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/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/mmu_notifier.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
71 #include <linux/psi.h>
72 #include <linux/padata.h>
73 #include <linux/khugepaged.h>
74 #include <linux/buffer_head.h>
75 #include <linux/delayacct.h>
76 #include <asm/sections.h>
77 #include <asm/tlbflush.h>
78 #include <asm/div64.h>
81 #include "page_reporting.h"
83 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
84 typedef int __bitwise fpi_t;
86 /* No special request */
87 #define FPI_NONE ((__force fpi_t)0)
90 * Skip free page reporting notification for the (possibly merged) page.
91 * This does not hinder free page reporting from grabbing the page,
92 * reporting it and marking it "reported" - it only skips notifying
93 * the free page reporting infrastructure about a newly freed page. For
94 * example, used when temporarily pulling a page from a freelist and
95 * putting it back unmodified.
97 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
100 * Place the (possibly merged) page to the tail of the freelist. Will ignore
101 * page shuffling (relevant code - e.g., memory onlining - is expected to
102 * shuffle the whole zone).
104 * Note: No code should rely on this flag for correctness - it's purely
105 * to allow for optimizations when handing back either fresh pages
106 * (memory onlining) or untouched pages (page isolation, free page
109 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
112 * Don't poison memory with KASAN (only for the tag-based modes).
113 * During boot, all non-reserved memblock memory is exposed to page_alloc.
114 * Poisoning all that memory lengthens boot time, especially on systems with
115 * large amount of RAM. This flag is used to skip that poisoning.
116 * This is only done for the tag-based KASAN modes, as those are able to
117 * detect memory corruptions with the memory tags assigned by default.
118 * All memory allocated normally after boot gets poisoned as usual.
120 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
122 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
123 static DEFINE_MUTEX(pcp_batch_high_lock);
124 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
129 static DEFINE_PER_CPU(struct pagesets, pagesets) = {
130 .lock = INIT_LOCAL_LOCK(lock),
133 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
134 DEFINE_PER_CPU(int, numa_node);
135 EXPORT_PER_CPU_SYMBOL(numa_node);
138 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
140 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
142 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
143 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
144 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
145 * defined in <linux/topology.h>.
147 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
148 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
151 /* work_structs for global per-cpu drains */
154 struct work_struct work;
156 static DEFINE_MUTEX(pcpu_drain_mutex);
157 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
159 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
160 volatile unsigned long latent_entropy __latent_entropy;
161 EXPORT_SYMBOL(latent_entropy);
165 * Array of node states.
167 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
168 [N_POSSIBLE] = NODE_MASK_ALL,
169 [N_ONLINE] = { { [0] = 1UL } },
171 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
172 #ifdef CONFIG_HIGHMEM
173 [N_HIGH_MEMORY] = { { [0] = 1UL } },
175 [N_MEMORY] = { { [0] = 1UL } },
176 [N_CPU] = { { [0] = 1UL } },
179 EXPORT_SYMBOL(node_states);
181 atomic_long_t _totalram_pages __read_mostly;
182 EXPORT_SYMBOL(_totalram_pages);
183 unsigned long totalreserve_pages __read_mostly;
184 unsigned long totalcma_pages __read_mostly;
186 int percpu_pagelist_high_fraction;
187 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
188 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
189 EXPORT_SYMBOL(init_on_alloc);
191 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
192 EXPORT_SYMBOL(init_on_free);
194 static bool _init_on_alloc_enabled_early __read_mostly
195 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
196 static int __init early_init_on_alloc(char *buf)
199 return kstrtobool(buf, &_init_on_alloc_enabled_early);
201 early_param("init_on_alloc", early_init_on_alloc);
203 static bool _init_on_free_enabled_early __read_mostly
204 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
205 static int __init early_init_on_free(char *buf)
207 return kstrtobool(buf, &_init_on_free_enabled_early);
209 early_param("init_on_free", early_init_on_free);
212 * A cached value of the page's pageblock's migratetype, used when the page is
213 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
214 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
215 * Also the migratetype set in the page does not necessarily match the pcplist
216 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
217 * other index - this ensures that it will be put on the correct CMA freelist.
219 static inline int get_pcppage_migratetype(struct page *page)
224 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
226 page->index = migratetype;
229 #ifdef CONFIG_PM_SLEEP
231 * The following functions are used by the suspend/hibernate code to temporarily
232 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
233 * while devices are suspended. To avoid races with the suspend/hibernate code,
234 * they should always be called with system_transition_mutex held
235 * (gfp_allowed_mask also should only be modified with system_transition_mutex
236 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
237 * with that modification).
240 static gfp_t saved_gfp_mask;
242 void pm_restore_gfp_mask(void)
244 WARN_ON(!mutex_is_locked(&system_transition_mutex));
245 if (saved_gfp_mask) {
246 gfp_allowed_mask = saved_gfp_mask;
251 void pm_restrict_gfp_mask(void)
253 WARN_ON(!mutex_is_locked(&system_transition_mutex));
254 WARN_ON(saved_gfp_mask);
255 saved_gfp_mask = gfp_allowed_mask;
256 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
259 bool pm_suspended_storage(void)
261 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
265 #endif /* CONFIG_PM_SLEEP */
267 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
268 unsigned int pageblock_order __read_mostly;
271 static void __free_pages_ok(struct page *page, unsigned int order,
275 * results with 256, 32 in the lowmem_reserve sysctl:
276 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
277 * 1G machine -> (16M dma, 784M normal, 224M high)
278 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
279 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
280 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
282 * TBD: should special case ZONE_DMA32 machines here - in those we normally
283 * don't need any ZONE_NORMAL reservation
285 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
286 #ifdef CONFIG_ZONE_DMA
289 #ifdef CONFIG_ZONE_DMA32
293 #ifdef CONFIG_HIGHMEM
299 static char * const zone_names[MAX_NR_ZONES] = {
300 #ifdef CONFIG_ZONE_DMA
303 #ifdef CONFIG_ZONE_DMA32
307 #ifdef CONFIG_HIGHMEM
311 #ifdef CONFIG_ZONE_DEVICE
316 const char * const migratetype_names[MIGRATE_TYPES] = {
324 #ifdef CONFIG_MEMORY_ISOLATION
329 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
330 [NULL_COMPOUND_DTOR] = NULL,
331 [COMPOUND_PAGE_DTOR] = free_compound_page,
332 #ifdef CONFIG_HUGETLB_PAGE
333 [HUGETLB_PAGE_DTOR] = free_huge_page,
335 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
336 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
340 int min_free_kbytes = 1024;
341 int user_min_free_kbytes = -1;
342 int watermark_boost_factor __read_mostly = 15000;
343 int watermark_scale_factor = 10;
345 static unsigned long nr_kernel_pages __initdata;
346 static unsigned long nr_all_pages __initdata;
347 static unsigned long dma_reserve __initdata;
349 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
350 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
351 static unsigned long required_kernelcore __initdata;
352 static unsigned long required_kernelcore_percent __initdata;
353 static unsigned long required_movablecore __initdata;
354 static unsigned long required_movablecore_percent __initdata;
355 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
356 static bool mirrored_kernelcore __meminitdata;
358 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
360 EXPORT_SYMBOL(movable_zone);
363 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
364 unsigned int nr_online_nodes __read_mostly = 1;
365 EXPORT_SYMBOL(nr_node_ids);
366 EXPORT_SYMBOL(nr_online_nodes);
369 int page_group_by_mobility_disabled __read_mostly;
371 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
373 * During boot we initialize deferred pages on-demand, as needed, but once
374 * page_alloc_init_late() has finished, the deferred pages are all initialized,
375 * and we can permanently disable that path.
377 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
380 * Calling kasan_poison_pages() only after deferred memory initialization
381 * has completed. Poisoning pages during deferred memory init will greatly
382 * lengthen the process and cause problem in large memory systems as the
383 * deferred pages initialization is done with interrupt disabled.
385 * Assuming that there will be no reference to those newly initialized
386 * pages before they are ever allocated, this should have no effect on
387 * KASAN memory tracking as the poison will be properly inserted at page
388 * allocation time. The only corner case is when pages are allocated by
389 * on-demand allocation and then freed again before the deferred pages
390 * initialization is done, but this is not likely to happen.
392 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
394 return static_branch_unlikely(&deferred_pages) ||
395 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
396 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
397 PageSkipKASanPoison(page);
400 /* Returns true if the struct page for the pfn is uninitialised */
401 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
403 int nid = early_pfn_to_nid(pfn);
405 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
412 * Returns true when the remaining initialisation should be deferred until
413 * later in the boot cycle when it can be parallelised.
415 static bool __meminit
416 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
418 static unsigned long prev_end_pfn, nr_initialised;
421 * prev_end_pfn static that contains the end of previous zone
422 * No need to protect because called very early in boot before smp_init.
424 if (prev_end_pfn != end_pfn) {
425 prev_end_pfn = end_pfn;
429 /* Always populate low zones for address-constrained allocations */
430 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
433 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
436 * We start only with one section of pages, more pages are added as
437 * needed until the rest of deferred pages are initialized.
440 if ((nr_initialised > PAGES_PER_SECTION) &&
441 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
442 NODE_DATA(nid)->first_deferred_pfn = pfn;
448 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
450 return (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
451 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
452 PageSkipKASanPoison(page);
455 static inline bool early_page_uninitialised(unsigned long pfn)
460 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
466 /* Return a pointer to the bitmap storing bits affecting a block of pages */
467 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
470 #ifdef CONFIG_SPARSEMEM
471 return section_to_usemap(__pfn_to_section(pfn));
473 return page_zone(page)->pageblock_flags;
474 #endif /* CONFIG_SPARSEMEM */
477 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
479 #ifdef CONFIG_SPARSEMEM
480 pfn &= (PAGES_PER_SECTION-1);
482 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
483 #endif /* CONFIG_SPARSEMEM */
484 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
487 static __always_inline
488 unsigned long __get_pfnblock_flags_mask(const struct page *page,
492 unsigned long *bitmap;
493 unsigned long bitidx, word_bitidx;
496 bitmap = get_pageblock_bitmap(page, pfn);
497 bitidx = pfn_to_bitidx(page, pfn);
498 word_bitidx = bitidx / BITS_PER_LONG;
499 bitidx &= (BITS_PER_LONG-1);
501 word = bitmap[word_bitidx];
502 return (word >> bitidx) & mask;
506 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
507 * @page: The page within the block of interest
508 * @pfn: The target page frame number
509 * @mask: mask of bits that the caller is interested in
511 * Return: pageblock_bits flags
513 unsigned long get_pfnblock_flags_mask(const struct page *page,
514 unsigned long pfn, unsigned long mask)
516 return __get_pfnblock_flags_mask(page, pfn, mask);
519 static __always_inline int get_pfnblock_migratetype(const struct page *page,
522 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
526 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
527 * @page: The page within the block of interest
528 * @flags: The flags to set
529 * @pfn: The target page frame number
530 * @mask: mask of bits that the caller is interested in
532 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
536 unsigned long *bitmap;
537 unsigned long bitidx, word_bitidx;
538 unsigned long old_word, word;
540 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
541 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
543 bitmap = get_pageblock_bitmap(page, pfn);
544 bitidx = pfn_to_bitidx(page, pfn);
545 word_bitidx = bitidx / BITS_PER_LONG;
546 bitidx &= (BITS_PER_LONG-1);
548 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
553 word = READ_ONCE(bitmap[word_bitidx]);
555 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
556 if (word == old_word)
562 void set_pageblock_migratetype(struct page *page, int migratetype)
564 if (unlikely(page_group_by_mobility_disabled &&
565 migratetype < MIGRATE_PCPTYPES))
566 migratetype = MIGRATE_UNMOVABLE;
568 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
569 page_to_pfn(page), MIGRATETYPE_MASK);
572 #ifdef CONFIG_DEBUG_VM
573 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
577 unsigned long pfn = page_to_pfn(page);
578 unsigned long sp, start_pfn;
581 seq = zone_span_seqbegin(zone);
582 start_pfn = zone->zone_start_pfn;
583 sp = zone->spanned_pages;
584 if (!zone_spans_pfn(zone, pfn))
586 } while (zone_span_seqretry(zone, seq));
589 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
590 pfn, zone_to_nid(zone), zone->name,
591 start_pfn, start_pfn + sp);
596 static int page_is_consistent(struct zone *zone, struct page *page)
598 if (zone != page_zone(page))
604 * Temporary debugging check for pages not lying within a given zone.
606 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
608 if (page_outside_zone_boundaries(zone, page))
610 if (!page_is_consistent(zone, page))
616 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
622 static void bad_page(struct page *page, const char *reason)
624 static unsigned long resume;
625 static unsigned long nr_shown;
626 static unsigned long nr_unshown;
629 * Allow a burst of 60 reports, then keep quiet for that minute;
630 * or allow a steady drip of one report per second.
632 if (nr_shown == 60) {
633 if (time_before(jiffies, resume)) {
639 "BUG: Bad page state: %lu messages suppressed\n",
646 resume = jiffies + 60 * HZ;
648 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
649 current->comm, page_to_pfn(page));
650 dump_page(page, reason);
655 /* Leave bad fields for debug, except PageBuddy could make trouble */
656 page_mapcount_reset(page); /* remove PageBuddy */
657 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
660 static inline unsigned int order_to_pindex(int migratetype, int order)
664 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
665 if (order > PAGE_ALLOC_COSTLY_ORDER) {
666 VM_BUG_ON(order != pageblock_order);
667 base = PAGE_ALLOC_COSTLY_ORDER + 1;
670 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
673 return (MIGRATE_PCPTYPES * base) + migratetype;
676 static inline int pindex_to_order(unsigned int pindex)
678 int order = pindex / MIGRATE_PCPTYPES;
680 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
681 if (order > PAGE_ALLOC_COSTLY_ORDER)
682 order = pageblock_order;
684 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
690 static inline bool pcp_allowed_order(unsigned int order)
692 if (order <= PAGE_ALLOC_COSTLY_ORDER)
694 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
695 if (order == pageblock_order)
701 static inline void free_the_page(struct page *page, unsigned int order)
703 if (pcp_allowed_order(order)) /* Via pcp? */
704 free_unref_page(page, order);
706 __free_pages_ok(page, order, FPI_NONE);
710 * Higher-order pages are called "compound pages". They are structured thusly:
712 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
714 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
715 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
717 * The first tail page's ->compound_dtor holds the offset in array of compound
718 * page destructors. See compound_page_dtors.
720 * The first tail page's ->compound_order holds the order of allocation.
721 * This usage means that zero-order pages may not be compound.
724 void free_compound_page(struct page *page)
726 mem_cgroup_uncharge(page_folio(page));
727 free_the_page(page, compound_order(page));
730 void prep_compound_page(struct page *page, unsigned int order)
733 int nr_pages = 1 << order;
736 for (i = 1; i < nr_pages; i++) {
737 struct page *p = page + i;
738 p->mapping = TAIL_MAPPING;
739 set_compound_head(p, page);
742 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
743 set_compound_order(page, order);
744 atomic_set(compound_mapcount_ptr(page), -1);
745 if (hpage_pincount_available(page))
746 atomic_set(compound_pincount_ptr(page), 0);
749 #ifdef CONFIG_DEBUG_PAGEALLOC
750 unsigned int _debug_guardpage_minorder;
752 bool _debug_pagealloc_enabled_early __read_mostly
753 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
754 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
755 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
756 EXPORT_SYMBOL(_debug_pagealloc_enabled);
758 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
760 static int __init early_debug_pagealloc(char *buf)
762 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
764 early_param("debug_pagealloc", early_debug_pagealloc);
766 static int __init debug_guardpage_minorder_setup(char *buf)
770 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
771 pr_err("Bad debug_guardpage_minorder value\n");
774 _debug_guardpage_minorder = res;
775 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
778 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
780 static inline bool set_page_guard(struct zone *zone, struct page *page,
781 unsigned int order, int migratetype)
783 if (!debug_guardpage_enabled())
786 if (order >= debug_guardpage_minorder())
789 __SetPageGuard(page);
790 INIT_LIST_HEAD(&page->lru);
791 set_page_private(page, order);
792 /* Guard pages are not available for any usage */
793 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
798 static inline void clear_page_guard(struct zone *zone, struct page *page,
799 unsigned int order, int migratetype)
801 if (!debug_guardpage_enabled())
804 __ClearPageGuard(page);
806 set_page_private(page, 0);
807 if (!is_migrate_isolate(migratetype))
808 __mod_zone_freepage_state(zone, (1 << order), migratetype);
811 static inline bool set_page_guard(struct zone *zone, struct page *page,
812 unsigned int order, int migratetype) { return false; }
813 static inline void clear_page_guard(struct zone *zone, struct page *page,
814 unsigned int order, int migratetype) {}
818 * Enable static keys related to various memory debugging and hardening options.
819 * Some override others, and depend on early params that are evaluated in the
820 * order of appearance. So we need to first gather the full picture of what was
821 * enabled, and then make decisions.
823 void init_mem_debugging_and_hardening(void)
825 bool page_poisoning_requested = false;
827 #ifdef CONFIG_PAGE_POISONING
829 * Page poisoning is debug page alloc for some arches. If
830 * either of those options are enabled, enable poisoning.
832 if (page_poisoning_enabled() ||
833 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
834 debug_pagealloc_enabled())) {
835 static_branch_enable(&_page_poisoning_enabled);
836 page_poisoning_requested = true;
840 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
841 page_poisoning_requested) {
842 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
843 "will take precedence over init_on_alloc and init_on_free\n");
844 _init_on_alloc_enabled_early = false;
845 _init_on_free_enabled_early = false;
848 if (_init_on_alloc_enabled_early)
849 static_branch_enable(&init_on_alloc);
851 static_branch_disable(&init_on_alloc);
853 if (_init_on_free_enabled_early)
854 static_branch_enable(&init_on_free);
856 static_branch_disable(&init_on_free);
858 #ifdef CONFIG_DEBUG_PAGEALLOC
859 if (!debug_pagealloc_enabled())
862 static_branch_enable(&_debug_pagealloc_enabled);
864 if (!debug_guardpage_minorder())
867 static_branch_enable(&_debug_guardpage_enabled);
871 static inline void set_buddy_order(struct page *page, unsigned int order)
873 set_page_private(page, order);
874 __SetPageBuddy(page);
878 * This function checks whether a page is free && is the buddy
879 * we can coalesce a page and its buddy if
880 * (a) the buddy is not in a hole (check before calling!) &&
881 * (b) the buddy is in the buddy system &&
882 * (c) a page and its buddy have the same order &&
883 * (d) a page and its buddy are in the same zone.
885 * For recording whether a page is in the buddy system, we set PageBuddy.
886 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
888 * For recording page's order, we use page_private(page).
890 static inline bool page_is_buddy(struct page *page, struct page *buddy,
893 if (!page_is_guard(buddy) && !PageBuddy(buddy))
896 if (buddy_order(buddy) != order)
900 * zone check is done late to avoid uselessly calculating
901 * zone/node ids for pages that could never merge.
903 if (page_zone_id(page) != page_zone_id(buddy))
906 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
911 #ifdef CONFIG_COMPACTION
912 static inline struct capture_control *task_capc(struct zone *zone)
914 struct capture_control *capc = current->capture_control;
916 return unlikely(capc) &&
917 !(current->flags & PF_KTHREAD) &&
919 capc->cc->zone == zone ? capc : NULL;
923 compaction_capture(struct capture_control *capc, struct page *page,
924 int order, int migratetype)
926 if (!capc || order != capc->cc->order)
929 /* Do not accidentally pollute CMA or isolated regions*/
930 if (is_migrate_cma(migratetype) ||
931 is_migrate_isolate(migratetype))
935 * Do not let lower order allocations pollute a movable pageblock.
936 * This might let an unmovable request use a reclaimable pageblock
937 * and vice-versa but no more than normal fallback logic which can
938 * have trouble finding a high-order free page.
940 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
948 static inline struct capture_control *task_capc(struct zone *zone)
954 compaction_capture(struct capture_control *capc, struct page *page,
955 int order, int migratetype)
959 #endif /* CONFIG_COMPACTION */
961 /* Used for pages not on another list */
962 static inline void add_to_free_list(struct page *page, struct zone *zone,
963 unsigned int order, int migratetype)
965 struct free_area *area = &zone->free_area[order];
967 list_add(&page->lru, &area->free_list[migratetype]);
971 /* Used for pages not on another list */
972 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
973 unsigned int order, int migratetype)
975 struct free_area *area = &zone->free_area[order];
977 list_add_tail(&page->lru, &area->free_list[migratetype]);
982 * Used for pages which are on another list. Move the pages to the tail
983 * of the list - so the moved pages won't immediately be considered for
984 * allocation again (e.g., optimization for memory onlining).
986 static inline void move_to_free_list(struct page *page, struct zone *zone,
987 unsigned int order, int migratetype)
989 struct free_area *area = &zone->free_area[order];
991 list_move_tail(&page->lru, &area->free_list[migratetype]);
994 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
997 /* clear reported state and update reported page count */
998 if (page_reported(page))
999 __ClearPageReported(page);
1001 list_del(&page->lru);
1002 __ClearPageBuddy(page);
1003 set_page_private(page, 0);
1004 zone->free_area[order].nr_free--;
1008 * If this is not the largest possible page, check if the buddy
1009 * of the next-highest order is free. If it is, it's possible
1010 * that pages are being freed that will coalesce soon. In case,
1011 * that is happening, add the free page to the tail of the list
1012 * so it's less likely to be used soon and more likely to be merged
1013 * as a higher order page
1016 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1017 struct page *page, unsigned int order)
1019 struct page *higher_page, *higher_buddy;
1020 unsigned long combined_pfn;
1022 if (order >= MAX_ORDER - 2)
1025 combined_pfn = buddy_pfn & pfn;
1026 higher_page = page + (combined_pfn - pfn);
1027 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1028 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1030 return page_is_buddy(higher_page, higher_buddy, order + 1);
1034 * Freeing function for a buddy system allocator.
1036 * The concept of a buddy system is to maintain direct-mapped table
1037 * (containing bit values) for memory blocks of various "orders".
1038 * The bottom level table contains the map for the smallest allocatable
1039 * units of memory (here, pages), and each level above it describes
1040 * pairs of units from the levels below, hence, "buddies".
1041 * At a high level, all that happens here is marking the table entry
1042 * at the bottom level available, and propagating the changes upward
1043 * as necessary, plus some accounting needed to play nicely with other
1044 * parts of the VM system.
1045 * At each level, we keep a list of pages, which are heads of continuous
1046 * free pages of length of (1 << order) and marked with PageBuddy.
1047 * Page's order is recorded in page_private(page) field.
1048 * So when we are allocating or freeing one, we can derive the state of the
1049 * other. That is, if we allocate a small block, and both were
1050 * free, the remainder of the region must be split into blocks.
1051 * If a block is freed, and its buddy is also free, then this
1052 * triggers coalescing into a block of larger size.
1057 static inline void __free_one_page(struct page *page,
1059 struct zone *zone, unsigned int order,
1060 int migratetype, fpi_t fpi_flags)
1062 struct capture_control *capc = task_capc(zone);
1063 unsigned long buddy_pfn;
1064 unsigned long combined_pfn;
1065 unsigned int max_order;
1069 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1071 VM_BUG_ON(!zone_is_initialized(zone));
1072 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1074 VM_BUG_ON(migratetype == -1);
1075 if (likely(!is_migrate_isolate(migratetype)))
1076 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1078 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1079 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1082 while (order < max_order) {
1083 if (compaction_capture(capc, page, order, migratetype)) {
1084 __mod_zone_freepage_state(zone, -(1 << order),
1088 buddy_pfn = __find_buddy_pfn(pfn, order);
1089 buddy = page + (buddy_pfn - pfn);
1091 if (!page_is_buddy(page, buddy, order))
1094 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1095 * merge with it and move up one order.
1097 if (page_is_guard(buddy))
1098 clear_page_guard(zone, buddy, order, migratetype);
1100 del_page_from_free_list(buddy, zone, order);
1101 combined_pfn = buddy_pfn & pfn;
1102 page = page + (combined_pfn - pfn);
1106 if (order < MAX_ORDER - 1) {
1107 /* If we are here, it means order is >= pageblock_order.
1108 * We want to prevent merge between freepages on isolate
1109 * pageblock and normal pageblock. Without this, pageblock
1110 * isolation could cause incorrect freepage or CMA accounting.
1112 * We don't want to hit this code for the more frequent
1113 * low-order merging.
1115 if (unlikely(has_isolate_pageblock(zone))) {
1118 buddy_pfn = __find_buddy_pfn(pfn, order);
1119 buddy = page + (buddy_pfn - pfn);
1120 buddy_mt = get_pageblock_migratetype(buddy);
1122 if (migratetype != buddy_mt
1123 && (is_migrate_isolate(migratetype) ||
1124 is_migrate_isolate(buddy_mt)))
1127 max_order = order + 1;
1128 goto continue_merging;
1132 set_buddy_order(page, order);
1134 if (fpi_flags & FPI_TO_TAIL)
1136 else if (is_shuffle_order(order))
1137 to_tail = shuffle_pick_tail();
1139 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1142 add_to_free_list_tail(page, zone, order, migratetype);
1144 add_to_free_list(page, zone, order, migratetype);
1146 /* Notify page reporting subsystem of freed page */
1147 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1148 page_reporting_notify_free(order);
1152 * A bad page could be due to a number of fields. Instead of multiple branches,
1153 * try and check multiple fields with one check. The caller must do a detailed
1154 * check if necessary.
1156 static inline bool page_expected_state(struct page *page,
1157 unsigned long check_flags)
1159 if (unlikely(atomic_read(&page->_mapcount) != -1))
1162 if (unlikely((unsigned long)page->mapping |
1163 page_ref_count(page) |
1167 (page->flags & check_flags)))
1173 static const char *page_bad_reason(struct page *page, unsigned long flags)
1175 const char *bad_reason = NULL;
1177 if (unlikely(atomic_read(&page->_mapcount) != -1))
1178 bad_reason = "nonzero mapcount";
1179 if (unlikely(page->mapping != NULL))
1180 bad_reason = "non-NULL mapping";
1181 if (unlikely(page_ref_count(page) != 0))
1182 bad_reason = "nonzero _refcount";
1183 if (unlikely(page->flags & flags)) {
1184 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1185 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1187 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1190 if (unlikely(page->memcg_data))
1191 bad_reason = "page still charged to cgroup";
1196 static void check_free_page_bad(struct page *page)
1199 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1202 static inline int check_free_page(struct page *page)
1204 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1207 /* Something has gone sideways, find it */
1208 check_free_page_bad(page);
1212 static int free_tail_pages_check(struct page *head_page, struct page *page)
1217 * We rely page->lru.next never has bit 0 set, unless the page
1218 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1220 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1222 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1226 switch (page - head_page) {
1228 /* the first tail page: ->mapping may be compound_mapcount() */
1229 if (unlikely(compound_mapcount(page))) {
1230 bad_page(page, "nonzero compound_mapcount");
1236 * the second tail page: ->mapping is
1237 * deferred_list.next -- ignore value.
