9166 order = buddy_order(pfn_to_page(outer_start));
9169 * outer_start page could be small order buddy page and
9170 * it doesn't include start page. Adjust outer_start
9171 * in this case to report failed page properly
9172 * on tracepoint in test_pages_isolated()
9174 if (outer_start + (1UL << order) <= start)
9175 outer_start = start;
9178 /* Make sure the range is really isolated. */
9179 if (test_pages_isolated(outer_start, end, 0)) {
9184 /* Grab isolated pages from freelists. */
9185 outer_end = isolate_freepages_range(&cc, outer_start, end);
9191 /* Free head and tail (if any) */
9192 if (start != outer_start)
9193 free_contig_range(outer_start, start - outer_start);
9194 if (end != outer_end)
9195 free_contig_range(end, outer_end - end);
9198 undo_isolate_page_range(start, end, migratetype);
9201 EXPORT_SYMBOL(alloc_contig_range);
9203 static int __alloc_contig_pages(unsigned long start_pfn,
9204 unsigned long nr_pages, gfp_t gfp_mask)
9206 unsigned long end_pfn = start_pfn + nr_pages;
9208 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9212 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9213 unsigned long nr_pages)
9215 unsigned long i, end_pfn = start_pfn + nr_pages;
9218 for (i = start_pfn; i < end_pfn; i++) {
9219 page = pfn_to_online_page(i);
9223 if (page_zone(page) != z)
9226 if (PageReserved(page))
9232 static bool zone_spans_last_pfn(const struct zone *zone,
9233 unsigned long start_pfn, unsigned long nr_pages)
9235 unsigned long last_pfn = start_pfn + nr_pages - 1;
9237 return zone_spans_pfn(zone, last_pfn);
9241 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9242 * @nr_pages: Number of contiguous pages to allocate
9243 * @gfp_mask: GFP mask to limit search and used during compaction
9245 * @nodemask: Mask for other possible nodes
9247 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9248 * on an applicable zonelist to find a contiguous pfn range which can then be
9249 * tried for allocation with alloc_contig_range(). This routine is intended
9250 * for allocation requests which can not be fulfilled with the buddy allocator.
9252 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9253 * power of two, then allocated range is also guaranteed to be aligned to same
9254 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9256 * Allocated pages can be freed with free_contig_range() or by manually calling
9257 * __free_page() on each allocated page.
9259 * Return: pointer to contiguous pages on success, or NULL if not successful.
9261 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9262 int nid, nodemask_t *nodemask)
9264 unsigned long ret, pfn, flags;
9265 struct zonelist *zonelist;
9269 zonelist = node_zonelist(nid, gfp_mask);
9270 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9271 gfp_zone(gfp_mask), nodemask) {
9272 spin_lock_irqsave(&zone->lock, flags);
9274 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9275 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9276 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9278 * We release the zone lock here because
9279 * alloc_contig_range() will also lock the zone
9280 * at some point. If there's an allocation
9281 * spinning on this lock, it may win the race
9282 * and cause alloc_contig_range() to fail...
9284 spin_unlock_irqrestore(&zone->lock, flags);
9285 ret = __alloc_contig_pages(pfn, nr_pages,
9288 return pfn_to_page(pfn);
9289 spin_lock_irqsave(&zone->lock, flags);
9293 spin_unlock_irqrestore(&zone->lock, flags);
9297 #endif /* CONFIG_CONTIG_ALLOC */
9299 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9301 unsigned long count = 0;
9303 for (; nr_pages--; pfn++) {
9304 struct page *page = pfn_to_page(pfn);
9306 count += page_count(page) != 1;
9309 WARN(count != 0, "%lu pages are still in use!\n", count);
9311 EXPORT_SYMBOL(free_contig_range);
9314 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9315 * page high values need to be recalculated.
9317 void zone_pcp_update(struct zone *zone, int cpu_online)
9319 mutex_lock(&pcp_batch_high_lock);
9320 zone_set_pageset_high_and_batch(zone, cpu_online);
9321 mutex_unlock(&pcp_batch_high_lock);
9325 * Effectively disable pcplists for the zone by setting the high limit to 0
9326 * and draining all cpus. A concurrent page freeing on another CPU that's about
9327 * to put the page on pcplist will either finish before the drain and the page
9328 * will be drained, or observe the new high limit and skip the pcplist.
9330 * Must be paired with a call to zone_pcp_enable().
