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 void destroy_large_folio(struct folio *folio)
749 enum compound_dtor_id dtor = folio_page(folio, 1)->compound_dtor;
751 VM_BUG_ON_FOLIO(dtor >= NR_COMPOUND_DTORS, folio);
752 compound_page_dtors[dtor](&folio->page);
755 #ifdef CONFIG_DEBUG_PAGEALLOC
756 unsigned int _debug_guardpage_minorder;
758 bool _debug_pagealloc_enabled_early __read_mostly
759 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
760 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
761 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
762 EXPORT_SYMBOL(_debug_pagealloc_enabled);
764 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
766 static int __init early_debug_pagealloc(char *buf)
768 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
770 early_param("debug_pagealloc", early_debug_pagealloc);
772 static int __init debug_guardpage_minorder_setup(char *buf)
776 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
777 pr_err("Bad debug_guardpage_minorder value\n");
780 _debug_guardpage_minorder = res;
781 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
784 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
786 static inline bool set_page_guard(struct zone *zone, struct page *page,
787 unsigned int order, int migratetype)
789 if (!debug_guardpage_enabled())
792 if (order >= debug_guardpage_minorder())
795 __SetPageGuard(page);
796 INIT_LIST_HEAD(&page->lru);
797 set_page_private(page, order);
798 /* Guard pages are not available for any usage */
799 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
804 static inline void clear_page_guard(struct zone *zone, struct page *page,
805 unsigned int order, int migratetype)
807 if (!debug_guardpage_enabled())
810 __ClearPageGuard(page);
812 set_page_private(page, 0);
813 if (!is_migrate_isolate(migratetype))
814 __mod_zone_freepage_state(zone, (1 << order), migratetype);
817 static inline bool set_page_guard(struct zone *zone, struct page *page,
818 unsigned int order, int migratetype) { return false; }
819 static inline void clear_page_guard(struct zone *zone, struct page *page,
820 unsigned int order, int migratetype) {}
824 * Enable static keys related to various memory debugging and hardening options.
825 * Some override others, and depend on early params that are evaluated in the
826 * order of appearance. So we need to first gather the full picture of what was
827 * enabled, and then make decisions.
829 void init_mem_debugging_and_hardening(void)
831 bool page_poisoning_requested = false;
833 #ifdef CONFIG_PAGE_POISONING
835 * Page poisoning is debug page alloc for some arches. If
836 * either of those options are enabled, enable poisoning.
838 if (page_poisoning_enabled() ||
839 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
840 debug_pagealloc_enabled())) {
841 static_branch_enable(&_page_poisoning_enabled);
842 page_poisoning_requested = true;
846 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
847 page_poisoning_requested) {
848 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
849 "will take precedence over init_on_alloc and init_on_free\n");
850 _init_on_alloc_enabled_early = false;
851 _init_on_free_enabled_early = false;
854 if (_init_on_alloc_enabled_early)
855 static_branch_enable(&init_on_alloc);
857 static_branch_disable(&init_on_alloc);
859 if (_init_on_free_enabled_early)
860 static_branch_enable(&init_on_free);
862 static_branch_disable(&init_on_free);
864 #ifdef CONFIG_DEBUG_PAGEALLOC
865 if (!debug_pagealloc_enabled())
868 static_branch_enable(&_debug_pagealloc_enabled);
870 if (!debug_guardpage_minorder())
873 static_branch_enable(&_debug_guardpage_enabled);
877 static inline void set_buddy_order(struct page *page, unsigned int order)
879 set_page_private(page, order);
880 __SetPageBuddy(page);
883 #ifdef CONFIG_COMPACTION
884 static inline struct capture_control *task_capc(struct zone *zone)
886 struct capture_control *capc = current->capture_control;
888 return unlikely(capc) &&
889 !(current->flags & PF_KTHREAD) &&
891 capc->cc->zone == zone ? capc : NULL;
895 compaction_capture(struct capture_control *capc, struct page *page,
896 int order, int migratetype)
898 if (!capc || order != capc->cc->order)
901 /* Do not accidentally pollute CMA or isolated regions*/
902 if (is_migrate_cma(migratetype) ||
903 is_migrate_isolate(migratetype))
907 * Do not let lower order allocations pollute a movable pageblock.
908 * This might let an unmovable request use a reclaimable pageblock
909 * and vice-versa but no more than normal fallback logic which can
910 * have trouble finding a high-order free page.
912 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
920 static inline struct capture_control *task_capc(struct zone *zone)
926 compaction_capture(struct capture_control *capc, struct page *page,
927 int order, int migratetype)
931 #endif /* CONFIG_COMPACTION */
933 /* Used for pages not on another list */
934 static inline void add_to_free_list(struct page *page, struct zone *zone,
935 unsigned int order, int migratetype)
937 struct free_area *area = &zone->free_area[order];
939 list_add(&page->lru, &area->free_list[migratetype]);
943 /* Used for pages not on another list */
944 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
945 unsigned int order, int migratetype)
947 struct free_area *area = &zone->free_area[order];
949 list_add_tail(&page->lru, &area->free_list[migratetype]);
954 * Used for pages which are on another list. Move the pages to the tail
955 * of the list - so the moved pages won't immediately be considered for
956 * allocation again (e.g., optimization for memory onlining).
958 static inline void move_to_free_list(struct page *page, struct zone *zone,
959 unsigned int order, int migratetype)
961 struct free_area *area = &zone->free_area[order];
963 list_move_tail(&page->lru, &area->free_list[migratetype]);
966 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
969 /* clear reported state and update reported page count */
970 if (page_reported(page))
971 __ClearPageReported(page);
973 list_del(&page->lru);
974 __ClearPageBuddy(page);
975 set_page_private(page, 0);
976 zone->free_area[order].nr_free--;
980 * If this is not the largest possible page, check if the buddy
981 * of the next-highest order is free. If it is, it's possible
982 * that pages are being freed that will coalesce soon. In case,
983 * that is happening, add the free page to the tail of the list
984 * so it's less likely to be used soon and more likely to be merged
985 * as a higher order page
988 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
989 struct page *page, unsigned int order)
991 unsigned long higher_page_pfn;
992 struct page *higher_page;
994 if (order >= MAX_ORDER - 2)
997 higher_page_pfn = buddy_pfn & pfn;
998 higher_page = page + (higher_page_pfn - pfn);
1000 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
1005 * Freeing function for a buddy system allocator.
1007 * The concept of a buddy system is to maintain direct-mapped table
1008 * (containing bit values) for memory blocks of various "orders".
1009 * The bottom level table contains the map for the smallest allocatable
1010 * units of memory (here, pages), and each level above it describes
1011 * pairs of units from the levels below, hence, "buddies".
1012 * At a high level, all that happens here is marking the table entry
1013 * at the bottom level available, and propagating the changes upward
1014 * as necessary, plus some accounting needed to play nicely with other
1015 * parts of the VM system.
1016 * At each level, we keep a list of pages, which are heads of continuous
1017 * free pages of length of (1 << order) and marked with PageBuddy.
1018 * Page's order is recorded in page_private(page) field.
1019 * So when we are allocating or freeing one, we can derive the state of the
1020 * other. That is, if we allocate a small block, and both were
1021 * free, the remainder of the region must be split into blocks.
1022 * If a block is freed, and its buddy is also free, then this
1023 * triggers coalescing into a block of larger size.
1028 static inline void __free_one_page(struct page *page,
1030 struct zone *zone, unsigned int order,
1031 int migratetype, fpi_t fpi_flags)
1033 struct capture_control *capc = task_capc(zone);
1034 unsigned long buddy_pfn;
1035 unsigned long combined_pfn;
1039 VM_BUG_ON(!zone_is_initialized(zone));
1040 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1042 VM_BUG_ON(migratetype == -1);
1043 if (likely(!is_migrate_isolate(migratetype)))
1044 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1046 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1047 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1049 while (order < MAX_ORDER - 1) {
1050 if (compaction_capture(capc, page, order, migratetype)) {
1051 __mod_zone_freepage_state(zone, -(1 << order),
1056 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
1060 if (unlikely(order >= pageblock_order)) {
1062 * We want to prevent merge between freepages on pageblock
1063 * without fallbacks and normal pageblock. Without this,
1064 * pageblock isolation could cause incorrect freepage or CMA
1065 * accounting or HIGHATOMIC accounting.
1067 int buddy_mt = get_pageblock_migratetype(buddy);
1069 if (migratetype != buddy_mt
1070 && (!migratetype_is_mergeable(migratetype) ||
1071 !migratetype_is_mergeable(buddy_mt)))
1076 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1077 * merge with it and move up one order.
1079 if (page_is_guard(buddy))
1080 clear_page_guard(zone, buddy, order, migratetype);
1082 del_page_from_free_list(buddy, zone, order);
1083 combined_pfn = buddy_pfn & pfn;
1084 page = page + (combined_pfn - pfn);
1090 set_buddy_order(page, order);
1092 if (fpi_flags & FPI_TO_TAIL)
1094 else if (is_shuffle_order(order))
1095 to_tail = shuffle_pick_tail();
1097 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1100 add_to_free_list_tail(page, zone, order, migratetype);
1102 add_to_free_list(page, zone, order, migratetype);
1104 /* Notify page reporting subsystem of freed page */
1105 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1106 page_reporting_notify_free(order);
1110 * split_free_page() -- split a free page at split_pfn_offset
1111 * @free_page: the original free page
1112 * @order: the order of the page
1113 * @split_pfn_offset: split offset within the page
1115 * Return -ENOENT if the free page is changed, otherwise 0
1117 * It is used when the free page crosses two pageblocks with different migratetypes
1118 * at split_pfn_offset within the page. The split free page will be put into
1119 * separate migratetype lists afterwards. Otherwise, the function achieves
1122 int split_free_page(struct page *free_page,
1123 unsigned int order, unsigned long split_pfn_offset)
1125 struct zone *zone = page_zone(free_page);
1126 unsigned long free_page_pfn = page_to_pfn(free_page);
1128 unsigned long flags;
1129 int free_page_order;
1133 if (split_pfn_offset == 0)
1136 spin_lock_irqsave(&zone->lock, flags);
1138 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
1143 mt = get_pageblock_migratetype(free_page);
1144 if (likely(!is_migrate_isolate(mt)))
1145 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1147 del_page_from_free_list(free_page, zone, order);
1148 for (pfn = free_page_pfn;
1149 pfn < free_page_pfn + (1UL << order);) {
1150 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
1152 free_page_order = min_t(unsigned int,
1153 pfn ? __ffs(pfn) : order,
1154 __fls(split_pfn_offset));
1155 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
1157 pfn += 1UL << free_page_order;
1158 split_pfn_offset -= (1UL << free_page_order);
1159 /* we have done the first part, now switch to second part */
1160 if (split_pfn_offset == 0)
1161 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
1164 spin_unlock_irqrestore(&zone->lock, flags);
1168 * A bad page could be due to a number of fields. Instead of multiple branches,
1169 * try and check multiple fields with one check. The caller must do a detailed
1170 * check if necessary.
1172 static inline bool page_expected_state(struct page *page,
1173 unsigned long check_flags)
1175 if (unlikely(atomic_read(&page->_mapcount) != -1))
1178 if (unlikely((unsigned long)page->mapping |
1179 page_ref_count(page) |
1183 (page->flags & check_flags)))
1189 static const char *page_bad_reason(struct page *page, unsigned long flags)
1191 const char *bad_reason = NULL;
1193 if (unlikely(atomic_read(&page->_mapcount) != -1))
1194 bad_reason = "nonzero mapcount";
1195 if (unlikely(page->mapping != NULL))
1196 bad_reason = "non-NULL mapping";
1197 if (unlikely(page_ref_count(page) != 0))
1198 bad_reason = "nonzero _refcount";
1199 if (unlikely(page->flags & flags)) {
1200 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1201 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1203 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1206 if (unlikely(page->memcg_data))
1207 bad_reason = "page still charged to cgroup";
1212 static void check_free_page_bad(struct page *page)
1215 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1218 static inline int check_free_page(struct page *page)
1220 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1223 /* Something has gone sideways, find it */
1224 check_free_page_bad(page);
1228 static int free_tail_pages_check(struct page *head_page, struct page *page)
1233 * We rely page->lru.next never has bit 0 set, unless the page
1234 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1236 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1238 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1242 switch (page - head_page) {
1244 /* the first tail page: ->mapping may be compound_mapcount() */
1245 if (unlikely(compound_mapcount(page))) {
1246 bad_page(page, "nonzero compound_mapcount");
1252 * the second tail page: ->mapping is
1253 * deferred_list.next -- ignore value.
1257 if (page->mapping != TAIL_MAPPING) {
1258 bad_page(page, "corrupted mapping in tail page");
1263 if (unlikely(!PageTail(page))) {
1264 bad_page(page, "PageTail not set");
1267 if (unlikely(compound_head(page) != head_page)) {
1268 bad_page(page, "compound_head not consistent");
1273 page->mapping = NULL;
1274 clear_compound_head(page);
1279 * Skip KASAN memory poisoning when either:
1281 * 1. Deferred memory initialization has not yet completed,
1282 * see the explanation below.
1283 * 2. Skipping poisoning is requested via FPI_SKIP_KASAN_POISON,
1284 * see the comment next to it.
1285 * 3. Skipping poisoning is requested via __GFP_SKIP_KASAN_POISON,
1286 * see the comment next to it.
1288 * Poisoning pages during deferred memory init will greatly lengthen the
1289 * process and cause problem in large memory systems as the deferred pages
1290 * initialization is done with interrupt disabled.
1292 * Assuming that there will be no reference to those newly initialized
1293 * pages before they are ever allocated, this should have no effect on
1294 * KASAN memory tracking as the poison will be properly inserted at page
1295 * allocation time. The only corner case is when pages are allocated by
1296 * on-demand allocation and then freed again before the deferred pages
1297 * initialization is done, but this is not likely to happen.
1299 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1301 return deferred_pages_enabled() ||
1302 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
1303 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
1304 PageSkipKASanPoison(page);
1307 static void kernel_init_pages(struct page *page, int numpages)
1311 /* s390's use of memset() could override KASAN redzones. */
1312 kasan_disable_current();
1313 for (i = 0; i < numpages; i++)
1314 clear_highpage_kasan_tagged(page + i);
1315 kasan_enable_current();
1318 static __always_inline bool free_pages_prepare(struct page *page,
1319 unsigned int order, bool check_free, fpi_t fpi_flags)
1322 bool init = want_init_on_free();
1324 VM_BUG_ON_PAGE(PageTail(page), page);
1326 trace_mm_page_free(page, order);
1328 if (unlikely(PageHWPoison(page)) && !order) {
1330 * Do not let hwpoison pages hit pcplists/buddy
1331 * Untie memcg state and reset page's owner
1333 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1334 __memcg_kmem_uncharge_page(page, order);
1335 reset_page_owner(page, order);
1336 page_table_check_free(page, order);
1341 * Check tail pages before head page information is cleared to
1342 * avoid checking PageCompound for order-0 pages.
1344 if (unlikely(order)) {
1345 bool compound = PageCompound(page);
1348 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1351 ClearPageDoubleMap(page);
1352 ClearPageHasHWPoisoned(page);
1354 for (i = 1; i < (1 << order); i++) {
1356 bad += free_tail_pages_check(page, page + i);
1357 if (unlikely(check_free_page(page + i))) {
1361 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1364 if (PageMappingFlags(page))
1365 page->mapping = NULL;
1366 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1367 __memcg_kmem_uncharge_page(page, order);
1369 bad += check_free_page(page);
1373 page_cpupid_reset_last(page);
1374 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1375 reset_page_owner(page, order);
1376 page_table_check_free(page, order);
1378 if (!PageHighMem(page)) {
1379 debug_check_no_locks_freed(page_address(page),
1380 PAGE_SIZE << order);
1381 debug_check_no_obj_freed(page_address(page),
1382 PAGE_SIZE << order);
1385 kernel_poison_pages(page, 1 << order);
1388 * As memory initialization might be integrated into KASAN,
1389 * KASAN poisoning and memory initialization code must be
1390 * kept together to avoid discrepancies in behavior.
1392 * With hardware tag-based KASAN, memory tags must be set before the
1393 * page becomes unavailable via debug_pagealloc or arch_free_page.
1395 if (!should_skip_kasan_poison(page, fpi_flags)) {
1396 kasan_poison_pages(page, order, init);
1398 /* Memory is already initialized if KASAN did it internally. */
1399 if (kasan_has_integrated_init())
1403 kernel_init_pages(page, 1 << order);
1406 * arch_free_page() can make the page's contents inaccessible. s390
1407 * does this. So nothing which can access the page's contents should
1408 * happen after this.
1410 arch_free_page(page, order);
1412 debug_pagealloc_unmap_pages(page, 1 << order);
1417 #ifdef CONFIG_DEBUG_VM
1419 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1420 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1421 * moved from pcp lists to free lists.
1423 static bool free_pcp_prepare(struct page *page, unsigned int order)
1425 return free_pages_prepare(page, order, true, FPI_NONE);
1428 static bool bulkfree_pcp_prepare(struct page *page)
1430 if (debug_pagealloc_enabled_static())
1431 return check_free_page(page);
1437 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1438 * moving from pcp lists to free list in order to reduce overhead. With
1439 * debug_pagealloc enabled, they are checked also immediately when being freed
1442 static bool free_pcp_prepare(struct page *page, unsigned int order)
1444 if (debug_pagealloc_enabled_static())
1445 return free_pages_prepare(page, order, true, FPI_NONE);
1447 return free_pages_prepare(page, order, false, FPI_NONE);
1450 static bool bulkfree_pcp_prepare(struct page *page)
1452 return check_free_page(page);
1454 #endif /* CONFIG_DEBUG_VM */
1457 * Frees a number of pages from the PCP lists
1458 * Assumes all pages on list are in same zone.
1459 * count is the number of pages to free.
1461 static void free_pcppages_bulk(struct zone *zone, int count,
1462 struct per_cpu_pages *pcp,
1466 int max_pindex = NR_PCP_LISTS - 1;
1468 bool isolated_pageblocks;
1472 * Ensure proper count is passed which otherwise would stuck in the
1473 * below while (list_empty(list)) loop.
1475 count = min(pcp->count, count);
1477 /* Ensure requested pindex is drained first. */
1478 pindex = pindex - 1;
1481 * local_lock_irq held so equivalent to spin_lock_irqsave for
1482 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1484 spin_lock(&zone->lock);
1485 isolated_pageblocks = has_isolate_pageblock(zone);
1488 struct list_head *list;
1491 /* Remove pages from lists in a round-robin fashion. */
1493 if (++pindex > max_pindex)
1494 pindex = min_pindex;
1495 list = &pcp->lists[pindex];
1496 if (!list_empty(list))
1499 if (pindex == max_pindex)
1501 if (pindex == min_pindex)
1505 order = pindex_to_order(pindex);
1506 nr_pages = 1 << order;
1507 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1511 page = list_last_entry(list, struct page, lru);
1512 mt = get_pcppage_migratetype(page);
1514 /* must delete to avoid corrupting pcp list */
1515 list_del(&page->lru);
1517 pcp->count -= nr_pages;
1519 if (bulkfree_pcp_prepare(page))
1522 /* MIGRATE_ISOLATE page should not go to pcplists */
1523 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1524 /* Pageblock could have been isolated meanwhile */
1525 if (unlikely(isolated_pageblocks))
1526 mt = get_pageblock_migratetype(page);
1528 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1529 trace_mm_page_pcpu_drain(page, order, mt);
1530 } while (count > 0 && !list_empty(list));
1533 spin_unlock(&zone->lock);
1536 static void free_one_page(struct zone *zone,
1537 struct page *page, unsigned long pfn,
1539 int migratetype, fpi_t fpi_flags)
1541 unsigned long flags;
1543 spin_lock_irqsave(&zone->lock, flags);
1544 if (unlikely(has_isolate_pageblock(zone) ||
1545 is_migrate_isolate(migratetype))) {
1546 migratetype = get_pfnblock_migratetype(page, pfn);
1548 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1549 spin_unlock_irqrestore(&zone->lock, flags);
1552 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1553 unsigned long zone, int nid)
1555 mm_zero_struct_page(page);
1556 set_page_links(page, zone, nid, pfn);
1557 init_page_count(page);
1558 page_mapcount_reset(page);
1559 page_cpupid_reset_last(page);
1560 page_kasan_tag_reset(page);
1562 INIT_LIST_HEAD(&page->lru);
1563 #ifdef WANT_PAGE_VIRTUAL
1564 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1565 if (!is_highmem_idx(zone))
1566 set_page_address(page, __va(pfn << PAGE_SHIFT));
1570 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1571 static void __meminit init_reserved_page(unsigned long pfn)
1576 if (!early_page_uninitialised(pfn))
1579 nid = early_pfn_to_nid(pfn);
1580 pgdat = NODE_DATA(nid);
1582 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1583 struct zone *zone = &pgdat->node_zones[zid];
1585 if (zone_spans_pfn(zone, pfn))
1588 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1591 static inline void init_reserved_page(unsigned long pfn)
1594 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1597 * Initialised pages do not have PageReserved set. This function is
1598 * called for each range allocated by the bootmem allocator and
1599 * marks the pages PageReserved. The remaining valid pages are later
1600 * sent to the buddy page allocator.
1602 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1604 unsigned long start_pfn = PFN_DOWN(start);
1605 unsigned long end_pfn = PFN_UP(end);
1607 for (; start_pfn < end_pfn; start_pfn++) {
1608 if (pfn_valid(start_pfn)) {
1609 struct page *page = pfn_to_page(start_pfn);
1611 init_reserved_page(start_pfn);
1613 /* Avoid false-positive PageTail() */
1614 INIT_LIST_HEAD(&page->lru);
1617 * no need for atomic set_bit because the struct
1618 * page is not visible yet so nobody should
1621 __SetPageReserved(page);
1626 static void __free_pages_ok(struct page *page, unsigned int order,
1629 unsigned long flags;
1631 unsigned long pfn = page_to_pfn(page);
1632 struct zone *zone = page_zone(page);
1634 if (!free_pages_prepare(page, order, true, fpi_flags))
1637 migratetype = get_pfnblock_migratetype(page, pfn);
1639 spin_lock_irqsave(&zone->lock, flags);
1640 if (unlikely(has_isolate_pageblock(zone) ||
1641 is_migrate_isolate(migratetype))) {
1642 migratetype = get_pfnblock_migratetype(page, pfn);
1644 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1645 spin_unlock_irqrestore(&zone->lock, flags);
1647 __count_vm_events(PGFREE, 1 << order);
1650 void __free_pages_core(struct page *page, unsigned int order)
1652 unsigned int nr_pages = 1 << order;
1653 struct page *p = page;
1657 * When initializing the memmap, __init_single_page() sets the refcount
1658 * of all pages to 1 ("allocated"/"not free"). We have to set the
1659 * refcount of all involved pages to 0.
1662 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1664 __ClearPageReserved(p);
1665 set_page_count(p, 0);
1667 __ClearPageReserved(p);
1668 set_page_count(p, 0);
1670 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1673 * Bypass PCP and place fresh pages right to the tail, primarily
1674 * relevant for memory onlining.
1676 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1682 * During memory init memblocks map pfns to nids. The search is expensive and
1683 * this caches recent lookups. The implementation of __early_pfn_to_nid
1684 * treats start/end as pfns.
1686 struct mminit_pfnnid_cache {
1687 unsigned long last_start;
1688 unsigned long last_end;
1692 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1695 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1697 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1698 struct mminit_pfnnid_cache *state)
1700 unsigned long start_pfn, end_pfn;
1703 if (state->last_start <= pfn && pfn < state->last_end)
1704 return state->last_nid;
1706 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1707 if (nid != NUMA_NO_NODE) {
1708 state->last_start = start_pfn;
1709 state->last_end = end_pfn;
1710 state->last_nid = nid;
1716 int __meminit early_pfn_to_nid(unsigned long pfn)
1718 static DEFINE_SPINLOCK(early_pfn_lock);
1721 spin_lock(&early_pfn_lock);
1722 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1724 nid = first_online_node;
1725 spin_unlock(&early_pfn_lock);
1729 #endif /* CONFIG_NUMA */
1731 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1734 if (early_page_uninitialised(pfn))
1736 __free_pages_core(page, order);
1740 * Check that the whole (or subset of) a pageblock given by the interval of
1741 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1742 * with the migration of free compaction scanner.
1744 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1746 * It's possible on some configurations to have a setup like node0 node1 node0
1747 * i.e. it's possible that all pages within a zones range of pages do not
1748 * belong to a single zone. We assume that a border between node0 and node1
1749 * can occur within a single pageblock, but not a node0 node1 node0
1750 * interleaving within a single pageblock. It is therefore sufficient to check
1751 * the first and last page of a pageblock and avoid checking each individual
1752 * page in a pageblock.
1754 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1755 unsigned long end_pfn, struct zone *zone)
1757 struct page *start_page;
1758 struct page *end_page;
1760 /* end_pfn is one past the range we are checking */
1763 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1766 start_page = pfn_to_online_page(start_pfn);
1770 if (page_zone(start_page) != zone)
1773 end_page = pfn_to_page(end_pfn);
1775 /* This gives a shorter code than deriving page_zone(end_page) */
1776 if (page_zone_id(start_page) != page_zone_id(end_page))
1782 void set_zone_contiguous(struct zone *zone)
1784 unsigned long block_start_pfn = zone->zone_start_pfn;
1785 unsigned long block_end_pfn;
1787 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1788 for (; block_start_pfn < zone_end_pfn(zone);
1789 block_start_pfn = block_end_pfn,
1790 block_end_pfn += pageblock_nr_pages) {
1792 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1794 if (!__pageblock_pfn_to_page(block_start_pfn,
1795 block_end_pfn, zone))
1800 /* We confirm that there is no hole */
1801 zone->contiguous = true;
1804 void clear_zone_contiguous(struct zone *zone)
1806 zone->contiguous = false;
1809 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1810 static void __init deferred_free_range(unsigned long pfn,
1811 unsigned long nr_pages)
1819 page = pfn_to_page(pfn);
1821 /* Free a large naturally-aligned chunk if possible */
1822 if (nr_pages == pageblock_nr_pages &&
1823 (pfn & (pageblock_nr_pages - 1)) == 0) {
1824 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1825 __free_pages_core(page, pageblock_order);
1829 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1830 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1831 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1832 __free_pages_core(page, 0);
1836 /* Completion tracking for deferred_init_memmap() threads */
1837 static atomic_t pgdat_init_n_undone __initdata;
1838 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1840 static inline void __init pgdat_init_report_one_done(void)
1842 if (atomic_dec_and_test(&pgdat_init_n_undone))
1843 complete(&pgdat_init_all_done_comp);
1847 * Returns true if page needs to be initialized or freed to buddy allocator.