1241 if (page->mapping != TAIL_MAPPING) {
1242 bad_page(page, "corrupted mapping in tail page");
1247 if (unlikely(!PageTail(page))) {
1248 bad_page(page, "PageTail not set");
1251 if (unlikely(compound_head(page) != head_page)) {
1252 bad_page(page, "compound_head not consistent");
1257 page->mapping = NULL;
1258 clear_compound_head(page);
1262 static void kernel_init_free_pages(struct page *page, int numpages, bool zero_tags)
1267 for (i = 0; i < numpages; i++)
1268 tag_clear_highpage(page + i);
1272 /* s390's use of memset() could override KASAN redzones. */
1273 kasan_disable_current();
1274 for (i = 0; i < numpages; i++) {
1275 u8 tag = page_kasan_tag(page + i);
1276 page_kasan_tag_reset(page + i);
1277 clear_highpage(page + i);
1278 page_kasan_tag_set(page + i, tag);
1280 kasan_enable_current();
1283 static __always_inline bool free_pages_prepare(struct page *page,
1284 unsigned int order, bool check_free, fpi_t fpi_flags)
1287 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1289 VM_BUG_ON_PAGE(PageTail(page), page);
1291 trace_mm_page_free(page, order);
1293 if (unlikely(PageHWPoison(page)) && !order) {
1295 * Do not let hwpoison pages hit pcplists/buddy
1296 * Untie memcg state and reset page's owner
1298 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1299 __memcg_kmem_uncharge_page(page, order);
1300 reset_page_owner(page, order);
1305 * Check tail pages before head page information is cleared to
1306 * avoid checking PageCompound for order-0 pages.
1308 if (unlikely(order)) {
1309 bool compound = PageCompound(page);
1312 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1315 ClearPageDoubleMap(page);
1316 ClearPageHasHWPoisoned(page);
1318 for (i = 1; i < (1 << order); i++) {
1320 bad += free_tail_pages_check(page, page + i);
1321 if (unlikely(check_free_page(page + i))) {
1325 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1328 if (PageMappingFlags(page))
1329 page->mapping = NULL;
1330 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1331 __memcg_kmem_uncharge_page(page, order);
1333 bad += check_free_page(page);
1337 page_cpupid_reset_last(page);
1338 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1339 reset_page_owner(page, order);
1341 if (!PageHighMem(page)) {
1342 debug_check_no_locks_freed(page_address(page),
1343 PAGE_SIZE << order);
1344 debug_check_no_obj_freed(page_address(page),
1345 PAGE_SIZE << order);
1348 kernel_poison_pages(page, 1 << order);
1351 * As memory initialization might be integrated into KASAN,
1352 * kasan_free_pages and kernel_init_free_pages must be
1353 * kept together to avoid discrepancies in behavior.
1355 * With hardware tag-based KASAN, memory tags must be set before the
1356 * page becomes unavailable via debug_pagealloc or arch_free_page.
1358 if (kasan_has_integrated_init()) {
1359 if (!skip_kasan_poison)
1360 kasan_free_pages(page, order);
1362 bool init = want_init_on_free();
1365 kernel_init_free_pages(page, 1 << order, false);
1366 if (!skip_kasan_poison)
1367 kasan_poison_pages(page, order, init);
1371 * arch_free_page() can make the page's contents inaccessible. s390
1372 * does this. So nothing which can access the page's contents should
1373 * happen after this.
1375 arch_free_page(page, order);
1377 debug_pagealloc_unmap_pages(page, 1 << order);
1382 #ifdef CONFIG_DEBUG_VM
1384 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1385 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1386 * moved from pcp lists to free lists.
1388 static bool free_pcp_prepare(struct page *page, unsigned int order)
1390 return free_pages_prepare(page, order, true, FPI_NONE);
1393 static bool bulkfree_pcp_prepare(struct page *page)
1395 if (debug_pagealloc_enabled_static())
1396 return check_free_page(page);
1402 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1403 * moving from pcp lists to free list in order to reduce overhead. With
1404 * debug_pagealloc enabled, they are checked also immediately when being freed
1407 static bool free_pcp_prepare(struct page *page, unsigned int order)
1409 if (debug_pagealloc_enabled_static())
1410 return free_pages_prepare(page, order, true, FPI_NONE);
1412 return free_pages_prepare(page, order, false, FPI_NONE);
1415 static bool bulkfree_pcp_prepare(struct page *page)
1417 return check_free_page(page);
1419 #endif /* CONFIG_DEBUG_VM */
1421 static inline void prefetch_buddy(struct page *page)
1423 unsigned long pfn = page_to_pfn(page);
1424 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1425 struct page *buddy = page + (buddy_pfn - pfn);
1431 * Frees a number of pages from the PCP lists
1432 * Assumes all pages on list are in same zone.
1433 * count is the number of pages to free.
1435 static void free_pcppages_bulk(struct zone *zone, int count,
1436 struct per_cpu_pages *pcp)
1442 int prefetch_nr = READ_ONCE(pcp->batch);
1443 bool isolated_pageblocks;
1444 struct page *page, *tmp;
1448 * Ensure proper count is passed which otherwise would stuck in the
1449 * below while (list_empty(list)) loop.
1451 count = min(pcp->count, count);
1453 struct list_head *list;
1456 * Remove pages from lists in a round-robin fashion. A
1457 * batch_free count is maintained that is incremented when an
1458 * empty list is encountered. This is so more pages are freed
1459 * off fuller lists instead of spinning excessively around empty
1464 if (++pindex == NR_PCP_LISTS)
1466 list = &pcp->lists[pindex];
1467 } while (list_empty(list));
1469 /* This is the only non-empty list. Free them all. */
1470 if (batch_free == NR_PCP_LISTS)
1473 order = pindex_to_order(pindex);
1474 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1476 page = list_last_entry(list, struct page, lru);
1477 /* must delete to avoid corrupting pcp list */
1478 list_del(&page->lru);
1479 nr_freed += 1 << order;
1480 count -= 1 << order;
1482 if (bulkfree_pcp_prepare(page))
1485 /* Encode order with the migratetype */
1486 page->index <<= NR_PCP_ORDER_WIDTH;
1487 page->index |= order;
1489 list_add_tail(&page->lru, &head);
1492 * We are going to put the page back to the global
1493 * pool, prefetch its buddy to speed up later access
1494 * under zone->lock. It is believed the overhead of
1495 * an additional test and calculating buddy_pfn here
1496 * can be offset by reduced memory latency later. To
1497 * avoid excessive prefetching due to large count, only
1498 * prefetch buddy for the first pcp->batch nr of pages.
1501 prefetch_buddy(page);
1504 } while (count > 0 && --batch_free && !list_empty(list));
1506 pcp->count -= nr_freed;
1509 * local_lock_irq held so equivalent to spin_lock_irqsave for
1510 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1512 spin_lock(&zone->lock);
1513 isolated_pageblocks = has_isolate_pageblock(zone);
1516 * Use safe version since after __free_one_page(),
1517 * page->lru.next will not point to original list.
1519 list_for_each_entry_safe(page, tmp, &head, lru) {
1520 int mt = get_pcppage_migratetype(page);
1522 /* mt has been encoded with the order (see above) */
1523 order = mt & NR_PCP_ORDER_MASK;
1524 mt >>= NR_PCP_ORDER_WIDTH;
1526 /* MIGRATE_ISOLATE page should not go to pcplists */
1527 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1528 /* Pageblock could have been isolated meanwhile */
1529 if (unlikely(isolated_pageblocks))
1530 mt = get_pageblock_migratetype(page);
1532 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1533 trace_mm_page_pcpu_drain(page, order, mt);
1535 spin_unlock(&zone->lock);
1538 static void free_one_page(struct zone *zone,
1539 struct page *page, unsigned long pfn,
1541 int migratetype, fpi_t fpi_flags)
1543 unsigned long flags;
1545 spin_lock_irqsave(&zone->lock, flags);
1546 if (unlikely(has_isolate_pageblock(zone) ||
1547 is_migrate_isolate(migratetype))) {
1548 migratetype = get_pfnblock_migratetype(page, pfn);
1550 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1551 spin_unlock_irqrestore(&zone->lock, flags);
1554 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1555 unsigned long zone, int nid)
1557 mm_zero_struct_page(page);
1558 set_page_links(page, zone, nid, pfn);
1559 init_page_count(page);
1560 page_mapcount_reset(page);
1561 page_cpupid_reset_last(page);
1562 page_kasan_tag_reset(page);
1564 INIT_LIST_HEAD(&page->lru);
1565 #ifdef WANT_PAGE_VIRTUAL
1566 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1567 if (!is_highmem_idx(zone))
1568 set_page_address(page, __va(pfn << PAGE_SHIFT));
1572 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1573 static void __meminit init_reserved_page(unsigned long pfn)
1578 if (!early_page_uninitialised(pfn))
1581 nid = early_pfn_to_nid(pfn);
1582 pgdat = NODE_DATA(nid);
1584 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1585 struct zone *zone = &pgdat->node_zones[zid];
1587 if (zone_spans_pfn(zone, pfn))
1590 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1593 static inline void init_reserved_page(unsigned long pfn)
1596 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1599 * Initialised pages do not have PageReserved set. This function is
1600 * called for each range allocated by the bootmem allocator and
1601 * marks the pages PageReserved. The remaining valid pages are later
1602 * sent to the buddy page allocator.
1604 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1606 unsigned long start_pfn = PFN_DOWN(start);
1607 unsigned long end_pfn = PFN_UP(end);
1609 for (; start_pfn < end_pfn; start_pfn++) {
1610 if (pfn_valid(start_pfn)) {
1611 struct page *page = pfn_to_page(start_pfn);
1613 init_reserved_page(start_pfn);
1615 /* Avoid false-positive PageTail() */
1616 INIT_LIST_HEAD(&page->lru);
1619 * no need for atomic set_bit because the struct
1620 * page is not visible yet so nobody should
1623 __SetPageReserved(page);
1628 static void __free_pages_ok(struct page *page, unsigned int order,
1631 unsigned long flags;
1633 unsigned long pfn = page_to_pfn(page);
1634 struct zone *zone = page_zone(page);
1636 if (!free_pages_prepare(page, order, true, fpi_flags))
1639 migratetype = get_pfnblock_migratetype(page, pfn);
1641 spin_lock_irqsave(&zone->lock, flags);
1642 if (unlikely(has_isolate_pageblock(zone) ||
1643 is_migrate_isolate(migratetype))) {
1644 migratetype = get_pfnblock_migratetype(page, pfn);
1646 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1647 spin_unlock_irqrestore(&zone->lock, flags);
1649 __count_vm_events(PGFREE, 1 << order);
1652 void __free_pages_core(struct page *page, unsigned int order)
1654 unsigned int nr_pages = 1 << order;
1655 struct page *p = page;
1659 * When initializing the memmap, __init_single_page() sets the refcount
1660 * of all pages to 1 ("allocated"/"not free"). We have to set the
1661 * refcount of all involved pages to 0.
1664 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1666 __ClearPageReserved(p);
1667 set_page_count(p, 0);
1669 __ClearPageReserved(p);
1670 set_page_count(p, 0);
1672 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1675 * Bypass PCP and place fresh pages right to the tail, primarily
1676 * relevant for memory onlining.
1678 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1684 * During memory init memblocks map pfns to nids. The search is expensive and
1685 * this caches recent lookups. The implementation of __early_pfn_to_nid
1686 * treats start/end as pfns.
1688 struct mminit_pfnnid_cache {
1689 unsigned long last_start;
1690 unsigned long last_end;
1694 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1697 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1699 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1700 struct mminit_pfnnid_cache *state)
1702 unsigned long start_pfn, end_pfn;
1705 if (state->last_start <= pfn && pfn < state->last_end)
1706 return state->last_nid;
1708 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1709 if (nid != NUMA_NO_NODE) {
1710 state->last_start = start_pfn;
1711 state->last_end = end_pfn;
1712 state->last_nid = nid;
1718 int __meminit early_pfn_to_nid(unsigned long pfn)
1720 static DEFINE_SPINLOCK(early_pfn_lock);
1723 spin_lock(&early_pfn_lock);
1724 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1726 nid = first_online_node;
1727 spin_unlock(&early_pfn_lock);
1731 #endif /* CONFIG_NUMA */
1733 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1736 if (early_page_uninitialised(pfn))
1738 __free_pages_core(page, order);
1742 * Check that the whole (or subset of) a pageblock given by the interval of
1743 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1744 * with the migration of free compaction scanner.
1746 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1748 * It's possible on some configurations to have a setup like node0 node1 node0
1749 * i.e. it's possible that all pages within a zones range of pages do not
1750 * belong to a single zone. We assume that a border between node0 and node1
1751 * can occur within a single pageblock, but not a node0 node1 node0
1752 * interleaving within a single pageblock. It is therefore sufficient to check
1753 * the first and last page of a pageblock and avoid checking each individual
1754 * page in a pageblock.
1756 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1757 unsigned long end_pfn, struct zone *zone)
1759 struct page *start_page;
1760 struct page *end_page;
1762 /* end_pfn is one past the range we are checking */
1765 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1768 start_page = pfn_to_online_page(start_pfn);
1772 if (page_zone(start_page) != zone)
1775 end_page = pfn_to_page(end_pfn);
1777 /* This gives a shorter code than deriving page_zone(end_page) */
1778 if (page_zone_id(start_page) != page_zone_id(end_page))
1784 void set_zone_contiguous(struct zone *zone)
1786 unsigned long block_start_pfn = zone->zone_start_pfn;
1787 unsigned long block_end_pfn;
1789 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1790 for (; block_start_pfn < zone_end_pfn(zone);
1791 block_start_pfn = block_end_pfn,
1792 block_end_pfn += pageblock_nr_pages) {
1794 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1796 if (!__pageblock_pfn_to_page(block_start_pfn,
1797 block_end_pfn, zone))
1802 /* We confirm that there is no hole */
1803 zone->contiguous = true;
1806 void clear_zone_contiguous(struct zone *zone)
1808 zone->contiguous = false;
1811 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1812 static void __init deferred_free_range(unsigned long pfn,
1813 unsigned long nr_pages)
1821 page = pfn_to_page(pfn);
1823 /* Free a large naturally-aligned chunk if possible */
1824 if (nr_pages == pageblock_nr_pages &&
1825 (pfn & (pageblock_nr_pages - 1)) == 0) {
1826 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1827 __free_pages_core(page, pageblock_order);
1831 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1832 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1833 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1834 __free_pages_core(page, 0);
1838 /* Completion tracking for deferred_init_memmap() threads */
1839 static atomic_t pgdat_init_n_undone __initdata;
1840 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1842 static inline void __init pgdat_init_report_one_done(void)
1844 if (atomic_dec_and_test(&pgdat_init_n_undone))
1845 complete(&pgdat_init_all_done_comp);
1849 * Returns true if page needs to be initialized or freed to buddy allocator.
1851 * First we check if pfn is valid on architectures where it is possible to have
1852 * holes within pageblock_nr_pages. On systems where it is not possible, this
1853 * function is optimized out.
1855 * Then, we check if a current large page is valid by only checking the validity
1858 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1860 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1866 * Free pages to buddy allocator. Try to free aligned pages in
1867 * pageblock_nr_pages sizes.
1869 static void __init deferred_free_pages(unsigned long pfn,
1870 unsigned long end_pfn)
1872 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1873 unsigned long nr_free = 0;
1875 for (; pfn < end_pfn; pfn++) {
1876 if (!deferred_pfn_valid(pfn)) {
1877 deferred_free_range(pfn - nr_free, nr_free);
1879 } else if (!(pfn & nr_pgmask)) {
1880 deferred_free_range(pfn - nr_free, nr_free);
1886 /* Free the last block of pages to allocator */
1887 deferred_free_range(pfn - nr_free, nr_free);
1891 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1892 * by performing it only once every pageblock_nr_pages.
1893 * Return number of pages initialized.
1895 static unsigned long __init deferred_init_pages(struct zone *zone,
1897 unsigned long end_pfn)
1899 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1900 int nid = zone_to_nid(zone);
1901 unsigned long nr_pages = 0;
1902 int zid = zone_idx(zone);
1903 struct page *page = NULL;
1905 for (; pfn < end_pfn; pfn++) {
1906 if (!deferred_pfn_valid(pfn)) {
1909 } else if (!page || !(pfn & nr_pgmask)) {
1910 page = pfn_to_page(pfn);
1914 __init_single_page(page, pfn, zid, nid);
1921 * This function is meant to pre-load the iterator for the zone init.
1922 * Specifically it walks through the ranges until we are caught up to the
1923 * first_init_pfn value and exits there. If we never encounter the value we
1924 * return false indicating there are no valid ranges left.
1927 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1928 unsigned long *spfn, unsigned long *epfn,
1929 unsigned long first_init_pfn)
1934 * Start out by walking through the ranges in this zone that have
1935 * already been initialized. We don't need to do anything with them
1936 * so we just need to flush them out of the system.
1938 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1939 if (*epfn <= first_init_pfn)
1941 if (*spfn < first_init_pfn)
1942 *spfn = first_init_pfn;
1951 * Initialize and free pages. We do it in two loops: first we initialize
1952 * struct page, then free to buddy allocator, because while we are
1953 * freeing pages we can access pages that are ahead (computing buddy
1954 * page in __free_one_page()).
1956 * In order to try and keep some memory in the cache we have the loop
1957 * broken along max page order boundaries. This way we will not cause
1958 * any issues with the buddy page computation.
1960 static unsigned long __init
1961 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1962 unsigned long *end_pfn)
1964 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1965 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1966 unsigned long nr_pages = 0;
1969 /* First we loop through and initialize the page values */
1970 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1973 if (mo_pfn <= *start_pfn)
1976 t = min(mo_pfn, *end_pfn);
1977 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1979 if (mo_pfn < *end_pfn) {
1980 *start_pfn = mo_pfn;
1985 /* Reset values and now loop through freeing pages as needed */
1988 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1994 t = min(mo_pfn, epfn);
1995 deferred_free_pages(spfn, t);
2005 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2008 unsigned long spfn, epfn;
2009 struct zone *zone = arg;
2012 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2015 * Initialize and free pages in MAX_ORDER sized increments so that we
2016 * can avoid introducing any issues with the buddy allocator.
2018 while (spfn < end_pfn) {
2019 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2024 /* An arch may override for more concurrency. */
2026 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2031 /* Initialise remaining memory on a node */
2032 static int __init deferred_init_memmap(void *data)
2034 pg_data_t *pgdat = data;
2035 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2036 unsigned long spfn = 0, epfn = 0;
2037 unsigned long first_init_pfn, flags;
2038 unsigned long start = jiffies;
2040 int zid, max_threads;
2043 /* Bind memory initialisation thread to a local node if possible */
2044 if (!cpumask_empty(cpumask))
2045 set_cpus_allowed_ptr(current, cpumask);
2047 pgdat_resize_lock(pgdat, &flags);
2048 first_init_pfn = pgdat->first_deferred_pfn;
2049 if (first_init_pfn == ULONG_MAX) {
2050 pgdat_resize_unlock(pgdat, &flags);
2051 pgdat_init_report_one_done();
2055 /* Sanity check boundaries */
2056 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2057 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2058 pgdat->first_deferred_pfn = ULONG_MAX;
2061 * Once we unlock here, the zone cannot be grown anymore, thus if an
2062 * interrupt thread must allocate this early in boot, zone must be
2063 * pre-grown prior to start of deferred page initialization.
2065 pgdat_resize_unlock(pgdat, &flags);
2067 /* Only the highest zone is deferred so find it */
2068 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2069 zone = pgdat->node_zones + zid;
2070 if (first_init_pfn < zone_end_pfn(zone))
2074 /* If the zone is empty somebody else may have cleared out the zone */
2075 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2079 max_threads = deferred_page_init_max_threads(cpumask);
2081 while (spfn < epfn) {
2082 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2083 struct padata_mt_job job = {
2084 .thread_fn = deferred_init_memmap_chunk,
2087 .size = epfn_align - spfn,
2088 .align = PAGES_PER_SECTION,
2089 .min_chunk = PAGES_PER_SECTION,
2090 .max_threads = max_threads,
2093 padata_do_multithreaded(&job);
2094 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2098 /* Sanity check that the next zone really is unpopulated */
2099 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2101 pr_info("node %d deferred pages initialised in %ums\n",
2102 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2104 pgdat_init_report_one_done();
2109 * If this zone has deferred pages, try to grow it by initializing enough
2110 * deferred pages to satisfy the allocation specified by order, rounded up to
2111 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2112 * of SECTION_SIZE bytes by initializing struct pages in increments of
2113 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2115 * Return true when zone was grown, otherwise return false. We return true even
2116 * when we grow less than requested, to let the caller decide if there are
2117 * enough pages to satisfy the allocation.
2119 * Note: We use noinline because this function is needed only during boot, and
2120 * it is called from a __ref function _deferred_grow_zone. This way we are
2121 * making sure that it is not inlined into permanent text section.
2123 static noinline bool __init
2124 deferred_grow_zone(struct zone *zone, unsigned int order)
2126 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2127 pg_data_t *pgdat = zone->zone_pgdat;
2128 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2129 unsigned long spfn, epfn, flags;
2130 unsigned long nr_pages = 0;
2133 /* Only the last zone may have deferred pages */
2134 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2137 pgdat_resize_lock(pgdat, &flags);
2140 * If someone grew this zone while we were waiting for spinlock, return
2141 * true, as there might be enough pages already.
2143 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2144 pgdat_resize_unlock(pgdat, &flags);
2148 /* If the zone is empty somebody else may have cleared out the zone */
2149 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2150 first_deferred_pfn)) {
2151 pgdat->first_deferred_pfn = ULONG_MAX;
2152 pgdat_resize_unlock(pgdat, &flags);
2153 /* Retry only once. */
2154 return first_deferred_pfn != ULONG_MAX;
2158 * Initialize and free pages in MAX_ORDER sized increments so
2159 * that we can avoid introducing any issues with the buddy
2162 while (spfn < epfn) {
2163 /* update our first deferred PFN for this section */
2164 first_deferred_pfn = spfn;
2166 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2167 touch_nmi_watchdog();
2169 /* We should only stop along section boundaries */
2170 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2173 /* If our quota has been met we can stop here */
2174 if (nr_pages >= nr_pages_needed)
2178 pgdat->first_deferred_pfn = spfn;
2179 pgdat_resize_unlock(pgdat, &flags);
2181 return nr_pages > 0;
2185 * deferred_grow_zone() is __init, but it is called from
2186 * get_page_from_freelist() during early boot until deferred_pages permanently
2187 * disables this call. This is why we have refdata wrapper to avoid warning,
2188 * and to ensure that the function body gets unloaded.
2191 _deferred_grow_zone(struct zone *zone, unsigned int order)
2193 return deferred_grow_zone(zone, order);
2196 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2198 void __init page_alloc_init_late(void)
2203 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2205 /* There will be num_node_state(N_MEMORY) threads */
2206 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2207 for_each_node_state(nid, N_MEMORY) {
2208 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2211 /* Block until all are initialised */
2212 wait_for_completion(&pgdat_init_all_done_comp);
2215 * We initialized the rest of the deferred pages. Permanently disable
2216 * on-demand struct page initialization.
2218 static_branch_disable(&deferred_pages);
2220 /* Reinit limits that are based on free pages after the kernel is up */
2221 files_maxfiles_init();
2226 /* Discard memblock private memory */
2229 for_each_node_state(nid, N_MEMORY)
2230 shuffle_free_memory(NODE_DATA(nid));
2232 for_each_populated_zone(zone)
2233 set_zone_contiguous(zone);
2237 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2238 void __init init_cma_reserved_pageblock(struct page *page)
2240 unsigned i = pageblock_nr_pages;
2241 struct page *p = page;
2244 __ClearPageReserved(p);
2245 set_page_count(p, 0);
2248 set_pageblock_migratetype(page, MIGRATE_CMA);
2250 if (pageblock_order >= MAX_ORDER) {
2251 i = pageblock_nr_pages;
2254 set_page_refcounted(p);
2255 __free_pages(p, MAX_ORDER - 1);
2256 p += MAX_ORDER_NR_PAGES;
2257 } while (i -= MAX_ORDER_NR_PAGES);
2259 set_page_refcounted(page);
2260 __free_pages(page, pageblock_order);
2263 adjust_managed_page_count(page, pageblock_nr_pages);
2264 page_zone(page)->cma_pages += pageblock_nr_pages;
2269 * The order of subdivision here is critical for the IO subsystem.
2270 * Please do not alter this order without good reasons and regression
2271 * testing. Specifically, as large blocks of memory are subdivided,
2272 * the order in which smaller blocks are delivered depends on the order
2273 * they're subdivided in this function. This is the primary factor
2274 * influencing the order in which pages are delivered to the IO
2275 * subsystem according to empirical testing, and this is also justified
2276 * by considering the behavior of a buddy system containing a single
2277 * large block of memory acted on by a series of small allocations.
2278 * This behavior is a critical factor in sglist merging's success.
2282 static inline void expand(struct zone *zone, struct page *page,
2283 int low, int high, int migratetype)
2285 unsigned long size = 1 << high;
2287 while (high > low) {
2290 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2293 * Mark as guard pages (or page), that will allow to
2294 * merge back to allocator when buddy will be freed.
2295 * Corresponding page table entries will not be touched,
2296 * pages will stay not present in virtual address space
2298 if (set_page_guard(zone, &page[size], high, migratetype))
2301 add_to_free_list(&page[size], zone, high, migratetype);
2302 set_buddy_order(&page[size], high);
2306 static void check_new_page_bad(struct page *page)
2308 if (unlikely(page->flags & __PG_HWPOISON)) {
2309 /* Don't complain about hwpoisoned pages */
2310 page_mapcount_reset(page); /* remove PageBuddy */
2315 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2319 * This page is about to be returned from the page allocator
2321 static inline int check_new_page(struct page *page)
2323 if (likely(page_expected_state(page,
2324 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2327 check_new_page_bad(page);
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)
2339 if (debug_pagealloc_enabled_static())
2340 return check_new_page(page);
2345 static inline bool check_new_pcp(struct page *page)
2347 return check_new_page(page);
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)
2357 return check_new_page(page);
2359 static inline bool check_new_pcp(struct page *page)
2361 if (debug_pagealloc_enabled_static())
2362 return check_new_page(page);
2366 #endif /* CONFIG_DEBUG_VM */
2368 static bool check_new_pages(struct page *page, unsigned int order)
2371 for (i = 0; i < (1 << order); i++) {
2372 struct page *p = page + i;
2374 if (unlikely(check_new_page(p)))
2381 inline void post_alloc_hook(struct page *page, unsigned int order,
2384 set_page_private(page, 0);
2385 set_page_refcounted(page);
2387 arch_alloc_page(page, order);
2388 debug_pagealloc_map_pages(page, 1 << order);
2391 * Page unpoisoning must happen before memory initialization.
2392 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2393 * allocations and the page unpoisoning code will complain.
2395 kernel_unpoison_pages(page, 1 << order);
2398 * As memory initialization might be integrated into KASAN,
2399 * kasan_alloc_pages and kernel_init_free_pages must be
2400 * kept together to avoid discrepancies in behavior.
2402 if (kasan_has_integrated_init()) {
2403 kasan_alloc_pages(page, order, gfp_flags);
2405 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2407 kasan_unpoison_pages(page, order, init);
2409 kernel_init_free_pages(page, 1 << order,
2410 gfp_flags & __GFP_ZEROTAGS);
2413 set_page_owner(page, order, gfp_flags);
2416 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2417 unsigned int alloc_flags)
2419 post_alloc_hook(page, order, gfp_flags);
2421 if (order && (gfp_flags & __GFP_COMP))
2422 prep_compound_page(page, order);
2425 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2426 * allocate the page. The expectation is that the caller is taking
2427 * steps that will free more memory. The caller should avoid the page
2428 * being used for !PFMEMALLOC purposes.
