1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/mmu_notifier.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
71 #include <linux/psi.h>
72 #include <linux/padata.h>
73 #include <linux/khugepaged.h>
74 #include <linux/buffer_head.h>
75 #include <linux/vmalloc.h>
77 #include <asm/sections.h>
78 #include <asm/tlbflush.h>
79 #include <asm/div64.h>
82 #include "page_reporting.h"
84 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
85 typedef int __bitwise fpi_t;
87 /* No special request */
88 #define FPI_NONE ((__force fpi_t)0)
91 * Skip free page reporting notification for the (possibly merged) page.
92 * This does not hinder free page reporting from grabbing the page,
93 * reporting it and marking it "reported" - it only skips notifying
94 * the free page reporting infrastructure about a newly freed page. For
95 * example, used when temporarily pulling a page from a freelist and
96 * putting it back unmodified.
98 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
101 * Place the (possibly merged) page to the tail of the freelist. Will ignore
102 * page shuffling (relevant code - e.g., memory onlining - is expected to
103 * shuffle the whole zone).
105 * Note: No code should rely on this flag for correctness - it's purely
106 * to allow for optimizations when handing back either fresh pages
107 * (memory onlining) or untouched pages (page isolation, free page
110 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
113 * Don't poison memory with KASAN (only for the tag-based modes).
114 * During boot, all non-reserved memblock memory is exposed to page_alloc.
115 * Poisoning all that memory lengthens boot time, especially on systems with
116 * large amount of RAM. This flag is used to skip that poisoning.
117 * This is only done for the tag-based KASAN modes, as those are able to
118 * detect memory corruptions with the memory tags assigned by default.
119 * All memory allocated normally after boot gets poisoned as usual.
121 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
123 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
124 static DEFINE_MUTEX(pcp_batch_high_lock);
125 #define MIN_PERCPU_PAGELIST_FRACTION (8)
127 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
128 DEFINE_PER_CPU(int, numa_node);
129 EXPORT_PER_CPU_SYMBOL(numa_node);
132 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
134 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
136 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
137 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
138 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
139 * defined in <linux/topology.h>.
141 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
142 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
145 /* work_structs for global per-cpu drains */
148 struct work_struct work;
150 static DEFINE_MUTEX(pcpu_drain_mutex);
151 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
153 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
154 volatile unsigned long latent_entropy __latent_entropy;
155 EXPORT_SYMBOL(latent_entropy);
159 * Array of node states.
161 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
162 [N_POSSIBLE] = NODE_MASK_ALL,
163 [N_ONLINE] = { { [0] = 1UL } },
165 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
166 #ifdef CONFIG_HIGHMEM
167 [N_HIGH_MEMORY] = { { [0] = 1UL } },
169 [N_MEMORY] = { { [0] = 1UL } },
170 [N_CPU] = { { [0] = 1UL } },
173 EXPORT_SYMBOL(node_states);
175 atomic_long_t _totalram_pages __read_mostly;
176 EXPORT_SYMBOL(_totalram_pages);
177 unsigned long totalreserve_pages __read_mostly;
178 unsigned long totalcma_pages __read_mostly;
180 int percpu_pagelist_fraction;
181 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
182 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
183 EXPORT_SYMBOL(init_on_alloc);
185 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
186 EXPORT_SYMBOL(init_on_free);
188 static bool _init_on_alloc_enabled_early __read_mostly
189 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
190 static int __init early_init_on_alloc(char *buf)
193 return kstrtobool(buf, &_init_on_alloc_enabled_early);
195 early_param("init_on_alloc", early_init_on_alloc);
197 static bool _init_on_free_enabled_early __read_mostly
198 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
199 static int __init early_init_on_free(char *buf)
201 return kstrtobool(buf, &_init_on_free_enabled_early);
203 early_param("init_on_free", early_init_on_free);
206 * A cached value of the page's pageblock's migratetype, used when the page is
207 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
208 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
209 * Also the migratetype set in the page does not necessarily match the pcplist
210 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
211 * other index - this ensures that it will be put on the correct CMA freelist.
213 static inline int get_pcppage_migratetype(struct page *page)
218 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
220 page->index = migratetype;
223 #ifdef CONFIG_PM_SLEEP
225 * The following functions are used by the suspend/hibernate code to temporarily
226 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
227 * while devices are suspended. To avoid races with the suspend/hibernate code,
228 * they should always be called with system_transition_mutex held
229 * (gfp_allowed_mask also should only be modified with system_transition_mutex
230 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
231 * with that modification).
234 static gfp_t saved_gfp_mask;
236 void pm_restore_gfp_mask(void)
238 WARN_ON(!mutex_is_locked(&system_transition_mutex));
239 if (saved_gfp_mask) {
240 gfp_allowed_mask = saved_gfp_mask;
245 void pm_restrict_gfp_mask(void)
247 WARN_ON(!mutex_is_locked(&system_transition_mutex));
248 WARN_ON(saved_gfp_mask);
249 saved_gfp_mask = gfp_allowed_mask;
250 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
253 bool pm_suspended_storage(void)
255 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
259 #endif /* CONFIG_PM_SLEEP */
261 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
262 unsigned int pageblock_order __read_mostly;
265 static void __free_pages_ok(struct page *page, unsigned int order,
269 * results with 256, 32 in the lowmem_reserve sysctl:
270 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
271 * 1G machine -> (16M dma, 784M normal, 224M high)
272 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
273 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
274 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
276 * TBD: should special case ZONE_DMA32 machines here - in those we normally
277 * don't need any ZONE_NORMAL reservation
279 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
280 #ifdef CONFIG_ZONE_DMA
283 #ifdef CONFIG_ZONE_DMA32
287 #ifdef CONFIG_HIGHMEM
293 static char * const zone_names[MAX_NR_ZONES] = {
294 #ifdef CONFIG_ZONE_DMA
297 #ifdef CONFIG_ZONE_DMA32
301 #ifdef CONFIG_HIGHMEM
305 #ifdef CONFIG_ZONE_DEVICE
310 const char * const migratetype_names[MIGRATE_TYPES] = {
318 #ifdef CONFIG_MEMORY_ISOLATION
323 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
324 [NULL_COMPOUND_DTOR] = NULL,
325 [COMPOUND_PAGE_DTOR] = free_compound_page,
326 #ifdef CONFIG_HUGETLB_PAGE
327 [HUGETLB_PAGE_DTOR] = free_huge_page,
329 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
330 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
334 int min_free_kbytes = 1024;
335 int user_min_free_kbytes = -1;
336 #ifdef CONFIG_DISCONTIGMEM
338 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
339 * are not on separate NUMA nodes. Functionally this works but with
340 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
341 * quite small. By default, do not boost watermarks on discontigmem as in
342 * many cases very high-order allocations like THP are likely to be
343 * unsupported and the premature reclaim offsets the advantage of long-term
344 * fragmentation avoidance.
346 int watermark_boost_factor __read_mostly;
348 int watermark_boost_factor __read_mostly = 15000;
350 int watermark_scale_factor = 10;
352 static unsigned long nr_kernel_pages __initdata;
353 static unsigned long nr_all_pages __initdata;
354 static unsigned long dma_reserve __initdata;
356 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
357 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
358 static unsigned long required_kernelcore __initdata;
359 static unsigned long required_kernelcore_percent __initdata;
360 static unsigned long required_movablecore __initdata;
361 static unsigned long required_movablecore_percent __initdata;
362 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
363 static bool mirrored_kernelcore __meminitdata;
365 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
367 EXPORT_SYMBOL(movable_zone);
370 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
371 unsigned int nr_online_nodes __read_mostly = 1;
372 EXPORT_SYMBOL(nr_node_ids);
373 EXPORT_SYMBOL(nr_online_nodes);
376 int page_group_by_mobility_disabled __read_mostly;
378 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
380 * During boot we initialize deferred pages on-demand, as needed, but once
381 * page_alloc_init_late() has finished, the deferred pages are all initialized,
382 * and we can permanently disable that path.
384 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
387 * Calling kasan_free_pages() only after deferred memory initialization
388 * has completed. Poisoning pages during deferred memory init will greatly
389 * lengthen the process and cause problem in large memory systems as the
390 * deferred pages initialization is done with interrupt disabled.
392 * Assuming that there will be no reference to those newly initialized
393 * pages before they are ever allocated, this should have no effect on
394 * KASAN memory tracking as the poison will be properly inserted at page
395 * allocation time. The only corner case is when pages are allocated by
396 * on-demand allocation and then freed again before the deferred pages
397 * initialization is done, but this is not likely to happen.
399 static inline void kasan_free_nondeferred_pages(struct page *page, int order,
400 bool init, fpi_t fpi_flags)
402 if (static_branch_unlikely(&deferred_pages))
404 if (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
405 (fpi_flags & FPI_SKIP_KASAN_POISON))
407 kasan_free_pages(page, order, init);
410 /* Returns true if the struct page for the pfn is uninitialised */
411 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
413 int nid = early_pfn_to_nid(pfn);
415 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
422 * Returns true when the remaining initialisation should be deferred until
423 * later in the boot cycle when it can be parallelised.
425 static bool __meminit
426 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
428 static unsigned long prev_end_pfn, nr_initialised;
431 * prev_end_pfn static that contains the end of previous zone
432 * No need to protect because called very early in boot before smp_init.
434 if (prev_end_pfn != end_pfn) {
435 prev_end_pfn = end_pfn;
439 /* Always populate low zones for address-constrained allocations */
440 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
443 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
446 * We start only with one section of pages, more pages are added as
447 * needed until the rest of deferred pages are initialized.