1849 * First we check if pfn is valid on architectures where it is possible to have
1850 * holes within pageblock_nr_pages. On systems where it is not possible, this
1851 * function is optimized out.
1853 * Then, we check if a current large page is valid by only checking the validity
1856 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1858 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1864 * Free pages to buddy allocator. Try to free aligned pages in
1865 * pageblock_nr_pages sizes.
1867 static void __init deferred_free_pages(unsigned long pfn,
1868 unsigned long end_pfn)
1870 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1871 unsigned long nr_free = 0;
1873 for (; pfn < end_pfn; pfn++) {
1874 if (!deferred_pfn_valid(pfn)) {
1875 deferred_free_range(pfn - nr_free, nr_free);
1877 } else if (!(pfn & nr_pgmask)) {
1878 deferred_free_range(pfn - nr_free, nr_free);
1884 /* Free the last block of pages to allocator */
1885 deferred_free_range(pfn - nr_free, nr_free);
1889 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1890 * by performing it only once every pageblock_nr_pages.
1891 * Return number of pages initialized.
1893 static unsigned long __init deferred_init_pages(struct zone *zone,
1895 unsigned long end_pfn)
1897 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1898 int nid = zone_to_nid(zone);
1899 unsigned long nr_pages = 0;
1900 int zid = zone_idx(zone);
1901 struct page *page = NULL;
1903 for (; pfn < end_pfn; pfn++) {
1904 if (!deferred_pfn_valid(pfn)) {
1907 } else if (!page || !(pfn & nr_pgmask)) {
1908 page = pfn_to_page(pfn);
1912 __init_single_page(page, pfn, zid, nid);
1919 * This function is meant to pre-load the iterator for the zone init.
1920 * Specifically it walks through the ranges until we are caught up to the
1921 * first_init_pfn value and exits there. If we never encounter the value we
1922 * return false indicating there are no valid ranges left.
1925 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1926 unsigned long *spfn, unsigned long *epfn,
1927 unsigned long first_init_pfn)
1932 * Start out by walking through the ranges in this zone that have
1933 * already been initialized. We don't need to do anything with them
1934 * so we just need to flush them out of the system.
1936 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1937 if (*epfn <= first_init_pfn)
1939 if (*spfn < first_init_pfn)
1940 *spfn = first_init_pfn;
1949 * Initialize and free pages. We do it in two loops: first we initialize
1950 * struct page, then free to buddy allocator, because while we are
1951 * freeing pages we can access pages that are ahead (computing buddy
1952 * page in __free_one_page()).
1954 * In order to try and keep some memory in the cache we have the loop
1955 * broken along max page order boundaries. This way we will not cause
1956 * any issues with the buddy page computation.
1958 static unsigned long __init
1959 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1960 unsigned long *end_pfn)
1962 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1963 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1964 unsigned long nr_pages = 0;
1967 /* First we loop through and initialize the page values */
1968 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1971 if (mo_pfn <= *start_pfn)
1974 t = min(mo_pfn, *end_pfn);
1975 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1977 if (mo_pfn < *end_pfn) {
1978 *start_pfn = mo_pfn;
1983 /* Reset values and now loop through freeing pages as needed */
1986 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1992 t = min(mo_pfn, epfn);
1993 deferred_free_pages(spfn, t);
2003 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2006 unsigned long spfn, epfn;
2007 struct zone *zone = arg;
2010 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2013 * Initialize and free pages in MAX_ORDER sized increments so that we
2014 * can avoid introducing any issues with the buddy allocator.
2016 while (spfn < end_pfn) {
2017 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2022 /* An arch may override for more concurrency. */
2024 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2029 /* Initialise remaining memory on a node */
2030 static int __init deferred_init_memmap(void *data)
2032 pg_data_t *pgdat = data;
2033 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2034 unsigned long spfn = 0, epfn = 0;
2035 unsigned long first_init_pfn, flags;
2036 unsigned long start = jiffies;
2038 int zid, max_threads;
2041 /* Bind memory initialisation thread to a local node if possible */
2042 if (!cpumask_empty(cpumask))
2043 set_cpus_allowed_ptr(current, cpumask);
2045 pgdat_resize_lock(pgdat, &flags);
2046 first_init_pfn = pgdat->first_deferred_pfn;
2047 if (first_init_pfn == ULONG_MAX) {
2048 pgdat_resize_unlock(pgdat, &flags);
2049 pgdat_init_report_one_done();
2053 /* Sanity check boundaries */
2054 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2055 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2056 pgdat->first_deferred_pfn = ULONG_MAX;
2059 * Once we unlock here, the zone cannot be grown anymore, thus if an
2060 * interrupt thread must allocate this early in boot, zone must be
2061 * pre-grown prior to start of deferred page initialization.
2063 pgdat_resize_unlock(pgdat, &flags);
2065 /* Only the highest zone is deferred so find it */
2066 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2067 zone = pgdat->node_zones + zid;
2068 if (first_init_pfn < zone_end_pfn(zone))
2072 /* If the zone is empty somebody else may have cleared out the zone */
2073 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2077 max_threads = deferred_page_init_max_threads(cpumask);
2079 while (spfn < epfn) {
2080 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2081 struct padata_mt_job job = {
2082 .thread_fn = deferred_init_memmap_chunk,
2085 .size = epfn_align - spfn,
2086 .align = PAGES_PER_SECTION,
2087 .min_chunk = PAGES_PER_SECTION,
2088 .max_threads = max_threads,
2091 padata_do_multithreaded(&job);
2092 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2096 /* Sanity check that the next zone really is unpopulated */
2097 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2099 pr_info("node %d deferred pages initialised in %ums\n",
2100 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2102 pgdat_init_report_one_done();
2107 * If this zone has deferred pages, try to grow it by initializing enough
2108 * deferred pages to satisfy the allocation specified by order, rounded up to
2109 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2110 * of SECTION_SIZE bytes by initializing struct pages in increments of
2111 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2113 * Return true when zone was grown, otherwise return false. We return true even
2114 * when we grow less than requested, to let the caller decide if there are
2115 * enough pages to satisfy the allocation.
2117 * Note: We use noinline because this function is needed only during boot, and
2118 * it is called from a __ref function _deferred_grow_zone. This way we are
2119 * making sure that it is not inlined into permanent text section.
2121 static noinline bool __init
2122 deferred_grow_zone(struct zone *zone, unsigned int order)
2124 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2125 pg_data_t *pgdat = zone->zone_pgdat;
2126 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2127 unsigned long spfn, epfn, flags;
2128 unsigned long nr_pages = 0;
2131 /* Only the last zone may have deferred pages */
2132 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2135 pgdat_resize_lock(pgdat, &flags);
2138 * If someone grew this zone while we were waiting for spinlock, return
2139 * true, as there might be enough pages already.
2141 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2142 pgdat_resize_unlock(pgdat, &flags);
2146 /* If the zone is empty somebody else may have cleared out the zone */
2147 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2148 first_deferred_pfn)) {
2149 pgdat->first_deferred_pfn = ULONG_MAX;
2150 pgdat_resize_unlock(pgdat, &flags);
2151 /* Retry only once. */
2152 return first_deferred_pfn != ULONG_MAX;
2156 * Initialize and free pages in MAX_ORDER sized increments so
2157 * that we can avoid introducing any issues with the buddy
2160 while (spfn < epfn) {
2161 /* update our first deferred PFN for this section */
2162 first_deferred_pfn = spfn;
2164 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2165 touch_nmi_watchdog();
2167 /* We should only stop along section boundaries */
2168 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2171 /* If our quota has been met we can stop here */
2172 if (nr_pages >= nr_pages_needed)
2176 pgdat->first_deferred_pfn = spfn;
2177 pgdat_resize_unlock(pgdat, &flags);
2179 return nr_pages > 0;
2183 * deferred_grow_zone() is __init, but it is called from
2184 * get_page_from_freelist() during early boot until deferred_pages permanently
2185 * disables this call. This is why we have refdata wrapper to avoid warning,
2186 * and to ensure that the function body gets unloaded.
2189 _deferred_grow_zone(struct zone *zone, unsigned int order)
2191 return deferred_grow_zone(zone, order);
2194 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2196 void __init page_alloc_init_late(void)
2201 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2203 /* There will be num_node_state(N_MEMORY) threads */
2204 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2205 for_each_node_state(nid, N_MEMORY) {
2206 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2209 /* Block until all are initialised */
2210 wait_for_completion(&pgdat_init_all_done_comp);
2213 * We initialized the rest of the deferred pages. Permanently disable
2214 * on-demand struct page initialization.
2216 static_branch_disable(&deferred_pages);
2218 /* Reinit limits that are based on free pages after the kernel is up */
2219 files_maxfiles_init();
2224 /* Discard memblock private memory */
2227 for_each_node_state(nid, N_MEMORY)
2228 shuffle_free_memory(NODE_DATA(nid));
2230 for_each_populated_zone(zone)
2231 set_zone_contiguous(zone);
2235 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2236 void __init init_cma_reserved_pageblock(struct page *page)
2238 unsigned i = pageblock_nr_pages;
2239 struct page *p = page;
2242 __ClearPageReserved(p);
2243 set_page_count(p, 0);
2246 set_pageblock_migratetype(page, MIGRATE_CMA);
2247 set_page_refcounted(page);
2248 __free_pages(page, pageblock_order);
2250 adjust_managed_page_count(page, pageblock_nr_pages);
2251 page_zone(page)->cma_pages += pageblock_nr_pages;
2256 * The order of subdivision here is critical for the IO subsystem.
2257 * Please do not alter this order without good reasons and regression
2258 * testing. Specifically, as large blocks of memory are subdivided,
2259 * the order in which smaller blocks are delivered depends on the order
2260 * they're subdivided in this function. This is the primary factor
2261 * influencing the order in which pages are delivered to the IO
2262 * subsystem according to empirical testing, and this is also justified
2263 * by considering the behavior of a buddy system containing a single
2264 * large block of memory acted on by a series of small allocations.
2265 * This behavior is a critical factor in sglist merging's success.
2269 static inline void expand(struct zone *zone, struct page *page,
2270 int low, int high, int migratetype)
2272 unsigned long size = 1 << high;
2274 while (high > low) {
2277 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2280 * Mark as guard pages (or page), that will allow to
2281 * merge back to allocator when buddy will be freed.
2282 * Corresponding page table entries will not be touched,
2283 * pages will stay not present in virtual address space
2285 if (set_page_guard(zone, &page[size], high, migratetype))
2288 add_to_free_list(&page[size], zone, high, migratetype);
2289 set_buddy_order(&page[size], high);
2293 static void check_new_page_bad(struct page *page)
2295 if (unlikely(page->flags & __PG_HWPOISON)) {
2296 /* Don't complain about hwpoisoned pages */
2297 page_mapcount_reset(page); /* remove PageBuddy */
2302 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2306 * This page is about to be returned from the page allocator
2308 static inline int check_new_page(struct page *page)
2310 if (likely(page_expected_state(page,
2311 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2314 check_new_page_bad(page);
2318 static bool check_new_pages(struct page *page, unsigned int order)
2321 for (i = 0; i < (1 << order); i++) {
2322 struct page *p = page + i;
2324 if (unlikely(check_new_page(p)))
2331 #ifdef CONFIG_DEBUG_VM
2333 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2334 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2335 * also checked when pcp lists are refilled from the free lists.
2337 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2339 if (debug_pagealloc_enabled_static())
2340 return check_new_pages(page, order);
2345 static inline bool check_new_pcp(struct page *page, unsigned int order)
2347 return check_new_pages(page, order);
2351 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2352 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2353 * enabled, they are also checked when being allocated from the pcp lists.
2355 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2357 return check_new_pages(page, order);
2359 static inline bool check_new_pcp(struct page *page, unsigned int order)
2361 if (debug_pagealloc_enabled_static())
2362 return check_new_pages(page, order);
2366 #endif /* CONFIG_DEBUG_VM */
2368 static inline bool should_skip_kasan_unpoison(gfp_t flags, bool init_tags)
2370 /* Don't skip if a software KASAN mode is enabled. */
2371 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
2372 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
2375 /* Skip, if hardware tag-based KASAN is not enabled. */
2376 if (!kasan_hw_tags_enabled())
2380 * With hardware tag-based KASAN enabled, skip if either:
2382 * 1. Memory tags have already been cleared via tag_clear_highpage().
2383 * 2. Skipping has been requested via __GFP_SKIP_KASAN_UNPOISON.
2385 return init_tags || (flags & __GFP_SKIP_KASAN_UNPOISON);
2388 static inline bool should_skip_init(gfp_t flags)
2390 /* Don't skip, if hardware tag-based KASAN is not enabled. */
2391 if (!kasan_hw_tags_enabled())
2394 /* For hardware tag-based KASAN, skip if requested. */
2395 return (flags & __GFP_SKIP_ZERO);
2398 inline void post_alloc_hook(struct page *page, unsigned int order,
2401 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
2402 !should_skip_init(gfp_flags);
2403 bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
2405 set_page_private(page, 0);
2406 set_page_refcounted(page);
2408 arch_alloc_page(page, order);
2409 debug_pagealloc_map_pages(page, 1 << order);
2412 * Page unpoisoning must happen before memory initialization.
2413 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2414 * allocations and the page unpoisoning code will complain.
2416 kernel_unpoison_pages(page, 1 << order);
2419 * As memory initialization might be integrated into KASAN,
2420 * KASAN unpoisoning and memory initializion code must be
2421 * kept together to avoid discrepancies in behavior.
2425 * If memory tags should be zeroed (which happens only when memory
2426 * should be initialized as well).
2431 /* Initialize both memory and tags. */
2432 for (i = 0; i != 1 << order; ++i)
2433 tag_clear_highpage(page + i);
2435 /* Note that memory is already initialized by the loop above. */
2438 if (!should_skip_kasan_unpoison(gfp_flags, init_tags)) {
2439 /* Unpoison shadow memory or set memory tags. */
2440 kasan_unpoison_pages(page, order, init);
2442 /* Note that memory is already initialized by KASAN. */
2443 if (kasan_has_integrated_init())
2446 /* If memory is still not initialized, do it now. */
2448 kernel_init_pages(page, 1 << order);
2449 /* Propagate __GFP_SKIP_KASAN_POISON to page flags. */
2450 if (kasan_hw_tags_enabled() && (gfp_flags & __GFP_SKIP_KASAN_POISON))
2451 SetPageSkipKASanPoison(page);
2453 set_page_owner(page, order, gfp_flags);
2454 page_table_check_alloc(page, order);
2457 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2458 unsigned int alloc_flags)
2460 post_alloc_hook(page, order, gfp_flags);
2462 if (order && (gfp_flags & __GFP_COMP))
2463 prep_compound_page(page, order);
2466 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2467 * allocate the page. The expectation is that the caller is taking
2468 * steps that will free more memory. The caller should avoid the page
2469 * being used for !PFMEMALLOC purposes.
2471 if (alloc_flags & ALLOC_NO_WATERMARKS)
2472 set_page_pfmemalloc(page);
2474 clear_page_pfmemalloc(page);
2478 * Go through the free lists for the given migratetype and remove
2479 * the smallest available page from the freelists
2481 static __always_inline
2482 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2485 unsigned int current_order;
2486 struct free_area *area;
2489 /* Find a page of the appropriate size in the preferred list */
2490 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2491 area = &(zone->free_area[current_order]);
2492 page = get_page_from_free_area(area, migratetype);
2495 del_page_from_free_list(page, zone, current_order);
2496 expand(zone, page, order, current_order, migratetype);
2497 set_pcppage_migratetype(page, migratetype);
2498 trace_mm_page_alloc_zone_locked(page, order, migratetype,
2499 pcp_allowed_order(order) &&
2500 migratetype < MIGRATE_PCPTYPES);
2509 * This array describes the order lists are fallen back to when
2510 * the free lists for the desirable migrate type are depleted
2512 * The other migratetypes do not have fallbacks.
2514 static int fallbacks[MIGRATE_TYPES][3] = {
2515 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2516 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2517 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2521 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2524 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2527 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2528 unsigned int order) { return NULL; }
2532 * Move the free pages in a range to the freelist tail of the requested type.
2533 * Note that start_page and end_pages are not aligned on a pageblock
2534 * boundary. If alignment is required, use move_freepages_block()
2536 static int move_freepages(struct zone *zone,
2537 unsigned long start_pfn, unsigned long end_pfn,
2538 int migratetype, int *num_movable)
2543 int pages_moved = 0;
2545 for (pfn = start_pfn; pfn <= end_pfn;) {
2546 page = pfn_to_page(pfn);
2547 if (!PageBuddy(page)) {
2549 * We assume that pages that could be isolated for
2550 * migration are movable. But we don't actually try
2551 * isolating, as that would be expensive.
2554 (PageLRU(page) || __PageMovable(page)))
2560 /* Make sure we are not inadvertently changing nodes */
2561 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2562 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2564 order = buddy_order(page);
2565 move_to_free_list(page, zone, order, migratetype);
2567 pages_moved += 1 << order;
2573 int move_freepages_block(struct zone *zone, struct page *page,
2574 int migratetype, int *num_movable)
2576 unsigned long start_pfn, end_pfn, pfn;
2581 pfn = page_to_pfn(page);
2582 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2583 end_pfn = start_pfn + pageblock_nr_pages - 1;
2585 /* Do not cross zone boundaries */
2586 if (!zone_spans_pfn(zone, start_pfn))
2588 if (!zone_spans_pfn(zone, end_pfn))
2591 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2595 static void change_pageblock_range(struct page *pageblock_page,
2596 int start_order, int migratetype)
2598 int nr_pageblocks = 1 << (start_order - pageblock_order);
2600 while (nr_pageblocks--) {
2601 set_pageblock_migratetype(pageblock_page, migratetype);
2602 pageblock_page += pageblock_nr_pages;
2607 * When we are falling back to another migratetype during allocation, try to
2608 * steal extra free pages from the same pageblocks to satisfy further
2609 * allocations, instead of polluting multiple pageblocks.
2611 * If we are stealing a relatively large buddy page, it is likely there will
2612 * be more free pages in the pageblock, so try to steal them all. For
2613 * reclaimable and unmovable allocations, we steal regardless of page size,
2614 * as fragmentation caused by those allocations polluting movable pageblocks
2615 * is worse than movable allocations stealing from unmovable and reclaimable
2618 static bool can_steal_fallback(unsigned int order, int start_mt)
2621 * Leaving this order check is intended, although there is
2622 * relaxed order check in next check. The reason is that
2623 * we can actually steal whole pageblock if this condition met,
2624 * but, below check doesn't guarantee it and that is just heuristic
2625 * so could be changed anytime.
2627 if (order >= pageblock_order)
2630 if (order >= pageblock_order / 2 ||
2631 start_mt == MIGRATE_RECLAIMABLE ||
2632 start_mt == MIGRATE_UNMOVABLE ||
2633 page_group_by_mobility_disabled)
2639 static inline bool boost_watermark(struct zone *zone)
2641 unsigned long max_boost;
2643 if (!watermark_boost_factor)
2646 * Don't bother in zones that are unlikely to produce results.
2647 * On small machines, including kdump capture kernels running
2648 * in a small area, boosting the watermark can cause an out of
2649 * memory situation immediately.
2651 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2654 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2655 watermark_boost_factor, 10000);
2658 * high watermark may be uninitialised if fragmentation occurs
2659 * very early in boot so do not boost. We do not fall
2660 * through and boost by pageblock_nr_pages as failing
2661 * allocations that early means that reclaim is not going
2662 * to help and it may even be impossible to reclaim the
2663 * boosted watermark resulting in a hang.
2668 max_boost = max(pageblock_nr_pages, max_boost);
2670 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2677 * This function implements actual steal behaviour. If order is large enough,
2678 * we can steal whole pageblock. If not, we first move freepages in this
2679 * pageblock to our migratetype and determine how many already-allocated pages
2680 * are there in the pageblock with a compatible migratetype. If at least half
2681 * of pages are free or compatible, we can change migratetype of the pageblock
2682 * itself, so pages freed in the future will be put on the correct free list.
2684 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2685 unsigned int alloc_flags, int start_type, bool whole_block)
2687 unsigned int current_order = buddy_order(page);
2688 int free_pages, movable_pages, alike_pages;
2691 old_block_type = get_pageblock_migratetype(page);
2694 * This can happen due to races and we want to prevent broken
2695 * highatomic accounting.
2697 if (is_migrate_highatomic(old_block_type))
2700 /* Take ownership for orders >= pageblock_order */
2701 if (current_order >= pageblock_order) {
2702 change_pageblock_range(page, current_order, start_type);
2707 * Boost watermarks to increase reclaim pressure to reduce the
2708 * likelihood of future fallbacks. Wake kswapd now as the node
2709 * may be balanced overall and kswapd will not wake naturally.
2711 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2712 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2714 /* We are not allowed to try stealing from the whole block */
2718 free_pages = move_freepages_block(zone, page, start_type,
2721 * Determine how many pages are compatible with our allocation.
2722 * For movable allocation, it's the number of movable pages which
2723 * we just obtained. For other types it's a bit more tricky.
2725 if (start_type == MIGRATE_MOVABLE) {
2726 alike_pages = movable_pages;
2729 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2730 * to MOVABLE pageblock, consider all non-movable pages as
2731 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2732 * vice versa, be conservative since we can't distinguish the
2733 * exact migratetype of non-movable pages.
2735 if (old_block_type == MIGRATE_MOVABLE)
2736 alike_pages = pageblock_nr_pages
2737 - (free_pages + movable_pages);
2742 /* moving whole block can fail due to zone boundary conditions */
2747 * If a sufficient number of pages in the block are either free or of
2748 * comparable migratability as our allocation, claim the whole block.
2750 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2751 page_group_by_mobility_disabled)
2752 set_pageblock_migratetype(page, start_type);
2757 move_to_free_list(page, zone, current_order, start_type);
2761 * Check whether there is a suitable fallback freepage with requested order.
2762 * If only_stealable is true, this function returns fallback_mt only if
2763 * we can steal other freepages all together. This would help to reduce
2764 * fragmentation due to mixed migratetype pages in one pageblock.
2766 int find_suitable_fallback(struct free_area *area, unsigned int order,
2767 int migratetype, bool only_stealable, bool *can_steal)
2772 if (area->nr_free == 0)
2777 fallback_mt = fallbacks[migratetype][i];
2778 if (fallback_mt == MIGRATE_TYPES)
2781 if (free_area_empty(area, fallback_mt))
2784 if (can_steal_fallback(order, migratetype))
2787 if (!only_stealable)
2798 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2799 * there are no empty page blocks that contain a page with a suitable order
2801 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2802 unsigned int alloc_order)
2805 unsigned long max_managed, flags;
2808 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2809 * Check is race-prone but harmless.
2811 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2812 if (zone->nr_reserved_highatomic >= max_managed)
2815 spin_lock_irqsave(&zone->lock, flags);
2817 /* Recheck the nr_reserved_highatomic limit under the lock */
2818 if (zone->nr_reserved_highatomic >= max_managed)
2822 mt = get_pageblock_migratetype(page);
2823 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2824 if (migratetype_is_mergeable(mt)) {
2825 zone->nr_reserved_highatomic += pageblock_nr_pages;
2826 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2827 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2831 spin_unlock_irqrestore(&zone->lock, flags);
2835 * Used when an allocation is about to fail under memory pressure. This
2836 * potentially hurts the reliability of high-order allocations when under
2837 * intense memory pressure but failed atomic allocations should be easier
2838 * to recover from than an OOM.
2840 * If @force is true, try to unreserve a pageblock even though highatomic
2841 * pageblock is exhausted.
2843 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2846 struct zonelist *zonelist = ac->zonelist;
2847 unsigned long flags;
2854 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2857 * Preserve at least one pageblock unless memory pressure
2860 if (!force && zone->nr_reserved_highatomic <=
2864 spin_lock_irqsave(&zone->lock, flags);
2865 for (order = 0; order < MAX_ORDER; order++) {
2866 struct free_area *area = &(zone->free_area[order]);
2868 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2873 * In page freeing path, migratetype change is racy so
2874 * we can counter several free pages in a pageblock
2875 * in this loop although we changed the pageblock type
2876 * from highatomic to ac->migratetype. So we should
2877 * adjust the count once.
2879 if (is_migrate_highatomic_page(page)) {
2881 * It should never happen but changes to
2882 * locking could inadvertently allow a per-cpu
2883 * drain to add pages to MIGRATE_HIGHATOMIC
2884 * while unreserving so be safe and watch for
2887 zone->nr_reserved_highatomic -= min(
2889 zone->nr_reserved_highatomic);
2893 * Convert to ac->migratetype and avoid the normal
2894 * pageblock stealing heuristics. Minimally, the caller
2895 * is doing the work and needs the pages. More
2896 * importantly, if the block was always converted to
2897 * MIGRATE_UNMOVABLE or another type then the number
2898 * of pageblocks that cannot be completely freed
2901 set_pageblock_migratetype(page, ac->migratetype);
2902 ret = move_freepages_block(zone, page, ac->migratetype,
2905 spin_unlock_irqrestore(&zone->lock, flags);
2909 spin_unlock_irqrestore(&zone->lock, flags);
2916 * Try finding a free buddy page on the fallback list and put it on the free
2917 * list of requested migratetype, possibly along with other pages from the same
2918 * block, depending on fragmentation avoidance heuristics. Returns true if
2919 * fallback was found so that __rmqueue_smallest() can grab it.
