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 bool page_poisoning_requested = false;
791 #ifdef CONFIG_PAGE_POISONING
793 * Page poisoning is debug page alloc for some arches. If
794 * either of those options are enabled, enable poisoning.
796 if (page_poisoning_enabled() ||
797 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
798 debug_pagealloc_enabled())) {
799 static_branch_enable(&_page_poisoning_enabled);
800 page_poisoning_requested = true;
804 if (_init_on_alloc_enabled_early) {
805 if (page_poisoning_requested)
806 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
807 "will take precedence over init_on_alloc\n");
809 static_branch_enable(&init_on_alloc);
811 if (_init_on_free_enabled_early) {
812 if (page_poisoning_requested)
813 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
814 "will take precedence over init_on_free\n");
816 static_branch_enable(&init_on_free);
819 #ifdef CONFIG_DEBUG_PAGEALLOC
820 if (!debug_pagealloc_enabled())
823 static_branch_enable(&_debug_pagealloc_enabled);
825 if (!debug_guardpage_minorder())
828 static_branch_enable(&_debug_guardpage_enabled);
832 static inline void set_buddy_order(struct page *page, unsigned int order)
834 set_page_private(page, order);
835 __SetPageBuddy(page);
839 * This function checks whether a page is free && is the buddy
840 * we can coalesce a page and its buddy if
841 * (a) the buddy is not in a hole (check before calling!) &&
842 * (b) the buddy is in the buddy system &&
843 * (c) a page and its buddy have the same order &&
844 * (d) a page and its buddy are in the same zone.
846 * For recording whether a page is in the buddy system, we set PageBuddy.
847 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
849 * For recording page's order, we use page_private(page).
851 static inline bool page_is_buddy(struct page *page, struct page *buddy,
854 if (!page_is_guard(buddy) && !PageBuddy(buddy))
857 if (buddy_order(buddy) != order)
861 * zone check is done late to avoid uselessly calculating
862 * zone/node ids for pages that could never merge.
864 if (page_zone_id(page) != page_zone_id(buddy))
867 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
872 #ifdef CONFIG_COMPACTION
873 static inline struct capture_control *task_capc(struct zone *zone)
875 struct capture_control *capc = current->capture_control;
877 return unlikely(capc) &&
878 !(current->flags & PF_KTHREAD) &&
880 capc->cc->zone == zone ? capc : NULL;
884 compaction_capture(struct capture_control *capc, struct page *page,
885 int order, int migratetype)
887 if (!capc || order != capc->cc->order)
890 /* Do not accidentally pollute CMA or isolated regions*/
891 if (is_migrate_cma(migratetype) ||
892 is_migrate_isolate(migratetype))
896 * Do not let lower order allocations polluate a movable pageblock.
897 * This might let an unmovable request use a reclaimable pageblock
898 * and vice-versa but no more than normal fallback logic which can
899 * have trouble finding a high-order free page.
901 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
909 static inline struct capture_control *task_capc(struct zone *zone)
915 compaction_capture(struct capture_control *capc, struct page *page,
916 int order, int migratetype)
920 #endif /* CONFIG_COMPACTION */
922 /* Used for pages not on another list */
923 static inline void add_to_free_list(struct page *page, struct zone *zone,
924 unsigned int order, int migratetype)
926 struct free_area *area = &zone->free_area[order];
928 list_add(&page->lru, &area->free_list[migratetype]);
932 /* Used for pages not on another list */
933 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
934 unsigned int order, int migratetype)
936 struct free_area *area = &zone->free_area[order];
938 list_add_tail(&page->lru, &area->free_list[migratetype]);
943 * Used for pages which are on another list. Move the pages to the tail
944 * of the list - so the moved pages won't immediately be considered for
945 * allocation again (e.g., optimization for memory onlining).
947 static inline void move_to_free_list(struct page *page, struct zone *zone,
948 unsigned int order, int migratetype)
950 struct free_area *area = &zone->free_area[order];
952 list_move_tail(&page->lru, &area->free_list[migratetype]);
955 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
958 /* clear reported state and update reported page count */
959 if (page_reported(page))
960 __ClearPageReported(page);
962 list_del(&page->lru);
963 __ClearPageBuddy(page);
964 set_page_private(page, 0);
965 zone->free_area[order].nr_free--;
969 * If this is not the largest possible page, check if the buddy
970 * of the next-highest order is free. If it is, it's possible
971 * that pages are being freed that will coalesce soon. In case,
972 * that is happening, add the free page to the tail of the list
973 * so it's less likely to be used soon and more likely to be merged
974 * as a higher order page
977 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
978 struct page *page, unsigned int order)
980 struct page *higher_page, *higher_buddy;
981 unsigned long combined_pfn;
983 if (order >= MAX_ORDER - 2)
986 if (!pfn_valid_within(buddy_pfn))
989 combined_pfn = buddy_pfn & pfn;
990 higher_page = page + (combined_pfn - pfn);
991 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
992 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
994 return pfn_valid_within(buddy_pfn) &&
995 page_is_buddy(higher_page, higher_buddy, order + 1);
999 * Freeing function for a buddy system allocator.
1001 * The concept of a buddy system is to maintain direct-mapped table
1002 * (containing bit values) for memory blocks of various "orders".
1003 * The bottom level table contains the map for the smallest allocatable
1004 * units of memory (here, pages), and each level above it describes
1005 * pairs of units from the levels below, hence, "buddies".
1006 * At a high level, all that happens here is marking the table entry
1007 * at the bottom level available, and propagating the changes upward
1008 * as necessary, plus some accounting needed to play nicely with other
1009 * parts of the VM system.
1010 * At each level, we keep a list of pages, which are heads of continuous
1011 * free pages of length of (1 << order) and marked with PageBuddy.
1012 * Page's order is recorded in page_private(page) field.
1013 * So when we are allocating or freeing one, we can derive the state of the
1014 * other. That is, if we allocate a small block, and both were
1015 * free, the remainder of the region must be split into blocks.
1016 * If a block is freed, and its buddy is also free, then this
1017 * triggers coalescing into a block of larger size.
1022 static inline void __free_one_page(struct page *page,
1024 struct zone *zone, unsigned int order,
1025 int migratetype, fpi_t fpi_flags)
1027 struct capture_control *capc = task_capc(zone);
1028 unsigned long buddy_pfn;
1029 unsigned long combined_pfn;
1030 unsigned int max_order;
1034 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1036 VM_BUG_ON(!zone_is_initialized(zone));
1037 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1039 VM_BUG_ON(migratetype == -1);
1040 if (likely(!is_migrate_isolate(migratetype)))
1041 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1043 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1044 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1047 while (order < max_order) {
1048 if (compaction_capture(capc, page, order, migratetype)) {
1049 __mod_zone_freepage_state(zone, -(1 << order),
1053 buddy_pfn = __find_buddy_pfn(pfn, order);
1054 buddy = page + (buddy_pfn - pfn);
1056 if (!pfn_valid_within(buddy_pfn))
1058 if (!page_is_buddy(page, buddy, order))
1061 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1062 * merge with it and move up one order.
1064 if (page_is_guard(buddy))
1065 clear_page_guard(zone, buddy, order, migratetype);
1067 del_page_from_free_list(buddy, zone, order);
1068 combined_pfn = buddy_pfn & pfn;
1069 page = page + (combined_pfn - pfn);
1073 if (order < MAX_ORDER - 1) {
1074 /* If we are here, it means order is >= pageblock_order.
1075 * We want to prevent merge between freepages on isolate
1076 * pageblock and normal pageblock. Without this, pageblock
1077 * isolation could cause incorrect freepage or CMA accounting.
1079 * We don't want to hit this code for the more frequent
1080 * low-order merging.
1082 if (unlikely(has_isolate_pageblock(zone))) {
1085 buddy_pfn = __find_buddy_pfn(pfn, order);
1086 buddy = page + (buddy_pfn - pfn);
1087 buddy_mt = get_pageblock_migratetype(buddy);
1089 if (migratetype != buddy_mt
1090 && (is_migrate_isolate(migratetype) ||
1091 is_migrate_isolate(buddy_mt)))
1094 max_order = order + 1;
1095 goto continue_merging;
1099 set_buddy_order(page, order);
1101 if (fpi_flags & FPI_TO_TAIL)
1103 else if (is_shuffle_order(order))
1104 to_tail = shuffle_pick_tail();
1106 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1109 add_to_free_list_tail(page, zone, order, migratetype);
1111 add_to_free_list(page, zone, order, migratetype);
1113 /* Notify page reporting subsystem of freed page */
1114 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1115 page_reporting_notify_free(order);
1119 * A bad page could be due to a number of fields. Instead of multiple branches,
1120 * try and check multiple fields with one check. The caller must do a detailed
1121 * check if necessary.
1123 static inline bool page_expected_state(struct page *page,
1124 unsigned long check_flags)
1126 if (unlikely(atomic_read(&page->_mapcount) != -1))
1129 if (unlikely((unsigned long)page->mapping |
1130 page_ref_count(page) |
1134 (page->flags & check_flags)))
1140 static const char *page_bad_reason(struct page *page, unsigned long flags)
1142 const char *bad_reason = NULL;
1144 if (unlikely(atomic_read(&page->_mapcount) != -1))
1145 bad_reason = "nonzero mapcount";
1146 if (unlikely(page->mapping != NULL))
1147 bad_reason = "non-NULL mapping";
1148 if (unlikely(page_ref_count(page) != 0))
1149 bad_reason = "nonzero _refcount";
1150 if (unlikely(page->flags & flags)) {
1151 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1152 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1154 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1157 if (unlikely(page->memcg_data))
1158 bad_reason = "page still charged to cgroup";
1163 static void check_free_page_bad(struct page *page)
1166 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1169 static inline int check_free_page(struct page *page)
1171 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1174 /* Something has gone sideways, find it */
1175 check_free_page_bad(page);
1179 static int free_tail_pages_check(struct page *head_page, struct page *page)
1184 * We rely page->lru.next never has bit 0 set, unless the page
1185 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1187 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1189 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1193 switch (page - head_page) {
1195 /* the first tail page: ->mapping may be compound_mapcount() */
1196 if (unlikely(compound_mapcount(page))) {
1197 bad_page(page, "nonzero compound_mapcount");
1203 * the second tail page: ->mapping is
1204 * deferred_list.next -- ignore value.
1208 if (page->mapping != TAIL_MAPPING) {
1209 bad_page(page, "corrupted mapping in tail page");
1214 if (unlikely(!PageTail(page))) {
1215 bad_page(page, "PageTail not set");
1218 if (unlikely(compound_head(page) != head_page)) {
1219 bad_page(page, "compound_head not consistent");
1224 page->mapping = NULL;
1225 clear_compound_head(page);
1229 static void kernel_init_free_pages(struct page *page, int numpages)
1233 /* s390's use of memset() could override KASAN redzones. */
1234 kasan_disable_current();
1235 for (i = 0; i < numpages; i++) {
1236 u8 tag = page_kasan_tag(page + i);
1237 page_kasan_tag_reset(page + i);
1238 clear_highpage(page + i);
1239 page_kasan_tag_set(page + i, tag);
1241 kasan_enable_current();
1244 static __always_inline bool free_pages_prepare(struct page *page,
1245 unsigned int order, bool check_free, fpi_t fpi_flags)
1250 VM_BUG_ON_PAGE(PageTail(page), page);
1252 trace_mm_page_free(page, order);
1254 if (unlikely(PageHWPoison(page)) && !order) {
1256 * Do not let hwpoison pages hit pcplists/buddy
1257 * Untie memcg state and reset page's owner
1259 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1260 __memcg_kmem_uncharge_page(page, order);
1261 reset_page_owner(page, order);
1266 * Check tail pages before head page information is cleared to
1267 * avoid checking PageCompound for order-0 pages.
1269 if (unlikely(order)) {
1270 bool compound = PageCompound(page);
1273 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1276 ClearPageDoubleMap(page);
1277 for (i = 1; i < (1 << order); i++) {
1279 bad += free_tail_pages_check(page, page + i);
1280 if (unlikely(check_free_page(page + i))) {
1284 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1287 if (PageMappingFlags(page))
1288 page->mapping = NULL;
1289 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1290 __memcg_kmem_uncharge_page(page, order);
1292 bad += check_free_page(page);
1296 page_cpupid_reset_last(page);
1297 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1298 reset_page_owner(page, order);
1300 if (!PageHighMem(page)) {
1301 debug_check_no_locks_freed(page_address(page),
1302 PAGE_SIZE << order);
1303 debug_check_no_obj_freed(page_address(page),
1304 PAGE_SIZE << order);
1307 kernel_poison_pages(page, 1 << order);
1310 * As memory initialization might be integrated into KASAN,
1311 * kasan_free_pages and kernel_init_free_pages must be
1312 * kept together to avoid discrepancies in behavior.
1314 * With hardware tag-based KASAN, memory tags must be set before the
1315 * page becomes unavailable via debug_pagealloc or arch_free_page.
1317 init = want_init_on_free();
1318 if (init && !kasan_has_integrated_init())
1319 kernel_init_free_pages(page, 1 << order);
1320 kasan_free_nondeferred_pages(page, order, init, fpi_flags);
1323 * arch_free_page() can make the page's contents inaccessible. s390
1324 * does this. So nothing which can access the page's contents should
1325 * happen after this.
1327 arch_free_page(page, order);
1329 debug_pagealloc_unmap_pages(page, 1 << order);
1334 #ifdef CONFIG_DEBUG_VM
1336 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1337 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1338 * moved from pcp lists to free lists.
1340 static bool free_pcp_prepare(struct page *page)
1342 return free_pages_prepare(page, 0, true, FPI_NONE);
1345 static bool bulkfree_pcp_prepare(struct page *page)
1347 if (debug_pagealloc_enabled_static())
1348 return check_free_page(page);
1354 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1355 * moving from pcp lists to free list in order to reduce overhead. With
1356 * debug_pagealloc enabled, they are checked also immediately when being freed
1359 static bool free_pcp_prepare(struct page *page)
1361 if (debug_pagealloc_enabled_static())
1362 return free_pages_prepare(page, 0, true, FPI_NONE);
1364 return free_pages_prepare(page, 0, false, FPI_NONE);
1367 static bool bulkfree_pcp_prepare(struct page *page)
1369 return check_free_page(page);
1371 #endif /* CONFIG_DEBUG_VM */
1373 static inline void prefetch_buddy(struct page *page)
1375 unsigned long pfn = page_to_pfn(page);
1376 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1377 struct page *buddy = page + (buddy_pfn - pfn);
1383 * Frees a number of pages from the PCP lists
1384 * Assumes all pages on list are in same zone, and of same order.
1385 * count is the number of pages to free.
1387 * If the zone was previously in an "all pages pinned" state then look to
1388 * see if this freeing clears that state.
1390 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1391 * pinned" detection logic.
1393 static void free_pcppages_bulk(struct zone *zone, int count,
1394 struct per_cpu_pages *pcp)
1396 int migratetype = 0;
1398 int prefetch_nr = READ_ONCE(pcp->batch);
1399 bool isolated_pageblocks;
1400 struct page *page, *tmp;
1404 * Ensure proper count is passed which otherwise would stuck in the
1405 * below while (list_empty(list)) loop.
1407 count = min(pcp->count, count);
1409 struct list_head *list;
1412 * Remove pages from lists in a round-robin fashion. A
1413 * batch_free count is maintained that is incremented when an
1414 * empty list is encountered. This is so more pages are freed
1415 * off fuller lists instead of spinning excessively around empty
1420 if (++migratetype == MIGRATE_PCPTYPES)
1422 list = &pcp->lists[migratetype];
1423 } while (list_empty(list));
1425 /* This is the only non-empty list. Free them all. */
1426 if (batch_free == MIGRATE_PCPTYPES)
1430 page = list_last_entry(list, struct page, lru);
1431 /* must delete to avoid corrupting pcp list */
1432 list_del(&page->lru);
1435 if (bulkfree_pcp_prepare(page))
1438 list_add_tail(&page->lru, &head);
1441 * We are going to put the page back to the global
1442 * pool, prefetch its buddy to speed up later access
1443 * under zone->lock. It is believed the overhead of
1444 * an additional test and calculating buddy_pfn here
1445 * can be offset by reduced memory latency later. To
1446 * avoid excessive prefetching due to large count, only
1447 * prefetch buddy for the first pcp->batch nr of pages.
1450 prefetch_buddy(page);
1453 } while (--count && --batch_free && !list_empty(list));
1456 spin_lock(&zone->lock);
1457 isolated_pageblocks = has_isolate_pageblock(zone);
1460 * Use safe version since after __free_one_page(),
1461 * page->lru.next will not point to original list.
1463 list_for_each_entry_safe(page, tmp, &head, lru) {
1464 int mt = get_pcppage_migratetype(page);
1465 /* MIGRATE_ISOLATE page should not go to pcplists */
1466 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1467 /* Pageblock could have been isolated meanwhile */
1468 if (unlikely(isolated_pageblocks))
1469 mt = get_pageblock_migratetype(page);
1471 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1472 trace_mm_page_pcpu_drain(page, 0, mt);
1474 spin_unlock(&zone->lock);
1477 static void free_one_page(struct zone *zone,
1478 struct page *page, unsigned long pfn,
1480 int migratetype, fpi_t fpi_flags)
1482 spin_lock(&zone->lock);
1483 if (unlikely(has_isolate_pageblock(zone) ||
1484 is_migrate_isolate(migratetype))) {
1485 migratetype = get_pfnblock_migratetype(page, pfn);
1487 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1488 spin_unlock(&zone->lock);
1491 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1492 unsigned long zone, int nid)
1494 mm_zero_struct_page(page);
1495 set_page_links(page, zone, nid, pfn);
1496 init_page_count(page);
1497 page_mapcount_reset(page);
1498 page_cpupid_reset_last(page);
1499 page_kasan_tag_reset(page);
1501 INIT_LIST_HEAD(&page->lru);
1502 #ifdef WANT_PAGE_VIRTUAL
1503 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1504 if (!is_highmem_idx(zone))
1505 set_page_address(page, __va(pfn << PAGE_SHIFT));
1509 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1510 static void __meminit init_reserved_page(unsigned long pfn)
1515 if (!early_page_uninitialised(pfn))
1518 nid = early_pfn_to_nid(pfn);
1519 pgdat = NODE_DATA(nid);
1521 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1522 struct zone *zone = &pgdat->node_zones[zid];
1524 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1527 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1530 static inline void init_reserved_page(unsigned long pfn)
1533 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1536 * Initialised pages do not have PageReserved set. This function is
1537 * called for each range allocated by the bootmem allocator and
1538 * marks the pages PageReserved. The remaining valid pages are later
1539 * sent to the buddy page allocator.
1541 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1543 unsigned long start_pfn = PFN_DOWN(start);
1544 unsigned long end_pfn = PFN_UP(end);
1546 for (; start_pfn < end_pfn; start_pfn++) {
1547 if (pfn_valid(start_pfn)) {
1548 struct page *page = pfn_to_page(start_pfn);
1550 init_reserved_page(start_pfn);
1552 /* Avoid false-positive PageTail() */
1553 INIT_LIST_HEAD(&page->lru);
1556 * no need for atomic set_bit because the struct
1557 * page is not visible yet so nobody should
1560 __SetPageReserved(page);
1565 static void __free_pages_ok(struct page *page, unsigned int order,
1568 unsigned long flags;
1570 unsigned long pfn = page_to_pfn(page);
1572 if (!free_pages_prepare(page, order, true, fpi_flags))
1575 migratetype = get_pfnblock_migratetype(page, pfn);
1576 local_irq_save(flags);
1577 __count_vm_events(PGFREE, 1 << order);
1578 free_one_page(page_zone(page), page, pfn, order, migratetype,
1580 local_irq_restore(flags);
1583 void __free_pages_core(struct page *page, unsigned int order)
1585 unsigned int nr_pages = 1 << order;
1586 struct page *p = page;
1590 * When initializing the memmap, __init_single_page() sets the refcount
1591 * of all pages to 1 ("allocated"/"not free"). We have to set the
1592 * refcount of all involved pages to 0.
1595 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1597 __ClearPageReserved(p);
1598 set_page_count(p, 0);
1600 __ClearPageReserved(p);
1601 set_page_count(p, 0);
1603 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1606 * Bypass PCP and place fresh pages right to the tail, primarily
1607 * relevant for memory onlining.
1609 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1612 #ifdef CONFIG_NEED_MULTIPLE_NODES
1615 * During memory init memblocks map pfns to nids. The search is expensive and
1616 * this caches recent lookups. The implementation of __early_pfn_to_nid
1617 * treats start/end as pfns.
1619 struct mminit_pfnnid_cache {
1620 unsigned long last_start;
1621 unsigned long last_end;
1625 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1628 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1630 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1631 struct mminit_pfnnid_cache *state)
1633 unsigned long start_pfn, end_pfn;
1636 if (state->last_start <= pfn && pfn < state->last_end)
1637 return state->last_nid;
1639 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1640 if (nid != NUMA_NO_NODE) {
1641 state->last_start = start_pfn;
1642 state->last_end = end_pfn;
1643 state->last_nid = nid;
1649 int __meminit early_pfn_to_nid(unsigned long pfn)
1651 static DEFINE_SPINLOCK(early_pfn_lock);
1654 spin_lock(&early_pfn_lock);
1655 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1657 nid = first_online_node;
1658 spin_unlock(&early_pfn_lock);
1662 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1664 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1667 if (early_page_uninitialised(pfn))
1669 __free_pages_core(page, order);
1673 * Check that the whole (or subset of) a pageblock given by the interval of
1674 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1675 * with the migration of free compaction scanner. The scanners then need to
1676 * use only pfn_valid_within() check for arches that allow holes within
1679 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1681 * It's possible on some configurations to have a setup like node0 node1 node0
1682 * i.e. it's possible that all pages within a zones range of pages do not
1683 * belong to a single zone. We assume that a border between node0 and node1
1684 * can occur within a single pageblock, but not a node0 node1 node0
1685 * interleaving within a single pageblock. It is therefore sufficient to check
1686 * the first and last page of a pageblock and avoid checking each individual
1687 * page in a pageblock.
1689 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1690 unsigned long end_pfn, struct zone *zone)
1692 struct page *start_page;
1693 struct page *end_page;
1695 /* end_pfn is one past the range we are checking */
1698 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1701 start_page = pfn_to_online_page(start_pfn);
1705 if (page_zone(start_page) != zone)
1708 end_page = pfn_to_page(end_pfn);
1710 /* This gives a shorter code than deriving page_zone(end_page) */
1711 if (page_zone_id(start_page) != page_zone_id(end_page))
1717 void set_zone_contiguous(struct zone *zone)
1719 unsigned long block_start_pfn = zone->zone_start_pfn;
1720 unsigned long block_end_pfn;
1722 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1723 for (; block_start_pfn < zone_end_pfn(zone);
1724 block_start_pfn = block_end_pfn,
1725 block_end_pfn += pageblock_nr_pages) {
1727 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1729 if (!__pageblock_pfn_to_page(block_start_pfn,
1730 block_end_pfn, zone))
1735 /* We confirm that there is no hole */
1736 zone->contiguous = true;
1739 void clear_zone_contiguous(struct zone *zone)
1741 zone->contiguous = false;
1744 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1745 static void __init deferred_free_range(unsigned long pfn,
1746 unsigned long nr_pages)
1754 page = pfn_to_page(pfn);
1756 /* Free a large naturally-aligned chunk if possible */
1757 if (nr_pages == pageblock_nr_pages &&
1758 (pfn & (pageblock_nr_pages - 1)) == 0) {
1759 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1760 __free_pages_core(page, pageblock_order);
1764 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1765 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1766 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1767 __free_pages_core(page, 0);
1771 /* Completion tracking for deferred_init_memmap() threads */
1772 static atomic_t pgdat_init_n_undone __initdata;
1773 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1775 static inline void __init pgdat_init_report_one_done(void)
1777 if (atomic_dec_and_test(&pgdat_init_n_undone))
1778 complete(&pgdat_init_all_done_comp);
1782 * Returns true if page needs to be initialized or freed to buddy allocator.
1784 * First we check if pfn is valid on architectures where it is possible to have
1785 * holes within pageblock_nr_pages. On systems where it is not possible, this
1786 * function is optimized out.
1788 * Then, we check if a current large page is valid by only checking the validity
1791 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1793 if (!pfn_valid_within(pfn))
1795 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1801 * Free pages to buddy allocator. Try to free aligned pages in
1802 * pageblock_nr_pages sizes.
1804 static void __init deferred_free_pages(unsigned long pfn,
1805 unsigned long end_pfn)
1807 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1808 unsigned long nr_free = 0;
1810 for (; pfn < end_pfn; pfn++) {
1811 if (!deferred_pfn_valid(pfn)) {
1812 deferred_free_range(pfn - nr_free, nr_free);
1814 } else if (!(pfn & nr_pgmask)) {
1815 deferred_free_range(pfn - nr_free, nr_free);
1821 /* Free the last block of pages to allocator */
1822 deferred_free_range(pfn - nr_free, nr_free);
1826 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1827 * by performing it only once every pageblock_nr_pages.
