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 <asm/sections.h>
76 #include <asm/tlbflush.h>
77 #include <asm/div64.h>
80 #include "page_reporting.h"
82 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
83 typedef int __bitwise fpi_t;
85 /* No special request */
86 #define FPI_NONE ((__force fpi_t)0)
89 * Skip free page reporting notification for the (possibly merged) page.
90 * This does not hinder free page reporting from grabbing the page,
91 * reporting it and marking it "reported" - it only skips notifying
92 * the free page reporting infrastructure about a newly freed page. For
93 * example, used when temporarily pulling a page from a freelist and
94 * putting it back unmodified.
96 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
99 * Place the (possibly merged) page to the tail of the freelist. Will ignore
100 * page shuffling (relevant code - e.g., memory onlining - is expected to
101 * shuffle the whole zone).
103 * Note: No code should rely on this flag for correctness - it's purely
104 * to allow for optimizations when handing back either fresh pages
105 * (memory onlining) or untouched pages (page isolation, free page
108 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
111 * Don't poison memory with KASAN (only for the tag-based modes).
112 * During boot, all non-reserved memblock memory is exposed to page_alloc.
113 * Poisoning all that memory lengthens boot time, especially on systems with
114 * large amount of RAM. This flag is used to skip that poisoning.
115 * This is only done for the tag-based KASAN modes, as those are able to
116 * detect memory corruptions with the memory tags assigned by default.
117 * All memory allocated normally after boot gets poisoned as usual.
119 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
121 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
122 static DEFINE_MUTEX(pcp_batch_high_lock);
123 #define MIN_PERCPU_PAGELIST_FRACTION (8)
125 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
126 DEFINE_PER_CPU(int, numa_node);
127 EXPORT_PER_CPU_SYMBOL(numa_node);
130 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
132 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
134 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
135 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
136 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
137 * defined in <linux/topology.h>.
139 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
140 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
143 /* work_structs for global per-cpu drains */
146 struct work_struct work;
148 static DEFINE_MUTEX(pcpu_drain_mutex);
149 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
151 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
152 volatile unsigned long latent_entropy __latent_entropy;
153 EXPORT_SYMBOL(latent_entropy);
157 * Array of node states.
159 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
160 [N_POSSIBLE] = NODE_MASK_ALL,
161 [N_ONLINE] = { { [0] = 1UL } },
163 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
164 #ifdef CONFIG_HIGHMEM
165 [N_HIGH_MEMORY] = { { [0] = 1UL } },
167 [N_MEMORY] = { { [0] = 1UL } },
168 [N_CPU] = { { [0] = 1UL } },
171 EXPORT_SYMBOL(node_states);
173 atomic_long_t _totalram_pages __read_mostly;
174 EXPORT_SYMBOL(_totalram_pages);
175 unsigned long totalreserve_pages __read_mostly;
176 unsigned long totalcma_pages __read_mostly;
178 int percpu_pagelist_fraction;
179 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
180 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
181 EXPORT_SYMBOL(init_on_alloc);
183 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
184 EXPORT_SYMBOL(init_on_free);
186 static bool _init_on_alloc_enabled_early __read_mostly
187 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
188 static int __init early_init_on_alloc(char *buf)
191 return kstrtobool(buf, &_init_on_alloc_enabled_early);
193 early_param("init_on_alloc", early_init_on_alloc);
195 static bool _init_on_free_enabled_early __read_mostly
196 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
197 static int __init early_init_on_free(char *buf)
199 return kstrtobool(buf, &_init_on_free_enabled_early);
201 early_param("init_on_free", early_init_on_free);
204 * A cached value of the page's pageblock's migratetype, used when the page is
205 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
206 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
207 * Also the migratetype set in the page does not necessarily match the pcplist
208 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
209 * other index - this ensures that it will be put on the correct CMA freelist.
211 static inline int get_pcppage_migratetype(struct page *page)
216 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
218 page->index = migratetype;
221 #ifdef CONFIG_PM_SLEEP
223 * The following functions are used by the suspend/hibernate code to temporarily
224 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
225 * while devices are suspended. To avoid races with the suspend/hibernate code,
226 * they should always be called with system_transition_mutex held
227 * (gfp_allowed_mask also should only be modified with system_transition_mutex
228 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
229 * with that modification).
232 static gfp_t saved_gfp_mask;
234 void pm_restore_gfp_mask(void)
236 WARN_ON(!mutex_is_locked(&system_transition_mutex));
237 if (saved_gfp_mask) {
238 gfp_allowed_mask = saved_gfp_mask;
243 void pm_restrict_gfp_mask(void)
245 WARN_ON(!mutex_is_locked(&system_transition_mutex));
246 WARN_ON(saved_gfp_mask);
247 saved_gfp_mask = gfp_allowed_mask;
248 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
251 bool pm_suspended_storage(void)
253 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
257 #endif /* CONFIG_PM_SLEEP */
259 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
260 unsigned int pageblock_order __read_mostly;
263 static void __free_pages_ok(struct page *page, unsigned int order,
267 * results with 256, 32 in the lowmem_reserve sysctl:
268 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
269 * 1G machine -> (16M dma, 784M normal, 224M high)
270 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
271 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
272 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
274 * TBD: should special case ZONE_DMA32 machines here - in those we normally
275 * don't need any ZONE_NORMAL reservation
277 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
278 #ifdef CONFIG_ZONE_DMA
281 #ifdef CONFIG_ZONE_DMA32
285 #ifdef CONFIG_HIGHMEM
291 static char * const zone_names[MAX_NR_ZONES] = {
292 #ifdef CONFIG_ZONE_DMA
295 #ifdef CONFIG_ZONE_DMA32
299 #ifdef CONFIG_HIGHMEM
303 #ifdef CONFIG_ZONE_DEVICE
308 const char * const migratetype_names[MIGRATE_TYPES] = {
316 #ifdef CONFIG_MEMORY_ISOLATION
321 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
322 [NULL_COMPOUND_DTOR] = NULL,
323 [COMPOUND_PAGE_DTOR] = free_compound_page,
324 #ifdef CONFIG_HUGETLB_PAGE
325 [HUGETLB_PAGE_DTOR] = free_huge_page,
327 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
328 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
332 int min_free_kbytes = 1024;
333 int user_min_free_kbytes = -1;
334 #ifdef CONFIG_DISCONTIGMEM
336 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
337 * are not on separate NUMA nodes. Functionally this works but with
338 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
339 * quite small. By default, do not boost watermarks on discontigmem as in
340 * many cases very high-order allocations like THP are likely to be
341 * unsupported and the premature reclaim offsets the advantage of long-term
342 * fragmentation avoidance.
344 int watermark_boost_factor __read_mostly;
346 int watermark_boost_factor __read_mostly = 15000;
348 int watermark_scale_factor = 10;
350 static unsigned long nr_kernel_pages __initdata;
351 static unsigned long nr_all_pages __initdata;
352 static unsigned long dma_reserve __initdata;
354 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
355 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
356 static unsigned long required_kernelcore __initdata;
357 static unsigned long required_kernelcore_percent __initdata;
358 static unsigned long required_movablecore __initdata;
359 static unsigned long required_movablecore_percent __initdata;
360 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
361 static bool mirrored_kernelcore __meminitdata;
363 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
365 EXPORT_SYMBOL(movable_zone);
368 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
369 unsigned int nr_online_nodes __read_mostly = 1;
370 EXPORT_SYMBOL(nr_node_ids);
371 EXPORT_SYMBOL(nr_online_nodes);
374 int page_group_by_mobility_disabled __read_mostly;
376 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
378 * During boot we initialize deferred pages on-demand, as needed, but once
379 * page_alloc_init_late() has finished, the deferred pages are all initialized,
380 * and we can permanently disable that path.
382 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
385 * Calling kasan_free_pages() only after deferred memory initialization
386 * has completed. Poisoning pages during deferred memory init will greatly
387 * lengthen the process and cause problem in large memory systems as the
388 * deferred pages initialization is done with interrupt disabled.
390 * Assuming that there will be no reference to those newly initialized
391 * pages before they are ever allocated, this should have no effect on
392 * KASAN memory tracking as the poison will be properly inserted at page
393 * allocation time. The only corner case is when pages are allocated by
394 * on-demand allocation and then freed again before the deferred pages
395 * initialization is done, but this is not likely to happen.
397 static inline void kasan_free_nondeferred_pages(struct page *page, int order,
398 bool init, fpi_t fpi_flags)
400 if (static_branch_unlikely(&deferred_pages))
402 if (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
403 (fpi_flags & FPI_SKIP_KASAN_POISON))
405 kasan_free_pages(page, order, init);
408 /* Returns true if the struct page for the pfn is uninitialised */
409 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
411 int nid = early_pfn_to_nid(pfn);
413 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
420 * Returns true when the remaining initialisation should be deferred until
421 * later in the boot cycle when it can be parallelised.
423 static bool __meminit
424 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
426 static unsigned long prev_end_pfn, nr_initialised;
429 * prev_end_pfn static that contains the end of previous zone
430 * No need to protect because called very early in boot before smp_init.
432 if (prev_end_pfn != end_pfn) {
433 prev_end_pfn = end_pfn;
437 /* Always populate low zones for address-constrained allocations */
438 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
441 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
444 * We start only with one section of pages, more pages are added as
445 * needed until the rest of deferred pages are initialized.
448 if ((nr_initialised > PAGES_PER_SECTION) &&
449 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
450 NODE_DATA(nid)->first_deferred_pfn = pfn;
456 static inline void kasan_free_nondeferred_pages(struct page *page, int order,
457 bool init, fpi_t fpi_flags)
459 if (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
460 (fpi_flags & FPI_SKIP_KASAN_POISON))
462 kasan_free_pages(page, order, init);
465 static inline bool early_page_uninitialised(unsigned long pfn)
470 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
476 /* Return a pointer to the bitmap storing bits affecting a block of pages */
477 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
480 #ifdef CONFIG_SPARSEMEM
481 return section_to_usemap(__pfn_to_section(pfn));
483 return page_zone(page)->pageblock_flags;
484 #endif /* CONFIG_SPARSEMEM */
487 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
489 #ifdef CONFIG_SPARSEMEM
490 pfn &= (PAGES_PER_SECTION-1);
492 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
493 #endif /* CONFIG_SPARSEMEM */
494 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
497 static __always_inline
498 unsigned long __get_pfnblock_flags_mask(const struct page *page,
502 unsigned long *bitmap;
503 unsigned long bitidx, word_bitidx;
506 bitmap = get_pageblock_bitmap(page, pfn);
507 bitidx = pfn_to_bitidx(page, pfn);
508 word_bitidx = bitidx / BITS_PER_LONG;
509 bitidx &= (BITS_PER_LONG-1);
511 word = bitmap[word_bitidx];
512 return (word >> bitidx) & mask;
516 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
517 * @page: The page within the block of interest
518 * @pfn: The target page frame number
519 * @mask: mask of bits that the caller is interested in
521 * Return: pageblock_bits flags
523 unsigned long get_pfnblock_flags_mask(const struct page *page,
524 unsigned long pfn, unsigned long mask)
526 return __get_pfnblock_flags_mask(page, pfn, mask);
529 static __always_inline int get_pfnblock_migratetype(const struct page *page,
532 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
536 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
537 * @page: The page within the block of interest
538 * @flags: The flags to set
539 * @pfn: The target page frame number
540 * @mask: mask of bits that the caller is interested in
542 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
546 unsigned long *bitmap;
547 unsigned long bitidx, word_bitidx;
548 unsigned long old_word, word;
550 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
551 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
553 bitmap = get_pageblock_bitmap(page, pfn);
554 bitidx = pfn_to_bitidx(page, pfn);
555 word_bitidx = bitidx / BITS_PER_LONG;
556 bitidx &= (BITS_PER_LONG-1);
558 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
563 word = READ_ONCE(bitmap[word_bitidx]);
565 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
566 if (word == old_word)
572 void set_pageblock_migratetype(struct page *page, int migratetype)
574 if (unlikely(page_group_by_mobility_disabled &&
575 migratetype < MIGRATE_PCPTYPES))
576 migratetype = MIGRATE_UNMOVABLE;
578 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
579 page_to_pfn(page), MIGRATETYPE_MASK);
582 #ifdef CONFIG_DEBUG_VM
583 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
587 unsigned long pfn = page_to_pfn(page);
588 unsigned long sp, start_pfn;
591 seq = zone_span_seqbegin(zone);
592 start_pfn = zone->zone_start_pfn;
593 sp = zone->spanned_pages;
594 if (!zone_spans_pfn(zone, pfn))
596 } while (zone_span_seqretry(zone, seq));
599 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
600 pfn, zone_to_nid(zone), zone->name,
601 start_pfn, start_pfn + sp);
606 static int page_is_consistent(struct zone *zone, struct page *page)
608 if (!pfn_valid_within(page_to_pfn(page)))
610 if (zone != page_zone(page))
616 * Temporary debugging check for pages not lying within a given zone.
618 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
620 if (page_outside_zone_boundaries(zone, page))
622 if (!page_is_consistent(zone, page))
628 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
634 static void bad_page(struct page *page, const char *reason)
636 static unsigned long resume;
637 static unsigned long nr_shown;
638 static unsigned long nr_unshown;
641 * Allow a burst of 60 reports, then keep quiet for that minute;
642 * or allow a steady drip of one report per second.
644 if (nr_shown == 60) {
645 if (time_before(jiffies, resume)) {
651 "BUG: Bad page state: %lu messages suppressed\n",
658 resume = jiffies + 60 * HZ;
660 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
661 current->comm, page_to_pfn(page));
662 dump_page(page, reason);
667 /* Leave bad fields for debug, except PageBuddy could make trouble */
668 page_mapcount_reset(page); /* remove PageBuddy */
669 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
673 * Higher-order pages are called "compound pages". They are structured thusly:
675 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
677 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
678 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
680 * The first tail page's ->compound_dtor holds the offset in array of compound
681 * page destructors. See compound_page_dtors.
683 * The first tail page's ->compound_order holds the order of allocation.
684 * This usage means that zero-order pages may not be compound.
687 void free_compound_page(struct page *page)
689 mem_cgroup_uncharge(page);
690 __free_pages_ok(page, compound_order(page), FPI_NONE);
693 void prep_compound_page(struct page *page, unsigned int order)
696 int nr_pages = 1 << order;
699 for (i = 1; i < nr_pages; i++) {
700 struct page *p = page + i;
701 set_page_count(p, 0);
702 p->mapping = TAIL_MAPPING;
703 set_compound_head(p, page);
706 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
707 set_compound_order(page, order);
708 atomic_set(compound_mapcount_ptr(page), -1);
709 if (hpage_pincount_available(page))
710 atomic_set(compound_pincount_ptr(page), 0);
713 #ifdef CONFIG_DEBUG_PAGEALLOC
714 unsigned int _debug_guardpage_minorder;
716 bool _debug_pagealloc_enabled_early __read_mostly
717 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
718 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
719 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
720 EXPORT_SYMBOL(_debug_pagealloc_enabled);
722 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
724 static int __init early_debug_pagealloc(char *buf)
726 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
728 early_param("debug_pagealloc", early_debug_pagealloc);
730 static int __init debug_guardpage_minorder_setup(char *buf)
734 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
735 pr_err("Bad debug_guardpage_minorder value\n");
738 _debug_guardpage_minorder = res;
739 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
742 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
744 static inline bool set_page_guard(struct zone *zone, struct page *page,
745 unsigned int order, int migratetype)
747 if (!debug_guardpage_enabled())
750 if (order >= debug_guardpage_minorder())
753 __SetPageGuard(page);
754 INIT_LIST_HEAD(&page->lru);
755 set_page_private(page, order);
756 /* Guard pages are not available for any usage */
757 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
762 static inline void clear_page_guard(struct zone *zone, struct page *page,
763 unsigned int order, int migratetype)
765 if (!debug_guardpage_enabled())
768 __ClearPageGuard(page);
770 set_page_private(page, 0);
771 if (!is_migrate_isolate(migratetype))
772 __mod_zone_freepage_state(zone, (1 << order), migratetype);
775 static inline bool set_page_guard(struct zone *zone, struct page *page,
776 unsigned int order, int migratetype) { return false; }
777 static inline void clear_page_guard(struct zone *zone, struct page *page,
778 unsigned int order, int migratetype) {}
782 * Enable static keys related to various memory debugging and hardening options.
783 * Some override others, and depend on early params that are evaluated in the
784 * order of appearance. So we need to first gather the full picture of what was
785 * enabled, and then make decisions.
787 void init_mem_debugging_and_hardening(void)
789 bool page_poisoning_requested = false;
791 #ifdef CONFIG_PAGE_POISONING
793 * Page poisoning is debug page alloc for some arches. If
794 * either of those options are enabled, enable poisoning.
796 if (page_poisoning_enabled() ||
797 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
798 debug_pagealloc_enabled())) {
799 static_branch_enable(&_page_poisoning_enabled);
800 page_poisoning_requested = true;
804 if (_init_on_alloc_enabled_early) {
805 if (page_poisoning_requested)
806 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
807 "will take precedence over init_on_alloc\n");
809 static_branch_enable(&init_on_alloc);
811 if (_init_on_free_enabled_early) {
812 if (page_poisoning_requested)
813 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
814 "will take precedence over init_on_free\n");
816 static_branch_enable(&init_on_free);
819 #ifdef CONFIG_DEBUG_PAGEALLOC
820 if (!debug_pagealloc_enabled())
823 static_branch_enable(&_debug_pagealloc_enabled);
825 if (!debug_guardpage_minorder())
828 static_branch_enable(&_debug_guardpage_enabled);
832 static inline void set_buddy_order(struct page *page, unsigned int order)
834 set_page_private(page, order);
835 __SetPageBuddy(page);
839 * This function checks whether a page is free && is the buddy
840 * we can coalesce a page and its buddy if
841 * (a) the buddy is not in a hole (check before calling!) &&
842 * (b) the buddy is in the buddy system &&
843 * (c) a page and its buddy have the same order &&
844 * (d) a page and its buddy are in the same zone.
846 * For recording whether a page is in the buddy system, we set PageBuddy.
847 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
849 * For recording page's order, we use page_private(page).
851 static inline bool page_is_buddy(struct page *page, struct page *buddy,
854 if (!page_is_guard(buddy) && !PageBuddy(buddy))
857 if (buddy_order(buddy) != order)
861 * zone check is done late to avoid uselessly calculating
862 * zone/node ids for pages that could never merge.
864 if (page_zone_id(page) != page_zone_id(buddy))
867 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
872 #ifdef CONFIG_COMPACTION
873 static inline struct capture_control *task_capc(struct zone *zone)
875 struct capture_control *capc = current->capture_control;
877 return unlikely(capc) &&
878 !(current->flags & PF_KTHREAD) &&
880 capc->cc->zone == zone ? capc : NULL;
884 compaction_capture(struct capture_control *capc, struct page *page,
885 int order, int migratetype)
887 if (!capc || order != capc->cc->order)
890 /* Do not accidentally pollute CMA or isolated regions*/
891 if (is_migrate_cma(migratetype) ||
892 is_migrate_isolate(migratetype))
896 * Do not let lower order allocations pollute a movable pageblock.
897 * This might let an unmovable request use a reclaimable pageblock
898 * and vice-versa but no more than normal fallback logic which can
899 * have trouble finding a high-order free page.
901 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
909 static inline struct capture_control *task_capc(struct zone *zone)
915 compaction_capture(struct capture_control *capc, struct page *page,
916 int order, int migratetype)
920 #endif /* CONFIG_COMPACTION */
922 /* Used for pages not on another list */
923 static inline void add_to_free_list(struct page *page, struct zone *zone,
924 unsigned int order, int migratetype)
926 struct free_area *area = &zone->free_area[order];
928 list_add(&page->lru, &area->free_list[migratetype]);
932 /* Used for pages not on another list */
933 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
934 unsigned int order, int migratetype)
936 struct free_area *area = &zone->free_area[order];
938 list_add_tail(&page->lru, &area->free_list[migratetype]);
943 * Used for pages which are on another list. Move the pages to the tail
944 * of the list - so the moved pages won't immediately be considered for
945 * allocation again (e.g., optimization for memory onlining).
947 static inline void move_to_free_list(struct page *page, struct zone *zone,
948 unsigned int order, int migratetype)
950 struct free_area *area = &zone->free_area[order];
952 list_move_tail(&page->lru, &area->free_list[migratetype]);
955 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
958 /* clear reported state and update reported page count */
959 if (page_reported(page))
960 __ClearPageReported(page);
962 list_del(&page->lru);
963 __ClearPageBuddy(page);
964 set_page_private(page, 0);
965 zone->free_area[order].nr_free--;
969 * If this is not the largest possible page, check if the buddy
970 * of the next-highest order is free. If it is, it's possible
971 * that pages are being freed that will coalesce soon. In case,
972 * that is happening, add the free page to the tail of the list
973 * so it's less likely to be used soon and more likely to be merged
974 * as a higher order page
977 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
978 struct page *page, unsigned int order)
980 struct page *higher_page, *higher_buddy;
981 unsigned long combined_pfn;
983 if (order >= MAX_ORDER - 2)
986 if (!pfn_valid_within(buddy_pfn))
989 combined_pfn = buddy_pfn & pfn;
990 higher_page = page + (combined_pfn - pfn);
991 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
992 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
994 return pfn_valid_within(buddy_pfn) &&
995 page_is_buddy(higher_page, higher_buddy, order + 1);
999 * Freeing function for a buddy system allocator.
1001 * The concept of a buddy system is to maintain direct-mapped table
1002 * (containing bit values) for memory blocks of various "orders".
1003 * The bottom level table contains the map for the smallest allocatable
1004 * units of memory (here, pages), and each level above it describes
1005 * pairs of units from the levels below, hence, "buddies".
1006 * At a high level, all that happens here is marking the table entry
1007 * at the bottom level available, and propagating the changes upward
1008 * as necessary, plus some accounting needed to play nicely with other
1009 * parts of the VM system.
1010 * At each level, we keep a list of pages, which are heads of continuous
1011 * free pages of length of (1 << order) and marked with PageBuddy.
1012 * Page's order is recorded in page_private(page) field.
1013 * So when we are allocating or freeing one, we can derive the state of the
1014 * other. That is, if we allocate a small block, and both were
1015 * free, the remainder of the region must be split into blocks.
1016 * If a block is freed, and its buddy is also free, then this
1017 * triggers coalescing into a block of larger size.
1022 static inline void __free_one_page(struct page *page,
1024 struct zone *zone, unsigned int order,
1025 int migratetype, fpi_t fpi_flags)
1027 struct capture_control *capc = task_capc(zone);
1028 unsigned long buddy_pfn;
1029 unsigned long combined_pfn;
1030 unsigned int max_order;
1034 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1036 VM_BUG_ON(!zone_is_initialized(zone));
1037 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1039 VM_BUG_ON(migratetype == -1);
1040 if (likely(!is_migrate_isolate(migratetype)))
1041 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1043 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1044 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1047 while (order < max_order) {
1048 if (compaction_capture(capc, page, order, migratetype)) {
1049 __mod_zone_freepage_state(zone, -(1 << order),
1053 buddy_pfn = __find_buddy_pfn(pfn, order);
1054 buddy = page + (buddy_pfn - pfn);
1056 if (!pfn_valid_within(buddy_pfn))
1058 if (!page_is_buddy(page, buddy, order))
1061 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1062 * merge with it and move up one order.
1064 if (page_is_guard(buddy))
1065 clear_page_guard(zone, buddy, order, migratetype);
1067 del_page_from_free_list(buddy, zone, order);
1068 combined_pfn = buddy_pfn & pfn;
1069 page = page + (combined_pfn - pfn);
1073 if (order < MAX_ORDER - 1) {
1074 /* If we are here, it means order is >= pageblock_order.
1075 * We want to prevent merge between freepages on isolate
1076 * pageblock and normal pageblock. Without this, pageblock
1077 * isolation could cause incorrect freepage or CMA accounting.
1079 * We don't want to hit this code for the more frequent
1080 * low-order merging.
1082 if (unlikely(has_isolate_pageblock(zone))) {
1085 buddy_pfn = __find_buddy_pfn(pfn, order);
1086 buddy = page + (buddy_pfn - pfn);
1087 buddy_mt = get_pageblock_migratetype(buddy);
1089 if (migratetype != buddy_mt
1090 && (is_migrate_isolate(migratetype) ||
1091 is_migrate_isolate(buddy_mt)))
1094 max_order = order + 1;
1095 goto continue_merging;
1099 set_buddy_order(page, order);
1101 if (fpi_flags & FPI_TO_TAIL)
1103 else if (is_shuffle_order(order))
1104 to_tail = shuffle_pick_tail();
1106 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1109 add_to_free_list_tail(page, zone, order, migratetype);
1111 add_to_free_list(page, zone, order, migratetype);
1113 /* Notify page reporting subsystem of freed page */
1114 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1115 page_reporting_notify_free(order);
1119 * A bad page could be due to a number of fields. Instead of multiple branches,
1120 * try and check multiple fields with one check. The caller must do a detailed
1121 * check if necessary.
1123 static inline bool page_expected_state(struct page *page,
1124 unsigned long check_flags)
1126 if (unlikely(atomic_read(&page->_mapcount) != -1))
1129 if (unlikely((unsigned long)page->mapping |
1130 page_ref_count(page) |
1134 (page->flags & check_flags)))
1140 static const char *page_bad_reason(struct page *page, unsigned long flags)
1142 const char *bad_reason = NULL;
1144 if (unlikely(atomic_read(&page->_mapcount) != -1))
1145 bad_reason = "nonzero mapcount";
1146 if (unlikely(page->mapping != NULL))
1147 bad_reason = "non-NULL mapping";
1148 if (unlikely(page_ref_count(page) != 0))
1149 bad_reason = "nonzero _refcount";
1150 if (unlikely(page->flags & flags)) {
1151 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1152 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1154 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1157 if (unlikely(page->memcg_data))
1158 bad_reason = "page still charged to cgroup";
1163 static void check_free_page_bad(struct page *page)
1166 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1169 static inline int check_free_page(struct page *page)
1171 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1174 /* Something has gone sideways, find it */
1175 check_free_page_bad(page);
1179 static int free_tail_pages_check(struct page *head_page, struct page *page)
1184 * We rely page->lru.next never has bit 0 set, unless the page
1185 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1187 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1189 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1193 switch (page - head_page) {
1195 /* the first tail page: ->mapping may be compound_mapcount() */
1196 if (unlikely(compound_mapcount(page))) {
1197 bad_page(page, "nonzero compound_mapcount");
1203 * the second tail page: ->mapping is
1204 * deferred_list.next -- ignore value.