2921 * The use of signed ints for order and current_order is a deliberate
2922 * deviation from the rest of this file, to make the for loop
2923 * condition simpler.
2925 static __always_inline bool
2926 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2927 unsigned int alloc_flags)
2929 struct free_area *area;
2931 int min_order = order;
2937 * Do not steal pages from freelists belonging to other pageblocks
2938 * i.e. orders < pageblock_order. If there are no local zones free,
2939 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2941 if (alloc_flags & ALLOC_NOFRAGMENT)
2942 min_order = pageblock_order;
2945 * Find the largest available free page in the other list. This roughly
2946 * approximates finding the pageblock with the most free pages, which
2947 * would be too costly to do exactly.
2949 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2951 area = &(zone->free_area[current_order]);
2952 fallback_mt = find_suitable_fallback(area, current_order,
2953 start_migratetype, false, &can_steal);
2954 if (fallback_mt == -1)
2958 * We cannot steal all free pages from the pageblock and the
2959 * requested migratetype is movable. In that case it's better to
2960 * steal and split the smallest available page instead of the
2961 * largest available page, because even if the next movable
2962 * allocation falls back into a different pageblock than this
2963 * one, it won't cause permanent fragmentation.
2965 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2966 && current_order > order)
2975 for (current_order = order; current_order < MAX_ORDER;
2977 area = &(zone->free_area[current_order]);
2978 fallback_mt = find_suitable_fallback(area, current_order,
2979 start_migratetype, false, &can_steal);
2980 if (fallback_mt != -1)
2985 * This should not happen - we already found a suitable fallback
2986 * when looking for the largest page.
2988 VM_BUG_ON(current_order == MAX_ORDER);
2991 page = get_page_from_free_area(area, fallback_mt);
2993 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2996 trace_mm_page_alloc_extfrag(page, order, current_order,
2997 start_migratetype, fallback_mt);
3004 * Do the hard work of removing an element from the buddy allocator.
3005 * Call me with the zone->lock already held.
3007 static __always_inline struct page *
3008 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
3009 unsigned int alloc_flags)
3013 if (IS_ENABLED(CONFIG_CMA)) {
3015 * Balance movable allocations between regular and CMA areas by
3016 * allocating from CMA when over half of the zone's free memory
3017 * is in the CMA area.
3019 if (alloc_flags & ALLOC_CMA &&
3020 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3021 zone_page_state(zone, NR_FREE_PAGES) / 2) {
3022 page = __rmqueue_cma_fallback(zone, order);
3028 page = __rmqueue_smallest(zone, order, migratetype);
3029 if (unlikely(!page)) {
3030 if (alloc_flags & ALLOC_CMA)
3031 page = __rmqueue_cma_fallback(zone, order);
3033 if (!page && __rmqueue_fallback(zone, order, migratetype,
3041 * Obtain a specified number of elements from the buddy allocator, all under
3042 * a single hold of the lock, for efficiency. Add them to the supplied list.
3043 * Returns the number of new pages which were placed at *list.
3045 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3046 unsigned long count, struct list_head *list,
3047 int migratetype, unsigned int alloc_flags)
3049 int i, allocated = 0;
3052 * local_lock_irq held so equivalent to spin_lock_irqsave for
3053 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3055 spin_lock(&zone->lock);
3056 for (i = 0; i < count; ++i) {
3057 struct page *page = __rmqueue(zone, order, migratetype,
3059 if (unlikely(page == NULL))
3062 if (unlikely(check_pcp_refill(page, order)))
3066 * Split buddy pages returned by expand() are received here in
3067 * physical page order. The page is added to the tail of
3068 * caller's list. From the callers perspective, the linked list
3069 * is ordered by page number under some conditions. This is
3070 * useful for IO devices that can forward direction from the
3071 * head, thus also in the physical page order. This is useful
3072 * for IO devices that can merge IO requests if the physical
3073 * pages are ordered properly.
3075 list_add_tail(&page->lru, list);
3077 if (is_migrate_cma(get_pcppage_migratetype(page)))
3078 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3083 * i pages were removed from the buddy list even if some leak due
3084 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3085 * on i. Do not confuse with 'allocated' which is the number of
3086 * pages added to the pcp list.
3088 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3089 spin_unlock(&zone->lock);
3095 * Called from the vmstat counter updater to drain pagesets of this
3096 * currently executing processor on remote nodes after they have
3099 * Note that this function must be called with the thread pinned to
3100 * a single processor.
3102 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3104 unsigned long flags;
3105 int to_drain, batch;
3107 local_lock_irqsave(&pagesets.lock, flags);
3108 batch = READ_ONCE(pcp->batch);
3109 to_drain = min(pcp->count, batch);
3111 free_pcppages_bulk(zone, to_drain, pcp, 0);
3112 local_unlock_irqrestore(&pagesets.lock, flags);
3117 * Drain pcplists of the indicated processor and zone.
3119 * The processor must either be the current processor and the
3120 * thread pinned to the current processor or a processor that
3123 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3125 unsigned long flags;
3126 struct per_cpu_pages *pcp;
3128 local_lock_irqsave(&pagesets.lock, flags);
3130 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3132 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3134 local_unlock_irqrestore(&pagesets.lock, flags);
3138 * Drain pcplists of all zones on the indicated processor.
3140 * The processor must either be the current processor and the
3141 * thread pinned to the current processor or a processor that
3144 static void drain_pages(unsigned int cpu)
3148 for_each_populated_zone(zone) {
3149 drain_pages_zone(cpu, zone);
3154 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3156 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3157 * the single zone's pages.
3159 void drain_local_pages(struct zone *zone)
3161 int cpu = smp_processor_id();
3164 drain_pages_zone(cpu, zone);
3169 static void drain_local_pages_wq(struct work_struct *work)
3171 struct pcpu_drain *drain;
3173 drain = container_of(work, struct pcpu_drain, work);
3176 * drain_all_pages doesn't use proper cpu hotplug protection so
3177 * we can race with cpu offline when the WQ can move this from
3178 * a cpu pinned worker to an unbound one. We can operate on a different
3179 * cpu which is alright but we also have to make sure to not move to
3183 drain_local_pages(drain->zone);
3188 * The implementation of drain_all_pages(), exposing an extra parameter to
3189 * drain on all cpus.
3191 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3192 * not empty. The check for non-emptiness can however race with a free to
3193 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3194 * that need the guarantee that every CPU has drained can disable the
3195 * optimizing racy check.
3197 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3202 * Allocate in the BSS so we won't require allocation in
3203 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3205 static cpumask_t cpus_with_pcps;
3208 * Make sure nobody triggers this path before mm_percpu_wq is fully
3211 if (WARN_ON_ONCE(!mm_percpu_wq))
3215 * Do not drain if one is already in progress unless it's specific to
3216 * a zone. Such callers are primarily CMA and memory hotplug and need
3217 * the drain to be complete when the call returns.
3219 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3222 mutex_lock(&pcpu_drain_mutex);
3226 * We don't care about racing with CPU hotplug event
3227 * as offline notification will cause the notified
3228 * cpu to drain that CPU pcps and on_each_cpu_mask
3229 * disables preemption as part of its processing
3231 for_each_online_cpu(cpu) {
3232 struct per_cpu_pages *pcp;
3234 bool has_pcps = false;
3236 if (force_all_cpus) {
3238 * The pcp.count check is racy, some callers need a
3239 * guarantee that no cpu is missed.
3243 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3247 for_each_populated_zone(z) {
3248 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3257 cpumask_set_cpu(cpu, &cpus_with_pcps);
3259 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3262 for_each_cpu(cpu, &cpus_with_pcps) {
3263 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3266 INIT_WORK(&drain->work, drain_local_pages_wq);
3267 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3269 for_each_cpu(cpu, &cpus_with_pcps)
3270 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3272 mutex_unlock(&pcpu_drain_mutex);
3276 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3278 * When zone parameter is non-NULL, spill just the single zone's pages.
3280 * Note that this can be extremely slow as the draining happens in a workqueue.
3282 void drain_all_pages(struct zone *zone)
3284 __drain_all_pages(zone, false);
3287 #ifdef CONFIG_HIBERNATION
3290 * Touch the watchdog for every WD_PAGE_COUNT pages.
3292 #define WD_PAGE_COUNT (128*1024)
3294 void mark_free_pages(struct zone *zone)
3296 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3297 unsigned long flags;
3298 unsigned int order, t;
3301 if (zone_is_empty(zone))
3304 spin_lock_irqsave(&zone->lock, flags);
3306 max_zone_pfn = zone_end_pfn(zone);
3307 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3308 if (pfn_valid(pfn)) {
3309 page = pfn_to_page(pfn);
3311 if (!--page_count) {
3312 touch_nmi_watchdog();
3313 page_count = WD_PAGE_COUNT;
3316 if (page_zone(page) != zone)
3319 if (!swsusp_page_is_forbidden(page))
3320 swsusp_unset_page_free(page);
3323 for_each_migratetype_order(order, t) {
3324 list_for_each_entry(page,
3325 &zone->free_area[order].free_list[t], lru) {
3328 pfn = page_to_pfn(page);
3329 for (i = 0; i < (1UL << order); i++) {
3330 if (!--page_count) {
3331 touch_nmi_watchdog();
3332 page_count = WD_PAGE_COUNT;
3334 swsusp_set_page_free(pfn_to_page(pfn + i));
3338 spin_unlock_irqrestore(&zone->lock, flags);
3340 #endif /* CONFIG_PM */
3342 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3347 if (!free_pcp_prepare(page, order))
3350 migratetype = get_pfnblock_migratetype(page, pfn);
3351 set_pcppage_migratetype(page, migratetype);
3355 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3358 int min_nr_free, max_nr_free;
3360 /* Free everything if batch freeing high-order pages. */
3361 if (unlikely(free_high))
3364 /* Check for PCP disabled or boot pageset */
3365 if (unlikely(high < batch))
3368 /* Leave at least pcp->batch pages on the list */
3369 min_nr_free = batch;
3370 max_nr_free = high - batch;
3373 * Double the number of pages freed each time there is subsequent
3374 * freeing of pages without any allocation.
3376 batch <<= pcp->free_factor;
3377 if (batch < max_nr_free)
3379 batch = clamp(batch, min_nr_free, max_nr_free);
3384 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3387 int high = READ_ONCE(pcp->high);
3389 if (unlikely(!high || free_high))
3392 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3396 * If reclaim is active, limit the number of pages that can be
3397 * stored on pcp lists
3399 return min(READ_ONCE(pcp->batch) << 2, high);
3402 static void free_unref_page_commit(struct page *page, int migratetype,
3405 struct zone *zone = page_zone(page);
3406 struct per_cpu_pages *pcp;
3411 __count_vm_event(PGFREE);
3412 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3413 pindex = order_to_pindex(migratetype, order);
3414 list_add(&page->lru, &pcp->lists[pindex]);
3415 pcp->count += 1 << order;
3418 * As high-order pages other than THP's stored on PCP can contribute
3419 * to fragmentation, limit the number stored when PCP is heavily
3420 * freeing without allocation. The remainder after bulk freeing
3421 * stops will be drained from vmstat refresh context.
3423 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3425 high = nr_pcp_high(pcp, zone, free_high);
3426 if (pcp->count >= high) {
3427 int batch = READ_ONCE(pcp->batch);
3429 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3436 void free_unref_page(struct page *page, unsigned int order)
3438 unsigned long flags;
3439 unsigned long pfn = page_to_pfn(page);
3442 if (!free_unref_page_prepare(page, pfn, order))
3446 * We only track unmovable, reclaimable and movable on pcp lists.
3447 * Place ISOLATE pages on the isolated list because they are being
3448 * offlined but treat HIGHATOMIC as movable pages so we can get those
3449 * areas back if necessary. Otherwise, we may have to free
3450 * excessively into the page allocator
3452 migratetype = get_pcppage_migratetype(page);
3453 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3454 if (unlikely(is_migrate_isolate(migratetype))) {
3455 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3458 migratetype = MIGRATE_MOVABLE;
3461 local_lock_irqsave(&pagesets.lock, flags);
3462 free_unref_page_commit(page, migratetype, order);
3463 local_unlock_irqrestore(&pagesets.lock, flags);
3467 * Free a list of 0-order pages
3469 void free_unref_page_list(struct list_head *list)
3471 struct page *page, *next;
3472 unsigned long flags;
3473 int batch_count = 0;
3476 /* Prepare pages for freeing */
3477 list_for_each_entry_safe(page, next, list, lru) {
3478 unsigned long pfn = page_to_pfn(page);
3479 if (!free_unref_page_prepare(page, pfn, 0)) {
3480 list_del(&page->lru);
3485 * Free isolated pages directly to the allocator, see
3486 * comment in free_unref_page.
3488 migratetype = get_pcppage_migratetype(page);
3489 if (unlikely(is_migrate_isolate(migratetype))) {
3490 list_del(&page->lru);
3491 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3496 local_lock_irqsave(&pagesets.lock, flags);
3497 list_for_each_entry_safe(page, next, list, lru) {
3499 * Non-isolated types over MIGRATE_PCPTYPES get added
3500 * to the MIGRATE_MOVABLE pcp list.
3502 migratetype = get_pcppage_migratetype(page);
3503 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3504 migratetype = MIGRATE_MOVABLE;
3506 trace_mm_page_free_batched(page);
3507 free_unref_page_commit(page, migratetype, 0);
3510 * Guard against excessive IRQ disabled times when we get
3511 * a large list of pages to free.
3513 if (++batch_count == SWAP_CLUSTER_MAX) {
3514 local_unlock_irqrestore(&pagesets.lock, flags);
3516 local_lock_irqsave(&pagesets.lock, flags);
3519 local_unlock_irqrestore(&pagesets.lock, flags);
3523 * split_page takes a non-compound higher-order page, and splits it into
3524 * n (1<<order) sub-pages: page[0..n]
3525 * Each sub-page must be freed individually.
3527 * Note: this is probably too low level an operation for use in drivers.
3528 * Please consult with lkml before using this in your driver.
3530 void split_page(struct page *page, unsigned int order)
3534 VM_BUG_ON_PAGE(PageCompound(page), page);
3535 VM_BUG_ON_PAGE(!page_count(page), page);
3537 for (i = 1; i < (1 << order); i++)
3538 set_page_refcounted(page + i);
3539 split_page_owner(page, 1 << order);
3540 split_page_memcg(page, 1 << order);
3542 EXPORT_SYMBOL_GPL(split_page);
3544 int __isolate_free_page(struct page *page, unsigned int order)
3546 unsigned long watermark;
3550 BUG_ON(!PageBuddy(page));
3552 zone = page_zone(page);
3553 mt = get_pageblock_migratetype(page);
3555 if (!is_migrate_isolate(mt)) {
3557 * Obey watermarks as if the page was being allocated. We can
3558 * emulate a high-order watermark check with a raised order-0
3559 * watermark, because we already know our high-order page
3562 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3563 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3566 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3569 /* Remove page from free list */
3571 del_page_from_free_list(page, zone, order);
3574 * Set the pageblock if the isolated page is at least half of a
3577 if (order >= pageblock_order - 1) {
3578 struct page *endpage = page + (1 << order) - 1;
3579 for (; page < endpage; page += pageblock_nr_pages) {
3580 int mt = get_pageblock_migratetype(page);
3582 * Only change normal pageblocks (i.e., they can merge
3585 if (migratetype_is_mergeable(mt))
3586 set_pageblock_migratetype(page,
3592 return 1UL << order;
3596 * __putback_isolated_page - Return a now-isolated page back where we got it
3597 * @page: Page that was isolated
3598 * @order: Order of the isolated page
3599 * @mt: The page's pageblock's migratetype
3601 * This function is meant to return a page pulled from the free lists via
3602 * __isolate_free_page back to the free lists they were pulled from.
3604 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3606 struct zone *zone = page_zone(page);
3608 /* zone lock should be held when this function is called */
3609 lockdep_assert_held(&zone->lock);
3611 /* Return isolated page to tail of freelist. */
3612 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3613 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3617 * Update NUMA hit/miss statistics
3619 * Must be called with interrupts disabled.
3621 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3625 enum numa_stat_item local_stat = NUMA_LOCAL;
3627 /* skip numa counters update if numa stats is disabled */
3628 if (!static_branch_likely(&vm_numa_stat_key))
3631 if (zone_to_nid(z) != numa_node_id())
3632 local_stat = NUMA_OTHER;
3634 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3635 __count_numa_events(z, NUMA_HIT, nr_account);
3637 __count_numa_events(z, NUMA_MISS, nr_account);
3638 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3640 __count_numa_events(z, local_stat, nr_account);
3644 /* Remove page from the per-cpu list, caller must protect the list */
3646 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3648 unsigned int alloc_flags,
3649 struct per_cpu_pages *pcp,
3650 struct list_head *list)
3655 if (list_empty(list)) {
3656 int batch = READ_ONCE(pcp->batch);
3660 * Scale batch relative to order if batch implies
3661 * free pages can be stored on the PCP. Batch can
3662 * be 1 for small zones or for boot pagesets which
3663 * should never store free pages as the pages may
3664 * belong to arbitrary zones.
3667 batch = max(batch >> order, 2);
3668 alloced = rmqueue_bulk(zone, order,
3670 migratetype, alloc_flags);
3672 pcp->count += alloced << order;
3673 if (unlikely(list_empty(list)))
3677 page = list_first_entry(list, struct page, lru);
3678 list_del(&page->lru);
3679 pcp->count -= 1 << order;
3680 } while (check_new_pcp(page, order));
3685 /* Lock and remove page from the per-cpu list */
3686 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3687 struct zone *zone, unsigned int order,
3688 gfp_t gfp_flags, int migratetype,
3689 unsigned int alloc_flags)
3691 struct per_cpu_pages *pcp;
3692 struct list_head *list;
3694 unsigned long flags;
3696 local_lock_irqsave(&pagesets.lock, flags);
3699 * On allocation, reduce the number of pages that are batch freed.
3700 * See nr_pcp_free() where free_factor is increased for subsequent
3703 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3704 pcp->free_factor >>= 1;
3705 list = &pcp->lists[order_to_pindex(migratetype, order)];
3706 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3707 local_unlock_irqrestore(&pagesets.lock, flags);
3709 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3710 zone_statistics(preferred_zone, zone, 1);
3716 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3719 struct page *rmqueue(struct zone *preferred_zone,
3720 struct zone *zone, unsigned int order,
3721 gfp_t gfp_flags, unsigned int alloc_flags,
3724 unsigned long flags;
3727 if (likely(pcp_allowed_order(order))) {
3729 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3730 * we need to skip it when CMA area isn't allowed.
3732 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3733 migratetype != MIGRATE_MOVABLE) {
3734 page = rmqueue_pcplist(preferred_zone, zone, order,
3735 gfp_flags, migratetype, alloc_flags);
3741 * We most definitely don't want callers attempting to
3742 * allocate greater than order-1 page units with __GFP_NOFAIL.
3744 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3748 spin_lock_irqsave(&zone->lock, flags);
3750 * order-0 request can reach here when the pcplist is skipped
3751 * due to non-CMA allocation context. HIGHATOMIC area is
3752 * reserved for high-order atomic allocation, so order-0
3753 * request should skip it.
3755 if (order > 0 && alloc_flags & ALLOC_HARDER)
3756 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3758 page = __rmqueue(zone, order, migratetype, alloc_flags);
3762 __mod_zone_freepage_state(zone, -(1 << order),
3763 get_pcppage_migratetype(page));
3764 spin_unlock_irqrestore(&zone->lock, flags);
3765 } while (check_new_pages(page, order));
3767 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3768 zone_statistics(preferred_zone, zone, 1);
3771 /* Separate test+clear to avoid unnecessary atomics */
3772 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3773 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3774 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3777 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3781 spin_unlock_irqrestore(&zone->lock, flags);
3785 #ifdef CONFIG_FAIL_PAGE_ALLOC
3788 struct fault_attr attr;
3790 bool ignore_gfp_highmem;
3791 bool ignore_gfp_reclaim;
3793 } fail_page_alloc = {
3794 .attr = FAULT_ATTR_INITIALIZER,
3795 .ignore_gfp_reclaim = true,
3796 .ignore_gfp_highmem = true,
3800 static int __init setup_fail_page_alloc(char *str)
3802 return setup_fault_attr(&fail_page_alloc.attr, str);
3804 __setup("fail_page_alloc=", setup_fail_page_alloc);
3806 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3808 if (order < fail_page_alloc.min_order)
3810 if (gfp_mask & __GFP_NOFAIL)
3812 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3814 if (fail_page_alloc.ignore_gfp_reclaim &&
3815 (gfp_mask & __GFP_DIRECT_RECLAIM))
3818 if (gfp_mask & __GFP_NOWARN)
3819 fail_page_alloc.attr.no_warn = true;
3821 return should_fail(&fail_page_alloc.attr, 1 << order);
3824 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3826 static int __init fail_page_alloc_debugfs(void)
3828 umode_t mode = S_IFREG | 0600;
3831 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3832 &fail_page_alloc.attr);
3834 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3835 &fail_page_alloc.ignore_gfp_reclaim);
3836 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3837 &fail_page_alloc.ignore_gfp_highmem);
3838 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3843 late_initcall(fail_page_alloc_debugfs);
3845 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3847 #else /* CONFIG_FAIL_PAGE_ALLOC */
3849 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3854 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3856 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3858 return __should_fail_alloc_page(gfp_mask, order);
3860 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3862 static inline long __zone_watermark_unusable_free(struct zone *z,
3863 unsigned int order, unsigned int alloc_flags)
3865 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3866 long unusable_free = (1 << order) - 1;
3869 * If the caller does not have rights to ALLOC_HARDER then subtract
3870 * the high-atomic reserves. This will over-estimate the size of the
3871 * atomic reserve but it avoids a search.
3873 if (likely(!alloc_harder))
3874 unusable_free += z->nr_reserved_highatomic;
3877 /* If allocation can't use CMA areas don't use free CMA pages */
3878 if (!(alloc_flags & ALLOC_CMA))
3879 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3882 return unusable_free;
3886 * Return true if free base pages are above 'mark'. For high-order checks it
3887 * will return true of the order-0 watermark is reached and there is at least
3888 * one free page of a suitable size. Checking now avoids taking the zone lock
3889 * to check in the allocation paths if no pages are free.
3891 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3892 int highest_zoneidx, unsigned int alloc_flags,
3897 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3899 /* free_pages may go negative - that's OK */
3900 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3902 if (alloc_flags & ALLOC_HIGH)
3905 if (unlikely(alloc_harder)) {
3907 * OOM victims can try even harder than normal ALLOC_HARDER
3908 * users on the grounds that it's definitely going to be in
3909 * the exit path shortly and free memory. Any allocation it
3910 * makes during the free path will be small and short-lived.
3912 if (alloc_flags & ALLOC_OOM)
3919 * Check watermarks for an order-0 allocation request. If these
3920 * are not met, then a high-order request also cannot go ahead
3921 * even if a suitable page happened to be free.
3923 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3926 /* If this is an order-0 request then the watermark is fine */
3930 /* For a high-order request, check at least one suitable page is free */
3931 for (o = order; o < MAX_ORDER; o++) {
3932 struct free_area *area = &z->free_area[o];
3938 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3939 if (!free_area_empty(area, mt))
3944 if ((alloc_flags & ALLOC_CMA) &&
3945 !free_area_empty(area, MIGRATE_CMA)) {
3949 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3955 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3956 int highest_zoneidx, unsigned int alloc_flags)
3958 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3959 zone_page_state(z, NR_FREE_PAGES));
3962 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3963 unsigned long mark, int highest_zoneidx,
3964 unsigned int alloc_flags, gfp_t gfp_mask)
3968 free_pages = zone_page_state(z, NR_FREE_PAGES);
3971 * Fast check for order-0 only. If this fails then the reserves
3972 * need to be calculated.
3977 fast_free = free_pages;
3978 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3979 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3983 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3987 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3988 * when checking the min watermark. The min watermark is the
3989 * point where boosting is ignored so that kswapd is woken up
3990 * when below the low watermark.
3992 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3993 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3994 mark = z->_watermark[WMARK_MIN];
3995 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3996 alloc_flags, free_pages);
4002 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
4003 unsigned long mark, int highest_zoneidx)
4005 long free_pages = zone_page_state(z, NR_FREE_PAGES);
4007 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
4008 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
4010 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
4015 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
4017 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4019 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
4020 node_reclaim_distance;
4022 #else /* CONFIG_NUMA */
4023 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4027 #endif /* CONFIG_NUMA */
4030 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4031 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4032 * premature use of a lower zone may cause lowmem pressure problems that
4033 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4034 * probably too small. It only makes sense to spread allocations to avoid
4035 * fragmentation between the Normal and DMA32 zones.
4037 static inline unsigned int
4038 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4040 unsigned int alloc_flags;
4043 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4046 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4048 #ifdef CONFIG_ZONE_DMA32
4052 if (zone_idx(zone) != ZONE_NORMAL)
4056 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4057 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4058 * on UMA that if Normal is populated then so is DMA32.
4060 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4061 if (nr_online_nodes > 1 && !populated_zone(--zone))
4064 alloc_flags |= ALLOC_NOFRAGMENT;
4065 #endif /* CONFIG_ZONE_DMA32 */
4069 /* Must be called after current_gfp_context() which can change gfp_mask */
4070 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4071 unsigned int alloc_flags)
4074 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4075 alloc_flags |= ALLOC_CMA;
4081 * get_page_from_freelist goes through the zonelist trying to allocate
4084 static struct page *
4085 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4086 const struct alloc_context *ac)
4090 struct pglist_data *last_pgdat = NULL;
4091 bool last_pgdat_dirty_ok = false;
4096 * Scan zonelist, looking for a zone with enough free.
4097 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4099 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4100 z = ac->preferred_zoneref;
4101 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4106 if (cpusets_enabled() &&
4107 (alloc_flags & ALLOC_CPUSET) &&
4108 !__cpuset_zone_allowed(zone, gfp_mask))
4111 * When allocating a page cache page for writing, we
4112 * want to get it from a node that is within its dirty
4113 * limit, such that no single node holds more than its
4114 * proportional share of globally allowed dirty pages.