450 if ((nr_initialised > PAGES_PER_SECTION) &&
451 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
452 NODE_DATA(nid)->first_deferred_pfn = pfn;
458 static inline void kasan_free_nondeferred_pages(struct page *page, int order,
459 bool init, fpi_t fpi_flags)
461 if (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
462 (fpi_flags & FPI_SKIP_KASAN_POISON))
464 kasan_free_pages(page, order, init);
467 static inline bool early_page_uninitialised(unsigned long pfn)
472 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
478 /* Return a pointer to the bitmap storing bits affecting a block of pages */
479 static inline unsigned long *get_pageblock_bitmap(struct page *page,
482 #ifdef CONFIG_SPARSEMEM
483 return section_to_usemap(__pfn_to_section(pfn));
485 return page_zone(page)->pageblock_flags;
486 #endif /* CONFIG_SPARSEMEM */
489 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
491 #ifdef CONFIG_SPARSEMEM
492 pfn &= (PAGES_PER_SECTION-1);
494 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
495 #endif /* CONFIG_SPARSEMEM */
496 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
499 static __always_inline
500 unsigned long __get_pfnblock_flags_mask(struct page *page,
504 unsigned long *bitmap;
505 unsigned long bitidx, word_bitidx;
508 bitmap = get_pageblock_bitmap(page, pfn);
509 bitidx = pfn_to_bitidx(page, pfn);
510 word_bitidx = bitidx / BITS_PER_LONG;
511 bitidx &= (BITS_PER_LONG-1);
513 word = bitmap[word_bitidx];
514 return (word >> bitidx) & mask;
518 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
519 * @page: The page within the block of interest
520 * @pfn: The target page frame number
521 * @mask: mask of bits that the caller is interested in
523 * Return: pageblock_bits flags
525 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
528 return __get_pfnblock_flags_mask(page, pfn, mask);
531 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
533 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
537 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
538 * @page: The page within the block of interest
539 * @flags: The flags to set
540 * @pfn: The target page frame number
541 * @mask: mask of bits that the caller is interested in
543 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
547 unsigned long *bitmap;
548 unsigned long bitidx, word_bitidx;
549 unsigned long old_word, word;
551 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
552 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
554 bitmap = get_pageblock_bitmap(page, pfn);
555 bitidx = pfn_to_bitidx(page, pfn);
556 word_bitidx = bitidx / BITS_PER_LONG;
557 bitidx &= (BITS_PER_LONG-1);
559 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
564 word = READ_ONCE(bitmap[word_bitidx]);
566 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
567 if (word == old_word)
573 void set_pageblock_migratetype(struct page *page, int migratetype)
575 if (unlikely(page_group_by_mobility_disabled &&
576 migratetype < MIGRATE_PCPTYPES))
577 migratetype = MIGRATE_UNMOVABLE;
579 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
580 page_to_pfn(page), MIGRATETYPE_MASK);
583 #ifdef CONFIG_DEBUG_VM
584 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
588 unsigned long pfn = page_to_pfn(page);
589 unsigned long sp, start_pfn;
592 seq = zone_span_seqbegin(zone);
593 start_pfn = zone->zone_start_pfn;
594 sp = zone->spanned_pages;
595 if (!zone_spans_pfn(zone, pfn))
597 } while (zone_span_seqretry(zone, seq));
600 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
601 pfn, zone_to_nid(zone), zone->name,
602 start_pfn, start_pfn + sp);
607 static int page_is_consistent(struct zone *zone, struct page *page)
609 if (!pfn_valid_within(page_to_pfn(page)))
611 if (zone != page_zone(page))
617 * Temporary debugging check for pages not lying within a given zone.
619 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
621 if (page_outside_zone_boundaries(zone, page))
623 if (!page_is_consistent(zone, page))
629 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
635 static void bad_page(struct page *page, const char *reason)
637 static unsigned long resume;
638 static unsigned long nr_shown;
639 static unsigned long nr_unshown;
642 * Allow a burst of 60 reports, then keep quiet for that minute;
643 * or allow a steady drip of one report per second.
645 if (nr_shown == 60) {
646 if (time_before(jiffies, resume)) {
652 "BUG: Bad page state: %lu messages suppressed\n",
659 resume = jiffies + 60 * HZ;
661 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
662 current->comm, page_to_pfn(page));
663 __dump_page(page, reason);
664 dump_page_owner(page);
669 /* Leave bad fields for debug, except PageBuddy could make trouble */
670 page_mapcount_reset(page); /* remove PageBuddy */
671 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
675 * Higher-order pages are called "compound pages". They are structured thusly:
677 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
679 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
680 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
682 * The first tail page's ->compound_dtor holds the offset in array of compound
683 * page destructors. See compound_page_dtors.
685 * The first tail page's ->compound_order holds the order of allocation.
686 * This usage means that zero-order pages may not be compound.
689 void free_compound_page(struct page *page)
691 mem_cgroup_uncharge(page);
692 __free_pages_ok(page, compound_order(page), FPI_NONE);
695 void prep_compound_page(struct page *page, unsigned int order)
698 int nr_pages = 1 << order;
701 for (i = 1; i < nr_pages; i++) {
702 struct page *p = page + i;
703 set_page_count(p, 0);
704 p->mapping = TAIL_MAPPING;
705 set_compound_head(p, page);
708 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
709 set_compound_order(page, order);
710 atomic_set(compound_mapcount_ptr(page), -1);
711 if (hpage_pincount_available(page))
712 atomic_set(compound_pincount_ptr(page), 0);
715 #ifdef CONFIG_DEBUG_PAGEALLOC
716 unsigned int _debug_guardpage_minorder;
718 bool _debug_pagealloc_enabled_early __read_mostly
719 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
720 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
721 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
722 EXPORT_SYMBOL(_debug_pagealloc_enabled);
724 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
726 static int __init early_debug_pagealloc(char *buf)
728 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
730 early_param("debug_pagealloc", early_debug_pagealloc);
732 static int __init debug_guardpage_minorder_setup(char *buf)
736 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
737 pr_err("Bad debug_guardpage_minorder value\n");
740 _debug_guardpage_minorder = res;
741 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
744 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
746 static inline bool set_page_guard(struct zone *zone, struct page *page,
747 unsigned int order, int migratetype)
749 if (!debug_guardpage_enabled())
752 if (order >= debug_guardpage_minorder())
755 __SetPageGuard(page);
756 INIT_LIST_HEAD(&page->lru);
757 set_page_private(page, order);
758 /* Guard pages are not available for any usage */
759 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
764 static inline void clear_page_guard(struct zone *zone, struct page *page,
765 unsigned int order, int migratetype)
767 if (!debug_guardpage_enabled())
770 __ClearPageGuard(page);
772 set_page_private(page, 0);
773 if (!is_migrate_isolate(migratetype))
774 __mod_zone_freepage_state(zone, (1 << order), migratetype);
777 static inline bool set_page_guard(struct zone *zone, struct page *page,
778 unsigned int order, int migratetype) { return false; }
779 static inline void clear_page_guard(struct zone *zone, struct page *page,
780 unsigned int order, int migratetype) {}
784 * Enable static keys related to various memory debugging and hardening options.
785 * Some override others, and depend on early params that are evaluated in the
786 * order of appearance. So we need to first gather the full picture of what was
787 * enabled, and then make decisions.
789 void init_mem_debugging_and_hardening(void)
791 if (_init_on_alloc_enabled_early) {
792 if (page_poisoning_enabled())
793 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
794 "will take precedence over init_on_alloc\n");
796 static_branch_enable(&init_on_alloc);
798 if (_init_on_free_enabled_early) {
799 if (page_poisoning_enabled())
800 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
801 "will take precedence over init_on_free\n");
803 static_branch_enable(&init_on_free);
806 #ifdef CONFIG_PAGE_POISONING
808 * Page poisoning is debug page alloc for some arches. If
809 * either of those options are enabled, enable poisoning.
811 if (page_poisoning_enabled() ||
812 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
813 debug_pagealloc_enabled()))
814 static_branch_enable(&_page_poisoning_enabled);
817 #ifdef CONFIG_DEBUG_PAGEALLOC
818 if (!debug_pagealloc_enabled())
821 static_branch_enable(&_debug_pagealloc_enabled);
823 if (!debug_guardpage_minorder())
826 static_branch_enable(&_debug_guardpage_enabled);
830 static inline void set_buddy_order(struct page *page, unsigned int order)
832 set_page_private(page, order);
833 __SetPageBuddy(page);
837 * This function checks whether a page is free && is the buddy
838 * we can coalesce a page and its buddy if
839 * (a) the buddy is not in a hole (check before calling!) &&
840 * (b) the buddy is in the buddy system &&
841 * (c) a page and its buddy have the same order &&
842 * (d) a page and its buddy are in the same zone.
844 * For recording whether a page is in the buddy system, we set PageBuddy.
845 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
847 * For recording page's order, we use page_private(page).
849 static inline bool page_is_buddy(struct page *page, struct page *buddy,
852 if (!page_is_guard(buddy) && !PageBuddy(buddy))
855 if (buddy_order(buddy) != order)
859 * zone check is done late to avoid uselessly calculating
860 * zone/node ids for pages that could never merge.
862 if (page_zone_id(page) != page_zone_id(buddy))
865 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
870 #ifdef CONFIG_COMPACTION
871 static inline struct capture_control *task_capc(struct zone *zone)
873 struct capture_control *capc = current->capture_control;
875 return unlikely(capc) &&
876 !(current->flags & PF_KTHREAD) &&
878 capc->cc->zone == zone ? capc : NULL;
882 compaction_capture(struct capture_control *capc, struct page *page,
883 int order, int migratetype)
885 if (!capc || order != capc->cc->order)
888 /* Do not accidentally pollute CMA or isolated regions*/
889 if (is_migrate_cma(migratetype) ||
890 is_migrate_isolate(migratetype))
894 * Do not let lower order allocations polluate a movable pageblock.
895 * This might let an unmovable request use a reclaimable pageblock
896 * and vice-versa but no more than normal fallback logic which can
897 * have trouble finding a high-order free page.
899 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
907 static inline struct capture_control *task_capc(struct zone *zone)
913 compaction_capture(struct capture_control *capc, struct page *page,
914 int order, int migratetype)
918 #endif /* CONFIG_COMPACTION */
920 /* Used for pages not on another list */
921 static inline void add_to_free_list(struct page *page, struct zone *zone,
922 unsigned int order, int migratetype)
924 struct free_area *area = &zone->free_area[order];
926 list_add(&page->lru, &area->free_list[migratetype]);
930 /* Used for pages not on another list */
931 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
932 unsigned int order, int migratetype)
934 struct free_area *area = &zone->free_area[order];
936 list_add_tail(&page->lru, &area->free_list[migratetype]);
941 * Used for pages which are on another list. Move the pages to the tail
942 * of the list - so the moved pages won't immediately be considered for
943 * allocation again (e.g., optimization for memory onlining).
945 static inline void move_to_free_list(struct page *page, struct zone *zone,
946 unsigned int order, int migratetype)
948 struct free_area *area = &zone->free_area[order];
950 list_move_tail(&page->lru, &area->free_list[migratetype]);
953 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
956 /* clear reported state and update reported page count */
957 if (page_reported(page))
958 __ClearPageReported(page);
960 list_del(&page->lru);
961 __ClearPageBuddy(page);
962 set_page_private(page, 0);
963 zone->free_area[order].nr_free--;
967 * If this is not the largest possible page, check if the buddy
968 * of the next-highest order is free. If it is, it's possible
969 * that pages are being freed that will coalesce soon. In case,
970 * that is happening, add the free page to the tail of the list
971 * so it's less likely to be used soon and more likely to be merged
972 * as a higher order page
975 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
976 struct page *page, unsigned int order)
978 struct page *higher_page, *higher_buddy;
979 unsigned long combined_pfn;
981 if (order >= MAX_ORDER - 2)
984 if (!pfn_valid_within(buddy_pfn))
987 combined_pfn = buddy_pfn & pfn;
988 higher_page = page + (combined_pfn - pfn);
989 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
990 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
992 return pfn_valid_within(buddy_pfn) &&
993 page_is_buddy(higher_page, higher_buddy, order + 1);
997 * Freeing function for a buddy system allocator.