1828 * Return number of pages initialized.
1830 static unsigned long __init deferred_init_pages(struct zone *zone,
1832 unsigned long end_pfn)
1834 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1835 int nid = zone_to_nid(zone);
1836 unsigned long nr_pages = 0;
1837 int zid = zone_idx(zone);
1838 struct page *page = NULL;
1840 for (; pfn < end_pfn; pfn++) {
1841 if (!deferred_pfn_valid(pfn)) {
1844 } else if (!page || !(pfn & nr_pgmask)) {
1845 page = pfn_to_page(pfn);
1849 __init_single_page(page, pfn, zid, nid);
1856 * This function is meant to pre-load the iterator for the zone init.
1857 * Specifically it walks through the ranges until we are caught up to the
1858 * first_init_pfn value and exits there. If we never encounter the value we
1859 * return false indicating there are no valid ranges left.
1862 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1863 unsigned long *spfn, unsigned long *epfn,
1864 unsigned long first_init_pfn)
1869 * Start out by walking through the ranges in this zone that have
1870 * already been initialized. We don't need to do anything with them
1871 * so we just need to flush them out of the system.
1873 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1874 if (*epfn <= first_init_pfn)
1876 if (*spfn < first_init_pfn)
1877 *spfn = first_init_pfn;
1886 * Initialize and free pages. We do it in two loops: first we initialize
1887 * struct page, then free to buddy allocator, because while we are
1888 * freeing pages we can access pages that are ahead (computing buddy
1889 * page in __free_one_page()).
1891 * In order to try and keep some memory in the cache we have the loop
1892 * broken along max page order boundaries. This way we will not cause
1893 * any issues with the buddy page computation.
1895 static unsigned long __init
1896 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1897 unsigned long *end_pfn)
1899 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1900 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1901 unsigned long nr_pages = 0;
1904 /* First we loop through and initialize the page values */
1905 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1908 if (mo_pfn <= *start_pfn)
1911 t = min(mo_pfn, *end_pfn);
1912 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1914 if (mo_pfn < *end_pfn) {
1915 *start_pfn = mo_pfn;
1920 /* Reset values and now loop through freeing pages as needed */
1923 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1929 t = min(mo_pfn, epfn);
1930 deferred_free_pages(spfn, t);
1940 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1943 unsigned long spfn, epfn;
1944 struct zone *zone = arg;
1947 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1950 * Initialize and free pages in MAX_ORDER sized increments so that we
1951 * can avoid introducing any issues with the buddy allocator.
1953 while (spfn < end_pfn) {
1954 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1959 /* An arch may override for more concurrency. */
1961 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1966 /* Initialise remaining memory on a node */
1967 static int __init deferred_init_memmap(void *data)
1969 pg_data_t *pgdat = data;
1970 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1971 unsigned long spfn = 0, epfn = 0;
1972 unsigned long first_init_pfn, flags;
1973 unsigned long start = jiffies;
1975 int zid, max_threads;
1978 /* Bind memory initialisation thread to a local node if possible */
1979 if (!cpumask_empty(cpumask))
1980 set_cpus_allowed_ptr(current, cpumask);
1982 pgdat_resize_lock(pgdat, &flags);
1983 first_init_pfn = pgdat->first_deferred_pfn;
1984 if (first_init_pfn == ULONG_MAX) {
1985 pgdat_resize_unlock(pgdat, &flags);
1986 pgdat_init_report_one_done();
1990 /* Sanity check boundaries */
1991 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1992 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1993 pgdat->first_deferred_pfn = ULONG_MAX;
1996 * Once we unlock here, the zone cannot be grown anymore, thus if an
1997 * interrupt thread must allocate this early in boot, zone must be
1998 * pre-grown prior to start of deferred page initialization.
2000 pgdat_resize_unlock(pgdat, &flags);
2002 /* Only the highest zone is deferred so find it */
2003 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2004 zone = pgdat->node_zones + zid;
2005 if (first_init_pfn < zone_end_pfn(zone))
2009 /* If the zone is empty somebody else may have cleared out the zone */
2010 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2014 max_threads = deferred_page_init_max_threads(cpumask);
2016 while (spfn < epfn) {
2017 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2018 struct padata_mt_job job = {
2019 .thread_fn = deferred_init_memmap_chunk,
2022 .size = epfn_align - spfn,
2023 .align = PAGES_PER_SECTION,
2024 .min_chunk = PAGES_PER_SECTION,
2025 .max_threads = max_threads,
2028 padata_do_multithreaded(&job);
2029 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2033 /* Sanity check that the next zone really is unpopulated */
2034 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2036 pr_info("node %d deferred pages initialised in %ums\n",
2037 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2039 pgdat_init_report_one_done();
2044 * If this zone has deferred pages, try to grow it by initializing enough
2045 * deferred pages to satisfy the allocation specified by order, rounded up to
2046 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2047 * of SECTION_SIZE bytes by initializing struct pages in increments of
2048 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2050 * Return true when zone was grown, otherwise return false. We return true even
2051 * when we grow less than requested, to let the caller decide if there are
2052 * enough pages to satisfy the allocation.
2054 * Note: We use noinline because this function is needed only during boot, and
2055 * it is called from a __ref function _deferred_grow_zone. This way we are
2056 * making sure that it is not inlined into permanent text section.
2058 static noinline bool __init
2059 deferred_grow_zone(struct zone *zone, unsigned int order)
2061 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2062 pg_data_t *pgdat = zone->zone_pgdat;
2063 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2064 unsigned long spfn, epfn, flags;
2065 unsigned long nr_pages = 0;
2068 /* Only the last zone may have deferred pages */
2069 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2072 pgdat_resize_lock(pgdat, &flags);
2075 * If someone grew this zone while we were waiting for spinlock, return
2076 * true, as there might be enough pages already.
2078 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2079 pgdat_resize_unlock(pgdat, &flags);
2083 /* If the zone is empty somebody else may have cleared out the zone */
2084 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2085 first_deferred_pfn)) {
2086 pgdat->first_deferred_pfn = ULONG_MAX;
2087 pgdat_resize_unlock(pgdat, &flags);
2088 /* Retry only once. */
2089 return first_deferred_pfn != ULONG_MAX;
2093 * Initialize and free pages in MAX_ORDER sized increments so
2094 * that we can avoid introducing any issues with the buddy
2097 while (spfn < epfn) {
2098 /* update our first deferred PFN for this section */
2099 first_deferred_pfn = spfn;
2101 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2102 touch_nmi_watchdog();
2104 /* We should only stop along section boundaries */
2105 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2108 /* If our quota has been met we can stop here */
2109 if (nr_pages >= nr_pages_needed)
2113 pgdat->first_deferred_pfn = spfn;
2114 pgdat_resize_unlock(pgdat, &flags);
2116 return nr_pages > 0;
2120 * deferred_grow_zone() is __init, but it is called from
2121 * get_page_from_freelist() during early boot until deferred_pages permanently
2122 * disables this call. This is why we have refdata wrapper to avoid warning,
2123 * and to ensure that the function body gets unloaded.
2126 _deferred_grow_zone(struct zone *zone, unsigned int order)
2128 return deferred_grow_zone(zone, order);
2131 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2133 void __init page_alloc_init_late(void)
2138 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2140 /* There will be num_node_state(N_MEMORY) threads */
2141 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2142 for_each_node_state(nid, N_MEMORY) {
2143 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2146 /* Block until all are initialised */
2147 wait_for_completion(&pgdat_init_all_done_comp);
2150 * The number of managed pages has changed due to the initialisation
2151 * so the pcpu batch and high limits needs to be updated or the limits
2152 * will be artificially small.
2154 for_each_populated_zone(zone)
2155 zone_pcp_update(zone);
2158 * We initialized the rest of the deferred pages. Permanently disable
2159 * on-demand struct page initialization.
2161 static_branch_disable(&deferred_pages);
2163 /* Reinit limits that are based on free pages after the kernel is up */
2164 files_maxfiles_init();
2169 /* Discard memblock private memory */
2172 for_each_node_state(nid, N_MEMORY)
2173 shuffle_free_memory(NODE_DATA(nid));
2175 for_each_populated_zone(zone)
2176 set_zone_contiguous(zone);
2180 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2181 void __init init_cma_reserved_pageblock(struct page *page)
2183 unsigned i = pageblock_nr_pages;
2184 struct page *p = page;
2187 __ClearPageReserved(p);
2188 set_page_count(p, 0);
2191 set_pageblock_migratetype(page, MIGRATE_CMA);
2193 if (pageblock_order >= MAX_ORDER) {
2194 i = pageblock_nr_pages;
2197 set_page_refcounted(p);
2198 __free_pages(p, MAX_ORDER - 1);
2199 p += MAX_ORDER_NR_PAGES;
2200 } while (i -= MAX_ORDER_NR_PAGES);
2202 set_page_refcounted(page);
2203 __free_pages(page, pageblock_order);
2206 adjust_managed_page_count(page, pageblock_nr_pages);
2207 page_zone(page)->cma_pages += pageblock_nr_pages;
2212 * The order of subdivision here is critical for the IO subsystem.
2213 * Please do not alter this order without good reasons and regression
2214 * testing. Specifically, as large blocks of memory are subdivided,
2215 * the order in which smaller blocks are delivered depends on the order
2216 * they're subdivided in this function. This is the primary factor
2217 * influencing the order in which pages are delivered to the IO
2218 * subsystem according to empirical testing, and this is also justified
2219 * by considering the behavior of a buddy system containing a single
2220 * large block of memory acted on by a series of small allocations.
2221 * This behavior is a critical factor in sglist merging's success.
2225 static inline void expand(struct zone *zone, struct page *page,
2226 int low, int high, int migratetype)
2228 unsigned long size = 1 << high;
2230 while (high > low) {
2233 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2236 * Mark as guard pages (or page), that will allow to
2237 * merge back to allocator when buddy will be freed.
2238 * Corresponding page table entries will not be touched,
2239 * pages will stay not present in virtual address space
2241 if (set_page_guard(zone, &page[size], high, migratetype))
2244 add_to_free_list(&page[size], zone, high, migratetype);
2245 set_buddy_order(&page[size], high);
2249 static void check_new_page_bad(struct page *page)
2251 if (unlikely(page->flags & __PG_HWPOISON)) {
2252 /* Don't complain about hwpoisoned pages */
2253 page_mapcount_reset(page); /* remove PageBuddy */
2258 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2262 * This page is about to be returned from the page allocator
2264 static inline int check_new_page(struct page *page)
2266 if (likely(page_expected_state(page,
2267 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2270 check_new_page_bad(page);
2274 #ifdef CONFIG_DEBUG_VM
2276 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2277 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2278 * also checked when pcp lists are refilled from the free lists.
2280 static inline bool check_pcp_refill(struct page *page)
2282 if (debug_pagealloc_enabled_static())
2283 return check_new_page(page);
2288 static inline bool check_new_pcp(struct page *page)
2290 return check_new_page(page);
2294 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2295 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2296 * enabled, they are also checked when being allocated from the pcp lists.
2298 static inline bool check_pcp_refill(struct page *page)
2300 return check_new_page(page);
2302 static inline bool check_new_pcp(struct page *page)
2304 if (debug_pagealloc_enabled_static())
2305 return check_new_page(page);
2309 #endif /* CONFIG_DEBUG_VM */
2311 static bool check_new_pages(struct page *page, unsigned int order)
2314 for (i = 0; i < (1 << order); i++) {
2315 struct page *p = page + i;
2317 if (unlikely(check_new_page(p)))
2324 inline void post_alloc_hook(struct page *page, unsigned int order,
2329 set_page_private(page, 0);
2330 set_page_refcounted(page);
2332 arch_alloc_page(page, order);
2333 debug_pagealloc_map_pages(page, 1 << order);
2336 * Page unpoisoning must happen before memory initialization.
2337 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2338 * allocations and the page unpoisoning code will complain.
2340 kernel_unpoison_pages(page, 1 << order);
2343 * As memory initialization might be integrated into KASAN,
2344 * kasan_alloc_pages and kernel_init_free_pages must be
2345 * kept together to avoid discrepancies in behavior.
2347 init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2348 kasan_alloc_pages(page, order, init);
2349 if (init && !kasan_has_integrated_init())
2350 kernel_init_free_pages(page, 1 << order);
2352 set_page_owner(page, order, gfp_flags);
2355 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2356 unsigned int alloc_flags)
2358 post_alloc_hook(page, order, gfp_flags);
2360 if (order && (gfp_flags & __GFP_COMP))
2361 prep_compound_page(page, order);
2364 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2365 * allocate the page. The expectation is that the caller is taking
2366 * steps that will free more memory. The caller should avoid the page
2367 * being used for !PFMEMALLOC purposes.
2369 if (alloc_flags & ALLOC_NO_WATERMARKS)
2370 set_page_pfmemalloc(page);
2372 clear_page_pfmemalloc(page);
2376 * Go through the free lists for the given migratetype and remove
2377 * the smallest available page from the freelists
2379 static __always_inline
2380 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2383 unsigned int current_order;
2384 struct free_area *area;
2387 /* Find a page of the appropriate size in the preferred list */
2388 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2389 area = &(zone->free_area[current_order]);
2390 page = get_page_from_free_area(area, migratetype);
2393 del_page_from_free_list(page, zone, current_order);
2394 expand(zone, page, order, current_order, migratetype);
2395 set_pcppage_migratetype(page, migratetype);
2404 * This array describes the order lists are fallen back to when
2405 * the free lists for the desirable migrate type are depleted
2407 static int fallbacks[MIGRATE_TYPES][3] = {
2408 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2409 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2410 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2412 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2414 #ifdef CONFIG_MEMORY_ISOLATION
2415 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2420 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2423 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2426 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2427 unsigned int order) { return NULL; }
2431 * Move the free pages in a range to the freelist tail of the requested type.
2432 * Note that start_page and end_pages are not aligned on a pageblock
2433 * boundary. If alignment is required, use move_freepages_block()
2435 static int move_freepages(struct zone *zone,
2436 unsigned long start_pfn, unsigned long end_pfn,
2437 int migratetype, int *num_movable)
2442 int pages_moved = 0;
2444 for (pfn = start_pfn; pfn <= end_pfn;) {
2445 if (!pfn_valid_within(pfn)) {
2450 page = pfn_to_page(pfn);
2451 if (!PageBuddy(page)) {
2453 * We assume that pages that could be isolated for
2454 * migration are movable. But we don't actually try
2455 * isolating, as that would be expensive.
2458 (PageLRU(page) || __PageMovable(page)))
2464 /* Make sure we are not inadvertently changing nodes */
2465 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2466 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2468 order = buddy_order(page);
2469 move_to_free_list(page, zone, order, migratetype);
2471 pages_moved += 1 << order;
2477 int move_freepages_block(struct zone *zone, struct page *page,
2478 int migratetype, int *num_movable)
2480 unsigned long start_pfn, end_pfn, pfn;
2485 pfn = page_to_pfn(page);
2486 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2487 end_pfn = start_pfn + pageblock_nr_pages - 1;
2489 /* Do not cross zone boundaries */
2490 if (!zone_spans_pfn(zone, start_pfn))
2492 if (!zone_spans_pfn(zone, end_pfn))
2495 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2499 static void change_pageblock_range(struct page *pageblock_page,
2500 int start_order, int migratetype)
2502 int nr_pageblocks = 1 << (start_order - pageblock_order);
2504 while (nr_pageblocks--) {
2505 set_pageblock_migratetype(pageblock_page, migratetype);
2506 pageblock_page += pageblock_nr_pages;
2511 * When we are falling back to another migratetype during allocation, try to
2512 * steal extra free pages from the same pageblocks to satisfy further
2513 * allocations, instead of polluting multiple pageblocks.
2515 * If we are stealing a relatively large buddy page, it is likely there will
2516 * be more free pages in the pageblock, so try to steal them all. For
2517 * reclaimable and unmovable allocations, we steal regardless of page size,
2518 * as fragmentation caused by those allocations polluting movable pageblocks
2519 * is worse than movable allocations stealing from unmovable and reclaimable
2522 static bool can_steal_fallback(unsigned int order, int start_mt)
2525 * Leaving this order check is intended, although there is
2526 * relaxed order check in next check. The reason is that
2527 * we can actually steal whole pageblock if this condition met,
2528 * but, below check doesn't guarantee it and that is just heuristic
2529 * so could be changed anytime.
2531 if (order >= pageblock_order)
2534 if (order >= pageblock_order / 2 ||
2535 start_mt == MIGRATE_RECLAIMABLE ||
2536 start_mt == MIGRATE_UNMOVABLE ||
2537 page_group_by_mobility_disabled)
2543 static inline bool boost_watermark(struct zone *zone)
2545 unsigned long max_boost;
2547 if (!watermark_boost_factor)
2550 * Don't bother in zones that are unlikely to produce results.
2551 * On small machines, including kdump capture kernels running
2552 * in a small area, boosting the watermark can cause an out of
2553 * memory situation immediately.
2555 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2558 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2559 watermark_boost_factor, 10000);
2562 * high watermark may be uninitialised if fragmentation occurs
2563 * very early in boot so do not boost. We do not fall
2564 * through and boost by pageblock_nr_pages as failing
2565 * allocations that early means that reclaim is not going
2566 * to help and it may even be impossible to reclaim the
2567 * boosted watermark resulting in a hang.
2572 max_boost = max(pageblock_nr_pages, max_boost);
2574 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2581 * This function implements actual steal behaviour. If order is large enough,
2582 * we can steal whole pageblock. If not, we first move freepages in this
2583 * pageblock to our migratetype and determine how many already-allocated pages
2584 * are there in the pageblock with a compatible migratetype. If at least half
2585 * of pages are free or compatible, we can change migratetype of the pageblock
2586 * itself, so pages freed in the future will be put on the correct free list.
2588 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2589 unsigned int alloc_flags, int start_type, bool whole_block)
2591 unsigned int current_order = buddy_order(page);
2592 int free_pages, movable_pages, alike_pages;
2595 old_block_type = get_pageblock_migratetype(page);
2598 * This can happen due to races and we want to prevent broken
2599 * highatomic accounting.
2601 if (is_migrate_highatomic(old_block_type))
2604 /* Take ownership for orders >= pageblock_order */
2605 if (current_order >= pageblock_order) {
2606 change_pageblock_range(page, current_order, start_type);
2611 * Boost watermarks to increase reclaim pressure to reduce the
2612 * likelihood of future fallbacks. Wake kswapd now as the node
2613 * may be balanced overall and kswapd will not wake naturally.
2615 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2616 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2618 /* We are not allowed to try stealing from the whole block */
2622 free_pages = move_freepages_block(zone, page, start_type,
2625 * Determine how many pages are compatible with our allocation.
2626 * For movable allocation, it's the number of movable pages which
2627 * we just obtained. For other types it's a bit more tricky.
2629 if (start_type == MIGRATE_MOVABLE) {
2630 alike_pages = movable_pages;
2633 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2634 * to MOVABLE pageblock, consider all non-movable pages as
2635 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2636 * vice versa, be conservative since we can't distinguish the
2637 * exact migratetype of non-movable pages.
2639 if (old_block_type == MIGRATE_MOVABLE)
2640 alike_pages = pageblock_nr_pages
2641 - (free_pages + movable_pages);
2646 /* moving whole block can fail due to zone boundary conditions */
2651 * If a sufficient number of pages in the block are either free or of
2652 * comparable migratability as our allocation, claim the whole block.
2654 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2655 page_group_by_mobility_disabled)
2656 set_pageblock_migratetype(page, start_type);
2661 move_to_free_list(page, zone, current_order, start_type);
2665 * Check whether there is a suitable fallback freepage with requested order.
2666 * If only_stealable is true, this function returns fallback_mt only if
2667 * we can steal other freepages all together. This would help to reduce
2668 * fragmentation due to mixed migratetype pages in one pageblock.
2670 int find_suitable_fallback(struct free_area *area, unsigned int order,
2671 int migratetype, bool only_stealable, bool *can_steal)
2676 if (area->nr_free == 0)
2681 fallback_mt = fallbacks[migratetype][i];
2682 if (fallback_mt == MIGRATE_TYPES)
2685 if (free_area_empty(area, fallback_mt))
2688 if (can_steal_fallback(order, migratetype))
2691 if (!only_stealable)
2702 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2703 * there are no empty page blocks that contain a page with a suitable order
2705 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2706 unsigned int alloc_order)
2709 unsigned long max_managed, flags;
2712 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2713 * Check is race-prone but harmless.
2715 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2716 if (zone->nr_reserved_highatomic >= max_managed)
2719 spin_lock_irqsave(&zone->lock, flags);
2721 /* Recheck the nr_reserved_highatomic limit under the lock */
2722 if (zone->nr_reserved_highatomic >= max_managed)
2726 mt = get_pageblock_migratetype(page);
2727 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2728 && !is_migrate_cma(mt)) {
2729 zone->nr_reserved_highatomic += pageblock_nr_pages;
2730 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2731 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2735 spin_unlock_irqrestore(&zone->lock, flags);
2739 * Used when an allocation is about to fail under memory pressure. This
2740 * potentially hurts the reliability of high-order allocations when under
2741 * intense memory pressure but failed atomic allocations should be easier
2742 * to recover from than an OOM.
2744 * If @force is true, try to unreserve a pageblock even though highatomic
2745 * pageblock is exhausted.
2747 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2750 struct zonelist *zonelist = ac->zonelist;
2751 unsigned long flags;
2758 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2761 * Preserve at least one pageblock unless memory pressure
2764 if (!force && zone->nr_reserved_highatomic <=
2768 spin_lock_irqsave(&zone->lock, flags);
2769 for (order = 0; order < MAX_ORDER; order++) {
2770 struct free_area *area = &(zone->free_area[order]);
2772 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2777 * In page freeing path, migratetype change is racy so
2778 * we can counter several free pages in a pageblock
2779 * in this loop althoug we changed the pageblock type
2780 * from highatomic to ac->migratetype. So we should
2781 * adjust the count once.
2783 if (is_migrate_highatomic_page(page)) {
2785 * It should never happen but changes to
2786 * locking could inadvertently allow a per-cpu
2787 * drain to add pages to MIGRATE_HIGHATOMIC
2788 * while unreserving so be safe and watch for
2791 zone->nr_reserved_highatomic -= min(
2793 zone->nr_reserved_highatomic);
2797 * Convert to ac->migratetype and avoid the normal
2798 * pageblock stealing heuristics. Minimally, the caller
2799 * is doing the work and needs the pages. More
2800 * importantly, if the block was always converted to
2801 * MIGRATE_UNMOVABLE or another type then the number
2802 * of pageblocks that cannot be completely freed
2805 set_pageblock_migratetype(page, ac->migratetype);
2806 ret = move_freepages_block(zone, page, ac->migratetype,
2809 spin_unlock_irqrestore(&zone->lock, flags);
2813 spin_unlock_irqrestore(&zone->lock, flags);
2820 * Try finding a free buddy page on the fallback list and put it on the free
2821 * list of requested migratetype, possibly along with other pages from the same
2822 * block, depending on fragmentation avoidance heuristics. Returns true if
2823 * fallback was found so that __rmqueue_smallest() can grab it.
2825 * The use of signed ints for order and current_order is a deliberate
2826 * deviation from the rest of this file, to make the for loop
2827 * condition simpler.
2829 static __always_inline bool
2830 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2831 unsigned int alloc_flags)
2833 struct free_area *area;
2835 int min_order = order;
2841 * Do not steal pages from freelists belonging to other pageblocks
2842 * i.e. orders < pageblock_order. If there are no local zones free,
2843 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2845 if (alloc_flags & ALLOC_NOFRAGMENT)
2846 min_order = pageblock_order;
2849 * Find the largest available free page in the other list. This roughly
2850 * approximates finding the pageblock with the most free pages, which
2851 * would be too costly to do exactly.
2853 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2855 area = &(zone->free_area[current_order]);
2856 fallback_mt = find_suitable_fallback(area, current_order,
2857 start_migratetype, false, &can_steal);
2858 if (fallback_mt == -1)
2862 * We cannot steal all free pages from the pageblock and the
2863 * requested migratetype is movable. In that case it's better to
2864 * steal and split the smallest available page instead of the
2865 * largest available page, because even if the next movable
2866 * allocation falls back into a different pageblock than this
2867 * one, it won't cause permanent fragmentation.
2869 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2870 && current_order > order)
2879 for (current_order = order; current_order < MAX_ORDER;
2881 area = &(zone->free_area[current_order]);
2882 fallback_mt = find_suitable_fallback(area, current_order,
2883 start_migratetype, false, &can_steal);
2884 if (fallback_mt != -1)
2889 * This should not happen - we already found a suitable fallback
2890 * when looking for the largest page.