2430 if (alloc_flags & ALLOC_NO_WATERMARKS)
2431 set_page_pfmemalloc(page);
2433 clear_page_pfmemalloc(page);
2437 * Go through the free lists for the given migratetype and remove
2438 * the smallest available page from the freelists
2440 static __always_inline
2441 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2444 unsigned int current_order;
2445 struct free_area *area;
2448 /* Find a page of the appropriate size in the preferred list */
2449 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2450 area = &(zone->free_area[current_order]);
2451 page = get_page_from_free_area(area, migratetype);
2454 del_page_from_free_list(page, zone, current_order);
2455 expand(zone, page, order, current_order, migratetype);
2456 set_pcppage_migratetype(page, migratetype);
2465 * This array describes the order lists are fallen back to when
2466 * the free lists for the desirable migrate type are depleted
2468 static int fallbacks[MIGRATE_TYPES][3] = {
2469 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2470 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2471 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2473 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2475 #ifdef CONFIG_MEMORY_ISOLATION
2476 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2481 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2484 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2487 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2488 unsigned int order) { return NULL; }
2492 * Move the free pages in a range to the freelist tail of the requested type.
2493 * Note that start_page and end_pages are not aligned on a pageblock
2494 * boundary. If alignment is required, use move_freepages_block()
2496 static int move_freepages(struct zone *zone,
2497 unsigned long start_pfn, unsigned long end_pfn,
2498 int migratetype, int *num_movable)
2503 int pages_moved = 0;
2505 for (pfn = start_pfn; pfn <= end_pfn;) {
2506 page = pfn_to_page(pfn);
2507 if (!PageBuddy(page)) {
2509 * We assume that pages that could be isolated for
2510 * migration are movable. But we don't actually try
2511 * isolating, as that would be expensive.
2514 (PageLRU(page) || __PageMovable(page)))
2520 /* Make sure we are not inadvertently changing nodes */
2521 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2522 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2524 order = buddy_order(page);
2525 move_to_free_list(page, zone, order, migratetype);
2527 pages_moved += 1 << order;
2533 int move_freepages_block(struct zone *zone, struct page *page,
2534 int migratetype, int *num_movable)
2536 unsigned long start_pfn, end_pfn, pfn;
2541 pfn = page_to_pfn(page);
2542 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2543 end_pfn = start_pfn + pageblock_nr_pages - 1;
2545 /* Do not cross zone boundaries */
2546 if (!zone_spans_pfn(zone, start_pfn))
2548 if (!zone_spans_pfn(zone, end_pfn))
2551 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2555 static void change_pageblock_range(struct page *pageblock_page,
2556 int start_order, int migratetype)
2558 int nr_pageblocks = 1 << (start_order - pageblock_order);
2560 while (nr_pageblocks--) {
2561 set_pageblock_migratetype(pageblock_page, migratetype);
2562 pageblock_page += pageblock_nr_pages;
2567 * When we are falling back to another migratetype during allocation, try to
2568 * steal extra free pages from the same pageblocks to satisfy further
2569 * allocations, instead of polluting multiple pageblocks.
2571 * If we are stealing a relatively large buddy page, it is likely there will
2572 * be more free pages in the pageblock, so try to steal them all. For
2573 * reclaimable and unmovable allocations, we steal regardless of page size,
2574 * as fragmentation caused by those allocations polluting movable pageblocks
2575 * is worse than movable allocations stealing from unmovable and reclaimable
2578 static bool can_steal_fallback(unsigned int order, int start_mt)
2581 * Leaving this order check is intended, although there is
2582 * relaxed order check in next check. The reason is that
2583 * we can actually steal whole pageblock if this condition met,
2584 * but, below check doesn't guarantee it and that is just heuristic
2585 * so could be changed anytime.
2587 if (order >= pageblock_order)
2590 if (order >= pageblock_order / 2 ||
2591 start_mt == MIGRATE_RECLAIMABLE ||
2592 start_mt == MIGRATE_UNMOVABLE ||
2593 page_group_by_mobility_disabled)
2599 static inline bool boost_watermark(struct zone *zone)
2601 unsigned long max_boost;
2603 if (!watermark_boost_factor)
2606 * Don't bother in zones that are unlikely to produce results.
2607 * On small machines, including kdump capture kernels running
2608 * in a small area, boosting the watermark can cause an out of
2609 * memory situation immediately.
2611 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2614 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2615 watermark_boost_factor, 10000);
2618 * high watermark may be uninitialised if fragmentation occurs
2619 * very early in boot so do not boost. We do not fall
2620 * through and boost by pageblock_nr_pages as failing
2621 * allocations that early means that reclaim is not going
2622 * to help and it may even be impossible to reclaim the
2623 * boosted watermark resulting in a hang.
2628 max_boost = max(pageblock_nr_pages, max_boost);
2630 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2637 * This function implements actual steal behaviour. If order is large enough,
2638 * we can steal whole pageblock. If not, we first move freepages in this
2639 * pageblock to our migratetype and determine how many already-allocated pages
2640 * are there in the pageblock with a compatible migratetype. If at least half
2641 * of pages are free or compatible, we can change migratetype of the pageblock
2642 * itself, so pages freed in the future will be put on the correct free list.
2644 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2645 unsigned int alloc_flags, int start_type, bool whole_block)
2647 unsigned int current_order = buddy_order(page);
2648 int free_pages, movable_pages, alike_pages;
2651 old_block_type = get_pageblock_migratetype(page);
2654 * This can happen due to races and we want to prevent broken
2655 * highatomic accounting.
2657 if (is_migrate_highatomic(old_block_type))
2660 /* Take ownership for orders >= pageblock_order */
2661 if (current_order >= pageblock_order) {
2662 change_pageblock_range(page, current_order, start_type);
2667 * Boost watermarks to increase reclaim pressure to reduce the
2668 * likelihood of future fallbacks. Wake kswapd now as the node
2669 * may be balanced overall and kswapd will not wake naturally.
2671 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2672 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2674 /* We are not allowed to try stealing from the whole block */
2678 free_pages = move_freepages_block(zone, page, start_type,
2681 * Determine how many pages are compatible with our allocation.
2682 * For movable allocation, it's the number of movable pages which
2683 * we just obtained. For other types it's a bit more tricky.
2685 if (start_type == MIGRATE_MOVABLE) {
2686 alike_pages = movable_pages;
2689 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2690 * to MOVABLE pageblock, consider all non-movable pages as
2691 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2692 * vice versa, be conservative since we can't distinguish the
2693 * exact migratetype of non-movable pages.
2695 if (old_block_type == MIGRATE_MOVABLE)
2696 alike_pages = pageblock_nr_pages
2697 - (free_pages + movable_pages);
2702 /* moving whole block can fail due to zone boundary conditions */
2707 * If a sufficient number of pages in the block are either free or of
2708 * comparable migratability as our allocation, claim the whole block.
2710 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2711 page_group_by_mobility_disabled)
2712 set_pageblock_migratetype(page, start_type);
2717 move_to_free_list(page, zone, current_order, start_type);
2721 * Check whether there is a suitable fallback freepage with requested order.
2722 * If only_stealable is true, this function returns fallback_mt only if
2723 * we can steal other freepages all together. This would help to reduce
2724 * fragmentation due to mixed migratetype pages in one pageblock.
2726 int find_suitable_fallback(struct free_area *area, unsigned int order,
2727 int migratetype, bool only_stealable, bool *can_steal)
2732 if (area->nr_free == 0)
2737 fallback_mt = fallbacks[migratetype][i];
2738 if (fallback_mt == MIGRATE_TYPES)
2741 if (free_area_empty(area, fallback_mt))
2744 if (can_steal_fallback(order, migratetype))
2747 if (!only_stealable)
2758 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2759 * there are no empty page blocks that contain a page with a suitable order
2761 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2762 unsigned int alloc_order)
2765 unsigned long max_managed, flags;
2768 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2769 * Check is race-prone but harmless.
2771 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2772 if (zone->nr_reserved_highatomic >= max_managed)
2775 spin_lock_irqsave(&zone->lock, flags);
2777 /* Recheck the nr_reserved_highatomic limit under the lock */
2778 if (zone->nr_reserved_highatomic >= max_managed)
2782 mt = get_pageblock_migratetype(page);
2783 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2784 && !is_migrate_cma(mt)) {
2785 zone->nr_reserved_highatomic += pageblock_nr_pages;
2786 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2787 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2791 spin_unlock_irqrestore(&zone->lock, flags);
2795 * Used when an allocation is about to fail under memory pressure. This
2796 * potentially hurts the reliability of high-order allocations when under
2797 * intense memory pressure but failed atomic allocations should be easier
2798 * to recover from than an OOM.
2800 * If @force is true, try to unreserve a pageblock even though highatomic
2801 * pageblock is exhausted.
2803 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2806 struct zonelist *zonelist = ac->zonelist;
2807 unsigned long flags;
2814 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2817 * Preserve at least one pageblock unless memory pressure
2820 if (!force && zone->nr_reserved_highatomic <=
2824 spin_lock_irqsave(&zone->lock, flags);
2825 for (order = 0; order < MAX_ORDER; order++) {
2826 struct free_area *area = &(zone->free_area[order]);
2828 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2833 * In page freeing path, migratetype change is racy so
2834 * we can counter several free pages in a pageblock
2835 * in this loop although we changed the pageblock type
2836 * from highatomic to ac->migratetype. So we should
2837 * adjust the count once.
2839 if (is_migrate_highatomic_page(page)) {
2841 * It should never happen but changes to
2842 * locking could inadvertently allow a per-cpu
2843 * drain to add pages to MIGRATE_HIGHATOMIC
2844 * while unreserving so be safe and watch for
2847 zone->nr_reserved_highatomic -= min(
2849 zone->nr_reserved_highatomic);
2853 * Convert to ac->migratetype and avoid the normal
2854 * pageblock stealing heuristics. Minimally, the caller
2855 * is doing the work and needs the pages. More
2856 * importantly, if the block was always converted to
2857 * MIGRATE_UNMOVABLE or another type then the number
2858 * of pageblocks that cannot be completely freed
2861 set_pageblock_migratetype(page, ac->migratetype);
2862 ret = move_freepages_block(zone, page, ac->migratetype,
2865 spin_unlock_irqrestore(&zone->lock, flags);
2869 spin_unlock_irqrestore(&zone->lock, flags);
2876 * Try finding a free buddy page on the fallback list and put it on the free
2877 * list of requested migratetype, possibly along with other pages from the same
2878 * block, depending on fragmentation avoidance heuristics. Returns true if
2879 * fallback was found so that __rmqueue_smallest() can grab it.
2881 * The use of signed ints for order and current_order is a deliberate
2882 * deviation from the rest of this file, to make the for loop
2883 * condition simpler.
2885 static __always_inline bool
2886 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2887 unsigned int alloc_flags)
2889 struct free_area *area;
2891 int min_order = order;
2897 * Do not steal pages from freelists belonging to other pageblocks
2898 * i.e. orders < pageblock_order. If there are no local zones free,
2899 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2901 if (alloc_flags & ALLOC_NOFRAGMENT)
2902 min_order = pageblock_order;
2905 * Find the largest available free page in the other list. This roughly
2906 * approximates finding the pageblock with the most free pages, which
2907 * would be too costly to do exactly.
2909 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2911 area = &(zone->free_area[current_order]);
2912 fallback_mt = find_suitable_fallback(area, current_order,
2913 start_migratetype, false, &can_steal);
2914 if (fallback_mt == -1)
2918 * We cannot steal all free pages from the pageblock and the
2919 * requested migratetype is movable. In that case it's better to
2920 * steal and split the smallest available page instead of the
2921 * largest available page, because even if the next movable
2922 * allocation falls back into a different pageblock than this
2923 * one, it won't cause permanent fragmentation.
2925 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2926 && current_order > order)
2935 for (current_order = order; current_order < MAX_ORDER;
2937 area = &(zone->free_area[current_order]);
2938 fallback_mt = find_suitable_fallback(area, current_order,
2939 start_migratetype, false, &can_steal);
2940 if (fallback_mt != -1)
2945 * This should not happen - we already found a suitable fallback
2946 * when looking for the largest page.
2948 VM_BUG_ON(current_order == MAX_ORDER);
2951 page = get_page_from_free_area(area, fallback_mt);
2953 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2956 trace_mm_page_alloc_extfrag(page, order, current_order,
2957 start_migratetype, fallback_mt);
2964 * Do the hard work of removing an element from the buddy allocator.
2965 * Call me with the zone->lock already held.
2967 static __always_inline struct page *
2968 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2969 unsigned int alloc_flags)
2973 if (IS_ENABLED(CONFIG_CMA)) {
2975 * Balance movable allocations between regular and CMA areas by
2976 * allocating from CMA when over half of the zone's free memory
2977 * is in the CMA area.
2979 if (alloc_flags & ALLOC_CMA &&
2980 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2981 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2982 page = __rmqueue_cma_fallback(zone, order);
2988 page = __rmqueue_smallest(zone, order, migratetype);
2989 if (unlikely(!page)) {
2990 if (alloc_flags & ALLOC_CMA)
2991 page = __rmqueue_cma_fallback(zone, order);
2993 if (!page && __rmqueue_fallback(zone, order, migratetype,
2999 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3004 * Obtain a specified number of elements from the buddy allocator, all under
3005 * a single hold of the lock, for efficiency. Add them to the supplied list.
3006 * Returns the number of new pages which were placed at *list.
3008 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3009 unsigned long count, struct list_head *list,
3010 int migratetype, unsigned int alloc_flags)
3012 int i, allocated = 0;
3015 * local_lock_irq held so equivalent to spin_lock_irqsave for
3016 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3018 spin_lock(&zone->lock);
3019 for (i = 0; i < count; ++i) {
3020 struct page *page = __rmqueue(zone, order, migratetype,
3022 if (unlikely(page == NULL))
3025 if (unlikely(check_pcp_refill(page)))
3029 * Split buddy pages returned by expand() are received here in
3030 * physical page order. The page is added to the tail of
3031 * caller's list. From the callers perspective, the linked list
3032 * is ordered by page number under some conditions. This is
3033 * useful for IO devices that can forward direction from the
3034 * head, thus also in the physical page order. This is useful
3035 * for IO devices that can merge IO requests if the physical
3036 * pages are ordered properly.
3038 list_add_tail(&page->lru, list);
3040 if (is_migrate_cma(get_pcppage_migratetype(page)))
3041 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3046 * i pages were removed from the buddy list even if some leak due
3047 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3048 * on i. Do not confuse with 'allocated' which is the number of
3049 * pages added to the pcp list.
3051 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3052 spin_unlock(&zone->lock);
3058 * Called from the vmstat counter updater to drain pagesets of this
3059 * currently executing processor on remote nodes after they have
3062 * Note that this function must be called with the thread pinned to
3063 * a single processor.
3065 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3067 unsigned long flags;
3068 int to_drain, batch;
3070 local_lock_irqsave(&pagesets.lock, flags);
3071 batch = READ_ONCE(pcp->batch);
3072 to_drain = min(pcp->count, batch);
3074 free_pcppages_bulk(zone, to_drain, pcp);
3075 local_unlock_irqrestore(&pagesets.lock, flags);
3080 * Drain pcplists of the indicated processor and zone.
3082 * The processor must either be the current processor and the
3083 * thread pinned to the current processor or a processor that
3086 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3088 unsigned long flags;
3089 struct per_cpu_pages *pcp;
3091 local_lock_irqsave(&pagesets.lock, flags);
3093 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3095 free_pcppages_bulk(zone, pcp->count, pcp);
3097 local_unlock_irqrestore(&pagesets.lock, flags);
3101 * Drain pcplists of all zones on the indicated processor.
3103 * The processor must either be the current processor and the
3104 * thread pinned to the current processor or a processor that
3107 static void drain_pages(unsigned int cpu)
3111 for_each_populated_zone(zone) {
3112 drain_pages_zone(cpu, zone);
3117 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3119 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3120 * the single zone's pages.
3122 void drain_local_pages(struct zone *zone)
3124 int cpu = smp_processor_id();
3127 drain_pages_zone(cpu, zone);
3132 static void drain_local_pages_wq(struct work_struct *work)
3134 struct pcpu_drain *drain;
3136 drain = container_of(work, struct pcpu_drain, work);
3139 * drain_all_pages doesn't use proper cpu hotplug protection so
3140 * we can race with cpu offline when the WQ can move this from
3141 * a cpu pinned worker to an unbound one. We can operate on a different
3142 * cpu which is alright but we also have to make sure to not move to
3146 drain_local_pages(drain->zone);
3151 * The implementation of drain_all_pages(), exposing an extra parameter to
3152 * drain on all cpus.
3154 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3155 * not empty. The check for non-emptiness can however race with a free to
3156 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3157 * that need the guarantee that every CPU has drained can disable the
3158 * optimizing racy check.
3160 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3165 * Allocate in the BSS so we won't require allocation in
3166 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3168 static cpumask_t cpus_with_pcps;
3171 * Make sure nobody triggers this path before mm_percpu_wq is fully
3174 if (WARN_ON_ONCE(!mm_percpu_wq))
3178 * Do not drain if one is already in progress unless it's specific to
3179 * a zone. Such callers are primarily CMA and memory hotplug and need
3180 * the drain to be complete when the call returns.
3182 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3185 mutex_lock(&pcpu_drain_mutex);
3189 * We don't care about racing with CPU hotplug event
3190 * as offline notification will cause the notified
3191 * cpu to drain that CPU pcps and on_each_cpu_mask
3192 * disables preemption as part of its processing
3194 for_each_online_cpu(cpu) {
3195 struct per_cpu_pages *pcp;
3197 bool has_pcps = false;
3199 if (force_all_cpus) {
3201 * The pcp.count check is racy, some callers need a
3202 * guarantee that no cpu is missed.
3206 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3210 for_each_populated_zone(z) {
3211 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3220 cpumask_set_cpu(cpu, &cpus_with_pcps);
3222 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3225 for_each_cpu(cpu, &cpus_with_pcps) {
3226 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3229 INIT_WORK(&drain->work, drain_local_pages_wq);
3230 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3232 for_each_cpu(cpu, &cpus_with_pcps)
3233 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3235 mutex_unlock(&pcpu_drain_mutex);
3239 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3241 * When zone parameter is non-NULL, spill just the single zone's pages.
3243 * Note that this can be extremely slow as the draining happens in a workqueue.
3245 void drain_all_pages(struct zone *zone)
3247 __drain_all_pages(zone, false);
3250 #ifdef CONFIG_HIBERNATION
3253 * Touch the watchdog for every WD_PAGE_COUNT pages.
3255 #define WD_PAGE_COUNT (128*1024)
3257 void mark_free_pages(struct zone *zone)
3259 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3260 unsigned long flags;
3261 unsigned int order, t;
3264 if (zone_is_empty(zone))
3267 spin_lock_irqsave(&zone->lock, flags);
3269 max_zone_pfn = zone_end_pfn(zone);
3270 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3271 if (pfn_valid(pfn)) {
3272 page = pfn_to_page(pfn);
3274 if (!--page_count) {
3275 touch_nmi_watchdog();
3276 page_count = WD_PAGE_COUNT;
3279 if (page_zone(page) != zone)
3282 if (!swsusp_page_is_forbidden(page))
3283 swsusp_unset_page_free(page);
3286 for_each_migratetype_order(order, t) {
3287 list_for_each_entry(page,
3288 &zone->free_area[order].free_list[t], lru) {
3291 pfn = page_to_pfn(page);
3292 for (i = 0; i < (1UL << order); i++) {
3293 if (!--page_count) {
3294 touch_nmi_watchdog();
3295 page_count = WD_PAGE_COUNT;
3297 swsusp_set_page_free(pfn_to_page(pfn + i));
3301 spin_unlock_irqrestore(&zone->lock, flags);
3303 #endif /* CONFIG_PM */
3305 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3310 if (!free_pcp_prepare(page, order))
3313 migratetype = get_pfnblock_migratetype(page, pfn);
3314 set_pcppage_migratetype(page, migratetype);
3318 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch)
3320 int min_nr_free, max_nr_free;
3322 /* Check for PCP disabled or boot pageset */
3323 if (unlikely(high < batch))
3326 /* Leave at least pcp->batch pages on the list */
3327 min_nr_free = batch;
3328 max_nr_free = high - batch;
3331 * Double the number of pages freed each time there is subsequent
3332 * freeing of pages without any allocation.
3334 batch <<= pcp->free_factor;
3335 if (batch < max_nr_free)
3337 batch = clamp(batch, min_nr_free, max_nr_free);
3342 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone)
3344 int high = READ_ONCE(pcp->high);
3346 if (unlikely(!high))
3349 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3353 * If reclaim is active, limit the number of pages that can be
3354 * stored on pcp lists
3356 return min(READ_ONCE(pcp->batch) << 2, high);
3359 static void free_unref_page_commit(struct page *page, unsigned long pfn,
3360 int migratetype, unsigned int order)
3362 struct zone *zone = page_zone(page);
3363 struct per_cpu_pages *pcp;
3367 __count_vm_event(PGFREE);
3368 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3369 pindex = order_to_pindex(migratetype, order);
3370 list_add(&page->lru, &pcp->lists[pindex]);
3371 pcp->count += 1 << order;
3372 high = nr_pcp_high(pcp, zone);
3373 if (pcp->count >= high) {
3374 int batch = READ_ONCE(pcp->batch);
3376 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch), pcp);
3383 void free_unref_page(struct page *page, unsigned int order)
3385 unsigned long flags;
3386 unsigned long pfn = page_to_pfn(page);
3389 if (!free_unref_page_prepare(page, pfn, order))
3393 * We only track unmovable, reclaimable and movable on pcp lists.
3394 * Place ISOLATE pages on the isolated list because they are being
3395 * offlined but treat HIGHATOMIC as movable pages so we can get those
3396 * areas back if necessary. Otherwise, we may have to free
3397 * excessively into the page allocator
3399 migratetype = get_pcppage_migratetype(page);
3400 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3401 if (unlikely(is_migrate_isolate(migratetype))) {
3402 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3405 migratetype = MIGRATE_MOVABLE;
3408 local_lock_irqsave(&pagesets.lock, flags);
3409 free_unref_page_commit(page, pfn, migratetype, order);
3410 local_unlock_irqrestore(&pagesets.lock, flags);
3414 * Free a list of 0-order pages
3416 void free_unref_page_list(struct list_head *list)
3418 struct page *page, *next;
3419 unsigned long flags, pfn;
3420 int batch_count = 0;
3423 /* Prepare pages for freeing */
3424 list_for_each_entry_safe(page, next, list, lru) {
3425 pfn = page_to_pfn(page);
3426 if (!free_unref_page_prepare(page, pfn, 0)) {
3427 list_del(&page->lru);
3432 * Free isolated pages directly to the allocator, see
3433 * comment in free_unref_page.
3435 migratetype = get_pcppage_migratetype(page);
3436 if (unlikely(is_migrate_isolate(migratetype))) {
3437 list_del(&page->lru);
3438 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3442 set_page_private(page, pfn);
3445 local_lock_irqsave(&pagesets.lock, flags);
3446 list_for_each_entry_safe(page, next, list, lru) {
3447 pfn = page_private(page);
3448 set_page_private(page, 0);
3451 * Non-isolated types over MIGRATE_PCPTYPES get added
3452 * to the MIGRATE_MOVABLE pcp list.
3454 migratetype = get_pcppage_migratetype(page);
3455 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3456 migratetype = MIGRATE_MOVABLE;
3458 trace_mm_page_free_batched(page);
3459 free_unref_page_commit(page, pfn, migratetype, 0);
3462 * Guard against excessive IRQ disabled times when we get
3463 * a large list of pages to free.
3465 if (++batch_count == SWAP_CLUSTER_MAX) {
3466 local_unlock_irqrestore(&pagesets.lock, flags);
3468 local_lock_irqsave(&pagesets.lock, flags);
3471 local_unlock_irqrestore(&pagesets.lock, flags);
3475 * split_page takes a non-compound higher-order page, and splits it into
3476 * n (1<<order) sub-pages: page[0..n]
3477 * Each sub-page must be freed individually.
3479 * Note: this is probably too low level an operation for use in drivers.
3480 * Please consult with lkml before using this in your driver.
3482 void split_page(struct page *page, unsigned int order)
3486 VM_BUG_ON_PAGE(PageCompound(page), page);
3487 VM_BUG_ON_PAGE(!page_count(page), page);
3489 for (i = 1; i < (1 << order); i++)
3490 set_page_refcounted(page + i);
3491 split_page_owner(page, 1 << order);
3492 split_page_memcg(page, 1 << order);
3494 EXPORT_SYMBOL_GPL(split_page);
3496 int __isolate_free_page(struct page *page, unsigned int order)
3498 unsigned long watermark;
3502 BUG_ON(!PageBuddy(page));
3504 zone = page_zone(page);
3505 mt = get_pageblock_migratetype(page);
3507 if (!is_migrate_isolate(mt)) {
3509 * Obey watermarks as if the page was being allocated. We can
3510 * emulate a high-order watermark check with a raised order-0
3511 * watermark, because we already know our high-order page
3514 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3515 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3518 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3521 /* Remove page from free list */
3523 del_page_from_free_list(page, zone, order);
3526 * Set the pageblock if the isolated page is at least half of a
3529 if (order >= pageblock_order - 1) {
3530 struct page *endpage = page + (1 << order) - 1;
3531 for (; page < endpage; page += pageblock_nr_pages) {
3532 int mt = get_pageblock_migratetype(page);
3533 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3534 && !is_migrate_highatomic(mt))
3535 set_pageblock_migratetype(page,
3541 return 1UL << order;
3545 * __putback_isolated_page - Return a now-isolated page back where we got it
3546 * @page: Page that was isolated
3547 * @order: Order of the isolated page
3548 * @mt: The page's pageblock's migratetype
3550 * This function is meant to return a page pulled from the free lists via
3551 * __isolate_free_page back to the free lists they were pulled from.
3553 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3555 struct zone *zone = page_zone(page);
3557 /* zone lock should be held when this function is called */
3558 lockdep_assert_held(&zone->lock);
3560 /* Return isolated page to tail of freelist. */
3561 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3562 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3566 * Update NUMA hit/miss statistics
3568 * Must be called with interrupts disabled.
3570 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3574 enum numa_stat_item local_stat = NUMA_LOCAL;
3576 /* skip numa counters update if numa stats is disabled */
3577 if (!static_branch_likely(&vm_numa_stat_key))
3580 if (zone_to_nid(z) != numa_node_id())
3581 local_stat = NUMA_OTHER;
3583 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3584 __count_numa_events(z, NUMA_HIT, nr_account);
3586 __count_numa_events(z, NUMA_MISS, nr_account);
3587 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3589 __count_numa_events(z, local_stat, nr_account);
3593 /* Remove page from the per-cpu list, caller must protect the list */
3595 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3597 unsigned int alloc_flags,
3598 struct per_cpu_pages *pcp,
3599 struct list_head *list)
3604 if (list_empty(list)) {
3605 int batch = READ_ONCE(pcp->batch);
3609 * Scale batch relative to order if batch implies
3610 * free pages can be stored on the PCP. Batch can
3611 * be 1 for small zones or for boot pagesets which
3612 * should never store free pages as the pages may
3613 * belong to arbitrary zones.