1208 if (page->mapping != TAIL_MAPPING) {
1209 bad_page(page, "corrupted mapping in tail page");
1214 if (unlikely(!PageTail(page))) {
1215 bad_page(page, "PageTail not set");
1218 if (unlikely(compound_head(page) != head_page)) {
1219 bad_page(page, "compound_head not consistent");
1224 page->mapping = NULL;
1225 clear_compound_head(page);
1229 static void kernel_init_free_pages(struct page *page, int numpages)
1233 /* s390's use of memset() could override KASAN redzones. */
1234 kasan_disable_current();
1235 for (i = 0; i < numpages; i++) {
1236 u8 tag = page_kasan_tag(page + i);
1237 page_kasan_tag_reset(page + i);
1238 clear_highpage(page + i);
1239 page_kasan_tag_set(page + i, tag);
1241 kasan_enable_current();
1244 static __always_inline bool free_pages_prepare(struct page *page,
1245 unsigned int order, bool check_free, fpi_t fpi_flags)
1250 VM_BUG_ON_PAGE(PageTail(page), page);
1252 trace_mm_page_free(page, order);
1254 if (unlikely(PageHWPoison(page)) && !order) {
1256 * Do not let hwpoison pages hit pcplists/buddy
1257 * Untie memcg state and reset page's owner
1259 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1260 __memcg_kmem_uncharge_page(page, order);
1261 reset_page_owner(page, order);
1266 * Check tail pages before head page information is cleared to
1267 * avoid checking PageCompound for order-0 pages.
1269 if (unlikely(order)) {
1270 bool compound = PageCompound(page);
1273 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1276 ClearPageDoubleMap(page);
1277 for (i = 1; i < (1 << order); i++) {
1279 bad += free_tail_pages_check(page, page + i);
1280 if (unlikely(check_free_page(page + i))) {
1284 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1287 if (PageMappingFlags(page))
1288 page->mapping = NULL;
1289 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1290 __memcg_kmem_uncharge_page(page, order);
1292 bad += check_free_page(page);
1296 page_cpupid_reset_last(page);
1297 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1298 reset_page_owner(page, order);
1300 if (!PageHighMem(page)) {
1301 debug_check_no_locks_freed(page_address(page),
1302 PAGE_SIZE << order);
1303 debug_check_no_obj_freed(page_address(page),
1304 PAGE_SIZE << order);
1307 kernel_poison_pages(page, 1 << order);
1310 * As memory initialization might be integrated into KASAN,
1311 * kasan_free_pages and kernel_init_free_pages must be
1312 * kept together to avoid discrepancies in behavior.
1314 * With hardware tag-based KASAN, memory tags must be set before the
1315 * page becomes unavailable via debug_pagealloc or arch_free_page.
1317 init = want_init_on_free();
1318 if (init && !kasan_has_integrated_init())
1319 kernel_init_free_pages(page, 1 << order);
1320 kasan_free_nondeferred_pages(page, order, init, fpi_flags);
1323 * arch_free_page() can make the page's contents inaccessible. s390
1324 * does this. So nothing which can access the page's contents should
1325 * happen after this.
1327 arch_free_page(page, order);
1329 debug_pagealloc_unmap_pages(page, 1 << order);
1334 #ifdef CONFIG_DEBUG_VM
1336 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1337 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1338 * moved from pcp lists to free lists.
1340 static bool free_pcp_prepare(struct page *page)
1342 return free_pages_prepare(page, 0, true, FPI_NONE);
1345 static bool bulkfree_pcp_prepare(struct page *page)
1347 if (debug_pagealloc_enabled_static())
1348 return check_free_page(page);
1354 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1355 * moving from pcp lists to free list in order to reduce overhead. With
1356 * debug_pagealloc enabled, they are checked also immediately when being freed
1359 static bool free_pcp_prepare(struct page *page)
1361 if (debug_pagealloc_enabled_static())
1362 return free_pages_prepare(page, 0, true, FPI_NONE);
1364 return free_pages_prepare(page, 0, false, FPI_NONE);
1367 static bool bulkfree_pcp_prepare(struct page *page)
1369 return check_free_page(page);
1371 #endif /* CONFIG_DEBUG_VM */
1373 static inline void prefetch_buddy(struct page *page)
1375 unsigned long pfn = page_to_pfn(page);
1376 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1377 struct page *buddy = page + (buddy_pfn - pfn);
1383 * Frees a number of pages from the PCP lists
1384 * Assumes all pages on list are in same zone, and of same order.
1385 * count is the number of pages to free.
1387 * If the zone was previously in an "all pages pinned" state then look to
1388 * see if this freeing clears that state.
1390 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1391 * pinned" detection logic.
1393 static void free_pcppages_bulk(struct zone *zone, int count,
1394 struct per_cpu_pages *pcp)
1396 int migratetype = 0;
1398 int prefetch_nr = READ_ONCE(pcp->batch);
1399 bool isolated_pageblocks;
1400 struct page *page, *tmp;
1404 * Ensure proper count is passed which otherwise would stuck in the
1405 * below while (list_empty(list)) loop.
1407 count = min(pcp->count, count);
1409 struct list_head *list;
1412 * Remove pages from lists in a round-robin fashion. A
1413 * batch_free count is maintained that is incremented when an
1414 * empty list is encountered. This is so more pages are freed
1415 * off fuller lists instead of spinning excessively around empty
1420 if (++migratetype == MIGRATE_PCPTYPES)
1422 list = &pcp->lists[migratetype];
1423 } while (list_empty(list));
1425 /* This is the only non-empty list. Free them all. */
1426 if (batch_free == MIGRATE_PCPTYPES)
1430 page = list_last_entry(list, struct page, lru);
1431 /* must delete to avoid corrupting pcp list */
1432 list_del(&page->lru);
1435 if (bulkfree_pcp_prepare(page))
1438 list_add_tail(&page->lru, &head);
1441 * We are going to put the page back to the global
1442 * pool, prefetch its buddy to speed up later access
1443 * under zone->lock. It is believed the overhead of
1444 * an additional test and calculating buddy_pfn here
1445 * can be offset by reduced memory latency later. To
1446 * avoid excessive prefetching due to large count, only
1447 * prefetch buddy for the first pcp->batch nr of pages.
1450 prefetch_buddy(page);
1453 } while (--count && --batch_free && !list_empty(list));
1456 spin_lock(&zone->lock);
1457 isolated_pageblocks = has_isolate_pageblock(zone);
1460 * Use safe version since after __free_one_page(),
1461 * page->lru.next will not point to original list.
1463 list_for_each_entry_safe(page, tmp, &head, lru) {
1464 int mt = get_pcppage_migratetype(page);
1465 /* MIGRATE_ISOLATE page should not go to pcplists */
1466 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1467 /* Pageblock could have been isolated meanwhile */
1468 if (unlikely(isolated_pageblocks))
1469 mt = get_pageblock_migratetype(page);
1471 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1472 trace_mm_page_pcpu_drain(page, 0, mt);
1474 spin_unlock(&zone->lock);
1477 static void free_one_page(struct zone *zone,
1478 struct page *page, unsigned long pfn,
1480 int migratetype, fpi_t fpi_flags)
1482 spin_lock(&zone->lock);
1483 if (unlikely(has_isolate_pageblock(zone) ||
1484 is_migrate_isolate(migratetype))) {
1485 migratetype = get_pfnblock_migratetype(page, pfn);
1487 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1488 spin_unlock(&zone->lock);
1491 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1492 unsigned long zone, int nid)
1494 mm_zero_struct_page(page);
1495 set_page_links(page, zone, nid, pfn);
1496 init_page_count(page);
1497 page_mapcount_reset(page);
1498 page_cpupid_reset_last(page);
1499 page_kasan_tag_reset(page);
1501 INIT_LIST_HEAD(&page->lru);
1502 #ifdef WANT_PAGE_VIRTUAL
1503 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1504 if (!is_highmem_idx(zone))
1505 set_page_address(page, __va(pfn << PAGE_SHIFT));
1509 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1510 static void __meminit init_reserved_page(unsigned long pfn)
1515 if (!early_page_uninitialised(pfn))
1518 nid = early_pfn_to_nid(pfn);
1519 pgdat = NODE_DATA(nid);
1521 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1522 struct zone *zone = &pgdat->node_zones[zid];
1524 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1527 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1530 static inline void init_reserved_page(unsigned long pfn)
1533 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1536 * Initialised pages do not have PageReserved set. This function is
1537 * called for each range allocated by the bootmem allocator and
1538 * marks the pages PageReserved. The remaining valid pages are later
1539 * sent to the buddy page allocator.
1541 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1543 unsigned long start_pfn = PFN_DOWN(start);
1544 unsigned long end_pfn = PFN_UP(end);
1546 for (; start_pfn < end_pfn; start_pfn++) {
1547 if (pfn_valid(start_pfn)) {
1548 struct page *page = pfn_to_page(start_pfn);
1550 init_reserved_page(start_pfn);
1552 /* Avoid false-positive PageTail() */
1553 INIT_LIST_HEAD(&page->lru);
1556 * no need for atomic set_bit because the struct
1557 * page is not visible yet so nobody should
1560 __SetPageReserved(page);
1565 static void __free_pages_ok(struct page *page, unsigned int order,
1568 unsigned long flags;
1570 unsigned long pfn = page_to_pfn(page);
1572 if (!free_pages_prepare(page, order, true, fpi_flags))
1575 migratetype = get_pfnblock_migratetype(page, pfn);
1576 local_irq_save(flags);
1577 __count_vm_events(PGFREE, 1 << order);
1578 free_one_page(page_zone(page), page, pfn, order, migratetype,
1580 local_irq_restore(flags);
1583 void __free_pages_core(struct page *page, unsigned int order)
1585 unsigned int nr_pages = 1 << order;
1586 struct page *p = page;
1590 * When initializing the memmap, __init_single_page() sets the refcount
1591 * of all pages to 1 ("allocated"/"not free"). We have to set the
1592 * refcount of all involved pages to 0.
1595 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1597 __ClearPageReserved(p);
1598 set_page_count(p, 0);
1600 __ClearPageReserved(p);
1601 set_page_count(p, 0);
1603 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1606 * Bypass PCP and place fresh pages right to the tail, primarily
1607 * relevant for memory onlining.
1609 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1612 #ifdef CONFIG_NEED_MULTIPLE_NODES
1615 * During memory init memblocks map pfns to nids. The search is expensive and
1616 * this caches recent lookups. The implementation of __early_pfn_to_nid
1617 * treats start/end as pfns.
1619 struct mminit_pfnnid_cache {
1620 unsigned long last_start;
1621 unsigned long last_end;
1625 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1628 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1630 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1631 struct mminit_pfnnid_cache *state)
1633 unsigned long start_pfn, end_pfn;
1636 if (state->last_start <= pfn && pfn < state->last_end)
1637 return state->last_nid;
1639 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1640 if (nid != NUMA_NO_NODE) {
1641 state->last_start = start_pfn;
1642 state->last_end = end_pfn;
1643 state->last_nid = nid;
1649 int __meminit early_pfn_to_nid(unsigned long pfn)
1651 static DEFINE_SPINLOCK(early_pfn_lock);
1654 spin_lock(&early_pfn_lock);
1655 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1657 nid = first_online_node;
1658 spin_unlock(&early_pfn_lock);
1662 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1664 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1667 if (early_page_uninitialised(pfn))
1669 __free_pages_core(page, order);
1673 * Check that the whole (or subset of) a pageblock given by the interval of
1674 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1675 * with the migration of free compaction scanner. The scanners then need to
1676 * use only pfn_valid_within() check for arches that allow holes within
1679 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1681 * It's possible on some configurations to have a setup like node0 node1 node0
1682 * i.e. it's possible that all pages within a zones range of pages do not
1683 * belong to a single zone. We assume that a border between node0 and node1
1684 * can occur within a single pageblock, but not a node0 node1 node0
1685 * interleaving within a single pageblock. It is therefore sufficient to check
1686 * the first and last page of a pageblock and avoid checking each individual
1687 * page in a pageblock.
1689 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1690 unsigned long end_pfn, struct zone *zone)
1692 struct page *start_page;
1693 struct page *end_page;
1695 /* end_pfn is one past the range we are checking */
1698 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1701 start_page = pfn_to_online_page(start_pfn);
1705 if (page_zone(start_page) != zone)
1708 end_page = pfn_to_page(end_pfn);
1710 /* This gives a shorter code than deriving page_zone(end_page) */
1711 if (page_zone_id(start_page) != page_zone_id(end_page))
1717 void set_zone_contiguous(struct zone *zone)
1719 unsigned long block_start_pfn = zone->zone_start_pfn;
1720 unsigned long block_end_pfn;
1722 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1723 for (; block_start_pfn < zone_end_pfn(zone);
1724 block_start_pfn = block_end_pfn,
1725 block_end_pfn += pageblock_nr_pages) {
1727 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1729 if (!__pageblock_pfn_to_page(block_start_pfn,
1730 block_end_pfn, zone))
1735 /* We confirm that there is no hole */
1736 zone->contiguous = true;
1739 void clear_zone_contiguous(struct zone *zone)
1741 zone->contiguous = false;
1744 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1745 static void __init deferred_free_range(unsigned long pfn,
1746 unsigned long nr_pages)
1754 page = pfn_to_page(pfn);
1756 /* Free a large naturally-aligned chunk if possible */
1757 if (nr_pages == pageblock_nr_pages &&
1758 (pfn & (pageblock_nr_pages - 1)) == 0) {
1759 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1760 __free_pages_core(page, pageblock_order);
1764 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1765 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1766 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1767 __free_pages_core(page, 0);
1771 /* Completion tracking for deferred_init_memmap() threads */
1772 static atomic_t pgdat_init_n_undone __initdata;
1773 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1775 static inline void __init pgdat_init_report_one_done(void)
1777 if (atomic_dec_and_test(&pgdat_init_n_undone))
1778 complete(&pgdat_init_all_done_comp);
1782 * Returns true if page needs to be initialized or freed to buddy allocator.
1784 * First we check if pfn is valid on architectures where it is possible to have
1785 * holes within pageblock_nr_pages. On systems where it is not possible, this
1786 * function is optimized out.
1788 * Then, we check if a current large page is valid by only checking the validity
1791 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1793 if (!pfn_valid_within(pfn))
1795 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1801 * Free pages to buddy allocator. Try to free aligned pages in
1802 * pageblock_nr_pages sizes.
1804 static void __init deferred_free_pages(unsigned long pfn,
1805 unsigned long end_pfn)
1807 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1808 unsigned long nr_free = 0;
1810 for (; pfn < end_pfn; pfn++) {
1811 if (!deferred_pfn_valid(pfn)) {
1812 deferred_free_range(pfn - nr_free, nr_free);
1814 } else if (!(pfn & nr_pgmask)) {
1815 deferred_free_range(pfn - nr_free, nr_free);
1821 /* Free the last block of pages to allocator */
1822 deferred_free_range(pfn - nr_free, nr_free);
1826 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1827 * by performing it only once every pageblock_nr_pages.
1828 * Return number of pages initialized.
1830 static unsigned long __init deferred_init_pages(struct zone *zone,
1832 unsigned long end_pfn)
1834 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1835 int nid = zone_to_nid(zone);
1836 unsigned long nr_pages = 0;
1837 int zid = zone_idx(zone);
1838 struct page *page = NULL;
1840 for (; pfn < end_pfn; pfn++) {
1841 if (!deferred_pfn_valid(pfn)) {
1844 } else if (!page || !(pfn & nr_pgmask)) {
1845 page = pfn_to_page(pfn);
1849 __init_single_page(page, pfn, zid, nid);
1856 * This function is meant to pre-load the iterator for the zone init.
1857 * Specifically it walks through the ranges until we are caught up to the
1858 * first_init_pfn value and exits there. If we never encounter the value we
1859 * return false indicating there are no valid ranges left.
1862 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1863 unsigned long *spfn, unsigned long *epfn,
1864 unsigned long first_init_pfn)
1869 * Start out by walking through the ranges in this zone that have
1870 * already been initialized. We don't need to do anything with them
1871 * so we just need to flush them out of the system.
1873 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1874 if (*epfn <= first_init_pfn)
1876 if (*spfn < first_init_pfn)
1877 *spfn = first_init_pfn;
1886 * Initialize and free pages. We do it in two loops: first we initialize
1887 * struct page, then free to buddy allocator, because while we are
1888 * freeing pages we can access pages that are ahead (computing buddy
1889 * page in __free_one_page()).
1891 * In order to try and keep some memory in the cache we have the loop
1892 * broken along max page order boundaries. This way we will not cause
1893 * any issues with the buddy page computation.
1895 static unsigned long __init
1896 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1897 unsigned long *end_pfn)
1899 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1900 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1901 unsigned long nr_pages = 0;
1904 /* First we loop through and initialize the page values */
1905 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1908 if (mo_pfn <= *start_pfn)
1911 t = min(mo_pfn, *end_pfn);
1912 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1914 if (mo_pfn < *end_pfn) {
1915 *start_pfn = mo_pfn;
1920 /* Reset values and now loop through freeing pages as needed */
1923 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1929 t = min(mo_pfn, epfn);
1930 deferred_free_pages(spfn, t);
1940 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1943 unsigned long spfn, epfn;
1944 struct zone *zone = arg;
1947 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1950 * Initialize and free pages in MAX_ORDER sized increments so that we
1951 * can avoid introducing any issues with the buddy allocator.
1953 while (spfn < end_pfn) {
1954 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1959 /* An arch may override for more concurrency. */
1961 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1966 /* Initialise remaining memory on a node */
1967 static int __init deferred_init_memmap(void *data)
1969 pg_data_t *pgdat = data;
1970 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1971 unsigned long spfn = 0, epfn = 0;
1972 unsigned long first_init_pfn, flags;
1973 unsigned long start = jiffies;
1975 int zid, max_threads;
1978 /* Bind memory initialisation thread to a local node if possible */
1979 if (!cpumask_empty(cpumask))
1980 set_cpus_allowed_ptr(current, cpumask);
1982 pgdat_resize_lock(pgdat, &flags);
1983 first_init_pfn = pgdat->first_deferred_pfn;
1984 if (first_init_pfn == ULONG_MAX) {
1985 pgdat_resize_unlock(pgdat, &flags);
1986 pgdat_init_report_one_done();
1990 /* Sanity check boundaries */
1991 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1992 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1993 pgdat->first_deferred_pfn = ULONG_MAX;
1996 * Once we unlock here, the zone cannot be grown anymore, thus if an
1997 * interrupt thread must allocate this early in boot, zone must be
1998 * pre-grown prior to start of deferred page initialization.
2000 pgdat_resize_unlock(pgdat, &flags);
2002 /* Only the highest zone is deferred so find it */
2003 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2004 zone = pgdat->node_zones + zid;
2005 if (first_init_pfn < zone_end_pfn(zone))
2009 /* If the zone is empty somebody else may have cleared out the zone */
2010 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2014 max_threads = deferred_page_init_max_threads(cpumask);
2016 while (spfn < epfn) {
2017 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2018 struct padata_mt_job job = {
2019 .thread_fn = deferred_init_memmap_chunk,
2022 .size = epfn_align - spfn,
2023 .align = PAGES_PER_SECTION,
2024 .min_chunk = PAGES_PER_SECTION,
2025 .max_threads = max_threads,
2028 padata_do_multithreaded(&job);
2029 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2033 /* Sanity check that the next zone really is unpopulated */
2034 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2036 pr_info("node %d deferred pages initialised in %ums\n",
2037 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2039 pgdat_init_report_one_done();
2044 * If this zone has deferred pages, try to grow it by initializing enough
2045 * deferred pages to satisfy the allocation specified by order, rounded up to
2046 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2047 * of SECTION_SIZE bytes by initializing struct pages in increments of
2048 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2050 * Return true when zone was grown, otherwise return false. We return true even
2051 * when we grow less than requested, to let the caller decide if there are
2052 * enough pages to satisfy the allocation.
2054 * Note: We use noinline because this function is needed only during boot, and
2055 * it is called from a __ref function _deferred_grow_zone. This way we are
2056 * making sure that it is not inlined into permanent text section.
2058 static noinline bool __init
2059 deferred_grow_zone(struct zone *zone, unsigned int order)
2061 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2062 pg_data_t *pgdat = zone->zone_pgdat;
2063 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2064 unsigned long spfn, epfn, flags;
2065 unsigned long nr_pages = 0;
2068 /* Only the last zone may have deferred pages */
2069 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2072 pgdat_resize_lock(pgdat, &flags);
2075 * If someone grew this zone while we were waiting for spinlock, return
2076 * true, as there might be enough pages already.
2078 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2079 pgdat_resize_unlock(pgdat, &flags);
2083 /* If the zone is empty somebody else may have cleared out the zone */
2084 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2085 first_deferred_pfn)) {
2086 pgdat->first_deferred_pfn = ULONG_MAX;
2087 pgdat_resize_unlock(pgdat, &flags);
2088 /* Retry only once. */
2089 return first_deferred_pfn != ULONG_MAX;
2093 * Initialize and free pages in MAX_ORDER sized increments so
2094 * that we can avoid introducing any issues with the buddy
2097 while (spfn < epfn) {
2098 /* update our first deferred PFN for this section */
2099 first_deferred_pfn = spfn;
2101 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2102 touch_nmi_watchdog();
2104 /* We should only stop along section boundaries */
2105 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2108 /* If our quota has been met we can stop here */
2109 if (nr_pages >= nr_pages_needed)
2113 pgdat->first_deferred_pfn = spfn;
2114 pgdat_resize_unlock(pgdat, &flags);
2116 return nr_pages > 0;
2120 * deferred_grow_zone() is __init, but it is called from
2121 * get_page_from_freelist() during early boot until deferred_pages permanently
2122 * disables this call. This is why we have refdata wrapper to avoid warning,
2123 * and to ensure that the function body gets unloaded.
2126 _deferred_grow_zone(struct zone *zone, unsigned int order)
2128 return deferred_grow_zone(zone, order);
2131 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2133 void __init page_alloc_init_late(void)
2138 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2140 /* There will be num_node_state(N_MEMORY) threads */
2141 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2142 for_each_node_state(nid, N_MEMORY) {
2143 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2146 /* Block until all are initialised */
2147 wait_for_completion(&pgdat_init_all_done_comp);
2150 * The number of managed pages has changed due to the initialisation
2151 * so the pcpu batch and high limits needs to be updated or the limits
2152 * will be artificially small.
2154 for_each_populated_zone(zone)
2155 zone_pcp_update(zone);
2158 * We initialized the rest of the deferred pages. Permanently disable
2159 * on-demand struct page initialization.
2161 static_branch_disable(&deferred_pages);
2163 /* Reinit limits that are based on free pages after the kernel is up */
2164 files_maxfiles_init();
2169 /* Discard memblock private memory */
2172 for_each_node_state(nid, N_MEMORY)
2173 shuffle_free_memory(NODE_DATA(nid));
2175 for_each_populated_zone(zone)
2176 set_zone_contiguous(zone);
2180 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2181 void __init init_cma_reserved_pageblock(struct page *page)
2183 unsigned i = pageblock_nr_pages;
2184 struct page *p = page;
2187 __ClearPageReserved(p);
2188 set_page_count(p, 0);
2191 set_pageblock_migratetype(page, MIGRATE_CMA);
2193 if (pageblock_order >= MAX_ORDER) {
2194 i = pageblock_nr_pages;
2197 set_page_refcounted(p);
2198 __free_pages(p, MAX_ORDER - 1);
2199 p += MAX_ORDER_NR_PAGES;
2200 } while (i -= MAX_ORDER_NR_PAGES);
2202 set_page_refcounted(page);
2203 __free_pages(page, pageblock_order);
2206 adjust_managed_page_count(page, pageblock_nr_pages);
2207 page_zone(page)->cma_pages += pageblock_nr_pages;
2212 * The order of subdivision here is critical for the IO subsystem.
2213 * Please do not alter this order without good reasons and regression
2214 * testing. Specifically, as large blocks of memory are subdivided,
2215 * the order in which smaller blocks are delivered depends on the order
2216 * they're subdivided in this function. This is the primary factor
2217 * influencing the order in which pages are delivered to the IO
2218 * subsystem according to empirical testing, and this is also justified
2219 * by considering the behavior of a buddy system containing a single
2220 * large block of memory acted on by a series of small allocations.
2221 * This behavior is a critical factor in sglist merging's success.
2225 static inline void expand(struct zone *zone, struct page *page,
2226 int low, int high, int migratetype)
2228 unsigned long size = 1 << high;
2230 while (high > low) {
2233 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2236 * Mark as guard pages (or page), that will allow to
2237 * merge back to allocator when buddy will be freed.
2238 * Corresponding page table entries will not be touched,
2239 * pages will stay not present in virtual address space
2241 if (set_page_guard(zone, &page[size], high, migratetype))
2244 add_to_free_list(&page[size], zone, high, migratetype);
2245 set_buddy_order(&page[size], high);
2249 static void check_new_page_bad(struct page *page)
2251 if (unlikely(page->flags & __PG_HWPOISON)) {
2252 /* Don't complain about hwpoisoned pages */
2253 page_mapcount_reset(page); /* remove PageBuddy */
2258 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2262 * This page is about to be returned from the page allocator
2264 static inline int check_new_page(struct page *page)
2266 if (likely(page_expected_state(page,
2267 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2270 check_new_page_bad(page);
2274 #ifdef CONFIG_DEBUG_VM
2276 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2277 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2278 * also checked when pcp lists are refilled from the free lists.
2280 static inline bool check_pcp_refill(struct page *page)
2282 if (debug_pagealloc_enabled_static())
2283 return check_new_page(page);
2288 static inline bool check_new_pcp(struct page *page)
2290 return check_new_page(page);
2294 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2295 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2296 * enabled, they are also checked when being allocated from the pcp lists.
2298 static inline bool check_pcp_refill(struct page *page)
2300 return check_new_page(page);
2302 static inline bool check_new_pcp(struct page *page)
2304 if (debug_pagealloc_enabled_static())
2305 return check_new_page(page);
2309 #endif /* CONFIG_DEBUG_VM */
2311 static bool check_new_pages(struct page *page, unsigned int order)
2314 for (i = 0; i < (1 << order); i++) {
2315 struct page *p = page + i;
2317 if (unlikely(check_new_page(p)))
2324 inline void post_alloc_hook(struct page *page, unsigned int order,
2329 set_page_private(page, 0);
2330 set_page_refcounted(page);
2332 arch_alloc_page(page, order);
2333 debug_pagealloc_map_pages(page, 1 << order);
2336 * Page unpoisoning must happen before memory initialization.
2337 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2338 * allocations and the page unpoisoning code will complain.
2340 kernel_unpoison_pages(page, 1 << order);
2343 * As memory initialization might be integrated into KASAN,
2344 * kasan_alloc_pages and kernel_init_free_pages must be
2345 * kept together to avoid discrepancies in behavior.
2347 init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2348 kasan_alloc_pages(page, order, init);
2349 if (init && !kasan_has_integrated_init())
2350 kernel_init_free_pages(page, 1 << order);
2352 set_page_owner(page, order, gfp_flags);
2355 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2356 unsigned int alloc_flags)
2358 post_alloc_hook(page, order, gfp_flags);
2360 if (order && (gfp_flags & __GFP_COMP))
2361 prep_compound_page(page, order);
2364 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2365 * allocate the page. The expectation is that the caller is taking
2366 * steps that will free more memory. The caller should avoid the page
2367 * being used for !PFMEMALLOC purposes.