2892 VM_BUG_ON(current_order == MAX_ORDER);
2895 page = get_page_from_free_area(area, fallback_mt);
2897 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2900 trace_mm_page_alloc_extfrag(page, order, current_order,
2901 start_migratetype, fallback_mt);
2908 * Do the hard work of removing an element from the buddy allocator.
2909 * Call me with the zone->lock already held.
2911 static __always_inline struct page *
2912 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2913 unsigned int alloc_flags)
2917 if (IS_ENABLED(CONFIG_CMA)) {
2919 * Balance movable allocations between regular and CMA areas by
2920 * allocating from CMA when over half of the zone's free memory
2921 * is in the CMA area.
2923 if (alloc_flags & ALLOC_CMA &&
2924 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2925 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2926 page = __rmqueue_cma_fallback(zone, order);
2932 page = __rmqueue_smallest(zone, order, migratetype);
2933 if (unlikely(!page)) {
2934 if (alloc_flags & ALLOC_CMA)
2935 page = __rmqueue_cma_fallback(zone, order);
2937 if (!page && __rmqueue_fallback(zone, order, migratetype,
2943 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2948 * Obtain a specified number of elements from the buddy allocator, all under
2949 * a single hold of the lock, for efficiency. Add them to the supplied list.
2950 * Returns the number of new pages which were placed at *list.
2952 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2953 unsigned long count, struct list_head *list,
2954 int migratetype, unsigned int alloc_flags)
2956 int i, allocated = 0;
2958 spin_lock(&zone->lock);
2959 for (i = 0; i < count; ++i) {
2960 struct page *page = __rmqueue(zone, order, migratetype,
2962 if (unlikely(page == NULL))
2965 if (unlikely(check_pcp_refill(page)))
2969 * Split buddy pages returned by expand() are received here in
2970 * physical page order. The page is added to the tail of
2971 * caller's list. From the callers perspective, the linked list
2972 * is ordered by page number under some conditions. This is
2973 * useful for IO devices that can forward direction from the
2974 * head, thus also in the physical page order. This is useful
2975 * for IO devices that can merge IO requests if the physical
2976 * pages are ordered properly.
2978 list_add_tail(&page->lru, list);
2980 if (is_migrate_cma(get_pcppage_migratetype(page)))
2981 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2986 * i pages were removed from the buddy list even if some leak due
2987 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2988 * on i. Do not confuse with 'allocated' which is the number of
2989 * pages added to the pcp list.
2991 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2992 spin_unlock(&zone->lock);
2998 * Called from the vmstat counter updater to drain pagesets of this
2999 * currently executing processor on remote nodes after they have
3002 * Note that this function must be called with the thread pinned to
3003 * a single processor.
3005 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3007 unsigned long flags;
3008 int to_drain, batch;
3010 local_irq_save(flags);
3011 batch = READ_ONCE(pcp->batch);
3012 to_drain = min(pcp->count, batch);
3014 free_pcppages_bulk(zone, to_drain, pcp);
3015 local_irq_restore(flags);
3020 * Drain pcplists of the indicated processor and zone.
3022 * The processor must either be the current processor and the
3023 * thread pinned to the current processor or a processor that
3026 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3028 unsigned long flags;
3029 struct per_cpu_pageset *pset;
3030 struct per_cpu_pages *pcp;
3032 local_irq_save(flags);
3033 pset = per_cpu_ptr(zone->pageset, cpu);
3037 free_pcppages_bulk(zone, pcp->count, pcp);
3038 local_irq_restore(flags);
3042 * Drain pcplists of all zones on the indicated processor.
3044 * The processor must either be the current processor and the
3045 * thread pinned to the current processor or a processor that
3048 static void drain_pages(unsigned int cpu)
3052 for_each_populated_zone(zone) {
3053 drain_pages_zone(cpu, zone);
3058 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3060 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3061 * the single zone's pages.
3063 void drain_local_pages(struct zone *zone)
3065 int cpu = smp_processor_id();
3068 drain_pages_zone(cpu, zone);
3073 static void drain_local_pages_wq(struct work_struct *work)
3075 struct pcpu_drain *drain;
3077 drain = container_of(work, struct pcpu_drain, work);
3080 * drain_all_pages doesn't use proper cpu hotplug protection so
3081 * we can race with cpu offline when the WQ can move this from
3082 * a cpu pinned worker to an unbound one. We can operate on a different
3083 * cpu which is allright but we also have to make sure to not move to
3087 drain_local_pages(drain->zone);
3092 * The implementation of drain_all_pages(), exposing an extra parameter to
3093 * drain on all cpus.
3095 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3096 * not empty. The check for non-emptiness can however race with a free to
3097 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3098 * that need the guarantee that every CPU has drained can disable the
3099 * optimizing racy check.
3101 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3106 * Allocate in the BSS so we wont require allocation in
3107 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3109 static cpumask_t cpus_with_pcps;
3112 * Make sure nobody triggers this path before mm_percpu_wq is fully
3115 if (WARN_ON_ONCE(!mm_percpu_wq))
3119 * Do not drain if one is already in progress unless it's specific to
3120 * a zone. Such callers are primarily CMA and memory hotplug and need
3121 * the drain to be complete when the call returns.
3123 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3126 mutex_lock(&pcpu_drain_mutex);
3130 * We don't care about racing with CPU hotplug event
3131 * as offline notification will cause the notified
3132 * cpu to drain that CPU pcps and on_each_cpu_mask
3133 * disables preemption as part of its processing
3135 for_each_online_cpu(cpu) {
3136 struct per_cpu_pageset *pcp;
3138 bool has_pcps = false;
3140 if (force_all_cpus) {
3142 * The pcp.count check is racy, some callers need a
3143 * guarantee that no cpu is missed.
3147 pcp = per_cpu_ptr(zone->pageset, cpu);
3151 for_each_populated_zone(z) {
3152 pcp = per_cpu_ptr(z->pageset, cpu);
3153 if (pcp->pcp.count) {
3161 cpumask_set_cpu(cpu, &cpus_with_pcps);
3163 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3166 for_each_cpu(cpu, &cpus_with_pcps) {
3167 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3170 INIT_WORK(&drain->work, drain_local_pages_wq);
3171 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3173 for_each_cpu(cpu, &cpus_with_pcps)
3174 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3176 mutex_unlock(&pcpu_drain_mutex);
3180 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3182 * When zone parameter is non-NULL, spill just the single zone's pages.
3184 * Note that this can be extremely slow as the draining happens in a workqueue.
3186 void drain_all_pages(struct zone *zone)
3188 __drain_all_pages(zone, false);
3191 #ifdef CONFIG_HIBERNATION
3194 * Touch the watchdog for every WD_PAGE_COUNT pages.
3196 #define WD_PAGE_COUNT (128*1024)
3198 void mark_free_pages(struct zone *zone)
3200 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3201 unsigned long flags;
3202 unsigned int order, t;
3205 if (zone_is_empty(zone))
3208 spin_lock_irqsave(&zone->lock, flags);
3210 max_zone_pfn = zone_end_pfn(zone);
3211 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3212 if (pfn_valid(pfn)) {
3213 page = pfn_to_page(pfn);
3215 if (!--page_count) {
3216 touch_nmi_watchdog();
3217 page_count = WD_PAGE_COUNT;
3220 if (page_zone(page) != zone)
3223 if (!swsusp_page_is_forbidden(page))
3224 swsusp_unset_page_free(page);
3227 for_each_migratetype_order(order, t) {
3228 list_for_each_entry(page,
3229 &zone->free_area[order].free_list[t], lru) {
3232 pfn = page_to_pfn(page);
3233 for (i = 0; i < (1UL << order); i++) {
3234 if (!--page_count) {
3235 touch_nmi_watchdog();
3236 page_count = WD_PAGE_COUNT;
3238 swsusp_set_page_free(pfn_to_page(pfn + i));
3242 spin_unlock_irqrestore(&zone->lock, flags);
3244 #endif /* CONFIG_PM */
3246 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3250 if (!free_pcp_prepare(page))
3253 migratetype = get_pfnblock_migratetype(page, pfn);
3254 set_pcppage_migratetype(page, migratetype);
3258 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3260 struct zone *zone = page_zone(page);
3261 struct per_cpu_pages *pcp;
3264 migratetype = get_pcppage_migratetype(page);
3265 __count_vm_event(PGFREE);
3268 * We only track unmovable, reclaimable and movable on pcp lists.
3269 * Free ISOLATE pages back to the allocator because they are being
3270 * offlined but treat HIGHATOMIC as movable pages so we can get those
3271 * areas back if necessary. Otherwise, we may have to free
3272 * excessively into the page allocator
3274 if (migratetype >= MIGRATE_PCPTYPES) {
3275 if (unlikely(is_migrate_isolate(migratetype))) {
3276 free_one_page(zone, page, pfn, 0, migratetype,
3280 migratetype = MIGRATE_MOVABLE;
3283 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3284 list_add(&page->lru, &pcp->lists[migratetype]);
3286 if (pcp->count >= READ_ONCE(pcp->high))
3287 free_pcppages_bulk(zone, READ_ONCE(pcp->batch), pcp);
3291 * Free a 0-order page
3293 void free_unref_page(struct page *page)
3295 unsigned long flags;
3296 unsigned long pfn = page_to_pfn(page);
3298 if (!free_unref_page_prepare(page, pfn))
3301 local_irq_save(flags);
3302 free_unref_page_commit(page, pfn);
3303 local_irq_restore(flags);
3307 * Free a list of 0-order pages
3309 void free_unref_page_list(struct list_head *list)
3311 struct page *page, *next;
3312 unsigned long flags, pfn;
3313 int batch_count = 0;
3315 /* Prepare pages for freeing */
3316 list_for_each_entry_safe(page, next, list, lru) {
3317 pfn = page_to_pfn(page);
3318 if (!free_unref_page_prepare(page, pfn))
3319 list_del(&page->lru);
3320 set_page_private(page, pfn);
3323 local_irq_save(flags);
3324 list_for_each_entry_safe(page, next, list, lru) {
3325 unsigned long pfn = page_private(page);
3327 set_page_private(page, 0);
3328 trace_mm_page_free_batched(page);
3329 free_unref_page_commit(page, pfn);
3332 * Guard against excessive IRQ disabled times when we get
3333 * a large list of pages to free.
3335 if (++batch_count == SWAP_CLUSTER_MAX) {
3336 local_irq_restore(flags);
3338 local_irq_save(flags);
3341 local_irq_restore(flags);
3345 * split_page takes a non-compound higher-order page, and splits it into
3346 * n (1<<order) sub-pages: page[0..n]
3347 * Each sub-page must be freed individually.
3349 * Note: this is probably too low level an operation for use in drivers.
3350 * Please consult with lkml before using this in your driver.
3352 void split_page(struct page *page, unsigned int order)
3356 VM_BUG_ON_PAGE(PageCompound(page), page);
3357 VM_BUG_ON_PAGE(!page_count(page), page);
3359 for (i = 1; i < (1 << order); i++)
3360 set_page_refcounted(page + i);
3361 split_page_owner(page, 1 << order);
3362 split_page_memcg(page, 1 << order);
3364 EXPORT_SYMBOL_GPL(split_page);
3366 int __isolate_free_page(struct page *page, unsigned int order)
3368 unsigned long watermark;
3372 BUG_ON(!PageBuddy(page));
3374 zone = page_zone(page);
3375 mt = get_pageblock_migratetype(page);
3377 if (!is_migrate_isolate(mt)) {
3379 * Obey watermarks as if the page was being allocated. We can
3380 * emulate a high-order watermark check with a raised order-0
3381 * watermark, because we already know our high-order page
3384 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3385 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3388 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3391 /* Remove page from free list */
3393 del_page_from_free_list(page, zone, order);
3396 * Set the pageblock if the isolated page is at least half of a
3399 if (order >= pageblock_order - 1) {
3400 struct page *endpage = page + (1 << order) - 1;
3401 for (; page < endpage; page += pageblock_nr_pages) {
3402 int mt = get_pageblock_migratetype(page);
3403 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3404 && !is_migrate_highatomic(mt))
3405 set_pageblock_migratetype(page,
3411 return 1UL << order;
3415 * __putback_isolated_page - Return a now-isolated page back where we got it
3416 * @page: Page that was isolated
3417 * @order: Order of the isolated page
3418 * @mt: The page's pageblock's migratetype
3420 * This function is meant to return a page pulled from the free lists via
3421 * __isolate_free_page back to the free lists they were pulled from.
3423 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3425 struct zone *zone = page_zone(page);
3427 /* zone lock should be held when this function is called */
3428 lockdep_assert_held(&zone->lock);
3430 /* Return isolated page to tail of freelist. */
3431 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3432 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3436 * Update NUMA hit/miss statistics
3438 * Must be called with interrupts disabled.
3440 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3443 enum numa_stat_item local_stat = NUMA_LOCAL;
3445 /* skip numa counters update if numa stats is disabled */
3446 if (!static_branch_likely(&vm_numa_stat_key))
3449 if (zone_to_nid(z) != numa_node_id())
3450 local_stat = NUMA_OTHER;
3452 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3453 __inc_numa_state(z, NUMA_HIT);
3455 __inc_numa_state(z, NUMA_MISS);
3456 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3458 __inc_numa_state(z, local_stat);
3462 /* Remove page from the per-cpu list, caller must protect the list */
3464 struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3465 unsigned int alloc_flags,
3466 struct per_cpu_pages *pcp,
3467 struct list_head *list)
3472 if (list_empty(list)) {
3473 pcp->count += rmqueue_bulk(zone, 0,
3474 READ_ONCE(pcp->batch), list,
3475 migratetype, alloc_flags);
3476 if (unlikely(list_empty(list)))
3480 page = list_first_entry(list, struct page, lru);
3481 list_del(&page->lru);
3483 } while (check_new_pcp(page));
3488 /* Lock and remove page from the per-cpu list */
3489 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3490 struct zone *zone, gfp_t gfp_flags,
3491 int migratetype, unsigned int alloc_flags)
3493 struct per_cpu_pages *pcp;
3494 struct list_head *list;
3496 unsigned long flags;
3498 local_irq_save(flags);
3499 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3500 list = &pcp->lists[migratetype];
3501 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3503 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3504 zone_statistics(preferred_zone, zone);
3506 local_irq_restore(flags);
3511 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3514 struct page *rmqueue(struct zone *preferred_zone,
3515 struct zone *zone, unsigned int order,
3516 gfp_t gfp_flags, unsigned int alloc_flags,
3519 unsigned long flags;
3522 if (likely(order == 0)) {
3524 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3525 * we need to skip it when CMA area isn't allowed.
3527 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3528 migratetype != MIGRATE_MOVABLE) {
3529 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3530 migratetype, alloc_flags);
3536 * We most definitely don't want callers attempting to
3537 * allocate greater than order-1 page units with __GFP_NOFAIL.
3539 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3540 spin_lock_irqsave(&zone->lock, flags);
3545 * order-0 request can reach here when the pcplist is skipped
3546 * due to non-CMA allocation context. HIGHATOMIC area is
3547 * reserved for high-order atomic allocation, so order-0
3548 * request should skip it.
3550 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3551 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3553 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3556 page = __rmqueue(zone, order, migratetype, alloc_flags);
3557 } while (page && check_new_pages(page, order));
3558 spin_unlock(&zone->lock);
3561 __mod_zone_freepage_state(zone, -(1 << order),
3562 get_pcppage_migratetype(page));
3564 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3565 zone_statistics(preferred_zone, zone);
3566 local_irq_restore(flags);
3569 /* Separate test+clear to avoid unnecessary atomics */
3570 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3571 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3572 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3575 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3579 local_irq_restore(flags);
3583 #ifdef CONFIG_FAIL_PAGE_ALLOC
3586 struct fault_attr attr;
3588 bool ignore_gfp_highmem;
3589 bool ignore_gfp_reclaim;
3591 } fail_page_alloc = {
3592 .attr = FAULT_ATTR_INITIALIZER,
3593 .ignore_gfp_reclaim = true,
3594 .ignore_gfp_highmem = true,
3598 static int __init setup_fail_page_alloc(char *str)
3600 return setup_fault_attr(&fail_page_alloc.attr, str);
3602 __setup("fail_page_alloc=", setup_fail_page_alloc);
3604 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3606 if (order < fail_page_alloc.min_order)
3608 if (gfp_mask & __GFP_NOFAIL)
3610 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3612 if (fail_page_alloc.ignore_gfp_reclaim &&
3613 (gfp_mask & __GFP_DIRECT_RECLAIM))
3616 return should_fail(&fail_page_alloc.attr, 1 << order);
3619 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3621 static int __init fail_page_alloc_debugfs(void)
3623 umode_t mode = S_IFREG | 0600;
3626 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3627 &fail_page_alloc.attr);
3629 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3630 &fail_page_alloc.ignore_gfp_reclaim);
3631 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3632 &fail_page_alloc.ignore_gfp_highmem);
3633 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3638 late_initcall(fail_page_alloc_debugfs);
3640 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3642 #else /* CONFIG_FAIL_PAGE_ALLOC */
3644 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3649 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3651 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3653 return __should_fail_alloc_page(gfp_mask, order);
3655 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3657 static inline long __zone_watermark_unusable_free(struct zone *z,
3658 unsigned int order, unsigned int alloc_flags)
3660 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3661 long unusable_free = (1 << order) - 1;
3664 * If the caller does not have rights to ALLOC_HARDER then subtract
3665 * the high-atomic reserves. This will over-estimate the size of the
3666 * atomic reserve but it avoids a search.
3668 if (likely(!alloc_harder))
3669 unusable_free += z->nr_reserved_highatomic;
3672 /* If allocation can't use CMA areas don't use free CMA pages */
3673 if (!(alloc_flags & ALLOC_CMA))
3674 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3677 return unusable_free;
3681 * Return true if free base pages are above 'mark'. For high-order checks it
3682 * will return true of the order-0 watermark is reached and there is at least
3683 * one free page of a suitable size. Checking now avoids taking the zone lock
3684 * to check in the allocation paths if no pages are free.
3686 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3687 int highest_zoneidx, unsigned int alloc_flags,
3692 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3694 /* free_pages may go negative - that's OK */
3695 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3697 if (alloc_flags & ALLOC_HIGH)
3700 if (unlikely(alloc_harder)) {
3702 * OOM victims can try even harder than normal ALLOC_HARDER
3703 * users on the grounds that it's definitely going to be in
3704 * the exit path shortly and free memory. Any allocation it
3705 * makes during the free path will be small and short-lived.
3707 if (alloc_flags & ALLOC_OOM)
3714 * Check watermarks for an order-0 allocation request. If these
3715 * are not met, then a high-order request also cannot go ahead
3716 * even if a suitable page happened to be free.
3718 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3721 /* If this is an order-0 request then the watermark is fine */
3725 /* For a high-order request, check at least one suitable page is free */
3726 for (o = order; o < MAX_ORDER; o++) {
3727 struct free_area *area = &z->free_area[o];
3733 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3734 if (!free_area_empty(area, mt))
3739 if ((alloc_flags & ALLOC_CMA) &&
3740 !free_area_empty(area, MIGRATE_CMA)) {
3744 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3750 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3751 int highest_zoneidx, unsigned int alloc_flags)
3753 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3754 zone_page_state(z, NR_FREE_PAGES));
3757 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3758 unsigned long mark, int highest_zoneidx,
3759 unsigned int alloc_flags, gfp_t gfp_mask)
3763 free_pages = zone_page_state(z, NR_FREE_PAGES);
3766 * Fast check for order-0 only. If this fails then the reserves
3767 * need to be calculated.
3772 fast_free = free_pages;
3773 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3774 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3778 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3782 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3783 * when checking the min watermark. The min watermark is the
3784 * point where boosting is ignored so that kswapd is woken up
3785 * when below the low watermark.
3787 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3788 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3789 mark = z->_watermark[WMARK_MIN];
3790 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3791 alloc_flags, free_pages);
3797 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3798 unsigned long mark, int highest_zoneidx)
3800 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3802 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3803 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3805 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3810 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3812 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3813 node_reclaim_distance;
3815 #else /* CONFIG_NUMA */
3816 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3820 #endif /* CONFIG_NUMA */
3823 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3824 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3825 * premature use of a lower zone may cause lowmem pressure problems that
3826 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3827 * probably too small. It only makes sense to spread allocations to avoid
3828 * fragmentation between the Normal and DMA32 zones.
3830 static inline unsigned int
3831 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3833 unsigned int alloc_flags;
3836 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3839 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3841 #ifdef CONFIG_ZONE_DMA32
3845 if (zone_idx(zone) != ZONE_NORMAL)
3849 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3850 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3851 * on UMA that if Normal is populated then so is DMA32.
3853 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3854 if (nr_online_nodes > 1 && !populated_zone(--zone))
3857 alloc_flags |= ALLOC_NOFRAGMENT;
3858 #endif /* CONFIG_ZONE_DMA32 */
3862 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3863 unsigned int alloc_flags)
3866 unsigned int pflags = current->flags;
3868 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3869 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3870 alloc_flags |= ALLOC_CMA;
3877 * get_page_from_freelist goes through the zonelist trying to allocate
3880 static struct page *
3881 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3882 const struct alloc_context *ac)
3886 struct pglist_data *last_pgdat_dirty_limit = NULL;
3891 * Scan zonelist, looking for a zone with enough free.
3892 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3894 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3895 z = ac->preferred_zoneref;
3896 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3901 if (cpusets_enabled() &&
3902 (alloc_flags & ALLOC_CPUSET) &&
3903 !__cpuset_zone_allowed(zone, gfp_mask))
3906 * When allocating a page cache page for writing, we
3907 * want to get it from a node that is within its dirty
3908 * limit, such that no single node holds more than its
3909 * proportional share of globally allowed dirty pages.
3910 * The dirty limits take into account the node's
3911 * lowmem reserves and high watermark so that kswapd
3912 * should be able to balance it without having to
3913 * write pages from its LRU list.
3915 * XXX: For now, allow allocations to potentially
3916 * exceed the per-node dirty limit in the slowpath
3917 * (spread_dirty_pages unset) before going into reclaim,
3918 * which is important when on a NUMA setup the allowed
3919 * nodes are together not big enough to reach the
3920 * global limit. The proper fix for these situations
3921 * will require awareness of nodes in the
3922 * dirty-throttling and the flusher threads.
3924 if (ac->spread_dirty_pages) {
3925 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3928 if (!node_dirty_ok(zone->zone_pgdat)) {
3929 last_pgdat_dirty_limit = zone->zone_pgdat;
3934 if (no_fallback && nr_online_nodes > 1 &&
3935 zone != ac->preferred_zoneref->zone) {
3939 * If moving to a remote node, retry but allow
3940 * fragmenting fallbacks. Locality is more important
3941 * than fragmentation avoidance.
3943 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3944 if (zone_to_nid(zone) != local_nid) {
3945 alloc_flags &= ~ALLOC_NOFRAGMENT;
3950 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3951 if (!zone_watermark_fast(zone, order, mark,
3952 ac->highest_zoneidx, alloc_flags,
3956 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3958 * Watermark failed for this zone, but see if we can
3959 * grow this zone if it contains deferred pages.
3961 if (static_branch_unlikely(&deferred_pages)) {
3962 if (_deferred_grow_zone(zone, order))
3966 /* Checked here to keep the fast path fast */
3967 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3968 if (alloc_flags & ALLOC_NO_WATERMARKS)
3971 if (node_reclaim_mode == 0 ||
3972 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3975 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3977 case NODE_RECLAIM_NOSCAN:
3980 case NODE_RECLAIM_FULL:
3981 /* scanned but unreclaimable */
3984 /* did we reclaim enough */
3985 if (zone_watermark_ok(zone, order, mark,
3986 ac->highest_zoneidx, alloc_flags))
3994 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3995 gfp_mask, alloc_flags, ac->migratetype);
3997 prep_new_page(page, order, gfp_mask, alloc_flags);
4000 * If this is a high-order atomic allocation then check
4001 * if the pageblock should be reserved for the future
4003 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4004 reserve_highatomic_pageblock(page, zone, order);
4008 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4009 /* Try again if zone has deferred pages */
4010 if (static_branch_unlikely(&deferred_pages)) {
4011 if (_deferred_grow_zone(zone, order))
4019 * It's possible on a UMA machine to get through all zones that are
4020 * fragmented. If avoiding fragmentation, reset and try again.