3616 batch = max(batch >> order, 2);
3617 alloced = rmqueue_bulk(zone, order,
3619 migratetype, alloc_flags);
3621 pcp->count += alloced << order;
3622 if (unlikely(list_empty(list)))
3626 page = list_first_entry(list, struct page, lru);
3627 list_del(&page->lru);
3628 pcp->count -= 1 << order;
3629 } while (check_new_pcp(page));
3634 /* Lock and remove page from the per-cpu list */
3635 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3636 struct zone *zone, unsigned int order,
3637 gfp_t gfp_flags, int migratetype,
3638 unsigned int alloc_flags)
3640 struct per_cpu_pages *pcp;
3641 struct list_head *list;
3643 unsigned long flags;
3645 local_lock_irqsave(&pagesets.lock, flags);
3648 * On allocation, reduce the number of pages that are batch freed.
3649 * See nr_pcp_free() where free_factor is increased for subsequent
3652 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3653 pcp->free_factor >>= 1;
3654 list = &pcp->lists[order_to_pindex(migratetype, order)];
3655 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3656 local_unlock_irqrestore(&pagesets.lock, flags);
3658 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3659 zone_statistics(preferred_zone, zone, 1);
3665 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3668 struct page *rmqueue(struct zone *preferred_zone,
3669 struct zone *zone, unsigned int order,
3670 gfp_t gfp_flags, unsigned int alloc_flags,
3673 unsigned long flags;
3676 if (likely(pcp_allowed_order(order))) {
3678 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3679 * we need to skip it when CMA area isn't allowed.
3681 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3682 migratetype != MIGRATE_MOVABLE) {
3683 page = rmqueue_pcplist(preferred_zone, zone, order,
3684 gfp_flags, migratetype, alloc_flags);
3690 * We most definitely don't want callers attempting to
3691 * allocate greater than order-1 page units with __GFP_NOFAIL.
3693 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3694 spin_lock_irqsave(&zone->lock, flags);
3699 * order-0 request can reach here when the pcplist is skipped
3700 * due to non-CMA allocation context. HIGHATOMIC area is
3701 * reserved for high-order atomic allocation, so order-0
3702 * request should skip it.
3704 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3705 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3707 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3710 page = __rmqueue(zone, order, migratetype, alloc_flags);
3711 } while (page && check_new_pages(page, order));
3715 __mod_zone_freepage_state(zone, -(1 << order),
3716 get_pcppage_migratetype(page));
3717 spin_unlock_irqrestore(&zone->lock, flags);
3719 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3720 zone_statistics(preferred_zone, zone, 1);
3723 /* Separate test+clear to avoid unnecessary atomics */
3724 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3725 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3726 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3729 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3733 spin_unlock_irqrestore(&zone->lock, flags);
3737 #ifdef CONFIG_FAIL_PAGE_ALLOC
3740 struct fault_attr attr;
3742 bool ignore_gfp_highmem;
3743 bool ignore_gfp_reclaim;
3745 } fail_page_alloc = {
3746 .attr = FAULT_ATTR_INITIALIZER,
3747 .ignore_gfp_reclaim = true,
3748 .ignore_gfp_highmem = true,
3752 static int __init setup_fail_page_alloc(char *str)
3754 return setup_fault_attr(&fail_page_alloc.attr, str);
3756 __setup("fail_page_alloc=", setup_fail_page_alloc);
3758 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3760 if (order < fail_page_alloc.min_order)
3762 if (gfp_mask & __GFP_NOFAIL)
3764 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3766 if (fail_page_alloc.ignore_gfp_reclaim &&
3767 (gfp_mask & __GFP_DIRECT_RECLAIM))
3770 return should_fail(&fail_page_alloc.attr, 1 << order);
3773 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3775 static int __init fail_page_alloc_debugfs(void)
3777 umode_t mode = S_IFREG | 0600;
3780 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3781 &fail_page_alloc.attr);
3783 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3784 &fail_page_alloc.ignore_gfp_reclaim);
3785 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3786 &fail_page_alloc.ignore_gfp_highmem);
3787 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3792 late_initcall(fail_page_alloc_debugfs);
3794 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3796 #else /* CONFIG_FAIL_PAGE_ALLOC */
3798 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3803 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3805 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3807 return __should_fail_alloc_page(gfp_mask, order);
3809 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3811 static inline long __zone_watermark_unusable_free(struct zone *z,
3812 unsigned int order, unsigned int alloc_flags)
3814 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3815 long unusable_free = (1 << order) - 1;
3818 * If the caller does not have rights to ALLOC_HARDER then subtract
3819 * the high-atomic reserves. This will over-estimate the size of the
3820 * atomic reserve but it avoids a search.
3822 if (likely(!alloc_harder))
3823 unusable_free += z->nr_reserved_highatomic;
3826 /* If allocation can't use CMA areas don't use free CMA pages */
3827 if (!(alloc_flags & ALLOC_CMA))
3828 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3831 return unusable_free;
3835 * Return true if free base pages are above 'mark'. For high-order checks it
3836 * will return true of the order-0 watermark is reached and there is at least
3837 * one free page of a suitable size. Checking now avoids taking the zone lock
3838 * to check in the allocation paths if no pages are free.
3840 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3841 int highest_zoneidx, unsigned int alloc_flags,
3846 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3848 /* free_pages may go negative - that's OK */
3849 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3851 if (alloc_flags & ALLOC_HIGH)
3854 if (unlikely(alloc_harder)) {
3856 * OOM victims can try even harder than normal ALLOC_HARDER
3857 * users on the grounds that it's definitely going to be in
3858 * the exit path shortly and free memory. Any allocation it
3859 * makes during the free path will be small and short-lived.
3861 if (alloc_flags & ALLOC_OOM)
3868 * Check watermarks for an order-0 allocation request. If these
3869 * are not met, then a high-order request also cannot go ahead
3870 * even if a suitable page happened to be free.
3872 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3875 /* If this is an order-0 request then the watermark is fine */
3879 /* For a high-order request, check at least one suitable page is free */
3880 for (o = order; o < MAX_ORDER; o++) {
3881 struct free_area *area = &z->free_area[o];
3887 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3888 if (!free_area_empty(area, mt))
3893 if ((alloc_flags & ALLOC_CMA) &&
3894 !free_area_empty(area, MIGRATE_CMA)) {
3898 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3904 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3905 int highest_zoneidx, unsigned int alloc_flags)
3907 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3908 zone_page_state(z, NR_FREE_PAGES));
3911 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3912 unsigned long mark, int highest_zoneidx,
3913 unsigned int alloc_flags, gfp_t gfp_mask)
3917 free_pages = zone_page_state(z, NR_FREE_PAGES);
3920 * Fast check for order-0 only. If this fails then the reserves
3921 * need to be calculated.
3926 fast_free = free_pages;
3927 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3928 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3932 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3936 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3937 * when checking the min watermark. The min watermark is the
3938 * point where boosting is ignored so that kswapd is woken up
3939 * when below the low watermark.
3941 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3942 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3943 mark = z->_watermark[WMARK_MIN];
3944 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3945 alloc_flags, free_pages);
3951 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3952 unsigned long mark, int highest_zoneidx)
3954 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3956 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3957 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3959 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3964 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3966 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3968 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3969 node_reclaim_distance;
3971 #else /* CONFIG_NUMA */
3972 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3976 #endif /* CONFIG_NUMA */
3979 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3980 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3981 * premature use of a lower zone may cause lowmem pressure problems that
3982 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3983 * probably too small. It only makes sense to spread allocations to avoid
3984 * fragmentation between the Normal and DMA32 zones.
3986 static inline unsigned int
3987 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3989 unsigned int alloc_flags;
3992 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3995 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3997 #ifdef CONFIG_ZONE_DMA32
4001 if (zone_idx(zone) != ZONE_NORMAL)
4005 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4006 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4007 * on UMA that if Normal is populated then so is DMA32.
4009 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4010 if (nr_online_nodes > 1 && !populated_zone(--zone))
4013 alloc_flags |= ALLOC_NOFRAGMENT;
4014 #endif /* CONFIG_ZONE_DMA32 */
4018 /* Must be called after current_gfp_context() which can change gfp_mask */
4019 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4020 unsigned int alloc_flags)
4023 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4024 alloc_flags |= ALLOC_CMA;
4030 * get_page_from_freelist goes through the zonelist trying to allocate
4033 static struct page *
4034 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4035 const struct alloc_context *ac)
4039 struct pglist_data *last_pgdat_dirty_limit = NULL;
4044 * Scan zonelist, looking for a zone with enough free.
4045 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4047 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4048 z = ac->preferred_zoneref;
4049 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4054 if (cpusets_enabled() &&
4055 (alloc_flags & ALLOC_CPUSET) &&
4056 !__cpuset_zone_allowed(zone, gfp_mask))
4059 * When allocating a page cache page for writing, we
4060 * want to get it from a node that is within its dirty
4061 * limit, such that no single node holds more than its
4062 * proportional share of globally allowed dirty pages.
4063 * The dirty limits take into account the node's
4064 * lowmem reserves and high watermark so that kswapd
4065 * should be able to balance it without having to
4066 * write pages from its LRU list.
4068 * XXX: For now, allow allocations to potentially
4069 * exceed the per-node dirty limit in the slowpath
4070 * (spread_dirty_pages unset) before going into reclaim,
4071 * which is important when on a NUMA setup the allowed
4072 * nodes are together not big enough to reach the
4073 * global limit. The proper fix for these situations
4074 * will require awareness of nodes in the
4075 * dirty-throttling and the flusher threads.
4077 if (ac->spread_dirty_pages) {
4078 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4081 if (!node_dirty_ok(zone->zone_pgdat)) {
4082 last_pgdat_dirty_limit = zone->zone_pgdat;
4087 if (no_fallback && nr_online_nodes > 1 &&
4088 zone != ac->preferred_zoneref->zone) {
4092 * If moving to a remote node, retry but allow
4093 * fragmenting fallbacks. Locality is more important
4094 * than fragmentation avoidance.
4096 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4097 if (zone_to_nid(zone) != local_nid) {
4098 alloc_flags &= ~ALLOC_NOFRAGMENT;
4103 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4104 if (!zone_watermark_fast(zone, order, mark,
4105 ac->highest_zoneidx, alloc_flags,
4109 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4111 * Watermark failed for this zone, but see if we can
4112 * grow this zone if it contains deferred pages.
4114 if (static_branch_unlikely(&deferred_pages)) {
4115 if (_deferred_grow_zone(zone, order))
4119 /* Checked here to keep the fast path fast */
4120 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4121 if (alloc_flags & ALLOC_NO_WATERMARKS)
4124 if (!node_reclaim_enabled() ||
4125 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4128 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4130 case NODE_RECLAIM_NOSCAN:
4133 case NODE_RECLAIM_FULL:
4134 /* scanned but unreclaimable */
4137 /* did we reclaim enough */
4138 if (zone_watermark_ok(zone, order, mark,
4139 ac->highest_zoneidx, alloc_flags))
4147 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4148 gfp_mask, alloc_flags, ac->migratetype);
4150 prep_new_page(page, order, gfp_mask, alloc_flags);
4153 * If this is a high-order atomic allocation then check
4154 * if the pageblock should be reserved for the future
4156 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4157 reserve_highatomic_pageblock(page, zone, order);
4161 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4162 /* Try again if zone has deferred pages */
4163 if (static_branch_unlikely(&deferred_pages)) {
4164 if (_deferred_grow_zone(zone, order))
4172 * It's possible on a UMA machine to get through all zones that are
4173 * fragmented. If avoiding fragmentation, reset and try again.
4176 alloc_flags &= ~ALLOC_NOFRAGMENT;
4183 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4185 unsigned int filter = SHOW_MEM_FILTER_NODES;
4188 * This documents exceptions given to allocations in certain
4189 * contexts that are allowed to allocate outside current's set
4192 if (!(gfp_mask & __GFP_NOMEMALLOC))
4193 if (tsk_is_oom_victim(current) ||
4194 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4195 filter &= ~SHOW_MEM_FILTER_NODES;
4196 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4197 filter &= ~SHOW_MEM_FILTER_NODES;
4199 show_mem(filter, nodemask);
4202 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4204 struct va_format vaf;
4206 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4208 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4211 va_start(args, fmt);
4214 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4215 current->comm, &vaf, gfp_mask, &gfp_mask,
4216 nodemask_pr_args(nodemask));
4219 cpuset_print_current_mems_allowed();
4222 warn_alloc_show_mem(gfp_mask, nodemask);
4225 static inline struct page *
4226 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4227 unsigned int alloc_flags,
4228 const struct alloc_context *ac)
4232 page = get_page_from_freelist(gfp_mask, order,
4233 alloc_flags|ALLOC_CPUSET, ac);
4235 * fallback to ignore cpuset restriction if our nodes
4239 page = get_page_from_freelist(gfp_mask, order,
4245 static inline struct page *
4246 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4247 const struct alloc_context *ac, unsigned long *did_some_progress)
4249 struct oom_control oc = {
4250 .zonelist = ac->zonelist,
4251 .nodemask = ac->nodemask,
4253 .gfp_mask = gfp_mask,
4258 *did_some_progress = 0;
4261 * Acquire the oom lock. If that fails, somebody else is
4262 * making progress for us.
4264 if (!mutex_trylock(&oom_lock)) {
4265 *did_some_progress = 1;
4266 schedule_timeout_uninterruptible(1);
4271 * Go through the zonelist yet one more time, keep very high watermark
4272 * here, this is only to catch a parallel oom killing, we must fail if
4273 * we're still under heavy pressure. But make sure that this reclaim
4274 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4275 * allocation which will never fail due to oom_lock already held.
4277 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4278 ~__GFP_DIRECT_RECLAIM, order,
4279 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4283 /* Coredumps can quickly deplete all memory reserves */
4284 if (current->flags & PF_DUMPCORE)
4286 /* The OOM killer will not help higher order allocs */
4287 if (order > PAGE_ALLOC_COSTLY_ORDER)
4290 * We have already exhausted all our reclaim opportunities without any
4291 * success so it is time to admit defeat. We will skip the OOM killer
4292 * because it is very likely that the caller has a more reasonable
4293 * fallback than shooting a random task.
4295 * The OOM killer may not free memory on a specific node.
4297 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4299 /* The OOM killer does not needlessly kill tasks for lowmem */
4300 if (ac->highest_zoneidx < ZONE_NORMAL)
4302 if (pm_suspended_storage())
4305 * XXX: GFP_NOFS allocations should rather fail than rely on
4306 * other request to make a forward progress.
4307 * We are in an unfortunate situation where out_of_memory cannot
4308 * do much for this context but let's try it to at least get
4309 * access to memory reserved if the current task is killed (see
4310 * out_of_memory). Once filesystems are ready to handle allocation
4311 * failures more gracefully we should just bail out here.
4314 /* Exhausted what can be done so it's blame time */
4315 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4316 *did_some_progress = 1;
4319 * Help non-failing allocations by giving them access to memory
4322 if (gfp_mask & __GFP_NOFAIL)
4323 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4324 ALLOC_NO_WATERMARKS, ac);
4327 mutex_unlock(&oom_lock);
4332 * Maximum number of compaction retries with a progress before OOM
4333 * killer is consider as the only way to move forward.
4335 #define MAX_COMPACT_RETRIES 16
4337 #ifdef CONFIG_COMPACTION
4338 /* Try memory compaction for high-order allocations before reclaim */
4339 static struct page *
4340 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4341 unsigned int alloc_flags, const struct alloc_context *ac,
4342 enum compact_priority prio, enum compact_result *compact_result)
4344 struct page *page = NULL;
4345 unsigned long pflags;
4346 unsigned int noreclaim_flag;
4351 psi_memstall_enter(&pflags);
4352 delayacct_compact_start();
4353 noreclaim_flag = memalloc_noreclaim_save();
4355 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4358 memalloc_noreclaim_restore(noreclaim_flag);
4359 psi_memstall_leave(&pflags);
4360 delayacct_compact_end();
4362 if (*compact_result == COMPACT_SKIPPED)
4365 * At least in one zone compaction wasn't deferred or skipped, so let's
4366 * count a compaction stall
4368 count_vm_event(COMPACTSTALL);
4370 /* Prep a captured page if available */
4372 prep_new_page(page, order, gfp_mask, alloc_flags);
4374 /* Try get a page from the freelist if available */
4376 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4379 struct zone *zone = page_zone(page);
4381 zone->compact_blockskip_flush = false;
4382 compaction_defer_reset(zone, order, true);
4383 count_vm_event(COMPACTSUCCESS);
4388 * It's bad if compaction run occurs and fails. The most likely reason
4389 * is that pages exist, but not enough to satisfy watermarks.
4391 count_vm_event(COMPACTFAIL);
4399 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4400 enum compact_result compact_result,
4401 enum compact_priority *compact_priority,
4402 int *compaction_retries)
4404 int max_retries = MAX_COMPACT_RETRIES;
4407 int retries = *compaction_retries;
4408 enum compact_priority priority = *compact_priority;
4413 if (fatal_signal_pending(current))
4416 if (compaction_made_progress(compact_result))
4417 (*compaction_retries)++;
4420 * compaction considers all the zone as desperately out of memory
4421 * so it doesn't really make much sense to retry except when the
4422 * failure could be caused by insufficient priority
4424 if (compaction_failed(compact_result))
4425 goto check_priority;
4428 * compaction was skipped because there are not enough order-0 pages
4429 * to work with, so we retry only if it looks like reclaim can help.
4431 if (compaction_needs_reclaim(compact_result)) {
4432 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4437 * make sure the compaction wasn't deferred or didn't bail out early
4438 * due to locks contention before we declare that we should give up.
4439 * But the next retry should use a higher priority if allowed, so
4440 * we don't just keep bailing out endlessly.
4442 if (compaction_withdrawn(compact_result)) {
4443 goto check_priority;
4447 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4448 * costly ones because they are de facto nofail and invoke OOM
4449 * killer to move on while costly can fail and users are ready
4450 * to cope with that. 1/4 retries is rather arbitrary but we
4451 * would need much more detailed feedback from compaction to
4452 * make a better decision.
4454 if (order > PAGE_ALLOC_COSTLY_ORDER)
4456 if (*compaction_retries <= max_retries) {
4462 * Make sure there are attempts at the highest priority if we exhausted
4463 * all retries or failed at the lower priorities.
4466 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4467 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4469 if (*compact_priority > min_priority) {
4470 (*compact_priority)--;
4471 *compaction_retries = 0;
4475 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4479 static inline struct page *
4480 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4481 unsigned int alloc_flags, const struct alloc_context *ac,
4482 enum compact_priority prio, enum compact_result *compact_result)
4484 *compact_result = COMPACT_SKIPPED;
4489 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4490 enum compact_result compact_result,
4491 enum compact_priority *compact_priority,
4492 int *compaction_retries)
4497 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4501 * There are setups with compaction disabled which would prefer to loop
4502 * inside the allocator rather than hit the oom killer prematurely.
4503 * Let's give them a good hope and keep retrying while the order-0
4504 * watermarks are OK.
4506 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4507 ac->highest_zoneidx, ac->nodemask) {
4508 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4509 ac->highest_zoneidx, alloc_flags))
4514 #endif /* CONFIG_COMPACTION */
4516 #ifdef CONFIG_LOCKDEP
4517 static struct lockdep_map __fs_reclaim_map =
4518 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4520 static bool __need_reclaim(gfp_t gfp_mask)
4522 /* no reclaim without waiting on it */
4523 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4526 /* this guy won't enter reclaim */
4527 if (current->flags & PF_MEMALLOC)
4530 if (gfp_mask & __GFP_NOLOCKDEP)
4536 void __fs_reclaim_acquire(unsigned long ip)
4538 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4541 void __fs_reclaim_release(unsigned long ip)
4543 lock_release(&__fs_reclaim_map, ip);
4546 void fs_reclaim_acquire(gfp_t gfp_mask)
4548 gfp_mask = current_gfp_context(gfp_mask);
4550 if (__need_reclaim(gfp_mask)) {
4551 if (gfp_mask & __GFP_FS)
4552 __fs_reclaim_acquire(_RET_IP_);
4554 #ifdef CONFIG_MMU_NOTIFIER
4555 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4556 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4561 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4563 void fs_reclaim_release(gfp_t gfp_mask)
4565 gfp_mask = current_gfp_context(gfp_mask);
4567 if (__need_reclaim(gfp_mask)) {
4568 if (gfp_mask & __GFP_FS)
4569 __fs_reclaim_release(_RET_IP_);
4572 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4575 /* Perform direct synchronous page reclaim */
4576 static unsigned long
4577 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4578 const struct alloc_context *ac)
4580 unsigned int noreclaim_flag;
4581 unsigned long pflags, progress;
4585 /* We now go into synchronous reclaim */
4586 cpuset_memory_pressure_bump();
4587 psi_memstall_enter(&pflags);
4588 fs_reclaim_acquire(gfp_mask);
4589 noreclaim_flag = memalloc_noreclaim_save();
4591 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4594 memalloc_noreclaim_restore(noreclaim_flag);
4595 fs_reclaim_release(gfp_mask);
4596 psi_memstall_leave(&pflags);
4603 /* The really slow allocator path where we enter direct reclaim */
4604 static inline struct page *
4605 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4606 unsigned int alloc_flags, const struct alloc_context *ac,
4607 unsigned long *did_some_progress)
4609 struct page *page = NULL;
4610 bool drained = false;
4612 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4613 if (unlikely(!(*did_some_progress)))
4617 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4620 * If an allocation failed after direct reclaim, it could be because
4621 * pages are pinned on the per-cpu lists or in high alloc reserves.
4622 * Shrink them and try again
4624 if (!page && !drained) {
4625 unreserve_highatomic_pageblock(ac, false);
4626 drain_all_pages(NULL);
4634 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4635 const struct alloc_context *ac)
4639 pg_data_t *last_pgdat = NULL;
4640 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4642 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4644 if (last_pgdat != zone->zone_pgdat)
4645 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4646 last_pgdat = zone->zone_pgdat;
4650 static inline unsigned int
4651 gfp_to_alloc_flags(gfp_t gfp_mask)
4653 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4656 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4657 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4658 * to save two branches.
4660 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4661 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4664 * The caller may dip into page reserves a bit more if the caller
4665 * cannot run direct reclaim, or if the caller has realtime scheduling
4666 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4667 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4669 alloc_flags |= (__force int)
4670 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4672 if (gfp_mask & __GFP_ATOMIC) {
4674 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4675 * if it can't schedule.
4677 if (!(gfp_mask & __GFP_NOMEMALLOC))
4678 alloc_flags |= ALLOC_HARDER;
4680 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4681 * comment for __cpuset_node_allowed().
4683 alloc_flags &= ~ALLOC_CPUSET;
4684 } else if (unlikely(rt_task(current)) && in_task())
4685 alloc_flags |= ALLOC_HARDER;
4687 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4692 static bool oom_reserves_allowed(struct task_struct *tsk)
4694 if (!tsk_is_oom_victim(tsk))
4698 * !MMU doesn't have oom reaper so give access to memory reserves
4699 * only to the thread with TIF_MEMDIE set
4701 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4708 * Distinguish requests which really need access to full memory
4709 * reserves from oom victims which can live with a portion of it
4711 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4713 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4715 if (gfp_mask & __GFP_MEMALLOC)
4716 return ALLOC_NO_WATERMARKS;
4717 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4718 return ALLOC_NO_WATERMARKS;
4719 if (!in_interrupt()) {
4720 if (current->flags & PF_MEMALLOC)
4721 return ALLOC_NO_WATERMARKS;
4722 else if (oom_reserves_allowed(current))
4729 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4731 return !!__gfp_pfmemalloc_flags(gfp_mask);
4735 * Checks whether it makes sense to retry the reclaim to make a forward progress
4736 * for the given allocation request.
4738 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4739 * without success, or when we couldn't even meet the watermark if we
4740 * reclaimed all remaining pages on the LRU lists.
4742 * Returns true if a retry is viable or false to enter the oom path.
4745 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4746 struct alloc_context *ac, int alloc_flags,
4747 bool did_some_progress, int *no_progress_loops)
4754 * Costly allocations might have made a progress but this doesn't mean
4755 * their order will become available due to high fragmentation so
4756 * always increment the no progress counter for them
4758 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4759 *no_progress_loops = 0;
4761 (*no_progress_loops)++;
4764 * Make sure we converge to OOM if we cannot make any progress
4765 * several times in the row.
4767 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4768 /* Before OOM, exhaust highatomic_reserve */
4769 return unreserve_highatomic_pageblock(ac, true);
4773 * Keep reclaiming pages while there is a chance this will lead
4774 * somewhere. If none of the target zones can satisfy our allocation
4775 * request even if all reclaimable pages are considered then we are
4776 * screwed and have to go OOM.
4778 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4779 ac->highest_zoneidx, ac->nodemask) {
4780 unsigned long available;
4781 unsigned long reclaimable;
4782 unsigned long min_wmark = min_wmark_pages(zone);
4785 available = reclaimable = zone_reclaimable_pages(zone);
4786 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4789 * Would the allocation succeed if we reclaimed all
4790 * reclaimable pages?
4792 wmark = __zone_watermark_ok(zone, order, min_wmark,
4793 ac->highest_zoneidx, alloc_flags, available);
4794 trace_reclaim_retry_zone(z, order, reclaimable,
4795 available, min_wmark, *no_progress_loops, wmark);
4803 * Memory allocation/reclaim might be called from a WQ context and the
4804 * current implementation of the WQ concurrency control doesn't
4805 * recognize that a particular WQ is congested if the worker thread is
4806 * looping without ever sleeping. Therefore we have to do a short sleep
4807 * here rather than calling cond_resched().
4809 if (current->flags & PF_WQ_WORKER)
4810 schedule_timeout_uninterruptible(1);
4817 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4820 * It's possible that cpuset's mems_allowed and the nodemask from
4821 * mempolicy don't intersect. This should be normally dealt with by
4822 * policy_nodemask(), but it's possible to race with cpuset update in
4823 * such a way the check therein was true, and then it became false
4824 * before we got our cpuset_mems_cookie here.
4825 * This assumes that for all allocations, ac->nodemask can come only
4826 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4827 * when it does not intersect with the cpuset restrictions) or the
4828 * caller can deal with a violated nodemask.
4830 if (cpusets_enabled() && ac->nodemask &&
4831 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4832 ac->nodemask = NULL;
4837 * When updating a task's mems_allowed or mempolicy nodemask, it is
4838 * possible to race with parallel threads in such a way that our
4839 * allocation can fail while the mask is being updated. If we are about
4840 * to fail, check if the cpuset changed during allocation and if so,
4843 if (read_mems_allowed_retry(cpuset_mems_cookie))
4849 static inline struct page *
4850 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4851 struct alloc_context *ac)
4853 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4854 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4855 struct page *page = NULL;
4856 unsigned int alloc_flags;
4857 unsigned long did_some_progress;
4858 enum compact_priority compact_priority;
4859 enum compact_result compact_result;
4860 int compaction_retries;
4861 int no_progress_loops;
4862 unsigned int cpuset_mems_cookie;
4866 * We also sanity check to catch abuse of atomic reserves being used by
4867 * callers that are not in atomic context.
4869 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4870 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4871 gfp_mask &= ~__GFP_ATOMIC;
4874 compaction_retries = 0;
4875 no_progress_loops = 0;
4876 compact_priority = DEF_COMPACT_PRIORITY;
4877 cpuset_mems_cookie = read_mems_allowed_begin();
4880 * The fast path uses conservative alloc_flags to succeed only until
4881 * kswapd needs to be woken up, and to avoid the cost of setting up
4882 * alloc_flags precisely. So we do that now.
4884 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4887 * We need to recalculate the starting point for the zonelist iterator
4888 * because we might have used different nodemask in the fast path, or
4889 * there was a cpuset modification and we are retrying - otherwise we
4890 * could end up iterating over non-eligible zones endlessly.