999 * The concept of a buddy system is to maintain direct-mapped table
1000 * (containing bit values) for memory blocks of various "orders".
1001 * The bottom level table contains the map for the smallest allocatable
1002 * units of memory (here, pages), and each level above it describes
1003 * pairs of units from the levels below, hence, "buddies".
1004 * At a high level, all that happens here is marking the table entry
1005 * at the bottom level available, and propagating the changes upward
1006 * as necessary, plus some accounting needed to play nicely with other
1007 * parts of the VM system.
1008 * At each level, we keep a list of pages, which are heads of continuous
1009 * free pages of length of (1 << order) and marked with PageBuddy.
1010 * Page's order is recorded in page_private(page) field.
1011 * So when we are allocating or freeing one, we can derive the state of the
1012 * other. That is, if we allocate a small block, and both were
1013 * free, the remainder of the region must be split into blocks.
1014 * If a block is freed, and its buddy is also free, then this
1015 * triggers coalescing into a block of larger size.
1020 static inline void __free_one_page(struct page *page,
1022 struct zone *zone, unsigned int order,
1023 int migratetype, fpi_t fpi_flags)
1025 struct capture_control *capc = task_capc(zone);
1026 unsigned long buddy_pfn;
1027 unsigned long combined_pfn;
1028 unsigned int max_order;
1032 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1034 VM_BUG_ON(!zone_is_initialized(zone));
1035 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1037 VM_BUG_ON(migratetype == -1);
1038 if (likely(!is_migrate_isolate(migratetype)))
1039 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1041 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1042 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1045 while (order < max_order) {
1046 if (compaction_capture(capc, page, order, migratetype)) {
1047 __mod_zone_freepage_state(zone, -(1 << order),
1051 buddy_pfn = __find_buddy_pfn(pfn, order);
1052 buddy = page + (buddy_pfn - pfn);
1054 if (!pfn_valid_within(buddy_pfn))
1056 if (!page_is_buddy(page, buddy, order))
1059 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1060 * merge with it and move up one order.
1062 if (page_is_guard(buddy))
1063 clear_page_guard(zone, buddy, order, migratetype);
1065 del_page_from_free_list(buddy, zone, order);
1066 combined_pfn = buddy_pfn & pfn;
1067 page = page + (combined_pfn - pfn);
1071 if (order < MAX_ORDER - 1) {
1072 /* If we are here, it means order is >= pageblock_order.
1073 * We want to prevent merge between freepages on isolate
1074 * pageblock and normal pageblock. Without this, pageblock
1075 * isolation could cause incorrect freepage or CMA accounting.
1077 * We don't want to hit this code for the more frequent
1078 * low-order merging.
1080 if (unlikely(has_isolate_pageblock(zone))) {
1083 buddy_pfn = __find_buddy_pfn(pfn, order);
1084 buddy = page + (buddy_pfn - pfn);
1085 buddy_mt = get_pageblock_migratetype(buddy);
1087 if (migratetype != buddy_mt
1088 && (is_migrate_isolate(migratetype) ||
1089 is_migrate_isolate(buddy_mt)))
1092 max_order = order + 1;
1093 goto continue_merging;
1097 set_buddy_order(page, order);
1099 if (fpi_flags & FPI_TO_TAIL)
1101 else if (is_shuffle_order(order))
1102 to_tail = shuffle_pick_tail();
1104 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1107 add_to_free_list_tail(page, zone, order, migratetype);
1109 add_to_free_list(page, zone, order, migratetype);
1111 /* Notify page reporting subsystem of freed page */
1112 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1113 page_reporting_notify_free(order);
1117 * A bad page could be due to a number of fields. Instead of multiple branches,
1118 * try and check multiple fields with one check. The caller must do a detailed
1119 * check if necessary.
1121 static inline bool page_expected_state(struct page *page,
1122 unsigned long check_flags)
1124 if (unlikely(atomic_read(&page->_mapcount) != -1))
1127 if (unlikely((unsigned long)page->mapping |
1128 page_ref_count(page) |
1132 (page->flags & check_flags)))
1138 static const char *page_bad_reason(struct page *page, unsigned long flags)
1140 const char *bad_reason = NULL;
1142 if (unlikely(atomic_read(&page->_mapcount) != -1))
1143 bad_reason = "nonzero mapcount";
1144 if (unlikely(page->mapping != NULL))
1145 bad_reason = "non-NULL mapping";
1146 if (unlikely(page_ref_count(page) != 0))
1147 bad_reason = "nonzero _refcount";
1148 if (unlikely(page->flags & flags)) {
1149 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1150 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1152 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1155 if (unlikely(page->memcg_data))
1156 bad_reason = "page still charged to cgroup";
1161 static void check_free_page_bad(struct page *page)
1164 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1167 static inline int check_free_page(struct page *page)
1169 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1172 /* Something has gone sideways, find it */
1173 check_free_page_bad(page);
1177 static int free_tail_pages_check(struct page *head_page, struct page *page)
1182 * We rely page->lru.next never has bit 0 set, unless the page
1183 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1185 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1187 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1191 switch (page - head_page) {
1193 /* the first tail page: ->mapping may be compound_mapcount() */
1194 if (unlikely(compound_mapcount(page))) {
1195 bad_page(page, "nonzero compound_mapcount");
1201 * the second tail page: ->mapping is
1202 * deferred_list.next -- ignore value.
1206 if (page->mapping != TAIL_MAPPING) {
1207 bad_page(page, "corrupted mapping in tail page");
1212 if (unlikely(!PageTail(page))) {
1213 bad_page(page, "PageTail not set");
1216 if (unlikely(compound_head(page) != head_page)) {
1217 bad_page(page, "compound_head not consistent");
1222 page->mapping = NULL;
1223 clear_compound_head(page);
1227 static void kernel_init_free_pages(struct page *page, int numpages)
1231 /* s390's use of memset() could override KASAN redzones. */
1232 kasan_disable_current();
1233 for (i = 0; i < numpages; i++) {
1234 u8 tag = page_kasan_tag(page + i);
1235 page_kasan_tag_reset(page + i);
1236 clear_highpage(page + i);
1237 page_kasan_tag_set(page + i, tag);
1239 kasan_enable_current();
1242 static __always_inline bool free_pages_prepare(struct page *page,
1243 unsigned int order, bool check_free, fpi_t fpi_flags)
1248 VM_BUG_ON_PAGE(PageTail(page), page);
1250 trace_mm_page_free(page, order);
1252 if (unlikely(PageHWPoison(page)) && !order) {
1254 * Do not let hwpoison pages hit pcplists/buddy
1255 * Untie memcg state and reset page's owner
1257 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1258 __memcg_kmem_uncharge_page(page, order);
1259 reset_page_owner(page, order);
1264 * Check tail pages before head page information is cleared to
1265 * avoid checking PageCompound for order-0 pages.
1267 if (unlikely(order)) {
1268 bool compound = PageCompound(page);
1271 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1274 ClearPageDoubleMap(page);
1275 for (i = 1; i < (1 << order); i++) {
1277 bad += free_tail_pages_check(page, page + i);
1278 if (unlikely(check_free_page(page + i))) {
1282 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1285 if (PageMappingFlags(page))
1286 page->mapping = NULL;
1287 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1288 __memcg_kmem_uncharge_page(page, order);
1290 bad += check_free_page(page);
1294 page_cpupid_reset_last(page);
1295 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1296 reset_page_owner(page, order);
1298 if (!PageHighMem(page)) {
1299 debug_check_no_locks_freed(page_address(page),
1300 PAGE_SIZE << order);
1301 debug_check_no_obj_freed(page_address(page),
1302 PAGE_SIZE << order);
1305 kernel_poison_pages(page, 1 << order);
1308 * As memory initialization might be integrated into KASAN,
1309 * kasan_free_pages and kernel_init_free_pages must be
1310 * kept together to avoid discrepancies in behavior.
1312 * With hardware tag-based KASAN, memory tags must be set before the
1313 * page becomes unavailable via debug_pagealloc or arch_free_page.
1315 init = want_init_on_free();
1316 if (init && !kasan_has_integrated_init())
1317 kernel_init_free_pages(page, 1 << order);
1318 kasan_free_nondeferred_pages(page, order, init, fpi_flags);
1321 * arch_free_page() can make the page's contents inaccessible. s390
1322 * does this. So nothing which can access the page's contents should
1323 * happen after this.
1325 arch_free_page(page, order);
1327 debug_pagealloc_unmap_pages(page, 1 << order);
1332 #ifdef CONFIG_DEBUG_VM
1334 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1335 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1336 * moved from pcp lists to free lists.
1338 static bool free_pcp_prepare(struct page *page)
1340 return free_pages_prepare(page, 0, true, FPI_NONE);
1343 static bool bulkfree_pcp_prepare(struct page *page)
1345 if (debug_pagealloc_enabled_static())
1346 return check_free_page(page);
1352 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1353 * moving from pcp lists to free list in order to reduce overhead. With
1354 * debug_pagealloc enabled, they are checked also immediately when being freed
1357 static bool free_pcp_prepare(struct page *page)
1359 if (debug_pagealloc_enabled_static())
1360 return free_pages_prepare(page, 0, true, FPI_NONE);
1362 return free_pages_prepare(page, 0, false, FPI_NONE);
1365 static bool bulkfree_pcp_prepare(struct page *page)
1367 return check_free_page(page);
1369 #endif /* CONFIG_DEBUG_VM */
1371 static inline void prefetch_buddy(struct page *page)
1373 unsigned long pfn = page_to_pfn(page);
1374 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1375 struct page *buddy = page + (buddy_pfn - pfn);
1381 * Frees a number of pages from the PCP lists
1382 * Assumes all pages on list are in same zone, and of same order.
1383 * count is the number of pages to free.
1385 * If the zone was previously in an "all pages pinned" state then look to
1386 * see if this freeing clears that state.
1388 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1389 * pinned" detection logic.
1391 static void free_pcppages_bulk(struct zone *zone, int count,
1392 struct per_cpu_pages *pcp)
1394 int migratetype = 0;
1396 int prefetch_nr = READ_ONCE(pcp->batch);
1397 bool isolated_pageblocks;
1398 struct page *page, *tmp;
1402 * Ensure proper count is passed which otherwise would stuck in the
1403 * below while (list_empty(list)) loop.
1405 count = min(pcp->count, count);
1407 struct list_head *list;
1410 * Remove pages from lists in a round-robin fashion. A
1411 * batch_free count is maintained that is incremented when an
1412 * empty list is encountered. This is so more pages are freed
1413 * off fuller lists instead of spinning excessively around empty
1418 if (++migratetype == MIGRATE_PCPTYPES)
1420 list = &pcp->lists[migratetype];
1421 } while (list_empty(list));
1423 /* This is the only non-empty list. Free them all. */
1424 if (batch_free == MIGRATE_PCPTYPES)
1428 page = list_last_entry(list, struct page, lru);
1429 /* must delete to avoid corrupting pcp list */
1430 list_del(&page->lru);
1433 if (bulkfree_pcp_prepare(page))
1436 list_add_tail(&page->lru, &head);
1439 * We are going to put the page back to the global
1440 * pool, prefetch its buddy to speed up later access
1441 * under zone->lock. It is believed the overhead of
1442 * an additional test and calculating buddy_pfn here
1443 * can be offset by reduced memory latency later. To
1444 * avoid excessive prefetching due to large count, only
1445 * prefetch buddy for the first pcp->batch nr of pages.
