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(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(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(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(struct page *page, unsigned long pfn,
526 return __get_pfnblock_flags_mask(page, pfn, mask);
529 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
531 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
535 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
536 * @page: The page within the block of interest
537 * @flags: The flags to set
538 * @pfn: The target page frame number
539 * @mask: mask of bits that the caller is interested in
541 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
545 unsigned long *bitmap;
546 unsigned long bitidx, word_bitidx;
547 unsigned long old_word, word;
549 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
550 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
552 bitmap = get_pageblock_bitmap(page, pfn);
553 bitidx = pfn_to_bitidx(page, pfn);
554 word_bitidx = bitidx / BITS_PER_LONG;
555 bitidx &= (BITS_PER_LONG-1);
557 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
562 word = READ_ONCE(bitmap[word_bitidx]);
564 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
565 if (word == old_word)
571 void set_pageblock_migratetype(struct page *page, int migratetype)
573 if (unlikely(page_group_by_mobility_disabled &&
574 migratetype < MIGRATE_PCPTYPES))
575 migratetype = MIGRATE_UNMOVABLE;
577 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
578 page_to_pfn(page), MIGRATETYPE_MASK);
581 #ifdef CONFIG_DEBUG_VM
582 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
586 unsigned long pfn = page_to_pfn(page);
587 unsigned long sp, start_pfn;
590 seq = zone_span_seqbegin(zone);
591 start_pfn = zone->zone_start_pfn;
592 sp = zone->spanned_pages;
593 if (!zone_spans_pfn(zone, pfn))
595 } while (zone_span_seqretry(zone, seq));
598 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
599 pfn, zone_to_nid(zone), zone->name,
600 start_pfn, start_pfn + sp);
605 static int page_is_consistent(struct zone *zone, struct page *page)
607 if (!pfn_valid_within(page_to_pfn(page)))
609 if (zone != page_zone(page))
615 * Temporary debugging check for pages not lying within a given zone.
617 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
619 if (page_outside_zone_boundaries(zone, page))
621 if (!page_is_consistent(zone, page))
627 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
633 static void bad_page(struct page *page, const char *reason)
635 static unsigned long resume;
636 static unsigned long nr_shown;
637 static unsigned long nr_unshown;
640 * Allow a burst of 60 reports, then keep quiet for that minute;
641 * or allow a steady drip of one report per second.
643 if (nr_shown == 60) {
644 if (time_before(jiffies, resume)) {
650 "BUG: Bad page state: %lu messages suppressed\n",
657 resume = jiffies + 60 * HZ;
659 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
660 current->comm, page_to_pfn(page));
661 __dump_page(page, reason);
662 dump_page_owner(page);
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 if (_init_on_alloc_enabled_early) {
790 if (page_poisoning_enabled())
791 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
792 "will take precedence over init_on_alloc\n");
794 static_branch_enable(&init_on_alloc);
796 if (_init_on_free_enabled_early) {
797 if (page_poisoning_enabled())
798 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
799 "will take precedence over init_on_free\n");
801 static_branch_enable(&init_on_free);
804 #ifdef CONFIG_PAGE_POISONING
806 * Page poisoning is debug page alloc for some arches. If
807 * either of those options are enabled, enable poisoning.
809 if (page_poisoning_enabled() ||
810 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
811 debug_pagealloc_enabled()))
812 static_branch_enable(&_page_poisoning_enabled);
815 #ifdef CONFIG_DEBUG_PAGEALLOC
816 if (!debug_pagealloc_enabled())
819 static_branch_enable(&_debug_pagealloc_enabled);
821 if (!debug_guardpage_minorder())
824 static_branch_enable(&_debug_guardpage_enabled);
828 static inline void set_buddy_order(struct page *page, unsigned int order)
830 set_page_private(page, order);
831 __SetPageBuddy(page);
835 * This function checks whether a page is free && is the buddy
836 * we can coalesce a page and its buddy if
837 * (a) the buddy is not in a hole (check before calling!) &&
838 * (b) the buddy is in the buddy system &&
839 * (c) a page and its buddy have the same order &&
840 * (d) a page and its buddy are in the same zone.
842 * For recording whether a page is in the buddy system, we set PageBuddy.
843 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
845 * For recording page's order, we use page_private(page).
847 static inline bool page_is_buddy(struct page *page, struct page *buddy,
850 if (!page_is_guard(buddy) && !PageBuddy(buddy))
853 if (buddy_order(buddy) != order)
857 * zone check is done late to avoid uselessly calculating
858 * zone/node ids for pages that could never merge.
860 if (page_zone_id(page) != page_zone_id(buddy))
863 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
868 #ifdef CONFIG_COMPACTION
869 static inline struct capture_control *task_capc(struct zone *zone)
871 struct capture_control *capc = current->capture_control;
873 return unlikely(capc) &&
874 !(current->flags & PF_KTHREAD) &&
876 capc->cc->zone == zone ? capc : NULL;
880 compaction_capture(struct capture_control *capc, struct page *page,
881 int order, int migratetype)
883 if (!capc || order != capc->cc->order)
886 /* Do not accidentally pollute CMA or isolated regions*/
887 if (is_migrate_cma(migratetype) ||
888 is_migrate_isolate(migratetype))
892 * Do not let lower order allocations polluate a movable pageblock.
893 * This might let an unmovable request use a reclaimable pageblock
894 * and vice-versa but no more than normal fallback logic which can
895 * have trouble finding a high-order free page.
897 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
905 static inline struct capture_control *task_capc(struct zone *zone)
911 compaction_capture(struct capture_control *capc, struct page *page,
912 int order, int migratetype)
916 #endif /* CONFIG_COMPACTION */
918 /* Used for pages not on another list */
919 static inline void add_to_free_list(struct page *page, struct zone *zone,
920 unsigned int order, int migratetype)
922 struct free_area *area = &zone->free_area[order];
924 list_add(&page->lru, &area->free_list[migratetype]);
928 /* Used for pages not on another list */
929 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
930 unsigned int order, int migratetype)
932 struct free_area *area = &zone->free_area[order];
934 list_add_tail(&page->lru, &area->free_list[migratetype]);
939 * Used for pages which are on another list. Move the pages to the tail
940 * of the list - so the moved pages won't immediately be considered for
941 * allocation again (e.g., optimization for memory onlining).
943 static inline void move_to_free_list(struct page *page, struct zone *zone,
944 unsigned int order, int migratetype)
946 struct free_area *area = &zone->free_area[order];
948 list_move_tail(&page->lru, &area->free_list[migratetype]);
951 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
954 /* clear reported state and update reported page count */
955 if (page_reported(page))
956 __ClearPageReported(page);
958 list_del(&page->lru);
959 __ClearPageBuddy(page);
960 set_page_private(page, 0);
961 zone->free_area[order].nr_free--;
965 * If this is not the largest possible page, check if the buddy
966 * of the next-highest order is free. If it is, it's possible
967 * that pages are being freed that will coalesce soon. In case,
968 * that is happening, add the free page to the tail of the list
969 * so it's less likely to be used soon and more likely to be merged
970 * as a higher order page
973 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
974 struct page *page, unsigned int order)
976 struct page *higher_page, *higher_buddy;
977 unsigned long combined_pfn;
979 if (order >= MAX_ORDER - 2)
982 if (!pfn_valid_within(buddy_pfn))
985 combined_pfn = buddy_pfn & pfn;
986 higher_page = page + (combined_pfn - pfn);
987 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
988 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
990 return pfn_valid_within(buddy_pfn) &&
991 page_is_buddy(higher_page, higher_buddy, order + 1);
995 * Freeing function for a buddy system allocator.
997 * The concept of a buddy system is to maintain direct-mapped table
998 * (containing bit values) for memory blocks of various "orders".
999 * The bottom level table contains the map for the smallest allocatable
1000 * units of memory (here, pages), and each level above it describes
1001 * pairs of units from the levels below, hence, "buddies".
1002 * At a high level, all that happens here is marking the table entry
1003 * at the bottom level available, and propagating the changes upward
1004 * as necessary, plus some accounting needed to play nicely with other
1005 * parts of the VM system.
1006 * At each level, we keep a list of pages, which are heads of continuous
1007 * free pages of length of (1 << order) and marked with PageBuddy.
1008 * Page's order is recorded in page_private(page) field.
1009 * So when we are allocating or freeing one, we can derive the state of the
1010 * other. That is, if we allocate a small block, and both were
1011 * free, the remainder of the region must be split into blocks.
1012 * If a block is freed, and its buddy is also free, then this
1013 * triggers coalescing into a block of larger size.
1018 static inline void __free_one_page(struct page *page,
1020 struct zone *zone, unsigned int order,
1021 int migratetype, fpi_t fpi_flags)
1023 struct capture_control *capc = task_capc(zone);
1024 unsigned long buddy_pfn;
1025 unsigned long combined_pfn;
1026 unsigned int max_order;
1030 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1032 VM_BUG_ON(!zone_is_initialized(zone));
1033 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1035 VM_BUG_ON(migratetype == -1);
1036 if (likely(!is_migrate_isolate(migratetype)))
1037 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1039 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1040 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1043 while (order < max_order) {
1044 if (compaction_capture(capc, page, order, migratetype)) {
1045 __mod_zone_freepage_state(zone, -(1 << order),
1049 buddy_pfn = __find_buddy_pfn(pfn, order);
1050 buddy = page + (buddy_pfn - pfn);
1052 if (!pfn_valid_within(buddy_pfn))
1054 if (!page_is_buddy(page, buddy, order))
1057 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1058 * merge with it and move up one order.
1060 if (page_is_guard(buddy))
1061 clear_page_guard(zone, buddy, order, migratetype);
1063 del_page_from_free_list(buddy, zone, order);
1064 combined_pfn = buddy_pfn & pfn;
1065 page = page + (combined_pfn - pfn);
1069 if (order < MAX_ORDER - 1) {
1070 /* If we are here, it means order is >= pageblock_order.
1071 * We want to prevent merge between freepages on isolate
1072 * pageblock and normal pageblock. Without this, pageblock
1073 * isolation could cause incorrect freepage or CMA accounting.
1075 * We don't want to hit this code for the more frequent
1076 * low-order merging.
1078 if (unlikely(has_isolate_pageblock(zone))) {
1081 buddy_pfn = __find_buddy_pfn(pfn, order);
1082 buddy = page + (buddy_pfn - pfn);
1083 buddy_mt = get_pageblock_migratetype(buddy);
1085 if (migratetype != buddy_mt
1086 && (is_migrate_isolate(migratetype) ||
1087 is_migrate_isolate(buddy_mt)))
1090 max_order = order + 1;
1091 goto continue_merging;
1095 set_buddy_order(page, order);
1097 if (fpi_flags & FPI_TO_TAIL)
1099 else if (is_shuffle_order(order))
1100 to_tail = shuffle_pick_tail();
1102 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1105 add_to_free_list_tail(page, zone, order, migratetype);
1107 add_to_free_list(page, zone, order, migratetype);
1109 /* Notify page reporting subsystem of freed page */
1110 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1111 page_reporting_notify_free(order);
1115 * A bad page could be due to a number of fields. Instead of multiple branches,
1116 * try and check multiple fields with one check. The caller must do a detailed
1117 * check if necessary.
1119 static inline bool page_expected_state(struct page *page,
1120 unsigned long check_flags)
1122 if (unlikely(atomic_read(&page->_mapcount) != -1))
1125 if (unlikely((unsigned long)page->mapping |
1126 page_ref_count(page) |
1130 (page->flags & check_flags)))
1136 static const char *page_bad_reason(struct page *page, unsigned long flags)
1138 const char *bad_reason = NULL;
1140 if (unlikely(atomic_read(&page->_mapcount) != -1))
1141 bad_reason = "nonzero mapcount";
1142 if (unlikely(page->mapping != NULL))
1143 bad_reason = "non-NULL mapping";
1144 if (unlikely(page_ref_count(page) != 0))
1145 bad_reason = "nonzero _refcount";
1146 if (unlikely(page->flags & flags)) {
1147 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1148 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1150 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1153 if (unlikely(page->memcg_data))
1154 bad_reason = "page still charged to cgroup";
1159 static void check_free_page_bad(struct page *page)
1162 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1165 static inline int check_free_page(struct page *page)
1167 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1170 /* Something has gone sideways, find it */
1171 check_free_page_bad(page);
1175 static int free_tail_pages_check(struct page *head_page, struct page *page)
1180 * We rely page->lru.next never has bit 0 set, unless the page
1181 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1183 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1185 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1189 switch (page - head_page) {
1191 /* the first tail page: ->mapping may be compound_mapcount() */
1192 if (unlikely(compound_mapcount(page))) {
1193 bad_page(page, "nonzero compound_mapcount");
1199 * the second tail page: ->mapping is
1200 * deferred_list.next -- ignore value.
1204 if (page->mapping != TAIL_MAPPING) {
1205 bad_page(page, "corrupted mapping in tail page");
1210 if (unlikely(!PageTail(page))) {
1211 bad_page(page, "PageTail not set");
1214 if (unlikely(compound_head(page) != head_page)) {
1215 bad_page(page, "compound_head not consistent");
1220 page->mapping = NULL;
1221 clear_compound_head(page);
1225 static void kernel_init_free_pages(struct page *page, int numpages)
1229 /* s390's use of memset() could override KASAN redzones. */
1230 kasan_disable_current();
1231 for (i = 0; i < numpages; i++) {
1232 u8 tag = page_kasan_tag(page + i);
1233 page_kasan_tag_reset(page + i);
1234 clear_highpage(page + i);
1235 page_kasan_tag_set(page + i, tag);
1237 kasan_enable_current();
1240 static __always_inline bool free_pages_prepare(struct page *page,
1241 unsigned int order, bool check_free, fpi_t fpi_flags)
1246 VM_BUG_ON_PAGE(PageTail(page), page);
1248 trace_mm_page_free(page, order);
1250 if (unlikely(PageHWPoison(page)) && !order) {
1252 * Do not let hwpoison pages hit pcplists/buddy
1253 * Untie memcg state and reset page's owner
1255 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1256 __memcg_kmem_uncharge_page(page, order);
1257 reset_page_owner(page, order);
1262 * Check tail pages before head page information is cleared to
1263 * avoid checking PageCompound for order-0 pages.
1265 if (unlikely(order)) {
1266 bool compound = PageCompound(page);
1269 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1272 ClearPageDoubleMap(page);
1273 for (i = 1; i < (1 << order); i++) {
1275 bad += free_tail_pages_check(page, page + i);
1276 if (unlikely(check_free_page(page + i))) {
1280 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1283 if (PageMappingFlags(page))
1284 page->mapping = NULL;
1285 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1286 __memcg_kmem_uncharge_page(page, order);
1288 bad += check_free_page(page);
1292 page_cpupid_reset_last(page);
1293 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1294 reset_page_owner(page, order);
1296 if (!PageHighMem(page)) {
1297 debug_check_no_locks_freed(page_address(page),
1298 PAGE_SIZE << order);
1299 debug_check_no_obj_freed(page_address(page),
1300 PAGE_SIZE << order);
1303 kernel_poison_pages(page, 1 << order);
1306 * As memory initialization might be integrated into KASAN,
1307 * kasan_free_pages and kernel_init_free_pages must be
1308 * kept together to avoid discrepancies in behavior.
1310 * With hardware tag-based KASAN, memory tags must be set before the
1311 * page becomes unavailable via debug_pagealloc or arch_free_page.
1313 init = want_init_on_free();
1314 if (init && !kasan_has_integrated_init())
1315 kernel_init_free_pages(page, 1 << order);
1316 kasan_free_nondeferred_pages(page, order, init, fpi_flags);
1319 * arch_free_page() can make the page's contents inaccessible. s390
1320 * does this. So nothing which can access the page's contents should
1321 * happen after this.
1323 arch_free_page(page, order);
1325 debug_pagealloc_unmap_pages(page, 1 << order);
1330 #ifdef CONFIG_DEBUG_VM
1332 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1333 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1334 * moved from pcp lists to free lists.
1336 static bool free_pcp_prepare(struct page *page)
1338 return free_pages_prepare(page, 0, true, FPI_NONE);
1341 static bool bulkfree_pcp_prepare(struct page *page)
1343 if (debug_pagealloc_enabled_static())
1344 return check_free_page(page);
1350 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1351 * moving from pcp lists to free list in order to reduce overhead. With
1352 * debug_pagealloc enabled, they are checked also immediately when being freed
1355 static bool free_pcp_prepare(struct page *page)
1357 if (debug_pagealloc_enabled_static())
1358 return free_pages_prepare(page, 0, true, FPI_NONE);
1360 return free_pages_prepare(page, 0, false, FPI_NONE);
1363 static bool bulkfree_pcp_prepare(struct page *page)
1365 return check_free_page(page);
1367 #endif /* CONFIG_DEBUG_VM */
1369 static inline void prefetch_buddy(struct page *page)
1371 unsigned long pfn = page_to_pfn(page);
1372 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1373 struct page *buddy = page + (buddy_pfn - pfn);
1379 * Frees a number of pages from the PCP lists
1380 * Assumes all pages on list are in same zone, and of same order.
1381 * count is the number of pages to free.
1383 * If the zone was previously in an "all pages pinned" state then look to
1384 * see if this freeing clears that state.
1386 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1387 * pinned" detection logic.
1389 static void free_pcppages_bulk(struct zone *zone, int count,
1390 struct per_cpu_pages *pcp)
1392 int migratetype = 0;
1394 int prefetch_nr = READ_ONCE(pcp->batch);
1395 bool isolated_pageblocks;
1396 struct page *page, *tmp;
1400 * Ensure proper count is passed which otherwise would stuck in the
1401 * below while (list_empty(list)) loop.
1403 count = min(pcp->count, count);
1405 struct list_head *list;
1408 * Remove pages from lists in a round-robin fashion. A
1409 * batch_free count is maintained that is incremented when an
1410 * empty list is encountered. This is so more pages are freed
1411 * off fuller lists instead of spinning excessively around empty
1416 if (++migratetype == MIGRATE_PCPTYPES)
1418 list = &pcp->lists[migratetype];
1419 } while (list_empty(list));
1421 /* This is the only non-empty list. Free them all. */
1422 if (batch_free == MIGRATE_PCPTYPES)
1426 page = list_last_entry(list, struct page, lru);
1427 /* must delete to avoid corrupting pcp list */
1428 list_del(&page->lru);
1431 if (bulkfree_pcp_prepare(page))
1434 list_add_tail(&page->lru, &head);
1437 * We are going to put the page back to the global
1438 * pool, prefetch its buddy to speed up later access
1439 * under zone->lock. It is believed the overhead of
1440 * an additional test and calculating buddy_pfn here
1441 * can be offset by reduced memory latency later. To
1442 * avoid excessive prefetching due to large count, only
1443 * prefetch buddy for the first pcp->batch nr of pages.
1446 prefetch_buddy(page);
1449 } while (--count && --batch_free && !list_empty(list));
1452 spin_lock(&zone->lock);
1453 isolated_pageblocks = has_isolate_pageblock(zone);
1456 * Use safe version since after __free_one_page(),
1457 * page->lru.next will not point to original list.
1459 list_for_each_entry_safe(page, tmp, &head, lru) {
1460 int mt = get_pcppage_migratetype(page);
1461 /* MIGRATE_ISOLATE page should not go to pcplists */
1462 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1463 /* Pageblock could have been isolated meanwhile */
1464 if (unlikely(isolated_pageblocks))
1465 mt = get_pageblock_migratetype(page);
1467 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1468 trace_mm_page_pcpu_drain(page, 0, mt);
1470 spin_unlock(&zone->lock);
1473 static void free_one_page(struct zone *zone,
1474 struct page *page, unsigned long pfn,
1476 int migratetype, fpi_t fpi_flags)
1478 spin_lock(&zone->lock);
1479 if (unlikely(has_isolate_pageblock(zone) ||
1480 is_migrate_isolate(migratetype))) {
1481 migratetype = get_pfnblock_migratetype(page, pfn);
1483 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1484 spin_unlock(&zone->lock);
1487 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1488 unsigned long zone, int nid)
1490 mm_zero_struct_page(page);
1491 set_page_links(page, zone, nid, pfn);
1492 init_page_count(page);
1493 page_mapcount_reset(page);
1494 page_cpupid_reset_last(page);
1495 page_kasan_tag_reset(page);
1497 INIT_LIST_HEAD(&page->lru);
1498 #ifdef WANT_PAGE_VIRTUAL
1499 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1500 if (!is_highmem_idx(zone))
1501 set_page_address(page, __va(pfn << PAGE_SHIFT));
1505 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1506 static void __meminit init_reserved_page(unsigned long pfn)
1511 if (!early_page_uninitialised(pfn))
1514 nid = early_pfn_to_nid(pfn);
1515 pgdat = NODE_DATA(nid);
1517 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1518 struct zone *zone = &pgdat->node_zones[zid];
1520 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1523 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1526 static inline void init_reserved_page(unsigned long pfn)
1529 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1532 * Initialised pages do not have PageReserved set. This function is
1533 * called for each range allocated by the bootmem allocator and
1534 * marks the pages PageReserved. The remaining valid pages are later
1535 * sent to the buddy page allocator.
1537 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1539 unsigned long start_pfn = PFN_DOWN(start);
1540 unsigned long end_pfn = PFN_UP(end);
1542 for (; start_pfn < end_pfn; start_pfn++) {
1543 if (pfn_valid(start_pfn)) {
1544 struct page *page = pfn_to_page(start_pfn);
1546 init_reserved_page(start_pfn);
1548 /* Avoid false-positive PageTail() */
1549 INIT_LIST_HEAD(&page->lru);
1552 * no need for atomic set_bit because the struct
1553 * page is not visible yet so nobody should
1556 __SetPageReserved(page);
1561 static void __free_pages_ok(struct page *page, unsigned int order,
1564 unsigned long flags;
1566 unsigned long pfn = page_to_pfn(page);
1568 if (!free_pages_prepare(page, order, true, fpi_flags))
1571 migratetype = get_pfnblock_migratetype(page, pfn);
1572 local_irq_save(flags);
1573 __count_vm_events(PGFREE, 1 << order);
1574 free_one_page(page_zone(page), page, pfn, order, migratetype,
1576 local_irq_restore(flags);
1579 void __free_pages_core(struct page *page, unsigned int order)
1581 unsigned int nr_pages = 1 << order;
1582 struct page *p = page;
1586 * When initializing the memmap, __init_single_page() sets the refcount
1587 * of all pages to 1 ("allocated"/"not free"). We have to set the
1588 * refcount of all involved pages to 0.
1591 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1593 __ClearPageReserved(p);
1594 set_page_count(p, 0);
1596 __ClearPageReserved(p);
1597 set_page_count(p, 0);
1599 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1602 * Bypass PCP and place fresh pages right to the tail, primarily
1603 * relevant for memory onlining.
1605 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1608 #ifdef CONFIG_NEED_MULTIPLE_NODES
1611 * During memory init memblocks map pfns to nids. The search is expensive and
1612 * this caches recent lookups. The implementation of __early_pfn_to_nid
1613 * treats start/end as pfns.
1615 struct mminit_pfnnid_cache {
1616 unsigned long last_start;
1617 unsigned long last_end;
1621 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1624 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1626 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1627 struct mminit_pfnnid_cache *state)
1629 unsigned long start_pfn, end_pfn;
1632 if (state->last_start <= pfn && pfn < state->last_end)
1633 return state->last_nid;
1635 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1636 if (nid != NUMA_NO_NODE) {
1637 state->last_start = start_pfn;
1638 state->last_end = end_pfn;
1639 state->last_nid = nid;
1645 int __meminit early_pfn_to_nid(unsigned long pfn)
1647 static DEFINE_SPINLOCK(early_pfn_lock);
1650 spin_lock(&early_pfn_lock);
1651 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1653 nid = first_online_node;
1654 spin_unlock(&early_pfn_lock);
1658 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1660 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1663 if (early_page_uninitialised(pfn))
1665 __free_pages_core(page, order);
1669 * Check that the whole (or subset of) a pageblock given by the interval of
1670 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1671 * with the migration of free compaction scanner. The scanners then need to
1672 * use only pfn_valid_within() check for arches that allow holes within
1675 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1677 * It's possible on some configurations to have a setup like node0 node1 node0
1678 * i.e. it's possible that all pages within a zones range of pages do not
1679 * belong to a single zone. We assume that a border between node0 and node1
1680 * can occur within a single pageblock, but not a node0 node1 node0
1681 * interleaving within a single pageblock. It is therefore sufficient to check
1682 * the first and last page of a pageblock and avoid checking each individual
1683 * page in a pageblock.