4023 alloc_flags &= ~ALLOC_NOFRAGMENT;
4030 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4032 unsigned int filter = SHOW_MEM_FILTER_NODES;
4035 * This documents exceptions given to allocations in certain
4036 * contexts that are allowed to allocate outside current's set
4039 if (!(gfp_mask & __GFP_NOMEMALLOC))
4040 if (tsk_is_oom_victim(current) ||
4041 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4042 filter &= ~SHOW_MEM_FILTER_NODES;
4043 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4044 filter &= ~SHOW_MEM_FILTER_NODES;
4046 show_mem(filter, nodemask);
4049 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4051 struct va_format vaf;
4053 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4055 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4058 va_start(args, fmt);
4061 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4062 current->comm, &vaf, gfp_mask, &gfp_mask,
4063 nodemask_pr_args(nodemask));
4066 cpuset_print_current_mems_allowed();
4069 warn_alloc_show_mem(gfp_mask, nodemask);
4072 static inline struct page *
4073 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4074 unsigned int alloc_flags,
4075 const struct alloc_context *ac)
4079 page = get_page_from_freelist(gfp_mask, order,
4080 alloc_flags|ALLOC_CPUSET, ac);
4082 * fallback to ignore cpuset restriction if our nodes
4086 page = get_page_from_freelist(gfp_mask, order,
4092 static inline struct page *
4093 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4094 const struct alloc_context *ac, unsigned long *did_some_progress)
4096 struct oom_control oc = {
4097 .zonelist = ac->zonelist,
4098 .nodemask = ac->nodemask,
4100 .gfp_mask = gfp_mask,
4105 *did_some_progress = 0;
4108 * Acquire the oom lock. If that fails, somebody else is
4109 * making progress for us.
4111 if (!mutex_trylock(&oom_lock)) {
4112 *did_some_progress = 1;
4113 schedule_timeout_uninterruptible(1);
4118 * Go through the zonelist yet one more time, keep very high watermark
4119 * here, this is only to catch a parallel oom killing, we must fail if
4120 * we're still under heavy pressure. But make sure that this reclaim
4121 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4122 * allocation which will never fail due to oom_lock already held.
4124 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4125 ~__GFP_DIRECT_RECLAIM, order,
4126 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4130 /* Coredumps can quickly deplete all memory reserves */
4131 if (current->flags & PF_DUMPCORE)
4133 /* The OOM killer will not help higher order allocs */
4134 if (order > PAGE_ALLOC_COSTLY_ORDER)
4137 * We have already exhausted all our reclaim opportunities without any
4138 * success so it is time to admit defeat. We will skip the OOM killer
4139 * because it is very likely that the caller has a more reasonable
4140 * fallback than shooting a random task.
4142 * The OOM killer may not free memory on a specific node.
4144 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4146 /* The OOM killer does not needlessly kill tasks for lowmem */
4147 if (ac->highest_zoneidx < ZONE_NORMAL)
4149 if (pm_suspended_storage())
4152 * XXX: GFP_NOFS allocations should rather fail than rely on
4153 * other request to make a forward progress.
4154 * We are in an unfortunate situation where out_of_memory cannot
4155 * do much for this context but let's try it to at least get
4156 * access to memory reserved if the current task is killed (see
4157 * out_of_memory). Once filesystems are ready to handle allocation
4158 * failures more gracefully we should just bail out here.
4161 /* Exhausted what can be done so it's blame time */
4162 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4163 *did_some_progress = 1;
4166 * Help non-failing allocations by giving them access to memory
4169 if (gfp_mask & __GFP_NOFAIL)
4170 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4171 ALLOC_NO_WATERMARKS, ac);
4174 mutex_unlock(&oom_lock);
4179 * Maximum number of compaction retries wit a progress before OOM
4180 * killer is consider as the only way to move forward.
4182 #define MAX_COMPACT_RETRIES 16
4184 #ifdef CONFIG_COMPACTION
4185 /* Try memory compaction for high-order allocations before reclaim */
4186 static struct page *
4187 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4188 unsigned int alloc_flags, const struct alloc_context *ac,
4189 enum compact_priority prio, enum compact_result *compact_result)
4191 struct page *page = NULL;
4192 unsigned long pflags;
4193 unsigned int noreclaim_flag;
4198 psi_memstall_enter(&pflags);
4199 noreclaim_flag = memalloc_noreclaim_save();
4201 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4204 memalloc_noreclaim_restore(noreclaim_flag);
4205 psi_memstall_leave(&pflags);
4208 * At least in one zone compaction wasn't deferred or skipped, so let's
4209 * count a compaction stall
4211 count_vm_event(COMPACTSTALL);
4213 /* Prep a captured page if available */
4215 prep_new_page(page, order, gfp_mask, alloc_flags);
4217 /* Try get a page from the freelist if available */
4219 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4222 struct zone *zone = page_zone(page);
4224 zone->compact_blockskip_flush = false;
4225 compaction_defer_reset(zone, order, true);
4226 count_vm_event(COMPACTSUCCESS);
4231 * It's bad if compaction run occurs and fails. The most likely reason
4232 * is that pages exist, but not enough to satisfy watermarks.
4234 count_vm_event(COMPACTFAIL);
4242 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4243 enum compact_result compact_result,
4244 enum compact_priority *compact_priority,
4245 int *compaction_retries)
4247 int max_retries = MAX_COMPACT_RETRIES;
4250 int retries = *compaction_retries;
4251 enum compact_priority priority = *compact_priority;
4256 if (compaction_made_progress(compact_result))
4257 (*compaction_retries)++;
4260 * compaction considers all the zone as desperately out of memory
4261 * so it doesn't really make much sense to retry except when the
4262 * failure could be caused by insufficient priority
4264 if (compaction_failed(compact_result))
4265 goto check_priority;
4268 * compaction was skipped because there are not enough order-0 pages
4269 * to work with, so we retry only if it looks like reclaim can help.
4271 if (compaction_needs_reclaim(compact_result)) {
4272 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4277 * make sure the compaction wasn't deferred or didn't bail out early
4278 * due to locks contention before we declare that we should give up.
4279 * But the next retry should use a higher priority if allowed, so
4280 * we don't just keep bailing out endlessly.
4282 if (compaction_withdrawn(compact_result)) {
4283 goto check_priority;
4287 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4288 * costly ones because they are de facto nofail and invoke OOM
4289 * killer to move on while costly can fail and users are ready
4290 * to cope with that. 1/4 retries is rather arbitrary but we
4291 * would need much more detailed feedback from compaction to
4292 * make a better decision.
4294 if (order > PAGE_ALLOC_COSTLY_ORDER)
4296 if (*compaction_retries <= max_retries) {
4302 * Make sure there are attempts at the highest priority if we exhausted
4303 * all retries or failed at the lower priorities.
4306 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4307 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4309 if (*compact_priority > min_priority) {
4310 (*compact_priority)--;
4311 *compaction_retries = 0;
4315 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4319 static inline struct page *
4320 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4321 unsigned int alloc_flags, const struct alloc_context *ac,
4322 enum compact_priority prio, enum compact_result *compact_result)
4324 *compact_result = COMPACT_SKIPPED;
4329 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4330 enum compact_result compact_result,
4331 enum compact_priority *compact_priority,
4332 int *compaction_retries)
4337 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4341 * There are setups with compaction disabled which would prefer to loop
4342 * inside the allocator rather than hit the oom killer prematurely.
4343 * Let's give them a good hope and keep retrying while the order-0
4344 * watermarks are OK.
4346 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4347 ac->highest_zoneidx, ac->nodemask) {
4348 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4349 ac->highest_zoneidx, alloc_flags))
4354 #endif /* CONFIG_COMPACTION */
4356 #ifdef CONFIG_LOCKDEP
4357 static struct lockdep_map __fs_reclaim_map =
4358 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4360 static bool __need_reclaim(gfp_t gfp_mask)
4362 /* no reclaim without waiting on it */
4363 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4366 /* this guy won't enter reclaim */
4367 if (current->flags & PF_MEMALLOC)
4370 if (gfp_mask & __GFP_NOLOCKDEP)
4376 void __fs_reclaim_acquire(void)
4378 lock_map_acquire(&__fs_reclaim_map);
4381 void __fs_reclaim_release(void)
4383 lock_map_release(&__fs_reclaim_map);
4386 void fs_reclaim_acquire(gfp_t gfp_mask)
4388 gfp_mask = current_gfp_context(gfp_mask);
4390 if (__need_reclaim(gfp_mask)) {
4391 if (gfp_mask & __GFP_FS)
4392 __fs_reclaim_acquire();
4394 #ifdef CONFIG_MMU_NOTIFIER
4395 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4396 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4401 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4403 void fs_reclaim_release(gfp_t gfp_mask)
4405 gfp_mask = current_gfp_context(gfp_mask);
4407 if (__need_reclaim(gfp_mask)) {
4408 if (gfp_mask & __GFP_FS)
4409 __fs_reclaim_release();
4412 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4415 /* Perform direct synchronous page reclaim */
4416 static unsigned long
4417 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4418 const struct alloc_context *ac)
4420 unsigned int noreclaim_flag;
4421 unsigned long pflags, progress;
4425 /* We now go into synchronous reclaim */
4426 cpuset_memory_pressure_bump();
4427 psi_memstall_enter(&pflags);
4428 fs_reclaim_acquire(gfp_mask);
4429 noreclaim_flag = memalloc_noreclaim_save();
4431 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4434 memalloc_noreclaim_restore(noreclaim_flag);
4435 fs_reclaim_release(gfp_mask);
4436 psi_memstall_leave(&pflags);
4443 /* The really slow allocator path where we enter direct reclaim */
4444 static inline struct page *
4445 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4446 unsigned int alloc_flags, const struct alloc_context *ac,
4447 unsigned long *did_some_progress)
4449 struct page *page = NULL;
4450 bool drained = false;
4452 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4453 if (unlikely(!(*did_some_progress)))
4457 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4460 * If an allocation failed after direct reclaim, it could be because
4461 * pages are pinned on the per-cpu lists or in high alloc reserves.
4462 * Shrink them and try again
4464 if (!page && !drained) {
4465 unreserve_highatomic_pageblock(ac, false);
4466 drain_all_pages(NULL);
4474 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4475 const struct alloc_context *ac)
4479 pg_data_t *last_pgdat = NULL;
4480 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4482 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4484 if (last_pgdat != zone->zone_pgdat)
4485 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4486 last_pgdat = zone->zone_pgdat;
4490 static inline unsigned int
4491 gfp_to_alloc_flags(gfp_t gfp_mask)
4493 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4496 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4497 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4498 * to save two branches.
4500 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4501 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4504 * The caller may dip into page reserves a bit more if the caller
4505 * cannot run direct reclaim, or if the caller has realtime scheduling
4506 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4507 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4509 alloc_flags |= (__force int)
4510 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4512 if (gfp_mask & __GFP_ATOMIC) {
4514 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4515 * if it can't schedule.
4517 if (!(gfp_mask & __GFP_NOMEMALLOC))
4518 alloc_flags |= ALLOC_HARDER;
4520 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4521 * comment for __cpuset_node_allowed().
4523 alloc_flags &= ~ALLOC_CPUSET;
4524 } else if (unlikely(rt_task(current)) && !in_interrupt())
4525 alloc_flags |= ALLOC_HARDER;
4527 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4532 static bool oom_reserves_allowed(struct task_struct *tsk)
4534 if (!tsk_is_oom_victim(tsk))
4538 * !MMU doesn't have oom reaper so give access to memory reserves
4539 * only to the thread with TIF_MEMDIE set
4541 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4548 * Distinguish requests which really need access to full memory
4549 * reserves from oom victims which can live with a portion of it
4551 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4553 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4555 if (gfp_mask & __GFP_MEMALLOC)
4556 return ALLOC_NO_WATERMARKS;
4557 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4558 return ALLOC_NO_WATERMARKS;
4559 if (!in_interrupt()) {
4560 if (current->flags & PF_MEMALLOC)
4561 return ALLOC_NO_WATERMARKS;
4562 else if (oom_reserves_allowed(current))
4569 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4571 return !!__gfp_pfmemalloc_flags(gfp_mask);
4575 * Checks whether it makes sense to retry the reclaim to make a forward progress
4576 * for the given allocation request.
4578 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4579 * without success, or when we couldn't even meet the watermark if we
4580 * reclaimed all remaining pages on the LRU lists.
4582 * Returns true if a retry is viable or false to enter the oom path.
4585 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4586 struct alloc_context *ac, int alloc_flags,
4587 bool did_some_progress, int *no_progress_loops)
4594 * Costly allocations might have made a progress but this doesn't mean
4595 * their order will become available due to high fragmentation so
4596 * always increment the no progress counter for them
4598 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4599 *no_progress_loops = 0;
4601 (*no_progress_loops)++;
4604 * Make sure we converge to OOM if we cannot make any progress
4605 * several times in the row.
4607 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4608 /* Before OOM, exhaust highatomic_reserve */
4609 return unreserve_highatomic_pageblock(ac, true);
4613 * Keep reclaiming pages while there is a chance this will lead
4614 * somewhere. If none of the target zones can satisfy our allocation
4615 * request even if all reclaimable pages are considered then we are
4616 * screwed and have to go OOM.
4618 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4619 ac->highest_zoneidx, ac->nodemask) {
4620 unsigned long available;
4621 unsigned long reclaimable;
4622 unsigned long min_wmark = min_wmark_pages(zone);
4625 available = reclaimable = zone_reclaimable_pages(zone);
4626 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4629 * Would the allocation succeed if we reclaimed all
4630 * reclaimable pages?
4632 wmark = __zone_watermark_ok(zone, order, min_wmark,
4633 ac->highest_zoneidx, alloc_flags, available);
4634 trace_reclaim_retry_zone(z, order, reclaimable,
4635 available, min_wmark, *no_progress_loops, wmark);
4638 * If we didn't make any progress and have a lot of
4639 * dirty + writeback pages then we should wait for
4640 * an IO to complete to slow down the reclaim and
4641 * prevent from pre mature OOM
4643 if (!did_some_progress) {
4644 unsigned long write_pending;
4646 write_pending = zone_page_state_snapshot(zone,
4647 NR_ZONE_WRITE_PENDING);
4649 if (2 * write_pending > reclaimable) {
4650 congestion_wait(BLK_RW_ASYNC, HZ/10);
4662 * Memory allocation/reclaim might be called from a WQ context and the
4663 * current implementation of the WQ concurrency control doesn't
4664 * recognize that a particular WQ is congested if the worker thread is
4665 * looping without ever sleeping. Therefore we have to do a short sleep
4666 * here rather than calling cond_resched().
4668 if (current->flags & PF_WQ_WORKER)
4669 schedule_timeout_uninterruptible(1);
4676 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4679 * It's possible that cpuset's mems_allowed and the nodemask from
4680 * mempolicy don't intersect. This should be normally dealt with by
4681 * policy_nodemask(), but it's possible to race with cpuset update in
4682 * such a way the check therein was true, and then it became false
4683 * before we got our cpuset_mems_cookie here.
4684 * This assumes that for all allocations, ac->nodemask can come only
4685 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4686 * when it does not intersect with the cpuset restrictions) or the
4687 * caller can deal with a violated nodemask.
4689 if (cpusets_enabled() && ac->nodemask &&
4690 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4691 ac->nodemask = NULL;
4696 * When updating a task's mems_allowed or mempolicy nodemask, it is
4697 * possible to race with parallel threads in such a way that our
4698 * allocation can fail while the mask is being updated. If we are about
4699 * to fail, check if the cpuset changed during allocation and if so,
4702 if (read_mems_allowed_retry(cpuset_mems_cookie))
4708 static inline struct page *
4709 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4710 struct alloc_context *ac)
4712 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4713 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4714 struct page *page = NULL;
4715 unsigned int alloc_flags;
4716 unsigned long did_some_progress;
4717 enum compact_priority compact_priority;
4718 enum compact_result compact_result;
4719 int compaction_retries;
4720 int no_progress_loops;
4721 unsigned int cpuset_mems_cookie;
4725 * We also sanity check to catch abuse of atomic reserves being used by
4726 * callers that are not in atomic context.
4728 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4729 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4730 gfp_mask &= ~__GFP_ATOMIC;
4733 compaction_retries = 0;
4734 no_progress_loops = 0;
4735 compact_priority = DEF_COMPACT_PRIORITY;
4736 cpuset_mems_cookie = read_mems_allowed_begin();
4739 * The fast path uses conservative alloc_flags to succeed only until
4740 * kswapd needs to be woken up, and to avoid the cost of setting up
4741 * alloc_flags precisely. So we do that now.
4743 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4746 * We need to recalculate the starting point for the zonelist iterator
4747 * because we might have used different nodemask in the fast path, or
4748 * there was a cpuset modification and we are retrying - otherwise we
4749 * could end up iterating over non-eligible zones endlessly.
4751 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4752 ac->highest_zoneidx, ac->nodemask);
4753 if (!ac->preferred_zoneref->zone)
4756 if (alloc_flags & ALLOC_KSWAPD)
4757 wake_all_kswapds(order, gfp_mask, ac);
4760 * The adjusted alloc_flags might result in immediate success, so try
4763 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4768 * For costly allocations, try direct compaction first, as it's likely
4769 * that we have enough base pages and don't need to reclaim. For non-
4770 * movable high-order allocations, do that as well, as compaction will
4771 * try prevent permanent fragmentation by migrating from blocks of the
4773 * Don't try this for allocations that are allowed to ignore
4774 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4776 if (can_direct_reclaim &&
4778 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4779 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4780 page = __alloc_pages_direct_compact(gfp_mask, order,
4782 INIT_COMPACT_PRIORITY,
4788 * Checks for costly allocations with __GFP_NORETRY, which
4789 * includes some THP page fault allocations
4791 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4793 * If allocating entire pageblock(s) and compaction
4794 * failed because all zones are below low watermarks
4795 * or is prohibited because it recently failed at this
4796 * order, fail immediately unless the allocator has
4797 * requested compaction and reclaim retry.
4800 * - potentially very expensive because zones are far
4801 * below their low watermarks or this is part of very
4802 * bursty high order allocations,
4803 * - not guaranteed to help because isolate_freepages()
4804 * may not iterate over freed pages as part of its
4806 * - unlikely to make entire pageblocks free on its
4809 if (compact_result == COMPACT_SKIPPED ||
4810 compact_result == COMPACT_DEFERRED)
4814 * Looks like reclaim/compaction is worth trying, but
4815 * sync compaction could be very expensive, so keep
4816 * using async compaction.
4818 compact_priority = INIT_COMPACT_PRIORITY;
4823 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4824 if (alloc_flags & ALLOC_KSWAPD)
4825 wake_all_kswapds(order, gfp_mask, ac);
4827 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4829 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4832 * Reset the nodemask and zonelist iterators if memory policies can be
4833 * ignored. These allocations are high priority and system rather than
4836 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4837 ac->nodemask = NULL;
4838 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4839 ac->highest_zoneidx, ac->nodemask);
4842 /* Attempt with potentially adjusted zonelist and alloc_flags */
4843 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4847 /* Caller is not willing to reclaim, we can't balance anything */
4848 if (!can_direct_reclaim)
4851 /* Avoid recursion of direct reclaim */
4852 if (current->flags & PF_MEMALLOC)
4855 /* Try direct reclaim and then allocating */
4856 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4857 &did_some_progress);
4861 /* Try direct compaction and then allocating */
4862 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4863 compact_priority, &compact_result);
4867 /* Do not loop if specifically requested */
4868 if (gfp_mask & __GFP_NORETRY)
4872 * Do not retry costly high order allocations unless they are
4873 * __GFP_RETRY_MAYFAIL
4875 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4878 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4879 did_some_progress > 0, &no_progress_loops))
4883 * It doesn't make any sense to retry for the compaction if the order-0
4884 * reclaim is not able to make any progress because the current
4885 * implementation of the compaction depends on the sufficient amount
4886 * of free memory (see __compaction_suitable)
4888 if (did_some_progress > 0 &&
4889 should_compact_retry(ac, order, alloc_flags,
4890 compact_result, &compact_priority,
4891 &compaction_retries))
4895 /* Deal with possible cpuset update races before we start OOM killing */
4896 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4899 /* Reclaim has failed us, start killing things */
4900 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4904 /* Avoid allocations with no watermarks from looping endlessly */
4905 if (tsk_is_oom_victim(current) &&
4906 (alloc_flags & ALLOC_OOM ||
4907 (gfp_mask & __GFP_NOMEMALLOC)))
4910 /* Retry as long as the OOM killer is making progress */
4911 if (did_some_progress) {
4912 no_progress_loops = 0;
4917 /* Deal with possible cpuset update races before we fail */
4918 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4922 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4925 if (gfp_mask & __GFP_NOFAIL) {
4927 * All existing users of the __GFP_NOFAIL are blockable, so warn
4928 * of any new users that actually require GFP_NOWAIT
4930 if (WARN_ON_ONCE(!can_direct_reclaim))
4934 * PF_MEMALLOC request from this context is rather bizarre
4935 * because we cannot reclaim anything and only can loop waiting
4936 * for somebody to do a work for us
4938 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4941 * non failing costly orders are a hard requirement which we
4942 * are not prepared for much so let's warn about these users
4943 * so that we can identify them and convert them to something
4946 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4949 * Help non-failing allocations by giving them access to memory
4950 * reserves but do not use ALLOC_NO_WATERMARKS because this
4951 * could deplete whole memory reserves which would just make
4952 * the situation worse
4954 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4962 warn_alloc(gfp_mask, ac->nodemask,
4963 "page allocation failure: order:%u", order);
4968 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4969 int preferred_nid, nodemask_t *nodemask,
4970 struct alloc_context *ac, gfp_t *alloc_gfp,
4971 unsigned int *alloc_flags)
4973 ac->highest_zoneidx = gfp_zone(gfp_mask);
4974 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4975 ac->nodemask = nodemask;
4976 ac->migratetype = gfp_migratetype(gfp_mask);
4978 if (cpusets_enabled()) {
4979 *alloc_gfp |= __GFP_HARDWALL;
4981 * When we are in the interrupt context, it is irrelevant
4982 * to the current task context. It means that any node ok.
4984 if (!in_interrupt() && !ac->nodemask)
4985 ac->nodemask = &cpuset_current_mems_allowed;
4987 *alloc_flags |= ALLOC_CPUSET;
4990 fs_reclaim_acquire(gfp_mask);
4991 fs_reclaim_release(gfp_mask);
4993 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4995 if (should_fail_alloc_page(gfp_mask, order))
4998 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
5000 /* Dirty zone balancing only done in the fast path */
5001 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5004 * The preferred zone is used for statistics but crucially it is
5005 * also used as the starting point for the zonelist iterator. It
5006 * may get reset for allocations that ignore memory policies.
5008 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5009 ac->highest_zoneidx, ac->nodemask);
5015 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5016 * @gfp: GFP flags for the allocation
5017 * @preferred_nid: The preferred NUMA node ID to allocate from
5018 * @nodemask: Set of nodes to allocate from, may be NULL
5019 * @nr_pages: The number of pages desired on the list or array
5020 * @page_list: Optional list to store the allocated pages
5021 * @page_array: Optional array to store the pages
5023 * This is a batched version of the page allocator that attempts to
5024 * allocate nr_pages quickly. Pages are added to page_list if page_list
5025 * is not NULL, otherwise it is assumed that the page_array is valid.
5027 * For lists, nr_pages is the number of pages that should be allocated.
5029 * For arrays, only NULL elements are populated with pages and nr_pages
5030 * is the maximum number of pages that will be stored in the array.
5032 * Returns the number of pages on the list or array.
5034 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5035 nodemask_t *nodemask, int nr_pages,
5036 struct list_head *page_list,
5037 struct page **page_array)
5040 unsigned long flags;
5043 struct per_cpu_pages *pcp;
5044 struct list_head *pcp_list;
5045 struct alloc_context ac;
5047 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5048 int nr_populated = 0;
5050 if (unlikely(nr_pages <= 0))
5054 * Skip populated array elements to determine if any pages need
5055 * to be allocated before disabling IRQs.
5057 while (page_array && page_array[nr_populated] && nr_populated < nr_pages)
5060 /* Use the single page allocator for one page. */
5061 if (nr_pages - nr_populated == 1)
5064 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5065 gfp &= gfp_allowed_mask;
5067 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5071 /* Find an allowed local zone that meets the low watermark. */
5072 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5075 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5076 !__cpuset_zone_allowed(zone, gfp)) {
5080 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5081 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5085 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5086 if (zone_watermark_fast(zone, 0, mark,
5087 zonelist_zone_idx(ac.preferred_zoneref),
5088 alloc_flags, gfp)) {
5094 * If there are no allowed local zones that meets the watermarks then
5095 * try to allocate a single page and reclaim if necessary.
5097 if (unlikely(!zone))
5100 /* Attempt the batch allocation */
5101 local_irq_save(flags);
5102 pcp = &this_cpu_ptr(zone->pageset)->pcp;
5103 pcp_list = &pcp->lists[ac.migratetype];
5105 while (nr_populated < nr_pages) {
5107 /* Skip existing pages */
5108 if (page_array && page_array[nr_populated]) {
5113 page = __rmqueue_pcplist(zone, ac.migratetype, alloc_flags,
5115 if (unlikely(!page)) {
5116 /* Try and get at least one page */
5123 * Ideally this would be batched but the best way to do
5124 * that cheaply is to first convert zone_statistics to
5125 * be inaccurate per-cpu counter like vm_events to avoid
5126 * a RMW cycle then do the accounting with IRQs enabled.