1448 prefetch_buddy(page);
1451 } while (--count && --batch_free && !list_empty(list));
1454 spin_lock(&zone->lock);
1455 isolated_pageblocks = has_isolate_pageblock(zone);
1458 * Use safe version since after __free_one_page(),
1459 * page->lru.next will not point to original list.
1461 list_for_each_entry_safe(page, tmp, &head, lru) {
1462 int mt = get_pcppage_migratetype(page);
1463 /* MIGRATE_ISOLATE page should not go to pcplists */
1464 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1465 /* Pageblock could have been isolated meanwhile */
1466 if (unlikely(isolated_pageblocks))
1467 mt = get_pageblock_migratetype(page);
1469 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1470 trace_mm_page_pcpu_drain(page, 0, mt);
1472 spin_unlock(&zone->lock);
1475 static void free_one_page(struct zone *zone,
1476 struct page *page, unsigned long pfn,
1478 int migratetype, fpi_t fpi_flags)
1480 spin_lock(&zone->lock);
1481 if (unlikely(has_isolate_pageblock(zone) ||
1482 is_migrate_isolate(migratetype))) {
1483 migratetype = get_pfnblock_migratetype(page, pfn);
1485 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1486 spin_unlock(&zone->lock);
1489 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1490 unsigned long zone, int nid)
1492 mm_zero_struct_page(page);
1493 set_page_links(page, zone, nid, pfn);
1494 init_page_count(page);
1495 page_mapcount_reset(page);
1496 page_cpupid_reset_last(page);
1497 page_kasan_tag_reset(page);
1499 INIT_LIST_HEAD(&page->lru);
1500 #ifdef WANT_PAGE_VIRTUAL
1501 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1502 if (!is_highmem_idx(zone))
1503 set_page_address(page, __va(pfn << PAGE_SHIFT));
1507 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1508 static void __meminit init_reserved_page(unsigned long pfn)
1513 if (!early_page_uninitialised(pfn))
1516 nid = early_pfn_to_nid(pfn);
1517 pgdat = NODE_DATA(nid);
1519 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1520 struct zone *zone = &pgdat->node_zones[zid];
1522 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1525 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1528 static inline void init_reserved_page(unsigned long pfn)
1531 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1534 * Initialised pages do not have PageReserved set. This function is
1535 * called for each range allocated by the bootmem allocator and
1536 * marks the pages PageReserved. The remaining valid pages are later
1537 * sent to the buddy page allocator.
1539 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1541 unsigned long start_pfn = PFN_DOWN(start);
1542 unsigned long end_pfn = PFN_UP(end);
1544 for (; start_pfn < end_pfn; start_pfn++) {
1545 if (pfn_valid(start_pfn)) {
1546 struct page *page = pfn_to_page(start_pfn);
1548 init_reserved_page(start_pfn);
1550 /* Avoid false-positive PageTail() */
1551 INIT_LIST_HEAD(&page->lru);
1554 * no need for atomic set_bit because the struct
1555 * page is not visible yet so nobody should
1558 __SetPageReserved(page);
1563 static void __free_pages_ok(struct page *page, unsigned int order,
1566 unsigned long flags;
1568 unsigned long pfn = page_to_pfn(page);
1570 if (!free_pages_prepare(page, order, true, fpi_flags))
1573 migratetype = get_pfnblock_migratetype(page, pfn);
1574 local_irq_save(flags);
1575 __count_vm_events(PGFREE, 1 << order);
1576 free_one_page(page_zone(page), page, pfn, order, migratetype,
1578 local_irq_restore(flags);
1581 void __free_pages_core(struct page *page, unsigned int order)
1583 unsigned int nr_pages = 1 << order;
1584 struct page *p = page;
1588 * When initializing the memmap, __init_single_page() sets the refcount
1589 * of all pages to 1 ("allocated"/"not free"). We have to set the
1590 * refcount of all involved pages to 0.
1593 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1595 __ClearPageReserved(p);
1596 set_page_count(p, 0);
1598 __ClearPageReserved(p);
1599 set_page_count(p, 0);
1601 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1604 * Bypass PCP and place fresh pages right to the tail, primarily
1605 * relevant for memory onlining.
1607 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1610 #ifdef CONFIG_NEED_MULTIPLE_NODES
1613 * During memory init memblocks map pfns to nids. The search is expensive and
1614 * this caches recent lookups. The implementation of __early_pfn_to_nid
1615 * treats start/end as pfns.
1617 struct mminit_pfnnid_cache {
1618 unsigned long last_start;
1619 unsigned long last_end;
1623 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1626 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1628 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1629 struct mminit_pfnnid_cache *state)
1631 unsigned long start_pfn, end_pfn;
1634 if (state->last_start <= pfn && pfn < state->last_end)
1635 return state->last_nid;
1637 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1638 if (nid != NUMA_NO_NODE) {
1639 state->last_start = start_pfn;
1640 state->last_end = end_pfn;
1641 state->last_nid = nid;
1647 int __meminit early_pfn_to_nid(unsigned long pfn)
1649 static DEFINE_SPINLOCK(early_pfn_lock);
1652 spin_lock(&early_pfn_lock);
1653 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1655 nid = first_online_node;
1656 spin_unlock(&early_pfn_lock);
1660 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1662 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1665 if (early_page_uninitialised(pfn))
1667 __free_pages_core(page, order);
1671 * Check that the whole (or subset of) a pageblock given by the interval of
1672 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1673 * with the migration of free compaction scanner. The scanners then need to
1674 * use only pfn_valid_within() check for arches that allow holes within
1677 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1679 * It's possible on some configurations to have a setup like node0 node1 node0
1680 * i.e. it's possible that all pages within a zones range of pages do not
1681 * belong to a single zone. We assume that a border between node0 and node1
1682 * can occur within a single pageblock, but not a node0 node1 node0
1683 * interleaving within a single pageblock. It is therefore sufficient to check
1684 * the first and last page of a pageblock and avoid checking each individual
1685 * page in a pageblock.
1687 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1688 unsigned long end_pfn, struct zone *zone)
1690 struct page *start_page;
1691 struct page *end_page;
1693 /* end_pfn is one past the range we are checking */
1696 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1699 start_page = pfn_to_online_page(start_pfn);
1703 if (page_zone(start_page) != zone)
1706 end_page = pfn_to_page(end_pfn);
1708 /* This gives a shorter code than deriving page_zone(end_page) */
1709 if (page_zone_id(start_page) != page_zone_id(end_page))
1715 void set_zone_contiguous(struct zone *zone)
1717 unsigned long block_start_pfn = zone->zone_start_pfn;
1718 unsigned long block_end_pfn;
1720 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1721 for (; block_start_pfn < zone_end_pfn(zone);
1722 block_start_pfn = block_end_pfn,
1723 block_end_pfn += pageblock_nr_pages) {
1725 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1727 if (!__pageblock_pfn_to_page(block_start_pfn,
1728 block_end_pfn, zone))
1733 /* We confirm that there is no hole */
1734 zone->contiguous = true;
1737 void clear_zone_contiguous(struct zone *zone)
1739 zone->contiguous = false;
1742 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1743 static void __init deferred_free_range(unsigned long pfn,
1744 unsigned long nr_pages)
1752 page = pfn_to_page(pfn);
1754 /* Free a large naturally-aligned chunk if possible */
1755 if (nr_pages == pageblock_nr_pages &&
1756 (pfn & (pageblock_nr_pages - 1)) == 0) {
1757 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1758 __free_pages_core(page, pageblock_order);
1762 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1763 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1764 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1765 __free_pages_core(page, 0);
1769 /* Completion tracking for deferred_init_memmap() threads */
1770 static atomic_t pgdat_init_n_undone __initdata;
1771 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1773 static inline void __init pgdat_init_report_one_done(void)
1775 if (atomic_dec_and_test(&pgdat_init_n_undone))
1776 complete(&pgdat_init_all_done_comp);
1780 * Returns true if page needs to be initialized or freed to buddy allocator.
1782 * First we check if pfn is valid on architectures where it is possible to have
1783 * holes within pageblock_nr_pages. On systems where it is not possible, this
1784 * function is optimized out.
1786 * Then, we check if a current large page is valid by only checking the validity
1789 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1791 if (!pfn_valid_within(pfn))
1793 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1799 * Free pages to buddy allocator. Try to free aligned pages in
1800 * pageblock_nr_pages sizes.
1802 static void __init deferred_free_pages(unsigned long pfn,
1803 unsigned long end_pfn)
1805 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1806 unsigned long nr_free = 0;
1808 for (; pfn < end_pfn; pfn++) {
1809 if (!deferred_pfn_valid(pfn)) {
1810 deferred_free_range(pfn - nr_free, nr_free);
1812 } else if (!(pfn & nr_pgmask)) {
1813 deferred_free_range(pfn - nr_free, nr_free);
1819 /* Free the last block of pages to allocator */
1820 deferred_free_range(pfn - nr_free, nr_free);
1824 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1825 * by performing it only once every pageblock_nr_pages.
1826 * Return number of pages initialized.
1828 static unsigned long __init deferred_init_pages(struct zone *zone,
1830 unsigned long end_pfn)
1832 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1833 int nid = zone_to_nid(zone);
1834 unsigned long nr_pages = 0;
1835 int zid = zone_idx(zone);
1836 struct page *page = NULL;
1838 for (; pfn < end_pfn; pfn++) {
1839 if (!deferred_pfn_valid(pfn)) {
1842 } else if (!page || !(pfn & nr_pgmask)) {
1843 page = pfn_to_page(pfn);
1847 __init_single_page(page, pfn, zid, nid);
1854 * This function is meant to pre-load the iterator for the zone init.
1855 * Specifically it walks through the ranges until we are caught up to the
1856 * first_init_pfn value and exits there. If we never encounter the value we
1857 * return false indicating there are no valid ranges left.
1860 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1861 unsigned long *spfn, unsigned long *epfn,
1862 unsigned long first_init_pfn)
1867 * Start out by walking through the ranges in this zone that have
1868 * already been initialized. We don't need to do anything with them
1869 * so we just need to flush them out of the system.
1871 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1872 if (*epfn <= first_init_pfn)
1874 if (*spfn < first_init_pfn)
1875 *spfn = first_init_pfn;
1884 * Initialize and free pages. We do it in two loops: first we initialize
1885 * struct page, then free to buddy allocator, because while we are
1886 * freeing pages we can access pages that are ahead (computing buddy
1887 * page in __free_one_page()).