1685 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1686 unsigned long end_pfn, struct zone *zone)
1688 struct page *start_page;
1689 struct page *end_page;
1691 /* end_pfn is one past the range we are checking */
1694 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1697 start_page = pfn_to_online_page(start_pfn);
1701 if (page_zone(start_page) != zone)
1704 end_page = pfn_to_page(end_pfn);
1706 /* This gives a shorter code than deriving page_zone(end_page) */
1707 if (page_zone_id(start_page) != page_zone_id(end_page))
1713 void set_zone_contiguous(struct zone *zone)
1715 unsigned long block_start_pfn = zone->zone_start_pfn;
1716 unsigned long block_end_pfn;
1718 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1719 for (; block_start_pfn < zone_end_pfn(zone);
1720 block_start_pfn = block_end_pfn,
1721 block_end_pfn += pageblock_nr_pages) {
1723 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1725 if (!__pageblock_pfn_to_page(block_start_pfn,
1726 block_end_pfn, zone))
1731 /* We confirm that there is no hole */
1732 zone->contiguous = true;
1735 void clear_zone_contiguous(struct zone *zone)
1737 zone->contiguous = false;
1740 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1741 static void __init deferred_free_range(unsigned long pfn,
1742 unsigned long nr_pages)
1750 page = pfn_to_page(pfn);
1752 /* Free a large naturally-aligned chunk if possible */
1753 if (nr_pages == pageblock_nr_pages &&
1754 (pfn & (pageblock_nr_pages - 1)) == 0) {
1755 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1756 __free_pages_core(page, pageblock_order);
1760 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1761 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1762 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1763 __free_pages_core(page, 0);
1767 /* Completion tracking for deferred_init_memmap() threads */
1768 static atomic_t pgdat_init_n_undone __initdata;
1769 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1771 static inline void __init pgdat_init_report_one_done(void)
1773 if (atomic_dec_and_test(&pgdat_init_n_undone))
1774 complete(&pgdat_init_all_done_comp);
1778 * Returns true if page needs to be initialized or freed to buddy allocator.
1780 * First we check if pfn is valid on architectures where it is possible to have
1781 * holes within pageblock_nr_pages. On systems where it is not possible, this
1782 * function is optimized out.
1784 * Then, we check if a current large page is valid by only checking the validity
1787 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1789 if (!pfn_valid_within(pfn))
1791 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1797 * Free pages to buddy allocator. Try to free aligned pages in
1798 * pageblock_nr_pages sizes.
1800 static void __init deferred_free_pages(unsigned long pfn,
1801 unsigned long end_pfn)
1803 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1804 unsigned long nr_free = 0;
1806 for (; pfn < end_pfn; pfn++) {
1807 if (!deferred_pfn_valid(pfn)) {
1808 deferred_free_range(pfn - nr_free, nr_free);
1810 } else if (!(pfn & nr_pgmask)) {
1811 deferred_free_range(pfn - nr_free, nr_free);
1817 /* Free the last block of pages to allocator */
1818 deferred_free_range(pfn - nr_free, nr_free);
1822 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1823 * by performing it only once every pageblock_nr_pages.
1824 * Return number of pages initialized.
1826 static unsigned long __init deferred_init_pages(struct zone *zone,
1828 unsigned long end_pfn)
1830 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1831 int nid = zone_to_nid(zone);
1832 unsigned long nr_pages = 0;
1833 int zid = zone_idx(zone);
1834 struct page *page = NULL;
1836 for (; pfn < end_pfn; pfn++) {
1837 if (!deferred_pfn_valid(pfn)) {
1840 } else if (!page || !(pfn & nr_pgmask)) {
1841 page = pfn_to_page(pfn);
1845 __init_single_page(page, pfn, zid, nid);
1852 * This function is meant to pre-load the iterator for the zone init.
1853 * Specifically it walks through the ranges until we are caught up to the
1854 * first_init_pfn value and exits there. If we never encounter the value we
1855 * return false indicating there are no valid ranges left.
1858 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1859 unsigned long *spfn, unsigned long *epfn,
1860 unsigned long first_init_pfn)
1865 * Start out by walking through the ranges in this zone that have
1866 * already been initialized. We don't need to do anything with them
1867 * so we just need to flush them out of the system.
1869 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1870 if (*epfn <= first_init_pfn)
1872 if (*spfn < first_init_pfn)
1873 *spfn = first_init_pfn;
1882 * Initialize and free pages. We do it in two loops: first we initialize
1883 * struct page, then free to buddy allocator, because while we are
1884 * freeing pages we can access pages that are ahead (computing buddy
1885 * page in __free_one_page()).
1887 * In order to try and keep some memory in the cache we have the loop
1888 * broken along max page order boundaries. This way we will not cause
1889 * any issues with the buddy page computation.
1891 static unsigned long __init
1892 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1893 unsigned long *end_pfn)
1895 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1896 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1897 unsigned long nr_pages = 0;
1900 /* First we loop through and initialize the page values */
1901 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1904 if (mo_pfn <= *start_pfn)
1907 t = min(mo_pfn, *end_pfn);
1908 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1910 if (mo_pfn < *end_pfn) {
1911 *start_pfn = mo_pfn;
1916 /* Reset values and now loop through freeing pages as needed */
1919 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1925 t = min(mo_pfn, epfn);
1926 deferred_free_pages(spfn, t);
1936 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1939 unsigned long spfn, epfn;
1940 struct zone *zone = arg;
1943 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1946 * Initialize and free pages in MAX_ORDER sized increments so that we
1947 * can avoid introducing any issues with the buddy allocator.
1949 while (spfn < end_pfn) {
1950 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1955 /* An arch may override for more concurrency. */
1957 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1962 /* Initialise remaining memory on a node */
1963 static int __init deferred_init_memmap(void *data)
1965 pg_data_t *pgdat = data;
1966 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1967 unsigned long spfn = 0, epfn = 0;
1968 unsigned long first_init_pfn, flags;
1969 unsigned long start = jiffies;
1971 int zid, max_threads;
1974 /* Bind memory initialisation thread to a local node if possible */
1975 if (!cpumask_empty(cpumask))
1976 set_cpus_allowed_ptr(current, cpumask);
1978 pgdat_resize_lock(pgdat, &flags);
1979 first_init_pfn = pgdat->first_deferred_pfn;
1980 if (first_init_pfn == ULONG_MAX) {
1981 pgdat_resize_unlock(pgdat, &flags);
1982 pgdat_init_report_one_done();
1986 /* Sanity check boundaries */
1987 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1988 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1989 pgdat->first_deferred_pfn = ULONG_MAX;
1992 * Once we unlock here, the zone cannot be grown anymore, thus if an
1993 * interrupt thread must allocate this early in boot, zone must be
1994 * pre-grown prior to start of deferred page initialization.
1996 pgdat_resize_unlock(pgdat, &flags);
1998 /* Only the highest zone is deferred so find it */
1999 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2000 zone = pgdat->node_zones + zid;
2001 if (first_init_pfn < zone_end_pfn(zone))
2005 /* If the zone is empty somebody else may have cleared out the zone */
2006 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2010 max_threads = deferred_page_init_max_threads(cpumask);
2012 while (spfn < epfn) {
2013 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2014 struct padata_mt_job job = {
2015 .thread_fn = deferred_init_memmap_chunk,
2018 .size = epfn_align - spfn,
2019 .align = PAGES_PER_SECTION,
2020 .min_chunk = PAGES_PER_SECTION,
2021 .max_threads = max_threads,
2024 padata_do_multithreaded(&job);
2025 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2029 /* Sanity check that the next zone really is unpopulated */
2030 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2032 pr_info("node %d deferred pages initialised in %ums\n",
2033 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2035 pgdat_init_report_one_done();
2040 * If this zone has deferred pages, try to grow it by initializing enough
2041 * deferred pages to satisfy the allocation specified by order, rounded up to
2042 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2043 * of SECTION_SIZE bytes by initializing struct pages in increments of
2044 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2046 * Return true when zone was grown, otherwise return false. We return true even
2047 * when we grow less than requested, to let the caller decide if there are
2048 * enough pages to satisfy the allocation.
2050 * Note: We use noinline because this function is needed only during boot, and
2051 * it is called from a __ref function _deferred_grow_zone. This way we are
2052 * making sure that it is not inlined into permanent text section.
2054 static noinline bool __init
2055 deferred_grow_zone(struct zone *zone, unsigned int order)
2057 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2058 pg_data_t *pgdat = zone->zone_pgdat;
2059 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2060 unsigned long spfn, epfn, flags;
2061 unsigned long nr_pages = 0;
2064 /* Only the last zone may have deferred pages */
2065 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2068 pgdat_resize_lock(pgdat, &flags);
2071 * If someone grew this zone while we were waiting for spinlock, return
2072 * true, as there might be enough pages already.
2074 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2075 pgdat_resize_unlock(pgdat, &flags);
2079 /* If the zone is empty somebody else may have cleared out the zone */
2080 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2081 first_deferred_pfn)) {
2082 pgdat->first_deferred_pfn = ULONG_MAX;
2083 pgdat_resize_unlock(pgdat, &flags);
2084 /* Retry only once. */
2085 return first_deferred_pfn != ULONG_MAX;
2089 * Initialize and free pages in MAX_ORDER sized increments so
2090 * that we can avoid introducing any issues with the buddy
2093 while (spfn < epfn) {
2094 /* update our first deferred PFN for this section */
2095 first_deferred_pfn = spfn;
2097 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2098 touch_nmi_watchdog();
2100 /* We should only stop along section boundaries */
2101 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2104 /* If our quota has been met we can stop here */
2105 if (nr_pages >= nr_pages_needed)
2109 pgdat->first_deferred_pfn = spfn;
2110 pgdat_resize_unlock(pgdat, &flags);
2112 return nr_pages > 0;
2116 * deferred_grow_zone() is __init, but it is called from
2117 * get_page_from_freelist() during early boot until deferred_pages permanently
2118 * disables this call. This is why we have refdata wrapper to avoid warning,
2119 * and to ensure that the function body gets unloaded.
2122 _deferred_grow_zone(struct zone *zone, unsigned int order)
2124 return deferred_grow_zone(zone, order);
2127 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2129 void __init page_alloc_init_late(void)
2134 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2136 /* There will be num_node_state(N_MEMORY) threads */
2137 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2138 for_each_node_state(nid, N_MEMORY) {
2139 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2142 /* Block until all are initialised */
2143 wait_for_completion(&pgdat_init_all_done_comp);
2146 * The number of managed pages has changed due to the initialisation
2147 * so the pcpu batch and high limits needs to be updated or the limits
2148 * will be artificially small.
2150 for_each_populated_zone(zone)
2151 zone_pcp_update(zone);
2154 * We initialized the rest of the deferred pages. Permanently disable
2155 * on-demand struct page initialization.
2157 static_branch_disable(&deferred_pages);
2159 /* Reinit limits that are based on free pages after the kernel is up */
2160 files_maxfiles_init();
2165 /* Discard memblock private memory */
2168 for_each_node_state(nid, N_MEMORY)
2169 shuffle_free_memory(NODE_DATA(nid));
2171 for_each_populated_zone(zone)
2172 set_zone_contiguous(zone);
2176 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2177 void __init init_cma_reserved_pageblock(struct page *page)
2179 unsigned i = pageblock_nr_pages;
2180 struct page *p = page;
2183 __ClearPageReserved(p);
2184 set_page_count(p, 0);
2187 set_pageblock_migratetype(page, MIGRATE_CMA);
2189 if (pageblock_order >= MAX_ORDER) {
2190 i = pageblock_nr_pages;
2193 set_page_refcounted(p);
2194 __free_pages(p, MAX_ORDER - 1);
2195 p += MAX_ORDER_NR_PAGES;
2196 } while (i -= MAX_ORDER_NR_PAGES);
2198 set_page_refcounted(page);
2199 __free_pages(page, pageblock_order);
2202 adjust_managed_page_count(page, pageblock_nr_pages);
2203 page_zone(page)->cma_pages += pageblock_nr_pages;
2208 * The order of subdivision here is critical for the IO subsystem.
2209 * Please do not alter this order without good reasons and regression
2210 * testing. Specifically, as large blocks of memory are subdivided,
2211 * the order in which smaller blocks are delivered depends on the order
2212 * they're subdivided in this function. This is the primary factor
2213 * influencing the order in which pages are delivered to the IO
2214 * subsystem according to empirical testing, and this is also justified
2215 * by considering the behavior of a buddy system containing a single
2216 * large block of memory acted on by a series of small allocations.
2217 * This behavior is a critical factor in sglist merging's success.
2221 static inline void expand(struct zone *zone, struct page *page,
2222 int low, int high, int migratetype)
2224 unsigned long size = 1 << high;
2226 while (high > low) {
2229 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2232 * Mark as guard pages (or page), that will allow to
2233 * merge back to allocator when buddy will be freed.
2234 * Corresponding page table entries will not be touched,
2235 * pages will stay not present in virtual address space
2237 if (set_page_guard(zone, &page[size], high, migratetype))
2240 add_to_free_list(&page[size], zone, high, migratetype);
2241 set_buddy_order(&page[size], high);
2245 static void check_new_page_bad(struct page *page)
2247 if (unlikely(page->flags & __PG_HWPOISON)) {
2248 /* Don't complain about hwpoisoned pages */
2249 page_mapcount_reset(page); /* remove PageBuddy */
2254 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2258 * This page is about to be returned from the page allocator
2260 static inline int check_new_page(struct page *page)
2262 if (likely(page_expected_state(page,
2263 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2266 check_new_page_bad(page);
2270 #ifdef CONFIG_DEBUG_VM
2272 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2273 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2274 * also checked when pcp lists are refilled from the free lists.
2276 static inline bool check_pcp_refill(struct page *page)
2278 if (debug_pagealloc_enabled_static())
2279 return check_new_page(page);
2284 static inline bool check_new_pcp(struct page *page)
2286 return check_new_page(page);
2290 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2291 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2292 * enabled, they are also checked when being allocated from the pcp lists.
2294 static inline bool check_pcp_refill(struct page *page)
2296 return check_new_page(page);
2298 static inline bool check_new_pcp(struct page *page)
2300 if (debug_pagealloc_enabled_static())
2301 return check_new_page(page);
2305 #endif /* CONFIG_DEBUG_VM */
2307 static bool check_new_pages(struct page *page, unsigned int order)
2310 for (i = 0; i < (1 << order); i++) {
2311 struct page *p = page + i;
2313 if (unlikely(check_new_page(p)))
2320 inline void post_alloc_hook(struct page *page, unsigned int order,
2325 set_page_private(page, 0);
2326 set_page_refcounted(page);
2328 arch_alloc_page(page, order);
2329 debug_pagealloc_map_pages(page, 1 << order);
2332 * Page unpoisoning must happen before memory initialization.
2333 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2334 * allocations and the page unpoisoning code will complain.
2336 kernel_unpoison_pages(page, 1 << order);
2339 * As memory initialization might be integrated into KASAN,
2340 * kasan_alloc_pages and kernel_init_free_pages must be
2341 * kept together to avoid discrepancies in behavior.
2343 init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2344 kasan_alloc_pages(page, order, init);
2345 if (init && !kasan_has_integrated_init())
2346 kernel_init_free_pages(page, 1 << order);
2348 set_page_owner(page, order, gfp_flags);
2351 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2352 unsigned int alloc_flags)
2354 post_alloc_hook(page, order, gfp_flags);
2356 if (order && (gfp_flags & __GFP_COMP))
2357 prep_compound_page(page, order);
2360 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2361 * allocate the page. The expectation is that the caller is taking
2362 * steps that will free more memory. The caller should avoid the page
2363 * being used for !PFMEMALLOC purposes.
2365 if (alloc_flags & ALLOC_NO_WATERMARKS)
2366 set_page_pfmemalloc(page);
2368 clear_page_pfmemalloc(page);
2372 * Go through the free lists for the given migratetype and remove
2373 * the smallest available page from the freelists
2375 static __always_inline
2376 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2379 unsigned int current_order;
2380 struct free_area *area;
2383 /* Find a page of the appropriate size in the preferred list */
2384 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2385 area = &(zone->free_area[current_order]);
2386 page = get_page_from_free_area(area, migratetype);
2389 del_page_from_free_list(page, zone, current_order);
2390 expand(zone, page, order, current_order, migratetype);
2391 set_pcppage_migratetype(page, migratetype);
2400 * This array describes the order lists are fallen back to when
2401 * the free lists for the desirable migrate type are depleted
2403 static int fallbacks[MIGRATE_TYPES][3] = {
2404 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2405 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2406 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2408 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2410 #ifdef CONFIG_MEMORY_ISOLATION
2411 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2416 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2419 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2422 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2423 unsigned int order) { return NULL; }
2427 * Move the free pages in a range to the freelist tail of the requested type.
2428 * Note that start_page and end_pages are not aligned on a pageblock
2429 * boundary. If alignment is required, use move_freepages_block()
2431 static int move_freepages(struct zone *zone,
2432 unsigned long start_pfn, unsigned long end_pfn,
2433 int migratetype, int *num_movable)
2438 int pages_moved = 0;
2440 for (pfn = start_pfn; pfn <= end_pfn;) {
2441 if (!pfn_valid_within(pfn)) {
2446 page = pfn_to_page(pfn);
2447 if (!PageBuddy(page)) {
2449 * We assume that pages that could be isolated for
2450 * migration are movable. But we don't actually try
2451 * isolating, as that would be expensive.
2454 (PageLRU(page) || __PageMovable(page)))
2460 /* Make sure we are not inadvertently changing nodes */
2461 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2462 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2464 order = buddy_order(page);
2465 move_to_free_list(page, zone, order, migratetype);
2467 pages_moved += 1 << order;
2473 int move_freepages_block(struct zone *zone, struct page *page,
2474 int migratetype, int *num_movable)
2476 unsigned long start_pfn, end_pfn, pfn;
2481 pfn = page_to_pfn(page);
2482 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2483 end_pfn = start_pfn + pageblock_nr_pages - 1;
2485 /* Do not cross zone boundaries */
2486 if (!zone_spans_pfn(zone, start_pfn))
2488 if (!zone_spans_pfn(zone, end_pfn))
2491 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2495 static void change_pageblock_range(struct page *pageblock_page,
2496 int start_order, int migratetype)
2498 int nr_pageblocks = 1 << (start_order - pageblock_order);
2500 while (nr_pageblocks--) {
2501 set_pageblock_migratetype(pageblock_page, migratetype);
2502 pageblock_page += pageblock_nr_pages;
2507 * When we are falling back to another migratetype during allocation, try to
2508 * steal extra free pages from the same pageblocks to satisfy further
2509 * allocations, instead of polluting multiple pageblocks.
2511 * If we are stealing a relatively large buddy page, it is likely there will
2512 * be more free pages in the pageblock, so try to steal them all. For
2513 * reclaimable and unmovable allocations, we steal regardless of page size,
2514 * as fragmentation caused by those allocations polluting movable pageblocks
2515 * is worse than movable allocations stealing from unmovable and reclaimable
2518 static bool can_steal_fallback(unsigned int order, int start_mt)
2521 * Leaving this order check is intended, although there is
2522 * relaxed order check in next check. The reason is that
2523 * we can actually steal whole pageblock if this condition met,
2524 * but, below check doesn't guarantee it and that is just heuristic
2525 * so could be changed anytime.
2527 if (order >= pageblock_order)
2530 if (order >= pageblock_order / 2 ||
2531 start_mt == MIGRATE_RECLAIMABLE ||
2532 start_mt == MIGRATE_UNMOVABLE ||
2533 page_group_by_mobility_disabled)
2539 static inline bool boost_watermark(struct zone *zone)
2541 unsigned long max_boost;
2543 if (!watermark_boost_factor)
2546 * Don't bother in zones that are unlikely to produce results.
2547 * On small machines, including kdump capture kernels running
2548 * in a small area, boosting the watermark can cause an out of
2549 * memory situation immediately.
2551 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2554 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2555 watermark_boost_factor, 10000);
2558 * high watermark may be uninitialised if fragmentation occurs
2559 * very early in boot so do not boost. We do not fall
2560 * through and boost by pageblock_nr_pages as failing
2561 * allocations that early means that reclaim is not going
2562 * to help and it may even be impossible to reclaim the
2563 * boosted watermark resulting in a hang.
2568 max_boost = max(pageblock_nr_pages, max_boost);
2570 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2577 * This function implements actual steal behaviour. If order is large enough,
2578 * we can steal whole pageblock. If not, we first move freepages in this
2579 * pageblock to our migratetype and determine how many already-allocated pages
2580 * are there in the pageblock with a compatible migratetype. If at least half
2581 * of pages are free or compatible, we can change migratetype of the pageblock
2582 * itself, so pages freed in the future will be put on the correct free list.
2584 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2585 unsigned int alloc_flags, int start_type, bool whole_block)
2587 unsigned int current_order = buddy_order(page);
2588 int free_pages, movable_pages, alike_pages;
2591 old_block_type = get_pageblock_migratetype(page);
2594 * This can happen due to races and we want to prevent broken
2595 * highatomic accounting.
2597 if (is_migrate_highatomic(old_block_type))
2600 /* Take ownership for orders >= pageblock_order */
2601 if (current_order >= pageblock_order) {
2602 change_pageblock_range(page, current_order, start_type);
2607 * Boost watermarks to increase reclaim pressure to reduce the
2608 * likelihood of future fallbacks. Wake kswapd now as the node
2609 * may be balanced overall and kswapd will not wake naturally.
2611 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2612 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2614 /* We are not allowed to try stealing from the whole block */
2618 free_pages = move_freepages_block(zone, page, start_type,
2621 * Determine how many pages are compatible with our allocation.
2622 * For movable allocation, it's the number of movable pages which
2623 * we just obtained. For other types it's a bit more tricky.
2625 if (start_type == MIGRATE_MOVABLE) {
2626 alike_pages = movable_pages;
2629 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2630 * to MOVABLE pageblock, consider all non-movable pages as
2631 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2632 * vice versa, be conservative since we can't distinguish the
2633 * exact migratetype of non-movable pages.
2635 if (old_block_type == MIGRATE_MOVABLE)
2636 alike_pages = pageblock_nr_pages
2637 - (free_pages + movable_pages);
2642 /* moving whole block can fail due to zone boundary conditions */
2647 * If a sufficient number of pages in the block are either free or of
2648 * comparable migratability as our allocation, claim the whole block.
2650 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2651 page_group_by_mobility_disabled)
2652 set_pageblock_migratetype(page, start_type);
2657 move_to_free_list(page, zone, current_order, start_type);
2661 * Check whether there is a suitable fallback freepage with requested order.
2662 * If only_stealable is true, this function returns fallback_mt only if
2663 * we can steal other freepages all together. This would help to reduce
2664 * fragmentation due to mixed migratetype pages in one pageblock.
2666 int find_suitable_fallback(struct free_area *area, unsigned int order,
2667 int migratetype, bool only_stealable, bool *can_steal)
2672 if (area->nr_free == 0)
2677 fallback_mt = fallbacks[migratetype][i];
2678 if (fallback_mt == MIGRATE_TYPES)
2681 if (free_area_empty(area, fallback_mt))
2684 if (can_steal_fallback(order, migratetype))
2687 if (!only_stealable)
2698 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2699 * there are no empty page blocks that contain a page with a suitable order
2701 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2702 unsigned int alloc_order)
2705 unsigned long max_managed, flags;
2708 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2709 * Check is race-prone but harmless.
2711 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2712 if (zone->nr_reserved_highatomic >= max_managed)
2715 spin_lock_irqsave(&zone->lock, flags);
2717 /* Recheck the nr_reserved_highatomic limit under the lock */
2718 if (zone->nr_reserved_highatomic >= max_managed)
2722 mt = get_pageblock_migratetype(page);
2723 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2724 && !is_migrate_cma(mt)) {
2725 zone->nr_reserved_highatomic += pageblock_nr_pages;
2726 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2727 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2731 spin_unlock_irqrestore(&zone->lock, flags);
2735 * Used when an allocation is about to fail under memory pressure. This
2736 * potentially hurts the reliability of high-order allocations when under
2737 * intense memory pressure but failed atomic allocations should be easier
2738 * to recover from than an OOM.
2740 * If @force is true, try to unreserve a pageblock even though highatomic
2741 * pageblock is exhausted.
2743 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2746 struct zonelist *zonelist = ac->zonelist;
2747 unsigned long flags;
2754 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2757 * Preserve at least one pageblock unless memory pressure
2760 if (!force && zone->nr_reserved_highatomic <=
2764 spin_lock_irqsave(&zone->lock, flags);
2765 for (order = 0; order < MAX_ORDER; order++) {
2766 struct free_area *area = &(zone->free_area[order]);
2768 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2773 * In page freeing path, migratetype change is racy so
2774 * we can counter several free pages in a pageblock
2775 * in this loop althoug we changed the pageblock type
2776 * from highatomic to ac->migratetype. So we should
2777 * adjust the count once.