4892 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4893 ac->highest_zoneidx, ac->nodemask);
4894 if (!ac->preferred_zoneref->zone)
4898 * Check for insane configurations where the cpuset doesn't contain
4899 * any suitable zone to satisfy the request - e.g. non-movable
4900 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4902 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4903 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4904 ac->highest_zoneidx,
4905 &cpuset_current_mems_allowed);
4910 if (alloc_flags & ALLOC_KSWAPD)
4911 wake_all_kswapds(order, gfp_mask, ac);
4914 * The adjusted alloc_flags might result in immediate success, so try
4917 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4922 * For costly allocations, try direct compaction first, as it's likely
4923 * that we have enough base pages and don't need to reclaim. For non-
4924 * movable high-order allocations, do that as well, as compaction will
4925 * try prevent permanent fragmentation by migrating from blocks of the
4927 * Don't try this for allocations that are allowed to ignore
4928 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4930 if (can_direct_reclaim &&
4932 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4933 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4934 page = __alloc_pages_direct_compact(gfp_mask, order,
4936 INIT_COMPACT_PRIORITY,
4942 * Checks for costly allocations with __GFP_NORETRY, which
4943 * includes some THP page fault allocations
4945 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4947 * If allocating entire pageblock(s) and compaction
4948 * failed because all zones are below low watermarks
4949 * or is prohibited because it recently failed at this
4950 * order, fail immediately unless the allocator has
4951 * requested compaction and reclaim retry.
4954 * - potentially very expensive because zones are far
4955 * below their low watermarks or this is part of very
4956 * bursty high order allocations,
4957 * - not guaranteed to help because isolate_freepages()
4958 * may not iterate over freed pages as part of its
4960 * - unlikely to make entire pageblocks free on its
4963 if (compact_result == COMPACT_SKIPPED ||
4964 compact_result == COMPACT_DEFERRED)
4968 * Looks like reclaim/compaction is worth trying, but
4969 * sync compaction could be very expensive, so keep
4970 * using async compaction.
4972 compact_priority = INIT_COMPACT_PRIORITY;
4977 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4978 if (alloc_flags & ALLOC_KSWAPD)
4979 wake_all_kswapds(order, gfp_mask, ac);
4981 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4983 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
4986 * Reset the nodemask and zonelist iterators if memory policies can be
4987 * ignored. These allocations are high priority and system rather than
4990 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4991 ac->nodemask = NULL;
4992 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4993 ac->highest_zoneidx, ac->nodemask);
4996 /* Attempt with potentially adjusted zonelist and alloc_flags */
4997 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5001 /* Caller is not willing to reclaim, we can't balance anything */
5002 if (!can_direct_reclaim)
5005 /* Avoid recursion of direct reclaim */
5006 if (current->flags & PF_MEMALLOC)
5009 /* Try direct reclaim and then allocating */
5010 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5011 &did_some_progress);
5015 /* Try direct compaction and then allocating */
5016 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5017 compact_priority, &compact_result);
5021 /* Do not loop if specifically requested */
5022 if (gfp_mask & __GFP_NORETRY)
5026 * Do not retry costly high order allocations unless they are
5027 * __GFP_RETRY_MAYFAIL
5029 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5032 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5033 did_some_progress > 0, &no_progress_loops))
5037 * It doesn't make any sense to retry for the compaction if the order-0
5038 * reclaim is not able to make any progress because the current
5039 * implementation of the compaction depends on the sufficient amount
5040 * of free memory (see __compaction_suitable)
5042 if (did_some_progress > 0 &&
5043 should_compact_retry(ac, order, alloc_flags,
5044 compact_result, &compact_priority,
5045 &compaction_retries))
5049 /* Deal with possible cpuset update races before we start OOM killing */
5050 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5053 /* Reclaim has failed us, start killing things */
5054 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5058 /* Avoid allocations with no watermarks from looping endlessly */
5059 if (tsk_is_oom_victim(current) &&
5060 (alloc_flags & ALLOC_OOM ||
5061 (gfp_mask & __GFP_NOMEMALLOC)))
5064 /* Retry as long as the OOM killer is making progress */
5065 if (did_some_progress) {
5066 no_progress_loops = 0;
5071 /* Deal with possible cpuset update races before we fail */
5072 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5076 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5079 if (gfp_mask & __GFP_NOFAIL) {
5081 * All existing users of the __GFP_NOFAIL are blockable, so warn
5082 * of any new users that actually require GFP_NOWAIT
5084 if (WARN_ON_ONCE(!can_direct_reclaim))
5088 * PF_MEMALLOC request from this context is rather bizarre
5089 * because we cannot reclaim anything and only can loop waiting
5090 * for somebody to do a work for us
5092 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5095 * non failing costly orders are a hard requirement which we
5096 * are not prepared for much so let's warn about these users
5097 * so that we can identify them and convert them to something
5100 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5103 * Help non-failing allocations by giving them access to memory
5104 * reserves but do not use ALLOC_NO_WATERMARKS because this
5105 * could deplete whole memory reserves which would just make
5106 * the situation worse
5108 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5116 warn_alloc(gfp_mask, ac->nodemask,
5117 "page allocation failure: order:%u", order);
5122 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5123 int preferred_nid, nodemask_t *nodemask,
5124 struct alloc_context *ac, gfp_t *alloc_gfp,
5125 unsigned int *alloc_flags)
5127 ac->highest_zoneidx = gfp_zone(gfp_mask);
5128 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5129 ac->nodemask = nodemask;
5130 ac->migratetype = gfp_migratetype(gfp_mask);
5132 if (cpusets_enabled()) {
5133 *alloc_gfp |= __GFP_HARDWALL;
5135 * When we are in the interrupt context, it is irrelevant
5136 * to the current task context. It means that any node ok.
5138 if (in_task() && !ac->nodemask)
5139 ac->nodemask = &cpuset_current_mems_allowed;
5141 *alloc_flags |= ALLOC_CPUSET;
5144 fs_reclaim_acquire(gfp_mask);
5145 fs_reclaim_release(gfp_mask);
5147 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5149 if (should_fail_alloc_page(gfp_mask, order))
5152 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5154 /* Dirty zone balancing only done in the fast path */
5155 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5158 * The preferred zone is used for statistics but crucially it is
5159 * also used as the starting point for the zonelist iterator. It
5160 * may get reset for allocations that ignore memory policies.
5162 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5163 ac->highest_zoneidx, ac->nodemask);
5169 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5170 * @gfp: GFP flags for the allocation
5171 * @preferred_nid: The preferred NUMA node ID to allocate from
5172 * @nodemask: Set of nodes to allocate from, may be NULL
5173 * @nr_pages: The number of pages desired on the list or array
5174 * @page_list: Optional list to store the allocated pages
5175 * @page_array: Optional array to store the pages
5177 * This is a batched version of the page allocator that attempts to
5178 * allocate nr_pages quickly. Pages are added to page_list if page_list
5179 * is not NULL, otherwise it is assumed that the page_array is valid.
5181 * For lists, nr_pages is the number of pages that should be allocated.
5183 * For arrays, only NULL elements are populated with pages and nr_pages
5184 * is the maximum number of pages that will be stored in the array.
5186 * Returns the number of pages on the list or array.
5188 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5189 nodemask_t *nodemask, int nr_pages,
5190 struct list_head *page_list,
5191 struct page **page_array)
5194 unsigned long flags;
5197 struct per_cpu_pages *pcp;
5198 struct list_head *pcp_list;
5199 struct alloc_context ac;
5201 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5202 int nr_populated = 0, nr_account = 0;
5205 * Skip populated array elements to determine if any pages need
5206 * to be allocated before disabling IRQs.
5208 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5211 /* No pages requested? */
5212 if (unlikely(nr_pages <= 0))
5215 /* Already populated array? */
5216 if (unlikely(page_array && nr_pages - nr_populated == 0))
5219 /* Bulk allocator does not support memcg accounting. */
5220 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5223 /* Use the single page allocator for one page. */
5224 if (nr_pages - nr_populated == 1)
5227 #ifdef CONFIG_PAGE_OWNER
5229 * PAGE_OWNER may recurse into the allocator to allocate space to
5230 * save the stack with pagesets.lock held. Releasing/reacquiring
5231 * removes much of the performance benefit of bulk allocation so
5232 * force the caller to allocate one page at a time as it'll have
5233 * similar performance to added complexity to the bulk allocator.
5235 if (static_branch_unlikely(&page_owner_inited))
5239 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5240 gfp &= gfp_allowed_mask;
5242 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5246 /* Find an allowed local zone that meets the low watermark. */
5247 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5250 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5251 !__cpuset_zone_allowed(zone, gfp)) {
5255 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5256 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5260 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5261 if (zone_watermark_fast(zone, 0, mark,
5262 zonelist_zone_idx(ac.preferred_zoneref),
5263 alloc_flags, gfp)) {
5269 * If there are no allowed local zones that meets the watermarks then
5270 * try to allocate a single page and reclaim if necessary.
5272 if (unlikely(!zone))
5275 /* Attempt the batch allocation */
5276 local_lock_irqsave(&pagesets.lock, flags);
5277 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5278 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5280 while (nr_populated < nr_pages) {
5282 /* Skip existing pages */
5283 if (page_array && page_array[nr_populated]) {
5288 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5290 if (unlikely(!page)) {
5291 /* Try and get at least one page */
5298 prep_new_page(page, 0, gfp, 0);
5300 list_add(&page->lru, page_list);
5302 page_array[nr_populated] = page;
5306 local_unlock_irqrestore(&pagesets.lock, flags);
5308 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5309 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5312 return nr_populated;
5315 local_unlock_irqrestore(&pagesets.lock, flags);
5318 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5321 list_add(&page->lru, page_list);
5323 page_array[nr_populated] = page;
5329 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5332 * This is the 'heart' of the zoned buddy allocator.
5334 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5335 nodemask_t *nodemask)
5338 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5339 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5340 struct alloc_context ac = { };
5343 * There are several places where we assume that the order value is sane
5344 * so bail out early if the request is out of bound.
5346 if (unlikely(order >= MAX_ORDER)) {
5347 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5351 gfp &= gfp_allowed_mask;
5353 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5354 * resp. GFP_NOIO which has to be inherited for all allocation requests
5355 * from a particular context which has been marked by
5356 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5357 * movable zones are not used during allocation.
5359 gfp = current_gfp_context(gfp);
5361 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5362 &alloc_gfp, &alloc_flags))
5366 * Forbid the first pass from falling back to types that fragment
5367 * memory until all local zones are considered.
5369 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5371 /* First allocation attempt */
5372 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5377 ac.spread_dirty_pages = false;
5380 * Restore the original nodemask if it was potentially replaced with
5381 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5383 ac.nodemask = nodemask;
5385 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5388 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5389 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5390 __free_pages(page, order);
5394 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5398 EXPORT_SYMBOL(__alloc_pages);
5400 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5401 nodemask_t *nodemask)
5403 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5404 preferred_nid, nodemask);
5406 if (page && order > 1)
5407 prep_transhuge_page(page);
5408 return (struct folio *)page;
5410 EXPORT_SYMBOL(__folio_alloc);
5413 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5414 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5415 * you need to access high mem.
5417 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5421 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5424 return (unsigned long) page_address(page);
5426 EXPORT_SYMBOL(__get_free_pages);
5428 unsigned long get_zeroed_page(gfp_t gfp_mask)
5430 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5432 EXPORT_SYMBOL(get_zeroed_page);
5435 * __free_pages - Free pages allocated with alloc_pages().
5436 * @page: The page pointer returned from alloc_pages().
5437 * @order: The order of the allocation.
5439 * This function can free multi-page allocations that are not compound
5440 * pages. It does not check that the @order passed in matches that of
5441 * the allocation, so it is easy to leak memory. Freeing more memory
5442 * than was allocated will probably emit a warning.
5444 * If the last reference to this page is speculative, it will be released
5445 * by put_page() which only frees the first page of a non-compound
5446 * allocation. To prevent the remaining pages from being leaked, we free
5447 * the subsequent pages here. If you want to use the page's reference
5448 * count to decide when to free the allocation, you should allocate a
5449 * compound page, and use put_page() instead of __free_pages().
5451 * Context: May be called in interrupt context or while holding a normal
5452 * spinlock, but not in NMI context or while holding a raw spinlock.
5454 void __free_pages(struct page *page, unsigned int order)
5456 if (put_page_testzero(page))
5457 free_the_page(page, order);
5458 else if (!PageHead(page))
5460 free_the_page(page + (1 << order), order);
5462 EXPORT_SYMBOL(__free_pages);
5464 void free_pages(unsigned long addr, unsigned int order)
5467 VM_BUG_ON(!virt_addr_valid((void *)addr));
5468 __free_pages(virt_to_page((void *)addr), order);
5472 EXPORT_SYMBOL(free_pages);
5476 * An arbitrary-length arbitrary-offset area of memory which resides
5477 * within a 0 or higher order page. Multiple fragments within that page
5478 * are individually refcounted, in the page's reference counter.
5480 * The page_frag functions below provide a simple allocation framework for
5481 * page fragments. This is used by the network stack and network device
5482 * drivers to provide a backing region of memory for use as either an
5483 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5485 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5488 struct page *page = NULL;
5489 gfp_t gfp = gfp_mask;
5491 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5492 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5494 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5495 PAGE_FRAG_CACHE_MAX_ORDER);
5496 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5498 if (unlikely(!page))
5499 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5501 nc->va = page ? page_address(page) : NULL;
5506 void __page_frag_cache_drain(struct page *page, unsigned int count)
5508 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5510 if (page_ref_sub_and_test(page, count))
5511 free_the_page(page, compound_order(page));
5513 EXPORT_SYMBOL(__page_frag_cache_drain);
5515 void *page_frag_alloc_align(struct page_frag_cache *nc,
5516 unsigned int fragsz, gfp_t gfp_mask,
5517 unsigned int align_mask)
5519 unsigned int size = PAGE_SIZE;
5523 if (unlikely(!nc->va)) {
5525 page = __page_frag_cache_refill(nc, gfp_mask);
5529 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5530 /* if size can vary use size else just use PAGE_SIZE */
5533 /* Even if we own the page, we do not use atomic_set().
5534 * This would break get_page_unless_zero() users.
5536 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5538 /* reset page count bias and offset to start of new frag */
5539 nc->pfmemalloc = page_is_pfmemalloc(page);
5540 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5544 offset = nc->offset - fragsz;
5545 if (unlikely(offset < 0)) {
5546 page = virt_to_page(nc->va);
5548 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5551 if (unlikely(nc->pfmemalloc)) {
5552 free_the_page(page, compound_order(page));
5556 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5557 /* if size can vary use size else just use PAGE_SIZE */
5560 /* OK, page count is 0, we can safely set it */
5561 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5563 /* reset page count bias and offset to start of new frag */
5564 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5565 offset = size - fragsz;
5569 offset &= align_mask;
5570 nc->offset = offset;
5572 return nc->va + offset;
5574 EXPORT_SYMBOL(page_frag_alloc_align);
5577 * Frees a page fragment allocated out of either a compound or order 0 page.
5579 void page_frag_free(void *addr)
5581 struct page *page = virt_to_head_page(addr);
5583 if (unlikely(put_page_testzero(page)))
5584 free_the_page(page, compound_order(page));
5586 EXPORT_SYMBOL(page_frag_free);
5588 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5592 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5593 unsigned long used = addr + PAGE_ALIGN(size);
5595 split_page(virt_to_page((void *)addr), order);
5596 while (used < alloc_end) {
5601 return (void *)addr;
5605 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5606 * @size: the number of bytes to allocate
5607 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5609 * This function is similar to alloc_pages(), except that it allocates the
5610 * minimum number of pages to satisfy the request. alloc_pages() can only
5611 * allocate memory in power-of-two pages.
5613 * This function is also limited by MAX_ORDER.
5615 * Memory allocated by this function must be released by free_pages_exact().
5617 * Return: pointer to the allocated area or %NULL in case of error.
5619 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5621 unsigned int order = get_order(size);
5624 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5625 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5627 addr = __get_free_pages(gfp_mask, order);
5628 return make_alloc_exact(addr, order, size);
5630 EXPORT_SYMBOL(alloc_pages_exact);
5633 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5635 * @nid: the preferred node ID where memory should be allocated
5636 * @size: the number of bytes to allocate
5637 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5639 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5642 * Return: pointer to the allocated area or %NULL in case of error.
5644 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5646 unsigned int order = get_order(size);
5649 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5650 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5652 p = alloc_pages_node(nid, gfp_mask, order);
5655 return make_alloc_exact((unsigned long)page_address(p), order, size);
5659 * free_pages_exact - release memory allocated via alloc_pages_exact()
5660 * @virt: the value returned by alloc_pages_exact.
5661 * @size: size of allocation, same value as passed to alloc_pages_exact().
5663 * Release the memory allocated by a previous call to alloc_pages_exact.
5665 void free_pages_exact(void *virt, size_t size)
5667 unsigned long addr = (unsigned long)virt;
5668 unsigned long end = addr + PAGE_ALIGN(size);
5670 while (addr < end) {
5675 EXPORT_SYMBOL(free_pages_exact);
5678 * nr_free_zone_pages - count number of pages beyond high watermark
5679 * @offset: The zone index of the highest zone
5681 * nr_free_zone_pages() counts the number of pages which are beyond the
5682 * high watermark within all zones at or below a given zone index. For each
5683 * zone, the number of pages is calculated as:
5685 * nr_free_zone_pages = managed_pages - high_pages
5687 * Return: number of pages beyond high watermark.
5689 static unsigned long nr_free_zone_pages(int offset)
5694 /* Just pick one node, since fallback list is circular */
5695 unsigned long sum = 0;
5697 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5699 for_each_zone_zonelist(zone, z, zonelist, offset) {
5700 unsigned long size = zone_managed_pages(zone);
5701 unsigned long high = high_wmark_pages(zone);
5710 * nr_free_buffer_pages - count number of pages beyond high watermark
5712 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5713 * watermark within ZONE_DMA and ZONE_NORMAL.
5715 * Return: number of pages beyond high watermark within ZONE_DMA and
5718 unsigned long nr_free_buffer_pages(void)
5720 return nr_free_zone_pages(gfp_zone(GFP_USER));
5722 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5724 static inline void show_node(struct zone *zone)
5726 if (IS_ENABLED(CONFIG_NUMA))
5727 printk("Node %d ", zone_to_nid(zone));
5730 long si_mem_available(void)
5733 unsigned long pagecache;
5734 unsigned long wmark_low = 0;
5735 unsigned long pages[NR_LRU_LISTS];
5736 unsigned long reclaimable;
5740 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5741 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5744 wmark_low += low_wmark_pages(zone);
5747 * Estimate the amount of memory available for userspace allocations,
5748 * without causing swapping.
5750 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5753 * Not all the page cache can be freed, otherwise the system will
5754 * start swapping. Assume at least half of the page cache, or the
5755 * low watermark worth of cache, needs to stay.
5757 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5758 pagecache -= min(pagecache / 2, wmark_low);
5759 available += pagecache;
5762 * Part of the reclaimable slab and other kernel memory consists of
5763 * items that are in use, and cannot be freed. Cap this estimate at the
5766 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5767 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5768 available += reclaimable - min(reclaimable / 2, wmark_low);
5774 EXPORT_SYMBOL_GPL(si_mem_available);
5776 void si_meminfo(struct sysinfo *val)
5778 val->totalram = totalram_pages();
5779 val->sharedram = global_node_page_state(NR_SHMEM);
5780 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5781 val->bufferram = nr_blockdev_pages();
5782 val->totalhigh = totalhigh_pages();
5783 val->freehigh = nr_free_highpages();
5784 val->mem_unit = PAGE_SIZE;
5787 EXPORT_SYMBOL(si_meminfo);
5790 void si_meminfo_node(struct sysinfo *val, int nid)
5792 int zone_type; /* needs to be signed */
5793 unsigned long managed_pages = 0;
5794 unsigned long managed_highpages = 0;
5795 unsigned long free_highpages = 0;
5796 pg_data_t *pgdat = NODE_DATA(nid);
5798 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5799 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5800 val->totalram = managed_pages;
5801 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5802 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5803 #ifdef CONFIG_HIGHMEM
5804 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5805 struct zone *zone = &pgdat->node_zones[zone_type];
5807 if (is_highmem(zone)) {
5808 managed_highpages += zone_managed_pages(zone);
5809 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5812 val->totalhigh = managed_highpages;
5813 val->freehigh = free_highpages;
5815 val->totalhigh = managed_highpages;
5816 val->freehigh = free_highpages;
5818 val->mem_unit = PAGE_SIZE;
5823 * Determine whether the node should be displayed or not, depending on whether
5824 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5826 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5828 if (!(flags & SHOW_MEM_FILTER_NODES))
5832 * no node mask - aka implicit memory numa policy. Do not bother with
5833 * the synchronization - read_mems_allowed_begin - because we do not
5834 * have to be precise here.
5837 nodemask = &cpuset_current_mems_allowed;
5839 return !node_isset(nid, *nodemask);
5842 #define K(x) ((x) << (PAGE_SHIFT-10))
5844 static void show_migration_types(unsigned char type)
5846 static const char types[MIGRATE_TYPES] = {
5847 [MIGRATE_UNMOVABLE] = 'U',
5848 [MIGRATE_MOVABLE] = 'M',
5849 [MIGRATE_RECLAIMABLE] = 'E',
5850 [MIGRATE_HIGHATOMIC] = 'H',
5852 [MIGRATE_CMA] = 'C',
5854 #ifdef CONFIG_MEMORY_ISOLATION
5855 [MIGRATE_ISOLATE] = 'I',
5858 char tmp[MIGRATE_TYPES + 1];
5862 for (i = 0; i < MIGRATE_TYPES; i++) {
5863 if (type & (1 << i))
5868 printk(KERN_CONT "(%s) ", tmp);
5872 * Show free area list (used inside shift_scroll-lock stuff)
5873 * We also calculate the percentage fragmentation. We do this by counting the
5874 * memory on each free list with the exception of the first item on the list.
5877 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5880 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5882 unsigned long free_pcp = 0;
5887 for_each_populated_zone(zone) {
5888 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5891 for_each_online_cpu(cpu)
5892 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5895 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5896 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5897 " unevictable:%lu dirty:%lu writeback:%lu\n"
5898 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5899 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5900 " kernel_misc_reclaimable:%lu\n"
5901 " free:%lu free_pcp:%lu free_cma:%lu\n",
5902 global_node_page_state(NR_ACTIVE_ANON),
5903 global_node_page_state(NR_INACTIVE_ANON),
5904 global_node_page_state(NR_ISOLATED_ANON),
5905 global_node_page_state(NR_ACTIVE_FILE),
5906 global_node_page_state(NR_INACTIVE_FILE),
5907 global_node_page_state(NR_ISOLATED_FILE),
5908 global_node_page_state(NR_UNEVICTABLE),
5909 global_node_page_state(NR_FILE_DIRTY),
5910 global_node_page_state(NR_WRITEBACK),
5911 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5912 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5913 global_node_page_state(NR_FILE_MAPPED),
5914 global_node_page_state(NR_SHMEM),
5915 global_node_page_state(NR_PAGETABLE),
5916 global_zone_page_state(NR_BOUNCE),
5917 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5918 global_zone_page_state(NR_FREE_PAGES),
5920 global_zone_page_state(NR_FREE_CMA_PAGES));
5922 for_each_online_pgdat(pgdat) {
5923 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5927 " active_anon:%lukB"
5928 " inactive_anon:%lukB"
5929 " active_file:%lukB"
5930 " inactive_file:%lukB"
5931 " unevictable:%lukB"
5932 " isolated(anon):%lukB"
5933 " isolated(file):%lukB"
5938 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5940 " shmem_pmdmapped: %lukB"
5943 " writeback_tmp:%lukB"
5944 " kernel_stack:%lukB"
5945 #ifdef CONFIG_SHADOW_CALL_STACK
5946 " shadow_call_stack:%lukB"
5949 " all_unreclaimable? %s"
5952 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5953 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5954 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5955 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5956 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5957 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5958 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5959 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5960 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5961 K(node_page_state(pgdat, NR_WRITEBACK)),
5962 K(node_page_state(pgdat, NR_SHMEM)),
5963 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5964 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5965 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5966 K(node_page_state(pgdat, NR_ANON_THPS)),
5968 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5969 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5970 #ifdef CONFIG_SHADOW_CALL_STACK
5971 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5973 K(node_page_state(pgdat, NR_PAGETABLE)),
5974 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5978 for_each_populated_zone(zone) {
5981 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5985 for_each_online_cpu(cpu)
5986 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5996 " reserved_highatomic:%luKB"
5997 " active_anon:%lukB"
5998 " inactive_anon:%lukB"
5999 " active_file:%lukB"
6000 " inactive_file:%lukB"
6001 " unevictable:%lukB"
6002 " writepending:%lukB"
6012 K(zone_page_state(zone, NR_FREE_PAGES)),
6013 K(zone->watermark_boost),
6014 K(min_wmark_pages(zone)),
6015 K(low_wmark_pages(zone)),
6016 K(high_wmark_pages(zone)),
6017 K(zone->nr_reserved_highatomic),
6018 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6019 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6020 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6021 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6022 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6023 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6024 K(zone->present_pages),
6025 K(zone_managed_pages(zone)),
6026 K(zone_page_state(zone, NR_MLOCK)),
6027 K(zone_page_state(zone, NR_BOUNCE)),
6029 K(this_cpu_read(zone->per_cpu_pageset->count)),
6030 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6031 printk("lowmem_reserve[]:");
6032 for (i = 0; i < MAX_NR_ZONES; i++)
6033 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6034 printk(KERN_CONT "\n");
6037 for_each_populated_zone(zone) {
6039 unsigned long nr[MAX_ORDER], flags, total = 0;
6040 unsigned char types[MAX_ORDER];
6042 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6045 printk(KERN_CONT "%s: ", zone->name);
6047 spin_lock_irqsave(&zone->lock, flags);
6048 for (order = 0; order < MAX_ORDER; order++) {
6049 struct free_area *area = &zone->free_area[order];
6052 nr[order] = area->nr_free;
6053 total += nr[order] << order;
6056 for (type = 0; type < MIGRATE_TYPES; type++) {
6057 if (!free_area_empty(area, type))
6058 types[order] |= 1 << type;
6061 spin_unlock_irqrestore(&zone->lock, flags);
6062 for (order = 0; order < MAX_ORDER; order++) {
6063 printk(KERN_CONT "%lu*%lukB ",
6064 nr[order], K(1UL) << order);
6066 show_migration_types(types[order]);
6068 printk(KERN_CONT "= %lukB\n", K(total));
6071 hugetlb_show_meminfo();
6073 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6075 show_swap_cache_info();
6078 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6080 zoneref->zone = zone;
6081 zoneref->zone_idx = zone_idx(zone);
6085 * Builds allocation fallback zone lists.
6087 * Add all populated zones of a node to the zonelist.