5128 __count_zid_vm_events(PGALLOC, zone_idx(zone), 1);
5129 zone_statistics(ac.preferred_zoneref->zone, zone);
5131 prep_new_page(page, 0, gfp, 0);
5133 list_add(&page->lru, page_list);
5135 page_array[nr_populated] = page;
5139 local_irq_restore(flags);
5141 return nr_populated;
5144 local_irq_restore(flags);
5147 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5150 list_add(&page->lru, page_list);
5152 page_array[nr_populated] = page;
5156 return nr_populated;
5158 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5161 * This is the 'heart' of the zoned buddy allocator.
5163 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5164 nodemask_t *nodemask)
5167 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5168 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5169 struct alloc_context ac = { };
5172 * There are several places where we assume that the order value is sane
5173 * so bail out early if the request is out of bound.
5175 if (unlikely(order >= MAX_ORDER)) {
5176 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5180 gfp &= gfp_allowed_mask;
5182 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5183 &alloc_gfp, &alloc_flags))
5187 * Forbid the first pass from falling back to types that fragment
5188 * memory until all local zones are considered.
5190 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5192 /* First allocation attempt */
5193 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5198 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5199 * resp. GFP_NOIO which has to be inherited for all allocation requests
5200 * from a particular context which has been marked by
5201 * memalloc_no{fs,io}_{save,restore}.
5203 alloc_gfp = current_gfp_context(gfp);
5204 ac.spread_dirty_pages = false;
5207 * Restore the original nodemask if it was potentially replaced with
5208 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5210 ac.nodemask = nodemask;
5212 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5215 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5216 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5217 __free_pages(page, order);
5221 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5225 EXPORT_SYMBOL(__alloc_pages);
5228 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5229 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5230 * you need to access high mem.
5232 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5236 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5239 return (unsigned long) page_address(page);
5241 EXPORT_SYMBOL(__get_free_pages);
5243 unsigned long get_zeroed_page(gfp_t gfp_mask)
5245 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5247 EXPORT_SYMBOL(get_zeroed_page);
5249 static inline void free_the_page(struct page *page, unsigned int order)
5251 if (order == 0) /* Via pcp? */
5252 free_unref_page(page);
5254 __free_pages_ok(page, order, FPI_NONE);
5258 * __free_pages - Free pages allocated with alloc_pages().
5259 * @page: The page pointer returned from alloc_pages().
5260 * @order: The order of the allocation.
5262 * This function can free multi-page allocations that are not compound
5263 * pages. It does not check that the @order passed in matches that of
5264 * the allocation, so it is easy to leak memory. Freeing more memory
5265 * than was allocated will probably emit a warning.
5267 * If the last reference to this page is speculative, it will be released
5268 * by put_page() which only frees the first page of a non-compound
5269 * allocation. To prevent the remaining pages from being leaked, we free
5270 * the subsequent pages here. If you want to use the page's reference
5271 * count to decide when to free the allocation, you should allocate a
5272 * compound page, and use put_page() instead of __free_pages().
5274 * Context: May be called in interrupt context or while holding a normal
5275 * spinlock, but not in NMI context or while holding a raw spinlock.
5277 void __free_pages(struct page *page, unsigned int order)
5279 if (put_page_testzero(page))
5280 free_the_page(page, order);
5281 else if (!PageHead(page))
5283 free_the_page(page + (1 << order), order);
5285 EXPORT_SYMBOL(__free_pages);
5287 void free_pages(unsigned long addr, unsigned int order)
5290 VM_BUG_ON(!virt_addr_valid((void *)addr));
5291 __free_pages(virt_to_page((void *)addr), order);
5295 EXPORT_SYMBOL(free_pages);
5299 * An arbitrary-length arbitrary-offset area of memory which resides
5300 * within a 0 or higher order page. Multiple fragments within that page
5301 * are individually refcounted, in the page's reference counter.
5303 * The page_frag functions below provide a simple allocation framework for
5304 * page fragments. This is used by the network stack and network device
5305 * drivers to provide a backing region of memory for use as either an
5306 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5308 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5311 struct page *page = NULL;
5312 gfp_t gfp = gfp_mask;
5314 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5315 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5317 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5318 PAGE_FRAG_CACHE_MAX_ORDER);
5319 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5321 if (unlikely(!page))
5322 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5324 nc->va = page ? page_address(page) : NULL;
5329 void __page_frag_cache_drain(struct page *page, unsigned int count)
5331 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5333 if (page_ref_sub_and_test(page, count))
5334 free_the_page(page, compound_order(page));
5336 EXPORT_SYMBOL(__page_frag_cache_drain);
5338 void *page_frag_alloc_align(struct page_frag_cache *nc,
5339 unsigned int fragsz, gfp_t gfp_mask,
5340 unsigned int align_mask)
5342 unsigned int size = PAGE_SIZE;
5346 if (unlikely(!nc->va)) {
5348 page = __page_frag_cache_refill(nc, gfp_mask);
5352 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5353 /* if size can vary use size else just use PAGE_SIZE */
5356 /* Even if we own the page, we do not use atomic_set().
5357 * This would break get_page_unless_zero() users.
5359 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5361 /* reset page count bias and offset to start of new frag */
5362 nc->pfmemalloc = page_is_pfmemalloc(page);
5363 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5367 offset = nc->offset - fragsz;
5368 if (unlikely(offset < 0)) {
5369 page = virt_to_page(nc->va);
5371 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5374 if (unlikely(nc->pfmemalloc)) {
5375 free_the_page(page, compound_order(page));
5379 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5380 /* if size can vary use size else just use PAGE_SIZE */
5383 /* OK, page count is 0, we can safely set it */
5384 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5386 /* reset page count bias and offset to start of new frag */
5387 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5388 offset = size - fragsz;
5392 offset &= align_mask;
5393 nc->offset = offset;
5395 return nc->va + offset;
5397 EXPORT_SYMBOL(page_frag_alloc_align);
5400 * Frees a page fragment allocated out of either a compound or order 0 page.
5402 void page_frag_free(void *addr)
5404 struct page *page = virt_to_head_page(addr);
5406 if (unlikely(put_page_testzero(page)))
5407 free_the_page(page, compound_order(page));
5409 EXPORT_SYMBOL(page_frag_free);
5411 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5415 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5416 unsigned long used = addr + PAGE_ALIGN(size);
5418 split_page(virt_to_page((void *)addr), order);
5419 while (used < alloc_end) {
5424 return (void *)addr;
5428 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5429 * @size: the number of bytes to allocate
5430 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5432 * This function is similar to alloc_pages(), except that it allocates the
5433 * minimum number of pages to satisfy the request. alloc_pages() can only
5434 * allocate memory in power-of-two pages.
5436 * This function is also limited by MAX_ORDER.
5438 * Memory allocated by this function must be released by free_pages_exact().
5440 * Return: pointer to the allocated area or %NULL in case of error.
5442 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5444 unsigned int order = get_order(size);
5447 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5448 gfp_mask &= ~__GFP_COMP;
5450 addr = __get_free_pages(gfp_mask, order);
5451 return make_alloc_exact(addr, order, size);
5453 EXPORT_SYMBOL(alloc_pages_exact);
5456 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5458 * @nid: the preferred node ID where memory should be allocated
5459 * @size: the number of bytes to allocate
5460 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5462 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5465 * Return: pointer to the allocated area or %NULL in case of error.
5467 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5469 unsigned int order = get_order(size);
5472 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5473 gfp_mask &= ~__GFP_COMP;
5475 p = alloc_pages_node(nid, gfp_mask, order);
5478 return make_alloc_exact((unsigned long)page_address(p), order, size);
5482 * free_pages_exact - release memory allocated via alloc_pages_exact()
5483 * @virt: the value returned by alloc_pages_exact.
5484 * @size: size of allocation, same value as passed to alloc_pages_exact().
5486 * Release the memory allocated by a previous call to alloc_pages_exact.
5488 void free_pages_exact(void *virt, size_t size)
5490 unsigned long addr = (unsigned long)virt;
5491 unsigned long end = addr + PAGE_ALIGN(size);
5493 while (addr < end) {
5498 EXPORT_SYMBOL(free_pages_exact);
5501 * nr_free_zone_pages - count number of pages beyond high watermark
5502 * @offset: The zone index of the highest zone
5504 * nr_free_zone_pages() counts the number of pages which are beyond the
5505 * high watermark within all zones at or below a given zone index. For each
5506 * zone, the number of pages is calculated as:
5508 * nr_free_zone_pages = managed_pages - high_pages
5510 * Return: number of pages beyond high watermark.
5512 static unsigned long nr_free_zone_pages(int offset)
5517 /* Just pick one node, since fallback list is circular */
5518 unsigned long sum = 0;
5520 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5522 for_each_zone_zonelist(zone, z, zonelist, offset) {
5523 unsigned long size = zone_managed_pages(zone);
5524 unsigned long high = high_wmark_pages(zone);
5533 * nr_free_buffer_pages - count number of pages beyond high watermark
5535 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5536 * watermark within ZONE_DMA and ZONE_NORMAL.
5538 * Return: number of pages beyond high watermark within ZONE_DMA and
5541 unsigned long nr_free_buffer_pages(void)
5543 return nr_free_zone_pages(gfp_zone(GFP_USER));
5545 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5547 static inline void show_node(struct zone *zone)
5549 if (IS_ENABLED(CONFIG_NUMA))
5550 printk("Node %d ", zone_to_nid(zone));
5553 long si_mem_available(void)
5556 unsigned long pagecache;
5557 unsigned long wmark_low = 0;
5558 unsigned long pages[NR_LRU_LISTS];
5559 unsigned long reclaimable;
5563 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5564 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5567 wmark_low += low_wmark_pages(zone);
5570 * Estimate the amount of memory available for userspace allocations,
5571 * without causing swapping.
5573 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5576 * Not all the page cache can be freed, otherwise the system will
5577 * start swapping. Assume at least half of the page cache, or the
5578 * low watermark worth of cache, needs to stay.
5580 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5581 pagecache -= min(pagecache / 2, wmark_low);
5582 available += pagecache;
5585 * Part of the reclaimable slab and other kernel memory consists of
5586 * items that are in use, and cannot be freed. Cap this estimate at the
5589 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5590 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5591 available += reclaimable - min(reclaimable / 2, wmark_low);
5597 EXPORT_SYMBOL_GPL(si_mem_available);
5599 void si_meminfo(struct sysinfo *val)
5601 val->totalram = totalram_pages();
5602 val->sharedram = global_node_page_state(NR_SHMEM);
5603 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5604 val->bufferram = nr_blockdev_pages();
5605 val->totalhigh = totalhigh_pages();
5606 val->freehigh = nr_free_highpages();
5607 val->mem_unit = PAGE_SIZE;
5610 EXPORT_SYMBOL(si_meminfo);
5613 void si_meminfo_node(struct sysinfo *val, int nid)
5615 int zone_type; /* needs to be signed */
5616 unsigned long managed_pages = 0;
5617 unsigned long managed_highpages = 0;
5618 unsigned long free_highpages = 0;
5619 pg_data_t *pgdat = NODE_DATA(nid);
5621 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5622 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5623 val->totalram = managed_pages;
5624 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5625 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5626 #ifdef CONFIG_HIGHMEM
5627 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5628 struct zone *zone = &pgdat->node_zones[zone_type];
5630 if (is_highmem(zone)) {
5631 managed_highpages += zone_managed_pages(zone);
5632 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5635 val->totalhigh = managed_highpages;
5636 val->freehigh = free_highpages;
5638 val->totalhigh = managed_highpages;
5639 val->freehigh = free_highpages;
5641 val->mem_unit = PAGE_SIZE;
5646 * Determine whether the node should be displayed or not, depending on whether
5647 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5649 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5651 if (!(flags & SHOW_MEM_FILTER_NODES))
5655 * no node mask - aka implicit memory numa policy. Do not bother with
5656 * the synchronization - read_mems_allowed_begin - because we do not
5657 * have to be precise here.
5660 nodemask = &cpuset_current_mems_allowed;
5662 return !node_isset(nid, *nodemask);
5665 #define K(x) ((x) << (PAGE_SHIFT-10))
5667 static void show_migration_types(unsigned char type)
5669 static const char types[MIGRATE_TYPES] = {
5670 [MIGRATE_UNMOVABLE] = 'U',
5671 [MIGRATE_MOVABLE] = 'M',
5672 [MIGRATE_RECLAIMABLE] = 'E',
5673 [MIGRATE_HIGHATOMIC] = 'H',
5675 [MIGRATE_CMA] = 'C',
5677 #ifdef CONFIG_MEMORY_ISOLATION
5678 [MIGRATE_ISOLATE] = 'I',
5681 char tmp[MIGRATE_TYPES + 1];
5685 for (i = 0; i < MIGRATE_TYPES; i++) {
5686 if (type & (1 << i))
5691 printk(KERN_CONT "(%s) ", tmp);
5695 * Show free area list (used inside shift_scroll-lock stuff)
5696 * We also calculate the percentage fragmentation. We do this by counting the
5697 * memory on each free list with the exception of the first item on the list.
5700 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5703 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5705 unsigned long free_pcp = 0;
5710 for_each_populated_zone(zone) {
5711 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5714 for_each_online_cpu(cpu)
5715 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5718 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5719 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5720 " unevictable:%lu dirty:%lu writeback:%lu\n"
5721 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5722 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5723 " free:%lu free_pcp:%lu free_cma:%lu\n",
5724 global_node_page_state(NR_ACTIVE_ANON),
5725 global_node_page_state(NR_INACTIVE_ANON),
5726 global_node_page_state(NR_ISOLATED_ANON),
5727 global_node_page_state(NR_ACTIVE_FILE),
5728 global_node_page_state(NR_INACTIVE_FILE),
5729 global_node_page_state(NR_ISOLATED_FILE),
5730 global_node_page_state(NR_UNEVICTABLE),
5731 global_node_page_state(NR_FILE_DIRTY),
5732 global_node_page_state(NR_WRITEBACK),
5733 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5734 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5735 global_node_page_state(NR_FILE_MAPPED),
5736 global_node_page_state(NR_SHMEM),
5737 global_node_page_state(NR_PAGETABLE),
5738 global_zone_page_state(NR_BOUNCE),
5739 global_zone_page_state(NR_FREE_PAGES),
5741 global_zone_page_state(NR_FREE_CMA_PAGES));
5743 for_each_online_pgdat(pgdat) {
5744 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5748 " active_anon:%lukB"
5749 " inactive_anon:%lukB"
5750 " active_file:%lukB"
5751 " inactive_file:%lukB"
5752 " unevictable:%lukB"
5753 " isolated(anon):%lukB"
5754 " isolated(file):%lukB"
5759 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5761 " shmem_pmdmapped: %lukB"
5764 " writeback_tmp:%lukB"
5765 " kernel_stack:%lukB"
5766 #ifdef CONFIG_SHADOW_CALL_STACK
5767 " shadow_call_stack:%lukB"
5770 " all_unreclaimable? %s"
5773 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5774 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5775 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5776 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5777 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5778 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5779 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5780 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5781 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5782 K(node_page_state(pgdat, NR_WRITEBACK)),
5783 K(node_page_state(pgdat, NR_SHMEM)),
5784 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5785 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5786 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5787 K(node_page_state(pgdat, NR_ANON_THPS)),
5789 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5790 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5791 #ifdef CONFIG_SHADOW_CALL_STACK
5792 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5794 K(node_page_state(pgdat, NR_PAGETABLE)),
5795 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5799 for_each_populated_zone(zone) {
5802 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5806 for_each_online_cpu(cpu)
5807 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5816 " reserved_highatomic:%luKB"
5817 " active_anon:%lukB"
5818 " inactive_anon:%lukB"
5819 " active_file:%lukB"
5820 " inactive_file:%lukB"
5821 " unevictable:%lukB"
5822 " writepending:%lukB"
5832 K(zone_page_state(zone, NR_FREE_PAGES)),
5833 K(min_wmark_pages(zone)),
5834 K(low_wmark_pages(zone)),
5835 K(high_wmark_pages(zone)),
5836 K(zone->nr_reserved_highatomic),
5837 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5838 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5839 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5840 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5841 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5842 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5843 K(zone->present_pages),
5844 K(zone_managed_pages(zone)),
5845 K(zone_page_state(zone, NR_MLOCK)),
5846 K(zone_page_state(zone, NR_BOUNCE)),
5848 K(this_cpu_read(zone->pageset->pcp.count)),
5849 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5850 printk("lowmem_reserve[]:");
5851 for (i = 0; i < MAX_NR_ZONES; i++)
5852 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5853 printk(KERN_CONT "\n");
5856 for_each_populated_zone(zone) {
5858 unsigned long nr[MAX_ORDER], flags, total = 0;
5859 unsigned char types[MAX_ORDER];
5861 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5864 printk(KERN_CONT "%s: ", zone->name);
5866 spin_lock_irqsave(&zone->lock, flags);
5867 for (order = 0; order < MAX_ORDER; order++) {
5868 struct free_area *area = &zone->free_area[order];
5871 nr[order] = area->nr_free;
5872 total += nr[order] << order;
5875 for (type = 0; type < MIGRATE_TYPES; type++) {
5876 if (!free_area_empty(area, type))
5877 types[order] |= 1 << type;
5880 spin_unlock_irqrestore(&zone->lock, flags);
5881 for (order = 0; order < MAX_ORDER; order++) {
5882 printk(KERN_CONT "%lu*%lukB ",
5883 nr[order], K(1UL) << order);
5885 show_migration_types(types[order]);
5887 printk(KERN_CONT "= %lukB\n", K(total));
5890 hugetlb_show_meminfo();
5892 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5894 show_swap_cache_info();
5897 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5899 zoneref->zone = zone;
5900 zoneref->zone_idx = zone_idx(zone);
5904 * Builds allocation fallback zone lists.
5906 * Add all populated zones of a node to the zonelist.
5908 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5911 enum zone_type zone_type = MAX_NR_ZONES;
5916 zone = pgdat->node_zones + zone_type;
5917 if (managed_zone(zone)) {
5918 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5919 check_highest_zone(zone_type);
5921 } while (zone_type);
5928 static int __parse_numa_zonelist_order(char *s)
5931 * We used to support different zonlists modes but they turned
5932 * out to be just not useful. Let's keep the warning in place
5933 * if somebody still use the cmd line parameter so that we do
5934 * not fail it silently
5936 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5937 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5943 char numa_zonelist_order[] = "Node";
5946 * sysctl handler for numa_zonelist_order
5948 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5949 void *buffer, size_t *length, loff_t *ppos)
5952 return __parse_numa_zonelist_order(buffer);
5953 return proc_dostring(table, write, buffer, length, ppos);
5957 #define MAX_NODE_LOAD (nr_online_nodes)
5958 static int node_load[MAX_NUMNODES];
5961 * find_next_best_node - find the next node that should appear in a given node's fallback list
5962 * @node: node whose fallback list we're appending
5963 * @used_node_mask: nodemask_t of already used nodes
5965 * We use a number of factors to determine which is the next node that should
5966 * appear on a given node's fallback list. The node should not have appeared
5967 * already in @node's fallback list, and it should be the next closest node
5968 * according to the distance array (which contains arbitrary distance values
5969 * from each node to each node in the system), and should also prefer nodes
5970 * with no CPUs, since presumably they'll have very little allocation pressure
5971 * on them otherwise.
5973 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5975 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5978 int min_val = INT_MAX;
5979 int best_node = NUMA_NO_NODE;
5981 /* Use the local node if we haven't already */
5982 if (!node_isset(node, *used_node_mask)) {
5983 node_set(node, *used_node_mask);
5987 for_each_node_state(n, N_MEMORY) {
5989 /* Don't want a node to appear more than once */
5990 if (node_isset(n, *used_node_mask))
5993 /* Use the distance array to find the distance */
5994 val = node_distance(node, n);
5996 /* Penalize nodes under us ("prefer the next node") */
5999 /* Give preference to headless and unused nodes */
6000 if (!cpumask_empty(cpumask_of_node(n)))
6001 val += PENALTY_FOR_NODE_WITH_CPUS;
6003 /* Slight preference for less loaded node */
6004 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6005 val += node_load[n];
6007 if (val < min_val) {
6014 node_set(best_node, *used_node_mask);
6021 * Build zonelists ordered by node and zones within node.
6022 * This results in maximum locality--normal zone overflows into local
6023 * DMA zone, if any--but risks exhausting DMA zone.
6025 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6028 struct zoneref *zonerefs;
6031 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6033 for (i = 0; i < nr_nodes; i++) {
6036 pg_data_t *node = NODE_DATA(node_order[i]);
6038 nr_zones = build_zonerefs_node(node, zonerefs);
6039 zonerefs += nr_zones;
6041 zonerefs->zone = NULL;
6042 zonerefs->zone_idx = 0;
6046 * Build gfp_thisnode zonelists
6048 static void build_thisnode_zonelists(pg_data_t *pgdat)
6050 struct zoneref *zonerefs;
6053 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6054 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6055 zonerefs += nr_zones;
6056 zonerefs->zone = NULL;
6057 zonerefs->zone_idx = 0;
6061 * Build zonelists ordered by zone and nodes within zones.
6062 * This results in conserving DMA zone[s] until all Normal memory is
6063 * exhausted, but results in overflowing to remote node while memory
6064 * may still exist in local DMA zone.
6067 static void build_zonelists(pg_data_t *pgdat)
6069 static int node_order[MAX_NUMNODES];
6070 int node, load, nr_nodes = 0;
6071 nodemask_t used_mask = NODE_MASK_NONE;
6072 int local_node, prev_node;
6074 /* NUMA-aware ordering of nodes */
6075 local_node = pgdat->node_id;
6076 load = nr_online_nodes;
6077 prev_node = local_node;
6079 memset(node_order, 0, sizeof(node_order));
6080 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6082 * We don't want to pressure a particular node.
6083 * So adding penalty to the first node in same
6084 * distance group to make it round-robin.
6086 if (node_distance(local_node, node) !=
6087 node_distance(local_node, prev_node))
6088 node_load[node] = load;
6090 node_order[nr_nodes++] = node;
6095 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6096 build_thisnode_zonelists(pgdat);
6099 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6101 * Return node id of node used for "local" allocations.
6102 * I.e., first node id of first zone in arg node's generic zonelist.
6103 * Used for initializing percpu 'numa_mem', which is used primarily
6104 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6106 int local_memory_node(int node)
6110 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6111 gfp_zone(GFP_KERNEL),
6113 return zone_to_nid(z->zone);
6117 static void setup_min_unmapped_ratio(void);
6118 static void setup_min_slab_ratio(void);
6119 #else /* CONFIG_NUMA */
6121 static void build_zonelists(pg_data_t *pgdat)
6123 int node, local_node;
6124 struct zoneref *zonerefs;
6127 local_node = pgdat->node_id;
6129 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6130 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6131 zonerefs += nr_zones;
6134 * Now we build the zonelist so that it contains the zones
6135 * of all the other nodes.
6136 * We don't want to pressure a particular node, so when
6137 * building the zones for node N, we make sure that the
6138 * zones coming right after the local ones are those from
6139 * node N+1 (modulo N)
6141 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6142 if (!node_online(node))
6144 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6145 zonerefs += nr_zones;
6147 for (node = 0; node < local_node; node++) {
6148 if (!node_online(node))
6150 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6151 zonerefs += nr_zones;
6154 zonerefs->zone = NULL;
6155 zonerefs->zone_idx = 0;
6158 #endif /* CONFIG_NUMA */
6161 * Boot pageset table. One per cpu which is going to be used for all
6162 * zones and all nodes. The parameters will be set in such a way
6163 * that an item put on a list will immediately be handed over to
6164 * the buddy list. This is safe since pageset manipulation is done
6165 * with interrupts disabled.
6167 * The boot_pagesets must be kept even after bootup is complete for
6168 * unused processors and/or zones. They do play a role for bootstrapping
6169 * hotplugged processors.
6171 * zoneinfo_show() and maybe other functions do
6172 * not check if the processor is online before following the pageset pointer.
6173 * Other parts of the kernel may not check if the zone is available.
6175 static void pageset_init(struct per_cpu_pageset *p);
6176 /* These effectively disable the pcplists in the boot pageset completely */
6177 #define BOOT_PAGESET_HIGH 0
6178 #define BOOT_PAGESET_BATCH 1
6179 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
6180 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6182 static void __build_all_zonelists(void *data)
6185 int __maybe_unused cpu;
6186 pg_data_t *self = data;
6187 static DEFINE_SPINLOCK(lock);
6192 memset(node_load, 0, sizeof(node_load));
6196 * This node is hotadded and no memory is yet present. So just
6197 * building zonelists is fine - no need to touch other nodes.