2369 if (alloc_flags & ALLOC_NO_WATERMARKS)
2370 set_page_pfmemalloc(page);
2372 clear_page_pfmemalloc(page);
2376 * Go through the free lists for the given migratetype and remove
2377 * the smallest available page from the freelists
2379 static __always_inline
2380 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2383 unsigned int current_order;
2384 struct free_area *area;
2387 /* Find a page of the appropriate size in the preferred list */
2388 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2389 area = &(zone->free_area[current_order]);
2390 page = get_page_from_free_area(area, migratetype);
2393 del_page_from_free_list(page, zone, current_order);
2394 expand(zone, page, order, current_order, migratetype);
2395 set_pcppage_migratetype(page, migratetype);
2404 * This array describes the order lists are fallen back to when
2405 * the free lists for the desirable migrate type are depleted
2407 static int fallbacks[MIGRATE_TYPES][3] = {
2408 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2409 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2410 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2412 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2414 #ifdef CONFIG_MEMORY_ISOLATION
2415 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2420 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2423 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2426 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2427 unsigned int order) { return NULL; }
2431 * Move the free pages in a range to the freelist tail of the requested type.
2432 * Note that start_page and end_pages are not aligned on a pageblock
2433 * boundary. If alignment is required, use move_freepages_block()
2435 static int move_freepages(struct zone *zone,
2436 unsigned long start_pfn, unsigned long end_pfn,
2437 int migratetype, int *num_movable)
2442 int pages_moved = 0;
2444 for (pfn = start_pfn; pfn <= end_pfn;) {
2445 if (!pfn_valid_within(pfn)) {
2450 page = pfn_to_page(pfn);
2451 if (!PageBuddy(page)) {
2453 * We assume that pages that could be isolated for
2454 * migration are movable. But we don't actually try
2455 * isolating, as that would be expensive.
2458 (PageLRU(page) || __PageMovable(page)))
2464 /* Make sure we are not inadvertently changing nodes */
2465 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2466 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2468 order = buddy_order(page);
2469 move_to_free_list(page, zone, order, migratetype);
2471 pages_moved += 1 << order;
2477 int move_freepages_block(struct zone *zone, struct page *page,
2478 int migratetype, int *num_movable)
2480 unsigned long start_pfn, end_pfn, pfn;
2485 pfn = page_to_pfn(page);
2486 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2487 end_pfn = start_pfn + pageblock_nr_pages - 1;
2489 /* Do not cross zone boundaries */
2490 if (!zone_spans_pfn(zone, start_pfn))
2492 if (!zone_spans_pfn(zone, end_pfn))
2495 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2499 static void change_pageblock_range(struct page *pageblock_page,
2500 int start_order, int migratetype)
2502 int nr_pageblocks = 1 << (start_order - pageblock_order);
2504 while (nr_pageblocks--) {
2505 set_pageblock_migratetype(pageblock_page, migratetype);
2506 pageblock_page += pageblock_nr_pages;
2511 * When we are falling back to another migratetype during allocation, try to
2512 * steal extra free pages from the same pageblocks to satisfy further
2513 * allocations, instead of polluting multiple pageblocks.
2515 * If we are stealing a relatively large buddy page, it is likely there will
2516 * be more free pages in the pageblock, so try to steal them all. For
2517 * reclaimable and unmovable allocations, we steal regardless of page size,
2518 * as fragmentation caused by those allocations polluting movable pageblocks
2519 * is worse than movable allocations stealing from unmovable and reclaimable
2522 static bool can_steal_fallback(unsigned int order, int start_mt)
2525 * Leaving this order check is intended, although there is
2526 * relaxed order check in next check. The reason is that
2527 * we can actually steal whole pageblock if this condition met,
2528 * but, below check doesn't guarantee it and that is just heuristic
2529 * so could be changed anytime.
2531 if (order >= pageblock_order)
2534 if (order >= pageblock_order / 2 ||
2535 start_mt == MIGRATE_RECLAIMABLE ||
2536 start_mt == MIGRATE_UNMOVABLE ||
2537 page_group_by_mobility_disabled)
2543 static inline bool boost_watermark(struct zone *zone)
2545 unsigned long max_boost;
2547 if (!watermark_boost_factor)
2550 * Don't bother in zones that are unlikely to produce results.
2551 * On small machines, including kdump capture kernels running
2552 * in a small area, boosting the watermark can cause an out of
2553 * memory situation immediately.
2555 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2558 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2559 watermark_boost_factor, 10000);
2562 * high watermark may be uninitialised if fragmentation occurs
2563 * very early in boot so do not boost. We do not fall
2564 * through and boost by pageblock_nr_pages as failing
2565 * allocations that early means that reclaim is not going
2566 * to help and it may even be impossible to reclaim the
2567 * boosted watermark resulting in a hang.
2572 max_boost = max(pageblock_nr_pages, max_boost);
2574 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2581 * This function implements actual steal behaviour. If order is large enough,
2582 * we can steal whole pageblock. If not, we first move freepages in this
2583 * pageblock to our migratetype and determine how many already-allocated pages
2584 * are there in the pageblock with a compatible migratetype. If at least half
2585 * of pages are free or compatible, we can change migratetype of the pageblock
2586 * itself, so pages freed in the future will be put on the correct free list.
2588 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2589 unsigned int alloc_flags, int start_type, bool whole_block)
2591 unsigned int current_order = buddy_order(page);
2592 int free_pages, movable_pages, alike_pages;
2595 old_block_type = get_pageblock_migratetype(page);
2598 * This can happen due to races and we want to prevent broken
2599 * highatomic accounting.
2601 if (is_migrate_highatomic(old_block_type))
2604 /* Take ownership for orders >= pageblock_order */
2605 if (current_order >= pageblock_order) {
2606 change_pageblock_range(page, current_order, start_type);
2611 * Boost watermarks to increase reclaim pressure to reduce the
2612 * likelihood of future fallbacks. Wake kswapd now as the node
2613 * may be balanced overall and kswapd will not wake naturally.
2615 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2616 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2618 /* We are not allowed to try stealing from the whole block */
2622 free_pages = move_freepages_block(zone, page, start_type,
2625 * Determine how many pages are compatible with our allocation.
2626 * For movable allocation, it's the number of movable pages which
2627 * we just obtained. For other types it's a bit more tricky.
2629 if (start_type == MIGRATE_MOVABLE) {
2630 alike_pages = movable_pages;
2633 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2634 * to MOVABLE pageblock, consider all non-movable pages as
2635 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2636 * vice versa, be conservative since we can't distinguish the
2637 * exact migratetype of non-movable pages.
2639 if (old_block_type == MIGRATE_MOVABLE)
2640 alike_pages = pageblock_nr_pages
2641 - (free_pages + movable_pages);
2646 /* moving whole block can fail due to zone boundary conditions */
2651 * If a sufficient number of pages in the block are either free or of
2652 * comparable migratability as our allocation, claim the whole block.
2654 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2655 page_group_by_mobility_disabled)
2656 set_pageblock_migratetype(page, start_type);
2661 move_to_free_list(page, zone, current_order, start_type);
2665 * Check whether there is a suitable fallback freepage with requested order.
2666 * If only_stealable is true, this function returns fallback_mt only if
2667 * we can steal other freepages all together. This would help to reduce
2668 * fragmentation due to mixed migratetype pages in one pageblock.
2670 int find_suitable_fallback(struct free_area *area, unsigned int order,
2671 int migratetype, bool only_stealable, bool *can_steal)
2676 if (area->nr_free == 0)
2681 fallback_mt = fallbacks[migratetype][i];
2682 if (fallback_mt == MIGRATE_TYPES)
2685 if (free_area_empty(area, fallback_mt))
2688 if (can_steal_fallback(order, migratetype))
2691 if (!only_stealable)
2702 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2703 * there are no empty page blocks that contain a page with a suitable order
2705 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2706 unsigned int alloc_order)
2709 unsigned long max_managed, flags;
2712 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2713 * Check is race-prone but harmless.
2715 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2716 if (zone->nr_reserved_highatomic >= max_managed)
2719 spin_lock_irqsave(&zone->lock, flags);
2721 /* Recheck the nr_reserved_highatomic limit under the lock */
2722 if (zone->nr_reserved_highatomic >= max_managed)
2726 mt = get_pageblock_migratetype(page);
2727 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2728 && !is_migrate_cma(mt)) {
2729 zone->nr_reserved_highatomic += pageblock_nr_pages;
2730 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2731 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2735 spin_unlock_irqrestore(&zone->lock, flags);
2739 * Used when an allocation is about to fail under memory pressure. This
2740 * potentially hurts the reliability of high-order allocations when under
2741 * intense memory pressure but failed atomic allocations should be easier
2742 * to recover from than an OOM.
2744 * If @force is true, try to unreserve a pageblock even though highatomic
2745 * pageblock is exhausted.
2747 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2750 struct zonelist *zonelist = ac->zonelist;
2751 unsigned long flags;
2758 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2761 * Preserve at least one pageblock unless memory pressure
2764 if (!force && zone->nr_reserved_highatomic <=
2768 spin_lock_irqsave(&zone->lock, flags);
2769 for (order = 0; order < MAX_ORDER; order++) {
2770 struct free_area *area = &(zone->free_area[order]);
2772 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2777 * In page freeing path, migratetype change is racy so
2778 * we can counter several free pages in a pageblock
2779 * in this loop although we changed the pageblock type
2780 * from highatomic to ac->migratetype. So we should
2781 * adjust the count once.
2783 if (is_migrate_highatomic_page(page)) {
2785 * It should never happen but changes to
2786 * locking could inadvertently allow a per-cpu
2787 * drain to add pages to MIGRATE_HIGHATOMIC
2788 * while unreserving so be safe and watch for
2791 zone->nr_reserved_highatomic -= min(
2793 zone->nr_reserved_highatomic);
2797 * Convert to ac->migratetype and avoid the normal
2798 * pageblock stealing heuristics. Minimally, the caller
2799 * is doing the work and needs the pages. More
2800 * importantly, if the block was always converted to
2801 * MIGRATE_UNMOVABLE or another type then the number
2802 * of pageblocks that cannot be completely freed
2805 set_pageblock_migratetype(page, ac->migratetype);
2806 ret = move_freepages_block(zone, page, ac->migratetype,
2809 spin_unlock_irqrestore(&zone->lock, flags);
2813 spin_unlock_irqrestore(&zone->lock, flags);
2820 * Try finding a free buddy page on the fallback list and put it on the free
2821 * list of requested migratetype, possibly along with other pages from the same
2822 * block, depending on fragmentation avoidance heuristics. Returns true if
2823 * fallback was found so that __rmqueue_smallest() can grab it.
2825 * The use of signed ints for order and current_order is a deliberate
2826 * deviation from the rest of this file, to make the for loop
2827 * condition simpler.
2829 static __always_inline bool
2830 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2831 unsigned int alloc_flags)
2833 struct free_area *area;
2835 int min_order = order;
2841 * Do not steal pages from freelists belonging to other pageblocks
2842 * i.e. orders < pageblock_order. If there are no local zones free,
2843 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2845 if (alloc_flags & ALLOC_NOFRAGMENT)
2846 min_order = pageblock_order;
2849 * Find the largest available free page in the other list. This roughly
2850 * approximates finding the pageblock with the most free pages, which
2851 * would be too costly to do exactly.
2853 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2855 area = &(zone->free_area[current_order]);
2856 fallback_mt = find_suitable_fallback(area, current_order,
2857 start_migratetype, false, &can_steal);
2858 if (fallback_mt == -1)
2862 * We cannot steal all free pages from the pageblock and the
2863 * requested migratetype is movable. In that case it's better to
2864 * steal and split the smallest available page instead of the
2865 * largest available page, because even if the next movable
2866 * allocation falls back into a different pageblock than this
2867 * one, it won't cause permanent fragmentation.
2869 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2870 && current_order > order)
2879 for (current_order = order; current_order < MAX_ORDER;
2881 area = &(zone->free_area[current_order]);
2882 fallback_mt = find_suitable_fallback(area, current_order,
2883 start_migratetype, false, &can_steal);
2884 if (fallback_mt != -1)
2889 * This should not happen - we already found a suitable fallback
2890 * when looking for the largest page.
2892 VM_BUG_ON(current_order == MAX_ORDER);
2895 page = get_page_from_free_area(area, fallback_mt);
2897 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2900 trace_mm_page_alloc_extfrag(page, order, current_order,
2901 start_migratetype, fallback_mt);
2908 * Do the hard work of removing an element from the buddy allocator.
2909 * Call me with the zone->lock already held.
2911 static __always_inline struct page *
2912 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2913 unsigned int alloc_flags)
2917 if (IS_ENABLED(CONFIG_CMA)) {
2919 * Balance movable allocations between regular and CMA areas by
2920 * allocating from CMA when over half of the zone's free memory
2921 * is in the CMA area.
2923 if (alloc_flags & ALLOC_CMA &&
2924 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2925 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2926 page = __rmqueue_cma_fallback(zone, order);
2932 page = __rmqueue_smallest(zone, order, migratetype);
2933 if (unlikely(!page)) {
2934 if (alloc_flags & ALLOC_CMA)
2935 page = __rmqueue_cma_fallback(zone, order);
2937 if (!page && __rmqueue_fallback(zone, order, migratetype,
2943 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2948 * Obtain a specified number of elements from the buddy allocator, all under
2949 * a single hold of the lock, for efficiency. Add them to the supplied list.
2950 * Returns the number of new pages which were placed at *list.
2952 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2953 unsigned long count, struct list_head *list,
2954 int migratetype, unsigned int alloc_flags)
2956 int i, allocated = 0;
2958 spin_lock(&zone->lock);
2959 for (i = 0; i < count; ++i) {
2960 struct page *page = __rmqueue(zone, order, migratetype,
2962 if (unlikely(page == NULL))
2965 if (unlikely(check_pcp_refill(page)))
2969 * Split buddy pages returned by expand() are received here in
2970 * physical page order. The page is added to the tail of
2971 * caller's list. From the callers perspective, the linked list
2972 * is ordered by page number under some conditions. This is
2973 * useful for IO devices that can forward direction from the
2974 * head, thus also in the physical page order. This is useful
2975 * for IO devices that can merge IO requests if the physical
2976 * pages are ordered properly.
2978 list_add_tail(&page->lru, list);
2980 if (is_migrate_cma(get_pcppage_migratetype(page)))
2981 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2986 * i pages were removed from the buddy list even if some leak due
2987 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2988 * on i. Do not confuse with 'allocated' which is the number of
2989 * pages added to the pcp list.
2991 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2992 spin_unlock(&zone->lock);
2998 * Called from the vmstat counter updater to drain pagesets of this
2999 * currently executing processor on remote nodes after they have
3002 * Note that this function must be called with the thread pinned to
3003 * a single processor.
3005 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3007 unsigned long flags;
3008 int to_drain, batch;
3010 local_irq_save(flags);
3011 batch = READ_ONCE(pcp->batch);
3012 to_drain = min(pcp->count, batch);
3014 free_pcppages_bulk(zone, to_drain, pcp);
3015 local_irq_restore(flags);
3020 * Drain pcplists of the indicated processor and zone.
3022 * The processor must either be the current processor and the
3023 * thread pinned to the current processor or a processor that
3026 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3028 unsigned long flags;
3029 struct per_cpu_pageset *pset;
3030 struct per_cpu_pages *pcp;
3032 local_irq_save(flags);
3033 pset = per_cpu_ptr(zone->pageset, cpu);
3037 free_pcppages_bulk(zone, pcp->count, pcp);
3038 local_irq_restore(flags);
3042 * Drain pcplists of all zones on the indicated processor.
3044 * The processor must either be the current processor and the
3045 * thread pinned to the current processor or a processor that
3048 static void drain_pages(unsigned int cpu)
3052 for_each_populated_zone(zone) {
3053 drain_pages_zone(cpu, zone);
3058 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3060 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3061 * the single zone's pages.
3063 void drain_local_pages(struct zone *zone)
3065 int cpu = smp_processor_id();
3068 drain_pages_zone(cpu, zone);
3073 static void drain_local_pages_wq(struct work_struct *work)
3075 struct pcpu_drain *drain;
3077 drain = container_of(work, struct pcpu_drain, work);
3080 * drain_all_pages doesn't use proper cpu hotplug protection so
3081 * we can race with cpu offline when the WQ can move this from
3082 * a cpu pinned worker to an unbound one. We can operate on a different
3083 * cpu which is alright but we also have to make sure to not move to
3087 drain_local_pages(drain->zone);
3092 * The implementation of drain_all_pages(), exposing an extra parameter to
3093 * drain on all cpus.
3095 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3096 * not empty. The check for non-emptiness can however race with a free to
3097 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3098 * that need the guarantee that every CPU has drained can disable the
3099 * optimizing racy check.
3101 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3106 * Allocate in the BSS so we wont require allocation in
3107 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3109 static cpumask_t cpus_with_pcps;
3112 * Make sure nobody triggers this path before mm_percpu_wq is fully
3115 if (WARN_ON_ONCE(!mm_percpu_wq))
3119 * Do not drain if one is already in progress unless it's specific to
3120 * a zone. Such callers are primarily CMA and memory hotplug and need
3121 * the drain to be complete when the call returns.
3123 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3126 mutex_lock(&pcpu_drain_mutex);
3130 * We don't care about racing with CPU hotplug event
3131 * as offline notification will cause the notified
3132 * cpu to drain that CPU pcps and on_each_cpu_mask
3133 * disables preemption as part of its processing
3135 for_each_online_cpu(cpu) {
3136 struct per_cpu_pageset *pcp;
3138 bool has_pcps = false;
3140 if (force_all_cpus) {
3142 * The pcp.count check is racy, some callers need a
3143 * guarantee that no cpu is missed.
3147 pcp = per_cpu_ptr(zone->pageset, cpu);
3151 for_each_populated_zone(z) {
3152 pcp = per_cpu_ptr(z->pageset, cpu);
3153 if (pcp->pcp.count) {
3161 cpumask_set_cpu(cpu, &cpus_with_pcps);
3163 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3166 for_each_cpu(cpu, &cpus_with_pcps) {
3167 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3170 INIT_WORK(&drain->work, drain_local_pages_wq);
3171 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3173 for_each_cpu(cpu, &cpus_with_pcps)
3174 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3176 mutex_unlock(&pcpu_drain_mutex);
3180 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3182 * When zone parameter is non-NULL, spill just the single zone's pages.
3184 * Note that this can be extremely slow as the draining happens in a workqueue.
3186 void drain_all_pages(struct zone *zone)
3188 __drain_all_pages(zone, false);
3191 #ifdef CONFIG_HIBERNATION
3194 * Touch the watchdog for every WD_PAGE_COUNT pages.
3196 #define WD_PAGE_COUNT (128*1024)
3198 void mark_free_pages(struct zone *zone)
3200 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3201 unsigned long flags;
3202 unsigned int order, t;
3205 if (zone_is_empty(zone))
3208 spin_lock_irqsave(&zone->lock, flags);
3210 max_zone_pfn = zone_end_pfn(zone);
3211 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3212 if (pfn_valid(pfn)) {
3213 page = pfn_to_page(pfn);
3215 if (!--page_count) {
3216 touch_nmi_watchdog();
3217 page_count = WD_PAGE_COUNT;
3220 if (page_zone(page) != zone)
3223 if (!swsusp_page_is_forbidden(page))
3224 swsusp_unset_page_free(page);
3227 for_each_migratetype_order(order, t) {
3228 list_for_each_entry(page,
3229 &zone->free_area[order].free_list[t], lru) {
3232 pfn = page_to_pfn(page);
3233 for (i = 0; i < (1UL << order); i++) {
3234 if (!--page_count) {
3235 touch_nmi_watchdog();
3236 page_count = WD_PAGE_COUNT;
3238 swsusp_set_page_free(pfn_to_page(pfn + i));
3242 spin_unlock_irqrestore(&zone->lock, flags);
3244 #endif /* CONFIG_PM */
3246 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3250 if (!free_pcp_prepare(page))
3253 migratetype = get_pfnblock_migratetype(page, pfn);
3254 set_pcppage_migratetype(page, migratetype);
3258 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3260 struct zone *zone = page_zone(page);
3261 struct per_cpu_pages *pcp;
3264 migratetype = get_pcppage_migratetype(page);
3265 __count_vm_event(PGFREE);
3268 * We only track unmovable, reclaimable and movable on pcp lists.
3269 * Free ISOLATE pages back to the allocator because they are being
3270 * offlined but treat HIGHATOMIC as movable pages so we can get those
3271 * areas back if necessary. Otherwise, we may have to free
3272 * excessively into the page allocator
3274 if (migratetype >= MIGRATE_PCPTYPES) {
3275 if (unlikely(is_migrate_isolate(migratetype))) {
3276 free_one_page(zone, page, pfn, 0, migratetype,
3280 migratetype = MIGRATE_MOVABLE;
3283 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3284 list_add(&page->lru, &pcp->lists[migratetype]);
3286 if (pcp->count >= READ_ONCE(pcp->high))
3287 free_pcppages_bulk(zone, READ_ONCE(pcp->batch), pcp);
3291 * Free a 0-order page
3293 void free_unref_page(struct page *page)
3295 unsigned long flags;
3296 unsigned long pfn = page_to_pfn(page);
3298 if (!free_unref_page_prepare(page, pfn))
3301 local_irq_save(flags);
3302 free_unref_page_commit(page, pfn);
3303 local_irq_restore(flags);
3307 * Free a list of 0-order pages
3309 void free_unref_page_list(struct list_head *list)
3311 struct page *page, *next;
3312 unsigned long flags, pfn;
3313 int batch_count = 0;
3315 /* Prepare pages for freeing */
3316 list_for_each_entry_safe(page, next, list, lru) {
3317 pfn = page_to_pfn(page);
3318 if (!free_unref_page_prepare(page, pfn))
3319 list_del(&page->lru);
3320 set_page_private(page, pfn);
3323 local_irq_save(flags);
3324 list_for_each_entry_safe(page, next, list, lru) {
3325 unsigned long pfn = page_private(page);
3327 set_page_private(page, 0);
3328 trace_mm_page_free_batched(page);
3329 free_unref_page_commit(page, pfn);
3332 * Guard against excessive IRQ disabled times when we get
3333 * a large list of pages to free.
3335 if (++batch_count == SWAP_CLUSTER_MAX) {
3336 local_irq_restore(flags);
3338 local_irq_save(flags);
3341 local_irq_restore(flags);
3345 * split_page takes a non-compound higher-order page, and splits it into
3346 * n (1<<order) sub-pages: page[0..n]
3347 * Each sub-page must be freed individually.
3349 * Note: this is probably too low level an operation for use in drivers.
3350 * Please consult with lkml before using this in your driver.
3352 void split_page(struct page *page, unsigned int order)
3356 VM_BUG_ON_PAGE(PageCompound(page), page);
3357 VM_BUG_ON_PAGE(!page_count(page), page);
3359 for (i = 1; i < (1 << order); i++)
3360 set_page_refcounted(page + i);
3361 split_page_owner(page, 1 << order);
3362 split_page_memcg(page, 1 << order);
3364 EXPORT_SYMBOL_GPL(split_page);
3366 int __isolate_free_page(struct page *page, unsigned int order)
3368 unsigned long watermark;
3372 BUG_ON(!PageBuddy(page));
3374 zone = page_zone(page);
3375 mt = get_pageblock_migratetype(page);
3377 if (!is_migrate_isolate(mt)) {
3379 * Obey watermarks as if the page was being allocated. We can
3380 * emulate a high-order watermark check with a raised order-0
3381 * watermark, because we already know our high-order page
3384 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3385 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3388 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3391 /* Remove page from free list */
3393 del_page_from_free_list(page, zone, order);
3396 * Set the pageblock if the isolated page is at least half of a
3399 if (order >= pageblock_order - 1) {
3400 struct page *endpage = page + (1 << order) - 1;
3401 for (; page < endpage; page += pageblock_nr_pages) {
3402 int mt = get_pageblock_migratetype(page);
3403 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3404 && !is_migrate_highatomic(mt))
3405 set_pageblock_migratetype(page,
3411 return 1UL << order;
3415 * __putback_isolated_page - Return a now-isolated page back where we got it
3416 * @page: Page that was isolated
3417 * @order: Order of the isolated page
3418 * @mt: The page's pageblock's migratetype
3420 * This function is meant to return a page pulled from the free lists via
3421 * __isolate_free_page back to the free lists they were pulled from.
3423 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3425 struct zone *zone = page_zone(page);
3427 /* zone lock should be held when this function is called */
3428 lockdep_assert_held(&zone->lock);
3430 /* Return isolated page to tail of freelist. */
3431 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3432 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3436 * Update NUMA hit/miss statistics
3438 * Must be called with interrupts disabled.
3440 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3443 enum numa_stat_item local_stat = NUMA_LOCAL;
3445 /* skip numa counters update if numa stats is disabled */
3446 if (!static_branch_likely(&vm_numa_stat_key))
3449 if (zone_to_nid(z) != numa_node_id())
3450 local_stat = NUMA_OTHER;
3452 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3453 __inc_numa_state(z, NUMA_HIT);
3455 __inc_numa_state(z, NUMA_MISS);
3456 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3458 __inc_numa_state(z, local_stat);
3462 /* Remove page from the per-cpu list, caller must protect the list */
3464 struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3465 unsigned int alloc_flags,
3466 struct per_cpu_pages *pcp,
3467 struct list_head *list)
3472 if (list_empty(list)) {
3473 pcp->count += rmqueue_bulk(zone, 0,
3474 READ_ONCE(pcp->batch), list,
3475 migratetype, alloc_flags);
3476 if (unlikely(list_empty(list)))
3480 page = list_first_entry(list, struct page, lru);
3481 list_del(&page->lru);
3483 } while (check_new_pcp(page));
3488 /* Lock and remove page from the per-cpu list */
3489 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3490 struct zone *zone, gfp_t gfp_flags,
3491 int migratetype, unsigned int alloc_flags)
3493 struct per_cpu_pages *pcp;
3494 struct list_head *list;
3496 unsigned long flags;
3498 local_irq_save(flags);
3499 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3500 list = &pcp->lists[migratetype];
3501 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3503 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3504 zone_statistics(preferred_zone, zone);
3506 local_irq_restore(flags);
3511 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3514 struct page *rmqueue(struct zone *preferred_zone,
3515 struct zone *zone, unsigned int order,
3516 gfp_t gfp_flags, unsigned int alloc_flags,
3519 unsigned long flags;
3522 if (likely(order == 0)) {
3524 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3525 * we need to skip it when CMA area isn't allowed.
3527 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3528 migratetype != MIGRATE_MOVABLE) {
3529 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3530 migratetype, alloc_flags);
3536 * We most definitely don't want callers attempting to
3537 * allocate greater than order-1 page units with __GFP_NOFAIL.