1889 * In order to try and keep some memory in the cache we have the loop
1890 * broken along max page order boundaries. This way we will not cause
1891 * any issues with the buddy page computation.
1893 static unsigned long __init
1894 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1895 unsigned long *end_pfn)
1897 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1898 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1899 unsigned long nr_pages = 0;
1902 /* First we loop through and initialize the page values */
1903 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1906 if (mo_pfn <= *start_pfn)
1909 t = min(mo_pfn, *end_pfn);
1910 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1912 if (mo_pfn < *end_pfn) {
1913 *start_pfn = mo_pfn;
1918 /* Reset values and now loop through freeing pages as needed */
1921 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1927 t = min(mo_pfn, epfn);
1928 deferred_free_pages(spfn, t);
1938 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1941 unsigned long spfn, epfn;
1942 struct zone *zone = arg;
1945 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1948 * Initialize and free pages in MAX_ORDER sized increments so that we
1949 * can avoid introducing any issues with the buddy allocator.
1951 while (spfn < end_pfn) {
1952 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1957 /* An arch may override for more concurrency. */
1959 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1964 /* Initialise remaining memory on a node */
1965 static int __init deferred_init_memmap(void *data)
1967 pg_data_t *pgdat = data;
1968 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1969 unsigned long spfn = 0, epfn = 0;
1970 unsigned long first_init_pfn, flags;
1971 unsigned long start = jiffies;
1973 int zid, max_threads;
1976 /* Bind memory initialisation thread to a local node if possible */
1977 if (!cpumask_empty(cpumask))
1978 set_cpus_allowed_ptr(current, cpumask);
1980 pgdat_resize_lock(pgdat, &flags);
1981 first_init_pfn = pgdat->first_deferred_pfn;
1982 if (first_init_pfn == ULONG_MAX) {
1983 pgdat_resize_unlock(pgdat, &flags);
1984 pgdat_init_report_one_done();
1988 /* Sanity check boundaries */
1989 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1990 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1991 pgdat->first_deferred_pfn = ULONG_MAX;
1994 * Once we unlock here, the zone cannot be grown anymore, thus if an
1995 * interrupt thread must allocate this early in boot, zone must be
1996 * pre-grown prior to start of deferred page initialization.
1998 pgdat_resize_unlock(pgdat, &flags);
2000 /* Only the highest zone is deferred so find it */
2001 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2002 zone = pgdat->node_zones + zid;
2003 if (first_init_pfn < zone_end_pfn(zone))
2007 /* If the zone is empty somebody else may have cleared out the zone */
2008 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2012 max_threads = deferred_page_init_max_threads(cpumask);
2014 while (spfn < epfn) {
2015 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2016 struct padata_mt_job job = {
2017 .thread_fn = deferred_init_memmap_chunk,
2020 .size = epfn_align - spfn,
2021 .align = PAGES_PER_SECTION,
2022 .min_chunk = PAGES_PER_SECTION,
2023 .max_threads = max_threads,
2026 padata_do_multithreaded(&job);
2027 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2031 /* Sanity check that the next zone really is unpopulated */
2032 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2034 pr_info("node %d deferred pages initialised in %ums\n",
2035 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2037 pgdat_init_report_one_done();
2042 * If this zone has deferred pages, try to grow it by initializing enough
2043 * deferred pages to satisfy the allocation specified by order, rounded up to
2044 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2045 * of SECTION_SIZE bytes by initializing struct pages in increments of
2046 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2048 * Return true when zone was grown, otherwise return false. We return true even
2049 * when we grow less than requested, to let the caller decide if there are
2050 * enough pages to satisfy the allocation.
2052 * Note: We use noinline because this function is needed only during boot, and
2053 * it is called from a __ref function _deferred_grow_zone. This way we are
2054 * making sure that it is not inlined into permanent text section.
2056 static noinline bool __init
2057 deferred_grow_zone(struct zone *zone, unsigned int order)
2059 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2060 pg_data_t *pgdat = zone->zone_pgdat;
2061 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2062 unsigned long spfn, epfn, flags;
2063 unsigned long nr_pages = 0;
2066 /* Only the last zone may have deferred pages */
2067 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2070 pgdat_resize_lock(pgdat, &flags);
2073 * If someone grew this zone while we were waiting for spinlock, return
2074 * true, as there might be enough pages already.
2076 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2077 pgdat_resize_unlock(pgdat, &flags);
2081 /* If the zone is empty somebody else may have cleared out the zone */
2082 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2083 first_deferred_pfn)) {
2084 pgdat->first_deferred_pfn = ULONG_MAX;
2085 pgdat_resize_unlock(pgdat, &flags);
2086 /* Retry only once. */
2087 return first_deferred_pfn != ULONG_MAX;
2091 * Initialize and free pages in MAX_ORDER sized increments so
2092 * that we can avoid introducing any issues with the buddy
2095 while (spfn < epfn) {
2096 /* update our first deferred PFN for this section */
2097 first_deferred_pfn = spfn;
2099 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2100 touch_nmi_watchdog();
2102 /* We should only stop along section boundaries */
2103 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2106 /* If our quota has been met we can stop here */
2107 if (nr_pages >= nr_pages_needed)
2111 pgdat->first_deferred_pfn = spfn;
2112 pgdat_resize_unlock(pgdat, &flags);
2114 return nr_pages > 0;
2118 * deferred_grow_zone() is __init, but it is called from
2119 * get_page_from_freelist() during early boot until deferred_pages permanently
2120 * disables this call. This is why we have refdata wrapper to avoid warning,
2121 * and to ensure that the function body gets unloaded.
2124 _deferred_grow_zone(struct zone *zone, unsigned int order)
2126 return deferred_grow_zone(zone, order);
2129 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2131 void __init page_alloc_init_late(void)
2136 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2138 /* There will be num_node_state(N_MEMORY) threads */
2139 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2140 for_each_node_state(nid, N_MEMORY) {
2141 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2144 /* Block until all are initialised */
2145 wait_for_completion(&pgdat_init_all_done_comp);
2148 * The number of managed pages has changed due to the initialisation
2149 * so the pcpu batch and high limits needs to be updated or the limits
2150 * will be artificially small.
2152 for_each_populated_zone(zone)
2153 zone_pcp_update(zone);
2156 * We initialized the rest of the deferred pages. Permanently disable
2157 * on-demand struct page initialization.
2159 static_branch_disable(&deferred_pages);
2161 /* Reinit limits that are based on free pages after the kernel is up */
2162 files_maxfiles_init();
2167 /* Discard memblock private memory */
2170 for_each_node_state(nid, N_MEMORY)
2171 shuffle_free_memory(NODE_DATA(nid));
2173 for_each_populated_zone(zone)
2174 set_zone_contiguous(zone);
2178 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2179 void __init init_cma_reserved_pageblock(struct page *page)
2181 unsigned i = pageblock_nr_pages;
2182 struct page *p = page;
2185 __ClearPageReserved(p);
2186 set_page_count(p, 0);
2189 set_pageblock_migratetype(page, MIGRATE_CMA);
2191 if (pageblock_order >= MAX_ORDER) {
2192 i = pageblock_nr_pages;
2195 set_page_refcounted(p);
2196 __free_pages(p, MAX_ORDER - 1);
2197 p += MAX_ORDER_NR_PAGES;
2198 } while (i -= MAX_ORDER_NR_PAGES);
2200 set_page_refcounted(page);
2201 __free_pages(page, pageblock_order);
2204 adjust_managed_page_count(page, pageblock_nr_pages);
2205 page_zone(page)->cma_pages += pageblock_nr_pages;
2210 * The order of subdivision here is critical for the IO subsystem.
2211 * Please do not alter this order without good reasons and regression
2212 * testing. Specifically, as large blocks of memory are subdivided,
2213 * the order in which smaller blocks are delivered depends on the order
2214 * they're subdivided in this function. This is the primary factor
2215 * influencing the order in which pages are delivered to the IO
2216 * subsystem according to empirical testing, and this is also justified
2217 * by considering the behavior of a buddy system containing a single
2218 * large block of memory acted on by a series of small allocations.
2219 * This behavior is a critical factor in sglist merging's success.
2223 static inline void expand(struct zone *zone, struct page *page,
2224 int low, int high, int migratetype)
2226 unsigned long size = 1 << high;
2228 while (high > low) {
2231 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2234 * Mark as guard pages (or page), that will allow to
2235 * merge back to allocator when buddy will be freed.
2236 * Corresponding page table entries will not be touched,
2237 * pages will stay not present in virtual address space
2239 if (set_page_guard(zone, &page[size], high, migratetype))
2242 add_to_free_list(&page[size], zone, high, migratetype);
2243 set_buddy_order(&page[size], high);
2247 static void check_new_page_bad(struct page *page)
2249 if (unlikely(page->flags & __PG_HWPOISON)) {
2250 /* Don't complain about hwpoisoned pages */
2251 page_mapcount_reset(page); /* remove PageBuddy */
2256 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2260 * This page is about to be returned from the page allocator
2262 static inline int check_new_page(struct page *page)
2264 if (likely(page_expected_state(page,
2265 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2268 check_new_page_bad(page);
2272 #ifdef CONFIG_DEBUG_VM
2274 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2275 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2276 * also checked when pcp lists are refilled from the free lists.
2278 static inline bool check_pcp_refill(struct page *page)
2280 if (debug_pagealloc_enabled_static())
2281 return check_new_page(page);
2286 static inline bool check_new_pcp(struct page *page)
2288 return check_new_page(page);
2292 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2293 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2294 * enabled, they are also checked when being allocated from the pcp lists.
2296 static inline bool check_pcp_refill(struct page *page)
2298 return check_new_page(page);
2300 static inline bool check_new_pcp(struct page *page)
2302 if (debug_pagealloc_enabled_static())
2303 return check_new_page(page);
2307 #endif /* CONFIG_DEBUG_VM */
2309 static bool check_new_pages(struct page *page, unsigned int order)
2312 for (i = 0; i < (1 << order); i++) {
2313 struct page *p = page + i;
2315 if (unlikely(check_new_page(p)))
2322 inline void post_alloc_hook(struct page *page, unsigned int order,
2327 set_page_private(page, 0);
2328 set_page_refcounted(page);
2330 arch_alloc_page(page, order);
2331 debug_pagealloc_map_pages(page, 1 << order);
2334 * Page unpoisoning must happen before memory initialization.
2335 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2336 * allocations and the page unpoisoning code will complain.
2338 kernel_unpoison_pages(page, 1 << order);
2341 * As memory initialization might be integrated into KASAN,
2342 * kasan_alloc_pages and kernel_init_free_pages must be
2343 * kept together to avoid discrepancies in behavior.