6089 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6092 enum zone_type zone_type = MAX_NR_ZONES;
6097 zone = pgdat->node_zones + zone_type;
6098 if (managed_zone(zone)) {
6099 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6100 check_highest_zone(zone_type);
6102 } while (zone_type);
6109 static int __parse_numa_zonelist_order(char *s)
6112 * We used to support different zonelists modes but they turned
6113 * out to be just not useful. Let's keep the warning in place
6114 * if somebody still use the cmd line parameter so that we do
6115 * not fail it silently
6117 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6118 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6124 char numa_zonelist_order[] = "Node";
6127 * sysctl handler for numa_zonelist_order
6129 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6130 void *buffer, size_t *length, loff_t *ppos)
6133 return __parse_numa_zonelist_order(buffer);
6134 return proc_dostring(table, write, buffer, length, ppos);
6138 #define MAX_NODE_LOAD (nr_online_nodes)
6139 static int node_load[MAX_NUMNODES];
6142 * find_next_best_node - find the next node that should appear in a given node's fallback list
6143 * @node: node whose fallback list we're appending
6144 * @used_node_mask: nodemask_t of already used nodes
6146 * We use a number of factors to determine which is the next node that should
6147 * appear on a given node's fallback list. The node should not have appeared
6148 * already in @node's fallback list, and it should be the next closest node
6149 * according to the distance array (which contains arbitrary distance values
6150 * from each node to each node in the system), and should also prefer nodes
6151 * with no CPUs, since presumably they'll have very little allocation pressure
6152 * on them otherwise.
6154 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6156 int find_next_best_node(int node, nodemask_t *used_node_mask)
6159 int min_val = INT_MAX;
6160 int best_node = NUMA_NO_NODE;
6162 /* Use the local node if we haven't already */
6163 if (!node_isset(node, *used_node_mask)) {
6164 node_set(node, *used_node_mask);
6168 for_each_node_state(n, N_MEMORY) {
6170 /* Don't want a node to appear more than once */
6171 if (node_isset(n, *used_node_mask))
6174 /* Use the distance array to find the distance */
6175 val = node_distance(node, n);
6177 /* Penalize nodes under us ("prefer the next node") */
6180 /* Give preference to headless and unused nodes */
6181 if (!cpumask_empty(cpumask_of_node(n)))
6182 val += PENALTY_FOR_NODE_WITH_CPUS;
6184 /* Slight preference for less loaded node */
6185 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6186 val += node_load[n];
6188 if (val < min_val) {
6195 node_set(best_node, *used_node_mask);
6202 * Build zonelists ordered by node and zones within node.
6203 * This results in maximum locality--normal zone overflows into local
6204 * DMA zone, if any--but risks exhausting DMA zone.
6206 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6209 struct zoneref *zonerefs;
6212 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6214 for (i = 0; i < nr_nodes; i++) {
6217 pg_data_t *node = NODE_DATA(node_order[i]);
6219 nr_zones = build_zonerefs_node(node, zonerefs);
6220 zonerefs += nr_zones;
6222 zonerefs->zone = NULL;
6223 zonerefs->zone_idx = 0;
6227 * Build gfp_thisnode zonelists
6229 static void build_thisnode_zonelists(pg_data_t *pgdat)
6231 struct zoneref *zonerefs;
6234 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6235 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6236 zonerefs += nr_zones;
6237 zonerefs->zone = NULL;
6238 zonerefs->zone_idx = 0;
6242 * Build zonelists ordered by zone and nodes within zones.
6243 * This results in conserving DMA zone[s] until all Normal memory is
6244 * exhausted, but results in overflowing to remote node while memory
6245 * may still exist in local DMA zone.
6248 static void build_zonelists(pg_data_t *pgdat)
6250 static int node_order[MAX_NUMNODES];
6251 int node, load, nr_nodes = 0;
6252 nodemask_t used_mask = NODE_MASK_NONE;
6253 int local_node, prev_node;
6255 /* NUMA-aware ordering of nodes */
6256 local_node = pgdat->node_id;
6257 load = nr_online_nodes;
6258 prev_node = local_node;
6260 memset(node_order, 0, sizeof(node_order));
6261 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6263 * We don't want to pressure a particular node.
6264 * So adding penalty to the first node in same
6265 * distance group to make it round-robin.
6267 if (node_distance(local_node, node) !=
6268 node_distance(local_node, prev_node))
6269 node_load[node] += load;
6271 node_order[nr_nodes++] = node;
6276 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6277 build_thisnode_zonelists(pgdat);
6278 pr_info("Fallback order for Node %d: ", local_node);
6279 for (node = 0; node < nr_nodes; node++)
6280 pr_cont("%d ", node_order[node]);
6284 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6286 * Return node id of node used for "local" allocations.
6287 * I.e., first node id of first zone in arg node's generic zonelist.
6288 * Used for initializing percpu 'numa_mem', which is used primarily
6289 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6291 int local_memory_node(int node)
6295 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6296 gfp_zone(GFP_KERNEL),
6298 return zone_to_nid(z->zone);
6302 static void setup_min_unmapped_ratio(void);
6303 static void setup_min_slab_ratio(void);
6304 #else /* CONFIG_NUMA */
6306 static void build_zonelists(pg_data_t *pgdat)
6308 int node, local_node;
6309 struct zoneref *zonerefs;
6312 local_node = pgdat->node_id;
6314 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6315 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6316 zonerefs += nr_zones;
6319 * Now we build the zonelist so that it contains the zones
6320 * of all the other nodes.
6321 * We don't want to pressure a particular node, so when
6322 * building the zones for node N, we make sure that the
6323 * zones coming right after the local ones are those from
6324 * node N+1 (modulo N)
6326 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6327 if (!node_online(node))
6329 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6330 zonerefs += nr_zones;
6332 for (node = 0; node < local_node; node++) {
6333 if (!node_online(node))
6335 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6336 zonerefs += nr_zones;
6339 zonerefs->zone = NULL;
6340 zonerefs->zone_idx = 0;
6343 #endif /* CONFIG_NUMA */
6346 * Boot pageset table. One per cpu which is going to be used for all
6347 * zones and all nodes. The parameters will be set in such a way
6348 * that an item put on a list will immediately be handed over to
6349 * the buddy list. This is safe since pageset manipulation is done
6350 * with interrupts disabled.
6352 * The boot_pagesets must be kept even after bootup is complete for
6353 * unused processors and/or zones. They do play a role for bootstrapping
6354 * hotplugged processors.
6356 * zoneinfo_show() and maybe other functions do
6357 * not check if the processor is online before following the pageset pointer.
6358 * Other parts of the kernel may not check if the zone is available.
6360 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6361 /* These effectively disable the pcplists in the boot pageset completely */
6362 #define BOOT_PAGESET_HIGH 0
6363 #define BOOT_PAGESET_BATCH 1
6364 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6365 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6366 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6368 static void __build_all_zonelists(void *data)
6371 int __maybe_unused cpu;
6372 pg_data_t *self = data;
6373 static DEFINE_SPINLOCK(lock);
6378 memset(node_load, 0, sizeof(node_load));
6382 * This node is hotadded and no memory is yet present. So just
6383 * building zonelists is fine - no need to touch other nodes.
6385 if (self && !node_online(self->node_id)) {
6386 build_zonelists(self);
6388 for_each_online_node(nid) {
6389 pg_data_t *pgdat = NODE_DATA(nid);
6391 build_zonelists(pgdat);
6394 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6396 * We now know the "local memory node" for each node--
6397 * i.e., the node of the first zone in the generic zonelist.
6398 * Set up numa_mem percpu variable for on-line cpus. During
6399 * boot, only the boot cpu should be on-line; we'll init the
6400 * secondary cpus' numa_mem as they come on-line. During
6401 * node/memory hotplug, we'll fixup all on-line cpus.
6403 for_each_online_cpu(cpu)
6404 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6411 static noinline void __init
6412 build_all_zonelists_init(void)
6416 __build_all_zonelists(NULL);
6419 * Initialize the boot_pagesets that are going to be used
6420 * for bootstrapping processors. The real pagesets for
6421 * each zone will be allocated later when the per cpu
6422 * allocator is available.
6424 * boot_pagesets are used also for bootstrapping offline
6425 * cpus if the system is already booted because the pagesets
6426 * are needed to initialize allocators on a specific cpu too.
6427 * F.e. the percpu allocator needs the page allocator which
6428 * needs the percpu allocator in order to allocate its pagesets
6429 * (a chicken-egg dilemma).
6431 for_each_possible_cpu(cpu)
6432 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6434 mminit_verify_zonelist();
6435 cpuset_init_current_mems_allowed();
6439 * unless system_state == SYSTEM_BOOTING.
6441 * __ref due to call of __init annotated helper build_all_zonelists_init
6442 * [protected by SYSTEM_BOOTING].
6444 void __ref build_all_zonelists(pg_data_t *pgdat)
6446 unsigned long vm_total_pages;
6448 if (system_state == SYSTEM_BOOTING) {
6449 build_all_zonelists_init();
6451 __build_all_zonelists(pgdat);
6452 /* cpuset refresh routine should be here */
6454 /* Get the number of free pages beyond high watermark in all zones. */
6455 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6457 * Disable grouping by mobility if the number of pages in the
6458 * system is too low to allow the mechanism to work. It would be
6459 * more accurate, but expensive to check per-zone. This check is
6460 * made on memory-hotadd so a system can start with mobility
6461 * disabled and enable it later
6463 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6464 page_group_by_mobility_disabled = 1;
6466 page_group_by_mobility_disabled = 0;
6468 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6470 page_group_by_mobility_disabled ? "off" : "on",
6473 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6477 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6478 static bool __meminit
6479 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6481 static struct memblock_region *r;
6483 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6484 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6485 for_each_mem_region(r) {
6486 if (*pfn < memblock_region_memory_end_pfn(r))
6490 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6491 memblock_is_mirror(r)) {
6492 *pfn = memblock_region_memory_end_pfn(r);
6500 * Initially all pages are reserved - free ones are freed
6501 * up by memblock_free_all() once the early boot process is
6502 * done. Non-atomic initialization, single-pass.
6504 * All aligned pageblocks are initialized to the specified migratetype
6505 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6506 * zone stats (e.g., nr_isolate_pageblock) are touched.
6508 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6509 unsigned long start_pfn, unsigned long zone_end_pfn,
6510 enum meminit_context context,
6511 struct vmem_altmap *altmap, int migratetype)
6513 unsigned long pfn, end_pfn = start_pfn + size;
6516 if (highest_memmap_pfn < end_pfn - 1)
6517 highest_memmap_pfn = end_pfn - 1;
6519 #ifdef CONFIG_ZONE_DEVICE
6521 * Honor reservation requested by the driver for this ZONE_DEVICE
6522 * memory. We limit the total number of pages to initialize to just
6523 * those that might contain the memory mapping. We will defer the
6524 * ZONE_DEVICE page initialization until after we have released
6527 if (zone == ZONE_DEVICE) {
6531 if (start_pfn == altmap->base_pfn)
6532 start_pfn += altmap->reserve;
6533 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6537 for (pfn = start_pfn; pfn < end_pfn; ) {
6539 * There can be holes in boot-time mem_map[]s handed to this
6540 * function. They do not exist on hotplugged memory.
6542 if (context == MEMINIT_EARLY) {
6543 if (overlap_memmap_init(zone, &pfn))
6545 if (defer_init(nid, pfn, zone_end_pfn))
6549 page = pfn_to_page(pfn);
6550 __init_single_page(page, pfn, zone, nid);
6551 if (context == MEMINIT_HOTPLUG)
6552 __SetPageReserved(page);
6555 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6556 * such that unmovable allocations won't be scattered all
6557 * over the place during system boot.
6559 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6560 set_pageblock_migratetype(page, migratetype);
6567 #ifdef CONFIG_ZONE_DEVICE
6568 void __ref memmap_init_zone_device(struct zone *zone,
6569 unsigned long start_pfn,
6570 unsigned long nr_pages,
6571 struct dev_pagemap *pgmap)
6573 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6574 struct pglist_data *pgdat = zone->zone_pgdat;
6575 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6576 unsigned long zone_idx = zone_idx(zone);
6577 unsigned long start = jiffies;
6578 int nid = pgdat->node_id;
6580 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6584 * The call to memmap_init should have already taken care
6585 * of the pages reserved for the memmap, so we can just jump to
6586 * the end of that region and start processing the device pages.
6589 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6590 nr_pages = end_pfn - start_pfn;
6593 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6594 struct page *page = pfn_to_page(pfn);
6596 __init_single_page(page, pfn, zone_idx, nid);
6599 * Mark page reserved as it will need to wait for onlining
6600 * phase for it to be fully associated with a zone.
6602 * We can use the non-atomic __set_bit operation for setting
6603 * the flag as we are still initializing the pages.
6605 __SetPageReserved(page);
6608 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6609 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6610 * ever freed or placed on a driver-private list.
6612 page->pgmap = pgmap;
6613 page->zone_device_data = NULL;
6616 * Mark the block movable so that blocks are reserved for
6617 * movable at startup. This will force kernel allocations
6618 * to reserve their blocks rather than leaking throughout
6619 * the address space during boot when many long-lived
6620 * kernel allocations are made.
6622 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6623 * because this is done early in section_activate()
6625 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6626 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6631 pr_info("%s initialised %lu pages in %ums\n", __func__,
6632 nr_pages, jiffies_to_msecs(jiffies - start));
6636 static void __meminit zone_init_free_lists(struct zone *zone)
6638 unsigned int order, t;
6639 for_each_migratetype_order(order, t) {
6640 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6641 zone->free_area[order].nr_free = 0;
6646 * Only struct pages that correspond to ranges defined by memblock.memory
6647 * are zeroed and initialized by going through __init_single_page() during
6648 * memmap_init_zone_range().
6650 * But, there could be struct pages that correspond to holes in
6651 * memblock.memory. This can happen because of the following reasons:
6652 * - physical memory bank size is not necessarily the exact multiple of the
6653 * arbitrary section size
6654 * - early reserved memory may not be listed in memblock.memory
6655 * - memory layouts defined with memmap= kernel parameter may not align
6656 * nicely with memmap sections
6658 * Explicitly initialize those struct pages so that:
6659 * - PG_Reserved is set
6660 * - zone and node links point to zone and node that span the page if the
6661 * hole is in the middle of a zone
6662 * - zone and node links point to adjacent zone/node if the hole falls on
6663 * the zone boundary; the pages in such holes will be prepended to the
6664 * zone/node above the hole except for the trailing pages in the last
6665 * section that will be appended to the zone/node below.
6667 static void __init init_unavailable_range(unsigned long spfn,
6674 for (pfn = spfn; pfn < epfn; pfn++) {
6675 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6676 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6677 + pageblock_nr_pages - 1;
6680 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6681 __SetPageReserved(pfn_to_page(pfn));
6686 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6687 node, zone_names[zone], pgcnt);
6690 static void __init memmap_init_zone_range(struct zone *zone,
6691 unsigned long start_pfn,
6692 unsigned long end_pfn,
6693 unsigned long *hole_pfn)
6695 unsigned long zone_start_pfn = zone->zone_start_pfn;
6696 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6697 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6699 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6700 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6702 if (start_pfn >= end_pfn)
6705 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6706 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6708 if (*hole_pfn < start_pfn)
6709 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6711 *hole_pfn = end_pfn;
6714 static void __init memmap_init(void)
6716 unsigned long start_pfn, end_pfn;
6717 unsigned long hole_pfn = 0;
6718 int i, j, zone_id = 0, nid;
6720 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6721 struct pglist_data *node = NODE_DATA(nid);
6723 for (j = 0; j < MAX_NR_ZONES; j++) {
6724 struct zone *zone = node->node_zones + j;
6726 if (!populated_zone(zone))
6729 memmap_init_zone_range(zone, start_pfn, end_pfn,
6735 #ifdef CONFIG_SPARSEMEM
6737 * Initialize the memory map for hole in the range [memory_end,
6739 * Append the pages in this hole to the highest zone in the last
6741 * The call to init_unavailable_range() is outside the ifdef to
6742 * silence the compiler warining about zone_id set but not used;
6743 * for FLATMEM it is a nop anyway
6745 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6746 if (hole_pfn < end_pfn)
6748 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6751 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6752 phys_addr_t min_addr, int nid, bool exact_nid)
6757 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6758 MEMBLOCK_ALLOC_ACCESSIBLE,
6761 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6762 MEMBLOCK_ALLOC_ACCESSIBLE,
6765 if (ptr && size > 0)
6766 page_init_poison(ptr, size);
6771 static int zone_batchsize(struct zone *zone)
6777 * The number of pages to batch allocate is either ~0.1%
6778 * of the zone or 1MB, whichever is smaller. The batch
6779 * size is striking a balance between allocation latency
6780 * and zone lock contention.
6782 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6783 batch /= 4; /* We effectively *= 4 below */
6788 * Clamp the batch to a 2^n - 1 value. Having a power
6789 * of 2 value was found to be more likely to have
6790 * suboptimal cache aliasing properties in some cases.
6792 * For example if 2 tasks are alternately allocating
6793 * batches of pages, one task can end up with a lot
6794 * of pages of one half of the possible page colors
6795 * and the other with pages of the other colors.
6797 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6802 /* The deferral and batching of frees should be suppressed under NOMMU
6805 * The problem is that NOMMU needs to be able to allocate large chunks
6806 * of contiguous memory as there's no hardware page translation to
6807 * assemble apparent contiguous memory from discontiguous pages.
6809 * Queueing large contiguous runs of pages for batching, however,
6810 * causes the pages to actually be freed in smaller chunks. As there
6811 * can be a significant delay between the individual batches being
6812 * recycled, this leads to the once large chunks of space being
6813 * fragmented and becoming unavailable for high-order allocations.
6819 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6824 unsigned long total_pages;
6826 if (!percpu_pagelist_high_fraction) {
6828 * By default, the high value of the pcp is based on the zone
6829 * low watermark so that if they are full then background
6830 * reclaim will not be started prematurely.
6832 total_pages = low_wmark_pages(zone);
6835 * If percpu_pagelist_high_fraction is configured, the high
6836 * value is based on a fraction of the managed pages in the
6839 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6843 * Split the high value across all online CPUs local to the zone. Note
6844 * that early in boot that CPUs may not be online yet and that during
6845 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6846 * onlined. For memory nodes that have no CPUs, split pcp->high across
6847 * all online CPUs to mitigate the risk that reclaim is triggered
6848 * prematurely due to pages stored on pcp lists.
6850 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6852 nr_split_cpus = num_online_cpus();
6853 high = total_pages / nr_split_cpus;
6856 * Ensure high is at least batch*4. The multiple is based on the
6857 * historical relationship between high and batch.
6859 high = max(high, batch << 2);
6868 * pcp->high and pcp->batch values are related and generally batch is lower
6869 * than high. They are also related to pcp->count such that count is lower
6870 * than high, and as soon as it reaches high, the pcplist is flushed.
6872 * However, guaranteeing these relations at all times would require e.g. write
6873 * barriers here but also careful usage of read barriers at the read side, and
6874 * thus be prone to error and bad for performance. Thus the update only prevents
6875 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6876 * can cope with those fields changing asynchronously, and fully trust only the
6877 * pcp->count field on the local CPU with interrupts disabled.
6879 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6880 * outside of boot time (or some other assurance that no concurrent updaters
6883 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6884 unsigned long batch)
6886 WRITE_ONCE(pcp->batch, batch);
6887 WRITE_ONCE(pcp->high, high);
6890 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6894 memset(pcp, 0, sizeof(*pcp));
6895 memset(pzstats, 0, sizeof(*pzstats));
6897 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6898 INIT_LIST_HEAD(&pcp->lists[pindex]);
6901 * Set batch and high values safe for a boot pageset. A true percpu
6902 * pageset's initialization will update them subsequently. Here we don't
6903 * need to be as careful as pageset_update() as nobody can access the
6906 pcp->high = BOOT_PAGESET_HIGH;
6907 pcp->batch = BOOT_PAGESET_BATCH;
6908 pcp->free_factor = 0;
6911 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6912 unsigned long batch)
6914 struct per_cpu_pages *pcp;
6917 for_each_possible_cpu(cpu) {
6918 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6919 pageset_update(pcp, high, batch);
6924 * Calculate and set new high and batch values for all per-cpu pagesets of a
6925 * zone based on the zone's size.
6927 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6929 int new_high, new_batch;
6931 new_batch = max(1, zone_batchsize(zone));
6932 new_high = zone_highsize(zone, new_batch, cpu_online);
6934 if (zone->pageset_high == new_high &&
6935 zone->pageset_batch == new_batch)
6938 zone->pageset_high = new_high;
6939 zone->pageset_batch = new_batch;
6941 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6944 void __meminit setup_zone_pageset(struct zone *zone)
6948 /* Size may be 0 on !SMP && !NUMA */
6949 if (sizeof(struct per_cpu_zonestat) > 0)
6950 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6952 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6953 for_each_possible_cpu(cpu) {
6954 struct per_cpu_pages *pcp;
6955 struct per_cpu_zonestat *pzstats;
6957 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6958 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6959 per_cpu_pages_init(pcp, pzstats);
6962 zone_set_pageset_high_and_batch(zone, 0);
6966 * Allocate per cpu pagesets and initialize them.
6967 * Before this call only boot pagesets were available.
6969 void __init setup_per_cpu_pageset(void)
6971 struct pglist_data *pgdat;
6973 int __maybe_unused cpu;
6975 for_each_populated_zone(zone)
6976 setup_zone_pageset(zone);
6980 * Unpopulated zones continue using the boot pagesets.
6981 * The numa stats for these pagesets need to be reset.
6982 * Otherwise, they will end up skewing the stats of
6983 * the nodes these zones are associated with.
6985 for_each_possible_cpu(cpu) {
6986 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
6987 memset(pzstats->vm_numa_event, 0,
6988 sizeof(pzstats->vm_numa_event));
6992 for_each_online_pgdat(pgdat)
6993 pgdat->per_cpu_nodestats =
6994 alloc_percpu(struct per_cpu_nodestat);
6997 static __meminit void zone_pcp_init(struct zone *zone)
7000 * per cpu subsystem is not up at this point. The following code
7001 * relies on the ability of the linker to provide the
7002 * offset of a (static) per cpu variable into the per cpu area.
7004 zone->per_cpu_pageset = &boot_pageset;
7005 zone->per_cpu_zonestats = &boot_zonestats;
7006 zone->pageset_high = BOOT_PAGESET_HIGH;
7007 zone->pageset_batch = BOOT_PAGESET_BATCH;
7009 if (populated_zone(zone))
7010 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7011 zone->present_pages, zone_batchsize(zone));
7014 void __meminit init_currently_empty_zone(struct zone *zone,
7015 unsigned long zone_start_pfn,
7018 struct pglist_data *pgdat = zone->zone_pgdat;
7019 int zone_idx = zone_idx(zone) + 1;
7021 if (zone_idx > pgdat->nr_zones)
7022 pgdat->nr_zones = zone_idx;
7024 zone->zone_start_pfn = zone_start_pfn;
7026 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7027 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7029 (unsigned long)zone_idx(zone),
7030 zone_start_pfn, (zone_start_pfn + size));
7032 zone_init_free_lists(zone);
7033 zone->initialized = 1;
7037 * get_pfn_range_for_nid - Return the start and end page frames for a node
7038 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7039 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7040 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7042 * It returns the start and end page frame of a node based on information
7043 * provided by memblock_set_node(). If called for a node
7044 * with no available memory, a warning is printed and the start and end
7047 void __init get_pfn_range_for_nid(unsigned int nid,
7048 unsigned long *start_pfn, unsigned long *end_pfn)
7050 unsigned long this_start_pfn, this_end_pfn;
7056 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7057 *start_pfn = min(*start_pfn, this_start_pfn);
7058 *end_pfn = max(*end_pfn, this_end_pfn);
7061 if (*start_pfn == -1UL)
7066 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7067 * assumption is made that zones within a node are ordered in monotonic
7068 * increasing memory addresses so that the "highest" populated zone is used
7070 static void __init find_usable_zone_for_movable(void)
7073 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7074 if (zone_index == ZONE_MOVABLE)
7077 if (arch_zone_highest_possible_pfn[zone_index] >
7078 arch_zone_lowest_possible_pfn[zone_index])
7082 VM_BUG_ON(zone_index == -1);
7083 movable_zone = zone_index;
7087 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7088 * because it is sized independent of architecture. Unlike the other zones,
7089 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7090 * in each node depending on the size of each node and how evenly kernelcore
7091 * is distributed. This helper function adjusts the zone ranges
7092 * provided by the architecture for a given node by using the end of the
7093 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7094 * zones within a node are in order of monotonic increases memory addresses
7096 static void __init adjust_zone_range_for_zone_movable(int nid,
7097 unsigned long zone_type,
7098 unsigned long node_start_pfn,
7099 unsigned long node_end_pfn,
7100 unsigned long *zone_start_pfn,
7101 unsigned long *zone_end_pfn)
7103 /* Only adjust if ZONE_MOVABLE is on this node */
7104 if (zone_movable_pfn[nid]) {
7105 /* Size ZONE_MOVABLE */
7106 if (zone_type == ZONE_MOVABLE) {
7107 *zone_start_pfn = zone_movable_pfn[nid];
7108 *zone_end_pfn = min(node_end_pfn,
7109 arch_zone_highest_possible_pfn[movable_zone]);
7111 /* Adjust for ZONE_MOVABLE starting within this range */
7112 } else if (!mirrored_kernelcore &&
7113 *zone_start_pfn < zone_movable_pfn[nid] &&
7114 *zone_end_pfn > zone_movable_pfn[nid]) {
7115 *zone_end_pfn = zone_movable_pfn[nid];
7117 /* Check if this whole range is within ZONE_MOVABLE */
7118 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7119 *zone_start_pfn = *zone_end_pfn;
7124 * Return the number of pages a zone spans in a node, including holes
7125 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7127 static unsigned long __init zone_spanned_pages_in_node(int nid,
7128 unsigned long zone_type,
7129 unsigned long node_start_pfn,
7130 unsigned long node_end_pfn,
7131 unsigned long *zone_start_pfn,
7132 unsigned long *zone_end_pfn)
7134 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7135 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7136 /* When hotadd a new node from cpu_up(), the node should be empty */
7137 if (!node_start_pfn && !node_end_pfn)
7140 /* Get the start and end of the zone */
7141 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7142 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7143 adjust_zone_range_for_zone_movable(nid, zone_type,
7144 node_start_pfn, node_end_pfn,
7145 zone_start_pfn, zone_end_pfn);
7147 /* Check that this node has pages within the zone's required range */
7148 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7151 /* Move the zone boundaries inside the node if necessary */
7152 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7153 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7155 /* Return the spanned pages */
7156 return *zone_end_pfn - *zone_start_pfn;
7160 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7161 * then all holes in the requested range will be accounted for.
7163 unsigned long __init __absent_pages_in_range(int nid,
7164 unsigned long range_start_pfn,
7165 unsigned long range_end_pfn)
7167 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7168 unsigned long start_pfn, end_pfn;
7171 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7172 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7173 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7174 nr_absent -= end_pfn - start_pfn;
7180 * absent_pages_in_range - Return number of page frames in holes within a range
7181 * @start_pfn: The start PFN to start searching for holes
7182 * @end_pfn: The end PFN to stop searching for holes
7184 * Return: the number of pages frames in memory holes within a range.
7186 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7187 unsigned long end_pfn)
7189 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7192 /* Return the number of page frames in holes in a zone on a node */
7193 static unsigned long __init zone_absent_pages_in_node(int nid,
7194 unsigned long zone_type,
7195 unsigned long node_start_pfn,
7196 unsigned long node_end_pfn)
7198 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7199 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7200 unsigned long zone_start_pfn, zone_end_pfn;
7201 unsigned long nr_absent;
7203 /* When hotadd a new node from cpu_up(), the node should be empty */
7204 if (!node_start_pfn && !node_end_pfn)
7207 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7208 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7210 adjust_zone_range_for_zone_movable(nid, zone_type,
7211 node_start_pfn, node_end_pfn,
7212 &zone_start_pfn, &zone_end_pfn);
7213 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7216 * ZONE_MOVABLE handling.