3539 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3540 spin_lock_irqsave(&zone->lock, flags);
3545 * order-0 request can reach here when the pcplist is skipped
3546 * due to non-CMA allocation context. HIGHATOMIC area is
3547 * reserved for high-order atomic allocation, so order-0
3548 * request should skip it.
3550 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3551 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3553 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3556 page = __rmqueue(zone, order, migratetype, alloc_flags);
3557 } while (page && check_new_pages(page, order));
3558 spin_unlock(&zone->lock);
3561 __mod_zone_freepage_state(zone, -(1 << order),
3562 get_pcppage_migratetype(page));
3564 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3565 zone_statistics(preferred_zone, zone);
3566 local_irq_restore(flags);
3569 /* Separate test+clear to avoid unnecessary atomics */
3570 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3571 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3572 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3575 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3579 local_irq_restore(flags);
3583 #ifdef CONFIG_FAIL_PAGE_ALLOC
3586 struct fault_attr attr;
3588 bool ignore_gfp_highmem;
3589 bool ignore_gfp_reclaim;
3591 } fail_page_alloc = {
3592 .attr = FAULT_ATTR_INITIALIZER,
3593 .ignore_gfp_reclaim = true,
3594 .ignore_gfp_highmem = true,
3598 static int __init setup_fail_page_alloc(char *str)
3600 return setup_fault_attr(&fail_page_alloc.attr, str);
3602 __setup("fail_page_alloc=", setup_fail_page_alloc);
3604 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3606 if (order < fail_page_alloc.min_order)
3608 if (gfp_mask & __GFP_NOFAIL)
3610 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3612 if (fail_page_alloc.ignore_gfp_reclaim &&
3613 (gfp_mask & __GFP_DIRECT_RECLAIM))
3616 return should_fail(&fail_page_alloc.attr, 1 << order);
3619 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3621 static int __init fail_page_alloc_debugfs(void)
3623 umode_t mode = S_IFREG | 0600;
3626 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3627 &fail_page_alloc.attr);
3629 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3630 &fail_page_alloc.ignore_gfp_reclaim);
3631 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3632 &fail_page_alloc.ignore_gfp_highmem);
3633 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3638 late_initcall(fail_page_alloc_debugfs);
3640 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3642 #else /* CONFIG_FAIL_PAGE_ALLOC */
3644 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3649 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3651 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3653 return __should_fail_alloc_page(gfp_mask, order);
3655 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3657 static inline long __zone_watermark_unusable_free(struct zone *z,
3658 unsigned int order, unsigned int alloc_flags)
3660 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3661 long unusable_free = (1 << order) - 1;
3664 * If the caller does not have rights to ALLOC_HARDER then subtract
3665 * the high-atomic reserves. This will over-estimate the size of the
3666 * atomic reserve but it avoids a search.
3668 if (likely(!alloc_harder))
3669 unusable_free += z->nr_reserved_highatomic;
3672 /* If allocation can't use CMA areas don't use free CMA pages */
3673 if (!(alloc_flags & ALLOC_CMA))
3674 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3677 return unusable_free;
3681 * Return true if free base pages are above 'mark'. For high-order checks it
3682 * will return true of the order-0 watermark is reached and there is at least
3683 * one free page of a suitable size. Checking now avoids taking the zone lock
3684 * to check in the allocation paths if no pages are free.
3686 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3687 int highest_zoneidx, unsigned int alloc_flags,
3692 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3694 /* free_pages may go negative - that's OK */
3695 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3697 if (alloc_flags & ALLOC_HIGH)
3700 if (unlikely(alloc_harder)) {
3702 * OOM victims can try even harder than normal ALLOC_HARDER
3703 * users on the grounds that it's definitely going to be in
3704 * the exit path shortly and free memory. Any allocation it
3705 * makes during the free path will be small and short-lived.
3707 if (alloc_flags & ALLOC_OOM)
3714 * Check watermarks for an order-0 allocation request. If these
3715 * are not met, then a high-order request also cannot go ahead
3716 * even if a suitable page happened to be free.
3718 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3721 /* If this is an order-0 request then the watermark is fine */
3725 /* For a high-order request, check at least one suitable page is free */
3726 for (o = order; o < MAX_ORDER; o++) {
3727 struct free_area *area = &z->free_area[o];
3733 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3734 if (!free_area_empty(area, mt))
3739 if ((alloc_flags & ALLOC_CMA) &&
3740 !free_area_empty(area, MIGRATE_CMA)) {
3744 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3750 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3751 int highest_zoneidx, unsigned int alloc_flags)
3753 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3754 zone_page_state(z, NR_FREE_PAGES));
3757 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3758 unsigned long mark, int highest_zoneidx,
3759 unsigned int alloc_flags, gfp_t gfp_mask)
3763 free_pages = zone_page_state(z, NR_FREE_PAGES);
3766 * Fast check for order-0 only. If this fails then the reserves
3767 * need to be calculated.
3772 fast_free = free_pages;
3773 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3774 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3778 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3782 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3783 * when checking the min watermark. The min watermark is the
3784 * point where boosting is ignored so that kswapd is woken up
3785 * when below the low watermark.
3787 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3788 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3789 mark = z->_watermark[WMARK_MIN];
3790 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3791 alloc_flags, free_pages);
3797 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3798 unsigned long mark, int highest_zoneidx)
3800 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3802 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3803 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3805 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3810 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3812 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3813 node_reclaim_distance;
3815 #else /* CONFIG_NUMA */
3816 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3820 #endif /* CONFIG_NUMA */
3823 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3824 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3825 * premature use of a lower zone may cause lowmem pressure problems that
3826 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3827 * probably too small. It only makes sense to spread allocations to avoid
3828 * fragmentation between the Normal and DMA32 zones.
3830 static inline unsigned int
3831 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3833 unsigned int alloc_flags;
3836 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3839 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3841 #ifdef CONFIG_ZONE_DMA32
3845 if (zone_idx(zone) != ZONE_NORMAL)
3849 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3850 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3851 * on UMA that if Normal is populated then so is DMA32.
3853 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3854 if (nr_online_nodes > 1 && !populated_zone(--zone))
3857 alloc_flags |= ALLOC_NOFRAGMENT;
3858 #endif /* CONFIG_ZONE_DMA32 */
3862 /* Must be called after current_gfp_context() which can change gfp_mask */
3863 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3864 unsigned int alloc_flags)
3867 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3868 alloc_flags |= ALLOC_CMA;
3874 * get_page_from_freelist goes through the zonelist trying to allocate
3877 static struct page *
3878 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3879 const struct alloc_context *ac)
3883 struct pglist_data *last_pgdat_dirty_limit = NULL;
3888 * Scan zonelist, looking for a zone with enough free.
3889 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3891 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3892 z = ac->preferred_zoneref;
3893 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3898 if (cpusets_enabled() &&
3899 (alloc_flags & ALLOC_CPUSET) &&
3900 !__cpuset_zone_allowed(zone, gfp_mask))
3903 * When allocating a page cache page for writing, we
3904 * want to get it from a node that is within its dirty
3905 * limit, such that no single node holds more than its
3906 * proportional share of globally allowed dirty pages.
3907 * The dirty limits take into account the node's
3908 * lowmem reserves and high watermark so that kswapd
3909 * should be able to balance it without having to
3910 * write pages from its LRU list.
3912 * XXX: For now, allow allocations to potentially
3913 * exceed the per-node dirty limit in the slowpath
3914 * (spread_dirty_pages unset) before going into reclaim,
3915 * which is important when on a NUMA setup the allowed
3916 * nodes are together not big enough to reach the
3917 * global limit. The proper fix for these situations
3918 * will require awareness of nodes in the
3919 * dirty-throttling and the flusher threads.
3921 if (ac->spread_dirty_pages) {
3922 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3925 if (!node_dirty_ok(zone->zone_pgdat)) {
3926 last_pgdat_dirty_limit = zone->zone_pgdat;
3931 if (no_fallback && nr_online_nodes > 1 &&
3932 zone != ac->preferred_zoneref->zone) {
3936 * If moving to a remote node, retry but allow
3937 * fragmenting fallbacks. Locality is more important
3938 * than fragmentation avoidance.
3940 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3941 if (zone_to_nid(zone) != local_nid) {
3942 alloc_flags &= ~ALLOC_NOFRAGMENT;
3947 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3948 if (!zone_watermark_fast(zone, order, mark,
3949 ac->highest_zoneidx, alloc_flags,
3953 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3955 * Watermark failed for this zone, but see if we can
3956 * grow this zone if it contains deferred pages.
3958 if (static_branch_unlikely(&deferred_pages)) {
3959 if (_deferred_grow_zone(zone, order))
3963 /* Checked here to keep the fast path fast */
3964 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3965 if (alloc_flags & ALLOC_NO_WATERMARKS)
3968 if (!node_reclaim_enabled() ||
3969 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3972 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3974 case NODE_RECLAIM_NOSCAN:
3977 case NODE_RECLAIM_FULL:
3978 /* scanned but unreclaimable */
3981 /* did we reclaim enough */
3982 if (zone_watermark_ok(zone, order, mark,
3983 ac->highest_zoneidx, alloc_flags))
3991 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3992 gfp_mask, alloc_flags, ac->migratetype);
3994 prep_new_page(page, order, gfp_mask, alloc_flags);
3997 * If this is a high-order atomic allocation then check
3998 * if the pageblock should be reserved for the future
4000 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4001 reserve_highatomic_pageblock(page, zone, order);
4005 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4006 /* Try again if zone has deferred pages */
4007 if (static_branch_unlikely(&deferred_pages)) {
4008 if (_deferred_grow_zone(zone, order))
4016 * It's possible on a UMA machine to get through all zones that are
4017 * fragmented. If avoiding fragmentation, reset and try again.
4020 alloc_flags &= ~ALLOC_NOFRAGMENT;
4027 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4029 unsigned int filter = SHOW_MEM_FILTER_NODES;
4032 * This documents exceptions given to allocations in certain
4033 * contexts that are allowed to allocate outside current's set
4036 if (!(gfp_mask & __GFP_NOMEMALLOC))
4037 if (tsk_is_oom_victim(current) ||
4038 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4039 filter &= ~SHOW_MEM_FILTER_NODES;
4040 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4041 filter &= ~SHOW_MEM_FILTER_NODES;
4043 show_mem(filter, nodemask);
4046 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4048 struct va_format vaf;
4050 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4052 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4055 va_start(args, fmt);
4058 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4059 current->comm, &vaf, gfp_mask, &gfp_mask,
4060 nodemask_pr_args(nodemask));
4063 cpuset_print_current_mems_allowed();
4066 warn_alloc_show_mem(gfp_mask, nodemask);
4069 static inline struct page *
4070 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4071 unsigned int alloc_flags,
4072 const struct alloc_context *ac)
4076 page = get_page_from_freelist(gfp_mask, order,
4077 alloc_flags|ALLOC_CPUSET, ac);
4079 * fallback to ignore cpuset restriction if our nodes
4083 page = get_page_from_freelist(gfp_mask, order,
4089 static inline struct page *
4090 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4091 const struct alloc_context *ac, unsigned long *did_some_progress)
4093 struct oom_control oc = {
4094 .zonelist = ac->zonelist,
4095 .nodemask = ac->nodemask,
4097 .gfp_mask = gfp_mask,
4102 *did_some_progress = 0;
4105 * Acquire the oom lock. If that fails, somebody else is
4106 * making progress for us.
4108 if (!mutex_trylock(&oom_lock)) {
4109 *did_some_progress = 1;
4110 schedule_timeout_uninterruptible(1);
4115 * Go through the zonelist yet one more time, keep very high watermark
4116 * here, this is only to catch a parallel oom killing, we must fail if
4117 * we're still under heavy pressure. But make sure that this reclaim
4118 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4119 * allocation which will never fail due to oom_lock already held.
4121 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4122 ~__GFP_DIRECT_RECLAIM, order,
4123 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4127 /* Coredumps can quickly deplete all memory reserves */
4128 if (current->flags & PF_DUMPCORE)
4130 /* The OOM killer will not help higher order allocs */
4131 if (order > PAGE_ALLOC_COSTLY_ORDER)
4134 * We have already exhausted all our reclaim opportunities without any
4135 * success so it is time to admit defeat. We will skip the OOM killer
4136 * because it is very likely that the caller has a more reasonable
4137 * fallback than shooting a random task.
4139 * The OOM killer may not free memory on a specific node.
4141 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4143 /* The OOM killer does not needlessly kill tasks for lowmem */
4144 if (ac->highest_zoneidx < ZONE_NORMAL)
4146 if (pm_suspended_storage())
4149 * XXX: GFP_NOFS allocations should rather fail than rely on
4150 * other request to make a forward progress.
4151 * We are in an unfortunate situation where out_of_memory cannot
4152 * do much for this context but let's try it to at least get
4153 * access to memory reserved if the current task is killed (see
4154 * out_of_memory). Once filesystems are ready to handle allocation
4155 * failures more gracefully we should just bail out here.
4158 /* Exhausted what can be done so it's blame time */
4159 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4160 *did_some_progress = 1;
4163 * Help non-failing allocations by giving them access to memory
4166 if (gfp_mask & __GFP_NOFAIL)
4167 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4168 ALLOC_NO_WATERMARKS, ac);
4171 mutex_unlock(&oom_lock);
4176 * Maximum number of compaction retries with a progress before OOM
4177 * killer is consider as the only way to move forward.
4179 #define MAX_COMPACT_RETRIES 16
4181 #ifdef CONFIG_COMPACTION
4182 /* Try memory compaction for high-order allocations before reclaim */
4183 static struct page *
4184 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4185 unsigned int alloc_flags, const struct alloc_context *ac,
4186 enum compact_priority prio, enum compact_result *compact_result)
4188 struct page *page = NULL;
4189 unsigned long pflags;
4190 unsigned int noreclaim_flag;
4195 psi_memstall_enter(&pflags);
4196 noreclaim_flag = memalloc_noreclaim_save();
4198 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4201 memalloc_noreclaim_restore(noreclaim_flag);
4202 psi_memstall_leave(&pflags);
4204 if (*compact_result == COMPACT_SKIPPED)
4207 * At least in one zone compaction wasn't deferred or skipped, so let's
4208 * count a compaction stall
4210 count_vm_event(COMPACTSTALL);
4212 /* Prep a captured page if available */
4214 prep_new_page(page, order, gfp_mask, alloc_flags);
4216 /* Try get a page from the freelist if available */
4218 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4221 struct zone *zone = page_zone(page);
4223 zone->compact_blockskip_flush = false;
4224 compaction_defer_reset(zone, order, true);
4225 count_vm_event(COMPACTSUCCESS);
4230 * It's bad if compaction run occurs and fails. The most likely reason
4231 * is that pages exist, but not enough to satisfy watermarks.
4233 count_vm_event(COMPACTFAIL);
4241 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4242 enum compact_result compact_result,
4243 enum compact_priority *compact_priority,
4244 int *compaction_retries)
4246 int max_retries = MAX_COMPACT_RETRIES;
4249 int retries = *compaction_retries;
4250 enum compact_priority priority = *compact_priority;
4255 if (fatal_signal_pending(current))
4258 if (compaction_made_progress(compact_result))
4259 (*compaction_retries)++;
4262 * compaction considers all the zone as desperately out of memory
4263 * so it doesn't really make much sense to retry except when the
4264 * failure could be caused by insufficient priority
4266 if (compaction_failed(compact_result))
4267 goto check_priority;
4270 * compaction was skipped because there are not enough order-0 pages
4271 * to work with, so we retry only if it looks like reclaim can help.
4273 if (compaction_needs_reclaim(compact_result)) {
4274 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4279 * make sure the compaction wasn't deferred or didn't bail out early
4280 * due to locks contention before we declare that we should give up.
4281 * But the next retry should use a higher priority if allowed, so
4282 * we don't just keep bailing out endlessly.
4284 if (compaction_withdrawn(compact_result)) {
4285 goto check_priority;
4289 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4290 * costly ones because they are de facto nofail and invoke OOM
4291 * killer to move on while costly can fail and users are ready
4292 * to cope with that. 1/4 retries is rather arbitrary but we
4293 * would need much more detailed feedback from compaction to
4294 * make a better decision.
4296 if (order > PAGE_ALLOC_COSTLY_ORDER)
4298 if (*compaction_retries <= max_retries) {
4304 * Make sure there are attempts at the highest priority if we exhausted
4305 * all retries or failed at the lower priorities.
4308 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4309 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4311 if (*compact_priority > min_priority) {
4312 (*compact_priority)--;
4313 *compaction_retries = 0;
4317 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4321 static inline struct page *
4322 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4323 unsigned int alloc_flags, const struct alloc_context *ac,
4324 enum compact_priority prio, enum compact_result *compact_result)
4326 *compact_result = COMPACT_SKIPPED;
4331 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4332 enum compact_result compact_result,
4333 enum compact_priority *compact_priority,
4334 int *compaction_retries)
4339 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4343 * There are setups with compaction disabled which would prefer to loop
4344 * inside the allocator rather than hit the oom killer prematurely.
4345 * Let's give them a good hope and keep retrying while the order-0
4346 * watermarks are OK.
4348 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4349 ac->highest_zoneidx, ac->nodemask) {
4350 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4351 ac->highest_zoneidx, alloc_flags))
4356 #endif /* CONFIG_COMPACTION */
4358 #ifdef CONFIG_LOCKDEP
4359 static struct lockdep_map __fs_reclaim_map =
4360 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4362 static bool __need_reclaim(gfp_t gfp_mask)
4364 /* no reclaim without waiting on it */
4365 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4368 /* this guy won't enter reclaim */
4369 if (current->flags & PF_MEMALLOC)
4372 if (gfp_mask & __GFP_NOLOCKDEP)
4378 void __fs_reclaim_acquire(void)
4380 lock_map_acquire(&__fs_reclaim_map);
4383 void __fs_reclaim_release(void)
4385 lock_map_release(&__fs_reclaim_map);
4388 void fs_reclaim_acquire(gfp_t gfp_mask)
4390 gfp_mask = current_gfp_context(gfp_mask);
4392 if (__need_reclaim(gfp_mask)) {
4393 if (gfp_mask & __GFP_FS)
4394 __fs_reclaim_acquire();
4396 #ifdef CONFIG_MMU_NOTIFIER
4397 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4398 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4403 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4405 void fs_reclaim_release(gfp_t gfp_mask)
4407 gfp_mask = current_gfp_context(gfp_mask);
4409 if (__need_reclaim(gfp_mask)) {
4410 if (gfp_mask & __GFP_FS)
4411 __fs_reclaim_release();
4414 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4417 /* Perform direct synchronous page reclaim */
4418 static unsigned long
4419 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4420 const struct alloc_context *ac)
4422 unsigned int noreclaim_flag;
4423 unsigned long pflags, progress;
4427 /* We now go into synchronous reclaim */
4428 cpuset_memory_pressure_bump();
4429 psi_memstall_enter(&pflags);
4430 fs_reclaim_acquire(gfp_mask);
4431 noreclaim_flag = memalloc_noreclaim_save();
4433 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4436 memalloc_noreclaim_restore(noreclaim_flag);
4437 fs_reclaim_release(gfp_mask);
4438 psi_memstall_leave(&pflags);
4445 /* The really slow allocator path where we enter direct reclaim */
4446 static inline struct page *
4447 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4448 unsigned int alloc_flags, const struct alloc_context *ac,
4449 unsigned long *did_some_progress)
4451 struct page *page = NULL;
4452 bool drained = false;
4454 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4455 if (unlikely(!(*did_some_progress)))
4459 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4462 * If an allocation failed after direct reclaim, it could be because
4463 * pages are pinned on the per-cpu lists or in high alloc reserves.
4464 * Shrink them and try again
4466 if (!page && !drained) {
4467 unreserve_highatomic_pageblock(ac, false);
4468 drain_all_pages(NULL);
4476 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4477 const struct alloc_context *ac)
4481 pg_data_t *last_pgdat = NULL;
4482 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4484 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4486 if (last_pgdat != zone->zone_pgdat)
4487 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4488 last_pgdat = zone->zone_pgdat;
4492 static inline unsigned int
4493 gfp_to_alloc_flags(gfp_t gfp_mask)
4495 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4498 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4499 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4500 * to save two branches.
4502 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4503 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4506 * The caller may dip into page reserves a bit more if the caller
4507 * cannot run direct reclaim, or if the caller has realtime scheduling
4508 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4509 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4511 alloc_flags |= (__force int)
4512 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4514 if (gfp_mask & __GFP_ATOMIC) {
4516 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4517 * if it can't schedule.
4519 if (!(gfp_mask & __GFP_NOMEMALLOC))
4520 alloc_flags |= ALLOC_HARDER;
4522 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4523 * comment for __cpuset_node_allowed().
4525 alloc_flags &= ~ALLOC_CPUSET;
4526 } else if (unlikely(rt_task(current)) && !in_interrupt())
4527 alloc_flags |= ALLOC_HARDER;
4529 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4534 static bool oom_reserves_allowed(struct task_struct *tsk)
4536 if (!tsk_is_oom_victim(tsk))
4540 * !MMU doesn't have oom reaper so give access to memory reserves
4541 * only to the thread with TIF_MEMDIE set
4543 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4550 * Distinguish requests which really need access to full memory
4551 * reserves from oom victims which can live with a portion of it
4553 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4555 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4557 if (gfp_mask & __GFP_MEMALLOC)
4558 return ALLOC_NO_WATERMARKS;
4559 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4560 return ALLOC_NO_WATERMARKS;
4561 if (!in_interrupt()) {
4562 if (current->flags & PF_MEMALLOC)
4563 return ALLOC_NO_WATERMARKS;
4564 else if (oom_reserves_allowed(current))
4571 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4573 return !!__gfp_pfmemalloc_flags(gfp_mask);
4577 * Checks whether it makes sense to retry the reclaim to make a forward progress
4578 * for the given allocation request.
4580 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4581 * without success, or when we couldn't even meet the watermark if we
4582 * reclaimed all remaining pages on the LRU lists.
4584 * Returns true if a retry is viable or false to enter the oom path.
4587 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4588 struct alloc_context *ac, int alloc_flags,
4589 bool did_some_progress, int *no_progress_loops)
4596 * Costly allocations might have made a progress but this doesn't mean
4597 * their order will become available due to high fragmentation so
4598 * always increment the no progress counter for them
4600 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4601 *no_progress_loops = 0;
4603 (*no_progress_loops)++;
4606 * Make sure we converge to OOM if we cannot make any progress
4607 * several times in the row.
4609 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4610 /* Before OOM, exhaust highatomic_reserve */
4611 return unreserve_highatomic_pageblock(ac, true);
4615 * Keep reclaiming pages while there is a chance this will lead
4616 * somewhere. If none of the target zones can satisfy our allocation
4617 * request even if all reclaimable pages are considered then we are
4618 * screwed and have to go OOM.
4620 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4621 ac->highest_zoneidx, ac->nodemask) {
4622 unsigned long available;
4623 unsigned long reclaimable;
4624 unsigned long min_wmark = min_wmark_pages(zone);
4627 available = reclaimable = zone_reclaimable_pages(zone);
4628 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4631 * Would the allocation succeed if we reclaimed all
4632 * reclaimable pages?
4634 wmark = __zone_watermark_ok(zone, order, min_wmark,
4635 ac->highest_zoneidx, alloc_flags, available);
4636 trace_reclaim_retry_zone(z, order, reclaimable,
4637 available, min_wmark, *no_progress_loops, wmark);
4640 * If we didn't make any progress and have a lot of
4641 * dirty + writeback pages then we should wait for
4642 * an IO to complete to slow down the reclaim and
4643 * prevent from pre mature OOM
4645 if (!did_some_progress) {
4646 unsigned long write_pending;
4648 write_pending = zone_page_state_snapshot(zone,
4649 NR_ZONE_WRITE_PENDING);
4651 if (2 * write_pending > reclaimable) {
4652 congestion_wait(BLK_RW_ASYNC, HZ/10);
4664 * Memory allocation/reclaim might be called from a WQ context and the
4665 * current implementation of the WQ concurrency control doesn't
4666 * recognize that a particular WQ is congested if the worker thread is
4667 * looping without ever sleeping. Therefore we have to do a short sleep
4668 * here rather than calling cond_resched().
4670 if (current->flags & PF_WQ_WORKER)
4671 schedule_timeout_uninterruptible(1);
4678 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4681 * It's possible that cpuset's mems_allowed and the nodemask from
4682 * mempolicy don't intersect. This should be normally dealt with by
4683 * policy_nodemask(), but it's possible to race with cpuset update in
4684 * such a way the check therein was true, and then it became false
4685 * before we got our cpuset_mems_cookie here.
4686 * This assumes that for all allocations, ac->nodemask can come only
4687 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4688 * when it does not intersect with the cpuset restrictions) or the
4689 * caller can deal with a violated nodemask.
4691 if (cpusets_enabled() && ac->nodemask &&
4692 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4693 ac->nodemask = NULL;
4698 * When updating a task's mems_allowed or mempolicy nodemask, it is
4699 * possible to race with parallel threads in such a way that our
4700 * allocation can fail while the mask is being updated. If we are about
4701 * to fail, check if the cpuset changed during allocation and if so,
4704 if (read_mems_allowed_retry(cpuset_mems_cookie))
4710 static inline struct page *
4711 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4712 struct alloc_context *ac)
4714 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4715 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4716 struct page *page = NULL;
4717 unsigned int alloc_flags;
4718 unsigned long did_some_progress;
4719 enum compact_priority compact_priority;
4720 enum compact_result compact_result;
4721 int compaction_retries;
4722 int no_progress_loops;
4723 unsigned int cpuset_mems_cookie;
4727 * We also sanity check to catch abuse of atomic reserves being used by
4728 * callers that are not in atomic context.
4730 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4731 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4732 gfp_mask &= ~__GFP_ATOMIC;
4735 compaction_retries = 0;
4736 no_progress_loops = 0;
4737 compact_priority = DEF_COMPACT_PRIORITY;
4738 cpuset_mems_cookie = read_mems_allowed_begin();
4741 * The fast path uses conservative alloc_flags to succeed only until
4742 * kswapd needs to be woken up, and to avoid the cost of setting up
4743 * alloc_flags precisely. So we do that now.
4745 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4748 * We need to recalculate the starting point for the zonelist iterator
4749 * because we might have used different nodemask in the fast path, or
4750 * there was a cpuset modification and we are retrying - otherwise we
4751 * could end up iterating over non-eligible zones endlessly.