2345 init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2346 kasan_alloc_pages(page, order, init);
2347 if (init && !kasan_has_integrated_init())
2348 kernel_init_free_pages(page, 1 << order);
2350 set_page_owner(page, order, gfp_flags);
2353 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2354 unsigned int alloc_flags)
2356 post_alloc_hook(page, order, gfp_flags);
2358 if (order && (gfp_flags & __GFP_COMP))
2359 prep_compound_page(page, order);
2362 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2363 * allocate the page. The expectation is that the caller is taking
2364 * steps that will free more memory. The caller should avoid the page
2365 * being used for !PFMEMALLOC purposes.
2367 if (alloc_flags & ALLOC_NO_WATERMARKS)
2368 set_page_pfmemalloc(page);
2370 clear_page_pfmemalloc(page);
2374 * Go through the free lists for the given migratetype and remove
2375 * the smallest available page from the freelists
2377 static __always_inline
2378 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2381 unsigned int current_order;
2382 struct free_area *area;
2385 /* Find a page of the appropriate size in the preferred list */
2386 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2387 area = &(zone->free_area[current_order]);
2388 page = get_page_from_free_area(area, migratetype);
2391 del_page_from_free_list(page, zone, current_order);
2392 expand(zone, page, order, current_order, migratetype);
2393 set_pcppage_migratetype(page, migratetype);
2402 * This array describes the order lists are fallen back to when
2403 * the free lists for the desirable migrate type are depleted
2405 static int fallbacks[MIGRATE_TYPES][3] = {
2406 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2407 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2408 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2410 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2412 #ifdef CONFIG_MEMORY_ISOLATION
2413 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2418 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2421 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2424 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2425 unsigned int order) { return NULL; }
2429 * Move the free pages in a range to the freelist tail of the requested type.
2430 * Note that start_page and end_pages are not aligned on a pageblock
2431 * boundary. If alignment is required, use move_freepages_block()
2433 static int move_freepages(struct zone *zone,
2434 struct page *start_page, struct page *end_page,
2435 int migratetype, int *num_movable)
2439 int pages_moved = 0;
2441 for (page = start_page; page <= end_page;) {
2442 if (!pfn_valid_within(page_to_pfn(page))) {
2447 if (!PageBuddy(page)) {
2449 * We assume that pages that could be isolated for
2450 * migration are movable. But we don't actually try
2451 * isolating, as that would be expensive.
2454 (PageLRU(page) || __PageMovable(page)))
2461 /* Make sure we are not inadvertently changing nodes */
2462 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2463 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2465 order = buddy_order(page);
2466 move_to_free_list(page, zone, order, migratetype);
2468 pages_moved += 1 << order;
2474 int move_freepages_block(struct zone *zone, struct page *page,
2475 int migratetype, int *num_movable)
2477 unsigned long start_pfn, end_pfn;
2478 struct page *start_page, *end_page;
2483 start_pfn = page_to_pfn(page);
2484 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2485 start_page = pfn_to_page(start_pfn);
2486 end_page = start_page + 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_page, end_page, 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)
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 'alloced' 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 */
3463 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3464 unsigned int alloc_flags,
3465 struct per_cpu_pages *pcp,
3466 struct list_head *list)
3471 if (list_empty(list)) {
3472 pcp->count += rmqueue_bulk(zone, 0,
3473 READ_ONCE(pcp->batch), list,
3474 migratetype, alloc_flags);
3475 if (unlikely(list_empty(list)))
3479 page = list_first_entry(list, struct page, lru);
3480 list_del(&page->lru);
3482 } while (check_new_pcp(page));
3487 /* Lock and remove page from the per-cpu list */
3488 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3489 struct zone *zone, gfp_t gfp_flags,
3490 int migratetype, unsigned int alloc_flags)
3492 struct per_cpu_pages *pcp;
3493 struct list_head *list;
3495 unsigned long flags;
3497 local_irq_save(flags);
3498 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3499 list = &pcp->lists[migratetype];
3500 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3502 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3503 zone_statistics(preferred_zone, zone);
3505 local_irq_restore(flags);
3510 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3513 struct page *rmqueue(struct zone *preferred_zone,
3514 struct zone *zone, unsigned int order,
3515 gfp_t gfp_flags, unsigned int alloc_flags,
3518 unsigned long flags;
3521 if (likely(order == 0)) {
3523 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3524 * we need to skip it when CMA area isn't allowed.
3526 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3527 migratetype != MIGRATE_MOVABLE) {
3528 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3529 migratetype, alloc_flags);
3535 * We most definitely don't want callers attempting to
3536 * allocate greater than order-1 page units with __GFP_NOFAIL.
3538 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3539 spin_lock_irqsave(&zone->lock, flags);
3544 * order-0 request can reach here when the pcplist is skipped
3545 * due to non-CMA allocation context. HIGHATOMIC area is
3546 * reserved for high-order atomic allocation, so order-0
3547 * request should skip it.
3549 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3550 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3552 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3555 page = __rmqueue(zone, order, migratetype, alloc_flags);
3556 } while (page && check_new_pages(page, order));
3557 spin_unlock(&zone->lock);
3560 __mod_zone_freepage_state(zone, -(1 << order),
3561 get_pcppage_migratetype(page));
3563 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3564 zone_statistics(preferred_zone, zone);
3565 local_irq_restore(flags);
3568 /* Separate test+clear to avoid unnecessary atomics */
3569 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3570 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3571 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3574 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3578 local_irq_restore(flags);
3582 #ifdef CONFIG_FAIL_PAGE_ALLOC
3585 struct fault_attr attr;
3587 bool ignore_gfp_highmem;
3588 bool ignore_gfp_reclaim;
3590 } fail_page_alloc = {
3591 .attr = FAULT_ATTR_INITIALIZER,
3592 .ignore_gfp_reclaim = true,
3593 .ignore_gfp_highmem = true,
3597 static int __init setup_fail_page_alloc(char *str)
3599 return setup_fault_attr(&fail_page_alloc.attr, str);
3601 __setup("fail_page_alloc=", setup_fail_page_alloc);
3603 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3605 if (order < fail_page_alloc.min_order)
3607 if (gfp_mask & __GFP_NOFAIL)
3609 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3611 if (fail_page_alloc.ignore_gfp_reclaim &&
3612 (gfp_mask & __GFP_DIRECT_RECLAIM))
3615 return should_fail(&fail_page_alloc.attr, 1 << order);
3618 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3620 static int __init fail_page_alloc_debugfs(void)
3622 umode_t mode = S_IFREG | 0600;
3625 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3626 &fail_page_alloc.attr);
3628 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3629 &fail_page_alloc.ignore_gfp_reclaim);
3630 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3631 &fail_page_alloc.ignore_gfp_highmem);
3632 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3637 late_initcall(fail_page_alloc_debugfs);
3639 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3641 #else /* CONFIG_FAIL_PAGE_ALLOC */
3643 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3648 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3650 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3652 return __should_fail_alloc_page(gfp_mask, order);
3654 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3656 static inline long __zone_watermark_unusable_free(struct zone *z,
3657 unsigned int order, unsigned int alloc_flags)
3659 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3660 long unusable_free = (1 << order) - 1;
3663 * If the caller does not have rights to ALLOC_HARDER then subtract
3664 * the high-atomic reserves. This will over-estimate the size of the
3665 * atomic reserve but it avoids a search.
3667 if (likely(!alloc_harder))
3668 unusable_free += z->nr_reserved_highatomic;
3671 /* If allocation can't use CMA areas don't use free CMA pages */
3672 if (!(alloc_flags & ALLOC_CMA))
3673 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3676 return unusable_free;
3680 * Return true if free base pages are above 'mark'. For high-order checks it
3681 * will return true of the order-0 watermark is reached and there is at least
3682 * one free page of a suitable size. Checking now avoids taking the zone lock
3683 * to check in the allocation paths if no pages are free.
3685 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3686 int highest_zoneidx, unsigned int alloc_flags,
3691 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3693 /* free_pages may go negative - that's OK */
3694 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3696 if (alloc_flags & ALLOC_HIGH)
3699 if (unlikely(alloc_harder)) {
3701 * OOM victims can try even harder than normal ALLOC_HARDER
3702 * users on the grounds that it's definitely going to be in
3703 * the exit path shortly and free memory. Any allocation it
3704 * makes during the free path will be small and short-lived.
3706 if (alloc_flags & ALLOC_OOM)
3713 * Check watermarks for an order-0 allocation request. If these
3714 * are not met, then a high-order request also cannot go ahead
3715 * even if a suitable page happened to be free.
3717 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3720 /* If this is an order-0 request then the watermark is fine */
3724 /* For a high-order request, check at least one suitable page is free */
3725 for (o = order; o < MAX_ORDER; o++) {
3726 struct free_area *area = &z->free_area[o];
3732 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3733 if (!free_area_empty(area, mt))
3738 if ((alloc_flags & ALLOC_CMA) &&
3739 !free_area_empty(area, MIGRATE_CMA)) {
3743 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3749 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3750 int highest_zoneidx, unsigned int alloc_flags)
3752 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3753 zone_page_state(z, NR_FREE_PAGES));
3756 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3757 unsigned long mark, int highest_zoneidx,
3758 unsigned int alloc_flags, gfp_t gfp_mask)
3762 free_pages = zone_page_state(z, NR_FREE_PAGES);
3765 * Fast check for order-0 only. If this fails then the reserves
3766 * need to be calculated.
3771 fast_free = free_pages;
3772 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3773 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3777 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3781 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3782 * when checking the min watermark. The min watermark is the
3783 * point where boosting is ignored so that kswapd is woken up
3784 * when below the low watermark.
3786 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3787 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3788 mark = z->_watermark[WMARK_MIN];
3789 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3790 alloc_flags, free_pages);
3796 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3797 unsigned long mark, int highest_zoneidx)
3799 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3801 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3802 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3804 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3809 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3811 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3812 node_reclaim_distance;
3814 #else /* CONFIG_NUMA */
3815 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3819 #endif /* CONFIG_NUMA */
3822 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3823 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3824 * premature use of a lower zone may cause lowmem pressure problems that
3825 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3826 * probably too small. It only makes sense to spread allocations to avoid
3827 * fragmentation between the Normal and DMA32 zones.
3829 static inline unsigned int
3830 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3832 unsigned int alloc_flags;
3835 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3838 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3840 #ifdef CONFIG_ZONE_DMA32
3844 if (zone_idx(zone) != ZONE_NORMAL)
3848 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3849 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3850 * on UMA that if Normal is populated then so is DMA32.
3852 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3853 if (nr_online_nodes > 1 && !populated_zone(--zone))
3856 alloc_flags |= ALLOC_NOFRAGMENT;
3857 #endif /* CONFIG_ZONE_DMA32 */
3861 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3862 unsigned int alloc_flags)
3865 unsigned int pflags = current->flags;
3867 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3868 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3869 alloc_flags |= ALLOC_CMA;
3876 * get_page_from_freelist goes through the zonelist trying to allocate
3879 static struct page *
3880 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3881 const struct alloc_context *ac)
3885 struct pglist_data *last_pgdat_dirty_limit = NULL;
3890 * Scan zonelist, looking for a zone with enough free.