6199 if (self && !node_online(self->node_id)) {
6200 build_zonelists(self);
6202 for_each_online_node(nid) {
6203 pg_data_t *pgdat = NODE_DATA(nid);
6205 build_zonelists(pgdat);
6208 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6210 * We now know the "local memory node" for each node--
6211 * i.e., the node of the first zone in the generic zonelist.
6212 * Set up numa_mem percpu variable for on-line cpus. During
6213 * boot, only the boot cpu should be on-line; we'll init the
6214 * secondary cpus' numa_mem as they come on-line. During
6215 * node/memory hotplug, we'll fixup all on-line cpus.
6217 for_each_online_cpu(cpu)
6218 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6225 static noinline void __init
6226 build_all_zonelists_init(void)
6230 __build_all_zonelists(NULL);
6233 * Initialize the boot_pagesets that are going to be used
6234 * for bootstrapping processors. The real pagesets for
6235 * each zone will be allocated later when the per cpu
6236 * allocator is available.
6238 * boot_pagesets are used also for bootstrapping offline
6239 * cpus if the system is already booted because the pagesets
6240 * are needed to initialize allocators on a specific cpu too.
6241 * F.e. the percpu allocator needs the page allocator which
6242 * needs the percpu allocator in order to allocate its pagesets
6243 * (a chicken-egg dilemma).
6245 for_each_possible_cpu(cpu)
6246 pageset_init(&per_cpu(boot_pageset, cpu));
6248 mminit_verify_zonelist();
6249 cpuset_init_current_mems_allowed();
6253 * unless system_state == SYSTEM_BOOTING.
6255 * __ref due to call of __init annotated helper build_all_zonelists_init
6256 * [protected by SYSTEM_BOOTING].
6258 void __ref build_all_zonelists(pg_data_t *pgdat)
6260 unsigned long vm_total_pages;
6262 if (system_state == SYSTEM_BOOTING) {
6263 build_all_zonelists_init();
6265 __build_all_zonelists(pgdat);
6266 /* cpuset refresh routine should be here */
6268 /* Get the number of free pages beyond high watermark in all zones. */
6269 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6271 * Disable grouping by mobility if the number of pages in the
6272 * system is too low to allow the mechanism to work. It would be
6273 * more accurate, but expensive to check per-zone. This check is
6274 * made on memory-hotadd so a system can start with mobility
6275 * disabled and enable it later
6277 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6278 page_group_by_mobility_disabled = 1;
6280 page_group_by_mobility_disabled = 0;
6282 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6284 page_group_by_mobility_disabled ? "off" : "on",
6287 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6291 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6292 static bool __meminit
6293 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6295 static struct memblock_region *r;
6297 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6298 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6299 for_each_mem_region(r) {
6300 if (*pfn < memblock_region_memory_end_pfn(r))
6304 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6305 memblock_is_mirror(r)) {
6306 *pfn = memblock_region_memory_end_pfn(r);
6314 * Initially all pages are reserved - free ones are freed
6315 * up by memblock_free_all() once the early boot process is
6316 * done. Non-atomic initialization, single-pass.
6318 * All aligned pageblocks are initialized to the specified migratetype
6319 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6320 * zone stats (e.g., nr_isolate_pageblock) are touched.
6322 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6323 unsigned long start_pfn, unsigned long zone_end_pfn,
6324 enum meminit_context context,
6325 struct vmem_altmap *altmap, int migratetype)
6327 unsigned long pfn, end_pfn = start_pfn + size;
6330 if (highest_memmap_pfn < end_pfn - 1)
6331 highest_memmap_pfn = end_pfn - 1;
6333 #ifdef CONFIG_ZONE_DEVICE
6335 * Honor reservation requested by the driver for this ZONE_DEVICE
6336 * memory. We limit the total number of pages to initialize to just
6337 * those that might contain the memory mapping. We will defer the
6338 * ZONE_DEVICE page initialization until after we have released
6341 if (zone == ZONE_DEVICE) {
6345 if (start_pfn == altmap->base_pfn)
6346 start_pfn += altmap->reserve;
6347 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6351 for (pfn = start_pfn; pfn < end_pfn; ) {
6353 * There can be holes in boot-time mem_map[]s handed to this
6354 * function. They do not exist on hotplugged memory.
6356 if (context == MEMINIT_EARLY) {
6357 if (overlap_memmap_init(zone, &pfn))
6359 if (defer_init(nid, pfn, zone_end_pfn))
6363 page = pfn_to_page(pfn);
6364 __init_single_page(page, pfn, zone, nid);
6365 if (context == MEMINIT_HOTPLUG)
6366 __SetPageReserved(page);
6369 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6370 * such that unmovable allocations won't be scattered all
6371 * over the place during system boot.
6373 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6374 set_pageblock_migratetype(page, migratetype);
6381 #ifdef CONFIG_ZONE_DEVICE
6382 void __ref memmap_init_zone_device(struct zone *zone,
6383 unsigned long start_pfn,
6384 unsigned long nr_pages,
6385 struct dev_pagemap *pgmap)
6387 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6388 struct pglist_data *pgdat = zone->zone_pgdat;
6389 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6390 unsigned long zone_idx = zone_idx(zone);
6391 unsigned long start = jiffies;
6392 int nid = pgdat->node_id;
6394 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6398 * The call to memmap_init_zone should have already taken care
6399 * of the pages reserved for the memmap, so we can just jump to
6400 * the end of that region and start processing the device pages.
6403 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6404 nr_pages = end_pfn - start_pfn;
6407 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6408 struct page *page = pfn_to_page(pfn);
6410 __init_single_page(page, pfn, zone_idx, nid);
6413 * Mark page reserved as it will need to wait for onlining
6414 * phase for it to be fully associated with a zone.
6416 * We can use the non-atomic __set_bit operation for setting
6417 * the flag as we are still initializing the pages.
6419 __SetPageReserved(page);
6422 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6423 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6424 * ever freed or placed on a driver-private list.
6426 page->pgmap = pgmap;
6427 page->zone_device_data = NULL;
6430 * Mark the block movable so that blocks are reserved for
6431 * movable at startup. This will force kernel allocations
6432 * to reserve their blocks rather than leaking throughout
6433 * the address space during boot when many long-lived
6434 * kernel allocations are made.
6436 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6437 * because this is done early in section_activate()
6439 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6440 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6445 pr_info("%s initialised %lu pages in %ums\n", __func__,
6446 nr_pages, jiffies_to_msecs(jiffies - start));
6450 static void __meminit zone_init_free_lists(struct zone *zone)
6452 unsigned int order, t;
6453 for_each_migratetype_order(order, t) {
6454 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6455 zone->free_area[order].nr_free = 0;
6459 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6461 * Only struct pages that correspond to ranges defined by memblock.memory
6462 * are zeroed and initialized by going through __init_single_page() during
6463 * memmap_init_zone().
6465 * But, there could be struct pages that correspond to holes in
6466 * memblock.memory. This can happen because of the following reasons:
6467 * - physical memory bank size is not necessarily the exact multiple of the
6468 * arbitrary section size
6469 * - early reserved memory may not be listed in memblock.memory
6470 * - memory layouts defined with memmap= kernel parameter may not align
6471 * nicely with memmap sections
6473 * Explicitly initialize those struct pages so that:
6474 * - PG_Reserved is set
6475 * - zone and node links point to zone and node that span the page if the
6476 * hole is in the middle of a zone
6477 * - zone and node links point to adjacent zone/node if the hole falls on
6478 * the zone boundary; the pages in such holes will be prepended to the
6479 * zone/node above the hole except for the trailing pages in the last
6480 * section that will be appended to the zone/node below.
6482 static u64 __meminit init_unavailable_range(unsigned long spfn,
6489 for (pfn = spfn; pfn < epfn; pfn++) {
6490 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6491 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6492 + pageblock_nr_pages - 1;
6495 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6496 __SetPageReserved(pfn_to_page(pfn));
6503 static inline u64 init_unavailable_range(unsigned long spfn, unsigned long epfn,
6510 void __meminit __weak memmap_init_zone(struct zone *zone)
6512 unsigned long zone_start_pfn = zone->zone_start_pfn;
6513 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6514 int i, nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6515 static unsigned long hole_pfn;
6516 unsigned long start_pfn, end_pfn;
6519 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6520 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6521 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6523 if (end_pfn > start_pfn)
6524 memmap_init_range(end_pfn - start_pfn, nid,
6525 zone_id, start_pfn, zone_end_pfn,
6526 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6528 if (hole_pfn < start_pfn)
6529 pgcnt += init_unavailable_range(hole_pfn, start_pfn,
6534 #ifdef CONFIG_SPARSEMEM
6536 * Initialize the hole in the range [zone_end_pfn, section_end].
6537 * If zone boundary falls in the middle of a section, this hole
6538 * will be re-initialized during the call to this function for the
6541 end_pfn = round_up(zone_end_pfn, PAGES_PER_SECTION);
6542 if (hole_pfn < end_pfn)
6543 pgcnt += init_unavailable_range(hole_pfn, end_pfn,
6548 pr_info(" %s zone: %llu pages in unavailable ranges\n",
6552 static int zone_batchsize(struct zone *zone)
6558 * The per-cpu-pages pools are set to around 1000th of the
6561 batch = zone_managed_pages(zone) / 1024;
6562 /* But no more than a meg. */
6563 if (batch * PAGE_SIZE > 1024 * 1024)
6564 batch = (1024 * 1024) / PAGE_SIZE;
6565 batch /= 4; /* We effectively *= 4 below */
6570 * Clamp the batch to a 2^n - 1 value. Having a power
6571 * of 2 value was found to be more likely to have
6572 * suboptimal cache aliasing properties in some cases.
6574 * For example if 2 tasks are alternately allocating
6575 * batches of pages, one task can end up with a lot
6576 * of pages of one half of the possible page colors
6577 * and the other with pages of the other colors.
6579 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6584 /* The deferral and batching of frees should be suppressed under NOMMU
6587 * The problem is that NOMMU needs to be able to allocate large chunks
6588 * of contiguous memory as there's no hardware page translation to
6589 * assemble apparent contiguous memory from discontiguous pages.
6591 * Queueing large contiguous runs of pages for batching, however,
6592 * causes the pages to actually be freed in smaller chunks. As there
6593 * can be a significant delay between the individual batches being
6594 * recycled, this leads to the once large chunks of space being
6595 * fragmented and becoming unavailable for high-order allocations.
6602 * pcp->high and pcp->batch values are related and generally batch is lower
6603 * than high. They are also related to pcp->count such that count is lower
6604 * than high, and as soon as it reaches high, the pcplist is flushed.
6606 * However, guaranteeing these relations at all times would require e.g. write
6607 * barriers here but also careful usage of read barriers at the read side, and
6608 * thus be prone to error and bad for performance. Thus the update only prevents
6609 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6610 * can cope with those fields changing asynchronously, and fully trust only the
6611 * pcp->count field on the local CPU with interrupts disabled.
6613 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6614 * outside of boot time (or some other assurance that no concurrent updaters
6617 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6618 unsigned long batch)
6620 WRITE_ONCE(pcp->batch, batch);
6621 WRITE_ONCE(pcp->high, high);
6624 static void pageset_init(struct per_cpu_pageset *p)
6626 struct per_cpu_pages *pcp;
6629 memset(p, 0, sizeof(*p));
6632 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6633 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6636 * Set batch and high values safe for a boot pageset. A true percpu
6637 * pageset's initialization will update them subsequently. Here we don't
6638 * need to be as careful as pageset_update() as nobody can access the
6641 pcp->high = BOOT_PAGESET_HIGH;
6642 pcp->batch = BOOT_PAGESET_BATCH;
6645 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6646 unsigned long batch)
6648 struct per_cpu_pageset *p;
6651 for_each_possible_cpu(cpu) {
6652 p = per_cpu_ptr(zone->pageset, cpu);
6653 pageset_update(&p->pcp, high, batch);
6658 * Calculate and set new high and batch values for all per-cpu pagesets of a
6659 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6661 static void zone_set_pageset_high_and_batch(struct zone *zone)
6663 unsigned long new_high, new_batch;
6665 if (percpu_pagelist_fraction) {
6666 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6667 new_batch = max(1UL, new_high / 4);
6668 if ((new_high / 4) > (PAGE_SHIFT * 8))
6669 new_batch = PAGE_SHIFT * 8;
6671 new_batch = zone_batchsize(zone);
6672 new_high = 6 * new_batch;
6673 new_batch = max(1UL, 1 * new_batch);
6676 if (zone->pageset_high == new_high &&
6677 zone->pageset_batch == new_batch)
6680 zone->pageset_high = new_high;
6681 zone->pageset_batch = new_batch;
6683 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6686 void __meminit setup_zone_pageset(struct zone *zone)
6688 struct per_cpu_pageset *p;
6691 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6692 for_each_possible_cpu(cpu) {
6693 p = per_cpu_ptr(zone->pageset, cpu);
6697 zone_set_pageset_high_and_batch(zone);
6701 * Allocate per cpu pagesets and initialize them.
6702 * Before this call only boot pagesets were available.
6704 void __init setup_per_cpu_pageset(void)
6706 struct pglist_data *pgdat;
6708 int __maybe_unused cpu;
6710 for_each_populated_zone(zone)
6711 setup_zone_pageset(zone);
6715 * Unpopulated zones continue using the boot pagesets.
6716 * The numa stats for these pagesets need to be reset.
6717 * Otherwise, they will end up skewing the stats of
6718 * the nodes these zones are associated with.
6720 for_each_possible_cpu(cpu) {
6721 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6722 memset(pcp->vm_numa_stat_diff, 0,
6723 sizeof(pcp->vm_numa_stat_diff));
6727 for_each_online_pgdat(pgdat)
6728 pgdat->per_cpu_nodestats =
6729 alloc_percpu(struct per_cpu_nodestat);
6732 static __meminit void zone_pcp_init(struct zone *zone)
6735 * per cpu subsystem is not up at this point. The following code
6736 * relies on the ability of the linker to provide the
6737 * offset of a (static) per cpu variable into the per cpu area.
6739 zone->pageset = &boot_pageset;
6740 zone->pageset_high = BOOT_PAGESET_HIGH;
6741 zone->pageset_batch = BOOT_PAGESET_BATCH;
6743 if (populated_zone(zone))
6744 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6745 zone->name, zone->present_pages,
6746 zone_batchsize(zone));
6749 void __meminit init_currently_empty_zone(struct zone *zone,
6750 unsigned long zone_start_pfn,
6753 struct pglist_data *pgdat = zone->zone_pgdat;
6754 int zone_idx = zone_idx(zone) + 1;
6756 if (zone_idx > pgdat->nr_zones)
6757 pgdat->nr_zones = zone_idx;
6759 zone->zone_start_pfn = zone_start_pfn;
6761 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6762 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6764 (unsigned long)zone_idx(zone),
6765 zone_start_pfn, (zone_start_pfn + size));
6767 zone_init_free_lists(zone);
6768 zone->initialized = 1;
6772 * get_pfn_range_for_nid - Return the start and end page frames for a node
6773 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6774 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6775 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6777 * It returns the start and end page frame of a node based on information
6778 * provided by memblock_set_node(). If called for a node
6779 * with no available memory, a warning is printed and the start and end
6782 void __init get_pfn_range_for_nid(unsigned int nid,
6783 unsigned long *start_pfn, unsigned long *end_pfn)
6785 unsigned long this_start_pfn, this_end_pfn;
6791 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6792 *start_pfn = min(*start_pfn, this_start_pfn);
6793 *end_pfn = max(*end_pfn, this_end_pfn);
6796 if (*start_pfn == -1UL)
6801 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6802 * assumption is made that zones within a node are ordered in monotonic
6803 * increasing memory addresses so that the "highest" populated zone is used
6805 static void __init find_usable_zone_for_movable(void)
6808 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6809 if (zone_index == ZONE_MOVABLE)
6812 if (arch_zone_highest_possible_pfn[zone_index] >
6813 arch_zone_lowest_possible_pfn[zone_index])
6817 VM_BUG_ON(zone_index == -1);
6818 movable_zone = zone_index;
6822 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6823 * because it is sized independent of architecture. Unlike the other zones,
6824 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6825 * in each node depending on the size of each node and how evenly kernelcore
6826 * is distributed. This helper function adjusts the zone ranges
6827 * provided by the architecture for a given node by using the end of the
6828 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6829 * zones within a node are in order of monotonic increases memory addresses
6831 static void __init adjust_zone_range_for_zone_movable(int nid,
6832 unsigned long zone_type,
6833 unsigned long node_start_pfn,
6834 unsigned long node_end_pfn,
6835 unsigned long *zone_start_pfn,
6836 unsigned long *zone_end_pfn)
6838 /* Only adjust if ZONE_MOVABLE is on this node */
6839 if (zone_movable_pfn[nid]) {
6840 /* Size ZONE_MOVABLE */
6841 if (zone_type == ZONE_MOVABLE) {
6842 *zone_start_pfn = zone_movable_pfn[nid];
6843 *zone_end_pfn = min(node_end_pfn,
6844 arch_zone_highest_possible_pfn[movable_zone]);
6846 /* Adjust for ZONE_MOVABLE starting within this range */
6847 } else if (!mirrored_kernelcore &&
6848 *zone_start_pfn < zone_movable_pfn[nid] &&
6849 *zone_end_pfn > zone_movable_pfn[nid]) {
6850 *zone_end_pfn = zone_movable_pfn[nid];
6852 /* Check if this whole range is within ZONE_MOVABLE */
6853 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6854 *zone_start_pfn = *zone_end_pfn;
6859 * Return the number of pages a zone spans in a node, including holes
6860 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6862 static unsigned long __init zone_spanned_pages_in_node(int nid,
6863 unsigned long zone_type,
6864 unsigned long node_start_pfn,
6865 unsigned long node_end_pfn,
6866 unsigned long *zone_start_pfn,
6867 unsigned long *zone_end_pfn)
6869 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6870 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6871 /* When hotadd a new node from cpu_up(), the node should be empty */
6872 if (!node_start_pfn && !node_end_pfn)
6875 /* Get the start and end of the zone */
6876 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6877 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6878 adjust_zone_range_for_zone_movable(nid, zone_type,
6879 node_start_pfn, node_end_pfn,
6880 zone_start_pfn, zone_end_pfn);
6882 /* Check that this node has pages within the zone's required range */
6883 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6886 /* Move the zone boundaries inside the node if necessary */
6887 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6888 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6890 /* Return the spanned pages */
6891 return *zone_end_pfn - *zone_start_pfn;
6895 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6896 * then all holes in the requested range will be accounted for.
6898 unsigned long __init __absent_pages_in_range(int nid,
6899 unsigned long range_start_pfn,
6900 unsigned long range_end_pfn)
6902 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6903 unsigned long start_pfn, end_pfn;
6906 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6907 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6908 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6909 nr_absent -= end_pfn - start_pfn;
6915 * absent_pages_in_range - Return number of page frames in holes within a range
6916 * @start_pfn: The start PFN to start searching for holes
6917 * @end_pfn: The end PFN to stop searching for holes
6919 * Return: the number of pages frames in memory holes within a range.
6921 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6922 unsigned long end_pfn)
6924 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6927 /* Return the number of page frames in holes in a zone on a node */
6928 static unsigned long __init zone_absent_pages_in_node(int nid,
6929 unsigned long zone_type,
6930 unsigned long node_start_pfn,
6931 unsigned long node_end_pfn)
6933 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6934 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6935 unsigned long zone_start_pfn, zone_end_pfn;
6936 unsigned long nr_absent;
6938 /* When hotadd a new node from cpu_up(), the node should be empty */
6939 if (!node_start_pfn && !node_end_pfn)
6942 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6943 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6945 adjust_zone_range_for_zone_movable(nid, zone_type,
6946 node_start_pfn, node_end_pfn,
6947 &zone_start_pfn, &zone_end_pfn);
6948 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6951 * ZONE_MOVABLE handling.
6952 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6955 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6956 unsigned long start_pfn, end_pfn;
6957 struct memblock_region *r;
6959 for_each_mem_region(r) {
6960 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6961 zone_start_pfn, zone_end_pfn);
6962 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6963 zone_start_pfn, zone_end_pfn);
6965 if (zone_type == ZONE_MOVABLE &&
6966 memblock_is_mirror(r))
6967 nr_absent += end_pfn - start_pfn;
6969 if (zone_type == ZONE_NORMAL &&
6970 !memblock_is_mirror(r))
6971 nr_absent += end_pfn - start_pfn;
6978 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6979 unsigned long node_start_pfn,
6980 unsigned long node_end_pfn)
6982 unsigned long realtotalpages = 0, totalpages = 0;
6985 for (i = 0; i < MAX_NR_ZONES; i++) {
6986 struct zone *zone = pgdat->node_zones + i;
6987 unsigned long zone_start_pfn, zone_end_pfn;
6988 unsigned long spanned, absent;
6989 unsigned long size, real_size;
6991 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6996 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7001 real_size = size - absent;
7004 zone->zone_start_pfn = zone_start_pfn;
7006 zone->zone_start_pfn = 0;
7007 zone->spanned_pages = size;
7008 zone->present_pages = real_size;
7011 realtotalpages += real_size;
7014 pgdat->node_spanned_pages = totalpages;
7015 pgdat->node_present_pages = realtotalpages;
7016 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
7020 #ifndef CONFIG_SPARSEMEM
7022 * Calculate the size of the zone->blockflags rounded to an unsigned long
7023 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7024 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7025 * round what is now in bits to nearest long in bits, then return it in
7028 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7030 unsigned long usemapsize;
7032 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7033 usemapsize = roundup(zonesize, pageblock_nr_pages);
7034 usemapsize = usemapsize >> pageblock_order;
7035 usemapsize *= NR_PAGEBLOCK_BITS;
7036 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7038 return usemapsize / 8;
7041 static void __ref setup_usemap(struct zone *zone)
7043 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7044 zone->spanned_pages);
7045 zone->pageblock_flags = NULL;
7047 zone->pageblock_flags =
7048 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7050 if (!zone->pageblock_flags)
7051 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7052 usemapsize, zone->name, zone_to_nid(zone));
7056 static inline void setup_usemap(struct zone *zone) {}
7057 #endif /* CONFIG_SPARSEMEM */
7059 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7061 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7062 void __init set_pageblock_order(void)
7066 /* Check that pageblock_nr_pages has not already been setup */
7067 if (pageblock_order)
7070 if (HPAGE_SHIFT > PAGE_SHIFT)
7071 order = HUGETLB_PAGE_ORDER;
7073 order = MAX_ORDER - 1;
7076 * Assume the largest contiguous order of interest is a huge page.
7077 * This value may be variable depending on boot parameters on IA64 and
7080 pageblock_order = order;
7082 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7085 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7086 * is unused as pageblock_order is set at compile-time. See
7087 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7090 void __init set_pageblock_order(void)
7094 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7096 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7097 unsigned long present_pages)
7099 unsigned long pages = spanned_pages;
7102 * Provide a more accurate estimation if there are holes within
7103 * the zone and SPARSEMEM is in use. If there are holes within the
7104 * zone, each populated memory region may cost us one or two extra
7105 * memmap pages due to alignment because memmap pages for each
7106 * populated regions may not be naturally aligned on page boundary.
7107 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7109 if (spanned_pages > present_pages + (present_pages >> 4) &&
7110 IS_ENABLED(CONFIG_SPARSEMEM))
7111 pages = present_pages;
7113 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7116 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7117 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7119 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7121 spin_lock_init(&ds_queue->split_queue_lock);
7122 INIT_LIST_HEAD(&ds_queue->split_queue);
7123 ds_queue->split_queue_len = 0;
7126 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7129 #ifdef CONFIG_COMPACTION
7130 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7132 init_waitqueue_head(&pgdat->kcompactd_wait);
7135 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7138 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7140 pgdat_resize_init(pgdat);
7142 pgdat_init_split_queue(pgdat);
7143 pgdat_init_kcompactd(pgdat);
7145 init_waitqueue_head(&pgdat->kswapd_wait);
7146 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7148 pgdat_page_ext_init(pgdat);
7149 lruvec_init(&pgdat->__lruvec);
7152 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7153 unsigned long remaining_pages)
7155 atomic_long_set(&zone->managed_pages, remaining_pages);
7156 zone_set_nid(zone, nid);
7157 zone->name = zone_names[idx];
7158 zone->zone_pgdat = NODE_DATA(nid);
7159 spin_lock_init(&zone->lock);
7160 zone_seqlock_init(zone);
7161 zone_pcp_init(zone);
7165 * Set up the zone data structures
7166 * - init pgdat internals
7167 * - init all zones belonging to this node
7169 * NOTE: this function is only called during memory hotplug
7171 #ifdef CONFIG_MEMORY_HOTPLUG
7172 void __ref free_area_init_core_hotplug(int nid)
7175 pg_data_t *pgdat = NODE_DATA(nid);
7177 pgdat_init_internals(pgdat);
7178 for (z = 0; z < MAX_NR_ZONES; z++)
7179 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7184 * Set up the zone data structures:
7185 * - mark all pages reserved
7186 * - mark all memory queues empty
7187 * - clear the memory bitmaps
7189 * NOTE: pgdat should get zeroed by caller.