4753 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4754 ac->highest_zoneidx, ac->nodemask);
4755 if (!ac->preferred_zoneref->zone)
4758 if (alloc_flags & ALLOC_KSWAPD)
4759 wake_all_kswapds(order, gfp_mask, ac);
4762 * The adjusted alloc_flags might result in immediate success, so try
4765 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4770 * For costly allocations, try direct compaction first, as it's likely
4771 * that we have enough base pages and don't need to reclaim. For non-
4772 * movable high-order allocations, do that as well, as compaction will
4773 * try prevent permanent fragmentation by migrating from blocks of the
4775 * Don't try this for allocations that are allowed to ignore
4776 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4778 if (can_direct_reclaim &&
4780 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4781 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4782 page = __alloc_pages_direct_compact(gfp_mask, order,
4784 INIT_COMPACT_PRIORITY,
4790 * Checks for costly allocations with __GFP_NORETRY, which
4791 * includes some THP page fault allocations
4793 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4795 * If allocating entire pageblock(s) and compaction
4796 * failed because all zones are below low watermarks
4797 * or is prohibited because it recently failed at this
4798 * order, fail immediately unless the allocator has
4799 * requested compaction and reclaim retry.
4802 * - potentially very expensive because zones are far
4803 * below their low watermarks or this is part of very
4804 * bursty high order allocations,
4805 * - not guaranteed to help because isolate_freepages()
4806 * may not iterate over freed pages as part of its
4808 * - unlikely to make entire pageblocks free on its
4811 if (compact_result == COMPACT_SKIPPED ||
4812 compact_result == COMPACT_DEFERRED)
4816 * Looks like reclaim/compaction is worth trying, but
4817 * sync compaction could be very expensive, so keep
4818 * using async compaction.
4820 compact_priority = INIT_COMPACT_PRIORITY;
4825 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4826 if (alloc_flags & ALLOC_KSWAPD)
4827 wake_all_kswapds(order, gfp_mask, ac);
4829 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4831 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
4834 * Reset the nodemask and zonelist iterators if memory policies can be
4835 * ignored. These allocations are high priority and system rather than
4838 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4839 ac->nodemask = NULL;
4840 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4841 ac->highest_zoneidx, ac->nodemask);
4844 /* Attempt with potentially adjusted zonelist and alloc_flags */
4845 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4849 /* Caller is not willing to reclaim, we can't balance anything */
4850 if (!can_direct_reclaim)
4853 /* Avoid recursion of direct reclaim */
4854 if (current->flags & PF_MEMALLOC)
4857 /* Try direct reclaim and then allocating */
4858 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4859 &did_some_progress);
4863 /* Try direct compaction and then allocating */
4864 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4865 compact_priority, &compact_result);
4869 /* Do not loop if specifically requested */
4870 if (gfp_mask & __GFP_NORETRY)
4874 * Do not retry costly high order allocations unless they are
4875 * __GFP_RETRY_MAYFAIL
4877 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4880 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4881 did_some_progress > 0, &no_progress_loops))
4885 * It doesn't make any sense to retry for the compaction if the order-0
4886 * reclaim is not able to make any progress because the current
4887 * implementation of the compaction depends on the sufficient amount
4888 * of free memory (see __compaction_suitable)
4890 if (did_some_progress > 0 &&
4891 should_compact_retry(ac, order, alloc_flags,
4892 compact_result, &compact_priority,
4893 &compaction_retries))
4897 /* Deal with possible cpuset update races before we start OOM killing */
4898 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4901 /* Reclaim has failed us, start killing things */
4902 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4906 /* Avoid allocations with no watermarks from looping endlessly */
4907 if (tsk_is_oom_victim(current) &&
4908 (alloc_flags & ALLOC_OOM ||
4909 (gfp_mask & __GFP_NOMEMALLOC)))
4912 /* Retry as long as the OOM killer is making progress */
4913 if (did_some_progress) {
4914 no_progress_loops = 0;
4919 /* Deal with possible cpuset update races before we fail */
4920 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4924 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4927 if (gfp_mask & __GFP_NOFAIL) {
4929 * All existing users of the __GFP_NOFAIL are blockable, so warn
4930 * of any new users that actually require GFP_NOWAIT
4932 if (WARN_ON_ONCE(!can_direct_reclaim))
4936 * PF_MEMALLOC request from this context is rather bizarre
4937 * because we cannot reclaim anything and only can loop waiting
4938 * for somebody to do a work for us
4940 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4943 * non failing costly orders are a hard requirement which we
4944 * are not prepared for much so let's warn about these users
4945 * so that we can identify them and convert them to something
4948 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4951 * Help non-failing allocations by giving them access to memory
4952 * reserves but do not use ALLOC_NO_WATERMARKS because this
4953 * could deplete whole memory reserves which would just make
4954 * the situation worse
4956 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4964 warn_alloc(gfp_mask, ac->nodemask,
4965 "page allocation failure: order:%u", order);
4970 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4971 int preferred_nid, nodemask_t *nodemask,
4972 struct alloc_context *ac, gfp_t *alloc_gfp,
4973 unsigned int *alloc_flags)
4975 ac->highest_zoneidx = gfp_zone(gfp_mask);
4976 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4977 ac->nodemask = nodemask;
4978 ac->migratetype = gfp_migratetype(gfp_mask);
4980 if (cpusets_enabled()) {
4981 *alloc_gfp |= __GFP_HARDWALL;
4983 * When we are in the interrupt context, it is irrelevant
4984 * to the current task context. It means that any node ok.
4986 if (!in_interrupt() && !ac->nodemask)
4987 ac->nodemask = &cpuset_current_mems_allowed;
4989 *alloc_flags |= ALLOC_CPUSET;
4992 fs_reclaim_acquire(gfp_mask);
4993 fs_reclaim_release(gfp_mask);
4995 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4997 if (should_fail_alloc_page(gfp_mask, order))
5000 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5002 /* Dirty zone balancing only done in the fast path */
5003 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5006 * The preferred zone is used for statistics but crucially it is
5007 * also used as the starting point for the zonelist iterator. It
5008 * may get reset for allocations that ignore memory policies.
5010 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5011 ac->highest_zoneidx, ac->nodemask);
5017 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5018 * @gfp: GFP flags for the allocation
5019 * @preferred_nid: The preferred NUMA node ID to allocate from
5020 * @nodemask: Set of nodes to allocate from, may be NULL
5021 * @nr_pages: The number of pages desired on the list or array
5022 * @page_list: Optional list to store the allocated pages
5023 * @page_array: Optional array to store the pages
5025 * This is a batched version of the page allocator that attempts to
5026 * allocate nr_pages quickly. Pages are added to page_list if page_list
5027 * is not NULL, otherwise it is assumed that the page_array is valid.
5029 * For lists, nr_pages is the number of pages that should be allocated.
5031 * For arrays, only NULL elements are populated with pages and nr_pages
5032 * is the maximum number of pages that will be stored in the array.
5034 * Returns the number of pages on the list or array.
5036 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5037 nodemask_t *nodemask, int nr_pages,
5038 struct list_head *page_list,
5039 struct page **page_array)
5042 unsigned long flags;
5045 struct per_cpu_pages *pcp;
5046 struct list_head *pcp_list;
5047 struct alloc_context ac;
5049 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5050 int nr_populated = 0;
5052 if (unlikely(nr_pages <= 0))
5056 * Skip populated array elements to determine if any pages need
5057 * to be allocated before disabling IRQs.
5059 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5062 /* Already populated array? */
5063 if (unlikely(page_array && nr_pages - nr_populated == 0))
5064 return nr_populated;
5066 /* Use the single page allocator for one page. */
5067 if (nr_pages - nr_populated == 1)
5070 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5071 gfp &= gfp_allowed_mask;
5073 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5077 /* Find an allowed local zone that meets the low watermark. */
5078 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5081 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5082 !__cpuset_zone_allowed(zone, gfp)) {
5086 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5087 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5091 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5092 if (zone_watermark_fast(zone, 0, mark,
5093 zonelist_zone_idx(ac.preferred_zoneref),
5094 alloc_flags, gfp)) {
5100 * If there are no allowed local zones that meets the watermarks then
5101 * try to allocate a single page and reclaim if necessary.
5103 if (unlikely(!zone))
5106 /* Attempt the batch allocation */
5107 local_irq_save(flags);
5108 pcp = &this_cpu_ptr(zone->pageset)->pcp;
5109 pcp_list = &pcp->lists[ac.migratetype];
5111 while (nr_populated < nr_pages) {
5113 /* Skip existing pages */
5114 if (page_array && page_array[nr_populated]) {
5119 page = __rmqueue_pcplist(zone, ac.migratetype, alloc_flags,
5121 if (unlikely(!page)) {
5122 /* Try and get at least one page */
5129 * Ideally this would be batched but the best way to do
5130 * that cheaply is to first convert zone_statistics to
5131 * be inaccurate per-cpu counter like vm_events to avoid
5132 * a RMW cycle then do the accounting with IRQs enabled.
5134 __count_zid_vm_events(PGALLOC, zone_idx(zone), 1);
5135 zone_statistics(ac.preferred_zoneref->zone, zone);
5137 prep_new_page(page, 0, gfp, 0);
5139 list_add(&page->lru, page_list);
5141 page_array[nr_populated] = page;
5145 local_irq_restore(flags);
5147 return nr_populated;
5150 local_irq_restore(flags);
5153 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5156 list_add(&page->lru, page_list);
5158 page_array[nr_populated] = page;
5162 return nr_populated;
5164 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5167 * This is the 'heart' of the zoned buddy allocator.
5169 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5170 nodemask_t *nodemask)
5173 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5174 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5175 struct alloc_context ac = { };
5178 * There are several places where we assume that the order value is sane
5179 * so bail out early if the request is out of bound.
5181 if (unlikely(order >= MAX_ORDER)) {
5182 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5186 gfp &= gfp_allowed_mask;
5188 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5189 * resp. GFP_NOIO which has to be inherited for all allocation requests
5190 * from a particular context which has been marked by
5191 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5192 * movable zones are not used during allocation.
5194 gfp = current_gfp_context(gfp);
5196 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5197 &alloc_gfp, &alloc_flags))
5201 * Forbid the first pass from falling back to types that fragment
5202 * memory until all local zones are considered.
5204 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5206 /* First allocation attempt */
5207 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5212 ac.spread_dirty_pages = false;
5215 * Restore the original nodemask if it was potentially replaced with
5216 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5218 ac.nodemask = nodemask;
5220 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5223 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5224 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5225 __free_pages(page, order);
5229 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5233 EXPORT_SYMBOL(__alloc_pages);
5236 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5237 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5238 * you need to access high mem.
5240 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5244 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5247 return (unsigned long) page_address(page);
5249 EXPORT_SYMBOL(__get_free_pages);
5251 unsigned long get_zeroed_page(gfp_t gfp_mask)
5253 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5255 EXPORT_SYMBOL(get_zeroed_page);
5257 static inline void free_the_page(struct page *page, unsigned int order)
5259 if (order == 0) /* Via pcp? */
5260 free_unref_page(page);
5262 __free_pages_ok(page, order, FPI_NONE);
5266 * __free_pages - Free pages allocated with alloc_pages().
5267 * @page: The page pointer returned from alloc_pages().
5268 * @order: The order of the allocation.
5270 * This function can free multi-page allocations that are not compound
5271 * pages. It does not check that the @order passed in matches that of
5272 * the allocation, so it is easy to leak memory. Freeing more memory
5273 * than was allocated will probably emit a warning.
5275 * If the last reference to this page is speculative, it will be released
5276 * by put_page() which only frees the first page of a non-compound
5277 * allocation. To prevent the remaining pages from being leaked, we free
5278 * the subsequent pages here. If you want to use the page's reference
5279 * count to decide when to free the allocation, you should allocate a
5280 * compound page, and use put_page() instead of __free_pages().
5282 * Context: May be called in interrupt context or while holding a normal
5283 * spinlock, but not in NMI context or while holding a raw spinlock.
5285 void __free_pages(struct page *page, unsigned int order)
5287 if (put_page_testzero(page))
5288 free_the_page(page, order);
5289 else if (!PageHead(page))
5291 free_the_page(page + (1 << order), order);
5293 EXPORT_SYMBOL(__free_pages);
5295 void free_pages(unsigned long addr, unsigned int order)
5298 VM_BUG_ON(!virt_addr_valid((void *)addr));
5299 __free_pages(virt_to_page((void *)addr), order);
5303 EXPORT_SYMBOL(free_pages);
5307 * An arbitrary-length arbitrary-offset area of memory which resides
5308 * within a 0 or higher order page. Multiple fragments within that page
5309 * are individually refcounted, in the page's reference counter.
5311 * The page_frag functions below provide a simple allocation framework for
5312 * page fragments. This is used by the network stack and network device
5313 * drivers to provide a backing region of memory for use as either an
5314 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5316 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5319 struct page *page = NULL;
5320 gfp_t gfp = gfp_mask;
5322 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5323 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5325 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5326 PAGE_FRAG_CACHE_MAX_ORDER);
5327 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5329 if (unlikely(!page))
5330 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5332 nc->va = page ? page_address(page) : NULL;
5337 void __page_frag_cache_drain(struct page *page, unsigned int count)
5339 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5341 if (page_ref_sub_and_test(page, count))
5342 free_the_page(page, compound_order(page));
5344 EXPORT_SYMBOL(__page_frag_cache_drain);
5346 void *page_frag_alloc_align(struct page_frag_cache *nc,
5347 unsigned int fragsz, gfp_t gfp_mask,
5348 unsigned int align_mask)
5350 unsigned int size = PAGE_SIZE;
5354 if (unlikely(!nc->va)) {
5356 page = __page_frag_cache_refill(nc, gfp_mask);
5360 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5361 /* if size can vary use size else just use PAGE_SIZE */
5364 /* Even if we own the page, we do not use atomic_set().
5365 * This would break get_page_unless_zero() users.
5367 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5369 /* reset page count bias and offset to start of new frag */
5370 nc->pfmemalloc = page_is_pfmemalloc(page);
5371 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5375 offset = nc->offset - fragsz;
5376 if (unlikely(offset < 0)) {
5377 page = virt_to_page(nc->va);
5379 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5382 if (unlikely(nc->pfmemalloc)) {
5383 free_the_page(page, compound_order(page));
5387 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5388 /* if size can vary use size else just use PAGE_SIZE */
5391 /* OK, page count is 0, we can safely set it */
5392 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5394 /* reset page count bias and offset to start of new frag */
5395 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5396 offset = size - fragsz;
5400 offset &= align_mask;
5401 nc->offset = offset;
5403 return nc->va + offset;
5405 EXPORT_SYMBOL(page_frag_alloc_align);
5408 * Frees a page fragment allocated out of either a compound or order 0 page.
5410 void page_frag_free(void *addr)
5412 struct page *page = virt_to_head_page(addr);
5414 if (unlikely(put_page_testzero(page)))
5415 free_the_page(page, compound_order(page));
5417 EXPORT_SYMBOL(page_frag_free);
5419 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5423 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5424 unsigned long used = addr + PAGE_ALIGN(size);
5426 split_page(virt_to_page((void *)addr), order);
5427 while (used < alloc_end) {
5432 return (void *)addr;
5436 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5437 * @size: the number of bytes to allocate
5438 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5440 * This function is similar to alloc_pages(), except that it allocates the
5441 * minimum number of pages to satisfy the request. alloc_pages() can only
5442 * allocate memory in power-of-two pages.
5444 * This function is also limited by MAX_ORDER.
5446 * Memory allocated by this function must be released by free_pages_exact().
5448 * Return: pointer to the allocated area or %NULL in case of error.
5450 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5452 unsigned int order = get_order(size);
5455 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5456 gfp_mask &= ~__GFP_COMP;
5458 addr = __get_free_pages(gfp_mask, order);
5459 return make_alloc_exact(addr, order, size);
5461 EXPORT_SYMBOL(alloc_pages_exact);
5464 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5466 * @nid: the preferred node ID where memory should be allocated
5467 * @size: the number of bytes to allocate
5468 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5470 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5473 * Return: pointer to the allocated area or %NULL in case of error.
5475 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5477 unsigned int order = get_order(size);
5480 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5481 gfp_mask &= ~__GFP_COMP;
5483 p = alloc_pages_node(nid, gfp_mask, order);
5486 return make_alloc_exact((unsigned long)page_address(p), order, size);
5490 * free_pages_exact - release memory allocated via alloc_pages_exact()
5491 * @virt: the value returned by alloc_pages_exact.
5492 * @size: size of allocation, same value as passed to alloc_pages_exact().
5494 * Release the memory allocated by a previous call to alloc_pages_exact.
5496 void free_pages_exact(void *virt, size_t size)
5498 unsigned long addr = (unsigned long)virt;
5499 unsigned long end = addr + PAGE_ALIGN(size);
5501 while (addr < end) {
5506 EXPORT_SYMBOL(free_pages_exact);
5509 * nr_free_zone_pages - count number of pages beyond high watermark
5510 * @offset: The zone index of the highest zone
5512 * nr_free_zone_pages() counts the number of pages which are beyond the
5513 * high watermark within all zones at or below a given zone index. For each
5514 * zone, the number of pages is calculated as:
5516 * nr_free_zone_pages = managed_pages - high_pages
5518 * Return: number of pages beyond high watermark.
5520 static unsigned long nr_free_zone_pages(int offset)
5525 /* Just pick one node, since fallback list is circular */
5526 unsigned long sum = 0;
5528 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5530 for_each_zone_zonelist(zone, z, zonelist, offset) {
5531 unsigned long size = zone_managed_pages(zone);
5532 unsigned long high = high_wmark_pages(zone);
5541 * nr_free_buffer_pages - count number of pages beyond high watermark
5543 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5544 * watermark within ZONE_DMA and ZONE_NORMAL.
5546 * Return: number of pages beyond high watermark within ZONE_DMA and
5549 unsigned long nr_free_buffer_pages(void)
5551 return nr_free_zone_pages(gfp_zone(GFP_USER));
5553 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5555 static inline void show_node(struct zone *zone)
5557 if (IS_ENABLED(CONFIG_NUMA))
5558 printk("Node %d ", zone_to_nid(zone));
5561 long si_mem_available(void)
5564 unsigned long pagecache;
5565 unsigned long wmark_low = 0;
5566 unsigned long pages[NR_LRU_LISTS];
5567 unsigned long reclaimable;
5571 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5572 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5575 wmark_low += low_wmark_pages(zone);
5578 * Estimate the amount of memory available for userspace allocations,
5579 * without causing swapping.
5581 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5584 * Not all the page cache can be freed, otherwise the system will
5585 * start swapping. Assume at least half of the page cache, or the
5586 * low watermark worth of cache, needs to stay.
5588 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5589 pagecache -= min(pagecache / 2, wmark_low);
5590 available += pagecache;
5593 * Part of the reclaimable slab and other kernel memory consists of
5594 * items that are in use, and cannot be freed. Cap this estimate at the
5597 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5598 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5599 available += reclaimable - min(reclaimable / 2, wmark_low);
5605 EXPORT_SYMBOL_GPL(si_mem_available);
5607 void si_meminfo(struct sysinfo *val)
5609 val->totalram = totalram_pages();
5610 val->sharedram = global_node_page_state(NR_SHMEM);
5611 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5612 val->bufferram = nr_blockdev_pages();
5613 val->totalhigh = totalhigh_pages();
5614 val->freehigh = nr_free_highpages();
5615 val->mem_unit = PAGE_SIZE;
5618 EXPORT_SYMBOL(si_meminfo);
5621 void si_meminfo_node(struct sysinfo *val, int nid)
5623 int zone_type; /* needs to be signed */
5624 unsigned long managed_pages = 0;
5625 unsigned long managed_highpages = 0;
5626 unsigned long free_highpages = 0;
5627 pg_data_t *pgdat = NODE_DATA(nid);
5629 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5630 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5631 val->totalram = managed_pages;
5632 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5633 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5634 #ifdef CONFIG_HIGHMEM
5635 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5636 struct zone *zone = &pgdat->node_zones[zone_type];
5638 if (is_highmem(zone)) {
5639 managed_highpages += zone_managed_pages(zone);
5640 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5643 val->totalhigh = managed_highpages;
5644 val->freehigh = free_highpages;
5646 val->totalhigh = managed_highpages;
5647 val->freehigh = free_highpages;
5649 val->mem_unit = PAGE_SIZE;
5654 * Determine whether the node should be displayed or not, depending on whether
5655 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5657 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5659 if (!(flags & SHOW_MEM_FILTER_NODES))
5663 * no node mask - aka implicit memory numa policy. Do not bother with
5664 * the synchronization - read_mems_allowed_begin - because we do not
5665 * have to be precise here.
5668 nodemask = &cpuset_current_mems_allowed;
5670 return !node_isset(nid, *nodemask);
5673 #define K(x) ((x) << (PAGE_SHIFT-10))
5675 static void show_migration_types(unsigned char type)
5677 static const char types[MIGRATE_TYPES] = {
5678 [MIGRATE_UNMOVABLE] = 'U',
5679 [MIGRATE_MOVABLE] = 'M',
5680 [MIGRATE_RECLAIMABLE] = 'E',
5681 [MIGRATE_HIGHATOMIC] = 'H',
5683 [MIGRATE_CMA] = 'C',
5685 #ifdef CONFIG_MEMORY_ISOLATION
5686 [MIGRATE_ISOLATE] = 'I',
5689 char tmp[MIGRATE_TYPES + 1];
5693 for (i = 0; i < MIGRATE_TYPES; i++) {
5694 if (type & (1 << i))
5699 printk(KERN_CONT "(%s) ", tmp);
5703 * Show free area list (used inside shift_scroll-lock stuff)
5704 * We also calculate the percentage fragmentation. We do this by counting the
5705 * memory on each free list with the exception of the first item on the list.
5708 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5711 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5713 unsigned long free_pcp = 0;
5718 for_each_populated_zone(zone) {
5719 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5722 for_each_online_cpu(cpu)
5723 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5726 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5727 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5728 " unevictable:%lu dirty:%lu writeback:%lu\n"
5729 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5730 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5731 " free:%lu free_pcp:%lu free_cma:%lu\n",
5732 global_node_page_state(NR_ACTIVE_ANON),
5733 global_node_page_state(NR_INACTIVE_ANON),
5734 global_node_page_state(NR_ISOLATED_ANON),
5735 global_node_page_state(NR_ACTIVE_FILE),
5736 global_node_page_state(NR_INACTIVE_FILE),
5737 global_node_page_state(NR_ISOLATED_FILE),
5738 global_node_page_state(NR_UNEVICTABLE),
5739 global_node_page_state(NR_FILE_DIRTY),
5740 global_node_page_state(NR_WRITEBACK),
5741 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5742 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5743 global_node_page_state(NR_FILE_MAPPED),
5744 global_node_page_state(NR_SHMEM),
5745 global_node_page_state(NR_PAGETABLE),
5746 global_zone_page_state(NR_BOUNCE),
5747 global_zone_page_state(NR_FREE_PAGES),
5749 global_zone_page_state(NR_FREE_CMA_PAGES));
5751 for_each_online_pgdat(pgdat) {
5752 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5756 " active_anon:%lukB"
5757 " inactive_anon:%lukB"
5758 " active_file:%lukB"
5759 " inactive_file:%lukB"
5760 " unevictable:%lukB"
5761 " isolated(anon):%lukB"
5762 " isolated(file):%lukB"
5767 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5769 " shmem_pmdmapped: %lukB"
5772 " writeback_tmp:%lukB"
5773 " kernel_stack:%lukB"
5774 #ifdef CONFIG_SHADOW_CALL_STACK
5775 " shadow_call_stack:%lukB"
5778 " all_unreclaimable? %s"
5781 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5782 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5783 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5784 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5785 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5786 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5787 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5788 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5789 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5790 K(node_page_state(pgdat, NR_WRITEBACK)),
5791 K(node_page_state(pgdat, NR_SHMEM)),
5792 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5793 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5794 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5795 K(node_page_state(pgdat, NR_ANON_THPS)),
5797 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5798 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5799 #ifdef CONFIG_SHADOW_CALL_STACK
5800 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5802 K(node_page_state(pgdat, NR_PAGETABLE)),
5803 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5807 for_each_populated_zone(zone) {
5810 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5814 for_each_online_cpu(cpu)
5815 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5824 " reserved_highatomic:%luKB"
5825 " active_anon:%lukB"
5826 " inactive_anon:%lukB"
5827 " active_file:%lukB"
5828 " inactive_file:%lukB"
5829 " unevictable:%lukB"
5830 " writepending:%lukB"
5840 K(zone_page_state(zone, NR_FREE_PAGES)),
5841 K(min_wmark_pages(zone)),
5842 K(low_wmark_pages(zone)),
5843 K(high_wmark_pages(zone)),
5844 K(zone->nr_reserved_highatomic),
5845 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5846 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5847 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5848 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5849 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5850 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5851 K(zone->present_pages),
5852 K(zone_managed_pages(zone)),
5853 K(zone_page_state(zone, NR_MLOCK)),
5854 K(zone_page_state(zone, NR_BOUNCE)),
5856 K(this_cpu_read(zone->pageset->pcp.count)),
5857 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5858 printk("lowmem_reserve[]:");
5859 for (i = 0; i < MAX_NR_ZONES; i++)
5860 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5861 printk(KERN_CONT "\n");
5864 for_each_populated_zone(zone) {
5866 unsigned long nr[MAX_ORDER], flags, total = 0;
5867 unsigned char types[MAX_ORDER];
5869 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5872 printk(KERN_CONT "%s: ", zone->name);
5874 spin_lock_irqsave(&zone->lock, flags);
5875 for (order = 0; order < MAX_ORDER; order++) {
5876 struct free_area *area = &zone->free_area[order];
5879 nr[order] = area->nr_free;
5880 total += nr[order] << order;
5883 for (type = 0; type < MIGRATE_TYPES; type++) {
5884 if (!free_area_empty(area, type))
5885 types[order] |= 1 << type;
5888 spin_unlock_irqrestore(&zone->lock, flags);
5889 for (order = 0; order < MAX_ORDER; order++) {
5890 printk(KERN_CONT "%lu*%lukB ",
5891 nr[order], K(1UL) << order);
5893 show_migration_types(types[order]);
5895 printk(KERN_CONT "= %lukB\n", K(total));
5898 hugetlb_show_meminfo();
5900 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5902 show_swap_cache_info();
5905 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5907 zoneref->zone = zone;
5908 zoneref->zone_idx = zone_idx(zone);
5912 * Builds allocation fallback zone lists.
5914 * Add all populated zones of a node to the zonelist.
5916 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5919 enum zone_type zone_type = MAX_NR_ZONES;
5924 zone = pgdat->node_zones + zone_type;
5925 if (managed_zone(zone)) {
5926 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5927 check_highest_zone(zone_type);
5929 } while (zone_type);
5936 static int __parse_numa_zonelist_order(char *s)
5939 * We used to support different zonelists modes but they turned
5940 * out to be just not useful. Let's keep the warning in place
5941 * if somebody still use the cmd line parameter so that we do
5942 * not fail it silently
5944 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5945 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5951 char numa_zonelist_order[] = "Node";
5954 * sysctl handler for numa_zonelist_order
5956 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5957 void *buffer, size_t *length, loff_t *ppos)
5960 return __parse_numa_zonelist_order(buffer);
5961 return proc_dostring(table, write, buffer, length, ppos);
5965 #define MAX_NODE_LOAD (nr_online_nodes)
5966 static int node_load[MAX_NUMNODES];
5969 * find_next_best_node - find the next node that should appear in a given node's fallback list
5970 * @node: node whose fallback list we're appending
5971 * @used_node_mask: nodemask_t of already used nodes
5973 * We use a number of factors to determine which is the next node that should
5974 * appear on a given node's fallback list. The node should not have appeared
5975 * already in @node's fallback list, and it should be the next closest node
5976 * according to the distance array (which contains arbitrary distance values
5977 * from each node to each node in the system), and should also prefer nodes
5978 * with no CPUs, since presumably they'll have very little allocation pressure
5979 * on them otherwise.