7217 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7220 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7221 unsigned long start_pfn, end_pfn;
7222 struct memblock_region *r;
7224 for_each_mem_region(r) {
7225 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7226 zone_start_pfn, zone_end_pfn);
7227 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7228 zone_start_pfn, zone_end_pfn);
7230 if (zone_type == ZONE_MOVABLE &&
7231 memblock_is_mirror(r))
7232 nr_absent += end_pfn - start_pfn;
7234 if (zone_type == ZONE_NORMAL &&
7235 !memblock_is_mirror(r))
7236 nr_absent += end_pfn - start_pfn;
7243 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7244 unsigned long node_start_pfn,
7245 unsigned long node_end_pfn)
7247 unsigned long realtotalpages = 0, totalpages = 0;
7250 for (i = 0; i < MAX_NR_ZONES; i++) {
7251 struct zone *zone = pgdat->node_zones + i;
7252 unsigned long zone_start_pfn, zone_end_pfn;
7253 unsigned long spanned, absent;
7254 unsigned long size, real_size;
7256 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7261 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7266 real_size = size - absent;
7269 zone->zone_start_pfn = zone_start_pfn;
7271 zone->zone_start_pfn = 0;
7272 zone->spanned_pages = size;
7273 zone->present_pages = real_size;
7274 #if defined(CONFIG_MEMORY_HOTPLUG)
7275 zone->present_early_pages = real_size;
7279 realtotalpages += real_size;
7282 pgdat->node_spanned_pages = totalpages;
7283 pgdat->node_present_pages = realtotalpages;
7284 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7287 #ifndef CONFIG_SPARSEMEM
7289 * Calculate the size of the zone->blockflags rounded to an unsigned long
7290 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7291 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7292 * round what is now in bits to nearest long in bits, then return it in
7295 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7297 unsigned long usemapsize;
7299 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7300 usemapsize = roundup(zonesize, pageblock_nr_pages);
7301 usemapsize = usemapsize >> pageblock_order;
7302 usemapsize *= NR_PAGEBLOCK_BITS;
7303 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7305 return usemapsize / 8;
7308 static void __ref setup_usemap(struct zone *zone)
7310 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7311 zone->spanned_pages);
7312 zone->pageblock_flags = NULL;
7314 zone->pageblock_flags =
7315 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7317 if (!zone->pageblock_flags)
7318 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7319 usemapsize, zone->name, zone_to_nid(zone));
7323 static inline void setup_usemap(struct zone *zone) {}
7324 #endif /* CONFIG_SPARSEMEM */
7326 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7328 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7329 void __init set_pageblock_order(void)
7333 /* Check that pageblock_nr_pages has not already been setup */
7334 if (pageblock_order)
7337 if (HPAGE_SHIFT > PAGE_SHIFT)
7338 order = HUGETLB_PAGE_ORDER;
7340 order = MAX_ORDER - 1;
7343 * Assume the largest contiguous order of interest is a huge page.
7344 * This value may be variable depending on boot parameters on IA64 and
7347 pageblock_order = order;
7349 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7352 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7353 * is unused as pageblock_order is set at compile-time. See
7354 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7357 void __init set_pageblock_order(void)
7361 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7363 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7364 unsigned long present_pages)
7366 unsigned long pages = spanned_pages;
7369 * Provide a more accurate estimation if there are holes within
7370 * the zone and SPARSEMEM is in use. If there are holes within the
7371 * zone, each populated memory region may cost us one or two extra
7372 * memmap pages due to alignment because memmap pages for each
7373 * populated regions may not be naturally aligned on page boundary.
7374 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7376 if (spanned_pages > present_pages + (present_pages >> 4) &&
7377 IS_ENABLED(CONFIG_SPARSEMEM))
7378 pages = present_pages;
7380 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7383 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7384 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7386 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7388 spin_lock_init(&ds_queue->split_queue_lock);
7389 INIT_LIST_HEAD(&ds_queue->split_queue);
7390 ds_queue->split_queue_len = 0;
7393 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7396 #ifdef CONFIG_COMPACTION
7397 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7399 init_waitqueue_head(&pgdat->kcompactd_wait);
7402 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7405 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7409 pgdat_resize_init(pgdat);
7411 pgdat_init_split_queue(pgdat);
7412 pgdat_init_kcompactd(pgdat);
7414 init_waitqueue_head(&pgdat->kswapd_wait);
7415 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7417 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7418 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7420 pgdat_page_ext_init(pgdat);
7421 lruvec_init(&pgdat->__lruvec);
7424 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7425 unsigned long remaining_pages)
7427 atomic_long_set(&zone->managed_pages, remaining_pages);
7428 zone_set_nid(zone, nid);
7429 zone->name = zone_names[idx];
7430 zone->zone_pgdat = NODE_DATA(nid);
7431 spin_lock_init(&zone->lock);
7432 zone_seqlock_init(zone);
7433 zone_pcp_init(zone);
7437 * Set up the zone data structures
7438 * - init pgdat internals
7439 * - init all zones belonging to this node
7441 * NOTE: this function is only called during memory hotplug
7443 #ifdef CONFIG_MEMORY_HOTPLUG
7444 void __ref free_area_init_core_hotplug(int nid)
7447 pg_data_t *pgdat = NODE_DATA(nid);
7449 pgdat_init_internals(pgdat);
7450 for (z = 0; z < MAX_NR_ZONES; z++)
7451 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7456 * Set up the zone data structures:
7457 * - mark all pages reserved
7458 * - mark all memory queues empty
7459 * - clear the memory bitmaps
7461 * NOTE: pgdat should get zeroed by caller.
7462 * NOTE: this function is only called during early init.
7464 static void __init free_area_init_core(struct pglist_data *pgdat)
7467 int nid = pgdat->node_id;
7469 pgdat_init_internals(pgdat);
7470 pgdat->per_cpu_nodestats = &boot_nodestats;
7472 for (j = 0; j < MAX_NR_ZONES; j++) {
7473 struct zone *zone = pgdat->node_zones + j;
7474 unsigned long size, freesize, memmap_pages;
7476 size = zone->spanned_pages;
7477 freesize = zone->present_pages;
7480 * Adjust freesize so that it accounts for how much memory
7481 * is used by this zone for memmap. This affects the watermark
7482 * and per-cpu initialisations
7484 memmap_pages = calc_memmap_size(size, freesize);
7485 if (!is_highmem_idx(j)) {
7486 if (freesize >= memmap_pages) {
7487 freesize -= memmap_pages;
7489 pr_debug(" %s zone: %lu pages used for memmap\n",
7490 zone_names[j], memmap_pages);
7492 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7493 zone_names[j], memmap_pages, freesize);
7496 /* Account for reserved pages */
7497 if (j == 0 && freesize > dma_reserve) {
7498 freesize -= dma_reserve;
7499 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7502 if (!is_highmem_idx(j))
7503 nr_kernel_pages += freesize;
7504 /* Charge for highmem memmap if there are enough kernel pages */
7505 else if (nr_kernel_pages > memmap_pages * 2)
7506 nr_kernel_pages -= memmap_pages;
7507 nr_all_pages += freesize;
7510 * Set an approximate value for lowmem here, it will be adjusted
7511 * when the bootmem allocator frees pages into the buddy system.
7512 * And all highmem pages will be managed by the buddy system.
7514 zone_init_internals(zone, j, nid, freesize);
7519 set_pageblock_order();
7521 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7525 #ifdef CONFIG_FLATMEM
7526 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7528 unsigned long __maybe_unused start = 0;
7529 unsigned long __maybe_unused offset = 0;
7531 /* Skip empty nodes */
7532 if (!pgdat->node_spanned_pages)
7535 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7536 offset = pgdat->node_start_pfn - start;
7537 /* ia64 gets its own node_mem_map, before this, without bootmem */
7538 if (!pgdat->node_mem_map) {
7539 unsigned long size, end;
7543 * The zone's endpoints aren't required to be MAX_ORDER
7544 * aligned but the node_mem_map endpoints must be in order
7545 * for the buddy allocator to function correctly.
7547 end = pgdat_end_pfn(pgdat);
7548 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7549 size = (end - start) * sizeof(struct page);
7550 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7551 pgdat->node_id, false);
7553 panic("Failed to allocate %ld bytes for node %d memory map\n",
7554 size, pgdat->node_id);
7555 pgdat->node_mem_map = map + offset;
7557 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7558 __func__, pgdat->node_id, (unsigned long)pgdat,
7559 (unsigned long)pgdat->node_mem_map);
7562 * With no DISCONTIG, the global mem_map is just set as node 0's
7564 if (pgdat == NODE_DATA(0)) {
7565 mem_map = NODE_DATA(0)->node_mem_map;
7566 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7572 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7573 #endif /* CONFIG_FLATMEM */
7575 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7576 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7578 pgdat->first_deferred_pfn = ULONG_MAX;
7581 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7584 static void __init free_area_init_node(int nid)
7586 pg_data_t *pgdat = NODE_DATA(nid);
7587 unsigned long start_pfn = 0;
7588 unsigned long end_pfn = 0;
7590 /* pg_data_t should be reset to zero when it's allocated */
7591 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7593 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7595 pgdat->node_id = nid;
7596 pgdat->node_start_pfn = start_pfn;
7597 pgdat->per_cpu_nodestats = NULL;
7599 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7600 (u64)start_pfn << PAGE_SHIFT,
7601 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7602 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7604 alloc_node_mem_map(pgdat);
7605 pgdat_set_deferred_range(pgdat);
7607 free_area_init_core(pgdat);
7610 void __init free_area_init_memoryless_node(int nid)
7612 free_area_init_node(nid);
7615 #if MAX_NUMNODES > 1
7617 * Figure out the number of possible node ids.
7619 void __init setup_nr_node_ids(void)
7621 unsigned int highest;
7623 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7624 nr_node_ids = highest + 1;
7629 * node_map_pfn_alignment - determine the maximum internode alignment
7631 * This function should be called after node map is populated and sorted.
7632 * It calculates the maximum power of two alignment which can distinguish
7635 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7636 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7637 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7638 * shifted, 1GiB is enough and this function will indicate so.
7640 * This is used to test whether pfn -> nid mapping of the chosen memory
7641 * model has fine enough granularity to avoid incorrect mapping for the
7642 * populated node map.
7644 * Return: the determined alignment in pfn's. 0 if there is no alignment
7645 * requirement (single node).
7647 unsigned long __init node_map_pfn_alignment(void)
7649 unsigned long accl_mask = 0, last_end = 0;
7650 unsigned long start, end, mask;
7651 int last_nid = NUMA_NO_NODE;
7654 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7655 if (!start || last_nid < 0 || last_nid == nid) {
7662 * Start with a mask granular enough to pin-point to the
7663 * start pfn and tick off bits one-by-one until it becomes
7664 * too coarse to separate the current node from the last.
7666 mask = ~((1 << __ffs(start)) - 1);
7667 while (mask && last_end <= (start & (mask << 1)))
7670 /* accumulate all internode masks */
7674 /* convert mask to number of pages */
7675 return ~accl_mask + 1;
7679 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7681 * Return: the minimum PFN based on information provided via
7682 * memblock_set_node().
7684 unsigned long __init find_min_pfn_with_active_regions(void)
7686 return PHYS_PFN(memblock_start_of_DRAM());
7690 * early_calculate_totalpages()
7691 * Sum pages in active regions for movable zone.
7692 * Populate N_MEMORY for calculating usable_nodes.
7694 static unsigned long __init early_calculate_totalpages(void)
7696 unsigned long totalpages = 0;
7697 unsigned long start_pfn, end_pfn;
7700 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7701 unsigned long pages = end_pfn - start_pfn;
7703 totalpages += pages;
7705 node_set_state(nid, N_MEMORY);
7711 * Find the PFN the Movable zone begins in each node. Kernel memory
7712 * is spread evenly between nodes as long as the nodes have enough
7713 * memory. When they don't, some nodes will have more kernelcore than
7716 static void __init find_zone_movable_pfns_for_nodes(void)
7719 unsigned long usable_startpfn;
7720 unsigned long kernelcore_node, kernelcore_remaining;
7721 /* save the state before borrow the nodemask */
7722 nodemask_t saved_node_state = node_states[N_MEMORY];
7723 unsigned long totalpages = early_calculate_totalpages();
7724 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7725 struct memblock_region *r;
7727 /* Need to find movable_zone earlier when movable_node is specified. */
7728 find_usable_zone_for_movable();
7731 * If movable_node is specified, ignore kernelcore and movablecore
7734 if (movable_node_is_enabled()) {
7735 for_each_mem_region(r) {
7736 if (!memblock_is_hotpluggable(r))
7739 nid = memblock_get_region_node(r);
7741 usable_startpfn = PFN_DOWN(r->base);
7742 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7743 min(usable_startpfn, zone_movable_pfn[nid]) :
7751 * If kernelcore=mirror is specified, ignore movablecore option
7753 if (mirrored_kernelcore) {
7754 bool mem_below_4gb_not_mirrored = false;
7756 for_each_mem_region(r) {
7757 if (memblock_is_mirror(r))
7760 nid = memblock_get_region_node(r);
7762 usable_startpfn = memblock_region_memory_base_pfn(r);
7764 if (usable_startpfn < 0x100000) {
7765 mem_below_4gb_not_mirrored = true;
7769 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7770 min(usable_startpfn, zone_movable_pfn[nid]) :
7774 if (mem_below_4gb_not_mirrored)
7775 pr_warn("This configuration results in unmirrored kernel memory.\n");
7781 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7782 * amount of necessary memory.
7784 if (required_kernelcore_percent)
7785 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7787 if (required_movablecore_percent)
7788 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7792 * If movablecore= was specified, calculate what size of
7793 * kernelcore that corresponds so that memory usable for
7794 * any allocation type is evenly spread. If both kernelcore
7795 * and movablecore are specified, then the value of kernelcore
7796 * will be used for required_kernelcore if it's greater than
7797 * what movablecore would have allowed.
7799 if (required_movablecore) {
7800 unsigned long corepages;
7803 * Round-up so that ZONE_MOVABLE is at least as large as what
7804 * was requested by the user
7806 required_movablecore =
7807 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7808 required_movablecore = min(totalpages, required_movablecore);
7809 corepages = totalpages - required_movablecore;
7811 required_kernelcore = max(required_kernelcore, corepages);
7815 * If kernelcore was not specified or kernelcore size is larger
7816 * than totalpages, there is no ZONE_MOVABLE.
7818 if (!required_kernelcore || required_kernelcore >= totalpages)
7821 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7822 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7825 /* Spread kernelcore memory as evenly as possible throughout nodes */
7826 kernelcore_node = required_kernelcore / usable_nodes;
7827 for_each_node_state(nid, N_MEMORY) {
7828 unsigned long start_pfn, end_pfn;
7831 * Recalculate kernelcore_node if the division per node
7832 * now exceeds what is necessary to satisfy the requested
7833 * amount of memory for the kernel
7835 if (required_kernelcore < kernelcore_node)
7836 kernelcore_node = required_kernelcore / usable_nodes;
7839 * As the map is walked, we track how much memory is usable
7840 * by the kernel using kernelcore_remaining. When it is
7841 * 0, the rest of the node is usable by ZONE_MOVABLE
7843 kernelcore_remaining = kernelcore_node;
7845 /* Go through each range of PFNs within this node */
7846 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7847 unsigned long size_pages;
7849 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7850 if (start_pfn >= end_pfn)
7853 /* Account for what is only usable for kernelcore */
7854 if (start_pfn < usable_startpfn) {
7855 unsigned long kernel_pages;
7856 kernel_pages = min(end_pfn, usable_startpfn)
7859 kernelcore_remaining -= min(kernel_pages,
7860 kernelcore_remaining);
7861 required_kernelcore -= min(kernel_pages,
7862 required_kernelcore);
7864 /* Continue if range is now fully accounted */
7865 if (end_pfn <= usable_startpfn) {
7868 * Push zone_movable_pfn to the end so
7869 * that if we have to rebalance
7870 * kernelcore across nodes, we will
7871 * not double account here
7873 zone_movable_pfn[nid] = end_pfn;
7876 start_pfn = usable_startpfn;
7880 * The usable PFN range for ZONE_MOVABLE is from
7881 * start_pfn->end_pfn. Calculate size_pages as the
7882 * number of pages used as kernelcore
7884 size_pages = end_pfn - start_pfn;
7885 if (size_pages > kernelcore_remaining)
7886 size_pages = kernelcore_remaining;
7887 zone_movable_pfn[nid] = start_pfn + size_pages;
7890 * Some kernelcore has been met, update counts and
7891 * break if the kernelcore for this node has been
7894 required_kernelcore -= min(required_kernelcore,
7896 kernelcore_remaining -= size_pages;
7897 if (!kernelcore_remaining)
7903 * If there is still required_kernelcore, we do another pass with one
7904 * less node in the count. This will push zone_movable_pfn[nid] further
7905 * along on the nodes that still have memory until kernelcore is
7909 if (usable_nodes && required_kernelcore > usable_nodes)
7913 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7914 for (nid = 0; nid < MAX_NUMNODES; nid++)
7915 zone_movable_pfn[nid] =
7916 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7919 /* restore the node_state */
7920 node_states[N_MEMORY] = saved_node_state;
7923 /* Any regular or high memory on that node ? */
7924 static void check_for_memory(pg_data_t *pgdat, int nid)
7926 enum zone_type zone_type;
7928 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7929 struct zone *zone = &pgdat->node_zones[zone_type];
7930 if (populated_zone(zone)) {
7931 if (IS_ENABLED(CONFIG_HIGHMEM))
7932 node_set_state(nid, N_HIGH_MEMORY);
7933 if (zone_type <= ZONE_NORMAL)
7934 node_set_state(nid, N_NORMAL_MEMORY);
7941 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7942 * such cases we allow max_zone_pfn sorted in the descending order
7944 bool __weak arch_has_descending_max_zone_pfns(void)
7950 * free_area_init - Initialise all pg_data_t and zone data
7951 * @max_zone_pfn: an array of max PFNs for each zone
7953 * This will call free_area_init_node() for each active node in the system.
7954 * Using the page ranges provided by memblock_set_node(), the size of each
7955 * zone in each node and their holes is calculated. If the maximum PFN
7956 * between two adjacent zones match, it is assumed that the zone is empty.
7957 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7958 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7959 * starts where the previous one ended. For example, ZONE_DMA32 starts
7960 * at arch_max_dma_pfn.
7962 void __init free_area_init(unsigned long *max_zone_pfn)
7964 unsigned long start_pfn, end_pfn;
7968 /* Record where the zone boundaries are */
7969 memset(arch_zone_lowest_possible_pfn, 0,
7970 sizeof(arch_zone_lowest_possible_pfn));
7971 memset(arch_zone_highest_possible_pfn, 0,
7972 sizeof(arch_zone_highest_possible_pfn));
7974 start_pfn = find_min_pfn_with_active_regions();
7975 descending = arch_has_descending_max_zone_pfns();
7977 for (i = 0; i < MAX_NR_ZONES; i++) {
7979 zone = MAX_NR_ZONES - i - 1;
7983 if (zone == ZONE_MOVABLE)
7986 end_pfn = max(max_zone_pfn[zone], start_pfn);
7987 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7988 arch_zone_highest_possible_pfn[zone] = end_pfn;
7990 start_pfn = end_pfn;
7993 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7994 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7995 find_zone_movable_pfns_for_nodes();
7997 /* Print out the zone ranges */
7998 pr_info("Zone ranges:\n");
7999 for (i = 0; i < MAX_NR_ZONES; i++) {
8000 if (i == ZONE_MOVABLE)
8002 pr_info(" %-8s ", zone_names[i]);
8003 if (arch_zone_lowest_possible_pfn[i] ==
8004 arch_zone_highest_possible_pfn[i])
8007 pr_cont("[mem %#018Lx-%#018Lx]\n",
8008 (u64)arch_zone_lowest_possible_pfn[i]
8010 ((u64)arch_zone_highest_possible_pfn[i]
8011 << PAGE_SHIFT) - 1);
8014 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8015 pr_info("Movable zone start for each node\n");
8016 for (i = 0; i < MAX_NUMNODES; i++) {
8017 if (zone_movable_pfn[i])
8018 pr_info(" Node %d: %#018Lx\n", i,
8019 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8023 * Print out the early node map, and initialize the
8024 * subsection-map relative to active online memory ranges to
8025 * enable future "sub-section" extensions of the memory map.
8027 pr_info("Early memory node ranges\n");
8028 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8029 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8030 (u64)start_pfn << PAGE_SHIFT,
8031 ((u64)end_pfn << PAGE_SHIFT) - 1);
8032 subsection_map_init(start_pfn, end_pfn - start_pfn);
8035 /* Initialise every node */
8036 mminit_verify_pageflags_layout();
8037 setup_nr_node_ids();
8038 for_each_online_node(nid) {
8039 pg_data_t *pgdat = NODE_DATA(nid);
8040 free_area_init_node(nid);
8042 /* Any memory on that node */
8043 if (pgdat->node_present_pages)
8044 node_set_state(nid, N_MEMORY);
8045 check_for_memory(pgdat, nid);
8051 static int __init cmdline_parse_core(char *p, unsigned long *core,
8052 unsigned long *percent)
8054 unsigned long long coremem;
8060 /* Value may be a percentage of total memory, otherwise bytes */
8061 coremem = simple_strtoull(p, &endptr, 0);
8062 if (*endptr == '%') {
8063 /* Paranoid check for percent values greater than 100 */
8064 WARN_ON(coremem > 100);
8068 coremem = memparse(p, &p);
8069 /* Paranoid check that UL is enough for the coremem value */
8070 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8072 *core = coremem >> PAGE_SHIFT;
8079 * kernelcore=size sets the amount of memory for use for allocations that
8080 * cannot be reclaimed or migrated.
8082 static int __init cmdline_parse_kernelcore(char *p)
8084 /* parse kernelcore=mirror */
8085 if (parse_option_str(p, "mirror")) {
8086 mirrored_kernelcore = true;
8090 return cmdline_parse_core(p, &required_kernelcore,
8091 &required_kernelcore_percent);
8095 * movablecore=size sets the amount of memory for use for allocations that
8096 * can be reclaimed or migrated.
8098 static int __init cmdline_parse_movablecore(char *p)
8100 return cmdline_parse_core(p, &required_movablecore,
8101 &required_movablecore_percent);
8104 early_param("kernelcore", cmdline_parse_kernelcore);
8105 early_param("movablecore", cmdline_parse_movablecore);
8107 void adjust_managed_page_count(struct page *page, long count)
8109 atomic_long_add(count, &page_zone(page)->managed_pages);
8110 totalram_pages_add(count);
8111 #ifdef CONFIG_HIGHMEM
8112 if (PageHighMem(page))
8113 totalhigh_pages_add(count);
8116 EXPORT_SYMBOL(adjust_managed_page_count);
8118 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8121 unsigned long pages = 0;
8123 start = (void *)PAGE_ALIGN((unsigned long)start);
8124 end = (void *)((unsigned long)end & PAGE_MASK);
8125 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8126 struct page *page = virt_to_page(pos);
8127 void *direct_map_addr;
8130 * 'direct_map_addr' might be different from 'pos'
8131 * because some architectures' virt_to_page()
8132 * work with aliases. Getting the direct map
8133 * address ensures that we get a _writeable_
8134 * alias for the memset().
8136 direct_map_addr = page_address(page);
8138 * Perform a kasan-unchecked memset() since this memory
8139 * has not been initialized.
8141 direct_map_addr = kasan_reset_tag(direct_map_addr);
8142 if ((unsigned int)poison <= 0xFF)
8143 memset(direct_map_addr, poison, PAGE_SIZE);
8145 free_reserved_page(page);
8149 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8154 void __init mem_init_print_info(void)
8156 unsigned long physpages, codesize, datasize, rosize, bss_size;
8157 unsigned long init_code_size, init_data_size;
8159 physpages = get_num_physpages();
8160 codesize = _etext - _stext;
8161 datasize = _edata - _sdata;
8162 rosize = __end_rodata - __start_rodata;
8163 bss_size = __bss_stop - __bss_start;
8164 init_data_size = __init_end - __init_begin;
8165 init_code_size = _einittext - _sinittext;
8168 * Detect special cases and adjust section sizes accordingly:
8169 * 1) .init.* may be embedded into .data sections
8170 * 2) .init.text.* may be out of [__init_begin, __init_end],
8171 * please refer to arch/tile/kernel/vmlinux.lds.S.
8172 * 3) .rodata.* may be embedded into .text or .data sections.
8174 #define adj_init_size(start, end, size, pos, adj) \
8176 if (start <= pos && pos < end && size > adj) \
8180 adj_init_size(__init_begin, __init_end, init_data_size,
8181 _sinittext, init_code_size);
8182 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8183 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8184 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8185 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8187 #undef adj_init_size
8189 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8190 #ifdef CONFIG_HIGHMEM
8194 K(nr_free_pages()), K(physpages),
8195 codesize >> 10, datasize >> 10, rosize >> 10,
8196 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8197 K(physpages - totalram_pages() - totalcma_pages),
8199 #ifdef CONFIG_HIGHMEM
8200 , K(totalhigh_pages())
8206 * set_dma_reserve - set the specified number of pages reserved in the first zone
8207 * @new_dma_reserve: The number of pages to mark reserved
8209 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8210 * In the DMA zone, a significant percentage may be consumed by kernel image
8211 * and other unfreeable allocations which can skew the watermarks badly. This
8212 * function may optionally be used to account for unfreeable pages in the
8213 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8214 * smaller per-cpu batchsize.
8216 void __init set_dma_reserve(unsigned long new_dma_reserve)
8218 dma_reserve = new_dma_reserve;
8221 static int page_alloc_cpu_dead(unsigned int cpu)
8225 lru_add_drain_cpu(cpu);
8229 * Spill the event counters of the dead processor
8230 * into the current processors event counters.
8231 * This artificially elevates the count of the current
8234 vm_events_fold_cpu(cpu);
8237 * Zero the differential counters of the dead processor
8238 * so that the vm statistics are consistent.
8240 * This is only okay since the processor is dead and cannot
8241 * race with what we are doing.
8243 cpu_vm_stats_fold(cpu);
8245 for_each_populated_zone(zone)
8246 zone_pcp_update(zone, 0);
8251 static int page_alloc_cpu_online(unsigned int cpu)
8255 for_each_populated_zone(zone)
8256 zone_pcp_update(zone, 1);
8261 int hashdist = HASHDIST_DEFAULT;
8263 static int __init set_hashdist(char *str)
8267 hashdist = simple_strtoul(str, &str, 0);
8270 __setup("hashdist=", set_hashdist);
8273 void __init page_alloc_init(void)
8278 if (num_node_state(N_MEMORY) == 1)
8282 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8283 "mm/page_alloc:pcp",
8284 page_alloc_cpu_online,
8285 page_alloc_cpu_dead);
8290 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8291 * or min_free_kbytes changes.