7190 * NOTE: this function is only called during early init.
7192 static void __init free_area_init_core(struct pglist_data *pgdat)
7195 int nid = pgdat->node_id;
7197 pgdat_init_internals(pgdat);
7198 pgdat->per_cpu_nodestats = &boot_nodestats;
7200 for (j = 0; j < MAX_NR_ZONES; j++) {
7201 struct zone *zone = pgdat->node_zones + j;
7202 unsigned long size, freesize, memmap_pages;
7204 size = zone->spanned_pages;
7205 freesize = zone->present_pages;
7208 * Adjust freesize so that it accounts for how much memory
7209 * is used by this zone for memmap. This affects the watermark
7210 * and per-cpu initialisations
7212 memmap_pages = calc_memmap_size(size, freesize);
7213 if (!is_highmem_idx(j)) {
7214 if (freesize >= memmap_pages) {
7215 freesize -= memmap_pages;
7218 " %s zone: %lu pages used for memmap\n",
7219 zone_names[j], memmap_pages);
7221 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
7222 zone_names[j], memmap_pages, freesize);
7225 /* Account for reserved pages */
7226 if (j == 0 && freesize > dma_reserve) {
7227 freesize -= dma_reserve;
7228 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
7229 zone_names[0], dma_reserve);
7232 if (!is_highmem_idx(j))
7233 nr_kernel_pages += freesize;
7234 /* Charge for highmem memmap if there are enough kernel pages */
7235 else if (nr_kernel_pages > memmap_pages * 2)
7236 nr_kernel_pages -= memmap_pages;
7237 nr_all_pages += freesize;
7240 * Set an approximate value for lowmem here, it will be adjusted
7241 * when the bootmem allocator frees pages into the buddy system.
7242 * And all highmem pages will be managed by the buddy system.
7244 zone_init_internals(zone, j, nid, freesize);
7249 set_pageblock_order();
7251 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7252 memmap_init_zone(zone);
7256 #ifdef CONFIG_FLAT_NODE_MEM_MAP
7257 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7259 unsigned long __maybe_unused start = 0;
7260 unsigned long __maybe_unused offset = 0;
7262 /* Skip empty nodes */
7263 if (!pgdat->node_spanned_pages)
7266 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7267 offset = pgdat->node_start_pfn - start;
7268 /* ia64 gets its own node_mem_map, before this, without bootmem */
7269 if (!pgdat->node_mem_map) {
7270 unsigned long size, end;
7274 * The zone's endpoints aren't required to be MAX_ORDER
7275 * aligned but the node_mem_map endpoints must be in order
7276 * for the buddy allocator to function correctly.
7278 end = pgdat_end_pfn(pgdat);
7279 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7280 size = (end - start) * sizeof(struct page);
7281 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7284 panic("Failed to allocate %ld bytes for node %d memory map\n",
7285 size, pgdat->node_id);
7286 pgdat->node_mem_map = map + offset;
7288 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7289 __func__, pgdat->node_id, (unsigned long)pgdat,
7290 (unsigned long)pgdat->node_mem_map);
7291 #ifndef CONFIG_NEED_MULTIPLE_NODES
7293 * With no DISCONTIG, the global mem_map is just set as node 0's
7295 if (pgdat == NODE_DATA(0)) {
7296 mem_map = NODE_DATA(0)->node_mem_map;
7297 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7303 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7304 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7306 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7307 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7309 pgdat->first_deferred_pfn = ULONG_MAX;
7312 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7315 static void __init free_area_init_node(int nid)
7317 pg_data_t *pgdat = NODE_DATA(nid);
7318 unsigned long start_pfn = 0;
7319 unsigned long end_pfn = 0;
7321 /* pg_data_t should be reset to zero when it's allocated */
7322 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7324 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7326 pgdat->node_id = nid;
7327 pgdat->node_start_pfn = start_pfn;
7328 pgdat->per_cpu_nodestats = NULL;
7330 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7331 (u64)start_pfn << PAGE_SHIFT,
7332 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7333 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7335 alloc_node_mem_map(pgdat);
7336 pgdat_set_deferred_range(pgdat);
7338 free_area_init_core(pgdat);
7341 void __init free_area_init_memoryless_node(int nid)
7343 free_area_init_node(nid);
7346 #if MAX_NUMNODES > 1
7348 * Figure out the number of possible node ids.
7350 void __init setup_nr_node_ids(void)
7352 unsigned int highest;
7354 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7355 nr_node_ids = highest + 1;
7360 * node_map_pfn_alignment - determine the maximum internode alignment
7362 * This function should be called after node map is populated and sorted.
7363 * It calculates the maximum power of two alignment which can distinguish
7366 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7367 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7368 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7369 * shifted, 1GiB is enough and this function will indicate so.
7371 * This is used to test whether pfn -> nid mapping of the chosen memory
7372 * model has fine enough granularity to avoid incorrect mapping for the
7373 * populated node map.
7375 * Return: the determined alignment in pfn's. 0 if there is no alignment
7376 * requirement (single node).
7378 unsigned long __init node_map_pfn_alignment(void)
7380 unsigned long accl_mask = 0, last_end = 0;
7381 unsigned long start, end, mask;
7382 int last_nid = NUMA_NO_NODE;
7385 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7386 if (!start || last_nid < 0 || last_nid == nid) {
7393 * Start with a mask granular enough to pin-point to the
7394 * start pfn and tick off bits one-by-one until it becomes
7395 * too coarse to separate the current node from the last.
7397 mask = ~((1 << __ffs(start)) - 1);
7398 while (mask && last_end <= (start & (mask << 1)))
7401 /* accumulate all internode masks */
7405 /* convert mask to number of pages */
7406 return ~accl_mask + 1;
7410 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7412 * Return: the minimum PFN based on information provided via
7413 * memblock_set_node().
7415 unsigned long __init find_min_pfn_with_active_regions(void)
7417 return PHYS_PFN(memblock_start_of_DRAM());
7421 * early_calculate_totalpages()
7422 * Sum pages in active regions for movable zone.
7423 * Populate N_MEMORY for calculating usable_nodes.
7425 static unsigned long __init early_calculate_totalpages(void)
7427 unsigned long totalpages = 0;
7428 unsigned long start_pfn, end_pfn;
7431 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7432 unsigned long pages = end_pfn - start_pfn;
7434 totalpages += pages;
7436 node_set_state(nid, N_MEMORY);
7442 * Find the PFN the Movable zone begins in each node. Kernel memory
7443 * is spread evenly between nodes as long as the nodes have enough
7444 * memory. When they don't, some nodes will have more kernelcore than
7447 static void __init find_zone_movable_pfns_for_nodes(void)
7450 unsigned long usable_startpfn;
7451 unsigned long kernelcore_node, kernelcore_remaining;
7452 /* save the state before borrow the nodemask */
7453 nodemask_t saved_node_state = node_states[N_MEMORY];
7454 unsigned long totalpages = early_calculate_totalpages();
7455 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7456 struct memblock_region *r;
7458 /* Need to find movable_zone earlier when movable_node is specified. */
7459 find_usable_zone_for_movable();
7462 * If movable_node is specified, ignore kernelcore and movablecore
7465 if (movable_node_is_enabled()) {
7466 for_each_mem_region(r) {
7467 if (!memblock_is_hotpluggable(r))
7470 nid = memblock_get_region_node(r);
7472 usable_startpfn = PFN_DOWN(r->base);
7473 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7474 min(usable_startpfn, zone_movable_pfn[nid]) :
7482 * If kernelcore=mirror is specified, ignore movablecore option
7484 if (mirrored_kernelcore) {
7485 bool mem_below_4gb_not_mirrored = false;
7487 for_each_mem_region(r) {
7488 if (memblock_is_mirror(r))
7491 nid = memblock_get_region_node(r);
7493 usable_startpfn = memblock_region_memory_base_pfn(r);
7495 if (usable_startpfn < 0x100000) {
7496 mem_below_4gb_not_mirrored = true;
7500 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7501 min(usable_startpfn, zone_movable_pfn[nid]) :
7505 if (mem_below_4gb_not_mirrored)
7506 pr_warn("This configuration results in unmirrored kernel memory.\n");
7512 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7513 * amount of necessary memory.
7515 if (required_kernelcore_percent)
7516 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7518 if (required_movablecore_percent)
7519 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7523 * If movablecore= was specified, calculate what size of
7524 * kernelcore that corresponds so that memory usable for
7525 * any allocation type is evenly spread. If both kernelcore
7526 * and movablecore are specified, then the value of kernelcore
7527 * will be used for required_kernelcore if it's greater than
7528 * what movablecore would have allowed.
7530 if (required_movablecore) {
7531 unsigned long corepages;
7534 * Round-up so that ZONE_MOVABLE is at least as large as what
7535 * was requested by the user
7537 required_movablecore =
7538 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7539 required_movablecore = min(totalpages, required_movablecore);
7540 corepages = totalpages - required_movablecore;
7542 required_kernelcore = max(required_kernelcore, corepages);
7546 * If kernelcore was not specified or kernelcore size is larger
7547 * than totalpages, there is no ZONE_MOVABLE.
7549 if (!required_kernelcore || required_kernelcore >= totalpages)
7552 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7553 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7556 /* Spread kernelcore memory as evenly as possible throughout nodes */
7557 kernelcore_node = required_kernelcore / usable_nodes;
7558 for_each_node_state(nid, N_MEMORY) {
7559 unsigned long start_pfn, end_pfn;
7562 * Recalculate kernelcore_node if the division per node
7563 * now exceeds what is necessary to satisfy the requested
7564 * amount of memory for the kernel
7566 if (required_kernelcore < kernelcore_node)
7567 kernelcore_node = required_kernelcore / usable_nodes;
7570 * As the map is walked, we track how much memory is usable
7571 * by the kernel using kernelcore_remaining. When it is
7572 * 0, the rest of the node is usable by ZONE_MOVABLE
7574 kernelcore_remaining = kernelcore_node;
7576 /* Go through each range of PFNs within this node */
7577 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7578 unsigned long size_pages;
7580 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7581 if (start_pfn >= end_pfn)
7584 /* Account for what is only usable for kernelcore */
7585 if (start_pfn < usable_startpfn) {
7586 unsigned long kernel_pages;
7587 kernel_pages = min(end_pfn, usable_startpfn)
7590 kernelcore_remaining -= min(kernel_pages,
7591 kernelcore_remaining);
7592 required_kernelcore -= min(kernel_pages,
7593 required_kernelcore);
7595 /* Continue if range is now fully accounted */
7596 if (end_pfn <= usable_startpfn) {
7599 * Push zone_movable_pfn to the end so
7600 * that if we have to rebalance
7601 * kernelcore across nodes, we will
7602 * not double account here
7604 zone_movable_pfn[nid] = end_pfn;
7607 start_pfn = usable_startpfn;
7611 * The usable PFN range for ZONE_MOVABLE is from
7612 * start_pfn->end_pfn. Calculate size_pages as the
7613 * number of pages used as kernelcore
7615 size_pages = end_pfn - start_pfn;
7616 if (size_pages > kernelcore_remaining)
7617 size_pages = kernelcore_remaining;
7618 zone_movable_pfn[nid] = start_pfn + size_pages;
7621 * Some kernelcore has been met, update counts and
7622 * break if the kernelcore for this node has been
7625 required_kernelcore -= min(required_kernelcore,
7627 kernelcore_remaining -= size_pages;
7628 if (!kernelcore_remaining)
7634 * If there is still required_kernelcore, we do another pass with one
7635 * less node in the count. This will push zone_movable_pfn[nid] further
7636 * along on the nodes that still have memory until kernelcore is
7640 if (usable_nodes && required_kernelcore > usable_nodes)
7644 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7645 for (nid = 0; nid < MAX_NUMNODES; nid++)
7646 zone_movable_pfn[nid] =
7647 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7650 /* restore the node_state */
7651 node_states[N_MEMORY] = saved_node_state;
7654 /* Any regular or high memory on that node ? */
7655 static void check_for_memory(pg_data_t *pgdat, int nid)
7657 enum zone_type zone_type;
7659 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7660 struct zone *zone = &pgdat->node_zones[zone_type];
7661 if (populated_zone(zone)) {
7662 if (IS_ENABLED(CONFIG_HIGHMEM))
7663 node_set_state(nid, N_HIGH_MEMORY);
7664 if (zone_type <= ZONE_NORMAL)
7665 node_set_state(nid, N_NORMAL_MEMORY);
7672 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7673 * such cases we allow max_zone_pfn sorted in the descending order
7675 bool __weak arch_has_descending_max_zone_pfns(void)
7681 * free_area_init - Initialise all pg_data_t and zone data
7682 * @max_zone_pfn: an array of max PFNs for each zone
7684 * This will call free_area_init_node() for each active node in the system.
7685 * Using the page ranges provided by memblock_set_node(), the size of each
7686 * zone in each node and their holes is calculated. If the maximum PFN
7687 * between two adjacent zones match, it is assumed that the zone is empty.
7688 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7689 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7690 * starts where the previous one ended. For example, ZONE_DMA32 starts
7691 * at arch_max_dma_pfn.
7693 void __init free_area_init(unsigned long *max_zone_pfn)
7695 unsigned long start_pfn, end_pfn;
7699 /* Record where the zone boundaries are */
7700 memset(arch_zone_lowest_possible_pfn, 0,
7701 sizeof(arch_zone_lowest_possible_pfn));
7702 memset(arch_zone_highest_possible_pfn, 0,
7703 sizeof(arch_zone_highest_possible_pfn));
7705 start_pfn = find_min_pfn_with_active_regions();
7706 descending = arch_has_descending_max_zone_pfns();
7708 for (i = 0; i < MAX_NR_ZONES; i++) {
7710 zone = MAX_NR_ZONES - i - 1;
7714 if (zone == ZONE_MOVABLE)
7717 end_pfn = max(max_zone_pfn[zone], start_pfn);
7718 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7719 arch_zone_highest_possible_pfn[zone] = end_pfn;
7721 start_pfn = end_pfn;
7724 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7725 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7726 find_zone_movable_pfns_for_nodes();
7728 /* Print out the zone ranges */
7729 pr_info("Zone ranges:\n");
7730 for (i = 0; i < MAX_NR_ZONES; i++) {
7731 if (i == ZONE_MOVABLE)
7733 pr_info(" %-8s ", zone_names[i]);
7734 if (arch_zone_lowest_possible_pfn[i] ==
7735 arch_zone_highest_possible_pfn[i])
7738 pr_cont("[mem %#018Lx-%#018Lx]\n",
7739 (u64)arch_zone_lowest_possible_pfn[i]
7741 ((u64)arch_zone_highest_possible_pfn[i]
7742 << PAGE_SHIFT) - 1);
7745 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7746 pr_info("Movable zone start for each node\n");
7747 for (i = 0; i < MAX_NUMNODES; i++) {
7748 if (zone_movable_pfn[i])
7749 pr_info(" Node %d: %#018Lx\n", i,
7750 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7754 * Print out the early node map, and initialize the
7755 * subsection-map relative to active online memory ranges to
7756 * enable future "sub-section" extensions of the memory map.
7758 pr_info("Early memory node ranges\n");
7759 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7760 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7761 (u64)start_pfn << PAGE_SHIFT,
7762 ((u64)end_pfn << PAGE_SHIFT) - 1);
7763 subsection_map_init(start_pfn, end_pfn - start_pfn);
7766 /* Initialise every node */
7767 mminit_verify_pageflags_layout();
7768 setup_nr_node_ids();
7769 for_each_online_node(nid) {
7770 pg_data_t *pgdat = NODE_DATA(nid);
7771 free_area_init_node(nid);
7773 /* Any memory on that node */
7774 if (pgdat->node_present_pages)
7775 node_set_state(nid, N_MEMORY);
7776 check_for_memory(pgdat, nid);
7780 static int __init cmdline_parse_core(char *p, unsigned long *core,
7781 unsigned long *percent)
7783 unsigned long long coremem;
7789 /* Value may be a percentage of total memory, otherwise bytes */
7790 coremem = simple_strtoull(p, &endptr, 0);
7791 if (*endptr == '%') {
7792 /* Paranoid check for percent values greater than 100 */
7793 WARN_ON(coremem > 100);
7797 coremem = memparse(p, &p);
7798 /* Paranoid check that UL is enough for the coremem value */
7799 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7801 *core = coremem >> PAGE_SHIFT;
7808 * kernelcore=size sets the amount of memory for use for allocations that
7809 * cannot be reclaimed or migrated.
7811 static int __init cmdline_parse_kernelcore(char *p)
7813 /* parse kernelcore=mirror */
7814 if (parse_option_str(p, "mirror")) {
7815 mirrored_kernelcore = true;
7819 return cmdline_parse_core(p, &required_kernelcore,
7820 &required_kernelcore_percent);
7824 * movablecore=size sets the amount of memory for use for allocations that
7825 * can be reclaimed or migrated.
7827 static int __init cmdline_parse_movablecore(char *p)
7829 return cmdline_parse_core(p, &required_movablecore,
7830 &required_movablecore_percent);
7833 early_param("kernelcore", cmdline_parse_kernelcore);
7834 early_param("movablecore", cmdline_parse_movablecore);
7836 void adjust_managed_page_count(struct page *page, long count)
7838 atomic_long_add(count, &page_zone(page)->managed_pages);
7839 totalram_pages_add(count);
7840 #ifdef CONFIG_HIGHMEM
7841 if (PageHighMem(page))
7842 totalhigh_pages_add(count);
7845 EXPORT_SYMBOL(adjust_managed_page_count);
7847 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7850 unsigned long pages = 0;
7852 start = (void *)PAGE_ALIGN((unsigned long)start);
7853 end = (void *)((unsigned long)end & PAGE_MASK);
7854 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7855 struct page *page = virt_to_page(pos);
7856 void *direct_map_addr;
7859 * 'direct_map_addr' might be different from 'pos'
7860 * because some architectures' virt_to_page()
7861 * work with aliases. Getting the direct map
7862 * address ensures that we get a _writeable_
7863 * alias for the memset().
7865 direct_map_addr = page_address(page);
7867 * Perform a kasan-unchecked memset() since this memory
7868 * has not been initialized.
7870 direct_map_addr = kasan_reset_tag(direct_map_addr);
7871 if ((unsigned int)poison <= 0xFF)
7872 memset(direct_map_addr, poison, PAGE_SIZE);
7874 free_reserved_page(page);
7878 pr_info("Freeing %s memory: %ldK\n",
7879 s, pages << (PAGE_SHIFT - 10));
7884 void __init mem_init_print_info(void)
7886 unsigned long physpages, codesize, datasize, rosize, bss_size;
7887 unsigned long init_code_size, init_data_size;
7889 physpages = get_num_physpages();
7890 codesize = _etext - _stext;
7891 datasize = _edata - _sdata;
7892 rosize = __end_rodata - __start_rodata;
7893 bss_size = __bss_stop - __bss_start;
7894 init_data_size = __init_end - __init_begin;
7895 init_code_size = _einittext - _sinittext;
7898 * Detect special cases and adjust section sizes accordingly:
7899 * 1) .init.* may be embedded into .data sections
7900 * 2) .init.text.* may be out of [__init_begin, __init_end],
7901 * please refer to arch/tile/kernel/vmlinux.lds.S.
7902 * 3) .rodata.* may be embedded into .text or .data sections.
7904 #define adj_init_size(start, end, size, pos, adj) \
7906 if (start <= pos && pos < end && size > adj) \
7910 adj_init_size(__init_begin, __init_end, init_data_size,
7911 _sinittext, init_code_size);
7912 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7913 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7914 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7915 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7917 #undef adj_init_size
7919 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7920 #ifdef CONFIG_HIGHMEM
7924 nr_free_pages() << (PAGE_SHIFT - 10),
7925 physpages << (PAGE_SHIFT - 10),
7926 codesize >> 10, datasize >> 10, rosize >> 10,
7927 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7928 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7929 totalcma_pages << (PAGE_SHIFT - 10)
7930 #ifdef CONFIG_HIGHMEM
7931 , totalhigh_pages() << (PAGE_SHIFT - 10)
7937 * set_dma_reserve - set the specified number of pages reserved in the first zone
7938 * @new_dma_reserve: The number of pages to mark reserved
7940 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7941 * In the DMA zone, a significant percentage may be consumed by kernel image
7942 * and other unfreeable allocations which can skew the watermarks badly. This
7943 * function may optionally be used to account for unfreeable pages in the
7944 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7945 * smaller per-cpu batchsize.
7947 void __init set_dma_reserve(unsigned long new_dma_reserve)
7949 dma_reserve = new_dma_reserve;
7952 static int page_alloc_cpu_dead(unsigned int cpu)
7955 lru_add_drain_cpu(cpu);
7959 * Spill the event counters of the dead processor
7960 * into the current processors event counters.
7961 * This artificially elevates the count of the current
7964 vm_events_fold_cpu(cpu);
7967 * Zero the differential counters of the dead processor
7968 * so that the vm statistics are consistent.
7970 * This is only okay since the processor is dead and cannot
7971 * race with what we are doing.
7973 cpu_vm_stats_fold(cpu);
7978 int hashdist = HASHDIST_DEFAULT;
7980 static int __init set_hashdist(char *str)
7984 hashdist = simple_strtoul(str, &str, 0);
7987 __setup("hashdist=", set_hashdist);
7990 void __init page_alloc_init(void)
7995 if (num_node_state(N_MEMORY) == 1)
7999 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
8000 "mm/page_alloc:dead", NULL,
8001 page_alloc_cpu_dead);
8006 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8007 * or min_free_kbytes changes.
8009 static void calculate_totalreserve_pages(void)
8011 struct pglist_data *pgdat;
8012 unsigned long reserve_pages = 0;
8013 enum zone_type i, j;
8015 for_each_online_pgdat(pgdat) {
8017 pgdat->totalreserve_pages = 0;
8019 for (i = 0; i < MAX_NR_ZONES; i++) {
8020 struct zone *zone = pgdat->node_zones + i;
8022 unsigned long managed_pages = zone_managed_pages(zone);
8024 /* Find valid and maximum lowmem_reserve in the zone */
8025 for (j = i; j < MAX_NR_ZONES; j++) {
8026 if (zone->lowmem_reserve[j] > max)
8027 max = zone->lowmem_reserve[j];
8030 /* we treat the high watermark as reserved pages. */
8031 max += high_wmark_pages(zone);
8033 if (max > managed_pages)
8034 max = managed_pages;
8036 pgdat->totalreserve_pages += max;
8038 reserve_pages += max;
8041 totalreserve_pages = reserve_pages;
8045 * setup_per_zone_lowmem_reserve - called whenever
8046 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8047 * has a correct pages reserved value, so an adequate number of
8048 * pages are left in the zone after a successful __alloc_pages().
8050 static void setup_per_zone_lowmem_reserve(void)
8052 struct pglist_data *pgdat;
8053 enum zone_type i, j;
8055 for_each_online_pgdat(pgdat) {
8056 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8057 struct zone *zone = &pgdat->node_zones[i];
8058 int ratio = sysctl_lowmem_reserve_ratio[i];
8059 bool clear = !ratio || !zone_managed_pages(zone);
8060 unsigned long managed_pages = 0;
8062 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8064 zone->lowmem_reserve[j] = 0;
8066 struct zone *upper_zone = &pgdat->node_zones[j];
8068 managed_pages += zone_managed_pages(upper_zone);
8069 zone->lowmem_reserve[j] = managed_pages / ratio;
8075 /* update totalreserve_pages */
8076 calculate_totalreserve_pages();
8079 static void __setup_per_zone_wmarks(void)
8081 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8082 unsigned long lowmem_pages = 0;
8084 unsigned long flags;
8086 /* Calculate total number of !ZONE_HIGHMEM pages */
8087 for_each_zone(zone) {
8088 if (!is_highmem(zone))
8089 lowmem_pages += zone_managed_pages(zone);
8092 for_each_zone(zone) {
8095 spin_lock_irqsave(&zone->lock, flags);
8096 tmp = (u64)pages_min * zone_managed_pages(zone);
8097 do_div(tmp, lowmem_pages);
8098 if (is_highmem(zone)) {
8100 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8101 * need highmem pages, so cap pages_min to a small
8104 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8105 * deltas control async page reclaim, and so should
8106 * not be capped for highmem.
8108 unsigned long min_pages;
8110 min_pages = zone_managed_pages(zone) / 1024;
8111 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8112 zone->_watermark[WMARK_MIN] = min_pages;
8115 * If it's a lowmem zone, reserve a number of pages
8116 * proportionate to the zone's size.