5981 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5983 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5986 int min_val = INT_MAX;
5987 int best_node = NUMA_NO_NODE;
5989 /* Use the local node if we haven't already */
5990 if (!node_isset(node, *used_node_mask)) {
5991 node_set(node, *used_node_mask);
5995 for_each_node_state(n, N_MEMORY) {
5997 /* Don't want a node to appear more than once */
5998 if (node_isset(n, *used_node_mask))
6001 /* Use the distance array to find the distance */
6002 val = node_distance(node, n);
6004 /* Penalize nodes under us ("prefer the next node") */
6007 /* Give preference to headless and unused nodes */
6008 if (!cpumask_empty(cpumask_of_node(n)))
6009 val += PENALTY_FOR_NODE_WITH_CPUS;
6011 /* Slight preference for less loaded node */
6012 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6013 val += node_load[n];
6015 if (val < min_val) {
6022 node_set(best_node, *used_node_mask);
6029 * Build zonelists ordered by node and zones within node.
6030 * This results in maximum locality--normal zone overflows into local
6031 * DMA zone, if any--but risks exhausting DMA zone.
6033 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6036 struct zoneref *zonerefs;
6039 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6041 for (i = 0; i < nr_nodes; i++) {
6044 pg_data_t *node = NODE_DATA(node_order[i]);
6046 nr_zones = build_zonerefs_node(node, zonerefs);
6047 zonerefs += nr_zones;
6049 zonerefs->zone = NULL;
6050 zonerefs->zone_idx = 0;
6054 * Build gfp_thisnode zonelists
6056 static void build_thisnode_zonelists(pg_data_t *pgdat)
6058 struct zoneref *zonerefs;
6061 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6062 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6063 zonerefs += nr_zones;
6064 zonerefs->zone = NULL;
6065 zonerefs->zone_idx = 0;
6069 * Build zonelists ordered by zone and nodes within zones.
6070 * This results in conserving DMA zone[s] until all Normal memory is
6071 * exhausted, but results in overflowing to remote node while memory
6072 * may still exist in local DMA zone.
6075 static void build_zonelists(pg_data_t *pgdat)
6077 static int node_order[MAX_NUMNODES];
6078 int node, load, nr_nodes = 0;
6079 nodemask_t used_mask = NODE_MASK_NONE;
6080 int local_node, prev_node;
6082 /* NUMA-aware ordering of nodes */
6083 local_node = pgdat->node_id;
6084 load = nr_online_nodes;
6085 prev_node = local_node;
6087 memset(node_order, 0, sizeof(node_order));
6088 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6090 * We don't want to pressure a particular node.
6091 * So adding penalty to the first node in same
6092 * distance group to make it round-robin.
6094 if (node_distance(local_node, node) !=
6095 node_distance(local_node, prev_node))
6096 node_load[node] = load;
6098 node_order[nr_nodes++] = node;
6103 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6104 build_thisnode_zonelists(pgdat);
6107 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6109 * Return node id of node used for "local" allocations.
6110 * I.e., first node id of first zone in arg node's generic zonelist.
6111 * Used for initializing percpu 'numa_mem', which is used primarily
6112 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6114 int local_memory_node(int node)
6118 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6119 gfp_zone(GFP_KERNEL),
6121 return zone_to_nid(z->zone);
6125 static void setup_min_unmapped_ratio(void);
6126 static void setup_min_slab_ratio(void);
6127 #else /* CONFIG_NUMA */
6129 static void build_zonelists(pg_data_t *pgdat)
6131 int node, local_node;
6132 struct zoneref *zonerefs;
6135 local_node = pgdat->node_id;
6137 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6138 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6139 zonerefs += nr_zones;
6142 * Now we build the zonelist so that it contains the zones
6143 * of all the other nodes.
6144 * We don't want to pressure a particular node, so when
6145 * building the zones for node N, we make sure that the
6146 * zones coming right after the local ones are those from
6147 * node N+1 (modulo N)
6149 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6150 if (!node_online(node))
6152 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6153 zonerefs += nr_zones;
6155 for (node = 0; node < local_node; node++) {
6156 if (!node_online(node))
6158 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6159 zonerefs += nr_zones;
6162 zonerefs->zone = NULL;
6163 zonerefs->zone_idx = 0;
6166 #endif /* CONFIG_NUMA */
6169 * Boot pageset table. One per cpu which is going to be used for all
6170 * zones and all nodes. The parameters will be set in such a way
6171 * that an item put on a list will immediately be handed over to
6172 * the buddy list. This is safe since pageset manipulation is done
6173 * with interrupts disabled.
6175 * The boot_pagesets must be kept even after bootup is complete for
6176 * unused processors and/or zones. They do play a role for bootstrapping
6177 * hotplugged processors.
6179 * zoneinfo_show() and maybe other functions do
6180 * not check if the processor is online before following the pageset pointer.
6181 * Other parts of the kernel may not check if the zone is available.
6183 static void pageset_init(struct per_cpu_pageset *p);
6184 /* These effectively disable the pcplists in the boot pageset completely */
6185 #define BOOT_PAGESET_HIGH 0
6186 #define BOOT_PAGESET_BATCH 1
6187 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
6188 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6190 static void __build_all_zonelists(void *data)
6193 int __maybe_unused cpu;
6194 pg_data_t *self = data;
6195 static DEFINE_SPINLOCK(lock);
6200 memset(node_load, 0, sizeof(node_load));
6204 * This node is hotadded and no memory is yet present. So just
6205 * building zonelists is fine - no need to touch other nodes.
6207 if (self && !node_online(self->node_id)) {
6208 build_zonelists(self);
6210 for_each_online_node(nid) {
6211 pg_data_t *pgdat = NODE_DATA(nid);
6213 build_zonelists(pgdat);
6216 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6218 * We now know the "local memory node" for each node--
6219 * i.e., the node of the first zone in the generic zonelist.
6220 * Set up numa_mem percpu variable for on-line cpus. During
6221 * boot, only the boot cpu should be on-line; we'll init the
6222 * secondary cpus' numa_mem as they come on-line. During
6223 * node/memory hotplug, we'll fixup all on-line cpus.
6225 for_each_online_cpu(cpu)
6226 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6233 static noinline void __init
6234 build_all_zonelists_init(void)
6238 __build_all_zonelists(NULL);
6241 * Initialize the boot_pagesets that are going to be used
6242 * for bootstrapping processors. The real pagesets for
6243 * each zone will be allocated later when the per cpu
6244 * allocator is available.
6246 * boot_pagesets are used also for bootstrapping offline
6247 * cpus if the system is already booted because the pagesets
6248 * are needed to initialize allocators on a specific cpu too.
6249 * F.e. the percpu allocator needs the page allocator which
6250 * needs the percpu allocator in order to allocate its pagesets
6251 * (a chicken-egg dilemma).
6253 for_each_possible_cpu(cpu)
6254 pageset_init(&per_cpu(boot_pageset, cpu));
6256 mminit_verify_zonelist();
6257 cpuset_init_current_mems_allowed();
6261 * unless system_state == SYSTEM_BOOTING.
6263 * __ref due to call of __init annotated helper build_all_zonelists_init
6264 * [protected by SYSTEM_BOOTING].
6266 void __ref build_all_zonelists(pg_data_t *pgdat)
6268 unsigned long vm_total_pages;
6270 if (system_state == SYSTEM_BOOTING) {
6271 build_all_zonelists_init();
6273 __build_all_zonelists(pgdat);
6274 /* cpuset refresh routine should be here */
6276 /* Get the number of free pages beyond high watermark in all zones. */
6277 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6279 * Disable grouping by mobility if the number of pages in the
6280 * system is too low to allow the mechanism to work. It would be
6281 * more accurate, but expensive to check per-zone. This check is
6282 * made on memory-hotadd so a system can start with mobility
6283 * disabled and enable it later
6285 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6286 page_group_by_mobility_disabled = 1;
6288 page_group_by_mobility_disabled = 0;
6290 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6292 page_group_by_mobility_disabled ? "off" : "on",
6295 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6299 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6300 static bool __meminit
6301 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6303 static struct memblock_region *r;
6305 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6306 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6307 for_each_mem_region(r) {
6308 if (*pfn < memblock_region_memory_end_pfn(r))
6312 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6313 memblock_is_mirror(r)) {
6314 *pfn = memblock_region_memory_end_pfn(r);
6322 * Initially all pages are reserved - free ones are freed
6323 * up by memblock_free_all() once the early boot process is
6324 * done. Non-atomic initialization, single-pass.
6326 * All aligned pageblocks are initialized to the specified migratetype
6327 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6328 * zone stats (e.g., nr_isolate_pageblock) are touched.
6330 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6331 unsigned long start_pfn, unsigned long zone_end_pfn,
6332 enum meminit_context context,
6333 struct vmem_altmap *altmap, int migratetype)
6335 unsigned long pfn, end_pfn = start_pfn + size;
6338 if (highest_memmap_pfn < end_pfn - 1)
6339 highest_memmap_pfn = end_pfn - 1;
6341 #ifdef CONFIG_ZONE_DEVICE
6343 * Honor reservation requested by the driver for this ZONE_DEVICE
6344 * memory. We limit the total number of pages to initialize to just
6345 * those that might contain the memory mapping. We will defer the
6346 * ZONE_DEVICE page initialization until after we have released
6349 if (zone == ZONE_DEVICE) {
6353 if (start_pfn == altmap->base_pfn)
6354 start_pfn += altmap->reserve;
6355 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6359 for (pfn = start_pfn; pfn < end_pfn; ) {
6361 * There can be holes in boot-time mem_map[]s handed to this
6362 * function. They do not exist on hotplugged memory.
6364 if (context == MEMINIT_EARLY) {
6365 if (overlap_memmap_init(zone, &pfn))
6367 if (defer_init(nid, pfn, zone_end_pfn))
6371 page = pfn_to_page(pfn);
6372 __init_single_page(page, pfn, zone, nid);
6373 if (context == MEMINIT_HOTPLUG)
6374 __SetPageReserved(page);
6377 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6378 * such that unmovable allocations won't be scattered all
6379 * over the place during system boot.
6381 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6382 set_pageblock_migratetype(page, migratetype);
6389 #ifdef CONFIG_ZONE_DEVICE
6390 void __ref memmap_init_zone_device(struct zone *zone,
6391 unsigned long start_pfn,
6392 unsigned long nr_pages,
6393 struct dev_pagemap *pgmap)
6395 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6396 struct pglist_data *pgdat = zone->zone_pgdat;
6397 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6398 unsigned long zone_idx = zone_idx(zone);
6399 unsigned long start = jiffies;
6400 int nid = pgdat->node_id;
6402 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6406 * The call to memmap_init should have already taken care
6407 * of the pages reserved for the memmap, so we can just jump to
6408 * the end of that region and start processing the device pages.
6411 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6412 nr_pages = end_pfn - start_pfn;
6415 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6416 struct page *page = pfn_to_page(pfn);
6418 __init_single_page(page, pfn, zone_idx, nid);
6421 * Mark page reserved as it will need to wait for onlining
6422 * phase for it to be fully associated with a zone.
6424 * We can use the non-atomic __set_bit operation for setting
6425 * the flag as we are still initializing the pages.
6427 __SetPageReserved(page);
6430 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6431 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6432 * ever freed or placed on a driver-private list.
6434 page->pgmap = pgmap;
6435 page->zone_device_data = NULL;
6438 * Mark the block movable so that blocks are reserved for
6439 * movable at startup. This will force kernel allocations
6440 * to reserve their blocks rather than leaking throughout
6441 * the address space during boot when many long-lived
6442 * kernel allocations are made.
6444 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6445 * because this is done early in section_activate()
6447 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6448 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6453 pr_info("%s initialised %lu pages in %ums\n", __func__,
6454 nr_pages, jiffies_to_msecs(jiffies - start));
6458 static void __meminit zone_init_free_lists(struct zone *zone)
6460 unsigned int order, t;
6461 for_each_migratetype_order(order, t) {
6462 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6463 zone->free_area[order].nr_free = 0;
6467 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6469 * Only struct pages that correspond to ranges defined by memblock.memory
6470 * are zeroed and initialized by going through __init_single_page() during
6471 * memmap_init_zone_range().
6473 * But, there could be struct pages that correspond to holes in
6474 * memblock.memory. This can happen because of the following reasons:
6475 * - physical memory bank size is not necessarily the exact multiple of the
6476 * arbitrary section size
6477 * - early reserved memory may not be listed in memblock.memory
6478 * - memory layouts defined with memmap= kernel parameter may not align
6479 * nicely with memmap sections
6481 * Explicitly initialize those struct pages so that:
6482 * - PG_Reserved is set
6483 * - zone and node links point to zone and node that span the page if the
6484 * hole is in the middle of a zone
6485 * - zone and node links point to adjacent zone/node if the hole falls on
6486 * the zone boundary; the pages in such holes will be prepended to the
6487 * zone/node above the hole except for the trailing pages in the last
6488 * section that will be appended to the zone/node below.
6490 static void __init init_unavailable_range(unsigned long spfn,
6497 for (pfn = spfn; pfn < epfn; pfn++) {
6498 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6499 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6500 + pageblock_nr_pages - 1;
6503 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6504 __SetPageReserved(pfn_to_page(pfn));
6509 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6510 node, zone_names[zone], pgcnt);
6513 static inline void init_unavailable_range(unsigned long spfn,
6520 static void __init memmap_init_zone_range(struct zone *zone,
6521 unsigned long start_pfn,
6522 unsigned long end_pfn,
6523 unsigned long *hole_pfn)
6525 unsigned long zone_start_pfn = zone->zone_start_pfn;
6526 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6527 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6529 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6530 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6532 if (start_pfn >= end_pfn)
6535 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6536 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6538 if (*hole_pfn < start_pfn)
6539 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6541 *hole_pfn = end_pfn;
6544 static void __init memmap_init(void)
6546 unsigned long start_pfn, end_pfn;
6547 unsigned long hole_pfn = 0;
6548 int i, j, zone_id, nid;
6550 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6551 struct pglist_data *node = NODE_DATA(nid);
6553 for (j = 0; j < MAX_NR_ZONES; j++) {
6554 struct zone *zone = node->node_zones + j;
6556 if (!populated_zone(zone))
6559 memmap_init_zone_range(zone, start_pfn, end_pfn,
6565 #ifdef CONFIG_SPARSEMEM
6567 * Initialize the memory map for hole in the range [memory_end,
6569 * Append the pages in this hole to the highest zone in the last
6571 * The call to init_unavailable_range() is outside the ifdef to
6572 * silence the compiler warining about zone_id set but not used;
6573 * for FLATMEM it is a nop anyway
6575 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6576 if (hole_pfn < end_pfn)
6578 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6581 static int zone_batchsize(struct zone *zone)
6587 * The per-cpu-pages pools are set to around 1000th of the
6590 batch = zone_managed_pages(zone) / 1024;
6591 /* But no more than a meg. */
6592 if (batch * PAGE_SIZE > 1024 * 1024)
6593 batch = (1024 * 1024) / PAGE_SIZE;
6594 batch /= 4; /* We effectively *= 4 below */
6599 * Clamp the batch to a 2^n - 1 value. Having a power
6600 * of 2 value was found to be more likely to have
6601 * suboptimal cache aliasing properties in some cases.
6603 * For example if 2 tasks are alternately allocating
6604 * batches of pages, one task can end up with a lot
6605 * of pages of one half of the possible page colors
6606 * and the other with pages of the other colors.
6608 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6613 /* The deferral and batching of frees should be suppressed under NOMMU
6616 * The problem is that NOMMU needs to be able to allocate large chunks
6617 * of contiguous memory as there's no hardware page translation to
6618 * assemble apparent contiguous memory from discontiguous pages.
6620 * Queueing large contiguous runs of pages for batching, however,
6621 * causes the pages to actually be freed in smaller chunks. As there
6622 * can be a significant delay between the individual batches being
6623 * recycled, this leads to the once large chunks of space being
6624 * fragmented and becoming unavailable for high-order allocations.
6631 * pcp->high and pcp->batch values are related and generally batch is lower
6632 * than high. They are also related to pcp->count such that count is lower
6633 * than high, and as soon as it reaches high, the pcplist is flushed.
6635 * However, guaranteeing these relations at all times would require e.g. write
6636 * barriers here but also careful usage of read barriers at the read side, and
6637 * thus be prone to error and bad for performance. Thus the update only prevents
6638 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6639 * can cope with those fields changing asynchronously, and fully trust only the
6640 * pcp->count field on the local CPU with interrupts disabled.
6642 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6643 * outside of boot time (or some other assurance that no concurrent updaters
6646 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6647 unsigned long batch)
6649 WRITE_ONCE(pcp->batch, batch);
6650 WRITE_ONCE(pcp->high, high);
6653 static void pageset_init(struct per_cpu_pageset *p)
6655 struct per_cpu_pages *pcp;
6658 memset(p, 0, sizeof(*p));
6661 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6662 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6665 * Set batch and high values safe for a boot pageset. A true percpu
6666 * pageset's initialization will update them subsequently. Here we don't
6667 * need to be as careful as pageset_update() as nobody can access the
6670 pcp->high = BOOT_PAGESET_HIGH;
6671 pcp->batch = BOOT_PAGESET_BATCH;
6674 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6675 unsigned long batch)
6677 struct per_cpu_pageset *p;
6680 for_each_possible_cpu(cpu) {
6681 p = per_cpu_ptr(zone->pageset, cpu);
6682 pageset_update(&p->pcp, high, batch);
6687 * Calculate and set new high and batch values for all per-cpu pagesets of a
6688 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6690 static void zone_set_pageset_high_and_batch(struct zone *zone)
6692 unsigned long new_high, new_batch;
6694 if (percpu_pagelist_fraction) {
6695 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6696 new_batch = max(1UL, new_high / 4);
6697 if ((new_high / 4) > (PAGE_SHIFT * 8))
6698 new_batch = PAGE_SHIFT * 8;
6700 new_batch = zone_batchsize(zone);
6701 new_high = 6 * new_batch;
6702 new_batch = max(1UL, 1 * new_batch);
6705 if (zone->pageset_high == new_high &&
6706 zone->pageset_batch == new_batch)
6709 zone->pageset_high = new_high;
6710 zone->pageset_batch = new_batch;
6712 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6715 void __meminit setup_zone_pageset(struct zone *zone)
6717 struct per_cpu_pageset *p;
6720 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6721 for_each_possible_cpu(cpu) {
6722 p = per_cpu_ptr(zone->pageset, cpu);
6726 zone_set_pageset_high_and_batch(zone);
6730 * Allocate per cpu pagesets and initialize them.
6731 * Before this call only boot pagesets were available.
6733 void __init setup_per_cpu_pageset(void)
6735 struct pglist_data *pgdat;
6737 int __maybe_unused cpu;
6739 for_each_populated_zone(zone)
6740 setup_zone_pageset(zone);
6744 * Unpopulated zones continue using the boot pagesets.
6745 * The numa stats for these pagesets need to be reset.
6746 * Otherwise, they will end up skewing the stats of
6747 * the nodes these zones are associated with.
6749 for_each_possible_cpu(cpu) {
6750 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6751 memset(pcp->vm_numa_stat_diff, 0,
6752 sizeof(pcp->vm_numa_stat_diff));
6756 for_each_online_pgdat(pgdat)
6757 pgdat->per_cpu_nodestats =
6758 alloc_percpu(struct per_cpu_nodestat);
6761 static __meminit void zone_pcp_init(struct zone *zone)
6764 * per cpu subsystem is not up at this point. The following code
6765 * relies on the ability of the linker to provide the
6766 * offset of a (static) per cpu variable into the per cpu area.
6768 zone->pageset = &boot_pageset;
6769 zone->pageset_high = BOOT_PAGESET_HIGH;
6770 zone->pageset_batch = BOOT_PAGESET_BATCH;
6772 if (populated_zone(zone))
6773 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6774 zone->name, zone->present_pages,
6775 zone_batchsize(zone));
6778 void __meminit init_currently_empty_zone(struct zone *zone,
6779 unsigned long zone_start_pfn,
6782 struct pglist_data *pgdat = zone->zone_pgdat;
6783 int zone_idx = zone_idx(zone) + 1;
6785 if (zone_idx > pgdat->nr_zones)
6786 pgdat->nr_zones = zone_idx;
6788 zone->zone_start_pfn = zone_start_pfn;
6790 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6791 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6793 (unsigned long)zone_idx(zone),
6794 zone_start_pfn, (zone_start_pfn + size));
6796 zone_init_free_lists(zone);
6797 zone->initialized = 1;
6801 * get_pfn_range_for_nid - Return the start and end page frames for a node
6802 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6803 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6804 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6806 * It returns the start and end page frame of a node based on information
6807 * provided by memblock_set_node(). If called for a node
6808 * with no available memory, a warning is printed and the start and end
6811 void __init get_pfn_range_for_nid(unsigned int nid,
6812 unsigned long *start_pfn, unsigned long *end_pfn)
6814 unsigned long this_start_pfn, this_end_pfn;
6820 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6821 *start_pfn = min(*start_pfn, this_start_pfn);
6822 *end_pfn = max(*end_pfn, this_end_pfn);
6825 if (*start_pfn == -1UL)
6830 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6831 * assumption is made that zones within a node are ordered in monotonic
6832 * increasing memory addresses so that the "highest" populated zone is used
6834 static void __init find_usable_zone_for_movable(void)
6837 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6838 if (zone_index == ZONE_MOVABLE)
6841 if (arch_zone_highest_possible_pfn[zone_index] >
6842 arch_zone_lowest_possible_pfn[zone_index])
6846 VM_BUG_ON(zone_index == -1);
6847 movable_zone = zone_index;
6851 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6852 * because it is sized independent of architecture. Unlike the other zones,
6853 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6854 * in each node depending on the size of each node and how evenly kernelcore
6855 * is distributed. This helper function adjusts the zone ranges
6856 * provided by the architecture for a given node by using the end of the
6857 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6858 * zones within a node are in order of monotonic increases memory addresses
6860 static void __init adjust_zone_range_for_zone_movable(int nid,
6861 unsigned long zone_type,
6862 unsigned long node_start_pfn,
6863 unsigned long node_end_pfn,
6864 unsigned long *zone_start_pfn,
6865 unsigned long *zone_end_pfn)
6867 /* Only adjust if ZONE_MOVABLE is on this node */
6868 if (zone_movable_pfn[nid]) {
6869 /* Size ZONE_MOVABLE */
6870 if (zone_type == ZONE_MOVABLE) {
6871 *zone_start_pfn = zone_movable_pfn[nid];
6872 *zone_end_pfn = min(node_end_pfn,
6873 arch_zone_highest_possible_pfn[movable_zone]);
6875 /* Adjust for ZONE_MOVABLE starting within this range */
6876 } else if (!mirrored_kernelcore &&
6877 *zone_start_pfn < zone_movable_pfn[nid] &&
6878 *zone_end_pfn > zone_movable_pfn[nid]) {
6879 *zone_end_pfn = zone_movable_pfn[nid];
6881 /* Check if this whole range is within ZONE_MOVABLE */
6882 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6883 *zone_start_pfn = *zone_end_pfn;
6888 * Return the number of pages a zone spans in a node, including holes
6889 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6891 static unsigned long __init zone_spanned_pages_in_node(int nid,
6892 unsigned long zone_type,
6893 unsigned long node_start_pfn,
6894 unsigned long node_end_pfn,
6895 unsigned long *zone_start_pfn,
6896 unsigned long *zone_end_pfn)
6898 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6899 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6900 /* When hotadd a new node from cpu_up(), the node should be empty */
6901 if (!node_start_pfn && !node_end_pfn)
6904 /* Get the start and end of the zone */
6905 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6906 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6907 adjust_zone_range_for_zone_movable(nid, zone_type,
6908 node_start_pfn, node_end_pfn,
6909 zone_start_pfn, zone_end_pfn);
6911 /* Check that this node has pages within the zone's required range */
6912 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6915 /* Move the zone boundaries inside the node if necessary */
6916 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6917 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6919 /* Return the spanned pages */
6920 return *zone_end_pfn - *zone_start_pfn;
6924 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6925 * then all holes in the requested range will be accounted for.
6927 unsigned long __init __absent_pages_in_range(int nid,
6928 unsigned long range_start_pfn,
6929 unsigned long range_end_pfn)
6931 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6932 unsigned long start_pfn, end_pfn;
6935 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6936 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6937 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6938 nr_absent -= end_pfn - start_pfn;
6944 * absent_pages_in_range - Return number of page frames in holes within a range
6945 * @start_pfn: The start PFN to start searching for holes
6946 * @end_pfn: The end PFN to stop searching for holes
6948 * Return: the number of pages frames in memory holes within a range.
6950 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6951 unsigned long end_pfn)
6953 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6956 /* Return the number of page frames in holes in a zone on a node */
6957 static unsigned long __init zone_absent_pages_in_node(int nid,
6958 unsigned long zone_type,
6959 unsigned long node_start_pfn,
6960 unsigned long node_end_pfn)
6962 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6963 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6964 unsigned long zone_start_pfn, zone_end_pfn;
6965 unsigned long nr_absent;
6967 /* When hotadd a new node from cpu_up(), the node should be empty */
6968 if (!node_start_pfn && !node_end_pfn)
6971 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6972 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6974 adjust_zone_range_for_zone_movable(nid, zone_type,
6975 node_start_pfn, node_end_pfn,
6976 &zone_start_pfn, &zone_end_pfn);
6977 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6980 * ZONE_MOVABLE handling.