3891 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3893 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3894 z = ac->preferred_zoneref;
3895 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3900 if (cpusets_enabled() &&
3901 (alloc_flags & ALLOC_CPUSET) &&
3902 !__cpuset_zone_allowed(zone, gfp_mask))
3905 * When allocating a page cache page for writing, we
3906 * want to get it from a node that is within its dirty
3907 * limit, such that no single node holds more than its
3908 * proportional share of globally allowed dirty pages.
3909 * The dirty limits take into account the node's
3910 * lowmem reserves and high watermark so that kswapd
3911 * should be able to balance it without having to
3912 * write pages from its LRU list.
3914 * XXX: For now, allow allocations to potentially
3915 * exceed the per-node dirty limit in the slowpath
3916 * (spread_dirty_pages unset) before going into reclaim,
3917 * which is important when on a NUMA setup the allowed
3918 * nodes are together not big enough to reach the
3919 * global limit. The proper fix for these situations
3920 * will require awareness of nodes in the
3921 * dirty-throttling and the flusher threads.
3923 if (ac->spread_dirty_pages) {
3924 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3927 if (!node_dirty_ok(zone->zone_pgdat)) {
3928 last_pgdat_dirty_limit = zone->zone_pgdat;
3933 if (no_fallback && nr_online_nodes > 1 &&
3934 zone != ac->preferred_zoneref->zone) {
3938 * If moving to a remote node, retry but allow
3939 * fragmenting fallbacks. Locality is more important
3940 * than fragmentation avoidance.
3942 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3943 if (zone_to_nid(zone) != local_nid) {
3944 alloc_flags &= ~ALLOC_NOFRAGMENT;
3949 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3950 if (!zone_watermark_fast(zone, order, mark,
3951 ac->highest_zoneidx, alloc_flags,
3955 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3957 * Watermark failed for this zone, but see if we can
3958 * grow this zone if it contains deferred pages.
3960 if (static_branch_unlikely(&deferred_pages)) {
3961 if (_deferred_grow_zone(zone, order))
3965 /* Checked here to keep the fast path fast */
3966 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3967 if (alloc_flags & ALLOC_NO_WATERMARKS)
3970 if (node_reclaim_mode == 0 ||
3971 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3974 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3976 case NODE_RECLAIM_NOSCAN:
3979 case NODE_RECLAIM_FULL:
3980 /* scanned but unreclaimable */
3983 /* did we reclaim enough */
3984 if (zone_watermark_ok(zone, order, mark,
3985 ac->highest_zoneidx, alloc_flags))
3993 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3994 gfp_mask, alloc_flags, ac->migratetype);
3996 prep_new_page(page, order, gfp_mask, alloc_flags);
3999 * If this is a high-order atomic allocation then check
4000 * if the pageblock should be reserved for the future
4002 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4003 reserve_highatomic_pageblock(page, zone, order);
4007 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4008 /* Try again if zone has deferred pages */
4009 if (static_branch_unlikely(&deferred_pages)) {
4010 if (_deferred_grow_zone(zone, order))
4018 * It's possible on a UMA machine to get through all zones that are
4019 * fragmented. If avoiding fragmentation, reset and try again.
4022 alloc_flags &= ~ALLOC_NOFRAGMENT;
4029 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4031 unsigned int filter = SHOW_MEM_FILTER_NODES;
4034 * This documents exceptions given to allocations in certain
4035 * contexts that are allowed to allocate outside current's set
4038 if (!(gfp_mask & __GFP_NOMEMALLOC))
4039 if (tsk_is_oom_victim(current) ||
4040 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4041 filter &= ~SHOW_MEM_FILTER_NODES;
4042 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4043 filter &= ~SHOW_MEM_FILTER_NODES;
4045 show_mem(filter, nodemask);
4048 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4050 struct va_format vaf;
4052 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4054 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4057 va_start(args, fmt);
4060 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4061 current->comm, &vaf, gfp_mask, &gfp_mask,
4062 nodemask_pr_args(nodemask));
4065 cpuset_print_current_mems_allowed();
4068 warn_alloc_show_mem(gfp_mask, nodemask);
4071 static inline struct page *
4072 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4073 unsigned int alloc_flags,
4074 const struct alloc_context *ac)
4078 page = get_page_from_freelist(gfp_mask, order,
4079 alloc_flags|ALLOC_CPUSET, ac);
4081 * fallback to ignore cpuset restriction if our nodes
4085 page = get_page_from_freelist(gfp_mask, order,
4091 static inline struct page *
4092 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4093 const struct alloc_context *ac, unsigned long *did_some_progress)
4095 struct oom_control oc = {
4096 .zonelist = ac->zonelist,
4097 .nodemask = ac->nodemask,
4099 .gfp_mask = gfp_mask,
4104 *did_some_progress = 0;
4107 * Acquire the oom lock. If that fails, somebody else is
4108 * making progress for us.
4110 if (!mutex_trylock(&oom_lock)) {
4111 *did_some_progress = 1;
4112 schedule_timeout_uninterruptible(1);
4117 * Go through the zonelist yet one more time, keep very high watermark
4118 * here, this is only to catch a parallel oom killing, we must fail if
4119 * we're still under heavy pressure. But make sure that this reclaim
4120 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4121 * allocation which will never fail due to oom_lock already held.
4123 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4124 ~__GFP_DIRECT_RECLAIM, order,
4125 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4129 /* Coredumps can quickly deplete all memory reserves */
4130 if (current->flags & PF_DUMPCORE)
4132 /* The OOM killer will not help higher order allocs */
4133 if (order > PAGE_ALLOC_COSTLY_ORDER)
4136 * We have already exhausted all our reclaim opportunities without any
4137 * success so it is time to admit defeat. We will skip the OOM killer
4138 * because it is very likely that the caller has a more reasonable
4139 * fallback than shooting a random task.
4141 * The OOM killer may not free memory on a specific node.
4143 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4145 /* The OOM killer does not needlessly kill tasks for lowmem */
4146 if (ac->highest_zoneidx < ZONE_NORMAL)
4148 if (pm_suspended_storage())
4151 * XXX: GFP_NOFS allocations should rather fail than rely on
4152 * other request to make a forward progress.
4153 * We are in an unfortunate situation where out_of_memory cannot
4154 * do much for this context but let's try it to at least get
4155 * access to memory reserved if the current task is killed (see
4156 * out_of_memory). Once filesystems are ready to handle allocation
4157 * failures more gracefully we should just bail out here.
4160 /* Exhausted what can be done so it's blame time */
4161 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4162 *did_some_progress = 1;
4165 * Help non-failing allocations by giving them access to memory
4168 if (gfp_mask & __GFP_NOFAIL)
4169 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4170 ALLOC_NO_WATERMARKS, ac);
4173 mutex_unlock(&oom_lock);
4178 * Maximum number of compaction retries wit a progress before OOM
4179 * killer is consider as the only way to move forward.
4181 #define MAX_COMPACT_RETRIES 16
4183 #ifdef CONFIG_COMPACTION
4184 /* Try memory compaction for high-order allocations before reclaim */
4185 static struct page *
4186 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4187 unsigned int alloc_flags, const struct alloc_context *ac,
4188 enum compact_priority prio, enum compact_result *compact_result)
4190 struct page *page = NULL;
4191 unsigned long pflags;
4192 unsigned int noreclaim_flag;
4197 psi_memstall_enter(&pflags);
4198 noreclaim_flag = memalloc_noreclaim_save();
4200 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4203 memalloc_noreclaim_restore(noreclaim_flag);
4204 psi_memstall_leave(&pflags);
4207 * At least in one zone compaction wasn't deferred or skipped, so let's
4208 * count a compaction stall
4210 count_vm_event(COMPACTSTALL);
4212 /* Prep a captured page if available */
4214 prep_new_page(page, order, gfp_mask, alloc_flags);
4216 /* Try get a page from the freelist if available */
4218 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4221 struct zone *zone = page_zone(page);
4223 zone->compact_blockskip_flush = false;
4224 compaction_defer_reset(zone, order, true);
4225 count_vm_event(COMPACTSUCCESS);
4230 * It's bad if compaction run occurs and fails. The most likely reason
4231 * is that pages exist, but not enough to satisfy watermarks.
4233 count_vm_event(COMPACTFAIL);
4241 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4242 enum compact_result compact_result,
4243 enum compact_priority *compact_priority,
4244 int *compaction_retries)
4246 int max_retries = MAX_COMPACT_RETRIES;
4249 int retries = *compaction_retries;
4250 enum compact_priority priority = *compact_priority;
4255 if (compaction_made_progress(compact_result))
4256 (*compaction_retries)++;
4259 * compaction considers all the zone as desperately out of memory
4260 * so it doesn't really make much sense to retry except when the
4261 * failure could be caused by insufficient priority
4263 if (compaction_failed(compact_result))
4264 goto check_priority;
4267 * compaction was skipped because there are not enough order-0 pages
4268 * to work with, so we retry only if it looks like reclaim can help.
4270 if (compaction_needs_reclaim(compact_result)) {
4271 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4276 * make sure the compaction wasn't deferred or didn't bail out early
4277 * due to locks contention before we declare that we should give up.
4278 * But the next retry should use a higher priority if allowed, so
4279 * we don't just keep bailing out endlessly.
4281 if (compaction_withdrawn(compact_result)) {
4282 goto check_priority;
4286 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4287 * costly ones because they are de facto nofail and invoke OOM
4288 * killer to move on while costly can fail and users are ready
4289 * to cope with that. 1/4 retries is rather arbitrary but we
4290 * would need much more detailed feedback from compaction to
4291 * make a better decision.
4293 if (order > PAGE_ALLOC_COSTLY_ORDER)
4295 if (*compaction_retries <= max_retries) {
4301 * Make sure there are attempts at the highest priority if we exhausted
4302 * all retries or failed at the lower priorities.
4305 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4306 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4308 if (*compact_priority > min_priority) {
4309 (*compact_priority)--;
4310 *compaction_retries = 0;
4314 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4318 static inline struct page *
4319 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4320 unsigned int alloc_flags, const struct alloc_context *ac,
4321 enum compact_priority prio, enum compact_result *compact_result)
4323 *compact_result = COMPACT_SKIPPED;
4328 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4329 enum compact_result compact_result,
4330 enum compact_priority *compact_priority,
4331 int *compaction_retries)
4336 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4340 * There are setups with compaction disabled which would prefer to loop
4341 * inside the allocator rather than hit the oom killer prematurely.
4342 * Let's give them a good hope and keep retrying while the order-0
4343 * watermarks are OK.