2779 if (is_migrate_highatomic_page(page)) {
2781 * It should never happen but changes to
2782 * locking could inadvertently allow a per-cpu
2783 * drain to add pages to MIGRATE_HIGHATOMIC
2784 * while unreserving so be safe and watch for
2787 zone->nr_reserved_highatomic -= min(
2789 zone->nr_reserved_highatomic);
2793 * Convert to ac->migratetype and avoid the normal
2794 * pageblock stealing heuristics. Minimally, the caller
2795 * is doing the work and needs the pages. More
2796 * importantly, if the block was always converted to
2797 * MIGRATE_UNMOVABLE or another type then the number
2798 * of pageblocks that cannot be completely freed
2801 set_pageblock_migratetype(page, ac->migratetype);
2802 ret = move_freepages_block(zone, page, ac->migratetype,
2805 spin_unlock_irqrestore(&zone->lock, flags);
2809 spin_unlock_irqrestore(&zone->lock, flags);
2816 * Try finding a free buddy page on the fallback list and put it on the free
2817 * list of requested migratetype, possibly along with other pages from the same
2818 * block, depending on fragmentation avoidance heuristics. Returns true if
2819 * fallback was found so that __rmqueue_smallest() can grab it.
2821 * The use of signed ints for order and current_order is a deliberate
2822 * deviation from the rest of this file, to make the for loop
2823 * condition simpler.
2825 static __always_inline bool
2826 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2827 unsigned int alloc_flags)
2829 struct free_area *area;
2831 int min_order = order;
2837 * Do not steal pages from freelists belonging to other pageblocks
2838 * i.e. orders < pageblock_order. If there are no local zones free,
2839 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2841 if (alloc_flags & ALLOC_NOFRAGMENT)
2842 min_order = pageblock_order;
2845 * Find the largest available free page in the other list. This roughly
2846 * approximates finding the pageblock with the most free pages, which
2847 * would be too costly to do exactly.
2849 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2851 area = &(zone->free_area[current_order]);
2852 fallback_mt = find_suitable_fallback(area, current_order,
2853 start_migratetype, false, &can_steal);
2854 if (fallback_mt == -1)
2858 * We cannot steal all free pages from the pageblock and the
2859 * requested migratetype is movable. In that case it's better to
2860 * steal and split the smallest available page instead of the
2861 * largest available page, because even if the next movable
2862 * allocation falls back into a different pageblock than this
2863 * one, it won't cause permanent fragmentation.
2865 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2866 && current_order > order)
2875 for (current_order = order; current_order < MAX_ORDER;
2877 area = &(zone->free_area[current_order]);
2878 fallback_mt = find_suitable_fallback(area, current_order,
2879 start_migratetype, false, &can_steal);
2880 if (fallback_mt != -1)
2885 * This should not happen - we already found a suitable fallback
2886 * when looking for the largest page.
2888 VM_BUG_ON(current_order == MAX_ORDER);
2891 page = get_page_from_free_area(area, fallback_mt);
2893 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2896 trace_mm_page_alloc_extfrag(page, order, current_order,
2897 start_migratetype, fallback_mt);
2904 * Do the hard work of removing an element from the buddy allocator.
2905 * Call me with the zone->lock already held.
2907 static __always_inline struct page *
2908 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2909 unsigned int alloc_flags)
2913 if (IS_ENABLED(CONFIG_CMA)) {
2915 * Balance movable allocations between regular and CMA areas by
2916 * allocating from CMA when over half of the zone's free memory
2917 * is in the CMA area.
2919 if (alloc_flags & ALLOC_CMA &&
2920 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2921 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2922 page = __rmqueue_cma_fallback(zone, order);
2928 page = __rmqueue_smallest(zone, order, migratetype);
2929 if (unlikely(!page)) {
2930 if (alloc_flags & ALLOC_CMA)
2931 page = __rmqueue_cma_fallback(zone, order);
2933 if (!page && __rmqueue_fallback(zone, order, migratetype,
2939 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2944 * Obtain a specified number of elements from the buddy allocator, all under
2945 * a single hold of the lock, for efficiency. Add them to the supplied list.
2946 * Returns the number of new pages which were placed at *list.
2948 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2949 unsigned long count, struct list_head *list,
2950 int migratetype, unsigned int alloc_flags)
2954 spin_lock(&zone->lock);
2955 for (i = 0; i < count; ++i) {
2956 struct page *page = __rmqueue(zone, order, migratetype,
2958 if (unlikely(page == NULL))
2961 if (unlikely(check_pcp_refill(page)))
2965 * Split buddy pages returned by expand() are received here in
2966 * physical page order. The page is added to the tail of
2967 * caller's list. From the callers perspective, the linked list
2968 * is ordered by page number under some conditions. This is
2969 * useful for IO devices that can forward direction from the
2970 * head, thus also in the physical page order. This is useful
2971 * for IO devices that can merge IO requests if the physical
2972 * pages are ordered properly.
2974 list_add_tail(&page->lru, list);
2976 if (is_migrate_cma(get_pcppage_migratetype(page)))
2977 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2982 * i pages were removed from the buddy list even if some leak due
2983 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2984 * on i. Do not confuse with 'alloced' which is the number of
2985 * pages added to the pcp list.
2987 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2988 spin_unlock(&zone->lock);
2994 * Called from the vmstat counter updater to drain pagesets of this
2995 * currently executing processor on remote nodes after they have
2998 * Note that this function must be called with the thread pinned to
2999 * a single processor.
3001 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3003 unsigned long flags;
3004 int to_drain, batch;
3006 local_irq_save(flags);
3007 batch = READ_ONCE(pcp->batch);
3008 to_drain = min(pcp->count, batch);
3010 free_pcppages_bulk(zone, to_drain, pcp);
3011 local_irq_restore(flags);
3016 * Drain pcplists of the indicated processor and zone.
3018 * The processor must either be the current processor and the
3019 * thread pinned to the current processor or a processor that
3022 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3024 unsigned long flags;
3025 struct per_cpu_pageset *pset;
3026 struct per_cpu_pages *pcp;
3028 local_irq_save(flags);
3029 pset = per_cpu_ptr(zone->pageset, cpu);
3033 free_pcppages_bulk(zone, pcp->count, pcp);
3034 local_irq_restore(flags);
3038 * Drain pcplists of all zones on the indicated processor.
3040 * The processor must either be the current processor and the
3041 * thread pinned to the current processor or a processor that
3044 static void drain_pages(unsigned int cpu)
3048 for_each_populated_zone(zone) {
3049 drain_pages_zone(cpu, zone);
3054 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3056 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3057 * the single zone's pages.
3059 void drain_local_pages(struct zone *zone)
3061 int cpu = smp_processor_id();
3064 drain_pages_zone(cpu, zone);
3069 static void drain_local_pages_wq(struct work_struct *work)
3071 struct pcpu_drain *drain;
3073 drain = container_of(work, struct pcpu_drain, work);
3076 * drain_all_pages doesn't use proper cpu hotplug protection so
3077 * we can race with cpu offline when the WQ can move this from
3078 * a cpu pinned worker to an unbound one. We can operate on a different
3079 * cpu which is allright but we also have to make sure to not move to
3083 drain_local_pages(drain->zone);
3088 * The implementation of drain_all_pages(), exposing an extra parameter to
3089 * drain on all cpus.
3091 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3092 * not empty. The check for non-emptiness can however race with a free to
3093 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3094 * that need the guarantee that every CPU has drained can disable the
3095 * optimizing racy check.
3097 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3102 * Allocate in the BSS so we wont require allocation in
3103 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3105 static cpumask_t cpus_with_pcps;
3108 * Make sure nobody triggers this path before mm_percpu_wq is fully
3111 if (WARN_ON_ONCE(!mm_percpu_wq))
3115 * Do not drain if one is already in progress unless it's specific to
3116 * a zone. Such callers are primarily CMA and memory hotplug and need
3117 * the drain to be complete when the call returns.
3119 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3122 mutex_lock(&pcpu_drain_mutex);
3126 * We don't care about racing with CPU hotplug event
3127 * as offline notification will cause the notified
3128 * cpu to drain that CPU pcps and on_each_cpu_mask
3129 * disables preemption as part of its processing
3131 for_each_online_cpu(cpu) {
3132 struct per_cpu_pageset *pcp;
3134 bool has_pcps = false;
3136 if (force_all_cpus) {
3138 * The pcp.count check is racy, some callers need a
3139 * guarantee that no cpu is missed.
3143 pcp = per_cpu_ptr(zone->pageset, cpu);
3147 for_each_populated_zone(z) {
3148 pcp = per_cpu_ptr(z->pageset, cpu);
3149 if (pcp->pcp.count) {
3157 cpumask_set_cpu(cpu, &cpus_with_pcps);
3159 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3162 for_each_cpu(cpu, &cpus_with_pcps) {
3163 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3166 INIT_WORK(&drain->work, drain_local_pages_wq);
3167 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3169 for_each_cpu(cpu, &cpus_with_pcps)
3170 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3172 mutex_unlock(&pcpu_drain_mutex);
3176 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3178 * When zone parameter is non-NULL, spill just the single zone's pages.
3180 * Note that this can be extremely slow as the draining happens in a workqueue.
3182 void drain_all_pages(struct zone *zone)
3184 __drain_all_pages(zone, false);
3187 #ifdef CONFIG_HIBERNATION
3190 * Touch the watchdog for every WD_PAGE_COUNT pages.
3192 #define WD_PAGE_COUNT (128*1024)
3194 void mark_free_pages(struct zone *zone)
3196 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3197 unsigned long flags;
3198 unsigned int order, t;
3201 if (zone_is_empty(zone))
3204 spin_lock_irqsave(&zone->lock, flags);
3206 max_zone_pfn = zone_end_pfn(zone);
3207 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3208 if (pfn_valid(pfn)) {
3209 page = pfn_to_page(pfn);
3211 if (!--page_count) {
3212 touch_nmi_watchdog();
3213 page_count = WD_PAGE_COUNT;
3216 if (page_zone(page) != zone)
3219 if (!swsusp_page_is_forbidden(page))
3220 swsusp_unset_page_free(page);
3223 for_each_migratetype_order(order, t) {
3224 list_for_each_entry(page,
3225 &zone->free_area[order].free_list[t], lru) {
3228 pfn = page_to_pfn(page);
3229 for (i = 0; i < (1UL << order); i++) {
3230 if (!--page_count) {
3231 touch_nmi_watchdog();
3232 page_count = WD_PAGE_COUNT;
3234 swsusp_set_page_free(pfn_to_page(pfn + i));
3238 spin_unlock_irqrestore(&zone->lock, flags);
3240 #endif /* CONFIG_PM */
3242 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3246 if (!free_pcp_prepare(page))
3249 migratetype = get_pfnblock_migratetype(page, pfn);
3250 set_pcppage_migratetype(page, migratetype);
3254 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3256 struct zone *zone = page_zone(page);
3257 struct per_cpu_pages *pcp;
3260 migratetype = get_pcppage_migratetype(page);
3261 __count_vm_event(PGFREE);
3264 * We only track unmovable, reclaimable and movable on pcp lists.
3265 * Free ISOLATE pages back to the allocator because they are being
3266 * offlined but treat HIGHATOMIC as movable pages so we can get those
3267 * areas back if necessary. Otherwise, we may have to free
3268 * excessively into the page allocator
3270 if (migratetype >= MIGRATE_PCPTYPES) {
3271 if (unlikely(is_migrate_isolate(migratetype))) {
3272 free_one_page(zone, page, pfn, 0, migratetype,
3276 migratetype = MIGRATE_MOVABLE;
3279 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3280 list_add(&page->lru, &pcp->lists[migratetype]);
3282 if (pcp->count >= READ_ONCE(pcp->high))
3283 free_pcppages_bulk(zone, READ_ONCE(pcp->batch), pcp);
3287 * Free a 0-order page
3289 void free_unref_page(struct page *page)
3291 unsigned long flags;
3292 unsigned long pfn = page_to_pfn(page);
3294 if (!free_unref_page_prepare(page, pfn))
3297 local_irq_save(flags);
3298 free_unref_page_commit(page, pfn);
3299 local_irq_restore(flags);
3303 * Free a list of 0-order pages
3305 void free_unref_page_list(struct list_head *list)
3307 struct page *page, *next;
3308 unsigned long flags, pfn;
3309 int batch_count = 0;
3311 /* Prepare pages for freeing */
3312 list_for_each_entry_safe(page, next, list, lru) {
3313 pfn = page_to_pfn(page);
3314 if (!free_unref_page_prepare(page, pfn))
3315 list_del(&page->lru);
3316 set_page_private(page, pfn);
3319 local_irq_save(flags);
3320 list_for_each_entry_safe(page, next, list, lru) {
3321 unsigned long pfn = page_private(page);
3323 set_page_private(page, 0);
3324 trace_mm_page_free_batched(page);
3325 free_unref_page_commit(page, pfn);
3328 * Guard against excessive IRQ disabled times when we get
3329 * a large list of pages to free.
3331 if (++batch_count == SWAP_CLUSTER_MAX) {
3332 local_irq_restore(flags);
3334 local_irq_save(flags);
3337 local_irq_restore(flags);
3341 * split_page takes a non-compound higher-order page, and splits it into
3342 * n (1<<order) sub-pages: page[0..n]
3343 * Each sub-page must be freed individually.
3345 * Note: this is probably too low level an operation for use in drivers.
3346 * Please consult with lkml before using this in your driver.
3348 void split_page(struct page *page, unsigned int order)
3352 VM_BUG_ON_PAGE(PageCompound(page), page);
3353 VM_BUG_ON_PAGE(!page_count(page), page);
3355 for (i = 1; i < (1 << order); i++)
3356 set_page_refcounted(page + i);
3357 split_page_owner(page, 1 << order);
3358 split_page_memcg(page, 1 << order);
3360 EXPORT_SYMBOL_GPL(split_page);
3362 int __isolate_free_page(struct page *page, unsigned int order)
3364 unsigned long watermark;
3368 BUG_ON(!PageBuddy(page));
3370 zone = page_zone(page);
3371 mt = get_pageblock_migratetype(page);
3373 if (!is_migrate_isolate(mt)) {
3375 * Obey watermarks as if the page was being allocated. We can
3376 * emulate a high-order watermark check with a raised order-0
3377 * watermark, because we already know our high-order page
3380 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3381 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3384 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3387 /* Remove page from free list */
3389 del_page_from_free_list(page, zone, order);
3392 * Set the pageblock if the isolated page is at least half of a
3395 if (order >= pageblock_order - 1) {
3396 struct page *endpage = page + (1 << order) - 1;
3397 for (; page < endpage; page += pageblock_nr_pages) {
3398 int mt = get_pageblock_migratetype(page);
3399 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3400 && !is_migrate_highatomic(mt))
3401 set_pageblock_migratetype(page,
3407 return 1UL << order;
3411 * __putback_isolated_page - Return a now-isolated page back where we got it
3412 * @page: Page that was isolated
3413 * @order: Order of the isolated page
3414 * @mt: The page's pageblock's migratetype
3416 * This function is meant to return a page pulled from the free lists via
3417 * __isolate_free_page back to the free lists they were pulled from.
3419 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3421 struct zone *zone = page_zone(page);
3423 /* zone lock should be held when this function is called */
3424 lockdep_assert_held(&zone->lock);
3426 /* Return isolated page to tail of freelist. */
3427 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3428 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3432 * Update NUMA hit/miss statistics
3434 * Must be called with interrupts disabled.
3436 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3439 enum numa_stat_item local_stat = NUMA_LOCAL;
3441 /* skip numa counters update if numa stats is disabled */
3442 if (!static_branch_likely(&vm_numa_stat_key))
3445 if (zone_to_nid(z) != numa_node_id())
3446 local_stat = NUMA_OTHER;
3448 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3449 __inc_numa_state(z, NUMA_HIT);
3451 __inc_numa_state(z, NUMA_MISS);
3452 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3454 __inc_numa_state(z, local_stat);
3458 /* Remove page from the per-cpu list, caller must protect the list */
3459 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3460 unsigned int alloc_flags,
3461 struct per_cpu_pages *pcp,
3462 struct list_head *list)
3467 if (list_empty(list)) {
3468 pcp->count += rmqueue_bulk(zone, 0,
3469 READ_ONCE(pcp->batch), list,
3470 migratetype, alloc_flags);
3471 if (unlikely(list_empty(list)))
3475 page = list_first_entry(list, struct page, lru);
3476 list_del(&page->lru);
3478 } while (check_new_pcp(page));
3483 /* Lock and remove page from the per-cpu list */
3484 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3485 struct zone *zone, gfp_t gfp_flags,
3486 int migratetype, unsigned int alloc_flags)
3488 struct per_cpu_pages *pcp;
3489 struct list_head *list;
3491 unsigned long flags;
3493 local_irq_save(flags);
3494 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3495 list = &pcp->lists[migratetype];
3496 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3498 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3499 zone_statistics(preferred_zone, zone);
3501 local_irq_restore(flags);
3506 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3509 struct page *rmqueue(struct zone *preferred_zone,
3510 struct zone *zone, unsigned int order,
3511 gfp_t gfp_flags, unsigned int alloc_flags,
3514 unsigned long flags;
3517 if (likely(order == 0)) {
3519 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3520 * we need to skip it when CMA area isn't allowed.
3522 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3523 migratetype != MIGRATE_MOVABLE) {
3524 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3525 migratetype, alloc_flags);
3531 * We most definitely don't want callers attempting to
3532 * allocate greater than order-1 page units with __GFP_NOFAIL.
3534 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3535 spin_lock_irqsave(&zone->lock, flags);
3540 * order-0 request can reach here when the pcplist is skipped
3541 * due to non-CMA allocation context. HIGHATOMIC area is
3542 * reserved for high-order atomic allocation, so order-0
3543 * request should skip it.
3545 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3546 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3548 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3551 page = __rmqueue(zone, order, migratetype, alloc_flags);
3552 } while (page && check_new_pages(page, order));
3553 spin_unlock(&zone->lock);
3556 __mod_zone_freepage_state(zone, -(1 << order),
3557 get_pcppage_migratetype(page));
3559 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3560 zone_statistics(preferred_zone, zone);
3561 local_irq_restore(flags);
3564 /* Separate test+clear to avoid unnecessary atomics */
3565 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3566 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3567 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3570 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3574 local_irq_restore(flags);
3578 #ifdef CONFIG_FAIL_PAGE_ALLOC
3581 struct fault_attr attr;
3583 bool ignore_gfp_highmem;
3584 bool ignore_gfp_reclaim;
3586 } fail_page_alloc = {
3587 .attr = FAULT_ATTR_INITIALIZER,
3588 .ignore_gfp_reclaim = true,
3589 .ignore_gfp_highmem = true,
3593 static int __init setup_fail_page_alloc(char *str)
3595 return setup_fault_attr(&fail_page_alloc.attr, str);
3597 __setup("fail_page_alloc=", setup_fail_page_alloc);
3599 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3601 if (order < fail_page_alloc.min_order)
3603 if (gfp_mask & __GFP_NOFAIL)
3605 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3607 if (fail_page_alloc.ignore_gfp_reclaim &&
3608 (gfp_mask & __GFP_DIRECT_RECLAIM))
3611 return should_fail(&fail_page_alloc.attr, 1 << order);
3614 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3616 static int __init fail_page_alloc_debugfs(void)
3618 umode_t mode = S_IFREG | 0600;
3621 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3622 &fail_page_alloc.attr);
3624 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3625 &fail_page_alloc.ignore_gfp_reclaim);
3626 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3627 &fail_page_alloc.ignore_gfp_highmem);
3628 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3633 late_initcall(fail_page_alloc_debugfs);
3635 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3637 #else /* CONFIG_FAIL_PAGE_ALLOC */
3639 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3644 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3646 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3648 return __should_fail_alloc_page(gfp_mask, order);
3650 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3652 static inline long __zone_watermark_unusable_free(struct zone *z,
3653 unsigned int order, unsigned int alloc_flags)
3655 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3656 long unusable_free = (1 << order) - 1;
3659 * If the caller does not have rights to ALLOC_HARDER then subtract
3660 * the high-atomic reserves. This will over-estimate the size of the
3661 * atomic reserve but it avoids a search.
3663 if (likely(!alloc_harder))
3664 unusable_free += z->nr_reserved_highatomic;
3667 /* If allocation can't use CMA areas don't use free CMA pages */
3668 if (!(alloc_flags & ALLOC_CMA))
3669 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3672 return unusable_free;
3676 * Return true if free base pages are above 'mark'. For high-order checks it
3677 * will return true of the order-0 watermark is reached and there is at least
3678 * one free page of a suitable size. Checking now avoids taking the zone lock
3679 * to check in the allocation paths if no pages are free.
3681 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3682 int highest_zoneidx, unsigned int alloc_flags,
3687 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3689 /* free_pages may go negative - that's OK */
3690 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3692 if (alloc_flags & ALLOC_HIGH)
3695 if (unlikely(alloc_harder)) {
3697 * OOM victims can try even harder than normal ALLOC_HARDER
3698 * users on the grounds that it's definitely going to be in
3699 * the exit path shortly and free memory. Any allocation it
3700 * makes during the free path will be small and short-lived.
3702 if (alloc_flags & ALLOC_OOM)
3709 * Check watermarks for an order-0 allocation request. If these
3710 * are not met, then a high-order request also cannot go ahead
3711 * even if a suitable page happened to be free.
3713 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3716 /* If this is an order-0 request then the watermark is fine */
3720 /* For a high-order request, check at least one suitable page is free */
3721 for (o = order; o < MAX_ORDER; o++) {
3722 struct free_area *area = &z->free_area[o];
3728 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3729 if (!free_area_empty(area, mt))
3734 if ((alloc_flags & ALLOC_CMA) &&
3735 !free_area_empty(area, MIGRATE_CMA)) {
3739 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3745 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3746 int highest_zoneidx, unsigned int alloc_flags)
3748 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3749 zone_page_state(z, NR_FREE_PAGES));
3752 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3753 unsigned long mark, int highest_zoneidx,
3754 unsigned int alloc_flags, gfp_t gfp_mask)
3758 free_pages = zone_page_state(z, NR_FREE_PAGES);
3761 * Fast check for order-0 only. If this fails then the reserves
3762 * need to be calculated.
3767 fast_free = free_pages;
3768 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3769 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3773 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3777 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3778 * when checking the min watermark. The min watermark is the
3779 * point where boosting is ignored so that kswapd is woken up
3780 * when below the low watermark.
3782 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3783 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3784 mark = z->_watermark[WMARK_MIN];
3785 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3786 alloc_flags, free_pages);
3792 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3793 unsigned long mark, int highest_zoneidx)
3795 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3797 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3798 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3800 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3805 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3807 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3808 node_reclaim_distance;
3810 #else /* CONFIG_NUMA */
3811 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3815 #endif /* CONFIG_NUMA */
3818 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3819 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3820 * premature use of a lower zone may cause lowmem pressure problems that
3821 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3822 * probably too small. It only makes sense to spread allocations to avoid
3823 * fragmentation between the Normal and DMA32 zones.
3825 static inline unsigned int
3826 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3828 unsigned int alloc_flags;
3831 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3834 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3836 #ifdef CONFIG_ZONE_DMA32
3840 if (zone_idx(zone) != ZONE_NORMAL)
3844 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3845 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3846 * on UMA that if Normal is populated then so is DMA32.
3848 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3849 if (nr_online_nodes > 1 && !populated_zone(--zone))
3852 alloc_flags |= ALLOC_NOFRAGMENT;
3853 #endif /* CONFIG_ZONE_DMA32 */
3857 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3858 unsigned int alloc_flags)
3861 unsigned int pflags = current->flags;
3863 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3864 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3865 alloc_flags |= ALLOC_CMA;
3872 * get_page_from_freelist goes through the zonelist trying to allocate
3875 static struct page *
3876 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3877 const struct alloc_context *ac)
3881 struct pglist_data *last_pgdat_dirty_limit = NULL;
3886 * Scan zonelist, looking for a zone with enough free.
3887 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3889 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3890 z = ac->preferred_zoneref;
3891 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3896 if (cpusets_enabled() &&
3897 (alloc_flags & ALLOC_CPUSET) &&
3898 !__cpuset_zone_allowed(zone, gfp_mask))
3901 * When allocating a page cache page for writing, we
3902 * want to get it from a node that is within its dirty
3903 * limit, such that no single node holds more than its
3904 * proportional share of globally allowed dirty pages.
3905 * The dirty limits take into account the node's
3906 * lowmem reserves and high watermark so that kswapd
3907 * should be able to balance it without having to
3908 * write pages from its LRU list.
3910 * XXX: For now, allow allocations to potentially
3911 * exceed the per-node dirty limit in the slowpath
3912 * (spread_dirty_pages unset) before going into reclaim,
3913 * which is important when on a NUMA setup the allowed
3914 * nodes are together not big enough to reach the
3915 * global limit. The proper fix for these situations
3916 * will require awareness of nodes in the
3917 * dirty-throttling and the flusher threads.