4115 * The dirty limits take into account the node's
4116 * lowmem reserves and high watermark so that kswapd
4117 * should be able to balance it without having to
4118 * write pages from its LRU list.
4120 * XXX: For now, allow allocations to potentially
4121 * exceed the per-node dirty limit in the slowpath
4122 * (spread_dirty_pages unset) before going into reclaim,
4123 * which is important when on a NUMA setup the allowed
4124 * nodes are together not big enough to reach the
4125 * global limit. The proper fix for these situations
4126 * will require awareness of nodes in the
4127 * dirty-throttling and the flusher threads.
4129 if (ac->spread_dirty_pages) {
4130 if (last_pgdat != zone->zone_pgdat) {
4131 last_pgdat = zone->zone_pgdat;
4132 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
4135 if (!last_pgdat_dirty_ok)
4139 if (no_fallback && nr_online_nodes > 1 &&
4140 zone != ac->preferred_zoneref->zone) {
4144 * If moving to a remote node, retry but allow
4145 * fragmenting fallbacks. Locality is more important
4146 * than fragmentation avoidance.
4148 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4149 if (zone_to_nid(zone) != local_nid) {
4150 alloc_flags &= ~ALLOC_NOFRAGMENT;
4155 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4156 if (!zone_watermark_fast(zone, order, mark,
4157 ac->highest_zoneidx, alloc_flags,
4161 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4163 * Watermark failed for this zone, but see if we can
4164 * grow this zone if it contains deferred pages.
4166 if (static_branch_unlikely(&deferred_pages)) {
4167 if (_deferred_grow_zone(zone, order))
4171 /* Checked here to keep the fast path fast */
4172 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4173 if (alloc_flags & ALLOC_NO_WATERMARKS)
4176 if (!node_reclaim_enabled() ||
4177 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4180 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4182 case NODE_RECLAIM_NOSCAN:
4185 case NODE_RECLAIM_FULL:
4186 /* scanned but unreclaimable */
4189 /* did we reclaim enough */
4190 if (zone_watermark_ok(zone, order, mark,
4191 ac->highest_zoneidx, alloc_flags))
4199 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4200 gfp_mask, alloc_flags, ac->migratetype);
4202 prep_new_page(page, order, gfp_mask, alloc_flags);
4205 * If this is a high-order atomic allocation then check
4206 * if the pageblock should be reserved for the future
4208 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4209 reserve_highatomic_pageblock(page, zone, order);
4213 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4214 /* Try again if zone has deferred pages */
4215 if (static_branch_unlikely(&deferred_pages)) {
4216 if (_deferred_grow_zone(zone, order))
4224 * It's possible on a UMA machine to get through all zones that are
4225 * fragmented. If avoiding fragmentation, reset and try again.
4228 alloc_flags &= ~ALLOC_NOFRAGMENT;
4235 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4237 unsigned int filter = SHOW_MEM_FILTER_NODES;
4240 * This documents exceptions given to allocations in certain
4241 * contexts that are allowed to allocate outside current's set
4244 if (!(gfp_mask & __GFP_NOMEMALLOC))
4245 if (tsk_is_oom_victim(current) ||
4246 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4247 filter &= ~SHOW_MEM_FILTER_NODES;
4248 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4249 filter &= ~SHOW_MEM_FILTER_NODES;
4251 show_mem(filter, nodemask);
4254 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4256 struct va_format vaf;
4258 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4260 if ((gfp_mask & __GFP_NOWARN) ||
4261 !__ratelimit(&nopage_rs) ||
4262 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4265 va_start(args, fmt);
4268 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4269 current->comm, &vaf, gfp_mask, &gfp_mask,
4270 nodemask_pr_args(nodemask));
4273 cpuset_print_current_mems_allowed();
4276 warn_alloc_show_mem(gfp_mask, nodemask);
4279 static inline struct page *
4280 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4281 unsigned int alloc_flags,
4282 const struct alloc_context *ac)
4286 page = get_page_from_freelist(gfp_mask, order,
4287 alloc_flags|ALLOC_CPUSET, ac);
4289 * fallback to ignore cpuset restriction if our nodes
4293 page = get_page_from_freelist(gfp_mask, order,
4299 static inline struct page *
4300 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4301 const struct alloc_context *ac, unsigned long *did_some_progress)
4303 struct oom_control oc = {
4304 .zonelist = ac->zonelist,
4305 .nodemask = ac->nodemask,
4307 .gfp_mask = gfp_mask,
4312 *did_some_progress = 0;
4315 * Acquire the oom lock. If that fails, somebody else is
4316 * making progress for us.
4318 if (!mutex_trylock(&oom_lock)) {
4319 *did_some_progress = 1;
4320 schedule_timeout_uninterruptible(1);
4325 * Go through the zonelist yet one more time, keep very high watermark
4326 * here, this is only to catch a parallel oom killing, we must fail if
4327 * we're still under heavy pressure. But make sure that this reclaim
4328 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4329 * allocation which will never fail due to oom_lock already held.
4331 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4332 ~__GFP_DIRECT_RECLAIM, order,
4333 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4337 /* Coredumps can quickly deplete all memory reserves */
4338 if (current->flags & PF_DUMPCORE)
4340 /* The OOM killer will not help higher order allocs */
4341 if (order > PAGE_ALLOC_COSTLY_ORDER)
4344 * We have already exhausted all our reclaim opportunities without any
4345 * success so it is time to admit defeat. We will skip the OOM killer
4346 * because it is very likely that the caller has a more reasonable
4347 * fallback than shooting a random task.
4349 * The OOM killer may not free memory on a specific node.
4351 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4353 /* The OOM killer does not needlessly kill tasks for lowmem */
4354 if (ac->highest_zoneidx < ZONE_NORMAL)
4356 if (pm_suspended_storage())
4359 * XXX: GFP_NOFS allocations should rather fail than rely on
4360 * other request to make a forward progress.
4361 * We are in an unfortunate situation where out_of_memory cannot
4362 * do much for this context but let's try it to at least get
4363 * access to memory reserved if the current task is killed (see
4364 * out_of_memory). Once filesystems are ready to handle allocation
4365 * failures more gracefully we should just bail out here.
4368 /* Exhausted what can be done so it's blame time */
4369 if (out_of_memory(&oc) ||
4370 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4371 *did_some_progress = 1;
4374 * Help non-failing allocations by giving them access to memory
4377 if (gfp_mask & __GFP_NOFAIL)
4378 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4379 ALLOC_NO_WATERMARKS, ac);
4382 mutex_unlock(&oom_lock);
4387 * Maximum number of compaction retries with a progress before OOM
4388 * killer is consider as the only way to move forward.
4390 #define MAX_COMPACT_RETRIES 16
4392 #ifdef CONFIG_COMPACTION
4393 /* Try memory compaction for high-order allocations before reclaim */
4394 static struct page *
4395 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4396 unsigned int alloc_flags, const struct alloc_context *ac,
4397 enum compact_priority prio, enum compact_result *compact_result)
4399 struct page *page = NULL;
4400 unsigned long pflags;
4401 unsigned int noreclaim_flag;
4406 psi_memstall_enter(&pflags);
4407 delayacct_compact_start();
4408 noreclaim_flag = memalloc_noreclaim_save();
4410 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4413 memalloc_noreclaim_restore(noreclaim_flag);
4414 psi_memstall_leave(&pflags);
4415 delayacct_compact_end();
4417 if (*compact_result == COMPACT_SKIPPED)
4420 * At least in one zone compaction wasn't deferred or skipped, so let's
4421 * count a compaction stall
4423 count_vm_event(COMPACTSTALL);
4425 /* Prep a captured page if available */
4427 prep_new_page(page, order, gfp_mask, alloc_flags);
4429 /* Try get a page from the freelist if available */
4431 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4434 struct zone *zone = page_zone(page);
4436 zone->compact_blockskip_flush = false;
4437 compaction_defer_reset(zone, order, true);
4438 count_vm_event(COMPACTSUCCESS);
4443 * It's bad if compaction run occurs and fails. The most likely reason
4444 * is that pages exist, but not enough to satisfy watermarks.
4446 count_vm_event(COMPACTFAIL);
4454 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4455 enum compact_result compact_result,
4456 enum compact_priority *compact_priority,
4457 int *compaction_retries)
4459 int max_retries = MAX_COMPACT_RETRIES;
4462 int retries = *compaction_retries;
4463 enum compact_priority priority = *compact_priority;
4468 if (fatal_signal_pending(current))
4471 if (compaction_made_progress(compact_result))
4472 (*compaction_retries)++;
4475 * compaction considers all the zone as desperately out of memory
4476 * so it doesn't really make much sense to retry except when the
4477 * failure could be caused by insufficient priority
4479 if (compaction_failed(compact_result))
4480 goto check_priority;
4483 * compaction was skipped because there are not enough order-0 pages
4484 * to work with, so we retry only if it looks like reclaim can help.
4486 if (compaction_needs_reclaim(compact_result)) {
4487 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4492 * make sure the compaction wasn't deferred or didn't bail out early
4493 * due to locks contention before we declare that we should give up.
4494 * But the next retry should use a higher priority if allowed, so
4495 * we don't just keep bailing out endlessly.
4497 if (compaction_withdrawn(compact_result)) {
4498 goto check_priority;
4502 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4503 * costly ones because they are de facto nofail and invoke OOM
4504 * killer to move on while costly can fail and users are ready
4505 * to cope with that. 1/4 retries is rather arbitrary but we
4506 * would need much more detailed feedback from compaction to
4507 * make a better decision.
4509 if (order > PAGE_ALLOC_COSTLY_ORDER)
4511 if (*compaction_retries <= max_retries) {
4517 * Make sure there are attempts at the highest priority if we exhausted
4518 * all retries or failed at the lower priorities.
4521 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4522 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4524 if (*compact_priority > min_priority) {
4525 (*compact_priority)--;
4526 *compaction_retries = 0;
4530 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4534 static inline struct page *
4535 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4536 unsigned int alloc_flags, const struct alloc_context *ac,
4537 enum compact_priority prio, enum compact_result *compact_result)
4539 *compact_result = COMPACT_SKIPPED;
4544 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4545 enum compact_result compact_result,
4546 enum compact_priority *compact_priority,
4547 int *compaction_retries)
4552 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4556 * There are setups with compaction disabled which would prefer to loop
4557 * inside the allocator rather than hit the oom killer prematurely.
4558 * Let's give them a good hope and keep retrying while the order-0
4559 * watermarks are OK.
4561 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4562 ac->highest_zoneidx, ac->nodemask) {
4563 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4564 ac->highest_zoneidx, alloc_flags))
4569 #endif /* CONFIG_COMPACTION */
4571 #ifdef CONFIG_LOCKDEP
4572 static struct lockdep_map __fs_reclaim_map =
4573 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4575 static bool __need_reclaim(gfp_t gfp_mask)
4577 /* no reclaim without waiting on it */
4578 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4581 /* this guy won't enter reclaim */
4582 if (current->flags & PF_MEMALLOC)
4585 if (gfp_mask & __GFP_NOLOCKDEP)
4591 void __fs_reclaim_acquire(unsigned long ip)
4593 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4596 void __fs_reclaim_release(unsigned long ip)
4598 lock_release(&__fs_reclaim_map, ip);
4601 void fs_reclaim_acquire(gfp_t gfp_mask)
4603 gfp_mask = current_gfp_context(gfp_mask);
4605 if (__need_reclaim(gfp_mask)) {
4606 if (gfp_mask & __GFP_FS)
4607 __fs_reclaim_acquire(_RET_IP_);
4609 #ifdef CONFIG_MMU_NOTIFIER
4610 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4611 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4616 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4618 void fs_reclaim_release(gfp_t gfp_mask)
4620 gfp_mask = current_gfp_context(gfp_mask);
4622 if (__need_reclaim(gfp_mask)) {
4623 if (gfp_mask & __GFP_FS)
4624 __fs_reclaim_release(_RET_IP_);
4627 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4630 /* Perform direct synchronous page reclaim */
4631 static unsigned long
4632 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4633 const struct alloc_context *ac)
4635 unsigned int noreclaim_flag;
4636 unsigned long progress;
4640 /* We now go into synchronous reclaim */
4641 cpuset_memory_pressure_bump();
4642 fs_reclaim_acquire(gfp_mask);
4643 noreclaim_flag = memalloc_noreclaim_save();
4645 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4648 memalloc_noreclaim_restore(noreclaim_flag);
4649 fs_reclaim_release(gfp_mask);
4656 /* The really slow allocator path where we enter direct reclaim */
4657 static inline struct page *
4658 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4659 unsigned int alloc_flags, const struct alloc_context *ac,
4660 unsigned long *did_some_progress)
4662 struct page *page = NULL;
4663 unsigned long pflags;
4664 bool drained = false;
4666 psi_memstall_enter(&pflags);
4667 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4668 if (unlikely(!(*did_some_progress)))
4672 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4675 * If an allocation failed after direct reclaim, it could be because
4676 * pages are pinned on the per-cpu lists or in high alloc reserves.
4677 * Shrink them and try again
4679 if (!page && !drained) {
4680 unreserve_highatomic_pageblock(ac, false);
4681 drain_all_pages(NULL);
4686 psi_memstall_leave(&pflags);
4691 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4692 const struct alloc_context *ac)
4696 pg_data_t *last_pgdat = NULL;
4697 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4699 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4701 if (!managed_zone(zone))
4703 if (last_pgdat != zone->zone_pgdat) {
4704 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4705 last_pgdat = zone->zone_pgdat;
4710 static inline unsigned int
4711 gfp_to_alloc_flags(gfp_t gfp_mask)
4713 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4716 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4717 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4718 * to save two branches.
4720 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4721 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4724 * The caller may dip into page reserves a bit more if the caller
4725 * cannot run direct reclaim, or if the caller has realtime scheduling
4726 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4727 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4729 alloc_flags |= (__force int)
4730 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4732 if (gfp_mask & __GFP_ATOMIC) {
4734 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4735 * if it can't schedule.
4737 if (!(gfp_mask & __GFP_NOMEMALLOC))
4738 alloc_flags |= ALLOC_HARDER;
4740 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4741 * comment for __cpuset_node_allowed().
4743 alloc_flags &= ~ALLOC_CPUSET;
4744 } else if (unlikely(rt_task(current)) && in_task())
4745 alloc_flags |= ALLOC_HARDER;
4747 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4752 static bool oom_reserves_allowed(struct task_struct *tsk)
4754 if (!tsk_is_oom_victim(tsk))
4758 * !MMU doesn't have oom reaper so give access to memory reserves
4759 * only to the thread with TIF_MEMDIE set
4761 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4768 * Distinguish requests which really need access to full memory
4769 * reserves from oom victims which can live with a portion of it
4771 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4773 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4775 if (gfp_mask & __GFP_MEMALLOC)
4776 return ALLOC_NO_WATERMARKS;
4777 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4778 return ALLOC_NO_WATERMARKS;
4779 if (!in_interrupt()) {
4780 if (current->flags & PF_MEMALLOC)
4781 return ALLOC_NO_WATERMARKS;
4782 else if (oom_reserves_allowed(current))
4789 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4791 return !!__gfp_pfmemalloc_flags(gfp_mask);
4795 * Checks whether it makes sense to retry the reclaim to make a forward progress
4796 * for the given allocation request.
4798 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4799 * without success, or when we couldn't even meet the watermark if we
4800 * reclaimed all remaining pages on the LRU lists.
4802 * Returns true if a retry is viable or false to enter the oom path.
4805 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4806 struct alloc_context *ac, int alloc_flags,
4807 bool did_some_progress, int *no_progress_loops)
4814 * Costly allocations might have made a progress but this doesn't mean
4815 * their order will become available due to high fragmentation so
4816 * always increment the no progress counter for them
4818 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4819 *no_progress_loops = 0;
4821 (*no_progress_loops)++;
4824 * Make sure we converge to OOM if we cannot make any progress
4825 * several times in the row.
4827 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4828 /* Before OOM, exhaust highatomic_reserve */
4829 return unreserve_highatomic_pageblock(ac, true);
4833 * Keep reclaiming pages while there is a chance this will lead
4834 * somewhere. If none of the target zones can satisfy our allocation
4835 * request even if all reclaimable pages are considered then we are
4836 * screwed and have to go OOM.
4838 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4839 ac->highest_zoneidx, ac->nodemask) {
4840 unsigned long available;
4841 unsigned long reclaimable;
4842 unsigned long min_wmark = min_wmark_pages(zone);
4845 available = reclaimable = zone_reclaimable_pages(zone);
4846 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4849 * Would the allocation succeed if we reclaimed all
4850 * reclaimable pages?
4852 wmark = __zone_watermark_ok(zone, order, min_wmark,
4853 ac->highest_zoneidx, alloc_flags, available);
4854 trace_reclaim_retry_zone(z, order, reclaimable,
4855 available, min_wmark, *no_progress_loops, wmark);
4863 * Memory allocation/reclaim might be called from a WQ context and the
4864 * current implementation of the WQ concurrency control doesn't
4865 * recognize that a particular WQ is congested if the worker thread is
4866 * looping without ever sleeping. Therefore we have to do a short sleep
4867 * here rather than calling cond_resched().
4869 if (current->flags & PF_WQ_WORKER)
4870 schedule_timeout_uninterruptible(1);
4877 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4880 * It's possible that cpuset's mems_allowed and the nodemask from
4881 * mempolicy don't intersect. This should be normally dealt with by
4882 * policy_nodemask(), but it's possible to race with cpuset update in
4883 * such a way the check therein was true, and then it became false
4884 * before we got our cpuset_mems_cookie here.
4885 * This assumes that for all allocations, ac->nodemask can come only
4886 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4887 * when it does not intersect with the cpuset restrictions) or the
4888 * caller can deal with a violated nodemask.
4890 if (cpusets_enabled() && ac->nodemask &&
4891 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4892 ac->nodemask = NULL;
4897 * When updating a task's mems_allowed or mempolicy nodemask, it is
4898 * possible to race with parallel threads in such a way that our
4899 * allocation can fail while the mask is being updated. If we are about
4900 * to fail, check if the cpuset changed during allocation and if so,
4903 if (read_mems_allowed_retry(cpuset_mems_cookie))
4909 static inline struct page *
4910 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4911 struct alloc_context *ac)
4913 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4914 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4915 struct page *page = NULL;
4916 unsigned int alloc_flags;
4917 unsigned long did_some_progress;
4918 enum compact_priority compact_priority;
4919 enum compact_result compact_result;
4920 int compaction_retries;
4921 int no_progress_loops;
4922 unsigned int cpuset_mems_cookie;
4926 * We also sanity check to catch abuse of atomic reserves being used by
4927 * callers that are not in atomic context.
4929 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4930 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4931 gfp_mask &= ~__GFP_ATOMIC;
4934 compaction_retries = 0;
4935 no_progress_loops = 0;
4936 compact_priority = DEF_COMPACT_PRIORITY;
4937 cpuset_mems_cookie = read_mems_allowed_begin();
4940 * The fast path uses conservative alloc_flags to succeed only until
4941 * kswapd needs to be woken up, and to avoid the cost of setting up
4942 * alloc_flags precisely. So we do that now.
4944 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4947 * We need to recalculate the starting point for the zonelist iterator
4948 * because we might have used different nodemask in the fast path, or
4949 * there was a cpuset modification and we are retrying - otherwise we
4950 * could end up iterating over non-eligible zones endlessly.
4952 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4953 ac->highest_zoneidx, ac->nodemask);
4954 if (!ac->preferred_zoneref->zone)
4958 * Check for insane configurations where the cpuset doesn't contain
4959 * any suitable zone to satisfy the request - e.g. non-movable
4960 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4962 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4963 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4964 ac->highest_zoneidx,
4965 &cpuset_current_mems_allowed);
4970 if (alloc_flags & ALLOC_KSWAPD)
4971 wake_all_kswapds(order, gfp_mask, ac);
4974 * The adjusted alloc_flags might result in immediate success, so try
4977 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4982 * For costly allocations, try direct compaction first, as it's likely
4983 * that we have enough base pages and don't need to reclaim. For non-
4984 * movable high-order allocations, do that as well, as compaction will
4985 * try prevent permanent fragmentation by migrating from blocks of the
4987 * Don't try this for allocations that are allowed to ignore
4988 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4990 if (can_direct_reclaim &&
4992 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4993 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4994 page = __alloc_pages_direct_compact(gfp_mask, order,
4996 INIT_COMPACT_PRIORITY,
5002 * Checks for costly allocations with __GFP_NORETRY, which
5003 * includes some THP page fault allocations
5005 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
5007 * If allocating entire pageblock(s) and compaction
5008 * failed because all zones are below low watermarks
5009 * or is prohibited because it recently failed at this
5010 * order, fail immediately unless the allocator has
5011 * requested compaction and reclaim retry.
5014 * - potentially very expensive because zones are far
5015 * below their low watermarks or this is part of very
5016 * bursty high order allocations,
5017 * - not guaranteed to help because isolate_freepages()
5018 * may not iterate over freed pages as part of its
5020 * - unlikely to make entire pageblocks free on its
5023 if (compact_result == COMPACT_SKIPPED ||
5024 compact_result == COMPACT_DEFERRED)
5028 * Looks like reclaim/compaction is worth trying, but
5029 * sync compaction could be very expensive, so keep
5030 * using async compaction.
5032 compact_priority = INIT_COMPACT_PRIORITY;
5037 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5038 if (alloc_flags & ALLOC_KSWAPD)
5039 wake_all_kswapds(order, gfp_mask, ac);
5041 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5043 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
5046 * Reset the nodemask and zonelist iterators if memory policies can be
5047 * ignored. These allocations are high priority and system rather than
5050 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5051 ac->nodemask = NULL;
5052 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5053 ac->highest_zoneidx, ac->nodemask);
5056 /* Attempt with potentially adjusted zonelist and alloc_flags */
5057 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5061 /* Caller is not willing to reclaim, we can't balance anything */
5062 if (!can_direct_reclaim)
5065 /* Avoid recursion of direct reclaim */
5066 if (current->flags & PF_MEMALLOC)
5069 /* Try direct reclaim and then allocating */
5070 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5071 &did_some_progress);
5075 /* Try direct compaction and then allocating */
5076 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5077 compact_priority, &compact_result);
5081 /* Do not loop if specifically requested */
5082 if (gfp_mask & __GFP_NORETRY)
5086 * Do not retry costly high order allocations unless they are
5087 * __GFP_RETRY_MAYFAIL
5089 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5092 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5093 did_some_progress > 0, &no_progress_loops))
5097 * It doesn't make any sense to retry for the compaction if the order-0
5098 * reclaim is not able to make any progress because the current
5099 * implementation of the compaction depends on the sufficient amount
5100 * of free memory (see __compaction_suitable)
5102 if (did_some_progress > 0 &&
5103 should_compact_retry(ac, order, alloc_flags,
5104 compact_result, &compact_priority,
5105 &compaction_retries))
5109 /* Deal with possible cpuset update races before we start OOM killing */
5110 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5113 /* Reclaim has failed us, start killing things */
5114 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5118 /* Avoid allocations with no watermarks from looping endlessly */
5119 if (tsk_is_oom_victim(current) &&
5120 (alloc_flags & ALLOC_OOM ||
5121 (gfp_mask & __GFP_NOMEMALLOC)))
5124 /* Retry as long as the OOM killer is making progress */
5125 if (did_some_progress) {
5126 no_progress_loops = 0;
5131 /* Deal with possible cpuset update races before we fail */
5132 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5136 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5139 if (gfp_mask & __GFP_NOFAIL) {
5141 * All existing users of the __GFP_NOFAIL are blockable, so warn
5142 * of any new users that actually require GFP_NOWAIT
5144 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
5148 * PF_MEMALLOC request from this context is rather bizarre
5149 * because we cannot reclaim anything and only can loop waiting
5150 * for somebody to do a work for us
5152 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
5155 * non failing costly orders are a hard requirement which we
5156 * are not prepared for much so let's warn about these users
5157 * so that we can identify them and convert them to something
5160 WARN_ON_ONCE_GFP(order > PAGE_ALLOC_COSTLY_ORDER, gfp_mask);
5163 * Help non-failing allocations by giving them access to memory
5164 * reserves but do not use ALLOC_NO_WATERMARKS because this
5165 * could deplete whole memory reserves which would just make
5166 * the situation worse
5168 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5176 warn_alloc(gfp_mask, ac->nodemask,
5177 "page allocation failure: order:%u", order);
5182 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5183 int preferred_nid, nodemask_t *nodemask,
5184 struct alloc_context *ac, gfp_t *alloc_gfp,
5185 unsigned int *alloc_flags)
5187 ac->highest_zoneidx = gfp_zone(gfp_mask);
5188 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5189 ac->nodemask = nodemask;
5190 ac->migratetype = gfp_migratetype(gfp_mask);
5192 if (cpusets_enabled()) {
5193 *alloc_gfp |= __GFP_HARDWALL;
5195 * When we are in the interrupt context, it is irrelevant
5196 * to the current task context. It means that any node ok.
5198 if (in_task() && !ac->nodemask)
5199 ac->nodemask = &cpuset_current_mems_allowed;
5201 *alloc_flags |= ALLOC_CPUSET;
5204 might_alloc(gfp_mask);
5206 if (should_fail_alloc_page(gfp_mask, order))
5209 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5211 /* Dirty zone balancing only done in the fast path */
5212 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5215 * The preferred zone is used for statistics but crucially it is
5216 * also used as the starting point for the zonelist iterator. It
5217 * may get reset for allocations that ignore memory policies.
5219 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5220 ac->highest_zoneidx, ac->nodemask);
5226 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5227 * @gfp: GFP flags for the allocation
5228 * @preferred_nid: The preferred NUMA node ID to allocate from
5229 * @nodemask: Set of nodes to allocate from, may be NULL
5230 * @nr_pages: The number of pages desired on the list or array
5231 * @page_list: Optional list to store the allocated pages
5232 * @page_array: Optional array to store the pages
5234 * This is a batched version of the page allocator that attempts to
5235 * allocate nr_pages quickly. Pages are added to page_list if page_list
5236 * is not NULL, otherwise it is assumed that the page_array is valid.