6981 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6984 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6985 unsigned long start_pfn, end_pfn;
6986 struct memblock_region *r;
6988 for_each_mem_region(r) {
6989 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6990 zone_start_pfn, zone_end_pfn);
6991 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6992 zone_start_pfn, zone_end_pfn);
6994 if (zone_type == ZONE_MOVABLE &&
6995 memblock_is_mirror(r))
6996 nr_absent += end_pfn - start_pfn;
6998 if (zone_type == ZONE_NORMAL &&
6999 !memblock_is_mirror(r))
7000 nr_absent += end_pfn - start_pfn;
7007 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7008 unsigned long node_start_pfn,
7009 unsigned long node_end_pfn)
7011 unsigned long realtotalpages = 0, totalpages = 0;
7014 for (i = 0; i < MAX_NR_ZONES; i++) {
7015 struct zone *zone = pgdat->node_zones + i;
7016 unsigned long zone_start_pfn, zone_end_pfn;
7017 unsigned long spanned, absent;
7018 unsigned long size, real_size;
7020 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7025 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7030 real_size = size - absent;
7033 zone->zone_start_pfn = zone_start_pfn;
7035 zone->zone_start_pfn = 0;
7036 zone->spanned_pages = size;
7037 zone->present_pages = real_size;
7040 realtotalpages += real_size;
7043 pgdat->node_spanned_pages = totalpages;
7044 pgdat->node_present_pages = realtotalpages;
7045 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
7049 #ifndef CONFIG_SPARSEMEM
7051 * Calculate the size of the zone->blockflags rounded to an unsigned long
7052 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7053 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7054 * round what is now in bits to nearest long in bits, then return it in
7057 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7059 unsigned long usemapsize;
7061 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7062 usemapsize = roundup(zonesize, pageblock_nr_pages);
7063 usemapsize = usemapsize >> pageblock_order;
7064 usemapsize *= NR_PAGEBLOCK_BITS;
7065 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7067 return usemapsize / 8;
7070 static void __ref setup_usemap(struct zone *zone)
7072 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7073 zone->spanned_pages);
7074 zone->pageblock_flags = NULL;
7076 zone->pageblock_flags =
7077 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7079 if (!zone->pageblock_flags)
7080 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7081 usemapsize, zone->name, zone_to_nid(zone));
7085 static inline void setup_usemap(struct zone *zone) {}
7086 #endif /* CONFIG_SPARSEMEM */
7088 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7090 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7091 void __init set_pageblock_order(void)
7095 /* Check that pageblock_nr_pages has not already been setup */
7096 if (pageblock_order)
7099 if (HPAGE_SHIFT > PAGE_SHIFT)
7100 order = HUGETLB_PAGE_ORDER;
7102 order = MAX_ORDER - 1;
7105 * Assume the largest contiguous order of interest is a huge page.
7106 * This value may be variable depending on boot parameters on IA64 and
7109 pageblock_order = order;
7111 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7114 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7115 * is unused as pageblock_order is set at compile-time. See
7116 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7119 void __init set_pageblock_order(void)
7123 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7125 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7126 unsigned long present_pages)
7128 unsigned long pages = spanned_pages;
7131 * Provide a more accurate estimation if there are holes within
7132 * the zone and SPARSEMEM is in use. If there are holes within the
7133 * zone, each populated memory region may cost us one or two extra
7134 * memmap pages due to alignment because memmap pages for each
7135 * populated regions may not be naturally aligned on page boundary.
7136 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7138 if (spanned_pages > present_pages + (present_pages >> 4) &&
7139 IS_ENABLED(CONFIG_SPARSEMEM))
7140 pages = present_pages;
7142 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7145 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7146 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7148 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7150 spin_lock_init(&ds_queue->split_queue_lock);
7151 INIT_LIST_HEAD(&ds_queue->split_queue);
7152 ds_queue->split_queue_len = 0;
7155 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7158 #ifdef CONFIG_COMPACTION
7159 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7161 init_waitqueue_head(&pgdat->kcompactd_wait);
7164 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7167 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7169 pgdat_resize_init(pgdat);
7171 pgdat_init_split_queue(pgdat);
7172 pgdat_init_kcompactd(pgdat);
7174 init_waitqueue_head(&pgdat->kswapd_wait);
7175 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7177 pgdat_page_ext_init(pgdat);
7178 lruvec_init(&pgdat->__lruvec);
7181 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7182 unsigned long remaining_pages)
7184 atomic_long_set(&zone->managed_pages, remaining_pages);
7185 zone_set_nid(zone, nid);
7186 zone->name = zone_names[idx];
7187 zone->zone_pgdat = NODE_DATA(nid);
7188 spin_lock_init(&zone->lock);
7189 zone_seqlock_init(zone);
7190 zone_pcp_init(zone);
7194 * Set up the zone data structures
7195 * - init pgdat internals
7196 * - init all zones belonging to this node
7198 * NOTE: this function is only called during memory hotplug
7200 #ifdef CONFIG_MEMORY_HOTPLUG
7201 void __ref free_area_init_core_hotplug(int nid)
7204 pg_data_t *pgdat = NODE_DATA(nid);
7206 pgdat_init_internals(pgdat);
7207 for (z = 0; z < MAX_NR_ZONES; z++)
7208 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7213 * Set up the zone data structures:
7214 * - mark all pages reserved
7215 * - mark all memory queues empty
7216 * - clear the memory bitmaps
7218 * NOTE: pgdat should get zeroed by caller.
7219 * NOTE: this function is only called during early init.
7221 static void __init free_area_init_core(struct pglist_data *pgdat)
7224 int nid = pgdat->node_id;
7226 pgdat_init_internals(pgdat);
7227 pgdat->per_cpu_nodestats = &boot_nodestats;
7229 for (j = 0; j < MAX_NR_ZONES; j++) {
7230 struct zone *zone = pgdat->node_zones + j;
7231 unsigned long size, freesize, memmap_pages;
7233 size = zone->spanned_pages;
7234 freesize = zone->present_pages;
7237 * Adjust freesize so that it accounts for how much memory
7238 * is used by this zone for memmap. This affects the watermark
7239 * and per-cpu initialisations
7241 memmap_pages = calc_memmap_size(size, freesize);
7242 if (!is_highmem_idx(j)) {
7243 if (freesize >= memmap_pages) {
7244 freesize -= memmap_pages;
7247 " %s zone: %lu pages used for memmap\n",
7248 zone_names[j], memmap_pages);
7250 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
7251 zone_names[j], memmap_pages, freesize);
7254 /* Account for reserved pages */
7255 if (j == 0 && freesize > dma_reserve) {
7256 freesize -= dma_reserve;
7257 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
7258 zone_names[0], dma_reserve);
7261 if (!is_highmem_idx(j))
7262 nr_kernel_pages += freesize;
7263 /* Charge for highmem memmap if there are enough kernel pages */
7264 else if (nr_kernel_pages > memmap_pages * 2)
7265 nr_kernel_pages -= memmap_pages;
7266 nr_all_pages += freesize;
7269 * Set an approximate value for lowmem here, it will be adjusted
7270 * when the bootmem allocator frees pages into the buddy system.
7271 * And all highmem pages will be managed by the buddy system.
7273 zone_init_internals(zone, j, nid, freesize);
7278 set_pageblock_order();
7280 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7284 #ifdef CONFIG_FLAT_NODE_MEM_MAP
7285 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7287 unsigned long __maybe_unused start = 0;
7288 unsigned long __maybe_unused offset = 0;
7290 /* Skip empty nodes */
7291 if (!pgdat->node_spanned_pages)
7294 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7295 offset = pgdat->node_start_pfn - start;
7296 /* ia64 gets its own node_mem_map, before this, without bootmem */
7297 if (!pgdat->node_mem_map) {
7298 unsigned long size, end;
7302 * The zone's endpoints aren't required to be MAX_ORDER
7303 * aligned but the node_mem_map endpoints must be in order
7304 * for the buddy allocator to function correctly.
7306 end = pgdat_end_pfn(pgdat);
7307 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7308 size = (end - start) * sizeof(struct page);
7309 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7312 panic("Failed to allocate %ld bytes for node %d memory map\n",
7313 size, pgdat->node_id);
7314 pgdat->node_mem_map = map + offset;
7316 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7317 __func__, pgdat->node_id, (unsigned long)pgdat,
7318 (unsigned long)pgdat->node_mem_map);
7319 #ifndef CONFIG_NEED_MULTIPLE_NODES
7321 * With no DISCONTIG, the global mem_map is just set as node 0's
7323 if (pgdat == NODE_DATA(0)) {
7324 mem_map = NODE_DATA(0)->node_mem_map;
7325 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7331 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7332 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7334 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7335 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7337 pgdat->first_deferred_pfn = ULONG_MAX;
7340 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7343 static void __init free_area_init_node(int nid)
7345 pg_data_t *pgdat = NODE_DATA(nid);
7346 unsigned long start_pfn = 0;
7347 unsigned long end_pfn = 0;
7349 /* pg_data_t should be reset to zero when it's allocated */
7350 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7352 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7354 pgdat->node_id = nid;
7355 pgdat->node_start_pfn = start_pfn;
7356 pgdat->per_cpu_nodestats = NULL;
7358 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7359 (u64)start_pfn << PAGE_SHIFT,
7360 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7361 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7363 alloc_node_mem_map(pgdat);
7364 pgdat_set_deferred_range(pgdat);
7366 free_area_init_core(pgdat);
7369 void __init free_area_init_memoryless_node(int nid)
7371 free_area_init_node(nid);
7374 #if MAX_NUMNODES > 1
7376 * Figure out the number of possible node ids.
7378 void __init setup_nr_node_ids(void)
7380 unsigned int highest;
7382 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7383 nr_node_ids = highest + 1;
7388 * node_map_pfn_alignment - determine the maximum internode alignment
7390 * This function should be called after node map is populated and sorted.
7391 * It calculates the maximum power of two alignment which can distinguish
7394 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7395 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7396 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7397 * shifted, 1GiB is enough and this function will indicate so.
7399 * This is used to test whether pfn -> nid mapping of the chosen memory
7400 * model has fine enough granularity to avoid incorrect mapping for the
7401 * populated node map.
7403 * Return: the determined alignment in pfn's. 0 if there is no alignment
7404 * requirement (single node).
7406 unsigned long __init node_map_pfn_alignment(void)
7408 unsigned long accl_mask = 0, last_end = 0;
7409 unsigned long start, end, mask;
7410 int last_nid = NUMA_NO_NODE;
7413 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7414 if (!start || last_nid < 0 || last_nid == nid) {
7421 * Start with a mask granular enough to pin-point to the
7422 * start pfn and tick off bits one-by-one until it becomes
7423 * too coarse to separate the current node from the last.
7425 mask = ~((1 << __ffs(start)) - 1);
7426 while (mask && last_end <= (start & (mask << 1)))
7429 /* accumulate all internode masks */
7433 /* convert mask to number of pages */
7434 return ~accl_mask + 1;
7438 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7440 * Return: the minimum PFN based on information provided via
7441 * memblock_set_node().
7443 unsigned long __init find_min_pfn_with_active_regions(void)
7445 return PHYS_PFN(memblock_start_of_DRAM());
7449 * early_calculate_totalpages()
7450 * Sum pages in active regions for movable zone.
7451 * Populate N_MEMORY for calculating usable_nodes.
7453 static unsigned long __init early_calculate_totalpages(void)
7455 unsigned long totalpages = 0;
7456 unsigned long start_pfn, end_pfn;
7459 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7460 unsigned long pages = end_pfn - start_pfn;
7462 totalpages += pages;
7464 node_set_state(nid, N_MEMORY);
7470 * Find the PFN the Movable zone begins in each node. Kernel memory
7471 * is spread evenly between nodes as long as the nodes have enough
7472 * memory. When they don't, some nodes will have more kernelcore than
7475 static void __init find_zone_movable_pfns_for_nodes(void)
7478 unsigned long usable_startpfn;
7479 unsigned long kernelcore_node, kernelcore_remaining;
7480 /* save the state before borrow the nodemask */
7481 nodemask_t saved_node_state = node_states[N_MEMORY];
7482 unsigned long totalpages = early_calculate_totalpages();
7483 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7484 struct memblock_region *r;
7486 /* Need to find movable_zone earlier when movable_node is specified. */
7487 find_usable_zone_for_movable();
7490 * If movable_node is specified, ignore kernelcore and movablecore
7493 if (movable_node_is_enabled()) {
7494 for_each_mem_region(r) {
7495 if (!memblock_is_hotpluggable(r))
7498 nid = memblock_get_region_node(r);
7500 usable_startpfn = PFN_DOWN(r->base);
7501 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7502 min(usable_startpfn, zone_movable_pfn[nid]) :
7510 * If kernelcore=mirror is specified, ignore movablecore option
7512 if (mirrored_kernelcore) {
7513 bool mem_below_4gb_not_mirrored = false;
7515 for_each_mem_region(r) {
7516 if (memblock_is_mirror(r))
7519 nid = memblock_get_region_node(r);
7521 usable_startpfn = memblock_region_memory_base_pfn(r);
7523 if (usable_startpfn < 0x100000) {
7524 mem_below_4gb_not_mirrored = true;
7528 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7529 min(usable_startpfn, zone_movable_pfn[nid]) :
7533 if (mem_below_4gb_not_mirrored)
7534 pr_warn("This configuration results in unmirrored kernel memory.\n");
7540 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7541 * amount of necessary memory.
7543 if (required_kernelcore_percent)
7544 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7546 if (required_movablecore_percent)
7547 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7551 * If movablecore= was specified, calculate what size of
7552 * kernelcore that corresponds so that memory usable for
7553 * any allocation type is evenly spread. If both kernelcore
7554 * and movablecore are specified, then the value of kernelcore
7555 * will be used for required_kernelcore if it's greater than
7556 * what movablecore would have allowed.
7558 if (required_movablecore) {
7559 unsigned long corepages;
7562 * Round-up so that ZONE_MOVABLE is at least as large as what
7563 * was requested by the user
7565 required_movablecore =
7566 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7567 required_movablecore = min(totalpages, required_movablecore);
7568 corepages = totalpages - required_movablecore;
7570 required_kernelcore = max(required_kernelcore, corepages);
7574 * If kernelcore was not specified or kernelcore size is larger
7575 * than totalpages, there is no ZONE_MOVABLE.
7577 if (!required_kernelcore || required_kernelcore >= totalpages)
7580 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7581 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7584 /* Spread kernelcore memory as evenly as possible throughout nodes */
7585 kernelcore_node = required_kernelcore / usable_nodes;
7586 for_each_node_state(nid, N_MEMORY) {
7587 unsigned long start_pfn, end_pfn;
7590 * Recalculate kernelcore_node if the division per node
7591 * now exceeds what is necessary to satisfy the requested
7592 * amount of memory for the kernel
7594 if (required_kernelcore < kernelcore_node)
7595 kernelcore_node = required_kernelcore / usable_nodes;
7598 * As the map is walked, we track how much memory is usable
7599 * by the kernel using kernelcore_remaining. When it is
7600 * 0, the rest of the node is usable by ZONE_MOVABLE
7602 kernelcore_remaining = kernelcore_node;
7604 /* Go through each range of PFNs within this node */
7605 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7606 unsigned long size_pages;
7608 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7609 if (start_pfn >= end_pfn)
7612 /* Account for what is only usable for kernelcore */
7613 if (start_pfn < usable_startpfn) {
7614 unsigned long kernel_pages;
7615 kernel_pages = min(end_pfn, usable_startpfn)
7618 kernelcore_remaining -= min(kernel_pages,
7619 kernelcore_remaining);
7620 required_kernelcore -= min(kernel_pages,
7621 required_kernelcore);
7623 /* Continue if range is now fully accounted */
7624 if (end_pfn <= usable_startpfn) {
7627 * Push zone_movable_pfn to the end so
7628 * that if we have to rebalance
7629 * kernelcore across nodes, we will
7630 * not double account here
7632 zone_movable_pfn[nid] = end_pfn;
7635 start_pfn = usable_startpfn;
7639 * The usable PFN range for ZONE_MOVABLE is from
7640 * start_pfn->end_pfn. Calculate size_pages as the
7641 * number of pages used as kernelcore
7643 size_pages = end_pfn - start_pfn;
7644 if (size_pages > kernelcore_remaining)
7645 size_pages = kernelcore_remaining;
7646 zone_movable_pfn[nid] = start_pfn + size_pages;
7649 * Some kernelcore has been met, update counts and
7650 * break if the kernelcore for this node has been
7653 required_kernelcore -= min(required_kernelcore,
7655 kernelcore_remaining -= size_pages;
7656 if (!kernelcore_remaining)
7662 * If there is still required_kernelcore, we do another pass with one
7663 * less node in the count. This will push zone_movable_pfn[nid] further
7664 * along on the nodes that still have memory until kernelcore is
7668 if (usable_nodes && required_kernelcore > usable_nodes)
7672 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7673 for (nid = 0; nid < MAX_NUMNODES; nid++)
7674 zone_movable_pfn[nid] =
7675 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7678 /* restore the node_state */
7679 node_states[N_MEMORY] = saved_node_state;
7682 /* Any regular or high memory on that node ? */
7683 static void check_for_memory(pg_data_t *pgdat, int nid)
7685 enum zone_type zone_type;
7687 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7688 struct zone *zone = &pgdat->node_zones[zone_type];
7689 if (populated_zone(zone)) {
7690 if (IS_ENABLED(CONFIG_HIGHMEM))
7691 node_set_state(nid, N_HIGH_MEMORY);
7692 if (zone_type <= ZONE_NORMAL)
7693 node_set_state(nid, N_NORMAL_MEMORY);
7700 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7701 * such cases we allow max_zone_pfn sorted in the descending order
7703 bool __weak arch_has_descending_max_zone_pfns(void)
7709 * free_area_init - Initialise all pg_data_t and zone data
7710 * @max_zone_pfn: an array of max PFNs for each zone
7712 * This will call free_area_init_node() for each active node in the system.
7713 * Using the page ranges provided by memblock_set_node(), the size of each
7714 * zone in each node and their holes is calculated. If the maximum PFN
7715 * between two adjacent zones match, it is assumed that the zone is empty.
7716 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7717 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7718 * starts where the previous one ended. For example, ZONE_DMA32 starts
7719 * at arch_max_dma_pfn.
7721 void __init free_area_init(unsigned long *max_zone_pfn)
7723 unsigned long start_pfn, end_pfn;
7727 /* Record where the zone boundaries are */
7728 memset(arch_zone_lowest_possible_pfn, 0,
7729 sizeof(arch_zone_lowest_possible_pfn));
7730 memset(arch_zone_highest_possible_pfn, 0,
7731 sizeof(arch_zone_highest_possible_pfn));
7733 start_pfn = find_min_pfn_with_active_regions();
7734 descending = arch_has_descending_max_zone_pfns();
7736 for (i = 0; i < MAX_NR_ZONES; i++) {
7738 zone = MAX_NR_ZONES - i - 1;
7742 if (zone == ZONE_MOVABLE)
7745 end_pfn = max(max_zone_pfn[zone], start_pfn);
7746 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7747 arch_zone_highest_possible_pfn[zone] = end_pfn;
7749 start_pfn = end_pfn;
7752 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7753 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7754 find_zone_movable_pfns_for_nodes();
7756 /* Print out the zone ranges */
7757 pr_info("Zone ranges:\n");
7758 for (i = 0; i < MAX_NR_ZONES; i++) {
7759 if (i == ZONE_MOVABLE)
7761 pr_info(" %-8s ", zone_names[i]);
7762 if (arch_zone_lowest_possible_pfn[i] ==
7763 arch_zone_highest_possible_pfn[i])
7766 pr_cont("[mem %#018Lx-%#018Lx]\n",
7767 (u64)arch_zone_lowest_possible_pfn[i]
7769 ((u64)arch_zone_highest_possible_pfn[i]
7770 << PAGE_SHIFT) - 1);
7773 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7774 pr_info("Movable zone start for each node\n");
7775 for (i = 0; i < MAX_NUMNODES; i++) {
7776 if (zone_movable_pfn[i])
7777 pr_info(" Node %d: %#018Lx\n", i,
7778 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7782 * Print out the early node map, and initialize the
7783 * subsection-map relative to active online memory ranges to
7784 * enable future "sub-section" extensions of the memory map.
7786 pr_info("Early memory node ranges\n");
7787 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7788 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7789 (u64)start_pfn << PAGE_SHIFT,
7790 ((u64)end_pfn << PAGE_SHIFT) - 1);
7791 subsection_map_init(start_pfn, end_pfn - start_pfn);
7794 /* Initialise every node */
7795 mminit_verify_pageflags_layout();
7796 setup_nr_node_ids();
7797 for_each_online_node(nid) {
7798 pg_data_t *pgdat = NODE_DATA(nid);
7799 free_area_init_node(nid);
7801 /* Any memory on that node */
7802 if (pgdat->node_present_pages)
7803 node_set_state(nid, N_MEMORY);
7804 check_for_memory(pgdat, nid);
7810 static int __init cmdline_parse_core(char *p, unsigned long *core,
7811 unsigned long *percent)
7813 unsigned long long coremem;
7819 /* Value may be a percentage of total memory, otherwise bytes */
7820 coremem = simple_strtoull(p, &endptr, 0);
7821 if (*endptr == '%') {
7822 /* Paranoid check for percent values greater than 100 */
7823 WARN_ON(coremem > 100);
7827 coremem = memparse(p, &p);
7828 /* Paranoid check that UL is enough for the coremem value */
7829 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7831 *core = coremem >> PAGE_SHIFT;
7838 * kernelcore=size sets the amount of memory for use for allocations that
7839 * cannot be reclaimed or migrated.
7841 static int __init cmdline_parse_kernelcore(char *p)
7843 /* parse kernelcore=mirror */
7844 if (parse_option_str(p, "mirror")) {
7845 mirrored_kernelcore = true;
7849 return cmdline_parse_core(p, &required_kernelcore,
7850 &required_kernelcore_percent);
7854 * movablecore=size sets the amount of memory for use for allocations that
7855 * can be reclaimed or migrated.
7857 static int __init cmdline_parse_movablecore(char *p)
7859 return cmdline_parse_core(p, &required_movablecore,
7860 &required_movablecore_percent);
7863 early_param("kernelcore", cmdline_parse_kernelcore);
7864 early_param("movablecore", cmdline_parse_movablecore);
7866 void adjust_managed_page_count(struct page *page, long count)
7868 atomic_long_add(count, &page_zone(page)->managed_pages);
7869 totalram_pages_add(count);
7870 #ifdef CONFIG_HIGHMEM
7871 if (PageHighMem(page))
7872 totalhigh_pages_add(count);
7875 EXPORT_SYMBOL(adjust_managed_page_count);
7877 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7880 unsigned long pages = 0;
7882 start = (void *)PAGE_ALIGN((unsigned long)start);
7883 end = (void *)((unsigned long)end & PAGE_MASK);
7884 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7885 struct page *page = virt_to_page(pos);
7886 void *direct_map_addr;
7889 * 'direct_map_addr' might be different from 'pos'
7890 * because some architectures' virt_to_page()
7891 * work with aliases. Getting the direct map
7892 * address ensures that we get a _writeable_
7893 * alias for the memset().
7895 direct_map_addr = page_address(page);
7897 * Perform a kasan-unchecked memset() since this memory
7898 * has not been initialized.
7900 direct_map_addr = kasan_reset_tag(direct_map_addr);
7901 if ((unsigned int)poison <= 0xFF)
7902 memset(direct_map_addr, poison, PAGE_SIZE);
7904 free_reserved_page(page);
7908 pr_info("Freeing %s memory: %ldK\n",
7909 s, pages << (PAGE_SHIFT - 10));
7914 void __init mem_init_print_info(void)
7916 unsigned long physpages, codesize, datasize, rosize, bss_size;
7917 unsigned long init_code_size, init_data_size;
7919 physpages = get_num_physpages();
7920 codesize = _etext - _stext;
7921 datasize = _edata - _sdata;
7922 rosize = __end_rodata - __start_rodata;
7923 bss_size = __bss_stop - __bss_start;
7924 init_data_size = __init_end - __init_begin;
7925 init_code_size = _einittext - _sinittext;
7928 * Detect special cases and adjust section sizes accordingly:
7929 * 1) .init.* may be embedded into .data sections
7930 * 2) .init.text.* may be out of [__init_begin, __init_end],
7931 * please refer to arch/tile/kernel/vmlinux.lds.S.
7932 * 3) .rodata.* may be embedded into .text or .data sections.
7934 #define adj_init_size(start, end, size, pos, adj) \
7936 if (start <= pos && pos < end && size > adj) \
7940 adj_init_size(__init_begin, __init_end, init_data_size,
7941 _sinittext, init_code_size);
7942 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7943 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7944 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7945 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7947 #undef adj_init_size
7949 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7950 #ifdef CONFIG_HIGHMEM
7954 nr_free_pages() << (PAGE_SHIFT - 10),
7955 physpages << (PAGE_SHIFT - 10),
7956 codesize >> 10, datasize >> 10, rosize >> 10,
7957 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7958 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7959 totalcma_pages << (PAGE_SHIFT - 10)
7960 #ifdef CONFIG_HIGHMEM
7961 , totalhigh_pages() << (PAGE_SHIFT - 10)
7967 * set_dma_reserve - set the specified number of pages reserved in the first zone
7968 * @new_dma_reserve: The number of pages to mark reserved
7970 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7971 * In the DMA zone, a significant percentage may be consumed by kernel image
7972 * and other unfreeable allocations which can skew the watermarks badly. This
7973 * function may optionally be used to account for unfreeable pages in the
7974 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7975 * smaller per-cpu batchsize.
7977 void __init set_dma_reserve(unsigned long new_dma_reserve)
7979 dma_reserve = new_dma_reserve;
7982 static int page_alloc_cpu_dead(unsigned int cpu)
7985 lru_add_drain_cpu(cpu);
7989 * Spill the event counters of the dead processor
7990 * into the current processors event counters.
7991 * This artificially elevates the count of the current
7994 vm_events_fold_cpu(cpu);
7997 * Zero the differential counters of the dead processor
7998 * so that the vm statistics are consistent.
8000 * This is only okay since the processor is dead and cannot
8001 * race with what we are doing.
8003 cpu_vm_stats_fold(cpu);
8008 int hashdist = HASHDIST_DEFAULT;
8010 static int __init set_hashdist(char *str)
8014 hashdist = simple_strtoul(str, &str, 0);
8017 __setup("hashdist=", set_hashdist);
8020 void __init page_alloc_init(void)
8025 if (num_node_state(N_MEMORY) == 1)
8029 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
8030 "mm/page_alloc:dead", NULL,
8031 page_alloc_cpu_dead);
8036 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8037 * or min_free_kbytes changes.
8039 static void calculate_totalreserve_pages(void)
8041 struct pglist_data *pgdat;
8042 unsigned long reserve_pages = 0;
8043 enum zone_type i, j;
8045 for_each_online_pgdat(pgdat) {
8047 pgdat->totalreserve_pages = 0;
8049 for (i = 0; i < MAX_NR_ZONES; i++) {
8050 struct zone *zone = pgdat->node_zones + i;
8052 unsigned long managed_pages = zone_managed_pages(zone);
8054 /* Find valid and maximum lowmem_reserve in the zone */
8055 for (j = i; j < MAX_NR_ZONES; j++) {
8056 if (zone->lowmem_reserve[j] > max)
8057 max = zone->lowmem_reserve[j];
8060 /* we treat the high watermark as reserved pages. */
8061 max += high_wmark_pages(zone);
8063 if (max > managed_pages)
8064 max = managed_pages;
8066 pgdat->totalreserve_pages += max;
8068 reserve_pages += max;
8071 totalreserve_pages = reserve_pages;
8075 * setup_per_zone_lowmem_reserve - called whenever
8076 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8077 * has a correct pages reserved value, so an adequate number of
8078 * pages are left in the zone after a successful __alloc_pages().