3919 if (ac->spread_dirty_pages) {
3920 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3923 if (!node_dirty_ok(zone->zone_pgdat)) {
3924 last_pgdat_dirty_limit = zone->zone_pgdat;
3929 if (no_fallback && nr_online_nodes > 1 &&
3930 zone != ac->preferred_zoneref->zone) {
3934 * If moving to a remote node, retry but allow
3935 * fragmenting fallbacks. Locality is more important
3936 * than fragmentation avoidance.
3938 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3939 if (zone_to_nid(zone) != local_nid) {
3940 alloc_flags &= ~ALLOC_NOFRAGMENT;
3945 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3946 if (!zone_watermark_fast(zone, order, mark,
3947 ac->highest_zoneidx, alloc_flags,
3951 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3953 * Watermark failed for this zone, but see if we can
3954 * grow this zone if it contains deferred pages.
3956 if (static_branch_unlikely(&deferred_pages)) {
3957 if (_deferred_grow_zone(zone, order))
3961 /* Checked here to keep the fast path fast */
3962 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3963 if (alloc_flags & ALLOC_NO_WATERMARKS)
3966 if (node_reclaim_mode == 0 ||
3967 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3970 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3972 case NODE_RECLAIM_NOSCAN:
3975 case NODE_RECLAIM_FULL:
3976 /* scanned but unreclaimable */
3979 /* did we reclaim enough */
3980 if (zone_watermark_ok(zone, order, mark,
3981 ac->highest_zoneidx, alloc_flags))
3989 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3990 gfp_mask, alloc_flags, ac->migratetype);
3992 prep_new_page(page, order, gfp_mask, alloc_flags);
3995 * If this is a high-order atomic allocation then check
3996 * if the pageblock should be reserved for the future
3998 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3999 reserve_highatomic_pageblock(page, zone, order);
4003 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4004 /* Try again if zone has deferred pages */
4005 if (static_branch_unlikely(&deferred_pages)) {
4006 if (_deferred_grow_zone(zone, order))
4014 * It's possible on a UMA machine to get through all zones that are
4015 * fragmented. If avoiding fragmentation, reset and try again.
4018 alloc_flags &= ~ALLOC_NOFRAGMENT;
4025 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4027 unsigned int filter = SHOW_MEM_FILTER_NODES;
4030 * This documents exceptions given to allocations in certain
4031 * contexts that are allowed to allocate outside current's set
4034 if (!(gfp_mask & __GFP_NOMEMALLOC))
4035 if (tsk_is_oom_victim(current) ||
4036 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4037 filter &= ~SHOW_MEM_FILTER_NODES;
4038 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4039 filter &= ~SHOW_MEM_FILTER_NODES;
4041 show_mem(filter, nodemask);
4044 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4046 struct va_format vaf;
4048 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4050 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4053 va_start(args, fmt);
4056 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4057 current->comm, &vaf, gfp_mask, &gfp_mask,
4058 nodemask_pr_args(nodemask));
4061 cpuset_print_current_mems_allowed();
4064 warn_alloc_show_mem(gfp_mask, nodemask);
4067 static inline struct page *
4068 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4069 unsigned int alloc_flags,
4070 const struct alloc_context *ac)
4074 page = get_page_from_freelist(gfp_mask, order,
4075 alloc_flags|ALLOC_CPUSET, ac);
4077 * fallback to ignore cpuset restriction if our nodes
4081 page = get_page_from_freelist(gfp_mask, order,
4087 static inline struct page *
4088 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4089 const struct alloc_context *ac, unsigned long *did_some_progress)
4091 struct oom_control oc = {
4092 .zonelist = ac->zonelist,
4093 .nodemask = ac->nodemask,
4095 .gfp_mask = gfp_mask,
4100 *did_some_progress = 0;
4103 * Acquire the oom lock. If that fails, somebody else is
4104 * making progress for us.
4106 if (!mutex_trylock(&oom_lock)) {
4107 *did_some_progress = 1;
4108 schedule_timeout_uninterruptible(1);
4113 * Go through the zonelist yet one more time, keep very high watermark
4114 * here, this is only to catch a parallel oom killing, we must fail if
4115 * we're still under heavy pressure. But make sure that this reclaim
4116 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4117 * allocation which will never fail due to oom_lock already held.
4119 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4120 ~__GFP_DIRECT_RECLAIM, order,
4121 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4125 /* Coredumps can quickly deplete all memory reserves */
4126 if (current->flags & PF_DUMPCORE)
4128 /* The OOM killer will not help higher order allocs */
4129 if (order > PAGE_ALLOC_COSTLY_ORDER)
4132 * We have already exhausted all our reclaim opportunities without any
4133 * success so it is time to admit defeat. We will skip the OOM killer
4134 * because it is very likely that the caller has a more reasonable
4135 * fallback than shooting a random task.
4137 * The OOM killer may not free memory on a specific node.
4139 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4141 /* The OOM killer does not needlessly kill tasks for lowmem */
4142 if (ac->highest_zoneidx < ZONE_NORMAL)
4144 if (pm_suspended_storage())
4147 * XXX: GFP_NOFS allocations should rather fail than rely on
4148 * other request to make a forward progress.
4149 * We are in an unfortunate situation where out_of_memory cannot
4150 * do much for this context but let's try it to at least get
4151 * access to memory reserved if the current task is killed (see
4152 * out_of_memory). Once filesystems are ready to handle allocation
4153 * failures more gracefully we should just bail out here.
4156 /* Exhausted what can be done so it's blame time */
4157 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4158 *did_some_progress = 1;
4161 * Help non-failing allocations by giving them access to memory
4164 if (gfp_mask & __GFP_NOFAIL)
4165 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4166 ALLOC_NO_WATERMARKS, ac);
4169 mutex_unlock(&oom_lock);
4174 * Maximum number of compaction retries wit a progress before OOM
4175 * killer is consider as the only way to move forward.
4177 #define MAX_COMPACT_RETRIES 16
4179 #ifdef CONFIG_COMPACTION
4180 /* Try memory compaction for high-order allocations before reclaim */
4181 static struct page *
4182 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4183 unsigned int alloc_flags, const struct alloc_context *ac,
4184 enum compact_priority prio, enum compact_result *compact_result)
4186 struct page *page = NULL;
4187 unsigned long pflags;
4188 unsigned int noreclaim_flag;
4193 psi_memstall_enter(&pflags);
4194 noreclaim_flag = memalloc_noreclaim_save();
4196 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4199 memalloc_noreclaim_restore(noreclaim_flag);
4200 psi_memstall_leave(&pflags);
4203 * At least in one zone compaction wasn't deferred or skipped, so let's
4204 * count a compaction stall
4206 count_vm_event(COMPACTSTALL);
4208 /* Prep a captured page if available */
4210 prep_new_page(page, order, gfp_mask, alloc_flags);
4212 /* Try get a page from the freelist if available */
4214 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4217 struct zone *zone = page_zone(page);
4219 zone->compact_blockskip_flush = false;
4220 compaction_defer_reset(zone, order, true);
4221 count_vm_event(COMPACTSUCCESS);
4226 * It's bad if compaction run occurs and fails. The most likely reason
4227 * is that pages exist, but not enough to satisfy watermarks.
4229 count_vm_event(COMPACTFAIL);
4237 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4238 enum compact_result compact_result,
4239 enum compact_priority *compact_priority,
4240 int *compaction_retries)
4242 int max_retries = MAX_COMPACT_RETRIES;
4245 int retries = *compaction_retries;
4246 enum compact_priority priority = *compact_priority;
4251 if (compaction_made_progress(compact_result))
4252 (*compaction_retries)++;
4255 * compaction considers all the zone as desperately out of memory
4256 * so it doesn't really make much sense to retry except when the
4257 * failure could be caused by insufficient priority
4259 if (compaction_failed(compact_result))
4260 goto check_priority;
4263 * compaction was skipped because there are not enough order-0 pages
4264 * to work with, so we retry only if it looks like reclaim can help.
4266 if (compaction_needs_reclaim(compact_result)) {
4267 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4272 * make sure the compaction wasn't deferred or didn't bail out early
4273 * due to locks contention before we declare that we should give up.
4274 * But the next retry should use a higher priority if allowed, so
4275 * we don't just keep bailing out endlessly.
4277 if (compaction_withdrawn(compact_result)) {
4278 goto check_priority;
4282 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4283 * costly ones because they are de facto nofail and invoke OOM
4284 * killer to move on while costly can fail and users are ready
4285 * to cope with that. 1/4 retries is rather arbitrary but we
4286 * would need much more detailed feedback from compaction to
4287 * make a better decision.
4289 if (order > PAGE_ALLOC_COSTLY_ORDER)
4291 if (*compaction_retries <= max_retries) {
4297 * Make sure there are attempts at the highest priority if we exhausted
4298 * all retries or failed at the lower priorities.
4301 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4302 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4304 if (*compact_priority > min_priority) {
4305 (*compact_priority)--;
4306 *compaction_retries = 0;
4310 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4314 static inline struct page *
4315 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4316 unsigned int alloc_flags, const struct alloc_context *ac,
4317 enum compact_priority prio, enum compact_result *compact_result)
4319 *compact_result = COMPACT_SKIPPED;
4324 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4325 enum compact_result compact_result,
4326 enum compact_priority *compact_priority,
4327 int *compaction_retries)
4332 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4336 * There are setups with compaction disabled which would prefer to loop
4337 * inside the allocator rather than hit the oom killer prematurely.
4338 * Let's give them a good hope and keep retrying while the order-0
4339 * watermarks are OK.
4341 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4342 ac->highest_zoneidx, ac->nodemask) {
4343 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4344 ac->highest_zoneidx, alloc_flags))
4349 #endif /* CONFIG_COMPACTION */
4351 #ifdef CONFIG_LOCKDEP
4352 static struct lockdep_map __fs_reclaim_map =
4353 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4355 static bool __need_reclaim(gfp_t gfp_mask)
4357 /* no reclaim without waiting on it */
4358 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4361 /* this guy won't enter reclaim */
4362 if (current->flags & PF_MEMALLOC)
4365 if (gfp_mask & __GFP_NOLOCKDEP)
4371 void __fs_reclaim_acquire(void)
4373 lock_map_acquire(&__fs_reclaim_map);
4376 void __fs_reclaim_release(void)
4378 lock_map_release(&__fs_reclaim_map);
4381 void fs_reclaim_acquire(gfp_t gfp_mask)
4383 gfp_mask = current_gfp_context(gfp_mask);
4385 if (__need_reclaim(gfp_mask)) {
4386 if (gfp_mask & __GFP_FS)
4387 __fs_reclaim_acquire();
4389 #ifdef CONFIG_MMU_NOTIFIER
4390 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4391 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4396 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4398 void fs_reclaim_release(gfp_t gfp_mask)
4400 gfp_mask = current_gfp_context(gfp_mask);
4402 if (__need_reclaim(gfp_mask)) {
4403 if (gfp_mask & __GFP_FS)
4404 __fs_reclaim_release();
4407 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4410 /* Perform direct synchronous page reclaim */
4411 static unsigned long
4412 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4413 const struct alloc_context *ac)
4415 unsigned int noreclaim_flag;
4416 unsigned long pflags, progress;
4420 /* We now go into synchronous reclaim */
4421 cpuset_memory_pressure_bump();
4422 psi_memstall_enter(&pflags);
4423 fs_reclaim_acquire(gfp_mask);
4424 noreclaim_flag = memalloc_noreclaim_save();
4426 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4429 memalloc_noreclaim_restore(noreclaim_flag);
4430 fs_reclaim_release(gfp_mask);
4431 psi_memstall_leave(&pflags);
4438 /* The really slow allocator path where we enter direct reclaim */
4439 static inline struct page *
4440 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4441 unsigned int alloc_flags, const struct alloc_context *ac,
4442 unsigned long *did_some_progress)
4444 struct page *page = NULL;
4445 bool drained = false;
4447 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4448 if (unlikely(!(*did_some_progress)))
4452 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4455 * If an allocation failed after direct reclaim, it could be because
4456 * pages are pinned on the per-cpu lists or in high alloc reserves.
4457 * Shrink them and try again
4459 if (!page && !drained) {
4460 unreserve_highatomic_pageblock(ac, false);
4461 drain_all_pages(NULL);
4469 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4470 const struct alloc_context *ac)
4474 pg_data_t *last_pgdat = NULL;
4475 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4477 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4479 if (last_pgdat != zone->zone_pgdat)
4480 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4481 last_pgdat = zone->zone_pgdat;
4485 static inline unsigned int
4486 gfp_to_alloc_flags(gfp_t gfp_mask)
4488 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4491 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4492 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4493 * to save two branches.
4495 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4496 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4499 * The caller may dip into page reserves a bit more if the caller
4500 * cannot run direct reclaim, or if the caller has realtime scheduling
4501 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4502 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4504 alloc_flags |= (__force int)
4505 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4507 if (gfp_mask & __GFP_ATOMIC) {
4509 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4510 * if it can't schedule.
4512 if (!(gfp_mask & __GFP_NOMEMALLOC))
4513 alloc_flags |= ALLOC_HARDER;
4515 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4516 * comment for __cpuset_node_allowed().
4518 alloc_flags &= ~ALLOC_CPUSET;
4519 } else if (unlikely(rt_task(current)) && !in_interrupt())
4520 alloc_flags |= ALLOC_HARDER;
4522 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4527 static bool oom_reserves_allowed(struct task_struct *tsk)
4529 if (!tsk_is_oom_victim(tsk))
4533 * !MMU doesn't have oom reaper so give access to memory reserves
4534 * only to the thread with TIF_MEMDIE set
4536 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4543 * Distinguish requests which really need access to full memory
4544 * reserves from oom victims which can live with a portion of it
4546 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4548 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4550 if (gfp_mask & __GFP_MEMALLOC)
4551 return ALLOC_NO_WATERMARKS;
4552 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4553 return ALLOC_NO_WATERMARKS;
4554 if (!in_interrupt()) {
4555 if (current->flags & PF_MEMALLOC)
4556 return ALLOC_NO_WATERMARKS;
4557 else if (oom_reserves_allowed(current))
4564 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4566 return !!__gfp_pfmemalloc_flags(gfp_mask);
4570 * Checks whether it makes sense to retry the reclaim to make a forward progress
4571 * for the given allocation request.
4573 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4574 * without success, or when we couldn't even meet the watermark if we
4575 * reclaimed all remaining pages on the LRU lists.
4577 * Returns true if a retry is viable or false to enter the oom path.
4580 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4581 struct alloc_context *ac, int alloc_flags,
4582 bool did_some_progress, int *no_progress_loops)
4589 * Costly allocations might have made a progress but this doesn't mean
4590 * their order will become available due to high fragmentation so
4591 * always increment the no progress counter for them
4593 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4594 *no_progress_loops = 0;
4596 (*no_progress_loops)++;
4599 * Make sure we converge to OOM if we cannot make any progress
4600 * several times in the row.
4602 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4603 /* Before OOM, exhaust highatomic_reserve */
4604 return unreserve_highatomic_pageblock(ac, true);
4608 * Keep reclaiming pages while there is a chance this will lead
4609 * somewhere. If none of the target zones can satisfy our allocation
4610 * request even if all reclaimable pages are considered then we are
4611 * screwed and have to go OOM.
4613 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4614 ac->highest_zoneidx, ac->nodemask) {
4615 unsigned long available;
4616 unsigned long reclaimable;
4617 unsigned long min_wmark = min_wmark_pages(zone);
4620 available = reclaimable = zone_reclaimable_pages(zone);
4621 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4624 * Would the allocation succeed if we reclaimed all
4625 * reclaimable pages?
4627 wmark = __zone_watermark_ok(zone, order, min_wmark,
4628 ac->highest_zoneidx, alloc_flags, available);
4629 trace_reclaim_retry_zone(z, order, reclaimable,
4630 available, min_wmark, *no_progress_loops, wmark);
4633 * If we didn't make any progress and have a lot of
4634 * dirty + writeback pages then we should wait for
4635 * an IO to complete to slow down the reclaim and
4636 * prevent from pre mature OOM
4638 if (!did_some_progress) {
4639 unsigned long write_pending;
4641 write_pending = zone_page_state_snapshot(zone,
4642 NR_ZONE_WRITE_PENDING);
4644 if (2 * write_pending > reclaimable) {
4645 congestion_wait(BLK_RW_ASYNC, HZ/10);
4657 * Memory allocation/reclaim might be called from a WQ context and the
4658 * current implementation of the WQ concurrency control doesn't
4659 * recognize that a particular WQ is congested if the worker thread is
4660 * looping without ever sleeping. Therefore we have to do a short sleep
4661 * here rather than calling cond_resched().
4663 if (current->flags & PF_WQ_WORKER)
4664 schedule_timeout_uninterruptible(1);
4671 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4674 * It's possible that cpuset's mems_allowed and the nodemask from
4675 * mempolicy don't intersect. This should be normally dealt with by
4676 * policy_nodemask(), but it's possible to race with cpuset update in
4677 * such a way the check therein was true, and then it became false
4678 * before we got our cpuset_mems_cookie here.
4679 * This assumes that for all allocations, ac->nodemask can come only
4680 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4681 * when it does not intersect with the cpuset restrictions) or the
4682 * caller can deal with a violated nodemask.
4684 if (cpusets_enabled() && ac->nodemask &&
4685 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4686 ac->nodemask = NULL;
4691 * When updating a task's mems_allowed or mempolicy nodemask, it is
4692 * possible to race with parallel threads in such a way that our
4693 * allocation can fail while the mask is being updated. If we are about
4694 * to fail, check if the cpuset changed during allocation and if so,
4697 if (read_mems_allowed_retry(cpuset_mems_cookie))
4703 static inline struct page *
4704 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4705 struct alloc_context *ac)
4707 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4708 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4709 struct page *page = NULL;
4710 unsigned int alloc_flags;
4711 unsigned long did_some_progress;
4712 enum compact_priority compact_priority;
4713 enum compact_result compact_result;
4714 int compaction_retries;
4715 int no_progress_loops;
4716 unsigned int cpuset_mems_cookie;
4720 * We also sanity check to catch abuse of atomic reserves being used by
4721 * callers that are not in atomic context.
4723 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4724 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4725 gfp_mask &= ~__GFP_ATOMIC;
4728 compaction_retries = 0;
4729 no_progress_loops = 0;
4730 compact_priority = DEF_COMPACT_PRIORITY;
4731 cpuset_mems_cookie = read_mems_allowed_begin();
4734 * The fast path uses conservative alloc_flags to succeed only until
4735 * kswapd needs to be woken up, and to avoid the cost of setting up
4736 * alloc_flags precisely. So we do that now.
4738 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4741 * We need to recalculate the starting point for the zonelist iterator
4742 * because we might have used different nodemask in the fast path, or
4743 * there was a cpuset modification and we are retrying - otherwise we
4744 * could end up iterating over non-eligible zones endlessly.
4746 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4747 ac->highest_zoneidx, ac->nodemask);
4748 if (!ac->preferred_zoneref->zone)
4751 if (alloc_flags & ALLOC_KSWAPD)
4752 wake_all_kswapds(order, gfp_mask, ac);
4755 * The adjusted alloc_flags might result in immediate success, so try
4758 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4763 * For costly allocations, try direct compaction first, as it's likely
4764 * that we have enough base pages and don't need to reclaim. For non-
4765 * movable high-order allocations, do that as well, as compaction will
4766 * try prevent permanent fragmentation by migrating from blocks of the
4768 * Don't try this for allocations that are allowed to ignore
4769 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4771 if (can_direct_reclaim &&
4773 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4774 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4775 page = __alloc_pages_direct_compact(gfp_mask, order,
4777 INIT_COMPACT_PRIORITY,
4783 * Checks for costly allocations with __GFP_NORETRY, which
4784 * includes some THP page fault allocations
4786 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4788 * If allocating entire pageblock(s) and compaction
4789 * failed because all zones are below low watermarks
4790 * or is prohibited because it recently failed at this
4791 * order, fail immediately unless the allocator has
4792 * requested compaction and reclaim retry.
4795 * - potentially very expensive because zones are far
4796 * below their low watermarks or this is part of very
4797 * bursty high order allocations,
4798 * - not guaranteed to help because isolate_freepages()
4799 * may not iterate over freed pages as part of its
4801 * - unlikely to make entire pageblocks free on its
4804 if (compact_result == COMPACT_SKIPPED ||
4805 compact_result == COMPACT_DEFERRED)
4809 * Looks like reclaim/compaction is worth trying, but
4810 * sync compaction could be very expensive, so keep
4811 * using async compaction.
4813 compact_priority = INIT_COMPACT_PRIORITY;
4818 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4819 if (alloc_flags & ALLOC_KSWAPD)
4820 wake_all_kswapds(order, gfp_mask, ac);
4822 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4824 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4827 * Reset the nodemask and zonelist iterators if memory policies can be
4828 * ignored. These allocations are high priority and system rather than
4831 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4832 ac->nodemask = NULL;
4833 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4834 ac->highest_zoneidx, ac->nodemask);
4837 /* Attempt with potentially adjusted zonelist and alloc_flags */
4838 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4842 /* Caller is not willing to reclaim, we can't balance anything */
4843 if (!can_direct_reclaim)
4846 /* Avoid recursion of direct reclaim */
4847 if (current->flags & PF_MEMALLOC)
4850 /* Try direct reclaim and then allocating */
4851 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4852 &did_some_progress);
4856 /* Try direct compaction and then allocating */
4857 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4858 compact_priority, &compact_result);
4862 /* Do not loop if specifically requested */
4863 if (gfp_mask & __GFP_NORETRY)
4867 * Do not retry costly high order allocations unless they are
4868 * __GFP_RETRY_MAYFAIL
4870 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4873 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4874 did_some_progress > 0, &no_progress_loops))
4878 * It doesn't make any sense to retry for the compaction if the order-0
4879 * reclaim is not able to make any progress because the current
4880 * implementation of the compaction depends on the sufficient amount
4881 * of free memory (see __compaction_suitable)
4883 if (did_some_progress > 0 &&
4884 should_compact_retry(ac, order, alloc_flags,
4885 compact_result, &compact_priority,
4886 &compaction_retries))
4890 /* Deal with possible cpuset update races before we start OOM killing */
4891 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4894 /* Reclaim has failed us, start killing things */
4895 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4899 /* Avoid allocations with no watermarks from looping endlessly */
4900 if (tsk_is_oom_victim(current) &&
4901 (alloc_flags & ALLOC_OOM ||
4902 (gfp_mask & __GFP_NOMEMALLOC)))
4905 /* Retry as long as the OOM killer is making progress */
4906 if (did_some_progress) {
4907 no_progress_loops = 0;
4912 /* Deal with possible cpuset update races before we fail */
4913 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4917 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4920 if (gfp_mask & __GFP_NOFAIL) {
4922 * All existing users of the __GFP_NOFAIL are blockable, so warn
4923 * of any new users that actually require GFP_NOWAIT
4925 if (WARN_ON_ONCE(!can_direct_reclaim))
4929 * PF_MEMALLOC request from this context is rather bizarre
4930 * because we cannot reclaim anything and only can loop waiting
4931 * for somebody to do a work for us
4933 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4936 * non failing costly orders are a hard requirement which we
4937 * are not prepared for much so let's warn about these users
4938 * so that we can identify them and convert them to something
4941 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4944 * Help non-failing allocations by giving them access to memory
4945 * reserves but do not use ALLOC_NO_WATERMARKS because this
4946 * could deplete whole memory reserves which would just make
4947 * the situation worse
4949 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4957 warn_alloc(gfp_mask, ac->nodemask,
4958 "page allocation failure: order:%u", order);
4963 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4964 int preferred_nid, nodemask_t *nodemask,
4965 struct alloc_context *ac, gfp_t *alloc_gfp,
4966 unsigned int *alloc_flags)
4968 ac->highest_zoneidx = gfp_zone(gfp_mask);
4969 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4970 ac->nodemask = nodemask;
4971 ac->migratetype = gfp_migratetype(gfp_mask);
4973 if (cpusets_enabled()) {
4974 *alloc_gfp |= __GFP_HARDWALL;
4976 * When we are in the interrupt context, it is irrelevant
4977 * to the current task context. It means that any node ok.
4979 if (!in_interrupt() && !ac->nodemask)
4980 ac->nodemask = &cpuset_current_mems_allowed;
4982 *alloc_flags |= ALLOC_CPUSET;
4985 fs_reclaim_acquire(gfp_mask);
4986 fs_reclaim_release(gfp_mask);
4988 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4990 if (should_fail_alloc_page(gfp_mask, order))
4993 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4995 /* Dirty zone balancing only done in the fast path */
4996 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4999 * The preferred zone is used for statistics but crucially it is
5000 * also used as the starting point for the zonelist iterator. It
5001 * may get reset for allocations that ignore memory policies.