5238 * For lists, nr_pages is the number of pages that should be allocated.
5240 * For arrays, only NULL elements are populated with pages and nr_pages
5241 * is the maximum number of pages that will be stored in the array.
5243 * Returns the number of pages on the list or array.
5245 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5246 nodemask_t *nodemask, int nr_pages,
5247 struct list_head *page_list,
5248 struct page **page_array)
5251 unsigned long flags;
5254 struct per_cpu_pages *pcp;
5255 struct list_head *pcp_list;
5256 struct alloc_context ac;
5258 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5259 int nr_populated = 0, nr_account = 0;
5262 * Skip populated array elements to determine if any pages need
5263 * to be allocated before disabling IRQs.
5265 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5268 /* No pages requested? */
5269 if (unlikely(nr_pages <= 0))
5272 /* Already populated array? */
5273 if (unlikely(page_array && nr_pages - nr_populated == 0))
5276 /* Bulk allocator does not support memcg accounting. */
5277 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5280 /* Use the single page allocator for one page. */
5281 if (nr_pages - nr_populated == 1)
5284 #ifdef CONFIG_PAGE_OWNER
5286 * PAGE_OWNER may recurse into the allocator to allocate space to
5287 * save the stack with pagesets.lock held. Releasing/reacquiring
5288 * removes much of the performance benefit of bulk allocation so
5289 * force the caller to allocate one page at a time as it'll have
5290 * similar performance to added complexity to the bulk allocator.
5292 if (static_branch_unlikely(&page_owner_inited))
5296 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5297 gfp &= gfp_allowed_mask;
5299 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5303 /* Find an allowed local zone that meets the low watermark. */
5304 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5307 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5308 !__cpuset_zone_allowed(zone, gfp)) {
5312 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5313 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5317 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5318 if (zone_watermark_fast(zone, 0, mark,
5319 zonelist_zone_idx(ac.preferred_zoneref),
5320 alloc_flags, gfp)) {
5326 * If there are no allowed local zones that meets the watermarks then
5327 * try to allocate a single page and reclaim if necessary.
5329 if (unlikely(!zone))
5332 /* Attempt the batch allocation */
5333 local_lock_irqsave(&pagesets.lock, flags);
5334 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5335 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5337 while (nr_populated < nr_pages) {
5339 /* Skip existing pages */
5340 if (page_array && page_array[nr_populated]) {
5345 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5347 if (unlikely(!page)) {
5348 /* Try and allocate at least one page */
5355 prep_new_page(page, 0, gfp, 0);
5357 list_add(&page->lru, page_list);
5359 page_array[nr_populated] = page;
5363 local_unlock_irqrestore(&pagesets.lock, flags);
5365 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5366 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5369 return nr_populated;
5372 local_unlock_irqrestore(&pagesets.lock, flags);
5375 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5378 list_add(&page->lru, page_list);
5380 page_array[nr_populated] = page;
5386 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5389 * This is the 'heart' of the zoned buddy allocator.
5391 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5392 nodemask_t *nodemask)
5395 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5396 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5397 struct alloc_context ac = { };
5400 * There are several places where we assume that the order value is sane
5401 * so bail out early if the request is out of bound.
5403 if (WARN_ON_ONCE_GFP(order >= MAX_ORDER, gfp))
5406 gfp &= gfp_allowed_mask;
5408 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5409 * resp. GFP_NOIO which has to be inherited for all allocation requests
5410 * from a particular context which has been marked by
5411 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5412 * movable zones are not used during allocation.
5414 gfp = current_gfp_context(gfp);
5416 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5417 &alloc_gfp, &alloc_flags))
5421 * Forbid the first pass from falling back to types that fragment
5422 * memory until all local zones are considered.
5424 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5426 /* First allocation attempt */
5427 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5432 ac.spread_dirty_pages = false;
5435 * Restore the original nodemask if it was potentially replaced with
5436 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5438 ac.nodemask = nodemask;
5440 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5443 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5444 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5445 __free_pages(page, order);
5449 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5453 EXPORT_SYMBOL(__alloc_pages);
5455 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5456 nodemask_t *nodemask)
5458 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5459 preferred_nid, nodemask);
5461 if (page && order > 1)
5462 prep_transhuge_page(page);
5463 return (struct folio *)page;
5465 EXPORT_SYMBOL(__folio_alloc);
5468 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5469 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5470 * you need to access high mem.
5472 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5476 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5479 return (unsigned long) page_address(page);
5481 EXPORT_SYMBOL(__get_free_pages);
5483 unsigned long get_zeroed_page(gfp_t gfp_mask)
5485 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5487 EXPORT_SYMBOL(get_zeroed_page);
5490 * __free_pages - Free pages allocated with alloc_pages().
5491 * @page: The page pointer returned from alloc_pages().
5492 * @order: The order of the allocation.
5494 * This function can free multi-page allocations that are not compound
5495 * pages. It does not check that the @order passed in matches that of
5496 * the allocation, so it is easy to leak memory. Freeing more memory
5497 * than was allocated will probably emit a warning.
5499 * If the last reference to this page is speculative, it will be released
5500 * by put_page() which only frees the first page of a non-compound
5501 * allocation. To prevent the remaining pages from being leaked, we free
5502 * the subsequent pages here. If you want to use the page's reference
5503 * count to decide when to free the allocation, you should allocate a
5504 * compound page, and use put_page() instead of __free_pages().
5506 * Context: May be called in interrupt context or while holding a normal
5507 * spinlock, but not in NMI context or while holding a raw spinlock.
5509 void __free_pages(struct page *page, unsigned int order)
5511 if (put_page_testzero(page))
5512 free_the_page(page, order);
5513 else if (!PageHead(page))
5515 free_the_page(page + (1 << order), order);
5517 EXPORT_SYMBOL(__free_pages);
5519 void free_pages(unsigned long addr, unsigned int order)
5522 VM_BUG_ON(!virt_addr_valid((void *)addr));
5523 __free_pages(virt_to_page((void *)addr), order);
5527 EXPORT_SYMBOL(free_pages);
5531 * An arbitrary-length arbitrary-offset area of memory which resides
5532 * within a 0 or higher order page. Multiple fragments within that page
5533 * are individually refcounted, in the page's reference counter.
5535 * The page_frag functions below provide a simple allocation framework for
5536 * page fragments. This is used by the network stack and network device
5537 * drivers to provide a backing region of memory for use as either an
5538 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5540 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5543 struct page *page = NULL;
5544 gfp_t gfp = gfp_mask;
5546 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5547 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5549 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5550 PAGE_FRAG_CACHE_MAX_ORDER);
5551 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5553 if (unlikely(!page))
5554 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5556 nc->va = page ? page_address(page) : NULL;
5561 void __page_frag_cache_drain(struct page *page, unsigned int count)
5563 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5565 if (page_ref_sub_and_test(page, count))
5566 free_the_page(page, compound_order(page));
5568 EXPORT_SYMBOL(__page_frag_cache_drain);
5570 void *page_frag_alloc_align(struct page_frag_cache *nc,
5571 unsigned int fragsz, gfp_t gfp_mask,
5572 unsigned int align_mask)
5574 unsigned int size = PAGE_SIZE;
5578 if (unlikely(!nc->va)) {
5580 page = __page_frag_cache_refill(nc, gfp_mask);
5584 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5585 /* if size can vary use size else just use PAGE_SIZE */
5588 /* Even if we own the page, we do not use atomic_set().
5589 * This would break get_page_unless_zero() users.
5591 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5593 /* reset page count bias and offset to start of new frag */
5594 nc->pfmemalloc = page_is_pfmemalloc(page);
5595 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5599 offset = nc->offset - fragsz;
5600 if (unlikely(offset < 0)) {
5601 page = virt_to_page(nc->va);
5603 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5606 if (unlikely(nc->pfmemalloc)) {
5607 free_the_page(page, compound_order(page));
5611 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5612 /* if size can vary use size else just use PAGE_SIZE */
5615 /* OK, page count is 0, we can safely set it */
5616 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5618 /* reset page count bias and offset to start of new frag */
5619 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5620 offset = size - fragsz;
5624 offset &= align_mask;
5625 nc->offset = offset;
5627 return nc->va + offset;
5629 EXPORT_SYMBOL(page_frag_alloc_align);
5632 * Frees a page fragment allocated out of either a compound or order 0 page.
5634 void page_frag_free(void *addr)
5636 struct page *page = virt_to_head_page(addr);
5638 if (unlikely(put_page_testzero(page)))
5639 free_the_page(page, compound_order(page));
5641 EXPORT_SYMBOL(page_frag_free);
5643 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5647 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5648 unsigned long used = addr + PAGE_ALIGN(size);
5650 split_page(virt_to_page((void *)addr), order);
5651 while (used < alloc_end) {
5656 return (void *)addr;
5660 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5661 * @size: the number of bytes to allocate
5662 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5664 * This function is similar to alloc_pages(), except that it allocates the
5665 * minimum number of pages to satisfy the request. alloc_pages() can only
5666 * allocate memory in power-of-two pages.
5668 * This function is also limited by MAX_ORDER.
5670 * Memory allocated by this function must be released by free_pages_exact().
5672 * Return: pointer to the allocated area or %NULL in case of error.
5674 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5676 unsigned int order = get_order(size);
5679 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5680 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5682 addr = __get_free_pages(gfp_mask, order);
5683 return make_alloc_exact(addr, order, size);
5685 EXPORT_SYMBOL(alloc_pages_exact);
5688 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5690 * @nid: the preferred node ID where memory should be allocated
5691 * @size: the number of bytes to allocate
5692 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5694 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5697 * Return: pointer to the allocated area or %NULL in case of error.
5699 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5701 unsigned int order = get_order(size);
5704 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5705 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5707 p = alloc_pages_node(nid, gfp_mask, order);
5710 return make_alloc_exact((unsigned long)page_address(p), order, size);
5714 * free_pages_exact - release memory allocated via alloc_pages_exact()
5715 * @virt: the value returned by alloc_pages_exact.
5716 * @size: size of allocation, same value as passed to alloc_pages_exact().
5718 * Release the memory allocated by a previous call to alloc_pages_exact.
5720 void free_pages_exact(void *virt, size_t size)
5722 unsigned long addr = (unsigned long)virt;
5723 unsigned long end = addr + PAGE_ALIGN(size);
5725 while (addr < end) {
5730 EXPORT_SYMBOL(free_pages_exact);
5733 * nr_free_zone_pages - count number of pages beyond high watermark
5734 * @offset: The zone index of the highest zone
5736 * nr_free_zone_pages() counts the number of pages which are beyond the
5737 * high watermark within all zones at or below a given zone index. For each
5738 * zone, the number of pages is calculated as:
5740 * nr_free_zone_pages = managed_pages - high_pages
5742 * Return: number of pages beyond high watermark.
5744 static unsigned long nr_free_zone_pages(int offset)
5749 /* Just pick one node, since fallback list is circular */
5750 unsigned long sum = 0;
5752 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5754 for_each_zone_zonelist(zone, z, zonelist, offset) {
5755 unsigned long size = zone_managed_pages(zone);
5756 unsigned long high = high_wmark_pages(zone);
5765 * nr_free_buffer_pages - count number of pages beyond high watermark
5767 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5768 * watermark within ZONE_DMA and ZONE_NORMAL.
5770 * Return: number of pages beyond high watermark within ZONE_DMA and
5773 unsigned long nr_free_buffer_pages(void)
5775 return nr_free_zone_pages(gfp_zone(GFP_USER));
5777 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5779 static inline void show_node(struct zone *zone)
5781 if (IS_ENABLED(CONFIG_NUMA))
5782 printk("Node %d ", zone_to_nid(zone));
5785 long si_mem_available(void)
5788 unsigned long pagecache;
5789 unsigned long wmark_low = 0;
5790 unsigned long pages[NR_LRU_LISTS];
5791 unsigned long reclaimable;
5795 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5796 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5799 wmark_low += low_wmark_pages(zone);
5802 * Estimate the amount of memory available for userspace allocations,
5803 * without causing swapping or OOM.
5805 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5808 * Not all the page cache can be freed, otherwise the system will
5809 * start swapping or thrashing. Assume at least half of the page
5810 * cache, or the low watermark worth of cache, needs to stay.
5812 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5813 pagecache -= min(pagecache / 2, wmark_low);
5814 available += pagecache;
5817 * Part of the reclaimable slab and other kernel memory consists of
5818 * items that are in use, and cannot be freed. Cap this estimate at the
5821 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5822 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5823 available += reclaimable - min(reclaimable / 2, wmark_low);
5829 EXPORT_SYMBOL_GPL(si_mem_available);
5831 void si_meminfo(struct sysinfo *val)
5833 val->totalram = totalram_pages();
5834 val->sharedram = global_node_page_state(NR_SHMEM);
5835 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5836 val->bufferram = nr_blockdev_pages();
5837 val->totalhigh = totalhigh_pages();
5838 val->freehigh = nr_free_highpages();
5839 val->mem_unit = PAGE_SIZE;
5842 EXPORT_SYMBOL(si_meminfo);
5845 void si_meminfo_node(struct sysinfo *val, int nid)
5847 int zone_type; /* needs to be signed */
5848 unsigned long managed_pages = 0;
5849 unsigned long managed_highpages = 0;
5850 unsigned long free_highpages = 0;
5851 pg_data_t *pgdat = NODE_DATA(nid);
5853 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5854 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5855 val->totalram = managed_pages;
5856 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5857 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5858 #ifdef CONFIG_HIGHMEM
5859 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5860 struct zone *zone = &pgdat->node_zones[zone_type];
5862 if (is_highmem(zone)) {
5863 managed_highpages += zone_managed_pages(zone);
5864 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5867 val->totalhigh = managed_highpages;
5868 val->freehigh = free_highpages;
5870 val->totalhigh = managed_highpages;
5871 val->freehigh = free_highpages;
5873 val->mem_unit = PAGE_SIZE;
5878 * Determine whether the node should be displayed or not, depending on whether
5879 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5881 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5883 if (!(flags & SHOW_MEM_FILTER_NODES))
5887 * no node mask - aka implicit memory numa policy. Do not bother with
5888 * the synchronization - read_mems_allowed_begin - because we do not
5889 * have to be precise here.
5892 nodemask = &cpuset_current_mems_allowed;
5894 return !node_isset(nid, *nodemask);
5897 #define K(x) ((x) << (PAGE_SHIFT-10))
5899 static void show_migration_types(unsigned char type)
5901 static const char types[MIGRATE_TYPES] = {
5902 [MIGRATE_UNMOVABLE] = 'U',
5903 [MIGRATE_MOVABLE] = 'M',
5904 [MIGRATE_RECLAIMABLE] = 'E',
5905 [MIGRATE_HIGHATOMIC] = 'H',
5907 [MIGRATE_CMA] = 'C',
5909 #ifdef CONFIG_MEMORY_ISOLATION
5910 [MIGRATE_ISOLATE] = 'I',
5913 char tmp[MIGRATE_TYPES + 1];
5917 for (i = 0; i < MIGRATE_TYPES; i++) {
5918 if (type & (1 << i))
5923 printk(KERN_CONT "(%s) ", tmp);
5927 * Show free area list (used inside shift_scroll-lock stuff)
5928 * We also calculate the percentage fragmentation. We do this by counting the
5929 * memory on each free list with the exception of the first item on the list.
5932 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5935 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5937 unsigned long free_pcp = 0;
5942 for_each_populated_zone(zone) {
5943 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5946 for_each_online_cpu(cpu)
5947 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5950 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5951 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5952 " unevictable:%lu dirty:%lu writeback:%lu\n"
5953 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5954 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5955 " kernel_misc_reclaimable:%lu\n"
5956 " free:%lu free_pcp:%lu free_cma:%lu\n",
5957 global_node_page_state(NR_ACTIVE_ANON),
5958 global_node_page_state(NR_INACTIVE_ANON),
5959 global_node_page_state(NR_ISOLATED_ANON),
5960 global_node_page_state(NR_ACTIVE_FILE),
5961 global_node_page_state(NR_INACTIVE_FILE),
5962 global_node_page_state(NR_ISOLATED_FILE),
5963 global_node_page_state(NR_UNEVICTABLE),
5964 global_node_page_state(NR_FILE_DIRTY),
5965 global_node_page_state(NR_WRITEBACK),
5966 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5967 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5968 global_node_page_state(NR_FILE_MAPPED),
5969 global_node_page_state(NR_SHMEM),
5970 global_node_page_state(NR_PAGETABLE),
5971 global_zone_page_state(NR_BOUNCE),
5972 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5973 global_zone_page_state(NR_FREE_PAGES),
5975 global_zone_page_state(NR_FREE_CMA_PAGES));
5977 for_each_online_pgdat(pgdat) {
5978 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5982 " active_anon:%lukB"
5983 " inactive_anon:%lukB"
5984 " active_file:%lukB"
5985 " inactive_file:%lukB"
5986 " unevictable:%lukB"
5987 " isolated(anon):%lukB"
5988 " isolated(file):%lukB"
5993 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5995 " shmem_pmdmapped: %lukB"
5998 " writeback_tmp:%lukB"
5999 " kernel_stack:%lukB"
6000 #ifdef CONFIG_SHADOW_CALL_STACK
6001 " shadow_call_stack:%lukB"
6004 " all_unreclaimable? %s"
6007 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
6008 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
6009 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
6010 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
6011 K(node_page_state(pgdat, NR_UNEVICTABLE)),
6012 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
6013 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
6014 K(node_page_state(pgdat, NR_FILE_MAPPED)),
6015 K(node_page_state(pgdat, NR_FILE_DIRTY)),
6016 K(node_page_state(pgdat, NR_WRITEBACK)),
6017 K(node_page_state(pgdat, NR_SHMEM)),
6018 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6019 K(node_page_state(pgdat, NR_SHMEM_THPS)),
6020 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6021 K(node_page_state(pgdat, NR_ANON_THPS)),
6023 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6024 node_page_state(pgdat, NR_KERNEL_STACK_KB),
6025 #ifdef CONFIG_SHADOW_CALL_STACK
6026 node_page_state(pgdat, NR_KERNEL_SCS_KB),
6028 K(node_page_state(pgdat, NR_PAGETABLE)),
6029 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6033 for_each_populated_zone(zone) {
6036 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6040 for_each_online_cpu(cpu)
6041 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6051 " reserved_highatomic:%luKB"
6052 " active_anon:%lukB"
6053 " inactive_anon:%lukB"
6054 " active_file:%lukB"
6055 " inactive_file:%lukB"
6056 " unevictable:%lukB"
6057 " writepending:%lukB"
6067 K(zone_page_state(zone, NR_FREE_PAGES)),
6068 K(zone->watermark_boost),
6069 K(min_wmark_pages(zone)),
6070 K(low_wmark_pages(zone)),
6071 K(high_wmark_pages(zone)),
6072 K(zone->nr_reserved_highatomic),
6073 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6074 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6075 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6076 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6077 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6078 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6079 K(zone->present_pages),
6080 K(zone_managed_pages(zone)),
6081 K(zone_page_state(zone, NR_MLOCK)),
6082 K(zone_page_state(zone, NR_BOUNCE)),
6084 K(this_cpu_read(zone->per_cpu_pageset->count)),
6085 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6086 printk("lowmem_reserve[]:");
6087 for (i = 0; i < MAX_NR_ZONES; i++)
6088 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6089 printk(KERN_CONT "\n");
6092 for_each_populated_zone(zone) {
6094 unsigned long nr[MAX_ORDER], flags, total = 0;
6095 unsigned char types[MAX_ORDER];
6097 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6100 printk(KERN_CONT "%s: ", zone->name);
6102 spin_lock_irqsave(&zone->lock, flags);
6103 for (order = 0; order < MAX_ORDER; order++) {
6104 struct free_area *area = &zone->free_area[order];
6107 nr[order] = area->nr_free;
6108 total += nr[order] << order;
6111 for (type = 0; type < MIGRATE_TYPES; type++) {
6112 if (!free_area_empty(area, type))
6113 types[order] |= 1 << type;
6116 spin_unlock_irqrestore(&zone->lock, flags);
6117 for (order = 0; order < MAX_ORDER; order++) {
6118 printk(KERN_CONT "%lu*%lukB ",
6119 nr[order], K(1UL) << order);
6121 show_migration_types(types[order]);
6123 printk(KERN_CONT "= %lukB\n", K(total));
6126 hugetlb_show_meminfo();
6128 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6130 show_swap_cache_info();
6133 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6135 zoneref->zone = zone;
6136 zoneref->zone_idx = zone_idx(zone);
6140 * Builds allocation fallback zone lists.
6142 * Add all populated zones of a node to the zonelist.
6144 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6147 enum zone_type zone_type = MAX_NR_ZONES;
6152 zone = pgdat->node_zones + zone_type;
6153 if (populated_zone(zone)) {
6154 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6155 check_highest_zone(zone_type);
6157 } while (zone_type);
6164 static int __parse_numa_zonelist_order(char *s)
6167 * We used to support different zonelists modes but they turned
6168 * out to be just not useful. Let's keep the warning in place
6169 * if somebody still use the cmd line parameter so that we do
6170 * not fail it silently
6172 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6173 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6179 char numa_zonelist_order[] = "Node";
6182 * sysctl handler for numa_zonelist_order
6184 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6185 void *buffer, size_t *length, loff_t *ppos)
6188 return __parse_numa_zonelist_order(buffer);
6189 return proc_dostring(table, write, buffer, length, ppos);
6193 static int node_load[MAX_NUMNODES];
6196 * find_next_best_node - find the next node that should appear in a given node's fallback list
6197 * @node: node whose fallback list we're appending
6198 * @used_node_mask: nodemask_t of already used nodes
6200 * We use a number of factors to determine which is the next node that should
6201 * appear on a given node's fallback list. The node should not have appeared
6202 * already in @node's fallback list, and it should be the next closest node
6203 * according to the distance array (which contains arbitrary distance values
6204 * from each node to each node in the system), and should also prefer nodes
6205 * with no CPUs, since presumably they'll have very little allocation pressure
6206 * on them otherwise.
6208 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6210 int find_next_best_node(int node, nodemask_t *used_node_mask)
6213 int min_val = INT_MAX;
6214 int best_node = NUMA_NO_NODE;
6216 /* Use the local node if we haven't already */
6217 if (!node_isset(node, *used_node_mask)) {
6218 node_set(node, *used_node_mask);
6222 for_each_node_state(n, N_MEMORY) {
6224 /* Don't want a node to appear more than once */
6225 if (node_isset(n, *used_node_mask))
6228 /* Use the distance array to find the distance */
6229 val = node_distance(node, n);
6231 /* Penalize nodes under us ("prefer the next node") */
6234 /* Give preference to headless and unused nodes */
6235 if (!cpumask_empty(cpumask_of_node(n)))
6236 val += PENALTY_FOR_NODE_WITH_CPUS;
6238 /* Slight preference for less loaded node */
6239 val *= MAX_NUMNODES;
6240 val += node_load[n];
6242 if (val < min_val) {
6249 node_set(best_node, *used_node_mask);
6256 * Build zonelists ordered by node and zones within node.
6257 * This results in maximum locality--normal zone overflows into local
6258 * DMA zone, if any--but risks exhausting DMA zone.
6260 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6263 struct zoneref *zonerefs;
6266 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6268 for (i = 0; i < nr_nodes; i++) {
6271 pg_data_t *node = NODE_DATA(node_order[i]);
6273 nr_zones = build_zonerefs_node(node, zonerefs);
6274 zonerefs += nr_zones;
6276 zonerefs->zone = NULL;
6277 zonerefs->zone_idx = 0;
6281 * Build gfp_thisnode zonelists
6283 static void build_thisnode_zonelists(pg_data_t *pgdat)
6285 struct zoneref *zonerefs;
6288 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6289 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6290 zonerefs += nr_zones;
6291 zonerefs->zone = NULL;
6292 zonerefs->zone_idx = 0;
6296 * Build zonelists ordered by zone and nodes within zones.
6297 * This results in conserving DMA zone[s] until all Normal memory is
6298 * exhausted, but results in overflowing to remote node while memory
6299 * may still exist in local DMA zone.
6302 static void build_zonelists(pg_data_t *pgdat)
6304 static int node_order[MAX_NUMNODES];
6305 int node, nr_nodes = 0;
6306 nodemask_t used_mask = NODE_MASK_NONE;
6307 int local_node, prev_node;
6309 /* NUMA-aware ordering of nodes */
6310 local_node = pgdat->node_id;
6311 prev_node = local_node;
6313 memset(node_order, 0, sizeof(node_order));
6314 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6316 * We don't want to pressure a particular node.
6317 * So adding penalty to the first node in same
6318 * distance group to make it round-robin.
6320 if (node_distance(local_node, node) !=
6321 node_distance(local_node, prev_node))
6322 node_load[node] += 1;
6324 node_order[nr_nodes++] = node;
6328 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6329 build_thisnode_zonelists(pgdat);
6330 pr_info("Fallback order for Node %d: ", local_node);
6331 for (node = 0; node < nr_nodes; node++)
6332 pr_cont("%d ", node_order[node]);
6336 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6338 * Return node id of node used for "local" allocations.
6339 * I.e., first node id of first zone in arg node's generic zonelist.
6340 * Used for initializing percpu 'numa_mem', which is used primarily
6341 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6343 int local_memory_node(int node)
6347 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6348 gfp_zone(GFP_KERNEL),
6350 return zone_to_nid(z->zone);
6354 static void setup_min_unmapped_ratio(void);
6355 static void setup_min_slab_ratio(void);
6356 #else /* CONFIG_NUMA */
6358 static void build_zonelists(pg_data_t *pgdat)
6360 int node, local_node;
6361 struct zoneref *zonerefs;
6364 local_node = pgdat->node_id;
6366 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6367 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6368 zonerefs += nr_zones;
6371 * Now we build the zonelist so that it contains the zones
6372 * of all the other nodes.