9332 void zone_pcp_disable(struct zone *zone)
9334 mutex_lock(&pcp_batch_high_lock);
9335 __zone_set_pageset_high_and_batch(zone, 0, 1);
9336 __drain_all_pages(zone, true);
9339 void zone_pcp_enable(struct zone *zone)
9341 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9342 mutex_unlock(&pcp_batch_high_lock);
9345 void zone_pcp_reset(struct zone *zone)
9348 struct per_cpu_zonestat *pzstats;
9350 if (zone->per_cpu_pageset != &boot_pageset) {
9351 for_each_online_cpu(cpu) {
9352 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9353 drain_zonestat(zone, pzstats);
9355 free_percpu(zone->per_cpu_pageset);
9356 free_percpu(zone->per_cpu_zonestats);
9357 zone->per_cpu_pageset = &boot_pageset;
9358 zone->per_cpu_zonestats = &boot_zonestats;
9362 #ifdef CONFIG_MEMORY_HOTREMOVE
9364 * All pages in the range must be in a single zone, must not contain holes,
9365 * must span full sections, and must be isolated before calling this function.
9367 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9369 unsigned long pfn = start_pfn;
9373 unsigned long flags;
9375 offline_mem_sections(pfn, end_pfn);
9376 zone = page_zone(pfn_to_page(pfn));
9377 spin_lock_irqsave(&zone->lock, flags);
9378 while (pfn < end_pfn) {
9379 page = pfn_to_page(pfn);
9381 * The HWPoisoned page may be not in buddy system, and
9382 * page_count() is not 0.
9384 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9389 * At this point all remaining PageOffline() pages have a
9390 * reference count of 0 and can simply be skipped.
9392 if (PageOffline(page)) {
9393 BUG_ON(page_count(page));
9394 BUG_ON(PageBuddy(page));
9399 BUG_ON(page_count(page));
9400 BUG_ON(!PageBuddy(page));
9401 order = buddy_order(page);
9402 del_page_from_free_list(page, zone, order);
9403 pfn += (1 << order);
9405 spin_unlock_irqrestore(&zone->lock, flags);
9410 * This function returns a stable result only if called under zone lock.
9412 bool is_free_buddy_page(struct page *page)
9414 unsigned long pfn = page_to_pfn(page);
9417 for (order = 0; order < MAX_ORDER; order++) {
9418 struct page *page_head = page - (pfn & ((1 << order) - 1));
9420 if (PageBuddy(page_head) &&
9421 buddy_order_unsafe(page_head) >= order)
9425 return order < MAX_ORDER;
9427 EXPORT_SYMBOL(is_free_buddy_page);
9429 #ifdef CONFIG_MEMORY_FAILURE
9431 * Break down a higher-order page in sub-pages, and keep our target out of
9434 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9435 struct page *target, int low, int high,
9438 unsigned long size = 1 << high;
9439 struct page *current_buddy, *next_page;
9441 while (high > low) {
9445 if (target >= &page[size]) {
9446 next_page = page + size;
9447 current_buddy = page;
9450 current_buddy = page + size;
9453 if (set_page_guard(zone, current_buddy, high, migratetype))
9456 if (current_buddy != target) {
9457 add_to_free_list(current_buddy, zone, high, migratetype);
9458 set_buddy_order(current_buddy, high);
9465 * Take a page that will be marked as poisoned off the buddy allocator.
9467 bool take_page_off_buddy(struct page *page)
9469 struct zone *zone = page_zone(page);
9470 unsigned long pfn = page_to_pfn(page);
9471 unsigned long flags;
9475 spin_lock_irqsave(&zone->lock, flags);
9476 for (order = 0; order < MAX_ORDER; order++) {
9477 struct page *page_head = page - (pfn & ((1 << order) - 1));
9478 int page_order = buddy_order(page_head);
9480 if (PageBuddy(page_head) && page_order >= order) {
9481 unsigned long pfn_head = page_to_pfn(page_head);
9482 int migratetype = get_pfnblock_migratetype(page_head,
9485 del_page_from_free_list(page_head, zone, page_order);
9486 break_down_buddy_pages(zone, page_head, page, 0,
9487 page_order, migratetype);
9488 SetPageHWPoisonTakenOff(page);
9489 if (!is_migrate_isolate(migratetype))
9490 __mod_zone_freepage_state(zone, -1, migratetype);
9494 if (page_count(page_head) > 0)
9497 spin_unlock_irqrestore(&zone->lock, flags);
9502 * Cancel takeoff done by take_page_off_buddy().
9504 bool put_page_back_buddy(struct page *page)
9506 struct zone *zone = page_zone(page);
9507 unsigned long pfn = page_to_pfn(page);
9508 unsigned long flags;
9509 int migratetype = get_pfnblock_migratetype(page, pfn);
9512 spin_lock_irqsave(&zone->lock, flags);
9513 if (put_page_testzero(page)) {
9514 ClearPageHWPoisonTakenOff(page);
9515 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9516 if (TestClearPageHWPoison(page)) {
9520 spin_unlock_irqrestore(&zone->lock, flags);
9526 #ifdef CONFIG_ZONE_DMA
9527 bool has_managed_dma(void)
9529 struct pglist_data *pgdat;
9531 for_each_online_pgdat(pgdat) {
9532 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9534 if (managed_zone(zone))
9539 #endif /* CONFIG_ZONE_DMA */