8293 static void calculate_totalreserve_pages(void)
8295 struct pglist_data *pgdat;
8296 unsigned long reserve_pages = 0;
8297 enum zone_type i, j;
8299 for_each_online_pgdat(pgdat) {
8301 pgdat->totalreserve_pages = 0;
8303 for (i = 0; i < MAX_NR_ZONES; i++) {
8304 struct zone *zone = pgdat->node_zones + i;
8306 unsigned long managed_pages = zone_managed_pages(zone);
8308 /* Find valid and maximum lowmem_reserve in the zone */
8309 for (j = i; j < MAX_NR_ZONES; j++) {
8310 if (zone->lowmem_reserve[j] > max)
8311 max = zone->lowmem_reserve[j];
8314 /* we treat the high watermark as reserved pages. */
8315 max += high_wmark_pages(zone);
8317 if (max > managed_pages)
8318 max = managed_pages;
8320 pgdat->totalreserve_pages += max;
8322 reserve_pages += max;
8325 totalreserve_pages = reserve_pages;
8329 * setup_per_zone_lowmem_reserve - called whenever
8330 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8331 * has a correct pages reserved value, so an adequate number of
8332 * pages are left in the zone after a successful __alloc_pages().
8334 static void setup_per_zone_lowmem_reserve(void)
8336 struct pglist_data *pgdat;
8337 enum zone_type i, j;
8339 for_each_online_pgdat(pgdat) {
8340 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8341 struct zone *zone = &pgdat->node_zones[i];
8342 int ratio = sysctl_lowmem_reserve_ratio[i];
8343 bool clear = !ratio || !zone_managed_pages(zone);
8344 unsigned long managed_pages = 0;
8346 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8347 struct zone *upper_zone = &pgdat->node_zones[j];
8349 managed_pages += zone_managed_pages(upper_zone);
8352 zone->lowmem_reserve[j] = 0;
8354 zone->lowmem_reserve[j] = managed_pages / ratio;
8359 /* update totalreserve_pages */
8360 calculate_totalreserve_pages();
8363 static void __setup_per_zone_wmarks(void)
8365 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8366 unsigned long lowmem_pages = 0;
8368 unsigned long flags;
8370 /* Calculate total number of !ZONE_HIGHMEM pages */
8371 for_each_zone(zone) {
8372 if (!is_highmem(zone))
8373 lowmem_pages += zone_managed_pages(zone);
8376 for_each_zone(zone) {
8379 spin_lock_irqsave(&zone->lock, flags);
8380 tmp = (u64)pages_min * zone_managed_pages(zone);
8381 do_div(tmp, lowmem_pages);
8382 if (is_highmem(zone)) {
8384 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8385 * need highmem pages, so cap pages_min to a small
8388 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8389 * deltas control async page reclaim, and so should
8390 * not be capped for highmem.
8392 unsigned long min_pages;
8394 min_pages = zone_managed_pages(zone) / 1024;
8395 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8396 zone->_watermark[WMARK_MIN] = min_pages;
8399 * If it's a lowmem zone, reserve a number of pages
8400 * proportionate to the zone's size.
8402 zone->_watermark[WMARK_MIN] = tmp;
8406 * Set the kswapd watermarks distance according to the
8407 * scale factor in proportion to available memory, but
8408 * ensure a minimum size on small systems.
8410 tmp = max_t(u64, tmp >> 2,
8411 mult_frac(zone_managed_pages(zone),
8412 watermark_scale_factor, 10000));
8414 zone->watermark_boost = 0;
8415 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8416 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8418 spin_unlock_irqrestore(&zone->lock, flags);
8421 /* update totalreserve_pages */
8422 calculate_totalreserve_pages();
8426 * setup_per_zone_wmarks - called when min_free_kbytes changes
8427 * or when memory is hot-{added|removed}
8429 * Ensures that the watermark[min,low,high] values for each zone are set
8430 * correctly with respect to min_free_kbytes.
8432 void setup_per_zone_wmarks(void)
8435 static DEFINE_SPINLOCK(lock);
8438 __setup_per_zone_wmarks();
8442 * The watermark size have changed so update the pcpu batch
8443 * and high limits or the limits may be inappropriate.
8446 zone_pcp_update(zone, 0);
8450 * Initialise min_free_kbytes.
8452 * For small machines we want it small (128k min). For large machines
8453 * we want it large (256MB max). But it is not linear, because network
8454 * bandwidth does not increase linearly with machine size. We use
8456 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8457 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8473 void calculate_min_free_kbytes(void)
8475 unsigned long lowmem_kbytes;
8476 int new_min_free_kbytes;
8478 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8479 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8481 if (new_min_free_kbytes > user_min_free_kbytes)
8482 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8484 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8485 new_min_free_kbytes, user_min_free_kbytes);
8489 int __meminit init_per_zone_wmark_min(void)
8491 calculate_min_free_kbytes();
8492 setup_per_zone_wmarks();
8493 refresh_zone_stat_thresholds();
8494 setup_per_zone_lowmem_reserve();
8497 setup_min_unmapped_ratio();
8498 setup_min_slab_ratio();
8501 khugepaged_min_free_kbytes_update();
8505 postcore_initcall(init_per_zone_wmark_min)
8508 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8509 * that we can call two helper functions whenever min_free_kbytes
8512 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8513 void *buffer, size_t *length, loff_t *ppos)
8517 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8522 user_min_free_kbytes = min_free_kbytes;
8523 setup_per_zone_wmarks();
8528 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8529 void *buffer, size_t *length, loff_t *ppos)
8533 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8538 setup_per_zone_wmarks();
8544 static void setup_min_unmapped_ratio(void)
8549 for_each_online_pgdat(pgdat)
8550 pgdat->min_unmapped_pages = 0;
8553 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8554 sysctl_min_unmapped_ratio) / 100;
8558 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8559 void *buffer, size_t *length, loff_t *ppos)
8563 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8567 setup_min_unmapped_ratio();
8572 static void setup_min_slab_ratio(void)
8577 for_each_online_pgdat(pgdat)
8578 pgdat->min_slab_pages = 0;
8581 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8582 sysctl_min_slab_ratio) / 100;
8585 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8586 void *buffer, size_t *length, loff_t *ppos)
8590 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8594 setup_min_slab_ratio();
8601 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8602 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8603 * whenever sysctl_lowmem_reserve_ratio changes.
8605 * The reserve ratio obviously has absolutely no relation with the
8606 * minimum watermarks. The lowmem reserve ratio can only make sense
8607 * if in function of the boot time zone sizes.
8609 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8610 void *buffer, size_t *length, loff_t *ppos)
8614 proc_dointvec_minmax(table, write, buffer, length, ppos);
8616 for (i = 0; i < MAX_NR_ZONES; i++) {
8617 if (sysctl_lowmem_reserve_ratio[i] < 1)
8618 sysctl_lowmem_reserve_ratio[i] = 0;
8621 setup_per_zone_lowmem_reserve();
8626 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8627 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8628 * pagelist can have before it gets flushed back to buddy allocator.
8630 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8631 int write, void *buffer, size_t *length, loff_t *ppos)
8634 int old_percpu_pagelist_high_fraction;
8637 mutex_lock(&pcp_batch_high_lock);
8638 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8640 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8641 if (!write || ret < 0)
8644 /* Sanity checking to avoid pcp imbalance */
8645 if (percpu_pagelist_high_fraction &&
8646 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8647 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8653 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8656 for_each_populated_zone(zone)
8657 zone_set_pageset_high_and_batch(zone, 0);
8659 mutex_unlock(&pcp_batch_high_lock);
8663 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8665 * Returns the number of pages that arch has reserved but
8666 * is not known to alloc_large_system_hash().
8668 static unsigned long __init arch_reserved_kernel_pages(void)
8675 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8676 * machines. As memory size is increased the scale is also increased but at
8677 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8678 * quadruples the scale is increased by one, which means the size of hash table
8679 * only doubles, instead of quadrupling as well.
8680 * Because 32-bit systems cannot have large physical memory, where this scaling
8681 * makes sense, it is disabled on such platforms.
8683 #if __BITS_PER_LONG > 32
8684 #define ADAPT_SCALE_BASE (64ul << 30)
8685 #define ADAPT_SCALE_SHIFT 2
8686 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8690 * allocate a large system hash table from bootmem
8691 * - it is assumed that the hash table must contain an exact power-of-2
8692 * quantity of entries
8693 * - limit is the number of hash buckets, not the total allocation size
8695 void *__init alloc_large_system_hash(const char *tablename,
8696 unsigned long bucketsize,
8697 unsigned long numentries,
8700 unsigned int *_hash_shift,
8701 unsigned int *_hash_mask,
8702 unsigned long low_limit,
8703 unsigned long high_limit)
8705 unsigned long long max = high_limit;
8706 unsigned long log2qty, size;
8712 /* allow the kernel cmdline to have a say */
8714 /* round applicable memory size up to nearest megabyte */
8715 numentries = nr_kernel_pages;
8716 numentries -= arch_reserved_kernel_pages();
8718 /* It isn't necessary when PAGE_SIZE >= 1MB */
8719 if (PAGE_SHIFT < 20)
8720 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8722 #if __BITS_PER_LONG > 32
8724 unsigned long adapt;
8726 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8727 adapt <<= ADAPT_SCALE_SHIFT)
8732 /* limit to 1 bucket per 2^scale bytes of low memory */
8733 if (scale > PAGE_SHIFT)
8734 numentries >>= (scale - PAGE_SHIFT);
8736 numentries <<= (PAGE_SHIFT - scale);
8738 /* Make sure we've got at least a 0-order allocation.. */
8739 if (unlikely(flags & HASH_SMALL)) {
8740 /* Makes no sense without HASH_EARLY */
8741 WARN_ON(!(flags & HASH_EARLY));
8742 if (!(numentries >> *_hash_shift)) {
8743 numentries = 1UL << *_hash_shift;
8744 BUG_ON(!numentries);
8746 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8747 numentries = PAGE_SIZE / bucketsize;
8749 numentries = roundup_pow_of_two(numentries);
8751 /* limit allocation size to 1/16 total memory by default */
8753 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8754 do_div(max, bucketsize);
8756 max = min(max, 0x80000000ULL);
8758 if (numentries < low_limit)
8759 numentries = low_limit;
8760 if (numentries > max)
8763 log2qty = ilog2(numentries);
8765 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8768 size = bucketsize << log2qty;
8769 if (flags & HASH_EARLY) {
8770 if (flags & HASH_ZERO)
8771 table = memblock_alloc(size, SMP_CACHE_BYTES);
8773 table = memblock_alloc_raw(size,
8775 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8776 table = __vmalloc(size, gfp_flags);
8779 huge = is_vm_area_hugepages(table);
8782 * If bucketsize is not a power-of-two, we may free
8783 * some pages at the end of hash table which
8784 * alloc_pages_exact() automatically does
8786 table = alloc_pages_exact(size, gfp_flags);
8787 kmemleak_alloc(table, size, 1, gfp_flags);
8789 } while (!table && size > PAGE_SIZE && --log2qty);
8792 panic("Failed to allocate %s hash table\n", tablename);
8794 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8795 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8796 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8799 *_hash_shift = log2qty;
8801 *_hash_mask = (1 << log2qty) - 1;
8807 * This function checks whether pageblock includes unmovable pages or not.
8809 * PageLRU check without isolation or lru_lock could race so that
8810 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8811 * check without lock_page also may miss some movable non-lru pages at
8812 * race condition. So you can't expect this function should be exact.
8814 * Returns a page without holding a reference. If the caller wants to
8815 * dereference that page (e.g., dumping), it has to make sure that it
8816 * cannot get removed (e.g., via memory unplug) concurrently.
8819 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8820 int migratetype, int flags)
8822 unsigned long iter = 0;
8823 unsigned long pfn = page_to_pfn(page);
8824 unsigned long offset = pfn % pageblock_nr_pages;
8826 if (is_migrate_cma_page(page)) {
8828 * CMA allocations (alloc_contig_range) really need to mark
8829 * isolate CMA pageblocks even when they are not movable in fact
8830 * so consider them movable here.
8832 if (is_migrate_cma(migratetype))
8838 for (; iter < pageblock_nr_pages - offset; iter++) {
8839 page = pfn_to_page(pfn + iter);
8842 * Both, bootmem allocations and memory holes are marked
8843 * PG_reserved and are unmovable. We can even have unmovable
8844 * allocations inside ZONE_MOVABLE, for example when
8845 * specifying "movablecore".
8847 if (PageReserved(page))
8851 * If the zone is movable and we have ruled out all reserved
8852 * pages then it should be reasonably safe to assume the rest
8855 if (zone_idx(zone) == ZONE_MOVABLE)
8859 * Hugepages are not in LRU lists, but they're movable.
8860 * THPs are on the LRU, but need to be counted as #small pages.
8861 * We need not scan over tail pages because we don't
8862 * handle each tail page individually in migration.
8864 if (PageHuge(page) || PageTransCompound(page)) {
8865 struct page *head = compound_head(page);
8866 unsigned int skip_pages;
8868 if (PageHuge(page)) {
8869 if (!hugepage_migration_supported(page_hstate(head)))
8871 } else if (!PageLRU(head) && !__PageMovable(head)) {
8875 skip_pages = compound_nr(head) - (page - head);
8876 iter += skip_pages - 1;
8881 * We can't use page_count without pin a page
8882 * because another CPU can free compound page.
8883 * This check already skips compound tails of THP
8884 * because their page->_refcount is zero at all time.
8886 if (!page_ref_count(page)) {
8887 if (PageBuddy(page))
8888 iter += (1 << buddy_order(page)) - 1;
8893 * The HWPoisoned page may be not in buddy system, and
8894 * page_count() is not 0.
8896 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8900 * We treat all PageOffline() pages as movable when offlining
8901 * to give drivers a chance to decrement their reference count
8902 * in MEM_GOING_OFFLINE in order to indicate that these pages
8903 * can be offlined as there are no direct references anymore.
8904 * For actually unmovable PageOffline() where the driver does
8905 * not support this, we will fail later when trying to actually
8906 * move these pages that still have a reference count > 0.
8907 * (false negatives in this function only)
8909 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8912 if (__PageMovable(page) || PageLRU(page))
8916 * If there are RECLAIMABLE pages, we need to check
8917 * it. But now, memory offline itself doesn't call
8918 * shrink_node_slabs() and it still to be fixed.
8925 #ifdef CONFIG_CONTIG_ALLOC
8926 static unsigned long pfn_max_align_down(unsigned long pfn)
8928 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8929 pageblock_nr_pages) - 1);
8932 static unsigned long pfn_max_align_up(unsigned long pfn)
8934 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8935 pageblock_nr_pages));
8938 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8939 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8940 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8941 static void alloc_contig_dump_pages(struct list_head *page_list)
8943 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8945 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8949 list_for_each_entry(page, page_list, lru)
8950 dump_page(page, "migration failure");
8954 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8959 /* [start, end) must belong to a single zone. */
8960 static int __alloc_contig_migrate_range(struct compact_control *cc,
8961 unsigned long start, unsigned long end)
8963 /* This function is based on compact_zone() from compaction.c. */
8964 unsigned int nr_reclaimed;
8965 unsigned long pfn = start;
8966 unsigned int tries = 0;
8968 struct migration_target_control mtc = {
8969 .nid = zone_to_nid(cc->zone),
8970 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8973 lru_cache_disable();
8975 while (pfn < end || !list_empty(&cc->migratepages)) {
8976 if (fatal_signal_pending(current)) {
8981 if (list_empty(&cc->migratepages)) {
8982 cc->nr_migratepages = 0;
8983 ret = isolate_migratepages_range(cc, pfn, end);
8984 if (ret && ret != -EAGAIN)
8986 pfn = cc->migrate_pfn;
8988 } else if (++tries == 5) {
8993 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8995 cc->nr_migratepages -= nr_reclaimed;
8997 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8998 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9001 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9002 * to retry again over this error, so do the same here.
9011 alloc_contig_dump_pages(&cc->migratepages);
9012 putback_movable_pages(&cc->migratepages);
9019 * alloc_contig_range() -- tries to allocate given range of pages
9020 * @start: start PFN to allocate
9021 * @end: one-past-the-last PFN to allocate
9022 * @migratetype: migratetype of the underlying pageblocks (either
9023 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9024 * in range must have the same migratetype and it must
9025 * be either of the two.
9026 * @gfp_mask: GFP mask to use during compaction
9028 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9029 * aligned. The PFN range must belong to a single zone.
9031 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9032 * pageblocks in the range. Once isolated, the pageblocks should not
9033 * be modified by others.
9035 * Return: zero on success or negative error code. On success all
9036 * pages which PFN is in [start, end) are allocated for the caller and
9037 * need to be freed with free_contig_range().
9039 int alloc_contig_range(unsigned long start, unsigned long end,
9040 unsigned migratetype, gfp_t gfp_mask)
9042 unsigned long outer_start, outer_end;
9046 struct compact_control cc = {
9047 .nr_migratepages = 0,
9049 .zone = page_zone(pfn_to_page(start)),
9050 .mode = MIGRATE_SYNC,
9051 .ignore_skip_hint = true,
9052 .no_set_skip_hint = true,
9053 .gfp_mask = current_gfp_context(gfp_mask),
9054 .alloc_contig = true,
9056 INIT_LIST_HEAD(&cc.migratepages);
9059 * What we do here is we mark all pageblocks in range as
9060 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9061 * have different sizes, and due to the way page allocator
9062 * work, we align the range to biggest of the two pages so
9063 * that page allocator won't try to merge buddies from
9064 * different pageblocks and change MIGRATE_ISOLATE to some
9065 * other migration type.
9067 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9068 * migrate the pages from an unaligned range (ie. pages that
9069 * we are interested in). This will put all the pages in
9070 * range back to page allocator as MIGRATE_ISOLATE.
9072 * When this is done, we take the pages in range from page
9073 * allocator removing them from the buddy system. This way
9074 * page allocator will never consider using them.
9076 * This lets us mark the pageblocks back as
9077 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9078 * aligned range but not in the unaligned, original range are
9079 * put back to page allocator so that buddy can use them.
9082 ret = start_isolate_page_range(pfn_max_align_down(start),
9083 pfn_max_align_up(end), migratetype, 0);
9087 drain_all_pages(cc.zone);
9090 * In case of -EBUSY, we'd like to know which page causes problem.
9091 * So, just fall through. test_pages_isolated() has a tracepoint
9092 * which will report the busy page.
9094 * It is possible that busy pages could become available before
9095 * the call to test_pages_isolated, and the range will actually be
9096 * allocated. So, if we fall through be sure to clear ret so that
9097 * -EBUSY is not accidentally used or returned to caller.
9099 ret = __alloc_contig_migrate_range(&cc, start, end);
9100 if (ret && ret != -EBUSY)
9105 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9106 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9107 * more, all pages in [start, end) are free in page allocator.
9108 * What we are going to do is to allocate all pages from
9109 * [start, end) (that is remove them from page allocator).
9111 * The only problem is that pages at the beginning and at the
9112 * end of interesting range may be not aligned with pages that
9113 * page allocator holds, ie. they can be part of higher order
9114 * pages. Because of this, we reserve the bigger range and
9115 * once this is done free the pages we are not interested in.
9117 * We don't have to hold zone->lock here because the pages are
9118 * isolated thus they won't get removed from buddy.
9122 outer_start = start;
9123 while (!PageBuddy(pfn_to_page(outer_start))) {
9124 if (++order >= MAX_ORDER) {
9125 outer_start = start;
9128 outer_start &= ~0UL << order;
9131 if (outer_start != start) {
9132 order = buddy_order(pfn_to_page(outer_start));
9135 * outer_start page could be small order buddy page and
9136 * it doesn't include start page. Adjust outer_start
9137 * in this case to report failed page properly
9138 * on tracepoint in test_pages_isolated()
9140 if (outer_start + (1UL << order) <= start)
9141 outer_start = start;
9144 /* Make sure the range is really isolated. */
9145 if (test_pages_isolated(outer_start, end, 0)) {
9150 /* Grab isolated pages from freelists. */
9151 outer_end = isolate_freepages_range(&cc, outer_start, end);
9157 /* Free head and tail (if any) */
9158 if (start != outer_start)
9159 free_contig_range(outer_start, start - outer_start);
9160 if (end != outer_end)
9161 free_contig_range(end, outer_end - end);
9164 undo_isolate_page_range(pfn_max_align_down(start),
9165 pfn_max_align_up(end), migratetype);
9168 EXPORT_SYMBOL(alloc_contig_range);
9170 static int __alloc_contig_pages(unsigned long start_pfn,
9171 unsigned long nr_pages, gfp_t gfp_mask)
9173 unsigned long end_pfn = start_pfn + nr_pages;
9175 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9179 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9180 unsigned long nr_pages)
9182 unsigned long i, end_pfn = start_pfn + nr_pages;
9185 for (i = start_pfn; i < end_pfn; i++) {
9186 page = pfn_to_online_page(i);
9190 if (page_zone(page) != z)
9193 if (PageReserved(page))
9199 static bool zone_spans_last_pfn(const struct zone *zone,
9200 unsigned long start_pfn, unsigned long nr_pages)
9202 unsigned long last_pfn = start_pfn + nr_pages - 1;
9204 return zone_spans_pfn(zone, last_pfn);
9208 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9209 * @nr_pages: Number of contiguous pages to allocate
9210 * @gfp_mask: GFP mask to limit search and used during compaction
9212 * @nodemask: Mask for other possible nodes
9214 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9215 * on an applicable zonelist to find a contiguous pfn range which can then be
9216 * tried for allocation with alloc_contig_range(). This routine is intended
9217 * for allocation requests which can not be fulfilled with the buddy allocator.
9219 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9220 * power of two then the alignment is guaranteed to be to the given nr_pages
9221 * (e.g. 1GB request would be aligned to 1GB).
9223 * Allocated pages can be freed with free_contig_range() or by manually calling
9224 * __free_page() on each allocated page.
9226 * Return: pointer to contiguous pages on success, or NULL if not successful.
9228 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9229 int nid, nodemask_t *nodemask)
9231 unsigned long ret, pfn, flags;
9232 struct zonelist *zonelist;
9236 zonelist = node_zonelist(nid, gfp_mask);
9237 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9238 gfp_zone(gfp_mask), nodemask) {
9239 spin_lock_irqsave(&zone->lock, flags);
9241 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9242 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9243 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9245 * We release the zone lock here because
9246 * alloc_contig_range() will also lock the zone
9247 * at some point. If there's an allocation
9248 * spinning on this lock, it may win the race
9249 * and cause alloc_contig_range() to fail...
9251 spin_unlock_irqrestore(&zone->lock, flags);
9252 ret = __alloc_contig_pages(pfn, nr_pages,
9255 return pfn_to_page(pfn);
9256 spin_lock_irqsave(&zone->lock, flags);
9260 spin_unlock_irqrestore(&zone->lock, flags);
9264 #endif /* CONFIG_CONTIG_ALLOC */
9266 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9268 unsigned long count = 0;
9270 for (; nr_pages--; pfn++) {
9271 struct page *page = pfn_to_page(pfn);
9273 count += page_count(page) != 1;
9276 WARN(count != 0, "%lu pages are still in use!\n", count);
9278 EXPORT_SYMBOL(free_contig_range);
9281 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9282 * page high values need to be recalculated.
9284 void zone_pcp_update(struct zone *zone, int cpu_online)
9286 mutex_lock(&pcp_batch_high_lock);
9287 zone_set_pageset_high_and_batch(zone, cpu_online);
9288 mutex_unlock(&pcp_batch_high_lock);
9292 * Effectively disable pcplists for the zone by setting the high limit to 0
9293 * and draining all cpus. A concurrent page freeing on another CPU that's about
9294 * to put the page on pcplist will either finish before the drain and the page
9295 * will be drained, or observe the new high limit and skip the pcplist.
9297 * Must be paired with a call to zone_pcp_enable().
9299 void zone_pcp_disable(struct zone *zone)
9301 mutex_lock(&pcp_batch_high_lock);
9302 __zone_set_pageset_high_and_batch(zone, 0, 1);
9303 __drain_all_pages(zone, true);
9306 void zone_pcp_enable(struct zone *zone)
9308 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9309 mutex_unlock(&pcp_batch_high_lock);
9312 void zone_pcp_reset(struct zone *zone)
9315 struct per_cpu_zonestat *pzstats;
9317 if (zone->per_cpu_pageset != &boot_pageset) {
9318 for_each_online_cpu(cpu) {
9319 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9320 drain_zonestat(zone, pzstats);
9322 free_percpu(zone->per_cpu_pageset);
9323 free_percpu(zone->per_cpu_zonestats);
9324 zone->per_cpu_pageset = &boot_pageset;
9325 zone->per_cpu_zonestats = &boot_zonestats;
9329 #ifdef CONFIG_MEMORY_HOTREMOVE
9331 * All pages in the range must be in a single zone, must not contain holes,
9332 * must span full sections, and must be isolated before calling this function.
9334 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9336 unsigned long pfn = start_pfn;
9340 unsigned long flags;
9342 offline_mem_sections(pfn, end_pfn);
9343 zone = page_zone(pfn_to_page(pfn));
9344 spin_lock_irqsave(&zone->lock, flags);
9345 while (pfn < end_pfn) {
9346 page = pfn_to_page(pfn);
9348 * The HWPoisoned page may be not in buddy system, and
9349 * page_count() is not 0.
9351 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9356 * At this point all remaining PageOffline() pages have a
9357 * reference count of 0 and can simply be skipped.
9359 if (PageOffline(page)) {
9360 BUG_ON(page_count(page));
9361 BUG_ON(PageBuddy(page));
9366 BUG_ON(page_count(page));
9367 BUG_ON(!PageBuddy(page));
9368 order = buddy_order(page);
9369 del_page_from_free_list(page, zone, order);
9370 pfn += (1 << order);
9372 spin_unlock_irqrestore(&zone->lock, flags);
9377 * This function returns a stable result only if called under zone lock.
9379 bool is_free_buddy_page(struct page *page)
9381 unsigned long pfn = page_to_pfn(page);
9384 for (order = 0; order < MAX_ORDER; order++) {
9385 struct page *page_head = page - (pfn & ((1 << order) - 1));
9387 if (PageBuddy(page_head) &&
9388 buddy_order_unsafe(page_head) >= order)
9392 return order < MAX_ORDER;
9395 #ifdef CONFIG_MEMORY_FAILURE
9397 * Break down a higher-order page in sub-pages, and keep our target out of
9400 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9401 struct page *target, int low, int high,
9404 unsigned long size = 1 << high;
9405 struct page *current_buddy, *next_page;
9407 while (high > low) {
9411 if (target >= &page[size]) {
9412 next_page = page + size;
9413 current_buddy = page;
9416 current_buddy = page + size;
9419 if (set_page_guard(zone, current_buddy, high, migratetype))
9422 if (current_buddy != target) {
9423 add_to_free_list(current_buddy, zone, high, migratetype);
9424 set_buddy_order(current_buddy, high);
9431 * Take a page that will be marked as poisoned off the buddy allocator.
9433 bool take_page_off_buddy(struct page *page)
9435 struct zone *zone = page_zone(page);
9436 unsigned long pfn = page_to_pfn(page);
9437 unsigned long flags;
9441 spin_lock_irqsave(&zone->lock, flags);
9442 for (order = 0; order < MAX_ORDER; order++) {
9443 struct page *page_head = page - (pfn & ((1 << order) - 1));
9444 int page_order = buddy_order(page_head);
9446 if (PageBuddy(page_head) && page_order >= order) {
9447 unsigned long pfn_head = page_to_pfn(page_head);
9448 int migratetype = get_pfnblock_migratetype(page_head,
9451 del_page_from_free_list(page_head, zone, page_order);
9452 break_down_buddy_pages(zone, page_head, page, 0,
9453 page_order, migratetype);
9454 if (!is_migrate_isolate(migratetype))
9455 __mod_zone_freepage_state(zone, -1, migratetype);
9459 if (page_count(page_head) > 0)
9462 spin_unlock_irqrestore(&zone->lock, flags);