8080 static void setup_per_zone_lowmem_reserve(void)
8082 struct pglist_data *pgdat;
8083 enum zone_type i, j;
8085 for_each_online_pgdat(pgdat) {
8086 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8087 struct zone *zone = &pgdat->node_zones[i];
8088 int ratio = sysctl_lowmem_reserve_ratio[i];
8089 bool clear = !ratio || !zone_managed_pages(zone);
8090 unsigned long managed_pages = 0;
8092 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8094 zone->lowmem_reserve[j] = 0;
8096 struct zone *upper_zone = &pgdat->node_zones[j];
8098 managed_pages += zone_managed_pages(upper_zone);
8099 zone->lowmem_reserve[j] = managed_pages / ratio;
8105 /* update totalreserve_pages */
8106 calculate_totalreserve_pages();
8109 static void __setup_per_zone_wmarks(void)
8111 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8112 unsigned long lowmem_pages = 0;
8114 unsigned long flags;
8116 /* Calculate total number of !ZONE_HIGHMEM pages */
8117 for_each_zone(zone) {
8118 if (!is_highmem(zone))
8119 lowmem_pages += zone_managed_pages(zone);
8122 for_each_zone(zone) {
8125 spin_lock_irqsave(&zone->lock, flags);
8126 tmp = (u64)pages_min * zone_managed_pages(zone);
8127 do_div(tmp, lowmem_pages);
8128 if (is_highmem(zone)) {
8130 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8131 * need highmem pages, so cap pages_min to a small
8134 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8135 * deltas control async page reclaim, and so should
8136 * not be capped for highmem.
8138 unsigned long min_pages;
8140 min_pages = zone_managed_pages(zone) / 1024;
8141 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8142 zone->_watermark[WMARK_MIN] = min_pages;
8145 * If it's a lowmem zone, reserve a number of pages
8146 * proportionate to the zone's size.
8148 zone->_watermark[WMARK_MIN] = tmp;
8152 * Set the kswapd watermarks distance according to the
8153 * scale factor in proportion to available memory, but
8154 * ensure a minimum size on small systems.
8156 tmp = max_t(u64, tmp >> 2,
8157 mult_frac(zone_managed_pages(zone),
8158 watermark_scale_factor, 10000));
8160 zone->watermark_boost = 0;
8161 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8162 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8164 spin_unlock_irqrestore(&zone->lock, flags);
8167 /* update totalreserve_pages */
8168 calculate_totalreserve_pages();
8172 * setup_per_zone_wmarks - called when min_free_kbytes changes
8173 * or when memory is hot-{added|removed}
8175 * Ensures that the watermark[min,low,high] values for each zone are set
8176 * correctly with respect to min_free_kbytes.
8178 void setup_per_zone_wmarks(void)
8180 static DEFINE_SPINLOCK(lock);
8183 __setup_per_zone_wmarks();
8188 * Initialise min_free_kbytes.
8190 * For small machines we want it small (128k min). For large machines
8191 * we want it large (256MB max). But it is not linear, because network
8192 * bandwidth does not increase linearly with machine size. We use
8194 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8195 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8211 int __meminit init_per_zone_wmark_min(void)
8213 unsigned long lowmem_kbytes;
8214 int new_min_free_kbytes;
8216 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8217 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8219 if (new_min_free_kbytes > user_min_free_kbytes) {
8220 min_free_kbytes = new_min_free_kbytes;
8221 if (min_free_kbytes < 128)
8222 min_free_kbytes = 128;
8223 if (min_free_kbytes > 262144)
8224 min_free_kbytes = 262144;
8226 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8227 new_min_free_kbytes, user_min_free_kbytes);
8229 setup_per_zone_wmarks();
8230 refresh_zone_stat_thresholds();
8231 setup_per_zone_lowmem_reserve();
8234 setup_min_unmapped_ratio();
8235 setup_min_slab_ratio();
8238 khugepaged_min_free_kbytes_update();
8242 postcore_initcall(init_per_zone_wmark_min)
8245 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8246 * that we can call two helper functions whenever min_free_kbytes
8249 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8250 void *buffer, size_t *length, loff_t *ppos)
8254 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8259 user_min_free_kbytes = min_free_kbytes;
8260 setup_per_zone_wmarks();
8265 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8266 void *buffer, size_t *length, loff_t *ppos)
8270 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8275 setup_per_zone_wmarks();
8281 static void setup_min_unmapped_ratio(void)
8286 for_each_online_pgdat(pgdat)
8287 pgdat->min_unmapped_pages = 0;
8290 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8291 sysctl_min_unmapped_ratio) / 100;
8295 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8296 void *buffer, size_t *length, loff_t *ppos)
8300 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8304 setup_min_unmapped_ratio();
8309 static void setup_min_slab_ratio(void)
8314 for_each_online_pgdat(pgdat)
8315 pgdat->min_slab_pages = 0;
8318 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8319 sysctl_min_slab_ratio) / 100;
8322 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8323 void *buffer, size_t *length, loff_t *ppos)
8327 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8331 setup_min_slab_ratio();
8338 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8339 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8340 * whenever sysctl_lowmem_reserve_ratio changes.
8342 * The reserve ratio obviously has absolutely no relation with the
8343 * minimum watermarks. The lowmem reserve ratio can only make sense
8344 * if in function of the boot time zone sizes.
8346 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8347 void *buffer, size_t *length, loff_t *ppos)
8351 proc_dointvec_minmax(table, write, buffer, length, ppos);
8353 for (i = 0; i < MAX_NR_ZONES; i++) {
8354 if (sysctl_lowmem_reserve_ratio[i] < 1)
8355 sysctl_lowmem_reserve_ratio[i] = 0;
8358 setup_per_zone_lowmem_reserve();
8363 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8364 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8365 * pagelist can have before it gets flushed back to buddy allocator.
8367 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8368 void *buffer, size_t *length, loff_t *ppos)
8371 int old_percpu_pagelist_fraction;
8374 mutex_lock(&pcp_batch_high_lock);
8375 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8377 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8378 if (!write || ret < 0)
8381 /* Sanity checking to avoid pcp imbalance */
8382 if (percpu_pagelist_fraction &&
8383 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8384 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8390 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8393 for_each_populated_zone(zone)
8394 zone_set_pageset_high_and_batch(zone);
8396 mutex_unlock(&pcp_batch_high_lock);
8400 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8402 * Returns the number of pages that arch has reserved but
8403 * is not known to alloc_large_system_hash().
8405 static unsigned long __init arch_reserved_kernel_pages(void)
8412 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8413 * machines. As memory size is increased the scale is also increased but at
8414 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8415 * quadruples the scale is increased by one, which means the size of hash table
8416 * only doubles, instead of quadrupling as well.
8417 * Because 32-bit systems cannot have large physical memory, where this scaling
8418 * makes sense, it is disabled on such platforms.
8420 #if __BITS_PER_LONG > 32
8421 #define ADAPT_SCALE_BASE (64ul << 30)
8422 #define ADAPT_SCALE_SHIFT 2
8423 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8427 * allocate a large system hash table from bootmem
8428 * - it is assumed that the hash table must contain an exact power-of-2
8429 * quantity of entries
8430 * - limit is the number of hash buckets, not the total allocation size
8432 void *__init alloc_large_system_hash(const char *tablename,
8433 unsigned long bucketsize,
8434 unsigned long numentries,
8437 unsigned int *_hash_shift,
8438 unsigned int *_hash_mask,
8439 unsigned long low_limit,
8440 unsigned long high_limit)
8442 unsigned long long max = high_limit;
8443 unsigned long log2qty, size;
8449 /* allow the kernel cmdline to have a say */
8451 /* round applicable memory size up to nearest megabyte */
8452 numentries = nr_kernel_pages;
8453 numentries -= arch_reserved_kernel_pages();
8455 /* It isn't necessary when PAGE_SIZE >= 1MB */
8456 if (PAGE_SHIFT < 20)
8457 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8459 #if __BITS_PER_LONG > 32
8461 unsigned long adapt;
8463 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8464 adapt <<= ADAPT_SCALE_SHIFT)
8469 /* limit to 1 bucket per 2^scale bytes of low memory */
8470 if (scale > PAGE_SHIFT)
8471 numentries >>= (scale - PAGE_SHIFT);
8473 numentries <<= (PAGE_SHIFT - scale);
8475 /* Make sure we've got at least a 0-order allocation.. */
8476 if (unlikely(flags & HASH_SMALL)) {
8477 /* Makes no sense without HASH_EARLY */
8478 WARN_ON(!(flags & HASH_EARLY));
8479 if (!(numentries >> *_hash_shift)) {
8480 numentries = 1UL << *_hash_shift;
8481 BUG_ON(!numentries);
8483 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8484 numentries = PAGE_SIZE / bucketsize;
8486 numentries = roundup_pow_of_two(numentries);
8488 /* limit allocation size to 1/16 total memory by default */
8490 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8491 do_div(max, bucketsize);
8493 max = min(max, 0x80000000ULL);
8495 if (numentries < low_limit)
8496 numentries = low_limit;
8497 if (numentries > max)
8500 log2qty = ilog2(numentries);
8502 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8505 size = bucketsize << log2qty;
8506 if (flags & HASH_EARLY) {
8507 if (flags & HASH_ZERO)
8508 table = memblock_alloc(size, SMP_CACHE_BYTES);
8510 table = memblock_alloc_raw(size,
8512 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8513 table = __vmalloc(size, gfp_flags);
8515 huge = is_vm_area_hugepages(table);
8518 * If bucketsize is not a power-of-two, we may free
8519 * some pages at the end of hash table which
8520 * alloc_pages_exact() automatically does
8522 table = alloc_pages_exact(size, gfp_flags);
8523 kmemleak_alloc(table, size, 1, gfp_flags);
8525 } while (!table && size > PAGE_SIZE && --log2qty);
8528 panic("Failed to allocate %s hash table\n", tablename);
8530 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8531 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8532 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8535 *_hash_shift = log2qty;
8537 *_hash_mask = (1 << log2qty) - 1;
8543 * This function checks whether pageblock includes unmovable pages or not.
8545 * PageLRU check without isolation or lru_lock could race so that
8546 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8547 * check without lock_page also may miss some movable non-lru pages at
8548 * race condition. So you can't expect this function should be exact.
8550 * Returns a page without holding a reference. If the caller wants to
8551 * dereference that page (e.g., dumping), it has to make sure that it
8552 * cannot get removed (e.g., via memory unplug) concurrently.
8555 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8556 int migratetype, int flags)
8558 unsigned long iter = 0;
8559 unsigned long pfn = page_to_pfn(page);
8560 unsigned long offset = pfn % pageblock_nr_pages;
8562 if (is_migrate_cma_page(page)) {
8564 * CMA allocations (alloc_contig_range) really need to mark
8565 * isolate CMA pageblocks even when they are not movable in fact
8566 * so consider them movable here.
8568 if (is_migrate_cma(migratetype))
8574 for (; iter < pageblock_nr_pages - offset; iter++) {
8575 if (!pfn_valid_within(pfn + iter))
8578 page = pfn_to_page(pfn + iter);
8581 * Both, bootmem allocations and memory holes are marked
8582 * PG_reserved and are unmovable. We can even have unmovable
8583 * allocations inside ZONE_MOVABLE, for example when
8584 * specifying "movablecore".
8586 if (PageReserved(page))
8590 * If the zone is movable and we have ruled out all reserved
8591 * pages then it should be reasonably safe to assume the rest
8594 if (zone_idx(zone) == ZONE_MOVABLE)
8598 * Hugepages are not in LRU lists, but they're movable.
8599 * THPs are on the LRU, but need to be counted as #small pages.
8600 * We need not scan over tail pages because we don't
8601 * handle each tail page individually in migration.
8603 if (PageHuge(page) || PageTransCompound(page)) {
8604 struct page *head = compound_head(page);
8605 unsigned int skip_pages;
8607 if (PageHuge(page)) {
8608 if (!hugepage_migration_supported(page_hstate(head)))
8610 } else if (!PageLRU(head) && !__PageMovable(head)) {
8614 skip_pages = compound_nr(head) - (page - head);
8615 iter += skip_pages - 1;
8620 * We can't use page_count without pin a page
8621 * because another CPU can free compound page.
8622 * This check already skips compound tails of THP
8623 * because their page->_refcount is zero at all time.
8625 if (!page_ref_count(page)) {
8626 if (PageBuddy(page))
8627 iter += (1 << buddy_order(page)) - 1;
8632 * The HWPoisoned page may be not in buddy system, and
8633 * page_count() is not 0.
8635 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8639 * We treat all PageOffline() pages as movable when offlining
8640 * to give drivers a chance to decrement their reference count
8641 * in MEM_GOING_OFFLINE in order to indicate that these pages
8642 * can be offlined as there are no direct references anymore.
8643 * For actually unmovable PageOffline() where the driver does
8644 * not support this, we will fail later when trying to actually
8645 * move these pages that still have a reference count > 0.
8646 * (false negatives in this function only)
8648 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8651 if (__PageMovable(page) || PageLRU(page))
8655 * If there are RECLAIMABLE pages, we need to check
8656 * it. But now, memory offline itself doesn't call
8657 * shrink_node_slabs() and it still to be fixed.
8664 #ifdef CONFIG_CONTIG_ALLOC
8665 static unsigned long pfn_max_align_down(unsigned long pfn)
8667 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8668 pageblock_nr_pages) - 1);
8671 static unsigned long pfn_max_align_up(unsigned long pfn)
8673 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8674 pageblock_nr_pages));
8677 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8678 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8679 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8680 static void alloc_contig_dump_pages(struct list_head *page_list)
8682 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8684 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8688 list_for_each_entry(page, page_list, lru)
8689 dump_page(page, "migration failure");
8693 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8698 /* [start, end) must belong to a single zone. */
8699 static int __alloc_contig_migrate_range(struct compact_control *cc,
8700 unsigned long start, unsigned long end)
8702 /* This function is based on compact_zone() from compaction.c. */
8703 unsigned int nr_reclaimed;
8704 unsigned long pfn = start;
8705 unsigned int tries = 0;
8707 struct migration_target_control mtc = {
8708 .nid = zone_to_nid(cc->zone),
8709 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8712 lru_cache_disable();
8714 while (pfn < end || !list_empty(&cc->migratepages)) {
8715 if (fatal_signal_pending(current)) {
8720 if (list_empty(&cc->migratepages)) {
8721 cc->nr_migratepages = 0;
8722 ret = isolate_migratepages_range(cc, pfn, end);
8723 if (ret && ret != -EAGAIN)
8725 pfn = cc->migrate_pfn;
8727 } else if (++tries == 5) {
8732 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8734 cc->nr_migratepages -= nr_reclaimed;
8736 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8737 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8740 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
8741 * to retry again over this error, so do the same here.
8749 alloc_contig_dump_pages(&cc->migratepages);
8750 putback_movable_pages(&cc->migratepages);
8757 * alloc_contig_range() -- tries to allocate given range of pages
8758 * @start: start PFN to allocate
8759 * @end: one-past-the-last PFN to allocate
8760 * @migratetype: migratetype of the underlying pageblocks (either
8761 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8762 * in range must have the same migratetype and it must
8763 * be either of the two.
8764 * @gfp_mask: GFP mask to use during compaction
8766 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8767 * aligned. The PFN range must belong to a single zone.
8769 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8770 * pageblocks in the range. Once isolated, the pageblocks should not
8771 * be modified by others.
8773 * Return: zero on success or negative error code. On success all
8774 * pages which PFN is in [start, end) are allocated for the caller and
8775 * need to be freed with free_contig_range().
8777 int alloc_contig_range(unsigned long start, unsigned long end,
8778 unsigned migratetype, gfp_t gfp_mask)
8780 unsigned long outer_start, outer_end;
8784 struct compact_control cc = {
8785 .nr_migratepages = 0,
8787 .zone = page_zone(pfn_to_page(start)),
8788 .mode = MIGRATE_SYNC,
8789 .ignore_skip_hint = true,
8790 .no_set_skip_hint = true,
8791 .gfp_mask = current_gfp_context(gfp_mask),
8792 .alloc_contig = true,
8794 INIT_LIST_HEAD(&cc.migratepages);
8797 * What we do here is we mark all pageblocks in range as
8798 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8799 * have different sizes, and due to the way page allocator
8800 * work, we align the range to biggest of the two pages so
8801 * that page allocator won't try to merge buddies from
8802 * different pageblocks and change MIGRATE_ISOLATE to some
8803 * other migration type.
8805 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8806 * migrate the pages from an unaligned range (ie. pages that
8807 * we are interested in). This will put all the pages in
8808 * range back to page allocator as MIGRATE_ISOLATE.
8810 * When this is done, we take the pages in range from page
8811 * allocator removing them from the buddy system. This way
8812 * page allocator will never consider using them.
8814 * This lets us mark the pageblocks back as
8815 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8816 * aligned range but not in the unaligned, original range are
8817 * put back to page allocator so that buddy can use them.
8820 ret = start_isolate_page_range(pfn_max_align_down(start),
8821 pfn_max_align_up(end), migratetype, 0);
8825 drain_all_pages(cc.zone);
8828 * In case of -EBUSY, we'd like to know which page causes problem.
8829 * So, just fall through. test_pages_isolated() has a tracepoint
8830 * which will report the busy page.
8832 * It is possible that busy pages could become available before
8833 * the call to test_pages_isolated, and the range will actually be
8834 * allocated. So, if we fall through be sure to clear ret so that
8835 * -EBUSY is not accidentally used or returned to caller.
8837 ret = __alloc_contig_migrate_range(&cc, start, end);
8838 if (ret && ret != -EBUSY)
8843 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8844 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8845 * more, all pages in [start, end) are free in page allocator.
8846 * What we are going to do is to allocate all pages from
8847 * [start, end) (that is remove them from page allocator).
8849 * The only problem is that pages at the beginning and at the
8850 * end of interesting range may be not aligned with pages that
8851 * page allocator holds, ie. they can be part of higher order
8852 * pages. Because of this, we reserve the bigger range and
8853 * once this is done free the pages we are not interested in.
8855 * We don't have to hold zone->lock here because the pages are
8856 * isolated thus they won't get removed from buddy.
8860 outer_start = start;
8861 while (!PageBuddy(pfn_to_page(outer_start))) {
8862 if (++order >= MAX_ORDER) {
8863 outer_start = start;
8866 outer_start &= ~0UL << order;
8869 if (outer_start != start) {
8870 order = buddy_order(pfn_to_page(outer_start));
8873 * outer_start page could be small order buddy page and
8874 * it doesn't include start page. Adjust outer_start
8875 * in this case to report failed page properly
8876 * on tracepoint in test_pages_isolated()
8878 if (outer_start + (1UL << order) <= start)
8879 outer_start = start;
8882 /* Make sure the range is really isolated. */
8883 if (test_pages_isolated(outer_start, end, 0)) {
8888 /* Grab isolated pages from freelists. */
8889 outer_end = isolate_freepages_range(&cc, outer_start, end);
8895 /* Free head and tail (if any) */
8896 if (start != outer_start)
8897 free_contig_range(outer_start, start - outer_start);
8898 if (end != outer_end)
8899 free_contig_range(end, outer_end - end);
8902 undo_isolate_page_range(pfn_max_align_down(start),
8903 pfn_max_align_up(end), migratetype);
8906 EXPORT_SYMBOL(alloc_contig_range);
8908 static int __alloc_contig_pages(unsigned long start_pfn,
8909 unsigned long nr_pages, gfp_t gfp_mask)
8911 unsigned long end_pfn = start_pfn + nr_pages;
8913 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8917 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8918 unsigned long nr_pages)
8920 unsigned long i, end_pfn = start_pfn + nr_pages;
8923 for (i = start_pfn; i < end_pfn; i++) {
8924 page = pfn_to_online_page(i);
8928 if (page_zone(page) != z)
8931 if (PageReserved(page))
8937 static bool zone_spans_last_pfn(const struct zone *zone,
8938 unsigned long start_pfn, unsigned long nr_pages)
8940 unsigned long last_pfn = start_pfn + nr_pages - 1;
8942 return zone_spans_pfn(zone, last_pfn);
8946 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8947 * @nr_pages: Number of contiguous pages to allocate
8948 * @gfp_mask: GFP mask to limit search and used during compaction
8950 * @nodemask: Mask for other possible nodes
8952 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8953 * on an applicable zonelist to find a contiguous pfn range which can then be
8954 * tried for allocation with alloc_contig_range(). This routine is intended
8955 * for allocation requests which can not be fulfilled with the buddy allocator.
8957 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8958 * power of two then the alignment is guaranteed to be to the given nr_pages
8959 * (e.g. 1GB request would be aligned to 1GB).
8961 * Allocated pages can be freed with free_contig_range() or by manually calling
8962 * __free_page() on each allocated page.
8964 * Return: pointer to contiguous pages on success, or NULL if not successful.
8966 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8967 int nid, nodemask_t *nodemask)
8969 unsigned long ret, pfn, flags;
8970 struct zonelist *zonelist;
8974 zonelist = node_zonelist(nid, gfp_mask);
8975 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8976 gfp_zone(gfp_mask), nodemask) {
8977 spin_lock_irqsave(&zone->lock, flags);
8979 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8980 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8981 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8983 * We release the zone lock here because
8984 * alloc_contig_range() will also lock the zone
8985 * at some point. If there's an allocation
8986 * spinning on this lock, it may win the race
8987 * and cause alloc_contig_range() to fail...
8989 spin_unlock_irqrestore(&zone->lock, flags);
8990 ret = __alloc_contig_pages(pfn, nr_pages,
8993 return pfn_to_page(pfn);
8994 spin_lock_irqsave(&zone->lock, flags);
8998 spin_unlock_irqrestore(&zone->lock, flags);
9002 #endif /* CONFIG_CONTIG_ALLOC */
9004 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9006 unsigned long count = 0;
9008 for (; nr_pages--; pfn++) {
9009 struct page *page = pfn_to_page(pfn);
9011 count += page_count(page) != 1;
9014 WARN(count != 0, "%lu pages are still in use!\n", count);
9016 EXPORT_SYMBOL(free_contig_range);
9019 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9020 * page high values need to be recalculated.
9022 void __meminit zone_pcp_update(struct zone *zone)
9024 mutex_lock(&pcp_batch_high_lock);
9025 zone_set_pageset_high_and_batch(zone);
9026 mutex_unlock(&pcp_batch_high_lock);
9030 * Effectively disable pcplists for the zone by setting the high limit to 0
9031 * and draining all cpus. A concurrent page freeing on another CPU that's about
9032 * to put the page on pcplist will either finish before the drain and the page
9033 * will be drained, or observe the new high limit and skip the pcplist.
9035 * Must be paired with a call to zone_pcp_enable().
9037 void zone_pcp_disable(struct zone *zone)
9039 mutex_lock(&pcp_batch_high_lock);
9040 __zone_set_pageset_high_and_batch(zone, 0, 1);
9041 __drain_all_pages(zone, true);
9044 void zone_pcp_enable(struct zone *zone)
9046 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9047 mutex_unlock(&pcp_batch_high_lock);
9050 void zone_pcp_reset(struct zone *zone)
9053 struct per_cpu_pageset *pset;
9055 if (zone->pageset != &boot_pageset) {
9056 for_each_online_cpu(cpu) {
9057 pset = per_cpu_ptr(zone->pageset, cpu);
9058 drain_zonestat(zone, pset);
9060 free_percpu(zone->pageset);
9061 zone->pageset = &boot_pageset;
9065 #ifdef CONFIG_MEMORY_HOTREMOVE
9067 * All pages in the range must be in a single zone, must not contain holes,
9068 * must span full sections, and must be isolated before calling this function.
9070 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9072 unsigned long pfn = start_pfn;
9076 unsigned long flags;
9078 offline_mem_sections(pfn, end_pfn);
9079 zone = page_zone(pfn_to_page(pfn));
9080 spin_lock_irqsave(&zone->lock, flags);
9081 while (pfn < end_pfn) {
9082 page = pfn_to_page(pfn);
9084 * The HWPoisoned page may be not in buddy system, and
9085 * page_count() is not 0.
9087 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9092 * At this point all remaining PageOffline() pages have a
9093 * reference count of 0 and can simply be skipped.
9095 if (PageOffline(page)) {
9096 BUG_ON(page_count(page));
9097 BUG_ON(PageBuddy(page));
9102 BUG_ON(page_count(page));
9103 BUG_ON(!PageBuddy(page));
9104 order = buddy_order(page);
9105 del_page_from_free_list(page, zone, order);
9106 pfn += (1 << order);
9108 spin_unlock_irqrestore(&zone->lock, flags);
9112 bool is_free_buddy_page(struct page *page)
9114 struct zone *zone = page_zone(page);
9115 unsigned long pfn = page_to_pfn(page);
9116 unsigned long flags;
9119 spin_lock_irqsave(&zone->lock, flags);
9120 for (order = 0; order < MAX_ORDER; order++) {
9121 struct page *page_head = page - (pfn & ((1 << order) - 1));
9123 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
9126 spin_unlock_irqrestore(&zone->lock, flags);
9128 return order < MAX_ORDER;
9131 #ifdef CONFIG_MEMORY_FAILURE
9133 * Break down a higher-order page in sub-pages, and keep our target out of
9136 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9137 struct page *target, int low, int high,
9140 unsigned long size = 1 << high;
9141 struct page *current_buddy, *next_page;
9143 while (high > low) {
9147 if (target >= &page[size]) {
9148 next_page = page + size;
9149 current_buddy = page;
9152 current_buddy = page + size;
9155 if (set_page_guard(zone, current_buddy, high, migratetype))
9158 if (current_buddy != target) {
9159 add_to_free_list(current_buddy, zone, high, migratetype);
9160 set_buddy_order(current_buddy, high);
9167 * Take a page that will be marked as poisoned off the buddy allocator.
9169 bool take_page_off_buddy(struct page *page)
9171 struct zone *zone = page_zone(page);
9172 unsigned long pfn = page_to_pfn(page);
9173 unsigned long flags;
9177 spin_lock_irqsave(&zone->lock, flags);
9178 for (order = 0; order < MAX_ORDER; order++) {
9179 struct page *page_head = page - (pfn & ((1 << order) - 1));
9180 int page_order = buddy_order(page_head);
9182 if (PageBuddy(page_head) && page_order >= order) {
9183 unsigned long pfn_head = page_to_pfn(page_head);
9184 int migratetype = get_pfnblock_migratetype(page_head,
9187 del_page_from_free_list(page_head, zone, page_order);
9188 break_down_buddy_pages(zone, page_head, page, 0,
9189 page_order, migratetype);
9190 if (!is_migrate_isolate(migratetype))
9191 __mod_zone_freepage_state(zone, -1, migratetype);
9195 if (page_count(page_head) > 0)
9198 spin_unlock_irqrestore(&zone->lock, flags);