5003 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5004 ac->highest_zoneidx, ac->nodemask);
5010 * This is the 'heart' of the zoned buddy allocator.
5012 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5013 nodemask_t *nodemask)
5016 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5017 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5018 struct alloc_context ac = { };
5021 * There are several places where we assume that the order value is sane
5022 * so bail out early if the request is out of bound.
5024 if (unlikely(order >= MAX_ORDER)) {
5025 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5029 gfp &= gfp_allowed_mask;
5031 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5032 &alloc_gfp, &alloc_flags))
5036 * Forbid the first pass from falling back to types that fragment
5037 * memory until all local zones are considered.
5039 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5041 /* First allocation attempt */
5042 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5047 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5048 * resp. GFP_NOIO which has to be inherited for all allocation requests
5049 * from a particular context which has been marked by
5050 * memalloc_no{fs,io}_{save,restore}.
5052 alloc_gfp = current_gfp_context(gfp);
5053 ac.spread_dirty_pages = false;
5056 * Restore the original nodemask if it was potentially replaced with
5057 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5059 ac.nodemask = nodemask;
5061 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5064 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5065 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5066 __free_pages(page, order);
5070 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5074 EXPORT_SYMBOL(__alloc_pages);
5077 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5078 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5079 * you need to access high mem.
5081 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5085 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5088 return (unsigned long) page_address(page);
5090 EXPORT_SYMBOL(__get_free_pages);
5092 unsigned long get_zeroed_page(gfp_t gfp_mask)
5094 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5096 EXPORT_SYMBOL(get_zeroed_page);
5098 static inline void free_the_page(struct page *page, unsigned int order)
5100 if (order == 0) /* Via pcp? */
5101 free_unref_page(page);
5103 __free_pages_ok(page, order, FPI_NONE);
5107 * __free_pages - Free pages allocated with alloc_pages().
5108 * @page: The page pointer returned from alloc_pages().
5109 * @order: The order of the allocation.
5111 * This function can free multi-page allocations that are not compound
5112 * pages. It does not check that the @order passed in matches that of
5113 * the allocation, so it is easy to leak memory. Freeing more memory
5114 * than was allocated will probably emit a warning.
5116 * If the last reference to this page is speculative, it will be released
5117 * by put_page() which only frees the first page of a non-compound
5118 * allocation. To prevent the remaining pages from being leaked, we free
5119 * the subsequent pages here. If you want to use the page's reference
5120 * count to decide when to free the allocation, you should allocate a
5121 * compound page, and use put_page() instead of __free_pages().
5123 * Context: May be called in interrupt context or while holding a normal
5124 * spinlock, but not in NMI context or while holding a raw spinlock.
5126 void __free_pages(struct page *page, unsigned int order)
5128 if (put_page_testzero(page))
5129 free_the_page(page, order);
5130 else if (!PageHead(page))
5132 free_the_page(page + (1 << order), order);
5134 EXPORT_SYMBOL(__free_pages);
5136 void free_pages(unsigned long addr, unsigned int order)
5139 VM_BUG_ON(!virt_addr_valid((void *)addr));
5140 __free_pages(virt_to_page((void *)addr), order);
5144 EXPORT_SYMBOL(free_pages);
5148 * An arbitrary-length arbitrary-offset area of memory which resides
5149 * within a 0 or higher order page. Multiple fragments within that page
5150 * are individually refcounted, in the page's reference counter.
5152 * The page_frag functions below provide a simple allocation framework for
5153 * page fragments. This is used by the network stack and network device
5154 * drivers to provide a backing region of memory for use as either an
5155 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5157 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5160 struct page *page = NULL;
5161 gfp_t gfp = gfp_mask;
5163 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5164 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5166 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5167 PAGE_FRAG_CACHE_MAX_ORDER);
5168 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5170 if (unlikely(!page))
5171 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5173 nc->va = page ? page_address(page) : NULL;
5178 void __page_frag_cache_drain(struct page *page, unsigned int count)
5180 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5182 if (page_ref_sub_and_test(page, count))
5183 free_the_page(page, compound_order(page));
5185 EXPORT_SYMBOL(__page_frag_cache_drain);
5187 void *page_frag_alloc_align(struct page_frag_cache *nc,
5188 unsigned int fragsz, gfp_t gfp_mask,
5189 unsigned int align_mask)
5191 unsigned int size = PAGE_SIZE;
5195 if (unlikely(!nc->va)) {
5197 page = __page_frag_cache_refill(nc, gfp_mask);
5201 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5202 /* if size can vary use size else just use PAGE_SIZE */
5205 /* Even if we own the page, we do not use atomic_set().
5206 * This would break get_page_unless_zero() users.
5208 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5210 /* reset page count bias and offset to start of new frag */
5211 nc->pfmemalloc = page_is_pfmemalloc(page);
5212 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5216 offset = nc->offset - fragsz;
5217 if (unlikely(offset < 0)) {
5218 page = virt_to_page(nc->va);
5220 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5223 if (unlikely(nc->pfmemalloc)) {
5224 free_the_page(page, compound_order(page));
5228 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5229 /* if size can vary use size else just use PAGE_SIZE */
5232 /* OK, page count is 0, we can safely set it */
5233 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5235 /* reset page count bias and offset to start of new frag */
5236 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5237 offset = size - fragsz;
5241 offset &= align_mask;
5242 nc->offset = offset;
5244 return nc->va + offset;
5246 EXPORT_SYMBOL(page_frag_alloc_align);
5249 * Frees a page fragment allocated out of either a compound or order 0 page.
5251 void page_frag_free(void *addr)
5253 struct page *page = virt_to_head_page(addr);
5255 if (unlikely(put_page_testzero(page)))
5256 free_the_page(page, compound_order(page));
5258 EXPORT_SYMBOL(page_frag_free);
5260 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5264 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5265 unsigned long used = addr + PAGE_ALIGN(size);
5267 split_page(virt_to_page((void *)addr), order);
5268 while (used < alloc_end) {
5273 return (void *)addr;
5277 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5278 * @size: the number of bytes to allocate
5279 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5281 * This function is similar to alloc_pages(), except that it allocates the
5282 * minimum number of pages to satisfy the request. alloc_pages() can only
5283 * allocate memory in power-of-two pages.
5285 * This function is also limited by MAX_ORDER.
5287 * Memory allocated by this function must be released by free_pages_exact().
5289 * Return: pointer to the allocated area or %NULL in case of error.
5291 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5293 unsigned int order = get_order(size);
5296 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5297 gfp_mask &= ~__GFP_COMP;
5299 addr = __get_free_pages(gfp_mask, order);
5300 return make_alloc_exact(addr, order, size);
5302 EXPORT_SYMBOL(alloc_pages_exact);
5305 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5307 * @nid: the preferred node ID where memory should be allocated
5308 * @size: the number of bytes to allocate
5309 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5311 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5314 * Return: pointer to the allocated area or %NULL in case of error.
5316 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5318 unsigned int order = get_order(size);
5321 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5322 gfp_mask &= ~__GFP_COMP;
5324 p = alloc_pages_node(nid, gfp_mask, order);
5327 return make_alloc_exact((unsigned long)page_address(p), order, size);
5331 * free_pages_exact - release memory allocated via alloc_pages_exact()
5332 * @virt: the value returned by alloc_pages_exact.
5333 * @size: size of allocation, same value as passed to alloc_pages_exact().
5335 * Release the memory allocated by a previous call to alloc_pages_exact.
5337 void free_pages_exact(void *virt, size_t size)
5339 unsigned long addr = (unsigned long)virt;
5340 unsigned long end = addr + PAGE_ALIGN(size);
5342 while (addr < end) {
5347 EXPORT_SYMBOL(free_pages_exact);
5350 * nr_free_zone_pages - count number of pages beyond high watermark
5351 * @offset: The zone index of the highest zone
5353 * nr_free_zone_pages() counts the number of pages which are beyond the
5354 * high watermark within all zones at or below a given zone index. For each
5355 * zone, the number of pages is calculated as:
5357 * nr_free_zone_pages = managed_pages - high_pages
5359 * Return: number of pages beyond high watermark.
5361 static unsigned long nr_free_zone_pages(int offset)
5366 /* Just pick one node, since fallback list is circular */
5367 unsigned long sum = 0;
5369 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5371 for_each_zone_zonelist(zone, z, zonelist, offset) {
5372 unsigned long size = zone_managed_pages(zone);
5373 unsigned long high = high_wmark_pages(zone);
5382 * nr_free_buffer_pages - count number of pages beyond high watermark
5384 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5385 * watermark within ZONE_DMA and ZONE_NORMAL.
5387 * Return: number of pages beyond high watermark within ZONE_DMA and
5390 unsigned long nr_free_buffer_pages(void)
5392 return nr_free_zone_pages(gfp_zone(GFP_USER));
5394 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5396 static inline void show_node(struct zone *zone)
5398 if (IS_ENABLED(CONFIG_NUMA))
5399 printk("Node %d ", zone_to_nid(zone));
5402 long si_mem_available(void)
5405 unsigned long pagecache;
5406 unsigned long wmark_low = 0;
5407 unsigned long pages[NR_LRU_LISTS];
5408 unsigned long reclaimable;
5412 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5413 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5416 wmark_low += low_wmark_pages(zone);
5419 * Estimate the amount of memory available for userspace allocations,
5420 * without causing swapping.
5422 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5425 * Not all the page cache can be freed, otherwise the system will
5426 * start swapping. Assume at least half of the page cache, or the
5427 * low watermark worth of cache, needs to stay.
5429 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5430 pagecache -= min(pagecache / 2, wmark_low);
5431 available += pagecache;
5434 * Part of the reclaimable slab and other kernel memory consists of
5435 * items that are in use, and cannot be freed. Cap this estimate at the
5438 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5439 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5440 available += reclaimable - min(reclaimable / 2, wmark_low);
5446 EXPORT_SYMBOL_GPL(si_mem_available);
5448 void si_meminfo(struct sysinfo *val)
5450 val->totalram = totalram_pages();
5451 val->sharedram = global_node_page_state(NR_SHMEM);
5452 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5453 val->bufferram = nr_blockdev_pages();
5454 val->totalhigh = totalhigh_pages();
5455 val->freehigh = nr_free_highpages();
5456 val->mem_unit = PAGE_SIZE;
5459 EXPORT_SYMBOL(si_meminfo);
5462 void si_meminfo_node(struct sysinfo *val, int nid)
5464 int zone_type; /* needs to be signed */
5465 unsigned long managed_pages = 0;
5466 unsigned long managed_highpages = 0;
5467 unsigned long free_highpages = 0;
5468 pg_data_t *pgdat = NODE_DATA(nid);
5470 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5471 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5472 val->totalram = managed_pages;
5473 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5474 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5475 #ifdef CONFIG_HIGHMEM
5476 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5477 struct zone *zone = &pgdat->node_zones[zone_type];
5479 if (is_highmem(zone)) {
5480 managed_highpages += zone_managed_pages(zone);
5481 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5484 val->totalhigh = managed_highpages;
5485 val->freehigh = free_highpages;
5487 val->totalhigh = managed_highpages;
5488 val->freehigh = free_highpages;
5490 val->mem_unit = PAGE_SIZE;
5495 * Determine whether the node should be displayed or not, depending on whether
5496 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5498 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5500 if (!(flags & SHOW_MEM_FILTER_NODES))
5504 * no node mask - aka implicit memory numa policy. Do not bother with
5505 * the synchronization - read_mems_allowed_begin - because we do not
5506 * have to be precise here.
5509 nodemask = &cpuset_current_mems_allowed;
5511 return !node_isset(nid, *nodemask);
5514 #define K(x) ((x) << (PAGE_SHIFT-10))
5516 static void show_migration_types(unsigned char type)
5518 static const char types[MIGRATE_TYPES] = {
5519 [MIGRATE_UNMOVABLE] = 'U',
5520 [MIGRATE_MOVABLE] = 'M',
5521 [MIGRATE_RECLAIMABLE] = 'E',
5522 [MIGRATE_HIGHATOMIC] = 'H',
5524 [MIGRATE_CMA] = 'C',
5526 #ifdef CONFIG_MEMORY_ISOLATION
5527 [MIGRATE_ISOLATE] = 'I',
5530 char tmp[MIGRATE_TYPES + 1];
5534 for (i = 0; i < MIGRATE_TYPES; i++) {
5535 if (type & (1 << i))
5540 printk(KERN_CONT "(%s) ", tmp);
5544 * Show free area list (used inside shift_scroll-lock stuff)
5545 * We also calculate the percentage fragmentation. We do this by counting the
5546 * memory on each free list with the exception of the first item on the list.
5549 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5552 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5554 unsigned long free_pcp = 0;
5559 for_each_populated_zone(zone) {
5560 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5563 for_each_online_cpu(cpu)
5564 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5567 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5568 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5569 " unevictable:%lu dirty:%lu writeback:%lu\n"
5570 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5571 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5572 " free:%lu free_pcp:%lu free_cma:%lu\n",
5573 global_node_page_state(NR_ACTIVE_ANON),
5574 global_node_page_state(NR_INACTIVE_ANON),
5575 global_node_page_state(NR_ISOLATED_ANON),
5576 global_node_page_state(NR_ACTIVE_FILE),
5577 global_node_page_state(NR_INACTIVE_FILE),
5578 global_node_page_state(NR_ISOLATED_FILE),
5579 global_node_page_state(NR_UNEVICTABLE),
5580 global_node_page_state(NR_FILE_DIRTY),
5581 global_node_page_state(NR_WRITEBACK),
5582 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5583 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5584 global_node_page_state(NR_FILE_MAPPED),
5585 global_node_page_state(NR_SHMEM),
5586 global_node_page_state(NR_PAGETABLE),
5587 global_zone_page_state(NR_BOUNCE),
5588 global_zone_page_state(NR_FREE_PAGES),
5590 global_zone_page_state(NR_FREE_CMA_PAGES));
5592 for_each_online_pgdat(pgdat) {
5593 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5597 " active_anon:%lukB"
5598 " inactive_anon:%lukB"
5599 " active_file:%lukB"
5600 " inactive_file:%lukB"
5601 " unevictable:%lukB"
5602 " isolated(anon):%lukB"
5603 " isolated(file):%lukB"
5608 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5610 " shmem_pmdmapped: %lukB"
5613 " writeback_tmp:%lukB"
5614 " kernel_stack:%lukB"
5615 #ifdef CONFIG_SHADOW_CALL_STACK
5616 " shadow_call_stack:%lukB"
5619 " all_unreclaimable? %s"
5622 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5623 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5624 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5625 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5626 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5627 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5628 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5629 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5630 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5631 K(node_page_state(pgdat, NR_WRITEBACK)),
5632 K(node_page_state(pgdat, NR_SHMEM)),
5633 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5634 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5635 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5636 K(node_page_state(pgdat, NR_ANON_THPS)),
5638 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5639 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5640 #ifdef CONFIG_SHADOW_CALL_STACK
5641 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5643 K(node_page_state(pgdat, NR_PAGETABLE)),
5644 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5648 for_each_populated_zone(zone) {
5651 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5655 for_each_online_cpu(cpu)
5656 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5665 " reserved_highatomic:%luKB"
5666 " active_anon:%lukB"
5667 " inactive_anon:%lukB"
5668 " active_file:%lukB"
5669 " inactive_file:%lukB"
5670 " unevictable:%lukB"
5671 " writepending:%lukB"
5681 K(zone_page_state(zone, NR_FREE_PAGES)),
5682 K(min_wmark_pages(zone)),
5683 K(low_wmark_pages(zone)),
5684 K(high_wmark_pages(zone)),
5685 K(zone->nr_reserved_highatomic),
5686 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5687 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5688 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5689 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5690 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5691 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5692 K(zone->present_pages),
5693 K(zone_managed_pages(zone)),
5694 K(zone_page_state(zone, NR_MLOCK)),
5695 K(zone_page_state(zone, NR_BOUNCE)),
5697 K(this_cpu_read(zone->pageset->pcp.count)),
5698 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5699 printk("lowmem_reserve[]:");
5700 for (i = 0; i < MAX_NR_ZONES; i++)
5701 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5702 printk(KERN_CONT "\n");
5705 for_each_populated_zone(zone) {
5707 unsigned long nr[MAX_ORDER], flags, total = 0;
5708 unsigned char types[MAX_ORDER];
5710 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5713 printk(KERN_CONT "%s: ", zone->name);
5715 spin_lock_irqsave(&zone->lock, flags);
5716 for (order = 0; order < MAX_ORDER; order++) {
5717 struct free_area *area = &zone->free_area[order];
5720 nr[order] = area->nr_free;
5721 total += nr[order] << order;
5724 for (type = 0; type < MIGRATE_TYPES; type++) {
5725 if (!free_area_empty(area, type))
5726 types[order] |= 1 << type;
5729 spin_unlock_irqrestore(&zone->lock, flags);
5730 for (order = 0; order < MAX_ORDER; order++) {
5731 printk(KERN_CONT "%lu*%lukB ",
5732 nr[order], K(1UL) << order);
5734 show_migration_types(types[order]);
5736 printk(KERN_CONT "= %lukB\n", K(total));
5739 hugetlb_show_meminfo();
5741 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5743 show_swap_cache_info();
5746 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5748 zoneref->zone = zone;
5749 zoneref->zone_idx = zone_idx(zone);
5753 * Builds allocation fallback zone lists.
5755 * Add all populated zones of a node to the zonelist.
5757 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5760 enum zone_type zone_type = MAX_NR_ZONES;
5765 zone = pgdat->node_zones + zone_type;
5766 if (managed_zone(zone)) {
5767 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5768 check_highest_zone(zone_type);
5770 } while (zone_type);
5777 static int __parse_numa_zonelist_order(char *s)
5780 * We used to support different zonlists modes but they turned
5781 * out to be just not useful. Let's keep the warning in place
5782 * if somebody still use the cmd line parameter so that we do
5783 * not fail it silently
5785 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5786 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5792 char numa_zonelist_order[] = "Node";
5795 * sysctl handler for numa_zonelist_order
5797 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5798 void *buffer, size_t *length, loff_t *ppos)
5801 return __parse_numa_zonelist_order(buffer);
5802 return proc_dostring(table, write, buffer, length, ppos);
5806 #define MAX_NODE_LOAD (nr_online_nodes)
5807 static int node_load[MAX_NUMNODES];
5810 * find_next_best_node - find the next node that should appear in a given node's fallback list
5811 * @node: node whose fallback list we're appending
5812 * @used_node_mask: nodemask_t of already used nodes
5814 * We use a number of factors to determine which is the next node that should
5815 * appear on a given node's fallback list. The node should not have appeared
5816 * already in @node's fallback list, and it should be the next closest node
5817 * according to the distance array (which contains arbitrary distance values
5818 * from each node to each node in the system), and should also prefer nodes
5819 * with no CPUs, since presumably they'll have very little allocation pressure
5820 * on them otherwise.
5822 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5824 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5827 int min_val = INT_MAX;
5828 int best_node = NUMA_NO_NODE;
5830 /* Use the local node if we haven't already */
5831 if (!node_isset(node, *used_node_mask)) {
5832 node_set(node, *used_node_mask);
5836 for_each_node_state(n, N_MEMORY) {
5838 /* Don't want a node to appear more than once */
5839 if (node_isset(n, *used_node_mask))
5842 /* Use the distance array to find the distance */
5843 val = node_distance(node, n);
5845 /* Penalize nodes under us ("prefer the next node") */
5848 /* Give preference to headless and unused nodes */
5849 if (!cpumask_empty(cpumask_of_node(n)))
5850 val += PENALTY_FOR_NODE_WITH_CPUS;
5852 /* Slight preference for less loaded node */
5853 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5854 val += node_load[n];
5856 if (val < min_val) {
5863 node_set(best_node, *used_node_mask);
5870 * Build zonelists ordered by node and zones within node.
5871 * This results in maximum locality--normal zone overflows into local
5872 * DMA zone, if any--but risks exhausting DMA zone.
5874 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5877 struct zoneref *zonerefs;
5880 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5882 for (i = 0; i < nr_nodes; i++) {
5885 pg_data_t *node = NODE_DATA(node_order[i]);
5887 nr_zones = build_zonerefs_node(node, zonerefs);
5888 zonerefs += nr_zones;
5890 zonerefs->zone = NULL;
5891 zonerefs->zone_idx = 0;
5895 * Build gfp_thisnode zonelists
5897 static void build_thisnode_zonelists(pg_data_t *pgdat)
5899 struct zoneref *zonerefs;
5902 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5903 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5904 zonerefs += nr_zones;
5905 zonerefs->zone = NULL;
5906 zonerefs->zone_idx = 0;
5910 * Build zonelists ordered by zone and nodes within zones.
5911 * This results in conserving DMA zone[s] until all Normal memory is
5912 * exhausted, but results in overflowing to remote node while memory
5913 * may still exist in local DMA zone.
5916 static void build_zonelists(pg_data_t *pgdat)
5918 static int node_order[MAX_NUMNODES];
5919 int node, load, nr_nodes = 0;
5920 nodemask_t used_mask = NODE_MASK_NONE;
5921 int local_node, prev_node;
5923 /* NUMA-aware ordering of nodes */
5924 local_node = pgdat->node_id;
5925 load = nr_online_nodes;
5926 prev_node = local_node;
5928 memset(node_order, 0, sizeof(node_order));
5929 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5931 * We don't want to pressure a particular node.
5932 * So adding penalty to the first node in same
5933 * distance group to make it round-robin.
5935 if (node_distance(local_node, node) !=
5936 node_distance(local_node, prev_node))
5937 node_load[node] = load;
5939 node_order[nr_nodes++] = node;
5944 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5945 build_thisnode_zonelists(pgdat);
5948 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5950 * Return node id of node used for "local" allocations.
5951 * I.e., first node id of first zone in arg node's generic zonelist.
5952 * Used for initializing percpu 'numa_mem', which is used primarily
5953 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5955 int local_memory_node(int node)
5959 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5960 gfp_zone(GFP_KERNEL),
5962 return zone_to_nid(z->zone);
5966 static void setup_min_unmapped_ratio(void);
5967 static void setup_min_slab_ratio(void);
5968 #else /* CONFIG_NUMA */
5970 static void build_zonelists(pg_data_t *pgdat)
5972 int node, local_node;
5973 struct zoneref *zonerefs;
5976 local_node = pgdat->node_id;
5978 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5979 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5980 zonerefs += nr_zones;
5983 * Now we build the zonelist so that it contains the zones
5984 * of all the other nodes.
5985 * We don't want to pressure a particular node, so when
5986 * building the zones for node N, we make sure that the
5987 * zones coming right after the local ones are those from
5988 * node N+1 (modulo N)
5990 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5991 if (!node_online(node))
5993 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5994 zonerefs += nr_zones;
5996 for (node = 0; node < local_node; node++) {
5997 if (!node_online(node))
5999 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6000 zonerefs += nr_zones;
6003 zonerefs->zone = NULL;
6004 zonerefs->zone_idx = 0;
6007 #endif /* CONFIG_NUMA */
6010 * Boot pageset table. One per cpu which is going to be used for all
6011 * zones and all nodes. The parameters will be set in such a way
6012 * that an item put on a list will immediately be handed over to
6013 * the buddy list. This is safe since pageset manipulation is done
6014 * with interrupts disabled.
6016 * The boot_pagesets must be kept even after bootup is complete for
6017 * unused processors and/or zones. They do play a role for bootstrapping
6018 * hotplugged processors.
6020 * zoneinfo_show() and maybe other functions do
6021 * not check if the processor is online before following the pageset pointer.
6022 * Other parts of the kernel may not check if the zone is available.
6024 static void pageset_init(struct per_cpu_pageset *p);
6025 /* These effectively disable the pcplists in the boot pageset completely */
6026 #define BOOT_PAGESET_HIGH 0
6027 #define BOOT_PAGESET_BATCH 1
6028 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
6029 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6031 static void __build_all_zonelists(void *data)
6034 int __maybe_unused cpu;
6035 pg_data_t *self = data;
6036 static DEFINE_SPINLOCK(lock);
6041 memset(node_load, 0, sizeof(node_load));
6045 * This node is hotadded and no memory is yet present. So just
6046 * building zonelists is fine - no need to touch other nodes.
6048 if (self && !node_online(self->node_id)) {
6049 build_zonelists(self);
6051 for_each_online_node(nid) {
6052 pg_data_t *pgdat = NODE_DATA(nid);
6054 build_zonelists(pgdat);
6057 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6059 * We now know the "local memory node" for each node--
6060 * i.e., the node of the first zone in the generic zonelist.