6373 * We don't want to pressure a particular node, so when
6374 * building the zones for node N, we make sure that the
6375 * zones coming right after the local ones are those from
6376 * node N+1 (modulo N)
6378 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6379 if (!node_online(node))
6381 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6382 zonerefs += nr_zones;
6384 for (node = 0; node < local_node; node++) {
6385 if (!node_online(node))
6387 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6388 zonerefs += nr_zones;
6391 zonerefs->zone = NULL;
6392 zonerefs->zone_idx = 0;
6395 #endif /* CONFIG_NUMA */
6398 * Boot pageset table. One per cpu which is going to be used for all
6399 * zones and all nodes. The parameters will be set in such a way
6400 * that an item put on a list will immediately be handed over to
6401 * the buddy list. This is safe since pageset manipulation is done
6402 * with interrupts disabled.
6404 * The boot_pagesets must be kept even after bootup is complete for
6405 * unused processors and/or zones. They do play a role for bootstrapping
6406 * hotplugged processors.
6408 * zoneinfo_show() and maybe other functions do
6409 * not check if the processor is online before following the pageset pointer.
6410 * Other parts of the kernel may not check if the zone is available.
6412 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6413 /* These effectively disable the pcplists in the boot pageset completely */
6414 #define BOOT_PAGESET_HIGH 0
6415 #define BOOT_PAGESET_BATCH 1
6416 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6417 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6418 DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6420 static void __build_all_zonelists(void *data)
6423 int __maybe_unused cpu;
6424 pg_data_t *self = data;
6425 static DEFINE_SPINLOCK(lock);
6430 memset(node_load, 0, sizeof(node_load));
6434 * This node is hotadded and no memory is yet present. So just
6435 * building zonelists is fine - no need to touch other nodes.
6437 if (self && !node_online(self->node_id)) {
6438 build_zonelists(self);
6441 * All possible nodes have pgdat preallocated
6444 for_each_node(nid) {
6445 pg_data_t *pgdat = NODE_DATA(nid);
6447 build_zonelists(pgdat);
6450 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6452 * We now know the "local memory node" for each node--
6453 * i.e., the node of the first zone in the generic zonelist.
6454 * Set up numa_mem percpu variable for on-line cpus. During
6455 * boot, only the boot cpu should be on-line; we'll init the
6456 * secondary cpus' numa_mem as they come on-line. During
6457 * node/memory hotplug, we'll fixup all on-line cpus.
6459 for_each_online_cpu(cpu)
6460 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6467 static noinline void __init
6468 build_all_zonelists_init(void)
6472 __build_all_zonelists(NULL);
6475 * Initialize the boot_pagesets that are going to be used
6476 * for bootstrapping processors. The real pagesets for
6477 * each zone will be allocated later when the per cpu
6478 * allocator is available.
6480 * boot_pagesets are used also for bootstrapping offline
6481 * cpus if the system is already booted because the pagesets
6482 * are needed to initialize allocators on a specific cpu too.
6483 * F.e. the percpu allocator needs the page allocator which
6484 * needs the percpu allocator in order to allocate its pagesets
6485 * (a chicken-egg dilemma).
6487 for_each_possible_cpu(cpu)
6488 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6490 mminit_verify_zonelist();
6491 cpuset_init_current_mems_allowed();
6495 * unless system_state == SYSTEM_BOOTING.
6497 * __ref due to call of __init annotated helper build_all_zonelists_init
6498 * [protected by SYSTEM_BOOTING].
6500 void __ref build_all_zonelists(pg_data_t *pgdat)
6502 unsigned long vm_total_pages;
6504 if (system_state == SYSTEM_BOOTING) {
6505 build_all_zonelists_init();
6507 __build_all_zonelists(pgdat);
6508 /* cpuset refresh routine should be here */
6510 /* Get the number of free pages beyond high watermark in all zones. */
6511 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6513 * Disable grouping by mobility if the number of pages in the
6514 * system is too low to allow the mechanism to work. It would be
6515 * more accurate, but expensive to check per-zone. This check is
6516 * made on memory-hotadd so a system can start with mobility
6517 * disabled and enable it later
6519 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6520 page_group_by_mobility_disabled = 1;
6522 page_group_by_mobility_disabled = 0;
6524 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6526 page_group_by_mobility_disabled ? "off" : "on",
6529 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6533 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6534 static bool __meminit
6535 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6537 static struct memblock_region *r;
6539 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6540 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6541 for_each_mem_region(r) {
6542 if (*pfn < memblock_region_memory_end_pfn(r))
6546 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6547 memblock_is_mirror(r)) {
6548 *pfn = memblock_region_memory_end_pfn(r);
6556 * Initially all pages are reserved - free ones are freed
6557 * up by memblock_free_all() once the early boot process is
6558 * done. Non-atomic initialization, single-pass.
6560 * All aligned pageblocks are initialized to the specified migratetype
6561 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6562 * zone stats (e.g., nr_isolate_pageblock) are touched.
6564 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6565 unsigned long start_pfn, unsigned long zone_end_pfn,
6566 enum meminit_context context,
6567 struct vmem_altmap *altmap, int migratetype)
6569 unsigned long pfn, end_pfn = start_pfn + size;
6572 if (highest_memmap_pfn < end_pfn - 1)
6573 highest_memmap_pfn = end_pfn - 1;
6575 #ifdef CONFIG_ZONE_DEVICE
6577 * Honor reservation requested by the driver for this ZONE_DEVICE
6578 * memory. We limit the total number of pages to initialize to just
6579 * those that might contain the memory mapping. We will defer the
6580 * ZONE_DEVICE page initialization until after we have released
6583 if (zone == ZONE_DEVICE) {
6587 if (start_pfn == altmap->base_pfn)
6588 start_pfn += altmap->reserve;
6589 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6593 for (pfn = start_pfn; pfn < end_pfn; ) {
6595 * There can be holes in boot-time mem_map[]s handed to this
6596 * function. They do not exist on hotplugged memory.
6598 if (context == MEMINIT_EARLY) {
6599 if (overlap_memmap_init(zone, &pfn))
6601 if (defer_init(nid, pfn, zone_end_pfn))
6605 page = pfn_to_page(pfn);
6606 __init_single_page(page, pfn, zone, nid);
6607 if (context == MEMINIT_HOTPLUG)
6608 __SetPageReserved(page);
6611 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6612 * such that unmovable allocations won't be scattered all
6613 * over the place during system boot.
6615 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6616 set_pageblock_migratetype(page, migratetype);
6623 #ifdef CONFIG_ZONE_DEVICE
6624 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6625 unsigned long zone_idx, int nid,
6626 struct dev_pagemap *pgmap)
6629 __init_single_page(page, pfn, zone_idx, nid);
6632 * Mark page reserved as it will need to wait for onlining
6633 * phase for it to be fully associated with a zone.
6635 * We can use the non-atomic __set_bit operation for setting
6636 * the flag as we are still initializing the pages.
6638 __SetPageReserved(page);
6641 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6642 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6643 * ever freed or placed on a driver-private list.
6645 page->pgmap = pgmap;
6646 page->zone_device_data = NULL;
6649 * Mark the block movable so that blocks are reserved for
6650 * movable at startup. This will force kernel allocations
6651 * to reserve their blocks rather than leaking throughout
6652 * the address space during boot when many long-lived
6653 * kernel allocations are made.
6655 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6656 * because this is done early in section_activate()
6658 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6659 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6665 * With compound page geometry and when struct pages are stored in ram most
6666 * tail pages are reused. Consequently, the amount of unique struct pages to
6667 * initialize is a lot smaller that the total amount of struct pages being
6668 * mapped. This is a paired / mild layering violation with explicit knowledge
6669 * of how the sparse_vmemmap internals handle compound pages in the lack
6670 * of an altmap. See vmemmap_populate_compound_pages().
6672 static inline unsigned long compound_nr_pages(struct vmem_altmap *altmap,
6673 unsigned long nr_pages)
6675 return is_power_of_2(sizeof(struct page)) &&
6676 !altmap ? 2 * (PAGE_SIZE / sizeof(struct page)) : nr_pages;
6679 static void __ref memmap_init_compound(struct page *head,
6680 unsigned long head_pfn,
6681 unsigned long zone_idx, int nid,
6682 struct dev_pagemap *pgmap,
6683 unsigned long nr_pages)
6685 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6686 unsigned int order = pgmap->vmemmap_shift;
6688 __SetPageHead(head);
6689 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6690 struct page *page = pfn_to_page(pfn);
6692 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6693 prep_compound_tail(head, pfn - head_pfn);
6694 set_page_count(page, 0);
6697 * The first tail page stores compound_mapcount_ptr() and
6698 * compound_order() and the second tail page stores
6699 * compound_pincount_ptr(). Call prep_compound_head() after
6700 * the first and second tail pages have been initialized to
6701 * not have the data overwritten.
6703 if (pfn == head_pfn + 2)
6704 prep_compound_head(head, order);
6708 void __ref memmap_init_zone_device(struct zone *zone,
6709 unsigned long start_pfn,
6710 unsigned long nr_pages,
6711 struct dev_pagemap *pgmap)
6713 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6714 struct pglist_data *pgdat = zone->zone_pgdat;
6715 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6716 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6717 unsigned long zone_idx = zone_idx(zone);
6718 unsigned long start = jiffies;
6719 int nid = pgdat->node_id;
6721 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6725 * The call to memmap_init should have already taken care
6726 * of the pages reserved for the memmap, so we can just jump to
6727 * the end of that region and start processing the device pages.
6730 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6731 nr_pages = end_pfn - start_pfn;
6734 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6735 struct page *page = pfn_to_page(pfn);
6737 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6739 if (pfns_per_compound == 1)
6742 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6743 compound_nr_pages(altmap, pfns_per_compound));
6746 pr_info("%s initialised %lu pages in %ums\n", __func__,
6747 nr_pages, jiffies_to_msecs(jiffies - start));
6751 static void __meminit zone_init_free_lists(struct zone *zone)
6753 unsigned int order, t;
6754 for_each_migratetype_order(order, t) {
6755 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6756 zone->free_area[order].nr_free = 0;
6761 * Only struct pages that correspond to ranges defined by memblock.memory
6762 * are zeroed and initialized by going through __init_single_page() during
6763 * memmap_init_zone_range().
6765 * But, there could be struct pages that correspond to holes in
6766 * memblock.memory. This can happen because of the following reasons:
6767 * - physical memory bank size is not necessarily the exact multiple of the
6768 * arbitrary section size
6769 * - early reserved memory may not be listed in memblock.memory
6770 * - memory layouts defined with memmap= kernel parameter may not align
6771 * nicely with memmap sections
6773 * Explicitly initialize those struct pages so that:
6774 * - PG_Reserved is set
6775 * - zone and node links point to zone and node that span the page if the
6776 * hole is in the middle of a zone
6777 * - zone and node links point to adjacent zone/node if the hole falls on
6778 * the zone boundary; the pages in such holes will be prepended to the
6779 * zone/node above the hole except for the trailing pages in the last
6780 * section that will be appended to the zone/node below.
6782 static void __init init_unavailable_range(unsigned long spfn,
6789 for (pfn = spfn; pfn < epfn; pfn++) {
6790 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6791 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6792 + pageblock_nr_pages - 1;
6795 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6796 __SetPageReserved(pfn_to_page(pfn));
6801 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6802 node, zone_names[zone], pgcnt);
6805 static void __init memmap_init_zone_range(struct zone *zone,
6806 unsigned long start_pfn,
6807 unsigned long end_pfn,
6808 unsigned long *hole_pfn)
6810 unsigned long zone_start_pfn = zone->zone_start_pfn;
6811 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6812 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6814 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6815 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6817 if (start_pfn >= end_pfn)
6820 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6821 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6823 if (*hole_pfn < start_pfn)
6824 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6826 *hole_pfn = end_pfn;
6829 static void __init memmap_init(void)
6831 unsigned long start_pfn, end_pfn;
6832 unsigned long hole_pfn = 0;
6833 int i, j, zone_id = 0, nid;
6835 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6836 struct pglist_data *node = NODE_DATA(nid);
6838 for (j = 0; j < MAX_NR_ZONES; j++) {
6839 struct zone *zone = node->node_zones + j;
6841 if (!populated_zone(zone))
6844 memmap_init_zone_range(zone, start_pfn, end_pfn,
6850 #ifdef CONFIG_SPARSEMEM
6852 * Initialize the memory map for hole in the range [memory_end,
6854 * Append the pages in this hole to the highest zone in the last
6856 * The call to init_unavailable_range() is outside the ifdef to
6857 * silence the compiler warining about zone_id set but not used;
6858 * for FLATMEM it is a nop anyway
6860 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6861 if (hole_pfn < end_pfn)
6863 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6866 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6867 phys_addr_t min_addr, int nid, bool exact_nid)
6872 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6873 MEMBLOCK_ALLOC_ACCESSIBLE,
6876 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6877 MEMBLOCK_ALLOC_ACCESSIBLE,
6880 if (ptr && size > 0)
6881 page_init_poison(ptr, size);
6886 static int zone_batchsize(struct zone *zone)
6892 * The number of pages to batch allocate is either ~0.1%
6893 * of the zone or 1MB, whichever is smaller. The batch
6894 * size is striking a balance between allocation latency
6895 * and zone lock contention.
6897 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6898 batch /= 4; /* We effectively *= 4 below */
6903 * Clamp the batch to a 2^n - 1 value. Having a power
6904 * of 2 value was found to be more likely to have
6905 * suboptimal cache aliasing properties in some cases.
6907 * For example if 2 tasks are alternately allocating
6908 * batches of pages, one task can end up with a lot
6909 * of pages of one half of the possible page colors
6910 * and the other with pages of the other colors.
6912 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6917 /* The deferral and batching of frees should be suppressed under NOMMU
6920 * The problem is that NOMMU needs to be able to allocate large chunks
6921 * of contiguous memory as there's no hardware page translation to
6922 * assemble apparent contiguous memory from discontiguous pages.
6924 * Queueing large contiguous runs of pages for batching, however,
6925 * causes the pages to actually be freed in smaller chunks. As there
6926 * can be a significant delay between the individual batches being
6927 * recycled, this leads to the once large chunks of space being
6928 * fragmented and becoming unavailable for high-order allocations.
6934 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6939 unsigned long total_pages;
6941 if (!percpu_pagelist_high_fraction) {
6943 * By default, the high value of the pcp is based on the zone
6944 * low watermark so that if they are full then background
6945 * reclaim will not be started prematurely.
6947 total_pages = low_wmark_pages(zone);
6950 * If percpu_pagelist_high_fraction is configured, the high
6951 * value is based on a fraction of the managed pages in the
6954 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6958 * Split the high value across all online CPUs local to the zone. Note
6959 * that early in boot that CPUs may not be online yet and that during
6960 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6961 * onlined. For memory nodes that have no CPUs, split pcp->high across
6962 * all online CPUs to mitigate the risk that reclaim is triggered
6963 * prematurely due to pages stored on pcp lists.
6965 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6967 nr_split_cpus = num_online_cpus();
6968 high = total_pages / nr_split_cpus;
6971 * Ensure high is at least batch*4. The multiple is based on the
6972 * historical relationship between high and batch.
6974 high = max(high, batch << 2);
6983 * pcp->high and pcp->batch values are related and generally batch is lower
6984 * than high. They are also related to pcp->count such that count is lower
6985 * than high, and as soon as it reaches high, the pcplist is flushed.
6987 * However, guaranteeing these relations at all times would require e.g. write
6988 * barriers here but also careful usage of read barriers at the read side, and
6989 * thus be prone to error and bad for performance. Thus the update only prevents
6990 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6991 * can cope with those fields changing asynchronously, and fully trust only the
6992 * pcp->count field on the local CPU with interrupts disabled.
6994 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6995 * outside of boot time (or some other assurance that no concurrent updaters
6998 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6999 unsigned long batch)
7001 WRITE_ONCE(pcp->batch, batch);
7002 WRITE_ONCE(pcp->high, high);
7005 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
7009 memset(pcp, 0, sizeof(*pcp));
7010 memset(pzstats, 0, sizeof(*pzstats));
7012 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
7013 INIT_LIST_HEAD(&pcp->lists[pindex]);
7016 * Set batch and high values safe for a boot pageset. A true percpu
7017 * pageset's initialization will update them subsequently. Here we don't
7018 * need to be as careful as pageset_update() as nobody can access the
7021 pcp->high = BOOT_PAGESET_HIGH;
7022 pcp->batch = BOOT_PAGESET_BATCH;
7023 pcp->free_factor = 0;
7026 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
7027 unsigned long batch)
7029 struct per_cpu_pages *pcp;
7032 for_each_possible_cpu(cpu) {
7033 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7034 pageset_update(pcp, high, batch);
7039 * Calculate and set new high and batch values for all per-cpu pagesets of a
7040 * zone based on the zone's size.
7042 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
7044 int new_high, new_batch;
7046 new_batch = max(1, zone_batchsize(zone));
7047 new_high = zone_highsize(zone, new_batch, cpu_online);
7049 if (zone->pageset_high == new_high &&
7050 zone->pageset_batch == new_batch)
7053 zone->pageset_high = new_high;
7054 zone->pageset_batch = new_batch;
7056 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7059 void __meminit setup_zone_pageset(struct zone *zone)
7063 /* Size may be 0 on !SMP && !NUMA */
7064 if (sizeof(struct per_cpu_zonestat) > 0)
7065 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7067 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7068 for_each_possible_cpu(cpu) {
7069 struct per_cpu_pages *pcp;
7070 struct per_cpu_zonestat *pzstats;
7072 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7073 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7074 per_cpu_pages_init(pcp, pzstats);
7077 zone_set_pageset_high_and_batch(zone, 0);
7081 * Allocate per cpu pagesets and initialize them.
7082 * Before this call only boot pagesets were available.
7084 void __init setup_per_cpu_pageset(void)
7086 struct pglist_data *pgdat;
7088 int __maybe_unused cpu;
7090 for_each_populated_zone(zone)
7091 setup_zone_pageset(zone);
7095 * Unpopulated zones continue using the boot pagesets.
7096 * The numa stats for these pagesets need to be reset.
7097 * Otherwise, they will end up skewing the stats of
7098 * the nodes these zones are associated with.
7100 for_each_possible_cpu(cpu) {
7101 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7102 memset(pzstats->vm_numa_event, 0,
7103 sizeof(pzstats->vm_numa_event));
7107 for_each_online_pgdat(pgdat)
7108 pgdat->per_cpu_nodestats =
7109 alloc_percpu(struct per_cpu_nodestat);
7112 static __meminit void zone_pcp_init(struct zone *zone)
7115 * per cpu subsystem is not up at this point. The following code
7116 * relies on the ability of the linker to provide the
7117 * offset of a (static) per cpu variable into the per cpu area.
7119 zone->per_cpu_pageset = &boot_pageset;
7120 zone->per_cpu_zonestats = &boot_zonestats;
7121 zone->pageset_high = BOOT_PAGESET_HIGH;
7122 zone->pageset_batch = BOOT_PAGESET_BATCH;
7124 if (populated_zone(zone))
7125 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7126 zone->present_pages, zone_batchsize(zone));
7129 void __meminit init_currently_empty_zone(struct zone *zone,
7130 unsigned long zone_start_pfn,
7133 struct pglist_data *pgdat = zone->zone_pgdat;
7134 int zone_idx = zone_idx(zone) + 1;
7136 if (zone_idx > pgdat->nr_zones)
7137 pgdat->nr_zones = zone_idx;
7139 zone->zone_start_pfn = zone_start_pfn;
7141 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7142 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7144 (unsigned long)zone_idx(zone),
7145 zone_start_pfn, (zone_start_pfn + size));
7147 zone_init_free_lists(zone);
7148 zone->initialized = 1;
7152 * get_pfn_range_for_nid - Return the start and end page frames for a node
7153 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7154 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7155 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7157 * It returns the start and end page frame of a node based on information
7158 * provided by memblock_set_node(). If called for a node
7159 * with no available memory, a warning is printed and the start and end
7162 void __init get_pfn_range_for_nid(unsigned int nid,
7163 unsigned long *start_pfn, unsigned long *end_pfn)
7165 unsigned long this_start_pfn, this_end_pfn;
7171 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7172 *start_pfn = min(*start_pfn, this_start_pfn);
7173 *end_pfn = max(*end_pfn, this_end_pfn);
7176 if (*start_pfn == -1UL)
7181 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7182 * assumption is made that zones within a node are ordered in monotonic
7183 * increasing memory addresses so that the "highest" populated zone is used
7185 static void __init find_usable_zone_for_movable(void)
7188 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7189 if (zone_index == ZONE_MOVABLE)
7192 if (arch_zone_highest_possible_pfn[zone_index] >
7193 arch_zone_lowest_possible_pfn[zone_index])
7197 VM_BUG_ON(zone_index == -1);
7198 movable_zone = zone_index;
7202 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7203 * because it is sized independent of architecture. Unlike the other zones,
7204 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7205 * in each node depending on the size of each node and how evenly kernelcore
7206 * is distributed. This helper function adjusts the zone ranges
7207 * provided by the architecture for a given node by using the end of the
7208 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7209 * zones within a node are in order of monotonic increases memory addresses
7211 static void __init adjust_zone_range_for_zone_movable(int nid,
7212 unsigned long zone_type,
7213 unsigned long node_start_pfn,
7214 unsigned long node_end_pfn,
7215 unsigned long *zone_start_pfn,
7216 unsigned long *zone_end_pfn)
7218 /* Only adjust if ZONE_MOVABLE is on this node */
7219 if (zone_movable_pfn[nid]) {
7220 /* Size ZONE_MOVABLE */
7221 if (zone_type == ZONE_MOVABLE) {
7222 *zone_start_pfn = zone_movable_pfn[nid];
7223 *zone_end_pfn = min(node_end_pfn,
7224 arch_zone_highest_possible_pfn[movable_zone]);
7226 /* Adjust for ZONE_MOVABLE starting within this range */
7227 } else if (!mirrored_kernelcore &&
7228 *zone_start_pfn < zone_movable_pfn[nid] &&
7229 *zone_end_pfn > zone_movable_pfn[nid]) {
7230 *zone_end_pfn = zone_movable_pfn[nid];
7232 /* Check if this whole range is within ZONE_MOVABLE */
7233 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7234 *zone_start_pfn = *zone_end_pfn;
7239 * Return the number of pages a zone spans in a node, including holes
7240 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7242 static unsigned long __init zone_spanned_pages_in_node(int nid,
7243 unsigned long zone_type,
7244 unsigned long node_start_pfn,
7245 unsigned long node_end_pfn,
7246 unsigned long *zone_start_pfn,
7247 unsigned long *zone_end_pfn)
7249 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7250 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7251 /* When hotadd a new node from cpu_up(), the node should be empty */
7252 if (!node_start_pfn && !node_end_pfn)
7255 /* Get the start and end of the zone */
7256 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7257 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7258 adjust_zone_range_for_zone_movable(nid, zone_type,
7259 node_start_pfn, node_end_pfn,
7260 zone_start_pfn, zone_end_pfn);
7262 /* Check that this node has pages within the zone's required range */
7263 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7266 /* Move the zone boundaries inside the node if necessary */
7267 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7268 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7270 /* Return the spanned pages */
7271 return *zone_end_pfn - *zone_start_pfn;
7275 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7276 * then all holes in the requested range will be accounted for.
7278 unsigned long __init __absent_pages_in_range(int nid,
7279 unsigned long range_start_pfn,
7280 unsigned long range_end_pfn)
7282 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7283 unsigned long start_pfn, end_pfn;
7286 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7287 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7288 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7289 nr_absent -= end_pfn - start_pfn;
7295 * absent_pages_in_range - Return number of page frames in holes within a range
7296 * @start_pfn: The start PFN to start searching for holes
7297 * @end_pfn: The end PFN to stop searching for holes
7299 * Return: the number of pages frames in memory holes within a range.
7301 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7302 unsigned long end_pfn)
7304 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7307 /* Return the number of page frames in holes in a zone on a node */
7308 static unsigned long __init zone_absent_pages_in_node(int nid,
7309 unsigned long zone_type,
7310 unsigned long node_start_pfn,
7311 unsigned long node_end_pfn)
7313 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7314 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7315 unsigned long zone_start_pfn, zone_end_pfn;
7316 unsigned long nr_absent;
7318 /* When hotadd a new node from cpu_up(), the node should be empty */
7319 if (!node_start_pfn && !node_end_pfn)
7322 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7323 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7325 adjust_zone_range_for_zone_movable(nid, zone_type,
7326 node_start_pfn, node_end_pfn,
7327 &zone_start_pfn, &zone_end_pfn);
7328 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7331 * ZONE_MOVABLE handling.
7332 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7335 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7336 unsigned long start_pfn, end_pfn;
7337 struct memblock_region *r;
7339 for_each_mem_region(r) {
7340 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7341 zone_start_pfn, zone_end_pfn);
7342 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7343 zone_start_pfn, zone_end_pfn);
7345 if (zone_type == ZONE_MOVABLE &&
7346 memblock_is_mirror(r))
7347 nr_absent += end_pfn - start_pfn;
7349 if (zone_type == ZONE_NORMAL &&
7350 !memblock_is_mirror(r))
7351 nr_absent += end_pfn - start_pfn;
7358 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7359 unsigned long node_start_pfn,
7360 unsigned long node_end_pfn)
7362 unsigned long realtotalpages = 0, totalpages = 0;
7365 for (i = 0; i < MAX_NR_ZONES; i++) {
7366 struct zone *zone = pgdat->node_zones + i;
7367 unsigned long zone_start_pfn, zone_end_pfn;
7368 unsigned long spanned, absent;
7369 unsigned long size, real_size;
7371 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7376 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7381 real_size = size - absent;
7384 zone->zone_start_pfn = zone_start_pfn;
7386 zone->zone_start_pfn = 0;
7387 zone->spanned_pages = size;
7388 zone->present_pages = real_size;
7389 #if defined(CONFIG_MEMORY_HOTPLUG)
7390 zone->present_early_pages = real_size;
7394 realtotalpages += real_size;
7397 pgdat->node_spanned_pages = totalpages;
7398 pgdat->node_present_pages = realtotalpages;
7399 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7402 #ifndef CONFIG_SPARSEMEM
7404 * Calculate the size of the zone->blockflags rounded to an unsigned long
7405 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7406 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7407 * round what is now in bits to nearest long in bits, then return it in
7410 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7412 unsigned long usemapsize;
7414 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7415 usemapsize = roundup(zonesize, pageblock_nr_pages);
7416 usemapsize = usemapsize >> pageblock_order;
7417 usemapsize *= NR_PAGEBLOCK_BITS;
7418 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7420 return usemapsize / 8;
7423 static void __ref setup_usemap(struct zone *zone)
7425 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7426 zone->spanned_pages);
7427 zone->pageblock_flags = NULL;
7429 zone->pageblock_flags =
7430 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7432 if (!zone->pageblock_flags)
7433 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7434 usemapsize, zone->name, zone_to_nid(zone));
7438 static inline void setup_usemap(struct zone *zone) {}
7439 #endif /* CONFIG_SPARSEMEM */
7441 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7443 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7444 void __init set_pageblock_order(void)
7446 unsigned int order = MAX_ORDER - 1;
7448 /* Check that pageblock_nr_pages has not already been setup */
7449 if (pageblock_order)
7452 /* Don't let pageblocks exceed the maximum allocation granularity. */
7453 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7454 order = HUGETLB_PAGE_ORDER;
7457 * Assume the largest contiguous order of interest is a huge page.