8118 zone->_watermark[WMARK_MIN] = tmp;
8122 * Set the kswapd watermarks distance according to the
8123 * scale factor in proportion to available memory, but
8124 * ensure a minimum size on small systems.
8126 tmp = max_t(u64, tmp >> 2,
8127 mult_frac(zone_managed_pages(zone),
8128 watermark_scale_factor, 10000));
8130 zone->watermark_boost = 0;
8131 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8132 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8134 spin_unlock_irqrestore(&zone->lock, flags);
8137 /* update totalreserve_pages */
8138 calculate_totalreserve_pages();
8142 * setup_per_zone_wmarks - called when min_free_kbytes changes
8143 * or when memory is hot-{added|removed}
8145 * Ensures that the watermark[min,low,high] values for each zone are set
8146 * correctly with respect to min_free_kbytes.
8148 void setup_per_zone_wmarks(void)
8150 static DEFINE_SPINLOCK(lock);
8153 __setup_per_zone_wmarks();
8158 * Initialise min_free_kbytes.
8160 * For small machines we want it small (128k min). For large machines
8161 * we want it large (256MB max). But it is not linear, because network
8162 * bandwidth does not increase linearly with machine size. We use
8164 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8165 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8181 int __meminit init_per_zone_wmark_min(void)
8183 unsigned long lowmem_kbytes;
8184 int new_min_free_kbytes;
8186 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8187 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8189 if (new_min_free_kbytes > user_min_free_kbytes) {
8190 min_free_kbytes = new_min_free_kbytes;
8191 if (min_free_kbytes < 128)
8192 min_free_kbytes = 128;
8193 if (min_free_kbytes > 262144)
8194 min_free_kbytes = 262144;
8196 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8197 new_min_free_kbytes, user_min_free_kbytes);
8199 setup_per_zone_wmarks();
8200 refresh_zone_stat_thresholds();
8201 setup_per_zone_lowmem_reserve();
8204 setup_min_unmapped_ratio();
8205 setup_min_slab_ratio();
8208 khugepaged_min_free_kbytes_update();
8212 postcore_initcall(init_per_zone_wmark_min)
8215 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8216 * that we can call two helper functions whenever min_free_kbytes
8219 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8220 void *buffer, size_t *length, loff_t *ppos)
8224 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8229 user_min_free_kbytes = min_free_kbytes;
8230 setup_per_zone_wmarks();
8235 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8236 void *buffer, size_t *length, loff_t *ppos)
8240 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8245 setup_per_zone_wmarks();
8251 static void setup_min_unmapped_ratio(void)
8256 for_each_online_pgdat(pgdat)
8257 pgdat->min_unmapped_pages = 0;
8260 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8261 sysctl_min_unmapped_ratio) / 100;
8265 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8266 void *buffer, size_t *length, loff_t *ppos)
8270 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8274 setup_min_unmapped_ratio();
8279 static void setup_min_slab_ratio(void)
8284 for_each_online_pgdat(pgdat)
8285 pgdat->min_slab_pages = 0;
8288 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8289 sysctl_min_slab_ratio) / 100;
8292 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8293 void *buffer, size_t *length, loff_t *ppos)
8297 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8301 setup_min_slab_ratio();
8308 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8309 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8310 * whenever sysctl_lowmem_reserve_ratio changes.
8312 * The reserve ratio obviously has absolutely no relation with the
8313 * minimum watermarks. The lowmem reserve ratio can only make sense
8314 * if in function of the boot time zone sizes.
8316 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8317 void *buffer, size_t *length, loff_t *ppos)
8321 proc_dointvec_minmax(table, write, buffer, length, ppos);
8323 for (i = 0; i < MAX_NR_ZONES; i++) {
8324 if (sysctl_lowmem_reserve_ratio[i] < 1)
8325 sysctl_lowmem_reserve_ratio[i] = 0;
8328 setup_per_zone_lowmem_reserve();
8333 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8334 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8335 * pagelist can have before it gets flushed back to buddy allocator.
8337 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8338 void *buffer, size_t *length, loff_t *ppos)
8341 int old_percpu_pagelist_fraction;
8344 mutex_lock(&pcp_batch_high_lock);
8345 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8347 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8348 if (!write || ret < 0)
8351 /* Sanity checking to avoid pcp imbalance */
8352 if (percpu_pagelist_fraction &&
8353 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8354 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8360 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8363 for_each_populated_zone(zone)
8364 zone_set_pageset_high_and_batch(zone);
8366 mutex_unlock(&pcp_batch_high_lock);
8370 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8372 * Returns the number of pages that arch has reserved but
8373 * is not known to alloc_large_system_hash().
8375 static unsigned long __init arch_reserved_kernel_pages(void)
8382 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8383 * machines. As memory size is increased the scale is also increased but at
8384 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8385 * quadruples the scale is increased by one, which means the size of hash table
8386 * only doubles, instead of quadrupling as well.
8387 * Because 32-bit systems cannot have large physical memory, where this scaling
8388 * makes sense, it is disabled on such platforms.
8390 #if __BITS_PER_LONG > 32
8391 #define ADAPT_SCALE_BASE (64ul << 30)
8392 #define ADAPT_SCALE_SHIFT 2
8393 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8397 * allocate a large system hash table from bootmem
8398 * - it is assumed that the hash table must contain an exact power-of-2
8399 * quantity of entries
8400 * - limit is the number of hash buckets, not the total allocation size
8402 void *__init alloc_large_system_hash(const char *tablename,
8403 unsigned long bucketsize,
8404 unsigned long numentries,
8407 unsigned int *_hash_shift,
8408 unsigned int *_hash_mask,
8409 unsigned long low_limit,
8410 unsigned long high_limit)
8412 unsigned long long max = high_limit;
8413 unsigned long log2qty, size;
8419 /* allow the kernel cmdline to have a say */
8421 /* round applicable memory size up to nearest megabyte */
8422 numentries = nr_kernel_pages;
8423 numentries -= arch_reserved_kernel_pages();
8425 /* It isn't necessary when PAGE_SIZE >= 1MB */
8426 if (PAGE_SHIFT < 20)
8427 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8429 #if __BITS_PER_LONG > 32
8431 unsigned long adapt;
8433 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8434 adapt <<= ADAPT_SCALE_SHIFT)
8439 /* limit to 1 bucket per 2^scale bytes of low memory */
8440 if (scale > PAGE_SHIFT)
8441 numentries >>= (scale - PAGE_SHIFT);
8443 numentries <<= (PAGE_SHIFT - scale);
8445 /* Make sure we've got at least a 0-order allocation.. */
8446 if (unlikely(flags & HASH_SMALL)) {
8447 /* Makes no sense without HASH_EARLY */
8448 WARN_ON(!(flags & HASH_EARLY));
8449 if (!(numentries >> *_hash_shift)) {
8450 numentries = 1UL << *_hash_shift;
8451 BUG_ON(!numentries);
8453 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8454 numentries = PAGE_SIZE / bucketsize;
8456 numentries = roundup_pow_of_two(numentries);
8458 /* limit allocation size to 1/16 total memory by default */
8460 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8461 do_div(max, bucketsize);
8463 max = min(max, 0x80000000ULL);
8465 if (numentries < low_limit)
8466 numentries = low_limit;
8467 if (numentries > max)
8470 log2qty = ilog2(numentries);
8472 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8475 size = bucketsize << log2qty;
8476 if (flags & HASH_EARLY) {
8477 if (flags & HASH_ZERO)
8478 table = memblock_alloc(size, SMP_CACHE_BYTES);
8480 table = memblock_alloc_raw(size,
8482 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8483 table = __vmalloc(size, gfp_flags);
8485 huge = is_vm_area_hugepages(table);
8488 * If bucketsize is not a power-of-two, we may free
8489 * some pages at the end of hash table which
8490 * alloc_pages_exact() automatically does
8492 table = alloc_pages_exact(size, gfp_flags);
8493 kmemleak_alloc(table, size, 1, gfp_flags);
8495 } while (!table && size > PAGE_SIZE && --log2qty);
8498 panic("Failed to allocate %s hash table\n", tablename);
8500 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8501 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8502 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8505 *_hash_shift = log2qty;
8507 *_hash_mask = (1 << log2qty) - 1;
8513 * This function checks whether pageblock includes unmovable pages or not.
8515 * PageLRU check without isolation or lru_lock could race so that
8516 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8517 * check without lock_page also may miss some movable non-lru pages at
8518 * race condition. So you can't expect this function should be exact.
8520 * Returns a page without holding a reference. If the caller wants to
8521 * dereference that page (e.g., dumping), it has to make sure that it
8522 * cannot get removed (e.g., via memory unplug) concurrently.
8525 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8526 int migratetype, int flags)
8528 unsigned long iter = 0;
8529 unsigned long pfn = page_to_pfn(page);
8530 unsigned long offset = pfn % pageblock_nr_pages;
8532 if (is_migrate_cma_page(page)) {
8534 * CMA allocations (alloc_contig_range) really need to mark
8535 * isolate CMA pageblocks even when they are not movable in fact
8536 * so consider them movable here.
8538 if (is_migrate_cma(migratetype))
8544 for (; iter < pageblock_nr_pages - offset; iter++) {
8545 if (!pfn_valid_within(pfn + iter))
8548 page = pfn_to_page(pfn + iter);
8551 * Both, bootmem allocations and memory holes are marked
8552 * PG_reserved and are unmovable. We can even have unmovable
8553 * allocations inside ZONE_MOVABLE, for example when
8554 * specifying "movablecore".
8556 if (PageReserved(page))
8560 * If the zone is movable and we have ruled out all reserved
8561 * pages then it should be reasonably safe to assume the rest
8564 if (zone_idx(zone) == ZONE_MOVABLE)
8568 * Hugepages are not in LRU lists, but they're movable.
8569 * THPs are on the LRU, but need to be counted as #small pages.
8570 * We need not scan over tail pages because we don't
8571 * handle each tail page individually in migration.
8573 if (PageHuge(page) || PageTransCompound(page)) {
8574 struct page *head = compound_head(page);
8575 unsigned int skip_pages;
8577 if (PageHuge(page)) {
8578 if (!hugepage_migration_supported(page_hstate(head)))
8580 } else if (!PageLRU(head) && !__PageMovable(head)) {
8584 skip_pages = compound_nr(head) - (page - head);
8585 iter += skip_pages - 1;
8590 * We can't use page_count without pin a page
8591 * because another CPU can free compound page.
8592 * This check already skips compound tails of THP
8593 * because their page->_refcount is zero at all time.
8595 if (!page_ref_count(page)) {
8596 if (PageBuddy(page))
8597 iter += (1 << buddy_order(page)) - 1;
8602 * The HWPoisoned page may be not in buddy system, and
8603 * page_count() is not 0.
8605 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8609 * We treat all PageOffline() pages as movable when offlining
8610 * to give drivers a chance to decrement their reference count
8611 * in MEM_GOING_OFFLINE in order to indicate that these pages
8612 * can be offlined as there are no direct references anymore.
8613 * For actually unmovable PageOffline() where the driver does
8614 * not support this, we will fail later when trying to actually
8615 * move these pages that still have a reference count > 0.
8616 * (false negatives in this function only)
8618 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8621 if (__PageMovable(page) || PageLRU(page))
8625 * If there are RECLAIMABLE pages, we need to check
8626 * it. But now, memory offline itself doesn't call
8627 * shrink_node_slabs() and it still to be fixed.
8634 #ifdef CONFIG_CONTIG_ALLOC
8635 static unsigned long pfn_max_align_down(unsigned long pfn)
8637 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8638 pageblock_nr_pages) - 1);
8641 static unsigned long pfn_max_align_up(unsigned long pfn)
8643 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8644 pageblock_nr_pages));
8647 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8648 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8649 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8650 static void alloc_contig_dump_pages(struct list_head *page_list)
8652 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8654 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8658 list_for_each_entry(page, page_list, lru)
8659 dump_page(page, "migration failure");
8663 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8668 /* [start, end) must belong to a single zone. */
8669 static int __alloc_contig_migrate_range(struct compact_control *cc,
8670 unsigned long start, unsigned long end)
8672 /* This function is based on compact_zone() from compaction.c. */
8673 unsigned int nr_reclaimed;
8674 unsigned long pfn = start;
8675 unsigned int tries = 0;
8677 struct migration_target_control mtc = {
8678 .nid = zone_to_nid(cc->zone),
8679 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8684 while (pfn < end || !list_empty(&cc->migratepages)) {
8685 if (fatal_signal_pending(current)) {
8690 if (list_empty(&cc->migratepages)) {
8691 cc->nr_migratepages = 0;
8692 pfn = isolate_migratepages_range(cc, pfn, end);
8698 } else if (++tries == 5) {
8699 ret = ret < 0 ? ret : -EBUSY;
8703 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8705 cc->nr_migratepages -= nr_reclaimed;
8707 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8708 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8711 alloc_contig_dump_pages(&cc->migratepages);
8712 putback_movable_pages(&cc->migratepages);
8719 * alloc_contig_range() -- tries to allocate given range of pages
8720 * @start: start PFN to allocate
8721 * @end: one-past-the-last PFN to allocate
8722 * @migratetype: migratetype of the underlaying pageblocks (either
8723 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8724 * in range must have the same migratetype and it must
8725 * be either of the two.
8726 * @gfp_mask: GFP mask to use during compaction
8728 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8729 * aligned. The PFN range must belong to a single zone.
8731 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8732 * pageblocks in the range. Once isolated, the pageblocks should not
8733 * be modified by others.
8735 * Return: zero on success or negative error code. On success all
8736 * pages which PFN is in [start, end) are allocated for the caller and
8737 * need to be freed with free_contig_range().
8739 int alloc_contig_range(unsigned long start, unsigned long end,
8740 unsigned migratetype, gfp_t gfp_mask)
8742 unsigned long outer_start, outer_end;
8746 struct compact_control cc = {
8747 .nr_migratepages = 0,
8749 .zone = page_zone(pfn_to_page(start)),
8750 .mode = MIGRATE_SYNC,
8751 .ignore_skip_hint = true,
8752 .no_set_skip_hint = true,
8753 .gfp_mask = current_gfp_context(gfp_mask),
8754 .alloc_contig = true,
8756 INIT_LIST_HEAD(&cc.migratepages);
8759 * What we do here is we mark all pageblocks in range as
8760 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8761 * have different sizes, and due to the way page allocator
8762 * work, we align the range to biggest of the two pages so
8763 * that page allocator won't try to merge buddies from
8764 * different pageblocks and change MIGRATE_ISOLATE to some
8765 * other migration type.
8767 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8768 * migrate the pages from an unaligned range (ie. pages that
8769 * we are interested in). This will put all the pages in
8770 * range back to page allocator as MIGRATE_ISOLATE.
8772 * When this is done, we take the pages in range from page
8773 * allocator removing them from the buddy system. This way
8774 * page allocator will never consider using them.
8776 * This lets us mark the pageblocks back as
8777 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8778 * aligned range but not in the unaligned, original range are
8779 * put back to page allocator so that buddy can use them.
8782 ret = start_isolate_page_range(pfn_max_align_down(start),
8783 pfn_max_align_up(end), migratetype, 0);
8787 drain_all_pages(cc.zone);
8790 * In case of -EBUSY, we'd like to know which page causes problem.
8791 * So, just fall through. test_pages_isolated() has a tracepoint
8792 * which will report the busy page.
8794 * It is possible that busy pages could become available before
8795 * the call to test_pages_isolated, and the range will actually be
8796 * allocated. So, if we fall through be sure to clear ret so that
8797 * -EBUSY is not accidentally used or returned to caller.
8799 ret = __alloc_contig_migrate_range(&cc, start, end);
8800 if (ret && ret != -EBUSY)
8805 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8806 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8807 * more, all pages in [start, end) are free in page allocator.
8808 * What we are going to do is to allocate all pages from
8809 * [start, end) (that is remove them from page allocator).
8811 * The only problem is that pages at the beginning and at the
8812 * end of interesting range may be not aligned with pages that
8813 * page allocator holds, ie. they can be part of higher order
8814 * pages. Because of this, we reserve the bigger range and
8815 * once this is done free the pages we are not interested in.
8817 * We don't have to hold zone->lock here because the pages are
8818 * isolated thus they won't get removed from buddy.
8822 outer_start = start;
8823 while (!PageBuddy(pfn_to_page(outer_start))) {
8824 if (++order >= MAX_ORDER) {
8825 outer_start = start;
8828 outer_start &= ~0UL << order;
8831 if (outer_start != start) {
8832 order = buddy_order(pfn_to_page(outer_start));
8835 * outer_start page could be small order buddy page and
8836 * it doesn't include start page. Adjust outer_start
8837 * in this case to report failed page properly
8838 * on tracepoint in test_pages_isolated()
8840 if (outer_start + (1UL << order) <= start)
8841 outer_start = start;
8844 /* Make sure the range is really isolated. */
8845 if (test_pages_isolated(outer_start, end, 0)) {
8850 /* Grab isolated pages from freelists. */
8851 outer_end = isolate_freepages_range(&cc, outer_start, end);
8857 /* Free head and tail (if any) */
8858 if (start != outer_start)
8859 free_contig_range(outer_start, start - outer_start);
8860 if (end != outer_end)
8861 free_contig_range(end, outer_end - end);
8864 undo_isolate_page_range(pfn_max_align_down(start),
8865 pfn_max_align_up(end), migratetype);
8868 EXPORT_SYMBOL(alloc_contig_range);
8870 static int __alloc_contig_pages(unsigned long start_pfn,
8871 unsigned long nr_pages, gfp_t gfp_mask)
8873 unsigned long end_pfn = start_pfn + nr_pages;
8875 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8879 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8880 unsigned long nr_pages)
8882 unsigned long i, end_pfn = start_pfn + nr_pages;
8885 for (i = start_pfn; i < end_pfn; i++) {
8886 page = pfn_to_online_page(i);
8890 if (page_zone(page) != z)
8893 if (PageReserved(page))
8896 if (page_count(page) > 0)
8905 static bool zone_spans_last_pfn(const struct zone *zone,
8906 unsigned long start_pfn, unsigned long nr_pages)
8908 unsigned long last_pfn = start_pfn + nr_pages - 1;
8910 return zone_spans_pfn(zone, last_pfn);
8914 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8915 * @nr_pages: Number of contiguous pages to allocate
8916 * @gfp_mask: GFP mask to limit search and used during compaction
8918 * @nodemask: Mask for other possible nodes
8920 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8921 * on an applicable zonelist to find a contiguous pfn range which can then be
8922 * tried for allocation with alloc_contig_range(). This routine is intended
8923 * for allocation requests which can not be fulfilled with the buddy allocator.
8925 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8926 * power of two then the alignment is guaranteed to be to the given nr_pages
8927 * (e.g. 1GB request would be aligned to 1GB).
8929 * Allocated pages can be freed with free_contig_range() or by manually calling
8930 * __free_page() on each allocated page.
8932 * Return: pointer to contiguous pages on success, or NULL if not successful.
8934 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8935 int nid, nodemask_t *nodemask)
8937 unsigned long ret, pfn, flags;
8938 struct zonelist *zonelist;
8942 zonelist = node_zonelist(nid, gfp_mask);
8943 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8944 gfp_zone(gfp_mask), nodemask) {
8945 spin_lock_irqsave(&zone->lock, flags);
8947 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8948 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8949 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8951 * We release the zone lock here because
8952 * alloc_contig_range() will also lock the zone
8953 * at some point. If there's an allocation
8954 * spinning on this lock, it may win the race
8955 * and cause alloc_contig_range() to fail...
8957 spin_unlock_irqrestore(&zone->lock, flags);
8958 ret = __alloc_contig_pages(pfn, nr_pages,
8961 return pfn_to_page(pfn);
8962 spin_lock_irqsave(&zone->lock, flags);
8966 spin_unlock_irqrestore(&zone->lock, flags);
8970 #endif /* CONFIG_CONTIG_ALLOC */
8972 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8974 unsigned int count = 0;
8976 for (; nr_pages--; pfn++) {
8977 struct page *page = pfn_to_page(pfn);
8979 count += page_count(page) != 1;
8982 WARN(count != 0, "%d pages are still in use!\n", count);
8984 EXPORT_SYMBOL(free_contig_range);
8987 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8988 * page high values need to be recalulated.
8990 void __meminit zone_pcp_update(struct zone *zone)
8992 mutex_lock(&pcp_batch_high_lock);
8993 zone_set_pageset_high_and_batch(zone);
8994 mutex_unlock(&pcp_batch_high_lock);
8998 * Effectively disable pcplists for the zone by setting the high limit to 0
8999 * and draining all cpus. A concurrent page freeing on another CPU that's about
9000 * to put the page on pcplist will either finish before the drain and the page
9001 * will be drained, or observe the new high limit and skip the pcplist.
9003 * Must be paired with a call to zone_pcp_enable().
9005 void zone_pcp_disable(struct zone *zone)
9007 mutex_lock(&pcp_batch_high_lock);
9008 __zone_set_pageset_high_and_batch(zone, 0, 1);
9009 __drain_all_pages(zone, true);
9012 void zone_pcp_enable(struct zone *zone)
9014 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9015 mutex_unlock(&pcp_batch_high_lock);
9018 void zone_pcp_reset(struct zone *zone)
9020 unsigned long flags;
9022 struct per_cpu_pageset *pset;
9024 /* avoid races with drain_pages() */
9025 local_irq_save(flags);
9026 if (zone->pageset != &boot_pageset) {
9027 for_each_online_cpu(cpu) {
9028 pset = per_cpu_ptr(zone->pageset, cpu);
9029 drain_zonestat(zone, pset);
9031 free_percpu(zone->pageset);
9032 zone->pageset = &boot_pageset;
9034 local_irq_restore(flags);
9037 #ifdef CONFIG_MEMORY_HOTREMOVE
9039 * All pages in the range must be in a single zone, must not contain holes,
9040 * must span full sections, and must be isolated before calling this function.
9042 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9044 unsigned long pfn = start_pfn;
9048 unsigned long flags;
9050 offline_mem_sections(pfn, end_pfn);
9051 zone = page_zone(pfn_to_page(pfn));
9052 spin_lock_irqsave(&zone->lock, flags);
9053 while (pfn < end_pfn) {
9054 page = pfn_to_page(pfn);
9056 * The HWPoisoned page may be not in buddy system, and
9057 * page_count() is not 0.
9059 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9064 * At this point all remaining PageOffline() pages have a
9065 * reference count of 0 and can simply be skipped.
9067 if (PageOffline(page)) {
9068 BUG_ON(page_count(page));
9069 BUG_ON(PageBuddy(page));
9074 BUG_ON(page_count(page));
9075 BUG_ON(!PageBuddy(page));
9076 order = buddy_order(page);
9077 del_page_from_free_list(page, zone, order);
9078 pfn += (1 << order);
9080 spin_unlock_irqrestore(&zone->lock, flags);
9084 bool is_free_buddy_page(struct page *page)
9086 struct zone *zone = page_zone(page);
9087 unsigned long pfn = page_to_pfn(page);
9088 unsigned long flags;
9091 spin_lock_irqsave(&zone->lock, flags);
9092 for (order = 0; order < MAX_ORDER; order++) {
9093 struct page *page_head = page - (pfn & ((1 << order) - 1));
9095 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
9098 spin_unlock_irqrestore(&zone->lock, flags);
9100 return order < MAX_ORDER;
9103 #ifdef CONFIG_MEMORY_FAILURE
9105 * Break down a higher-order page in sub-pages, and keep our target out of
9108 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9109 struct page *target, int low, int high,
9112 unsigned long size = 1 << high;
9113 struct page *current_buddy, *next_page;
9115 while (high > low) {
9119 if (target >= &page[size]) {
9120 next_page = page + size;
9121 current_buddy = page;
9124 current_buddy = page + size;
9127 if (set_page_guard(zone, current_buddy, high, migratetype))
9130 if (current_buddy != target) {
9131 add_to_free_list(current_buddy, zone, high, migratetype);
9132 set_buddy_order(current_buddy, high);
9139 * Take a page that will be marked as poisoned off the buddy allocator.
9141 bool take_page_off_buddy(struct page *page)
9143 struct zone *zone = page_zone(page);
9144 unsigned long pfn = page_to_pfn(page);
9145 unsigned long flags;
9149 spin_lock_irqsave(&zone->lock, flags);
9150 for (order = 0; order < MAX_ORDER; order++) {
9151 struct page *page_head = page - (pfn & ((1 << order) - 1));
9152 int page_order = buddy_order(page_head);
9154 if (PageBuddy(page_head) && page_order >= order) {
9155 unsigned long pfn_head = page_to_pfn(page_head);
9156 int migratetype = get_pfnblock_migratetype(page_head,
9159 del_page_from_free_list(page_head, zone, page_order);
9160 break_down_buddy_pages(zone, page_head, page, 0,
9161 page_order, migratetype);
9165 if (page_count(page_head) > 0)
9168 spin_unlock_irqrestore(&zone->lock, flags);