6061 * Set up numa_mem percpu variable for on-line cpus. During
6062 * boot, only the boot cpu should be on-line; we'll init the
6063 * secondary cpus' numa_mem as they come on-line. During
6064 * node/memory hotplug, we'll fixup all on-line cpus.
6066 for_each_online_cpu(cpu)
6067 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6074 static noinline void __init
6075 build_all_zonelists_init(void)
6079 __build_all_zonelists(NULL);
6082 * Initialize the boot_pagesets that are going to be used
6083 * for bootstrapping processors. The real pagesets for
6084 * each zone will be allocated later when the per cpu
6085 * allocator is available.
6087 * boot_pagesets are used also for bootstrapping offline
6088 * cpus if the system is already booted because the pagesets
6089 * are needed to initialize allocators on a specific cpu too.
6090 * F.e. the percpu allocator needs the page allocator which
6091 * needs the percpu allocator in order to allocate its pagesets
6092 * (a chicken-egg dilemma).
6094 for_each_possible_cpu(cpu)
6095 pageset_init(&per_cpu(boot_pageset, cpu));
6097 mminit_verify_zonelist();
6098 cpuset_init_current_mems_allowed();
6102 * unless system_state == SYSTEM_BOOTING.
6104 * __ref due to call of __init annotated helper build_all_zonelists_init
6105 * [protected by SYSTEM_BOOTING].
6107 void __ref build_all_zonelists(pg_data_t *pgdat)
6109 unsigned long vm_total_pages;
6111 if (system_state == SYSTEM_BOOTING) {
6112 build_all_zonelists_init();
6114 __build_all_zonelists(pgdat);
6115 /* cpuset refresh routine should be here */
6117 /* Get the number of free pages beyond high watermark in all zones. */
6118 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6120 * Disable grouping by mobility if the number of pages in the
6121 * system is too low to allow the mechanism to work. It would be
6122 * more accurate, but expensive to check per-zone. This check is
6123 * made on memory-hotadd so a system can start with mobility
6124 * disabled and enable it later
6126 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6127 page_group_by_mobility_disabled = 1;
6129 page_group_by_mobility_disabled = 0;
6131 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6133 page_group_by_mobility_disabled ? "off" : "on",
6136 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6140 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6141 static bool __meminit
6142 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6144 static struct memblock_region *r;
6146 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6147 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6148 for_each_mem_region(r) {
6149 if (*pfn < memblock_region_memory_end_pfn(r))
6153 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6154 memblock_is_mirror(r)) {
6155 *pfn = memblock_region_memory_end_pfn(r);
6163 * Initially all pages are reserved - free ones are freed
6164 * up by memblock_free_all() once the early boot process is
6165 * done. Non-atomic initialization, single-pass.
6167 * All aligned pageblocks are initialized to the specified migratetype
6168 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6169 * zone stats (e.g., nr_isolate_pageblock) are touched.
6171 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6172 unsigned long start_pfn, unsigned long zone_end_pfn,
6173 enum meminit_context context,
6174 struct vmem_altmap *altmap, int migratetype)
6176 unsigned long pfn, end_pfn = start_pfn + size;
6179 if (highest_memmap_pfn < end_pfn - 1)
6180 highest_memmap_pfn = end_pfn - 1;
6182 #ifdef CONFIG_ZONE_DEVICE
6184 * Honor reservation requested by the driver for this ZONE_DEVICE
6185 * memory. We limit the total number of pages to initialize to just
6186 * those that might contain the memory mapping. We will defer the
6187 * ZONE_DEVICE page initialization until after we have released
6190 if (zone == ZONE_DEVICE) {
6194 if (start_pfn == altmap->base_pfn)
6195 start_pfn += altmap->reserve;
6196 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6200 for (pfn = start_pfn; pfn < end_pfn; ) {
6202 * There can be holes in boot-time mem_map[]s handed to this
6203 * function. They do not exist on hotplugged memory.
6205 if (context == MEMINIT_EARLY) {
6206 if (overlap_memmap_init(zone, &pfn))
6208 if (defer_init(nid, pfn, zone_end_pfn))
6212 page = pfn_to_page(pfn);
6213 __init_single_page(page, pfn, zone, nid);
6214 if (context == MEMINIT_HOTPLUG)
6215 __SetPageReserved(page);
6218 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6219 * such that unmovable allocations won't be scattered all
6220 * over the place during system boot.
6222 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6223 set_pageblock_migratetype(page, migratetype);
6230 #ifdef CONFIG_ZONE_DEVICE
6231 void __ref memmap_init_zone_device(struct zone *zone,
6232 unsigned long start_pfn,
6233 unsigned long nr_pages,
6234 struct dev_pagemap *pgmap)
6236 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6237 struct pglist_data *pgdat = zone->zone_pgdat;
6238 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6239 unsigned long zone_idx = zone_idx(zone);
6240 unsigned long start = jiffies;
6241 int nid = pgdat->node_id;
6243 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6247 * The call to memmap_init_zone should have already taken care
6248 * of the pages reserved for the memmap, so we can just jump to
6249 * the end of that region and start processing the device pages.
6252 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6253 nr_pages = end_pfn - start_pfn;
6256 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6257 struct page *page = pfn_to_page(pfn);
6259 __init_single_page(page, pfn, zone_idx, nid);
6262 * Mark page reserved as it will need to wait for onlining
6263 * phase for it to be fully associated with a zone.
6265 * We can use the non-atomic __set_bit operation for setting
6266 * the flag as we are still initializing the pages.
6268 __SetPageReserved(page);
6271 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6272 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6273 * ever freed or placed on a driver-private list.
6275 page->pgmap = pgmap;
6276 page->zone_device_data = NULL;
6279 * Mark the block movable so that blocks are reserved for
6280 * movable at startup. This will force kernel allocations
6281 * to reserve their blocks rather than leaking throughout
6282 * the address space during boot when many long-lived
6283 * kernel allocations are made.
6285 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6286 * because this is done early in section_activate()
6288 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6289 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6294 pr_info("%s initialised %lu pages in %ums\n", __func__,
6295 nr_pages, jiffies_to_msecs(jiffies - start));
6299 static void __meminit zone_init_free_lists(struct zone *zone)
6301 unsigned int order, t;
6302 for_each_migratetype_order(order, t) {
6303 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6304 zone->free_area[order].nr_free = 0;
6308 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6310 * Only struct pages that correspond to ranges defined by memblock.memory
6311 * are zeroed and initialized by going through __init_single_page() during
6312 * memmap_init_zone().
6314 * But, there could be struct pages that correspond to holes in
6315 * memblock.memory. This can happen because of the following reasons:
6316 * - physical memory bank size is not necessarily the exact multiple of the
6317 * arbitrary section size
6318 * - early reserved memory may not be listed in memblock.memory
6319 * - memory layouts defined with memmap= kernel parameter may not align
6320 * nicely with memmap sections
6322 * Explicitly initialize those struct pages so that:
6323 * - PG_Reserved is set
6324 * - zone and node links point to zone and node that span the page if the
6325 * hole is in the middle of a zone
6326 * - zone and node links point to adjacent zone/node if the hole falls on
6327 * the zone boundary; the pages in such holes will be prepended to the
6328 * zone/node above the hole except for the trailing pages in the last
6329 * section that will be appended to the zone/node below.
6331 static u64 __meminit init_unavailable_range(unsigned long spfn,
6338 for (pfn = spfn; pfn < epfn; pfn++) {
6339 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6340 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6341 + pageblock_nr_pages - 1;
6344 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6345 __SetPageReserved(pfn_to_page(pfn));
6352 static inline u64 init_unavailable_range(unsigned long spfn, unsigned long epfn,
6359 void __meminit __weak memmap_init_zone(struct zone *zone)
6361 unsigned long zone_start_pfn = zone->zone_start_pfn;
6362 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6363 int i, nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6364 static unsigned long hole_pfn;
6365 unsigned long start_pfn, end_pfn;
6368 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6369 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6370 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6372 if (end_pfn > start_pfn)
6373 memmap_init_range(end_pfn - start_pfn, nid,
6374 zone_id, start_pfn, zone_end_pfn,
6375 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6377 if (hole_pfn < start_pfn)
6378 pgcnt += init_unavailable_range(hole_pfn, start_pfn,
6383 #ifdef CONFIG_SPARSEMEM
6385 * Initialize the hole in the range [zone_end_pfn, section_end].
6386 * If zone boundary falls in the middle of a section, this hole
6387 * will be re-initialized during the call to this function for the
6390 end_pfn = round_up(zone_end_pfn, PAGES_PER_SECTION);
6391 if (hole_pfn < end_pfn)
6392 pgcnt += init_unavailable_range(hole_pfn, end_pfn,
6397 pr_info(" %s zone: %llu pages in unavailable ranges\n",
6401 static int zone_batchsize(struct zone *zone)
6407 * The per-cpu-pages pools are set to around 1000th of the
6410 batch = zone_managed_pages(zone) / 1024;
6411 /* But no more than a meg. */
6412 if (batch * PAGE_SIZE > 1024 * 1024)
6413 batch = (1024 * 1024) / PAGE_SIZE;
6414 batch /= 4; /* We effectively *= 4 below */
6419 * Clamp the batch to a 2^n - 1 value. Having a power
6420 * of 2 value was found to be more likely to have
6421 * suboptimal cache aliasing properties in some cases.
6423 * For example if 2 tasks are alternately allocating
6424 * batches of pages, one task can end up with a lot
6425 * of pages of one half of the possible page colors
6426 * and the other with pages of the other colors.
6428 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6433 /* The deferral and batching of frees should be suppressed under NOMMU
6436 * The problem is that NOMMU needs to be able to allocate large chunks
6437 * of contiguous memory as there's no hardware page translation to
6438 * assemble apparent contiguous memory from discontiguous pages.
6440 * Queueing large contiguous runs of pages for batching, however,
6441 * causes the pages to actually be freed in smaller chunks. As there
6442 * can be a significant delay between the individual batches being
6443 * recycled, this leads to the once large chunks of space being
6444 * fragmented and becoming unavailable for high-order allocations.
6451 * pcp->high and pcp->batch values are related and generally batch is lower
6452 * than high. They are also related to pcp->count such that count is lower
6453 * than high, and as soon as it reaches high, the pcplist is flushed.
6455 * However, guaranteeing these relations at all times would require e.g. write
6456 * barriers here but also careful usage of read barriers at the read side, and
6457 * thus be prone to error and bad for performance. Thus the update only prevents
6458 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6459 * can cope with those fields changing asynchronously, and fully trust only the
6460 * pcp->count field on the local CPU with interrupts disabled.
6462 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6463 * outside of boot time (or some other assurance that no concurrent updaters
6466 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6467 unsigned long batch)
6469 WRITE_ONCE(pcp->batch, batch);
6470 WRITE_ONCE(pcp->high, high);
6473 static void pageset_init(struct per_cpu_pageset *p)
6475 struct per_cpu_pages *pcp;
6478 memset(p, 0, sizeof(*p));
6481 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6482 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6485 * Set batch and high values safe for a boot pageset. A true percpu
6486 * pageset's initialization will update them subsequently. Here we don't
6487 * need to be as careful as pageset_update() as nobody can access the
6490 pcp->high = BOOT_PAGESET_HIGH;
6491 pcp->batch = BOOT_PAGESET_BATCH;
6494 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6495 unsigned long batch)
6497 struct per_cpu_pageset *p;
6500 for_each_possible_cpu(cpu) {
6501 p = per_cpu_ptr(zone->pageset, cpu);
6502 pageset_update(&p->pcp, high, batch);
6507 * Calculate and set new high and batch values for all per-cpu pagesets of a
6508 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6510 static void zone_set_pageset_high_and_batch(struct zone *zone)
6512 unsigned long new_high, new_batch;
6514 if (percpu_pagelist_fraction) {
6515 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6516 new_batch = max(1UL, new_high / 4);
6517 if ((new_high / 4) > (PAGE_SHIFT * 8))
6518 new_batch = PAGE_SHIFT * 8;
6520 new_batch = zone_batchsize(zone);
6521 new_high = 6 * new_batch;
6522 new_batch = max(1UL, 1 * new_batch);
6525 if (zone->pageset_high == new_high &&
6526 zone->pageset_batch == new_batch)
6529 zone->pageset_high = new_high;
6530 zone->pageset_batch = new_batch;
6532 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6535 void __meminit setup_zone_pageset(struct zone *zone)
6537 struct per_cpu_pageset *p;
6540 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6541 for_each_possible_cpu(cpu) {
6542 p = per_cpu_ptr(zone->pageset, cpu);
6546 zone_set_pageset_high_and_batch(zone);
6550 * Allocate per cpu pagesets and initialize them.
6551 * Before this call only boot pagesets were available.
6553 void __init setup_per_cpu_pageset(void)
6555 struct pglist_data *pgdat;
6557 int __maybe_unused cpu;
6559 for_each_populated_zone(zone)
6560 setup_zone_pageset(zone);
6564 * Unpopulated zones continue using the boot pagesets.
6565 * The numa stats for these pagesets need to be reset.
6566 * Otherwise, they will end up skewing the stats of
6567 * the nodes these zones are associated with.
6569 for_each_possible_cpu(cpu) {
6570 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6571 memset(pcp->vm_numa_stat_diff, 0,
6572 sizeof(pcp->vm_numa_stat_diff));
6576 for_each_online_pgdat(pgdat)
6577 pgdat->per_cpu_nodestats =
6578 alloc_percpu(struct per_cpu_nodestat);
6581 static __meminit void zone_pcp_init(struct zone *zone)
6584 * per cpu subsystem is not up at this point. The following code
6585 * relies on the ability of the linker to provide the
6586 * offset of a (static) per cpu variable into the per cpu area.
6588 zone->pageset = &boot_pageset;
6589 zone->pageset_high = BOOT_PAGESET_HIGH;
6590 zone->pageset_batch = BOOT_PAGESET_BATCH;
6592 if (populated_zone(zone))
6593 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6594 zone->name, zone->present_pages,
6595 zone_batchsize(zone));
6598 void __meminit init_currently_empty_zone(struct zone *zone,
6599 unsigned long zone_start_pfn,
6602 struct pglist_data *pgdat = zone->zone_pgdat;
6603 int zone_idx = zone_idx(zone) + 1;
6605 if (zone_idx > pgdat->nr_zones)
6606 pgdat->nr_zones = zone_idx;
6608 zone->zone_start_pfn = zone_start_pfn;
6610 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6611 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6613 (unsigned long)zone_idx(zone),
6614 zone_start_pfn, (zone_start_pfn + size));
6616 zone_init_free_lists(zone);
6617 zone->initialized = 1;
6621 * get_pfn_range_for_nid - Return the start and end page frames for a node
6622 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6623 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6624 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6626 * It returns the start and end page frame of a node based on information
6627 * provided by memblock_set_node(). If called for a node
6628 * with no available memory, a warning is printed and the start and end
6631 void __init get_pfn_range_for_nid(unsigned int nid,
6632 unsigned long *start_pfn, unsigned long *end_pfn)
6634 unsigned long this_start_pfn, this_end_pfn;
6640 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6641 *start_pfn = min(*start_pfn, this_start_pfn);
6642 *end_pfn = max(*end_pfn, this_end_pfn);
6645 if (*start_pfn == -1UL)
6650 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6651 * assumption is made that zones within a node are ordered in monotonic
6652 * increasing memory addresses so that the "highest" populated zone is used
6654 static void __init find_usable_zone_for_movable(void)
6657 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6658 if (zone_index == ZONE_MOVABLE)
6661 if (arch_zone_highest_possible_pfn[zone_index] >
6662 arch_zone_lowest_possible_pfn[zone_index])
6666 VM_BUG_ON(zone_index == -1);
6667 movable_zone = zone_index;
6671 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6672 * because it is sized independent of architecture. Unlike the other zones,
6673 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6674 * in each node depending on the size of each node and how evenly kernelcore
6675 * is distributed. This helper function adjusts the zone ranges
6676 * provided by the architecture for a given node by using the end of the
6677 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6678 * zones within a node are in order of monotonic increases memory addresses
6680 static void __init adjust_zone_range_for_zone_movable(int nid,
6681 unsigned long zone_type,
6682 unsigned long node_start_pfn,
6683 unsigned long node_end_pfn,
6684 unsigned long *zone_start_pfn,
6685 unsigned long *zone_end_pfn)
6687 /* Only adjust if ZONE_MOVABLE is on this node */
6688 if (zone_movable_pfn[nid]) {
6689 /* Size ZONE_MOVABLE */
6690 if (zone_type == ZONE_MOVABLE) {
6691 *zone_start_pfn = zone_movable_pfn[nid];
6692 *zone_end_pfn = min(node_end_pfn,
6693 arch_zone_highest_possible_pfn[movable_zone]);
6695 /* Adjust for ZONE_MOVABLE starting within this range */
6696 } else if (!mirrored_kernelcore &&
6697 *zone_start_pfn < zone_movable_pfn[nid] &&
6698 *zone_end_pfn > zone_movable_pfn[nid]) {
6699 *zone_end_pfn = zone_movable_pfn[nid];
6701 /* Check if this whole range is within ZONE_MOVABLE */
6702 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6703 *zone_start_pfn = *zone_end_pfn;
6708 * Return the number of pages a zone spans in a node, including holes
6709 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6711 static unsigned long __init zone_spanned_pages_in_node(int nid,
6712 unsigned long zone_type,
6713 unsigned long node_start_pfn,
6714 unsigned long node_end_pfn,
6715 unsigned long *zone_start_pfn,
6716 unsigned long *zone_end_pfn)
6718 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6719 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6720 /* When hotadd a new node from cpu_up(), the node should be empty */
6721 if (!node_start_pfn && !node_end_pfn)
6724 /* Get the start and end of the zone */
6725 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6726 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6727 adjust_zone_range_for_zone_movable(nid, zone_type,
6728 node_start_pfn, node_end_pfn,
6729 zone_start_pfn, zone_end_pfn);
6731 /* Check that this node has pages within the zone's required range */
6732 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6735 /* Move the zone boundaries inside the node if necessary */
6736 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6737 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6739 /* Return the spanned pages */
6740 return *zone_end_pfn - *zone_start_pfn;
6744 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6745 * then all holes in the requested range will be accounted for.
6747 unsigned long __init __absent_pages_in_range(int nid,
6748 unsigned long range_start_pfn,
6749 unsigned long range_end_pfn)
6751 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6752 unsigned long start_pfn, end_pfn;
6755 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6756 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6757 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6758 nr_absent -= end_pfn - start_pfn;
6764 * absent_pages_in_range - Return number of page frames in holes within a range
6765 * @start_pfn: The start PFN to start searching for holes
6766 * @end_pfn: The end PFN to stop searching for holes
6768 * Return: the number of pages frames in memory holes within a range.
6770 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6771 unsigned long end_pfn)
6773 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6776 /* Return the number of page frames in holes in a zone on a node */
6777 static unsigned long __init zone_absent_pages_in_node(int nid,
6778 unsigned long zone_type,
6779 unsigned long node_start_pfn,
6780 unsigned long node_end_pfn)
6782 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6783 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6784 unsigned long zone_start_pfn, zone_end_pfn;
6785 unsigned long nr_absent;
6787 /* When hotadd a new node from cpu_up(), the node should be empty */
6788 if (!node_start_pfn && !node_end_pfn)
6791 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6792 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6794 adjust_zone_range_for_zone_movable(nid, zone_type,
6795 node_start_pfn, node_end_pfn,
6796 &zone_start_pfn, &zone_end_pfn);
6797 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6800 * ZONE_MOVABLE handling.
6801 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6804 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6805 unsigned long start_pfn, end_pfn;
6806 struct memblock_region *r;
6808 for_each_mem_region(r) {
6809 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6810 zone_start_pfn, zone_end_pfn);
6811 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6812 zone_start_pfn, zone_end_pfn);
6814 if (zone_type == ZONE_MOVABLE &&
6815 memblock_is_mirror(r))
6816 nr_absent += end_pfn - start_pfn;
6818 if (zone_type == ZONE_NORMAL &&
6819 !memblock_is_mirror(r))
6820 nr_absent += end_pfn - start_pfn;
6827 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6828 unsigned long node_start_pfn,
6829 unsigned long node_end_pfn)
6831 unsigned long realtotalpages = 0, totalpages = 0;
6834 for (i = 0; i < MAX_NR_ZONES; i++) {
6835 struct zone *zone = pgdat->node_zones + i;
6836 unsigned long zone_start_pfn, zone_end_pfn;
6837 unsigned long spanned, absent;
6838 unsigned long size, real_size;
6840 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6845 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6850 real_size = size - absent;
6853 zone->zone_start_pfn = zone_start_pfn;
6855 zone->zone_start_pfn = 0;
6856 zone->spanned_pages = size;
6857 zone->present_pages = real_size;
6860 realtotalpages += real_size;
6863 pgdat->node_spanned_pages = totalpages;
6864 pgdat->node_present_pages = realtotalpages;
6865 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6869 #ifndef CONFIG_SPARSEMEM
6871 * Calculate the size of the zone->blockflags rounded to an unsigned long
6872 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6873 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6874 * round what is now in bits to nearest long in bits, then return it in
6877 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6879 unsigned long usemapsize;
6881 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6882 usemapsize = roundup(zonesize, pageblock_nr_pages);
6883 usemapsize = usemapsize >> pageblock_order;
6884 usemapsize *= NR_PAGEBLOCK_BITS;
6885 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6887 return usemapsize / 8;
6890 static void __ref setup_usemap(struct zone *zone)
6892 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
6893 zone->spanned_pages);
6894 zone->pageblock_flags = NULL;
6896 zone->pageblock_flags =
6897 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6899 if (!zone->pageblock_flags)
6900 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6901 usemapsize, zone->name, zone_to_nid(zone));
6905 static inline void setup_usemap(struct zone *zone) {}
6906 #endif /* CONFIG_SPARSEMEM */
6908 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6910 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6911 void __init set_pageblock_order(void)
6915 /* Check that pageblock_nr_pages has not already been setup */
6916 if (pageblock_order)
6919 if (HPAGE_SHIFT > PAGE_SHIFT)
6920 order = HUGETLB_PAGE_ORDER;
6922 order = MAX_ORDER - 1;
6925 * Assume the largest contiguous order of interest is a huge page.
6926 * This value may be variable depending on boot parameters on IA64 and
6929 pageblock_order = order;
6931 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6934 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6935 * is unused as pageblock_order is set at compile-time. See
6936 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6939 void __init set_pageblock_order(void)
6943 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6945 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6946 unsigned long present_pages)
6948 unsigned long pages = spanned_pages;
6951 * Provide a more accurate estimation if there are holes within
6952 * the zone and SPARSEMEM is in use. If there are holes within the
6953 * zone, each populated memory region may cost us one or two extra
6954 * memmap pages due to alignment because memmap pages for each
6955 * populated regions may not be naturally aligned on page boundary.
6956 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6958 if (spanned_pages > present_pages + (present_pages >> 4) &&
6959 IS_ENABLED(CONFIG_SPARSEMEM))
6960 pages = present_pages;
6962 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6965 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6966 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6968 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6970 spin_lock_init(&ds_queue->split_queue_lock);
6971 INIT_LIST_HEAD(&ds_queue->split_queue);
6972 ds_queue->split_queue_len = 0;
6975 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6978 #ifdef CONFIG_COMPACTION
6979 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6981 init_waitqueue_head(&pgdat->kcompactd_wait);
6984 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6987 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6989 pgdat_resize_init(pgdat);
6991 pgdat_init_split_queue(pgdat);
6992 pgdat_init_kcompactd(pgdat);
6994 init_waitqueue_head(&pgdat->kswapd_wait);
6995 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6997 pgdat_page_ext_init(pgdat);
6998 lruvec_init(&pgdat->__lruvec);
7001 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7002 unsigned long remaining_pages)
7004 atomic_long_set(&zone->managed_pages, remaining_pages);
7005 zone_set_nid(zone, nid);
7006 zone->name = zone_names[idx];
7007 zone->zone_pgdat = NODE_DATA(nid);
7008 spin_lock_init(&zone->lock);
7009 zone_seqlock_init(zone);
7010 zone_pcp_init(zone);
7014 * Set up the zone data structures
7015 * - init pgdat internals
7016 * - init all zones belonging to this node
7018 * NOTE: this function is only called during memory hotplug
7020 #ifdef CONFIG_MEMORY_HOTPLUG
7021 void __ref free_area_init_core_hotplug(int nid)
7024 pg_data_t *pgdat = NODE_DATA(nid);
7026 pgdat_init_internals(pgdat);
7027 for (z = 0; z < MAX_NR_ZONES; z++)
7028 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7033 * Set up the zone data structures:
7034 * - mark all pages reserved
7035 * - mark all memory queues empty
7036 * - clear the memory bitmaps
7038 * NOTE: pgdat should get zeroed by caller.