7458 * This value may be variable depending on boot parameters on IA64 and
7461 pageblock_order = order;
7463 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7466 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7467 * is unused as pageblock_order is set at compile-time. See
7468 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7471 void __init set_pageblock_order(void)
7475 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7477 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7478 unsigned long present_pages)
7480 unsigned long pages = spanned_pages;
7483 * Provide a more accurate estimation if there are holes within
7484 * the zone and SPARSEMEM is in use. If there are holes within the
7485 * zone, each populated memory region may cost us one or two extra
7486 * memmap pages due to alignment because memmap pages for each
7487 * populated regions may not be naturally aligned on page boundary.
7488 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7490 if (spanned_pages > present_pages + (present_pages >> 4) &&
7491 IS_ENABLED(CONFIG_SPARSEMEM))
7492 pages = present_pages;
7494 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7497 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7498 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7500 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7502 spin_lock_init(&ds_queue->split_queue_lock);
7503 INIT_LIST_HEAD(&ds_queue->split_queue);
7504 ds_queue->split_queue_len = 0;
7507 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7510 #ifdef CONFIG_COMPACTION
7511 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7513 init_waitqueue_head(&pgdat->kcompactd_wait);
7516 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7519 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7523 pgdat_resize_init(pgdat);
7525 pgdat_init_split_queue(pgdat);
7526 pgdat_init_kcompactd(pgdat);
7528 init_waitqueue_head(&pgdat->kswapd_wait);
7529 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7531 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7532 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7534 pgdat_page_ext_init(pgdat);
7535 lruvec_init(&pgdat->__lruvec);
7538 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7539 unsigned long remaining_pages)
7541 atomic_long_set(&zone->managed_pages, remaining_pages);
7542 zone_set_nid(zone, nid);
7543 zone->name = zone_names[idx];
7544 zone->zone_pgdat = NODE_DATA(nid);
7545 spin_lock_init(&zone->lock);
7546 zone_seqlock_init(zone);
7547 zone_pcp_init(zone);
7551 * Set up the zone data structures
7552 * - init pgdat internals
7553 * - init all zones belonging to this node
7555 * NOTE: this function is only called during memory hotplug
7557 #ifdef CONFIG_MEMORY_HOTPLUG
7558 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7560 int nid = pgdat->node_id;
7564 pgdat_init_internals(pgdat);
7566 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7567 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7570 * Reset the nr_zones, order and highest_zoneidx before reuse.
7571 * Note that kswapd will init kswapd_highest_zoneidx properly
7572 * when it starts in the near future.
7574 pgdat->nr_zones = 0;
7575 pgdat->kswapd_order = 0;
7576 pgdat->kswapd_highest_zoneidx = 0;
7577 pgdat->node_start_pfn = 0;
7578 for_each_online_cpu(cpu) {
7579 struct per_cpu_nodestat *p;
7581 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7582 memset(p, 0, sizeof(*p));
7585 for (z = 0; z < MAX_NR_ZONES; z++)
7586 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7591 * Set up the zone data structures:
7592 * - mark all pages reserved
7593 * - mark all memory queues empty
7594 * - clear the memory bitmaps
7596 * NOTE: pgdat should get zeroed by caller.
7597 * NOTE: this function is only called during early init.
7599 static void __init free_area_init_core(struct pglist_data *pgdat)
7602 int nid = pgdat->node_id;
7604 pgdat_init_internals(pgdat);
7605 pgdat->per_cpu_nodestats = &boot_nodestats;
7607 for (j = 0; j < MAX_NR_ZONES; j++) {
7608 struct zone *zone = pgdat->node_zones + j;
7609 unsigned long size, freesize, memmap_pages;
7611 size = zone->spanned_pages;
7612 freesize = zone->present_pages;
7615 * Adjust freesize so that it accounts for how much memory
7616 * is used by this zone for memmap. This affects the watermark
7617 * and per-cpu initialisations
7619 memmap_pages = calc_memmap_size(size, freesize);
7620 if (!is_highmem_idx(j)) {
7621 if (freesize >= memmap_pages) {
7622 freesize -= memmap_pages;
7624 pr_debug(" %s zone: %lu pages used for memmap\n",
7625 zone_names[j], memmap_pages);
7627 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7628 zone_names[j], memmap_pages, freesize);
7631 /* Account for reserved pages */
7632 if (j == 0 && freesize > dma_reserve) {
7633 freesize -= dma_reserve;
7634 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7637 if (!is_highmem_idx(j))
7638 nr_kernel_pages += freesize;
7639 /* Charge for highmem memmap if there are enough kernel pages */
7640 else if (nr_kernel_pages > memmap_pages * 2)
7641 nr_kernel_pages -= memmap_pages;
7642 nr_all_pages += freesize;
7645 * Set an approximate value for lowmem here, it will be adjusted
7646 * when the bootmem allocator frees pages into the buddy system.
7647 * And all highmem pages will be managed by the buddy system.
7649 zone_init_internals(zone, j, nid, freesize);
7654 set_pageblock_order();
7656 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7660 #ifdef CONFIG_FLATMEM
7661 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7663 unsigned long __maybe_unused start = 0;
7664 unsigned long __maybe_unused offset = 0;
7666 /* Skip empty nodes */
7667 if (!pgdat->node_spanned_pages)
7670 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7671 offset = pgdat->node_start_pfn - start;
7672 /* ia64 gets its own node_mem_map, before this, without bootmem */
7673 if (!pgdat->node_mem_map) {
7674 unsigned long size, end;
7678 * The zone's endpoints aren't required to be MAX_ORDER
7679 * aligned but the node_mem_map endpoints must be in order
7680 * for the buddy allocator to function correctly.
7682 end = pgdat_end_pfn(pgdat);
7683 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7684 size = (end - start) * sizeof(struct page);
7685 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7686 pgdat->node_id, false);
7688 panic("Failed to allocate %ld bytes for node %d memory map\n",
7689 size, pgdat->node_id);
7690 pgdat->node_mem_map = map + offset;
7692 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7693 __func__, pgdat->node_id, (unsigned long)pgdat,
7694 (unsigned long)pgdat->node_mem_map);
7697 * With no DISCONTIG, the global mem_map is just set as node 0's
7699 if (pgdat == NODE_DATA(0)) {
7700 mem_map = NODE_DATA(0)->node_mem_map;
7701 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7707 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7708 #endif /* CONFIG_FLATMEM */
7710 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7711 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7713 pgdat->first_deferred_pfn = ULONG_MAX;
7716 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7719 static void __init free_area_init_node(int nid)
7721 pg_data_t *pgdat = NODE_DATA(nid);
7722 unsigned long start_pfn = 0;
7723 unsigned long end_pfn = 0;
7725 /* pg_data_t should be reset to zero when it's allocated */
7726 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7728 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7730 pgdat->node_id = nid;
7731 pgdat->node_start_pfn = start_pfn;
7732 pgdat->per_cpu_nodestats = NULL;
7734 if (start_pfn != end_pfn) {
7735 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7736 (u64)start_pfn << PAGE_SHIFT,
7737 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7739 pr_info("Initmem setup node %d as memoryless\n", nid);
7742 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7744 alloc_node_mem_map(pgdat);
7745 pgdat_set_deferred_range(pgdat);
7747 free_area_init_core(pgdat);
7750 static void __init free_area_init_memoryless_node(int nid)
7752 free_area_init_node(nid);
7755 #if MAX_NUMNODES > 1
7757 * Figure out the number of possible node ids.
7759 void __init setup_nr_node_ids(void)
7761 unsigned int highest;
7763 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7764 nr_node_ids = highest + 1;
7769 * node_map_pfn_alignment - determine the maximum internode alignment
7771 * This function should be called after node map is populated and sorted.
7772 * It calculates the maximum power of two alignment which can distinguish
7775 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7776 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7777 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7778 * shifted, 1GiB is enough and this function will indicate so.
7780 * This is used to test whether pfn -> nid mapping of the chosen memory
7781 * model has fine enough granularity to avoid incorrect mapping for the
7782 * populated node map.
7784 * Return: the determined alignment in pfn's. 0 if there is no alignment
7785 * requirement (single node).
7787 unsigned long __init node_map_pfn_alignment(void)
7789 unsigned long accl_mask = 0, last_end = 0;
7790 unsigned long start, end, mask;
7791 int last_nid = NUMA_NO_NODE;
7794 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7795 if (!start || last_nid < 0 || last_nid == nid) {
7802 * Start with a mask granular enough to pin-point to the
7803 * start pfn and tick off bits one-by-one until it becomes
7804 * too coarse to separate the current node from the last.
7806 mask = ~((1 << __ffs(start)) - 1);
7807 while (mask && last_end <= (start & (mask << 1)))
7810 /* accumulate all internode masks */
7814 /* convert mask to number of pages */
7815 return ~accl_mask + 1;
7819 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7821 * Return: the minimum PFN based on information provided via
7822 * memblock_set_node().
7824 unsigned long __init find_min_pfn_with_active_regions(void)
7826 return PHYS_PFN(memblock_start_of_DRAM());
7830 * early_calculate_totalpages()
7831 * Sum pages in active regions for movable zone.
7832 * Populate N_MEMORY for calculating usable_nodes.
7834 static unsigned long __init early_calculate_totalpages(void)
7836 unsigned long totalpages = 0;
7837 unsigned long start_pfn, end_pfn;
7840 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7841 unsigned long pages = end_pfn - start_pfn;
7843 totalpages += pages;
7845 node_set_state(nid, N_MEMORY);
7851 * Find the PFN the Movable zone begins in each node. Kernel memory
7852 * is spread evenly between nodes as long as the nodes have enough
7853 * memory. When they don't, some nodes will have more kernelcore than
7856 static void __init find_zone_movable_pfns_for_nodes(void)
7859 unsigned long usable_startpfn;
7860 unsigned long kernelcore_node, kernelcore_remaining;
7861 /* save the state before borrow the nodemask */
7862 nodemask_t saved_node_state = node_states[N_MEMORY];
7863 unsigned long totalpages = early_calculate_totalpages();
7864 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7865 struct memblock_region *r;
7867 /* Need to find movable_zone earlier when movable_node is specified. */
7868 find_usable_zone_for_movable();
7871 * If movable_node is specified, ignore kernelcore and movablecore
7874 if (movable_node_is_enabled()) {
7875 for_each_mem_region(r) {
7876 if (!memblock_is_hotpluggable(r))
7879 nid = memblock_get_region_node(r);
7881 usable_startpfn = PFN_DOWN(r->base);
7882 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7883 min(usable_startpfn, zone_movable_pfn[nid]) :
7891 * If kernelcore=mirror is specified, ignore movablecore option
7893 if (mirrored_kernelcore) {
7894 bool mem_below_4gb_not_mirrored = false;
7896 for_each_mem_region(r) {
7897 if (memblock_is_mirror(r))
7900 nid = memblock_get_region_node(r);
7902 usable_startpfn = memblock_region_memory_base_pfn(r);
7904 if (usable_startpfn < PHYS_PFN(SZ_4G)) {
7905 mem_below_4gb_not_mirrored = true;
7909 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7910 min(usable_startpfn, zone_movable_pfn[nid]) :
7914 if (mem_below_4gb_not_mirrored)
7915 pr_warn("This configuration results in unmirrored kernel memory.\n");
7921 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7922 * amount of necessary memory.
7924 if (required_kernelcore_percent)
7925 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7927 if (required_movablecore_percent)
7928 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7932 * If movablecore= was specified, calculate what size of
7933 * kernelcore that corresponds so that memory usable for
7934 * any allocation type is evenly spread. If both kernelcore
7935 * and movablecore are specified, then the value of kernelcore
7936 * will be used for required_kernelcore if it's greater than
7937 * what movablecore would have allowed.
7939 if (required_movablecore) {
7940 unsigned long corepages;
7943 * Round-up so that ZONE_MOVABLE is at least as large as what
7944 * was requested by the user
7946 required_movablecore =
7947 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7948 required_movablecore = min(totalpages, required_movablecore);
7949 corepages = totalpages - required_movablecore;
7951 required_kernelcore = max(required_kernelcore, corepages);
7955 * If kernelcore was not specified or kernelcore size is larger
7956 * than totalpages, there is no ZONE_MOVABLE.
7958 if (!required_kernelcore || required_kernelcore >= totalpages)
7961 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7962 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7965 /* Spread kernelcore memory as evenly as possible throughout nodes */
7966 kernelcore_node = required_kernelcore / usable_nodes;
7967 for_each_node_state(nid, N_MEMORY) {
7968 unsigned long start_pfn, end_pfn;
7971 * Recalculate kernelcore_node if the division per node
7972 * now exceeds what is necessary to satisfy the requested
7973 * amount of memory for the kernel
7975 if (required_kernelcore < kernelcore_node)
7976 kernelcore_node = required_kernelcore / usable_nodes;
7979 * As the map is walked, we track how much memory is usable
7980 * by the kernel using kernelcore_remaining. When it is
7981 * 0, the rest of the node is usable by ZONE_MOVABLE
7983 kernelcore_remaining = kernelcore_node;
7985 /* Go through each range of PFNs within this node */
7986 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7987 unsigned long size_pages;
7989 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7990 if (start_pfn >= end_pfn)
7993 /* Account for what is only usable for kernelcore */
7994 if (start_pfn < usable_startpfn) {
7995 unsigned long kernel_pages;
7996 kernel_pages = min(end_pfn, usable_startpfn)
7999 kernelcore_remaining -= min(kernel_pages,
8000 kernelcore_remaining);
8001 required_kernelcore -= min(kernel_pages,
8002 required_kernelcore);
8004 /* Continue if range is now fully accounted */
8005 if (end_pfn <= usable_startpfn) {
8008 * Push zone_movable_pfn to the end so
8009 * that if we have to rebalance
8010 * kernelcore across nodes, we will
8011 * not double account here
8013 zone_movable_pfn[nid] = end_pfn;
8016 start_pfn = usable_startpfn;
8020 * The usable PFN range for ZONE_MOVABLE is from
8021 * start_pfn->end_pfn. Calculate size_pages as the
8022 * number of pages used as kernelcore
8024 size_pages = end_pfn - start_pfn;
8025 if (size_pages > kernelcore_remaining)
8026 size_pages = kernelcore_remaining;
8027 zone_movable_pfn[nid] = start_pfn + size_pages;
8030 * Some kernelcore has been met, update counts and
8031 * break if the kernelcore for this node has been
8034 required_kernelcore -= min(required_kernelcore,
8036 kernelcore_remaining -= size_pages;
8037 if (!kernelcore_remaining)
8043 * If there is still required_kernelcore, we do another pass with one
8044 * less node in the count. This will push zone_movable_pfn[nid] further
8045 * along on the nodes that still have memory until kernelcore is
8049 if (usable_nodes && required_kernelcore > usable_nodes)
8053 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
8054 for (nid = 0; nid < MAX_NUMNODES; nid++) {
8055 unsigned long start_pfn, end_pfn;
8057 zone_movable_pfn[nid] =
8058 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
8060 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
8061 if (zone_movable_pfn[nid] >= end_pfn)
8062 zone_movable_pfn[nid] = 0;
8066 /* restore the node_state */
8067 node_states[N_MEMORY] = saved_node_state;
8070 /* Any regular or high memory on that node ? */
8071 static void check_for_memory(pg_data_t *pgdat, int nid)
8073 enum zone_type zone_type;
8075 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8076 struct zone *zone = &pgdat->node_zones[zone_type];
8077 if (populated_zone(zone)) {
8078 if (IS_ENABLED(CONFIG_HIGHMEM))
8079 node_set_state(nid, N_HIGH_MEMORY);
8080 if (zone_type <= ZONE_NORMAL)
8081 node_set_state(nid, N_NORMAL_MEMORY);
8088 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8089 * such cases we allow max_zone_pfn sorted in the descending order
8091 bool __weak arch_has_descending_max_zone_pfns(void)
8097 * free_area_init - Initialise all pg_data_t and zone data
8098 * @max_zone_pfn: an array of max PFNs for each zone
8100 * This will call free_area_init_node() for each active node in the system.
8101 * Using the page ranges provided by memblock_set_node(), the size of each
8102 * zone in each node and their holes is calculated. If the maximum PFN
8103 * between two adjacent zones match, it is assumed that the zone is empty.
8104 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8105 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8106 * starts where the previous one ended. For example, ZONE_DMA32 starts
8107 * at arch_max_dma_pfn.
8109 void __init free_area_init(unsigned long *max_zone_pfn)
8111 unsigned long start_pfn, end_pfn;
8115 /* Record where the zone boundaries are */
8116 memset(arch_zone_lowest_possible_pfn, 0,
8117 sizeof(arch_zone_lowest_possible_pfn));
8118 memset(arch_zone_highest_possible_pfn, 0,
8119 sizeof(arch_zone_highest_possible_pfn));
8121 start_pfn = find_min_pfn_with_active_regions();
8122 descending = arch_has_descending_max_zone_pfns();
8124 for (i = 0; i < MAX_NR_ZONES; i++) {
8126 zone = MAX_NR_ZONES - i - 1;
8130 if (zone == ZONE_MOVABLE)
8133 end_pfn = max(max_zone_pfn[zone], start_pfn);
8134 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8135 arch_zone_highest_possible_pfn[zone] = end_pfn;
8137 start_pfn = end_pfn;
8140 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8141 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8142 find_zone_movable_pfns_for_nodes();
8144 /* Print out the zone ranges */
8145 pr_info("Zone ranges:\n");
8146 for (i = 0; i < MAX_NR_ZONES; i++) {
8147 if (i == ZONE_MOVABLE)
8149 pr_info(" %-8s ", zone_names[i]);
8150 if (arch_zone_lowest_possible_pfn[i] ==
8151 arch_zone_highest_possible_pfn[i])
8154 pr_cont("[mem %#018Lx-%#018Lx]\n",
8155 (u64)arch_zone_lowest_possible_pfn[i]
8157 ((u64)arch_zone_highest_possible_pfn[i]
8158 << PAGE_SHIFT) - 1);
8161 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8162 pr_info("Movable zone start for each node\n");
8163 for (i = 0; i < MAX_NUMNODES; i++) {
8164 if (zone_movable_pfn[i])
8165 pr_info(" Node %d: %#018Lx\n", i,
8166 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8170 * Print out the early node map, and initialize the
8171 * subsection-map relative to active online memory ranges to
8172 * enable future "sub-section" extensions of the memory map.
8174 pr_info("Early memory node ranges\n");
8175 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8176 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8177 (u64)start_pfn << PAGE_SHIFT,
8178 ((u64)end_pfn << PAGE_SHIFT) - 1);
8179 subsection_map_init(start_pfn, end_pfn - start_pfn);
8182 /* Initialise every node */
8183 mminit_verify_pageflags_layout();
8184 setup_nr_node_ids();
8185 for_each_node(nid) {
8188 if (!node_online(nid)) {
8189 pr_info("Initializing node %d as memoryless\n", nid);
8191 /* Allocator not initialized yet */
8192 pgdat = arch_alloc_nodedata(nid);
8194 pr_err("Cannot allocate %zuB for node %d.\n",
8195 sizeof(*pgdat), nid);
8198 arch_refresh_nodedata(nid, pgdat);
8199 free_area_init_memoryless_node(nid);
8202 * We do not want to confuse userspace by sysfs
8203 * files/directories for node without any memory
8204 * attached to it, so this node is not marked as
8205 * N_MEMORY and not marked online so that no sysfs
8206 * hierarchy will be created via register_one_node for
8207 * it. The pgdat will get fully initialized by
8208 * hotadd_init_pgdat() when memory is hotplugged into
8214 pgdat = NODE_DATA(nid);
8215 free_area_init_node(nid);
8217 /* Any memory on that node */
8218 if (pgdat->node_present_pages)
8219 node_set_state(nid, N_MEMORY);
8220 check_for_memory(pgdat, nid);
8226 static int __init cmdline_parse_core(char *p, unsigned long *core,
8227 unsigned long *percent)
8229 unsigned long long coremem;
8235 /* Value may be a percentage of total memory, otherwise bytes */
8236 coremem = simple_strtoull(p, &endptr, 0);
8237 if (*endptr == '%') {
8238 /* Paranoid check for percent values greater than 100 */
8239 WARN_ON(coremem > 100);
8243 coremem = memparse(p, &p);
8244 /* Paranoid check that UL is enough for the coremem value */
8245 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8247 *core = coremem >> PAGE_SHIFT;
8254 * kernelcore=size sets the amount of memory for use for allocations that
8255 * cannot be reclaimed or migrated.
8257 static int __init cmdline_parse_kernelcore(char *p)
8259 /* parse kernelcore=mirror */
8260 if (parse_option_str(p, "mirror")) {
8261 mirrored_kernelcore = true;
8265 return cmdline_parse_core(p, &required_kernelcore,
8266 &required_kernelcore_percent);
8270 * movablecore=size sets the amount of memory for use for allocations that
8271 * can be reclaimed or migrated.
8273 static int __init cmdline_parse_movablecore(char *p)
8275 return cmdline_parse_core(p, &required_movablecore,
8276 &required_movablecore_percent);
8279 early_param("kernelcore", cmdline_parse_kernelcore);
8280 early_param("movablecore", cmdline_parse_movablecore);
8282 void adjust_managed_page_count(struct page *page, long count)
8284 atomic_long_add(count, &page_zone(page)->managed_pages);
8285 totalram_pages_add(count);
8286 #ifdef CONFIG_HIGHMEM
8287 if (PageHighMem(page))
8288 totalhigh_pages_add(count);
8291 EXPORT_SYMBOL(adjust_managed_page_count);
8293 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8296 unsigned long pages = 0;
8298 start = (void *)PAGE_ALIGN((unsigned long)start);
8299 end = (void *)((unsigned long)end & PAGE_MASK);
8300 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8301 struct page *page = virt_to_page(pos);
8302 void *direct_map_addr;
8305 * 'direct_map_addr' might be different from 'pos'
8306 * because some architectures' virt_to_page()
8307 * work with aliases. Getting the direct map
8308 * address ensures that we get a _writeable_
8309 * alias for the memset().
8311 direct_map_addr = page_address(page);
8313 * Perform a kasan-unchecked memset() since this memory
8314 * has not been initialized.
8316 direct_map_addr = kasan_reset_tag(direct_map_addr);
8317 if ((unsigned int)poison <= 0xFF)
8318 memset(direct_map_addr, poison, PAGE_SIZE);
8320 free_reserved_page(page);
8324 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8329 void __init mem_init_print_info(void)
8331 unsigned long physpages, codesize, datasize, rosize, bss_size;
8332 unsigned long init_code_size, init_data_size;
8334 physpages = get_num_physpages();
8335 codesize = _etext - _stext;
8336 datasize = _edata - _sdata;
8337 rosize = __end_rodata - __start_rodata;
8338 bss_size = __bss_stop - __bss_start;
8339 init_data_size = __init_end - __init_begin;
8340 init_code_size = _einittext - _sinittext;
8343 * Detect special cases and adjust section sizes accordingly:
8344 * 1) .init.* may be embedded into .data sections
8345 * 2) .init.text.* may be out of [__init_begin, __init_end],
8346 * please refer to arch/tile/kernel/vmlinux.lds.S.
8347 * 3) .rodata.* may be embedded into .text or .data sections.
8349 #define adj_init_size(start, end, size, pos, adj) \
8351 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8355 adj_init_size(__init_begin, __init_end, init_data_size,
8356 _sinittext, init_code_size);
8357 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8358 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8359 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8360 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8362 #undef adj_init_size
8364 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8365 #ifdef CONFIG_HIGHMEM
8369 K(nr_free_pages()), K(physpages),
8370 codesize >> 10, datasize >> 10, rosize >> 10,
8371 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8372 K(physpages - totalram_pages() - totalcma_pages),
8374 #ifdef CONFIG_HIGHMEM
8375 , K(totalhigh_pages())
8381 * set_dma_reserve - set the specified number of pages reserved in the first zone
8382 * @new_dma_reserve: The number of pages to mark reserved
8384 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8385 * In the DMA zone, a significant percentage may be consumed by kernel image
8386 * and other unfreeable allocations which can skew the watermarks badly. This
8387 * function may optionally be used to account for unfreeable pages in the
8388 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8389 * smaller per-cpu batchsize.
8391 void __init set_dma_reserve(unsigned long new_dma_reserve)
8393 dma_reserve = new_dma_reserve;
8396 static int page_alloc_cpu_dead(unsigned int cpu)
8400 lru_add_drain_cpu(cpu);
8401 mlock_page_drain_remote(cpu);
8405 * Spill the event counters of the dead processor
8406 * into the current processors event counters.
8407 * This artificially elevates the count of the current
8410 vm_events_fold_cpu(cpu);
8413 * Zero the differential counters of the dead processor
8414 * so that the vm statistics are consistent.
8416 * This is only okay since the processor is dead and cannot
8417 * race with what we are doing.