7039 * NOTE: this function is only called during early init.
7041 static void __init free_area_init_core(struct pglist_data *pgdat)
7044 int nid = pgdat->node_id;
7046 pgdat_init_internals(pgdat);
7047 pgdat->per_cpu_nodestats = &boot_nodestats;
7049 for (j = 0; j < MAX_NR_ZONES; j++) {
7050 struct zone *zone = pgdat->node_zones + j;
7051 unsigned long size, freesize, memmap_pages;
7053 size = zone->spanned_pages;
7054 freesize = zone->present_pages;
7057 * Adjust freesize so that it accounts for how much memory
7058 * is used by this zone for memmap. This affects the watermark
7059 * and per-cpu initialisations
7061 memmap_pages = calc_memmap_size(size, freesize);
7062 if (!is_highmem_idx(j)) {
7063 if (freesize >= memmap_pages) {
7064 freesize -= memmap_pages;
7067 " %s zone: %lu pages used for memmap\n",
7068 zone_names[j], memmap_pages);
7070 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
7071 zone_names[j], memmap_pages, freesize);
7074 /* Account for reserved pages */
7075 if (j == 0 && freesize > dma_reserve) {
7076 freesize -= dma_reserve;
7077 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
7078 zone_names[0], dma_reserve);
7081 if (!is_highmem_idx(j))
7082 nr_kernel_pages += freesize;
7083 /* Charge for highmem memmap if there are enough kernel pages */
7084 else if (nr_kernel_pages > memmap_pages * 2)
7085 nr_kernel_pages -= memmap_pages;
7086 nr_all_pages += freesize;
7089 * Set an approximate value for lowmem here, it will be adjusted
7090 * when the bootmem allocator frees pages into the buddy system.
7091 * And all highmem pages will be managed by the buddy system.
7093 zone_init_internals(zone, j, nid, freesize);
7098 set_pageblock_order();
7100 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7101 memmap_init_zone(zone);
7105 #ifdef CONFIG_FLAT_NODE_MEM_MAP
7106 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7108 unsigned long __maybe_unused start = 0;
7109 unsigned long __maybe_unused offset = 0;
7111 /* Skip empty nodes */
7112 if (!pgdat->node_spanned_pages)
7115 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7116 offset = pgdat->node_start_pfn - start;
7117 /* ia64 gets its own node_mem_map, before this, without bootmem */
7118 if (!pgdat->node_mem_map) {
7119 unsigned long size, end;
7123 * The zone's endpoints aren't required to be MAX_ORDER
7124 * aligned but the node_mem_map endpoints must be in order
7125 * for the buddy allocator to function correctly.
7127 end = pgdat_end_pfn(pgdat);
7128 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7129 size = (end - start) * sizeof(struct page);
7130 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7133 panic("Failed to allocate %ld bytes for node %d memory map\n",
7134 size, pgdat->node_id);
7135 pgdat->node_mem_map = map + offset;
7137 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7138 __func__, pgdat->node_id, (unsigned long)pgdat,
7139 (unsigned long)pgdat->node_mem_map);
7140 #ifndef CONFIG_NEED_MULTIPLE_NODES
7142 * With no DISCONTIG, the global mem_map is just set as node 0's
7144 if (pgdat == NODE_DATA(0)) {
7145 mem_map = NODE_DATA(0)->node_mem_map;
7146 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7152 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7153 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7155 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7156 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7158 pgdat->first_deferred_pfn = ULONG_MAX;
7161 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7164 static void __init free_area_init_node(int nid)
7166 pg_data_t *pgdat = NODE_DATA(nid);
7167 unsigned long start_pfn = 0;
7168 unsigned long end_pfn = 0;
7170 /* pg_data_t should be reset to zero when it's allocated */
7171 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7173 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7175 pgdat->node_id = nid;
7176 pgdat->node_start_pfn = start_pfn;
7177 pgdat->per_cpu_nodestats = NULL;
7179 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7180 (u64)start_pfn << PAGE_SHIFT,
7181 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7182 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7184 alloc_node_mem_map(pgdat);
7185 pgdat_set_deferred_range(pgdat);
7187 free_area_init_core(pgdat);
7190 void __init free_area_init_memoryless_node(int nid)
7192 free_area_init_node(nid);
7195 #if MAX_NUMNODES > 1
7197 * Figure out the number of possible node ids.
7199 void __init setup_nr_node_ids(void)
7201 unsigned int highest;
7203 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7204 nr_node_ids = highest + 1;
7209 * node_map_pfn_alignment - determine the maximum internode alignment
7211 * This function should be called after node map is populated and sorted.
7212 * It calculates the maximum power of two alignment which can distinguish
7215 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7216 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7217 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7218 * shifted, 1GiB is enough and this function will indicate so.
7220 * This is used to test whether pfn -> nid mapping of the chosen memory
7221 * model has fine enough granularity to avoid incorrect mapping for the
7222 * populated node map.
7224 * Return: the determined alignment in pfn's. 0 if there is no alignment
7225 * requirement (single node).
7227 unsigned long __init node_map_pfn_alignment(void)
7229 unsigned long accl_mask = 0, last_end = 0;
7230 unsigned long start, end, mask;
7231 int last_nid = NUMA_NO_NODE;
7234 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7235 if (!start || last_nid < 0 || last_nid == nid) {
7242 * Start with a mask granular enough to pin-point to the
7243 * start pfn and tick off bits one-by-one until it becomes
7244 * too coarse to separate the current node from the last.
7246 mask = ~((1 << __ffs(start)) - 1);
7247 while (mask && last_end <= (start & (mask << 1)))
7250 /* accumulate all internode masks */
7254 /* convert mask to number of pages */
7255 return ~accl_mask + 1;
7259 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7261 * Return: the minimum PFN based on information provided via
7262 * memblock_set_node().
7264 unsigned long __init find_min_pfn_with_active_regions(void)
7266 return PHYS_PFN(memblock_start_of_DRAM());
7270 * early_calculate_totalpages()
7271 * Sum pages in active regions for movable zone.
7272 * Populate N_MEMORY for calculating usable_nodes.
7274 static unsigned long __init early_calculate_totalpages(void)
7276 unsigned long totalpages = 0;
7277 unsigned long start_pfn, end_pfn;
7280 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7281 unsigned long pages = end_pfn - start_pfn;
7283 totalpages += pages;
7285 node_set_state(nid, N_MEMORY);
7291 * Find the PFN the Movable zone begins in each node. Kernel memory
7292 * is spread evenly between nodes as long as the nodes have enough
7293 * memory. When they don't, some nodes will have more kernelcore than
7296 static void __init find_zone_movable_pfns_for_nodes(void)
7299 unsigned long usable_startpfn;
7300 unsigned long kernelcore_node, kernelcore_remaining;
7301 /* save the state before borrow the nodemask */
7302 nodemask_t saved_node_state = node_states[N_MEMORY];
7303 unsigned long totalpages = early_calculate_totalpages();
7304 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7305 struct memblock_region *r;
7307 /* Need to find movable_zone earlier when movable_node is specified. */
7308 find_usable_zone_for_movable();
7311 * If movable_node is specified, ignore kernelcore and movablecore
7314 if (movable_node_is_enabled()) {
7315 for_each_mem_region(r) {
7316 if (!memblock_is_hotpluggable(r))
7319 nid = memblock_get_region_node(r);
7321 usable_startpfn = PFN_DOWN(r->base);
7322 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7323 min(usable_startpfn, zone_movable_pfn[nid]) :
7331 * If kernelcore=mirror is specified, ignore movablecore option
7333 if (mirrored_kernelcore) {
7334 bool mem_below_4gb_not_mirrored = false;
7336 for_each_mem_region(r) {
7337 if (memblock_is_mirror(r))
7340 nid = memblock_get_region_node(r);
7342 usable_startpfn = memblock_region_memory_base_pfn(r);
7344 if (usable_startpfn < 0x100000) {
7345 mem_below_4gb_not_mirrored = true;
7349 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7350 min(usable_startpfn, zone_movable_pfn[nid]) :
7354 if (mem_below_4gb_not_mirrored)
7355 pr_warn("This configuration results in unmirrored kernel memory.\n");
7361 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7362 * amount of necessary memory.
7364 if (required_kernelcore_percent)
7365 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7367 if (required_movablecore_percent)
7368 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7372 * If movablecore= was specified, calculate what size of
7373 * kernelcore that corresponds so that memory usable for
7374 * any allocation type is evenly spread. If both kernelcore
7375 * and movablecore are specified, then the value of kernelcore
7376 * will be used for required_kernelcore if it's greater than
7377 * what movablecore would have allowed.
7379 if (required_movablecore) {
7380 unsigned long corepages;
7383 * Round-up so that ZONE_MOVABLE is at least as large as what
7384 * was requested by the user
7386 required_movablecore =
7387 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7388 required_movablecore = min(totalpages, required_movablecore);
7389 corepages = totalpages - required_movablecore;
7391 required_kernelcore = max(required_kernelcore, corepages);
7395 * If kernelcore was not specified or kernelcore size is larger
7396 * than totalpages, there is no ZONE_MOVABLE.
7398 if (!required_kernelcore || required_kernelcore >= totalpages)
7401 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7402 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7405 /* Spread kernelcore memory as evenly as possible throughout nodes */
7406 kernelcore_node = required_kernelcore / usable_nodes;
7407 for_each_node_state(nid, N_MEMORY) {
7408 unsigned long start_pfn, end_pfn;
7411 * Recalculate kernelcore_node if the division per node
7412 * now exceeds what is necessary to satisfy the requested
7413 * amount of memory for the kernel
7415 if (required_kernelcore < kernelcore_node)
7416 kernelcore_node = required_kernelcore / usable_nodes;
7419 * As the map is walked, we track how much memory is usable
7420 * by the kernel using kernelcore_remaining. When it is
7421 * 0, the rest of the node is usable by ZONE_MOVABLE
7423 kernelcore_remaining = kernelcore_node;
7425 /* Go through each range of PFNs within this node */
7426 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7427 unsigned long size_pages;
7429 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7430 if (start_pfn >= end_pfn)
7433 /* Account for what is only usable for kernelcore */
7434 if (start_pfn < usable_startpfn) {
7435 unsigned long kernel_pages;
7436 kernel_pages = min(end_pfn, usable_startpfn)
7439 kernelcore_remaining -= min(kernel_pages,
7440 kernelcore_remaining);
7441 required_kernelcore -= min(kernel_pages,
7442 required_kernelcore);
7444 /* Continue if range is now fully accounted */
7445 if (end_pfn <= usable_startpfn) {
7448 * Push zone_movable_pfn to the end so
7449 * that if we have to rebalance
7450 * kernelcore across nodes, we will
7451 * not double account here
7453 zone_movable_pfn[nid] = end_pfn;
7456 start_pfn = usable_startpfn;
7460 * The usable PFN range for ZONE_MOVABLE is from
7461 * start_pfn->end_pfn. Calculate size_pages as the
7462 * number of pages used as kernelcore
7464 size_pages = end_pfn - start_pfn;
7465 if (size_pages > kernelcore_remaining)
7466 size_pages = kernelcore_remaining;
7467 zone_movable_pfn[nid] = start_pfn + size_pages;
7470 * Some kernelcore has been met, update counts and
7471 * break if the kernelcore for this node has been
7474 required_kernelcore -= min(required_kernelcore,
7476 kernelcore_remaining -= size_pages;
7477 if (!kernelcore_remaining)
7483 * If there is still required_kernelcore, we do another pass with one
7484 * less node in the count. This will push zone_movable_pfn[nid] further
7485 * along on the nodes that still have memory until kernelcore is
7489 if (usable_nodes && required_kernelcore > usable_nodes)
7493 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7494 for (nid = 0; nid < MAX_NUMNODES; nid++)
7495 zone_movable_pfn[nid] =
7496 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7499 /* restore the node_state */
7500 node_states[N_MEMORY] = saved_node_state;
7503 /* Any regular or high memory on that node ? */
7504 static void check_for_memory(pg_data_t *pgdat, int nid)
7506 enum zone_type zone_type;
7508 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7509 struct zone *zone = &pgdat->node_zones[zone_type];
7510 if (populated_zone(zone)) {
7511 if (IS_ENABLED(CONFIG_HIGHMEM))
7512 node_set_state(nid, N_HIGH_MEMORY);
7513 if (zone_type <= ZONE_NORMAL)
7514 node_set_state(nid, N_NORMAL_MEMORY);
7521 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7522 * such cases we allow max_zone_pfn sorted in the descending order
7524 bool __weak arch_has_descending_max_zone_pfns(void)
7530 * free_area_init - Initialise all pg_data_t and zone data
7531 * @max_zone_pfn: an array of max PFNs for each zone
7533 * This will call free_area_init_node() for each active node in the system.
7534 * Using the page ranges provided by memblock_set_node(), the size of each
7535 * zone in each node and their holes is calculated. If the maximum PFN
7536 * between two adjacent zones match, it is assumed that the zone is empty.
7537 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7538 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7539 * starts where the previous one ended. For example, ZONE_DMA32 starts
7540 * at arch_max_dma_pfn.
7542 void __init free_area_init(unsigned long *max_zone_pfn)
7544 unsigned long start_pfn, end_pfn;
7548 /* Record where the zone boundaries are */
7549 memset(arch_zone_lowest_possible_pfn, 0,
7550 sizeof(arch_zone_lowest_possible_pfn));
7551 memset(arch_zone_highest_possible_pfn, 0,
7552 sizeof(arch_zone_highest_possible_pfn));
7554 start_pfn = find_min_pfn_with_active_regions();
7555 descending = arch_has_descending_max_zone_pfns();
7557 for (i = 0; i < MAX_NR_ZONES; i++) {
7559 zone = MAX_NR_ZONES - i - 1;
7563 if (zone == ZONE_MOVABLE)
7566 end_pfn = max(max_zone_pfn[zone], start_pfn);
7567 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7568 arch_zone_highest_possible_pfn[zone] = end_pfn;
7570 start_pfn = end_pfn;
7573 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7574 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7575 find_zone_movable_pfns_for_nodes();
7577 /* Print out the zone ranges */
7578 pr_info("Zone ranges:\n");
7579 for (i = 0; i < MAX_NR_ZONES; i++) {
7580 if (i == ZONE_MOVABLE)
7582 pr_info(" %-8s ", zone_names[i]);
7583 if (arch_zone_lowest_possible_pfn[i] ==
7584 arch_zone_highest_possible_pfn[i])
7587 pr_cont("[mem %#018Lx-%#018Lx]\n",
7588 (u64)arch_zone_lowest_possible_pfn[i]
7590 ((u64)arch_zone_highest_possible_pfn[i]
7591 << PAGE_SHIFT) - 1);
7594 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7595 pr_info("Movable zone start for each node\n");
7596 for (i = 0; i < MAX_NUMNODES; i++) {
7597 if (zone_movable_pfn[i])
7598 pr_info(" Node %d: %#018Lx\n", i,
7599 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7603 * Print out the early node map, and initialize the
7604 * subsection-map relative to active online memory ranges to
7605 * enable future "sub-section" extensions of the memory map.
7607 pr_info("Early memory node ranges\n");
7608 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7609 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7610 (u64)start_pfn << PAGE_SHIFT,
7611 ((u64)end_pfn << PAGE_SHIFT) - 1);
7612 subsection_map_init(start_pfn, end_pfn - start_pfn);
7615 /* Initialise every node */
7616 mminit_verify_pageflags_layout();
7617 setup_nr_node_ids();
7618 for_each_online_node(nid) {
7619 pg_data_t *pgdat = NODE_DATA(nid);
7620 free_area_init_node(nid);
7622 /* Any memory on that node */
7623 if (pgdat->node_present_pages)
7624 node_set_state(nid, N_MEMORY);
7625 check_for_memory(pgdat, nid);
7629 static int __init cmdline_parse_core(char *p, unsigned long *core,
7630 unsigned long *percent)
7632 unsigned long long coremem;
7638 /* Value may be a percentage of total memory, otherwise bytes */
7639 coremem = simple_strtoull(p, &endptr, 0);
7640 if (*endptr == '%') {
7641 /* Paranoid check for percent values greater than 100 */
7642 WARN_ON(coremem > 100);
7646 coremem = memparse(p, &p);
7647 /* Paranoid check that UL is enough for the coremem value */
7648 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7650 *core = coremem >> PAGE_SHIFT;
7657 * kernelcore=size sets the amount of memory for use for allocations that
7658 * cannot be reclaimed or migrated.
7660 static int __init cmdline_parse_kernelcore(char *p)
7662 /* parse kernelcore=mirror */
7663 if (parse_option_str(p, "mirror")) {
7664 mirrored_kernelcore = true;
7668 return cmdline_parse_core(p, &required_kernelcore,
7669 &required_kernelcore_percent);
7673 * movablecore=size sets the amount of memory for use for allocations that
7674 * can be reclaimed or migrated.
7676 static int __init cmdline_parse_movablecore(char *p)
7678 return cmdline_parse_core(p, &required_movablecore,
7679 &required_movablecore_percent);
7682 early_param("kernelcore", cmdline_parse_kernelcore);
7683 early_param("movablecore", cmdline_parse_movablecore);
7685 void adjust_managed_page_count(struct page *page, long count)
7687 atomic_long_add(count, &page_zone(page)->managed_pages);
7688 totalram_pages_add(count);
7689 #ifdef CONFIG_HIGHMEM
7690 if (PageHighMem(page))
7691 totalhigh_pages_add(count);
7694 EXPORT_SYMBOL(adjust_managed_page_count);
7696 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7699 unsigned long pages = 0;
7701 start = (void *)PAGE_ALIGN((unsigned long)start);
7702 end = (void *)((unsigned long)end & PAGE_MASK);
7703 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7704 struct page *page = virt_to_page(pos);
7705 void *direct_map_addr;
7708 * 'direct_map_addr' might be different from 'pos'
7709 * because some architectures' virt_to_page()
7710 * work with aliases. Getting the direct map
7711 * address ensures that we get a _writeable_
7712 * alias for the memset().
7714 direct_map_addr = page_address(page);
7716 * Perform a kasan-unchecked memset() since this memory
7717 * has not been initialized.
7719 direct_map_addr = kasan_reset_tag(direct_map_addr);
7720 if ((unsigned int)poison <= 0xFF)
7721 memset(direct_map_addr, poison, PAGE_SIZE);
7723 free_reserved_page(page);
7727 pr_info("Freeing %s memory: %ldK\n",
7728 s, pages << (PAGE_SHIFT - 10));
7733 void __init mem_init_print_info(void)
7735 unsigned long physpages, codesize, datasize, rosize, bss_size;
7736 unsigned long init_code_size, init_data_size;
7738 physpages = get_num_physpages();
7739 codesize = _etext - _stext;
7740 datasize = _edata - _sdata;
7741 rosize = __end_rodata - __start_rodata;
7742 bss_size = __bss_stop - __bss_start;
7743 init_data_size = __init_end - __init_begin;
7744 init_code_size = _einittext - _sinittext;
7747 * Detect special cases and adjust section sizes accordingly:
7748 * 1) .init.* may be embedded into .data sections
7749 * 2) .init.text.* may be out of [__init_begin, __init_end],
7750 * please refer to arch/tile/kernel/vmlinux.lds.S.
7751 * 3) .rodata.* may be embedded into .text or .data sections.
7753 #define adj_init_size(start, end, size, pos, adj) \
7755 if (start <= pos && pos < end && size > adj) \
7759 adj_init_size(__init_begin, __init_end, init_data_size,
7760 _sinittext, init_code_size);
7761 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7762 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7763 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7764 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7766 #undef adj_init_size
7768 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7769 #ifdef CONFIG_HIGHMEM
7773 nr_free_pages() << (PAGE_SHIFT - 10),
7774 physpages << (PAGE_SHIFT - 10),
7775 codesize >> 10, datasize >> 10, rosize >> 10,
7776 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7777 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7778 totalcma_pages << (PAGE_SHIFT - 10)
7779 #ifdef CONFIG_HIGHMEM
7780 , totalhigh_pages() << (PAGE_SHIFT - 10)
7786 * set_dma_reserve - set the specified number of pages reserved in the first zone
7787 * @new_dma_reserve: The number of pages to mark reserved
7789 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7790 * In the DMA zone, a significant percentage may be consumed by kernel image
7791 * and other unfreeable allocations which can skew the watermarks badly. This
7792 * function may optionally be used to account for unfreeable pages in the
7793 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7794 * smaller per-cpu batchsize.
7796 void __init set_dma_reserve(unsigned long new_dma_reserve)
7798 dma_reserve = new_dma_reserve;
7801 static int page_alloc_cpu_dead(unsigned int cpu)
7804 lru_add_drain_cpu(cpu);
7808 * Spill the event counters of the dead processor
7809 * into the current processors event counters.
7810 * This artificially elevates the count of the current
7813 vm_events_fold_cpu(cpu);
7816 * Zero the differential counters of the dead processor
7817 * so that the vm statistics are consistent.
7819 * This is only okay since the processor is dead and cannot
7820 * race with what we are doing.
7822 cpu_vm_stats_fold(cpu);
7827 int hashdist = HASHDIST_DEFAULT;
7829 static int __init set_hashdist(char *str)
7833 hashdist = simple_strtoul(str, &str, 0);
7836 __setup("hashdist=", set_hashdist);
7839 void __init page_alloc_init(void)
7844 if (num_node_state(N_MEMORY) == 1)
7848 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7849 "mm/page_alloc:dead", NULL,
7850 page_alloc_cpu_dead);
7855 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7856 * or min_free_kbytes changes.
7858 static void calculate_totalreserve_pages(void)
7860 struct pglist_data *pgdat;
7861 unsigned long reserve_pages = 0;
7862 enum zone_type i, j;
7864 for_each_online_pgdat(pgdat) {
7866 pgdat->totalreserve_pages = 0;
7868 for (i = 0; i < MAX_NR_ZONES; i++) {
7869 struct zone *zone = pgdat->node_zones + i;
7871 unsigned long managed_pages = zone_managed_pages(zone);
7873 /* Find valid and maximum lowmem_reserve in the zone */
7874 for (j = i; j < MAX_NR_ZONES; j++) {
7875 if (zone->lowmem_reserve[j] > max)
7876 max = zone->lowmem_reserve[j];
7879 /* we treat the high watermark as reserved pages. */
7880 max += high_wmark_pages(zone);
7882 if (max > managed_pages)
7883 max = managed_pages;
7885 pgdat->totalreserve_pages += max;
7887 reserve_pages += max;
7890 totalreserve_pages = reserve_pages;
7894 * setup_per_zone_lowmem_reserve - called whenever
7895 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7896 * has a correct pages reserved value, so an adequate number of
7897 * pages are left in the zone after a successful __alloc_pages().
7899 static void setup_per_zone_lowmem_reserve(void)
7901 struct pglist_data *pgdat;
7902 enum zone_type i, j;
7904 for_each_online_pgdat(pgdat) {
7905 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
7906 struct zone *zone = &pgdat->node_zones[i];
7907 int ratio = sysctl_lowmem_reserve_ratio[i];
7908 bool clear = !ratio || !zone_managed_pages(zone);
7909 unsigned long managed_pages = 0;
7911 for (j = i + 1; j < MAX_NR_ZONES; j++) {
7913 zone->lowmem_reserve[j] = 0;
7915 struct zone *upper_zone = &pgdat->node_zones[j];
7917 managed_pages += zone_managed_pages(upper_zone);
7918 zone->lowmem_reserve[j] = managed_pages / ratio;
7924 /* update totalreserve_pages */
7925 calculate_totalreserve_pages();
7928 static void __setup_per_zone_wmarks(void)
7930 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7931 unsigned long lowmem_pages = 0;
7933 unsigned long flags;
7935 /* Calculate total number of !ZONE_HIGHMEM pages */
7936 for_each_zone(zone) {
7937 if (!is_highmem(zone))
7938 lowmem_pages += zone_managed_pages(zone);
7941 for_each_zone(zone) {
7944 spin_lock_irqsave(&zone->lock, flags);
7945 tmp = (u64)pages_min * zone_managed_pages(zone);
7946 do_div(tmp, lowmem_pages);
7947 if (is_highmem(zone)) {
7949 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7950 * need highmem pages, so cap pages_min to a small
7953 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7954 * deltas control async page reclaim, and so should
7955 * not be capped for highmem.
7957 unsigned long min_pages;
7959 min_pages = zone_managed_pages(zone) / 1024;
7960 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7961 zone->_watermark[WMARK_MIN] = min_pages;
7964 * If it's a lowmem zone, reserve a number of pages
7965 * proportionate to the zone's size.
7967 zone->_watermark[WMARK_MIN] = tmp;
7971 * Set the kswapd watermarks distance according to the
7972 * scale factor in proportion to available memory, but
7973 * ensure a minimum size on small systems.
7975 tmp = max_t(u64, tmp >> 2,
7976 mult_frac(zone_managed_pages(zone),
7977 watermark_scale_factor, 10000));
7979 zone->watermark_boost = 0;
7980 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7981 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7983 spin_unlock_irqrestore(&zone->lock, flags);
7986 /* update totalreserve_pages */
7987 calculate_totalreserve_pages();
7991 * setup_per_zone_wmarks - called when min_free_kbytes changes
7992 * or when memory is hot-{added|removed}
7994 * Ensures that the watermark[min,low,high] values for each zone are set
7995 * correctly with respect to min_free_kbytes.
7997 void setup_per_zone_wmarks(void)
7999 static DEFINE_SPINLOCK(lock);
8002 __setup_per_zone_wmarks();
8007 * Initialise min_free_kbytes.
8009 * For small machines we want it small (128k min). For large machines
8010 * we want it large (256MB max). But it is not linear, because network
8011 * bandwidth does not increase linearly with machine size. We use
8013 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8014 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8030 int __meminit init_per_zone_wmark_min(void)
8032 unsigned long lowmem_kbytes;
8033 int new_min_free_kbytes;
8035 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8036 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8038 if (new_min_free_kbytes > user_min_free_kbytes) {
8039 min_free_kbytes = new_min_free_kbytes;
8040 if (min_free_kbytes < 128)
8041 min_free_kbytes = 128;
8042 if (min_free_kbytes > 262144)
8043 min_free_kbytes = 262144;
8045 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8046 new_min_free_kbytes, user_min_free_kbytes);
8048 setup_per_zone_wmarks();
8049 refresh_zone_stat_thresholds();
8050 setup_per_zone_lowmem_reserve();
8053 setup_min_unmapped_ratio();
8054 setup_min_slab_ratio();
8057 khugepaged_min_free_kbytes_update();
8061 postcore_initcall(init_per_zone_wmark_min)
8064 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8065 * that we can call two helper functions whenever min_free_kbytes
8068 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8069 void *buffer, size_t *length, loff_t *ppos)
8073 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8078 user_min_free_kbytes = min_free_kbytes;
8079 setup_per_zone_wmarks();
8084 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8085 void *buffer, size_t *length, loff_t *ppos)
8089 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8094 setup_per_zone_wmarks();
8100 static void setup_min_unmapped_ratio(void)
8105 for_each_online_pgdat(pgdat)
8106 pgdat->min_unmapped_pages = 0;
8109 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8110 sysctl_min_unmapped_ratio) / 100;
8114 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8115 void *buffer, size_t *length, loff_t *ppos)
8119 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8123 setup_min_unmapped_ratio();
8128 static void setup_min_slab_ratio(void)
8133 for_each_online_pgdat(pgdat)
8134 pgdat->min_slab_pages = 0;
8137 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8138 sysctl_min_slab_ratio) / 100;
8141 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8142 void *buffer, size_t *length, loff_t *ppos)
8146 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8150 setup_min_slab_ratio();
8157 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8158 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8159 * whenever sysctl_lowmem_reserve_ratio changes.
8161 * The reserve ratio obviously has absolutely no relation with the
8162 * minimum watermarks. The lowmem reserve ratio can only make sense
8163 * if in function of the boot time zone sizes.
8165 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8166 void *buffer, size_t *length, loff_t *ppos)
8170 proc_dointvec_minmax(table, write, buffer, length, ppos);
8172 for (i = 0; i < MAX_NR_ZONES; i++) {
8173 if (sysctl_lowmem_reserve_ratio[i] < 1)
8174 sysctl_lowmem_reserve_ratio[i] = 0;
8177 setup_per_zone_lowmem_reserve();
8182 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8183 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8184 * pagelist can have before it gets flushed back to buddy allocator.
8186 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8187 void *buffer, size_t *length, loff_t *ppos)
8190 int old_percpu_pagelist_fraction;
8193 mutex_lock(&pcp_batch_high_lock);
8194 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8196 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8197 if (!write || ret < 0)
8200 /* Sanity checking to avoid pcp imbalance */
8201 if (percpu_pagelist_fraction &&
8202 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8203 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8209 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8212 for_each_populated_zone(zone)
8213 zone_set_pageset_high_and_batch(zone);
8215 mutex_unlock(&pcp_batch_high_lock);
8219 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8221 * Returns the number of pages that arch has reserved but
8222 * is not known to alloc_large_system_hash().
8224 static unsigned long __init arch_reserved_kernel_pages(void)
8231 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8232 * machines. As memory size is increased the scale is also increased but at
8233 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8234 * quadruples the scale is increased by one, which means the size of hash table
8235 * only doubles, instead of quadrupling as well.
8236 * Because 32-bit systems cannot have large physical memory, where this scaling
8237 * makes sense, it is disabled on such platforms.
8239 #if __BITS_PER_LONG > 32
8240 #define ADAPT_SCALE_BASE (64ul << 30)
8241 #define ADAPT_SCALE_SHIFT 2
8242 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8246 * allocate a large system hash table from bootmem
8247 * - it is assumed that the hash table must contain an exact power-of-2
8248 * quantity of entries
8249 * - limit is the number of hash buckets, not the total allocation size
8251 void *__init alloc_large_system_hash(const char *tablename,
8252 unsigned long bucketsize,
8253 unsigned long numentries,
8256 unsigned int *_hash_shift,
8257 unsigned int *_hash_mask,
8258 unsigned long low_limit,
8259 unsigned long high_limit)
8261 unsigned long long max = high_limit;
8262 unsigned long log2qty, size;
8268 /* allow the kernel cmdline to have a say */
8270 /* round applicable memory size up to nearest megabyte */
8271 numentries = nr_kernel_pages;
8272 numentries -= arch_reserved_kernel_pages();
8274 /* It isn't necessary when PAGE_SIZE >= 1MB */
8275 if (PAGE_SHIFT < 20)
8276 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8278 #if __BITS_PER_LONG > 32
8280 unsigned long adapt;
8282 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8283 adapt <<= ADAPT_SCALE_SHIFT)
8288 /* limit to 1 bucket per 2^scale bytes of low memory */
8289 if (scale > PAGE_SHIFT)
8290 numentries >>= (scale - PAGE_SHIFT);
8292 numentries <<= (PAGE_SHIFT - scale);
8294 /* Make sure we've got at least a 0-order allocation.. */
8295 if (unlikely(flags & HASH_SMALL)) {
8296 /* Makes no sense without HASH_EARLY */
8297 WARN_ON(!(flags & HASH_EARLY));
8298 if (!(numentries >> *_hash_shift)) {
8299 numentries = 1UL << *_hash_shift;
8300 BUG_ON(!numentries);
8302 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8303 numentries = PAGE_SIZE / bucketsize;
8305 numentries = roundup_pow_of_two(numentries);
8307 /* limit allocation size to 1/16 total memory by default */
8309 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8310 do_div(max, bucketsize);
8312 max = min(max, 0x80000000ULL);
8314 if (numentries < low_limit)
8315 numentries = low_limit;
8316 if (numentries > max)
8319 log2qty = ilog2(numentries);
8321 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8324 size = bucketsize << log2qty;
8325 if (flags & HASH_EARLY) {
8326 if (flags & HASH_ZERO)
8327 table = memblock_alloc(size, SMP_CACHE_BYTES);
8329 table = memblock_alloc_raw(size,
8331 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8332 table = __vmalloc(size, gfp_flags);
8334 huge = is_vm_area_hugepages(table);
8337 * If bucketsize is not a power-of-two, we may free
8338 * some pages at the end of hash table which
8339 * alloc_pages_exact() automatically does
8341 table = alloc_pages_exact(size, gfp_flags);
8342 kmemleak_alloc(table, size, 1, gfp_flags);
8344 } while (!table && size > PAGE_SIZE && --log2qty);
8347 panic("Failed to allocate %s hash table\n", tablename);
8349 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8350 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8351 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8354 *_hash_shift = log2qty;
8356 *_hash_mask = (1 << log2qty) - 1;
8362 * This function checks whether pageblock includes unmovable pages or not.
8364 * PageLRU check without isolation or lru_lock could race so that
8365 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8366 * check without lock_page also may miss some movable non-lru pages at
8367 * race condition. So you can't expect this function should be exact.
8369 * Returns a page without holding a reference. If the caller wants to
8370 * dereference that page (e.g., dumping), it has to make sure that it
8371 * cannot get removed (e.g., via memory unplug) concurrently.
8374 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8375 int migratetype, int flags)
8377 unsigned long iter = 0;
8378 unsigned long pfn = page_to_pfn(page);
8379 unsigned long offset = pfn % pageblock_nr_pages;
8381 if (is_migrate_cma_page(page)) {
8383 * CMA allocations (alloc_contig_range) really need to mark
8384 * isolate CMA pageblocks even when they are not movable in fact
8385 * so consider them movable here.
8387 if (is_migrate_cma(migratetype))
8393 for (; iter < pageblock_nr_pages - offset; iter++) {
8394 if (!pfn_valid_within(pfn + iter))
8397 page = pfn_to_page(pfn + iter);
8400 * Both, bootmem allocations and memory holes are marked
8401 * PG_reserved and are unmovable. We can even have unmovable
8402 * allocations inside ZONE_MOVABLE, for example when
8403 * specifying "movablecore".
8405 if (PageReserved(page))
8409 * If the zone is movable and we have ruled out all reserved
8410 * pages then it should be reasonably safe to assume the rest
8413 if (zone_idx(zone) == ZONE_MOVABLE)
8417 * Hugepages are not in LRU lists, but they're movable.
8418 * THPs are on the LRU, but need to be counted as #small pages.
8419 * We need not scan over tail pages because we don't
8420 * handle each tail page individually in migration.
8422 if (PageHuge(page) || PageTransCompound(page)) {
8423 struct page *head = compound_head(page);
8424 unsigned int skip_pages;
8426 if (PageHuge(page)) {
8427 if (!hugepage_migration_supported(page_hstate(head)))
8429 } else if (!PageLRU(head) && !__PageMovable(head)) {
8433 skip_pages = compound_nr(head) - (page - head);
8434 iter += skip_pages - 1;
8439 * We can't use page_count without pin a page
8440 * because another CPU can free compound page.
8441 * This check already skips compound tails of THP
8442 * because their page->_refcount is zero at all time.
8444 if (!page_ref_count(page)) {
8445 if (PageBuddy(page))
8446 iter += (1 << buddy_order(page)) - 1;
8451 * The HWPoisoned page may be not in buddy system, and
8452 * page_count() is not 0.
8454 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8458 * We treat all PageOffline() pages as movable when offlining
8459 * to give drivers a chance to decrement their reference count
8460 * in MEM_GOING_OFFLINE in order to indicate that these pages
8461 * can be offlined as there are no direct references anymore.
8462 * For actually unmovable PageOffline() where the driver does
8463 * not support this, we will fail later when trying to actually
8464 * move these pages that still have a reference count > 0.
8465 * (false negatives in this function only)
8467 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8470 if (__PageMovable(page) || PageLRU(page))
8474 * If there are RECLAIMABLE pages, we need to check
8475 * it. But now, memory offline itself doesn't call
8476 * shrink_node_slabs() and it still to be fixed.
8483 #ifdef CONFIG_CONTIG_ALLOC
8484 static unsigned long pfn_max_align_down(unsigned long pfn)
8486 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8487 pageblock_nr_pages) - 1);
8490 static unsigned long pfn_max_align_up(unsigned long pfn)
8492 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8493 pageblock_nr_pages));
8496 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8497 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8498 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8499 static void alloc_contig_dump_pages(struct list_head *page_list)
8501 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8503 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8507 list_for_each_entry(page, page_list, lru)
8508 dump_page(page, "migration failure");
8512 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8517 /* [start, end) must belong to a single zone. */
8518 static int __alloc_contig_migrate_range(struct compact_control *cc,
8519 unsigned long start, unsigned long end)
8521 /* This function is based on compact_zone() from compaction.c. */
8522 unsigned int nr_reclaimed;
8523 unsigned long pfn = start;
8524 unsigned int tries = 0;
8526 struct migration_target_control mtc = {
8527 .nid = zone_to_nid(cc->zone),
8528 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8533 while (pfn < end || !list_empty(&cc->migratepages)) {
8534 if (fatal_signal_pending(current)) {
8539 if (list_empty(&cc->migratepages)) {
8540 cc->nr_migratepages = 0;
8541 pfn = isolate_migratepages_range(cc, pfn, end);
8547 } else if (++tries == 5) {
8548 ret = ret < 0 ? ret : -EBUSY;
8552 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8554 cc->nr_migratepages -= nr_reclaimed;
8556 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8557 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8560 alloc_contig_dump_pages(&cc->migratepages);
8561 putback_movable_pages(&cc->migratepages);
8568 * alloc_contig_range() -- tries to allocate given range of pages
8569 * @start: start PFN to allocate
8570 * @end: one-past-the-last PFN to allocate
8571 * @migratetype: migratetype of the underlaying pageblocks (either
8572 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8573 * in range must have the same migratetype and it must
8574 * be either of the two.
8575 * @gfp_mask: GFP mask to use during compaction
8577 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8578 * aligned. The PFN range must belong to a single zone.
8580 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8581 * pageblocks in the range. Once isolated, the pageblocks should not
8582 * be modified by others.
8584 * Return: zero on success or negative error code. On success all
8585 * pages which PFN is in [start, end) are allocated for the caller and
8586 * need to be freed with free_contig_range().
8588 int alloc_contig_range(unsigned long start, unsigned long end,
8589 unsigned migratetype, gfp_t gfp_mask)
8591 unsigned long outer_start, outer_end;
8595 struct compact_control cc = {
8596 .nr_migratepages = 0,
8598 .zone = page_zone(pfn_to_page(start)),
8599 .mode = MIGRATE_SYNC,
8600 .ignore_skip_hint = true,
8601 .no_set_skip_hint = true,
8602 .gfp_mask = current_gfp_context(gfp_mask),
8603 .alloc_contig = true,
8605 INIT_LIST_HEAD(&cc.migratepages);
8608 * What we do here is we mark all pageblocks in range as
8609 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8610 * have different sizes, and due to the way page allocator
8611 * work, we align the range to biggest of the two pages so
8612 * that page allocator won't try to merge buddies from
8613 * different pageblocks and change MIGRATE_ISOLATE to some
8614 * other migration type.
8616 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8617 * migrate the pages from an unaligned range (ie. pages that
8618 * we are interested in). This will put all the pages in
8619 * range back to page allocator as MIGRATE_ISOLATE.
8621 * When this is done, we take the pages in range from page
8622 * allocator removing them from the buddy system. This way
8623 * page allocator will never consider using them.
8625 * This lets us mark the pageblocks back as
8626 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8627 * aligned range but not in the unaligned, original range are
8628 * put back to page allocator so that buddy can use them.
8631 ret = start_isolate_page_range(pfn_max_align_down(start),
8632 pfn_max_align_up(end), migratetype, 0);
8636 drain_all_pages(cc.zone);
8639 * In case of -EBUSY, we'd like to know which page causes problem.
8640 * So, just fall through. test_pages_isolated() has a tracepoint
8641 * which will report the busy page.
8643 * It is possible that busy pages could become available before
8644 * the call to test_pages_isolated, and the range will actually be
8645 * allocated. So, if we fall through be sure to clear ret so that
8646 * -EBUSY is not accidentally used or returned to caller.
8648 ret = __alloc_contig_migrate_range(&cc, start, end);
8649 if (ret && ret != -EBUSY)
8654 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8655 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8656 * more, all pages in [start, end) are free in page allocator.
8657 * What we are going to do is to allocate all pages from
8658 * [start, end) (that is remove them from page allocator).
8660 * The only problem is that pages at the beginning and at the
8661 * end of interesting range may be not aligned with pages that
8662 * page allocator holds, ie. they can be part of higher order
8663 * pages. Because of this, we reserve the bigger range and
8664 * once this is done free the pages we are not interested in.
8666 * We don't have to hold zone->lock here because the pages are
8667 * isolated thus they won't get removed from buddy.
8671 outer_start = start;
8672 while (!PageBuddy(pfn_to_page(outer_start))) {
8673 if (++order >= MAX_ORDER) {
8674 outer_start = start;
8677 outer_start &= ~0UL << order;
8680 if (outer_start != start) {
8681 order = buddy_order(pfn_to_page(outer_start));
8684 * outer_start page could be small order buddy page and
8685 * it doesn't include start page. Adjust outer_start
8686 * in this case to report failed page properly
8687 * on tracepoint in test_pages_isolated()
8689 if (outer_start + (1UL << order) <= start)
8690 outer_start = start;
8693 /* Make sure the range is really isolated. */
8694 if (test_pages_isolated(outer_start, end, 0)) {
8699 /* Grab isolated pages from freelists. */
8700 outer_end = isolate_freepages_range(&cc, outer_start, end);
8706 /* Free head and tail (if any) */
8707 if (start != outer_start)
8708 free_contig_range(outer_start, start - outer_start);
8709 if (end != outer_end)
8710 free_contig_range(end, outer_end - end);
8713 undo_isolate_page_range(pfn_max_align_down(start),
8714 pfn_max_align_up(end), migratetype);
8717 EXPORT_SYMBOL(alloc_contig_range);
8719 static int __alloc_contig_pages(unsigned long start_pfn,
8720 unsigned long nr_pages, gfp_t gfp_mask)
8722 unsigned long end_pfn = start_pfn + nr_pages;
8724 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8728 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8729 unsigned long nr_pages)
8731 unsigned long i, end_pfn = start_pfn + nr_pages;
8734 for (i = start_pfn; i < end_pfn; i++) {
8735 page = pfn_to_online_page(i);
8739 if (page_zone(page) != z)
8742 if (PageReserved(page))
8745 if (page_count(page) > 0)
8754 static bool zone_spans_last_pfn(const struct zone *zone,
8755 unsigned long start_pfn, unsigned long nr_pages)
8757 unsigned long last_pfn = start_pfn + nr_pages - 1;
8759 return zone_spans_pfn(zone, last_pfn);
8763 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8764 * @nr_pages: Number of contiguous pages to allocate
8765 * @gfp_mask: GFP mask to limit search and used during compaction
8767 * @nodemask: Mask for other possible nodes
8769 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8770 * on an applicable zonelist to find a contiguous pfn range which can then be
8771 * tried for allocation with alloc_contig_range(). This routine is intended
8772 * for allocation requests which can not be fulfilled with the buddy allocator.
8774 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8775 * power of two then the alignment is guaranteed to be to the given nr_pages
8776 * (e.g. 1GB request would be aligned to 1GB).
8778 * Allocated pages can be freed with free_contig_range() or by manually calling
8779 * __free_page() on each allocated page.
8781 * Return: pointer to contiguous pages on success, or NULL if not successful.
8783 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8784 int nid, nodemask_t *nodemask)
8786 unsigned long ret, pfn, flags;
8787 struct zonelist *zonelist;
8791 zonelist = node_zonelist(nid, gfp_mask);
8792 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8793 gfp_zone(gfp_mask), nodemask) {
8794 spin_lock_irqsave(&zone->lock, flags);
8796 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8797 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8798 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8800 * We release the zone lock here because
8801 * alloc_contig_range() will also lock the zone
8802 * at some point. If there's an allocation
8803 * spinning on this lock, it may win the race
8804 * and cause alloc_contig_range() to fail...
8806 spin_unlock_irqrestore(&zone->lock, flags);
8807 ret = __alloc_contig_pages(pfn, nr_pages,
8810 return pfn_to_page(pfn);
8811 spin_lock_irqsave(&zone->lock, flags);
8815 spin_unlock_irqrestore(&zone->lock, flags);
8819 #endif /* CONFIG_CONTIG_ALLOC */
8821 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8823 unsigned int count = 0;
8825 for (; nr_pages--; pfn++) {
8826 struct page *page = pfn_to_page(pfn);
8828 count += page_count(page) != 1;
8831 WARN(count != 0, "%d pages are still in use!\n", count);
8833 EXPORT_SYMBOL(free_contig_range);
8836 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8837 * page high values need to be recalulated.
8839 void __meminit zone_pcp_update(struct zone *zone)
8841 mutex_lock(&pcp_batch_high_lock);
8842 zone_set_pageset_high_and_batch(zone);
8843 mutex_unlock(&pcp_batch_high_lock);
8847 * Effectively disable pcplists for the zone by setting the high limit to 0
8848 * and draining all cpus. A concurrent page freeing on another CPU that's about
8849 * to put the page on pcplist will either finish before the drain and the page
8850 * will be drained, or observe the new high limit and skip the pcplist.
8852 * Must be paired with a call to zone_pcp_enable().
8854 void zone_pcp_disable(struct zone *zone)
8856 mutex_lock(&pcp_batch_high_lock);
8857 __zone_set_pageset_high_and_batch(zone, 0, 1);
8858 __drain_all_pages(zone, true);
8861 void zone_pcp_enable(struct zone *zone)
8863 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
8864 mutex_unlock(&pcp_batch_high_lock);
8867 void zone_pcp_reset(struct zone *zone)
8869 unsigned long flags;
8871 struct per_cpu_pageset *pset;
8873 /* avoid races with drain_pages() */
8874 local_irq_save(flags);
8875 if (zone->pageset != &boot_pageset) {
8876 for_each_online_cpu(cpu) {
8877 pset = per_cpu_ptr(zone->pageset, cpu);
8878 drain_zonestat(zone, pset);
8880 free_percpu(zone->pageset);
8881 zone->pageset = &boot_pageset;
8883 local_irq_restore(flags);
8886 #ifdef CONFIG_MEMORY_HOTREMOVE
8888 * All pages in the range must be in a single zone, must not contain holes,
8889 * must span full sections, and must be isolated before calling this function.
8891 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8893 unsigned long pfn = start_pfn;
8897 unsigned long flags;
8899 offline_mem_sections(pfn, end_pfn);
8900 zone = page_zone(pfn_to_page(pfn));
8901 spin_lock_irqsave(&zone->lock, flags);
8902 while (pfn < end_pfn) {
8903 page = pfn_to_page(pfn);
8905 * The HWPoisoned page may be not in buddy system, and
8906 * page_count() is not 0.
8908 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8913 * At this point all remaining PageOffline() pages have a
8914 * reference count of 0 and can simply be skipped.
8916 if (PageOffline(page)) {
8917 BUG_ON(page_count(page));
8918 BUG_ON(PageBuddy(page));
8923 BUG_ON(page_count(page));
8924 BUG_ON(!PageBuddy(page));
8925 order = buddy_order(page);
8926 del_page_from_free_list(page, zone, order);
8927 pfn += (1 << order);
8929 spin_unlock_irqrestore(&zone->lock, flags);
8933 bool is_free_buddy_page(struct page *page)
8935 struct zone *zone = page_zone(page);
8936 unsigned long pfn = page_to_pfn(page);
8937 unsigned long flags;
8940 spin_lock_irqsave(&zone->lock, flags);
8941 for (order = 0; order < MAX_ORDER; order++) {
8942 struct page *page_head = page - (pfn & ((1 << order) - 1));
8944 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
8947 spin_unlock_irqrestore(&zone->lock, flags);
8949 return order < MAX_ORDER;
8952 #ifdef CONFIG_MEMORY_FAILURE
8954 * Break down a higher-order page in sub-pages, and keep our target out of
8957 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8958 struct page *target, int low, int high,
8961 unsigned long size = 1 << high;
8962 struct page *current_buddy, *next_page;
8964 while (high > low) {
8968 if (target >= &page[size]) {
8969 next_page = page + size;
8970 current_buddy = page;
8973 current_buddy = page + size;
8976 if (set_page_guard(zone, current_buddy, high, migratetype))
8979 if (current_buddy != target) {
8980 add_to_free_list(current_buddy, zone, high, migratetype);
8981 set_buddy_order(current_buddy, high);
8988 * Take a page that will be marked as poisoned off the buddy allocator.
8990 bool take_page_off_buddy(struct page *page)
8992 struct zone *zone = page_zone(page);
8993 unsigned long pfn = page_to_pfn(page);
8994 unsigned long flags;
8998 spin_lock_irqsave(&zone->lock, flags);
8999 for (order = 0; order < MAX_ORDER; order++) {
9000 struct page *page_head = page - (pfn & ((1 << order) - 1));
9001 int page_order = buddy_order(page_head);
9003 if (PageBuddy(page_head) && page_order >= order) {
9004 unsigned long pfn_head = page_to_pfn(page_head);
9005 int migratetype = get_pfnblock_migratetype(page_head,
9008 del_page_from_free_list(page_head, zone, page_order);
9009 break_down_buddy_pages(zone, page_head, page, 0,
9010 page_order, migratetype);
9014 if (page_count(page_head) > 0)
9017 spin_unlock_irqrestore(&zone->lock, flags);