8419 cpu_vm_stats_fold(cpu);
8421 for_each_populated_zone(zone)
8422 zone_pcp_update(zone, 0);
8427 static int page_alloc_cpu_online(unsigned int cpu)
8431 for_each_populated_zone(zone)
8432 zone_pcp_update(zone, 1);
8437 int hashdist = HASHDIST_DEFAULT;
8439 static int __init set_hashdist(char *str)
8443 hashdist = simple_strtoul(str, &str, 0);
8446 __setup("hashdist=", set_hashdist);
8449 void __init page_alloc_init(void)
8454 if (num_node_state(N_MEMORY) == 1)
8458 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8459 "mm/page_alloc:pcp",
8460 page_alloc_cpu_online,
8461 page_alloc_cpu_dead);
8466 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8467 * or min_free_kbytes changes.
8469 static void calculate_totalreserve_pages(void)
8471 struct pglist_data *pgdat;
8472 unsigned long reserve_pages = 0;
8473 enum zone_type i, j;
8475 for_each_online_pgdat(pgdat) {
8477 pgdat->totalreserve_pages = 0;
8479 for (i = 0; i < MAX_NR_ZONES; i++) {
8480 struct zone *zone = pgdat->node_zones + i;
8482 unsigned long managed_pages = zone_managed_pages(zone);
8484 /* Find valid and maximum lowmem_reserve in the zone */
8485 for (j = i; j < MAX_NR_ZONES; j++) {
8486 if (zone->lowmem_reserve[j] > max)
8487 max = zone->lowmem_reserve[j];
8490 /* we treat the high watermark as reserved pages. */
8491 max += high_wmark_pages(zone);
8493 if (max > managed_pages)
8494 max = managed_pages;
8496 pgdat->totalreserve_pages += max;
8498 reserve_pages += max;
8501 totalreserve_pages = reserve_pages;
8505 * setup_per_zone_lowmem_reserve - called whenever
8506 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8507 * has a correct pages reserved value, so an adequate number of
8508 * pages are left in the zone after a successful __alloc_pages().
8510 static void setup_per_zone_lowmem_reserve(void)
8512 struct pglist_data *pgdat;
8513 enum zone_type i, j;
8515 for_each_online_pgdat(pgdat) {
8516 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8517 struct zone *zone = &pgdat->node_zones[i];
8518 int ratio = sysctl_lowmem_reserve_ratio[i];
8519 bool clear = !ratio || !zone_managed_pages(zone);
8520 unsigned long managed_pages = 0;
8522 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8523 struct zone *upper_zone = &pgdat->node_zones[j];
8525 managed_pages += zone_managed_pages(upper_zone);
8528 zone->lowmem_reserve[j] = 0;
8530 zone->lowmem_reserve[j] = managed_pages / ratio;
8535 /* update totalreserve_pages */
8536 calculate_totalreserve_pages();
8539 static void __setup_per_zone_wmarks(void)
8541 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8542 unsigned long lowmem_pages = 0;
8544 unsigned long flags;
8546 /* Calculate total number of !ZONE_HIGHMEM pages */
8547 for_each_zone(zone) {
8548 if (!is_highmem(zone))
8549 lowmem_pages += zone_managed_pages(zone);
8552 for_each_zone(zone) {
8555 spin_lock_irqsave(&zone->lock, flags);
8556 tmp = (u64)pages_min * zone_managed_pages(zone);
8557 do_div(tmp, lowmem_pages);
8558 if (is_highmem(zone)) {
8560 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8561 * need highmem pages, so cap pages_min to a small
8564 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8565 * deltas control async page reclaim, and so should
8566 * not be capped for highmem.
8568 unsigned long min_pages;
8570 min_pages = zone_managed_pages(zone) / 1024;
8571 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8572 zone->_watermark[WMARK_MIN] = min_pages;
8575 * If it's a lowmem zone, reserve a number of pages
8576 * proportionate to the zone's size.
8578 zone->_watermark[WMARK_MIN] = tmp;
8582 * Set the kswapd watermarks distance according to the
8583 * scale factor in proportion to available memory, but
8584 * ensure a minimum size on small systems.
8586 tmp = max_t(u64, tmp >> 2,
8587 mult_frac(zone_managed_pages(zone),
8588 watermark_scale_factor, 10000));
8590 zone->watermark_boost = 0;
8591 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8592 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8593 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8595 spin_unlock_irqrestore(&zone->lock, flags);
8598 /* update totalreserve_pages */
8599 calculate_totalreserve_pages();
8603 * setup_per_zone_wmarks - called when min_free_kbytes changes
8604 * or when memory is hot-{added|removed}
8606 * Ensures that the watermark[min,low,high] values for each zone are set
8607 * correctly with respect to min_free_kbytes.
8609 void setup_per_zone_wmarks(void)
8612 static DEFINE_SPINLOCK(lock);
8615 __setup_per_zone_wmarks();
8619 * The watermark size have changed so update the pcpu batch
8620 * and high limits or the limits may be inappropriate.
8623 zone_pcp_update(zone, 0);
8627 * Initialise min_free_kbytes.
8629 * For small machines we want it small (128k min). For large machines
8630 * we want it large (256MB max). But it is not linear, because network
8631 * bandwidth does not increase linearly with machine size. We use
8633 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8634 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8650 void calculate_min_free_kbytes(void)
8652 unsigned long lowmem_kbytes;
8653 int new_min_free_kbytes;
8655 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8656 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8658 if (new_min_free_kbytes > user_min_free_kbytes)
8659 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8661 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8662 new_min_free_kbytes, user_min_free_kbytes);
8666 int __meminit init_per_zone_wmark_min(void)
8668 calculate_min_free_kbytes();
8669 setup_per_zone_wmarks();
8670 refresh_zone_stat_thresholds();
8671 setup_per_zone_lowmem_reserve();
8674 setup_min_unmapped_ratio();
8675 setup_min_slab_ratio();
8678 khugepaged_min_free_kbytes_update();
8682 postcore_initcall(init_per_zone_wmark_min)
8685 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8686 * that we can call two helper functions whenever min_free_kbytes
8689 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8690 void *buffer, size_t *length, loff_t *ppos)
8694 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8699 user_min_free_kbytes = min_free_kbytes;
8700 setup_per_zone_wmarks();
8705 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8706 void *buffer, size_t *length, loff_t *ppos)
8710 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8715 setup_per_zone_wmarks();
8721 static void setup_min_unmapped_ratio(void)
8726 for_each_online_pgdat(pgdat)
8727 pgdat->min_unmapped_pages = 0;
8730 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8731 sysctl_min_unmapped_ratio) / 100;
8735 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8736 void *buffer, size_t *length, loff_t *ppos)
8740 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8744 setup_min_unmapped_ratio();
8749 static void setup_min_slab_ratio(void)
8754 for_each_online_pgdat(pgdat)
8755 pgdat->min_slab_pages = 0;
8758 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8759 sysctl_min_slab_ratio) / 100;
8762 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8763 void *buffer, size_t *length, loff_t *ppos)
8767 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8771 setup_min_slab_ratio();
8778 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8779 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8780 * whenever sysctl_lowmem_reserve_ratio changes.
8782 * The reserve ratio obviously has absolutely no relation with the
8783 * minimum watermarks. The lowmem reserve ratio can only make sense
8784 * if in function of the boot time zone sizes.
8786 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8787 void *buffer, size_t *length, loff_t *ppos)
8791 proc_dointvec_minmax(table, write, buffer, length, ppos);
8793 for (i = 0; i < MAX_NR_ZONES; i++) {
8794 if (sysctl_lowmem_reserve_ratio[i] < 1)
8795 sysctl_lowmem_reserve_ratio[i] = 0;
8798 setup_per_zone_lowmem_reserve();
8803 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8804 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8805 * pagelist can have before it gets flushed back to buddy allocator.
8807 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8808 int write, void *buffer, size_t *length, loff_t *ppos)
8811 int old_percpu_pagelist_high_fraction;
8814 mutex_lock(&pcp_batch_high_lock);
8815 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8817 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8818 if (!write || ret < 0)
8821 /* Sanity checking to avoid pcp imbalance */
8822 if (percpu_pagelist_high_fraction &&
8823 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8824 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8830 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8833 for_each_populated_zone(zone)
8834 zone_set_pageset_high_and_batch(zone, 0);
8836 mutex_unlock(&pcp_batch_high_lock);
8840 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8842 * Returns the number of pages that arch has reserved but
8843 * is not known to alloc_large_system_hash().
8845 static unsigned long __init arch_reserved_kernel_pages(void)
8852 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8853 * machines. As memory size is increased the scale is also increased but at
8854 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8855 * quadruples the scale is increased by one, which means the size of hash table
8856 * only doubles, instead of quadrupling as well.
8857 * Because 32-bit systems cannot have large physical memory, where this scaling
8858 * makes sense, it is disabled on such platforms.
8860 #if __BITS_PER_LONG > 32
8861 #define ADAPT_SCALE_BASE (64ul << 30)
8862 #define ADAPT_SCALE_SHIFT 2
8863 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8867 * allocate a large system hash table from bootmem
8868 * - it is assumed that the hash table must contain an exact power-of-2
8869 * quantity of entries
8870 * - limit is the number of hash buckets, not the total allocation size
8872 void *__init alloc_large_system_hash(const char *tablename,
8873 unsigned long bucketsize,
8874 unsigned long numentries,
8877 unsigned int *_hash_shift,
8878 unsigned int *_hash_mask,
8879 unsigned long low_limit,
8880 unsigned long high_limit)
8882 unsigned long long max = high_limit;
8883 unsigned long log2qty, size;
8889 /* allow the kernel cmdline to have a say */
8891 /* round applicable memory size up to nearest megabyte */
8892 numentries = nr_kernel_pages;
8893 numentries -= arch_reserved_kernel_pages();
8895 /* It isn't necessary when PAGE_SIZE >= 1MB */
8896 if (PAGE_SHIFT < 20)
8897 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8899 #if __BITS_PER_LONG > 32
8901 unsigned long adapt;
8903 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8904 adapt <<= ADAPT_SCALE_SHIFT)
8909 /* limit to 1 bucket per 2^scale bytes of low memory */
8910 if (scale > PAGE_SHIFT)
8911 numentries >>= (scale - PAGE_SHIFT);
8913 numentries <<= (PAGE_SHIFT - scale);
8915 /* Make sure we've got at least a 0-order allocation.. */
8916 if (unlikely(flags & HASH_SMALL)) {
8917 /* Makes no sense without HASH_EARLY */
8918 WARN_ON(!(flags & HASH_EARLY));
8919 if (!(numentries >> *_hash_shift)) {
8920 numentries = 1UL << *_hash_shift;
8921 BUG_ON(!numentries);
8923 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8924 numentries = PAGE_SIZE / bucketsize;
8926 numentries = roundup_pow_of_two(numentries);
8928 /* limit allocation size to 1/16 total memory by default */
8930 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8931 do_div(max, bucketsize);
8933 max = min(max, 0x80000000ULL);
8935 if (numentries < low_limit)
8936 numentries = low_limit;
8937 if (numentries > max)
8940 log2qty = ilog2(numentries);
8942 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8945 size = bucketsize << log2qty;
8946 if (flags & HASH_EARLY) {
8947 if (flags & HASH_ZERO)
8948 table = memblock_alloc(size, SMP_CACHE_BYTES);
8950 table = memblock_alloc_raw(size,
8952 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8953 table = vmalloc_huge(size, gfp_flags);
8956 huge = is_vm_area_hugepages(table);
8959 * If bucketsize is not a power-of-two, we may free
8960 * some pages at the end of hash table which
8961 * alloc_pages_exact() automatically does
8963 table = alloc_pages_exact(size, gfp_flags);
8964 kmemleak_alloc(table, size, 1, gfp_flags);
8966 } while (!table && size > PAGE_SIZE && --log2qty);
8969 panic("Failed to allocate %s hash table\n", tablename);
8971 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8972 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8973 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8976 *_hash_shift = log2qty;
8978 *_hash_mask = (1 << log2qty) - 1;
8983 #ifdef CONFIG_CONTIG_ALLOC
8984 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8985 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8986 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8987 static void alloc_contig_dump_pages(struct list_head *page_list)
8989 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8991 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8995 list_for_each_entry(page, page_list, lru)
8996 dump_page(page, "migration failure");
9000 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9005 /* [start, end) must belong to a single zone. */
9006 int __alloc_contig_migrate_range(struct compact_control *cc,
9007 unsigned long start, unsigned long end)
9009 /* This function is based on compact_zone() from compaction.c. */
9010 unsigned int nr_reclaimed;
9011 unsigned long pfn = start;
9012 unsigned int tries = 0;
9014 struct migration_target_control mtc = {
9015 .nid = zone_to_nid(cc->zone),
9016 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9019 lru_cache_disable();
9021 while (pfn < end || !list_empty(&cc->migratepages)) {
9022 if (fatal_signal_pending(current)) {
9027 if (list_empty(&cc->migratepages)) {
9028 cc->nr_migratepages = 0;
9029 ret = isolate_migratepages_range(cc, pfn, end);
9030 if (ret && ret != -EAGAIN)
9032 pfn = cc->migrate_pfn;
9034 } else if (++tries == 5) {
9039 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9041 cc->nr_migratepages -= nr_reclaimed;
9043 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9044 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9047 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9048 * to retry again over this error, so do the same here.
9056 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
9057 alloc_contig_dump_pages(&cc->migratepages);
9058 putback_movable_pages(&cc->migratepages);
9065 * alloc_contig_range() -- tries to allocate given range of pages
9066 * @start: start PFN to allocate
9067 * @end: one-past-the-last PFN to allocate
9068 * @migratetype: migratetype of the underlying pageblocks (either
9069 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9070 * in range must have the same migratetype and it must
9071 * be either of the two.
9072 * @gfp_mask: GFP mask to use during compaction
9074 * The PFN range does not have to be pageblock aligned. The PFN range must
9075 * belong to a single zone.
9077 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9078 * pageblocks in the range. Once isolated, the pageblocks should not
9079 * be modified by others.
9081 * Return: zero on success or negative error code. On success all
9082 * pages which PFN is in [start, end) are allocated for the caller and
9083 * need to be freed with free_contig_range().
9085 int alloc_contig_range(unsigned long start, unsigned long end,
9086 unsigned migratetype, gfp_t gfp_mask)
9088 unsigned long outer_start, outer_end;
9092 struct compact_control cc = {
9093 .nr_migratepages = 0,
9095 .zone = page_zone(pfn_to_page(start)),
9096 .mode = MIGRATE_SYNC,
9097 .ignore_skip_hint = true,
9098 .no_set_skip_hint = true,
9099 .gfp_mask = current_gfp_context(gfp_mask),
9100 .alloc_contig = true,
9102 INIT_LIST_HEAD(&cc.migratepages);
9105 * What we do here is we mark all pageblocks in range as
9106 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9107 * have different sizes, and due to the way page allocator
9108 * work, start_isolate_page_range() has special handlings for this.
9110 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9111 * migrate the pages from an unaligned range (ie. pages that
9112 * we are interested in). This will put all the pages in
9113 * range back to page allocator as MIGRATE_ISOLATE.
9115 * When this is done, we take the pages in range from page
9116 * allocator removing them from the buddy system. This way
9117 * page allocator will never consider using them.
9119 * This lets us mark the pageblocks back as
9120 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9121 * aligned range but not in the unaligned, original range are
9122 * put back to page allocator so that buddy can use them.
9125 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
9129 drain_all_pages(cc.zone);
9132 * In case of -EBUSY, we'd like to know which page causes problem.
9133 * So, just fall through. test_pages_isolated() has a tracepoint
9134 * which will report the busy page.
9136 * It is possible that busy pages could become available before
9137 * the call to test_pages_isolated, and the range will actually be
9138 * allocated. So, if we fall through be sure to clear ret so that
9139 * -EBUSY is not accidentally used or returned to caller.
9141 ret = __alloc_contig_migrate_range(&cc, start, end);
9142 if (ret && ret != -EBUSY)
9147 * Pages from [start, end) are within a pageblock_nr_pages
9148 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9149 * more, all pages in [start, end) are free in page allocator.
9150 * What we are going to do is to allocate all pages from
9151 * [start, end) (that is remove them from page allocator).
9153 * The only problem is that pages at the beginning and at the
9154 * end of interesting range may be not aligned with pages that
9155 * page allocator holds, ie. they can be part of higher order
9156 * pages. Because of this, we reserve the bigger range and
9157 * once this is done free the pages we are not interested in.
9159 * We don't have to hold zone->lock here because the pages are
9160 * isolated thus they won't get removed from buddy.
9164 outer_start = start;
9165 while (!PageBuddy(pfn_to_page(outer_start))) {
9166 if (++order >= MAX_ORDER) {
9167 outer_start = start;
9170 outer_start &= ~0UL << order;
9173 if (outer_start != start) {
9174 order = buddy_order(pfn_to_page(outer_start));
9177 * outer_start page could be small order buddy page and
9178 * it doesn't include start page. Adjust outer_start
9179 * in this case to report failed page properly
9180 * on tracepoint in test_pages_isolated()
9182 if (outer_start + (1UL << order) <= start)
9183 outer_start = start;
9186 /* Make sure the range is really isolated. */
9187 if (test_pages_isolated(outer_start, end, 0)) {
9192 /* Grab isolated pages from freelists. */
9193 outer_end = isolate_freepages_range(&cc, outer_start, end);
9199 /* Free head and tail (if any) */
9200 if (start != outer_start)
9201 free_contig_range(outer_start, start - outer_start);
9202 if (end != outer_end)
9203 free_contig_range(end, outer_end - end);
9206 undo_isolate_page_range(start, end, migratetype);
9209 EXPORT_SYMBOL(alloc_contig_range);
9211 static int __alloc_contig_pages(unsigned long start_pfn,
9212 unsigned long nr_pages, gfp_t gfp_mask)
9214 unsigned long end_pfn = start_pfn + nr_pages;
9216 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9220 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9221 unsigned long nr_pages)
9223 unsigned long i, end_pfn = start_pfn + nr_pages;
9226 for (i = start_pfn; i < end_pfn; i++) {
9227 page = pfn_to_online_page(i);
9231 if (page_zone(page) != z)
9234 if (PageReserved(page))
9240 static bool zone_spans_last_pfn(const struct zone *zone,
9241 unsigned long start_pfn, unsigned long nr_pages)
9243 unsigned long last_pfn = start_pfn + nr_pages - 1;
9245 return zone_spans_pfn(zone, last_pfn);
9249 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9250 * @nr_pages: Number of contiguous pages to allocate
9251 * @gfp_mask: GFP mask to limit search and used during compaction
9253 * @nodemask: Mask for other possible nodes
9255 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9256 * on an applicable zonelist to find a contiguous pfn range which can then be
9257 * tried for allocation with alloc_contig_range(). This routine is intended
9258 * for allocation requests which can not be fulfilled with the buddy allocator.
9260 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9261 * power of two, then allocated range is also guaranteed to be aligned to same
9262 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9264 * Allocated pages can be freed with free_contig_range() or by manually calling
9265 * __free_page() on each allocated page.
9267 * Return: pointer to contiguous pages on success, or NULL if not successful.
9269 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9270 int nid, nodemask_t *nodemask)
9272 unsigned long ret, pfn, flags;
9273 struct zonelist *zonelist;
9277 zonelist = node_zonelist(nid, gfp_mask);
9278 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9279 gfp_zone(gfp_mask), nodemask) {
9280 spin_lock_irqsave(&zone->lock, flags);
9282 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9283 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9284 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9286 * We release the zone lock here because
9287 * alloc_contig_range() will also lock the zone
9288 * at some point. If there's an allocation
9289 * spinning on this lock, it may win the race
9290 * and cause alloc_contig_range() to fail...
9292 spin_unlock_irqrestore(&zone->lock, flags);
9293 ret = __alloc_contig_pages(pfn, nr_pages,
9296 return pfn_to_page(pfn);
9297 spin_lock_irqsave(&zone->lock, flags);
9301 spin_unlock_irqrestore(&zone->lock, flags);
9305 #endif /* CONFIG_CONTIG_ALLOC */
9307 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9309 unsigned long count = 0;
9311 for (; nr_pages--; pfn++) {
9312 struct page *page = pfn_to_page(pfn);
9314 count += page_count(page) != 1;
9317 WARN(count != 0, "%lu pages are still in use!\n", count);
9319 EXPORT_SYMBOL(free_contig_range);
9322 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9323 * page high values need to be recalculated.
9325 void zone_pcp_update(struct zone *zone, int cpu_online)
9327 mutex_lock(&pcp_batch_high_lock);
9328 zone_set_pageset_high_and_batch(zone, cpu_online);
9329 mutex_unlock(&pcp_batch_high_lock);
9333 * Effectively disable pcplists for the zone by setting the high limit to 0
9334 * and draining all cpus. A concurrent page freeing on another CPU that's about
9335 * to put the page on pcplist will either finish before the drain and the page
9336 * will be drained, or observe the new high limit and skip the pcplist.
9338 * Must be paired with a call to zone_pcp_enable().
9340 void zone_pcp_disable(struct zone *zone)
9342 mutex_lock(&pcp_batch_high_lock);
9343 __zone_set_pageset_high_and_batch(zone, 0, 1);
9344 __drain_all_pages(zone, true);
9347 void zone_pcp_enable(struct zone *zone)
9349 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9350 mutex_unlock(&pcp_batch_high_lock);
9353 void zone_pcp_reset(struct zone *zone)
9356 struct per_cpu_zonestat *pzstats;
9358 if (zone->per_cpu_pageset != &boot_pageset) {
9359 for_each_online_cpu(cpu) {
9360 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9361 drain_zonestat(zone, pzstats);
9363 free_percpu(zone->per_cpu_pageset);
9364 free_percpu(zone->per_cpu_zonestats);
9365 zone->per_cpu_pageset = &boot_pageset;
9366 zone->per_cpu_zonestats = &boot_zonestats;
9370 #ifdef CONFIG_MEMORY_HOTREMOVE
9372 * All pages in the range must be in a single zone, must not contain holes,
9373 * must span full sections, and must be isolated before calling this function.
9375 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9377 unsigned long pfn = start_pfn;
9381 unsigned long flags;
9383 offline_mem_sections(pfn, end_pfn);
9384 zone = page_zone(pfn_to_page(pfn));
9385 spin_lock_irqsave(&zone->lock, flags);
9386 while (pfn < end_pfn) {
9387 page = pfn_to_page(pfn);
9389 * The HWPoisoned page may be not in buddy system, and
9390 * page_count() is not 0.
9392 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9397 * At this point all remaining PageOffline() pages have a
9398 * reference count of 0 and can simply be skipped.
9400 if (PageOffline(page)) {
9401 BUG_ON(page_count(page));
9402 BUG_ON(PageBuddy(page));
9407 BUG_ON(page_count(page));
9408 BUG_ON(!PageBuddy(page));
9409 order = buddy_order(page);
9410 del_page_from_free_list(page, zone, order);
9411 pfn += (1 << order);
9413 spin_unlock_irqrestore(&zone->lock, flags);
9418 * This function returns a stable result only if called under zone lock.
9420 bool is_free_buddy_page(struct page *page)
9422 unsigned long pfn = page_to_pfn(page);
9425 for (order = 0; order < MAX_ORDER; order++) {
9426 struct page *page_head = page - (pfn & ((1 << order) - 1));
9428 if (PageBuddy(page_head) &&
9429 buddy_order_unsafe(page_head) >= order)
9433 return order < MAX_ORDER;
9435 EXPORT_SYMBOL(is_free_buddy_page);
9437 #ifdef CONFIG_MEMORY_FAILURE
9439 * Break down a higher-order page in sub-pages, and keep our target out of
9442 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9443 struct page *target, int low, int high,
9446 unsigned long size = 1 << high;
9447 struct page *current_buddy, *next_page;
9449 while (high > low) {
9453 if (target >= &page[size]) {
9454 next_page = page + size;
9455 current_buddy = page;
9458 current_buddy = page + size;
9461 if (set_page_guard(zone, current_buddy, high, migratetype))
9464 if (current_buddy != target) {
9465 add_to_free_list(current_buddy, zone, high, migratetype);
9466 set_buddy_order(current_buddy, high);
9473 * Take a page that will be marked as poisoned off the buddy allocator.
9475 bool take_page_off_buddy(struct page *page)
9477 struct zone *zone = page_zone(page);
9478 unsigned long pfn = page_to_pfn(page);
9479 unsigned long flags;
9483 spin_lock_irqsave(&zone->lock, flags);
9484 for (order = 0; order < MAX_ORDER; order++) {
9485 struct page *page_head = page - (pfn & ((1 << order) - 1));
9486 int page_order = buddy_order(page_head);
9488 if (PageBuddy(page_head) && page_order >= order) {
9489 unsigned long pfn_head = page_to_pfn(page_head);
9490 int migratetype = get_pfnblock_migratetype(page_head,
9493 del_page_from_free_list(page_head, zone, page_order);
9494 break_down_buddy_pages(zone, page_head, page, 0,
9495 page_order, migratetype);
9496 SetPageHWPoisonTakenOff(page);
9497 if (!is_migrate_isolate(migratetype))
9498 __mod_zone_freepage_state(zone, -1, migratetype);
9502 if (page_count(page_head) > 0)
9505 spin_unlock_irqrestore(&zone->lock, flags);
9510 * Cancel takeoff done by take_page_off_buddy().
9512 bool put_page_back_buddy(struct page *page)
9514 struct zone *zone = page_zone(page);
9515 unsigned long pfn = page_to_pfn(page);
9516 unsigned long flags;
9517 int migratetype = get_pfnblock_migratetype(page, pfn);
9520 spin_lock_irqsave(&zone->lock, flags);
9521 if (put_page_testzero(page)) {
9522 ClearPageHWPoisonTakenOff(page);
9523 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9524 if (TestClearPageHWPoison(page)) {
9528 spin_unlock_irqrestore(&zone->lock, flags);
9534 #ifdef CONFIG_ZONE_DMA
9535 bool has_managed_dma(void)
9537 struct pglist_data *pgdat;
9539 for_each_online_pgdat(pgdat) {
9540 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9542 if (managed_zone(zone))
9547 #endif /* CONFIG_ZONE_DMA */