4345 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4346 ac->highest_zoneidx, ac->nodemask) {
4347 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4348 ac->highest_zoneidx, alloc_flags))
4353 #endif /* CONFIG_COMPACTION */
4355 #ifdef CONFIG_LOCKDEP
4356 static struct lockdep_map __fs_reclaim_map =
4357 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4359 static bool __need_reclaim(gfp_t gfp_mask)
4361 /* no reclaim without waiting on it */
4362 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4365 /* this guy won't enter reclaim */
4366 if (current->flags & PF_MEMALLOC)
4369 if (gfp_mask & __GFP_NOLOCKDEP)
4375 void __fs_reclaim_acquire(void)
4377 lock_map_acquire(&__fs_reclaim_map);
4380 void __fs_reclaim_release(void)
4382 lock_map_release(&__fs_reclaim_map);
4385 void fs_reclaim_acquire(gfp_t gfp_mask)
4387 gfp_mask = current_gfp_context(gfp_mask);
4389 if (__need_reclaim(gfp_mask)) {
4390 if (gfp_mask & __GFP_FS)
4391 __fs_reclaim_acquire();
4393 #ifdef CONFIG_MMU_NOTIFIER
4394 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4395 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4400 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4402 void fs_reclaim_release(gfp_t gfp_mask)
4404 gfp_mask = current_gfp_context(gfp_mask);
4406 if (__need_reclaim(gfp_mask)) {
4407 if (gfp_mask & __GFP_FS)
4408 __fs_reclaim_release();
4411 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4414 /* Perform direct synchronous page reclaim */
4415 static unsigned long
4416 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4417 const struct alloc_context *ac)
4419 unsigned int noreclaim_flag;
4420 unsigned long pflags, progress;
4424 /* We now go into synchronous reclaim */
4425 cpuset_memory_pressure_bump();
4426 psi_memstall_enter(&pflags);
4427 fs_reclaim_acquire(gfp_mask);
4428 noreclaim_flag = memalloc_noreclaim_save();
4430 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4433 memalloc_noreclaim_restore(noreclaim_flag);
4434 fs_reclaim_release(gfp_mask);
4435 psi_memstall_leave(&pflags);
4442 /* The really slow allocator path where we enter direct reclaim */
4443 static inline struct page *
4444 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4445 unsigned int alloc_flags, const struct alloc_context *ac,
4446 unsigned long *did_some_progress)
4448 struct page *page = NULL;
4449 bool drained = false;
4451 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4452 if (unlikely(!(*did_some_progress)))
4456 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4459 * If an allocation failed after direct reclaim, it could be because
4460 * pages are pinned on the per-cpu lists or in high alloc reserves.
4461 * Shrink them and try again
4463 if (!page && !drained) {
4464 unreserve_highatomic_pageblock(ac, false);
4465 drain_all_pages(NULL);
4473 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4474 const struct alloc_context *ac)
4478 pg_data_t *last_pgdat = NULL;
4479 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4481 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4483 if (last_pgdat != zone->zone_pgdat)
4484 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4485 last_pgdat = zone->zone_pgdat;
4489 static inline unsigned int
4490 gfp_to_alloc_flags(gfp_t gfp_mask)
4492 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4495 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4496 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4497 * to save two branches.
4499 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4500 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4503 * The caller may dip into page reserves a bit more if the caller
4504 * cannot run direct reclaim, or if the caller has realtime scheduling
4505 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4506 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4508 alloc_flags |= (__force int)
4509 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4511 if (gfp_mask & __GFP_ATOMIC) {
4513 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4514 * if it can't schedule.
4516 if (!(gfp_mask & __GFP_NOMEMALLOC))
4517 alloc_flags |= ALLOC_HARDER;
4519 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4520 * comment for __cpuset_node_allowed().
4522 alloc_flags &= ~ALLOC_CPUSET;
4523 } else if (unlikely(rt_task(current)) && !in_interrupt())
4524 alloc_flags |= ALLOC_HARDER;
4526 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4531 static bool oom_reserves_allowed(struct task_struct *tsk)
4533 if (!tsk_is_oom_victim(tsk))
4537 * !MMU doesn't have oom reaper so give access to memory reserves
4538 * only to the thread with TIF_MEMDIE set
4540 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4547 * Distinguish requests which really need access to full memory
4548 * reserves from oom victims which can live with a portion of it
4550 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4552 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4554 if (gfp_mask & __GFP_MEMALLOC)
4555 return ALLOC_NO_WATERMARKS;
4556 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4557 return ALLOC_NO_WATERMARKS;
4558 if (!in_interrupt()) {
4559 if (current->flags & PF_MEMALLOC)
4560 return ALLOC_NO_WATERMARKS;
4561 else if (oom_reserves_allowed(current))
4568 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4570 return !!__gfp_pfmemalloc_flags(gfp_mask);
4574 * Checks whether it makes sense to retry the reclaim to make a forward progress
4575 * for the given allocation request.
4577 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4578 * without success, or when we couldn't even meet the watermark if we
4579 * reclaimed all remaining pages on the LRU lists.
4581 * Returns true if a retry is viable or false to enter the oom path.
4584 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4585 struct alloc_context *ac, int alloc_flags,
4586 bool did_some_progress, int *no_progress_loops)
4593 * Costly allocations might have made a progress but this doesn't mean
4594 * their order will become available due to high fragmentation so
4595 * always increment the no progress counter for them
4597 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4598 *no_progress_loops = 0;
4600 (*no_progress_loops)++;
4603 * Make sure we converge to OOM if we cannot make any progress
4604 * several times in the row.
4606 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4607 /* Before OOM, exhaust highatomic_reserve */
4608 return unreserve_highatomic_pageblock(ac, true);
4612 * Keep reclaiming pages while there is a chance this will lead
4613 * somewhere. If none of the target zones can satisfy our allocation
4614 * request even if all reclaimable pages are considered then we are
4615 * screwed and have to go OOM.
4617 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4618 ac->highest_zoneidx, ac->nodemask) {
4619 unsigned long available;
4620 unsigned long reclaimable;
4621 unsigned long min_wmark = min_wmark_pages(zone);
4624 available = reclaimable = zone_reclaimable_pages(zone);
4625 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4628 * Would the allocation succeed if we reclaimed all
4629 * reclaimable pages?
4631 wmark = __zone_watermark_ok(zone, order, min_wmark,
4632 ac->highest_zoneidx, alloc_flags, available);
4633 trace_reclaim_retry_zone(z, order, reclaimable,
4634 available, min_wmark, *no_progress_loops, wmark);
4637 * If we didn't make any progress and have a lot of
4638 * dirty + writeback pages then we should wait for
4639 * an IO to complete to slow down the reclaim and
4640 * prevent from pre mature OOM
4642 if (!did_some_progress) {
4643 unsigned long write_pending;
4645 write_pending = zone_page_state_snapshot(zone,
4646 NR_ZONE_WRITE_PENDING);
4648 if (2 * write_pending > reclaimable) {
4649 congestion_wait(BLK_RW_ASYNC, HZ/10);
4661 * Memory allocation/reclaim might be called from a WQ context and the
4662 * current implementation of the WQ concurrency control doesn't
4663 * recognize that a particular WQ is congested if the worker thread is
4664 * looping without ever sleeping. Therefore we have to do a short sleep
4665 * here rather than calling cond_resched().
4667 if (current->flags & PF_WQ_WORKER)
4668 schedule_timeout_uninterruptible(1);
4675 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4678 * It's possible that cpuset's mems_allowed and the nodemask from
4679 * mempolicy don't intersect. This should be normally dealt with by
4680 * policy_nodemask(), but it's possible to race with cpuset update in
4681 * such a way the check therein was true, and then it became false
4682 * before we got our cpuset_mems_cookie here.
4683 * This assumes that for all allocations, ac->nodemask can come only
4684 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4685 * when it does not intersect with the cpuset restrictions) or the
4686 * caller can deal with a violated nodemask.
4688 if (cpusets_enabled() && ac->nodemask &&
4689 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4690 ac->nodemask = NULL;
4695 * When updating a task's mems_allowed or mempolicy nodemask, it is
4696 * possible to race with parallel threads in such a way that our
4697 * allocation can fail while the mask is being updated. If we are about
4698 * to fail, check if the cpuset changed during allocation and if so,
4701 if (read_mems_allowed_retry(cpuset_mems_cookie))
4707 static inline struct page *
4708 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4709 struct alloc_context *ac)
4711 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4712 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4713 struct page *page = NULL;
4714 unsigned int alloc_flags;
4715 unsigned long did_some_progress;
4716 enum compact_priority compact_priority;
4717 enum compact_result compact_result;
4718 int compaction_retries;
4719 int no_progress_loops;
4720 unsigned int cpuset_mems_cookie;
4724 * We also sanity check to catch abuse of atomic reserves being used by
4725 * callers that are not in atomic context.
4727 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4728 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4729 gfp_mask &= ~__GFP_ATOMIC;
4732 compaction_retries = 0;
4733 no_progress_loops = 0;
4734 compact_priority = DEF_COMPACT_PRIORITY;
4735 cpuset_mems_cookie = read_mems_allowed_begin();
4738 * The fast path uses conservative alloc_flags to succeed only until
4739 * kswapd needs to be woken up, and to avoid the cost of setting up
4740 * alloc_flags precisely. So we do that now.
4742 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4745 * We need to recalculate the starting point for the zonelist iterator
4746 * because we might have used different nodemask in the fast path, or
4747 * there was a cpuset modification and we are retrying - otherwise we
4748 * could end up iterating over non-eligible zones endlessly.
4750 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4751 ac->highest_zoneidx, ac->nodemask);
4752 if (!ac->preferred_zoneref->zone)
4755 if (alloc_flags & ALLOC_KSWAPD)
4756 wake_all_kswapds(order, gfp_mask, ac);
4759 * The adjusted alloc_flags might result in immediate success, so try
4762 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4767 * For costly allocations, try direct compaction first, as it's likely
4768 * that we have enough base pages and don't need to reclaim. For non-
4769 * movable high-order allocations, do that as well, as compaction will
4770 * try prevent permanent fragmentation by migrating from blocks of the
4772 * Don't try this for allocations that are allowed to ignore
4773 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4775 if (can_direct_reclaim &&
4777 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4778 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4779 page = __alloc_pages_direct_compact(gfp_mask, order,
4781 INIT_COMPACT_PRIORITY,
4787 * Checks for costly allocations with __GFP_NORETRY, which
4788 * includes some THP page fault allocations
4790 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4792 * If allocating entire pageblock(s) and compaction
4793 * failed because all zones are below low watermarks
4794 * or is prohibited because it recently failed at this
4795 * order, fail immediately unless the allocator has
4796 * requested compaction and reclaim retry.
4799 * - potentially very expensive because zones are far
4800 * below their low watermarks or this is part of very
4801 * bursty high order allocations,
4802 * - not guaranteed to help because isolate_freepages()
4803 * may not iterate over freed pages as part of its
4805 * - unlikely to make entire pageblocks free on its
4808 if (compact_result == COMPACT_SKIPPED ||
4809 compact_result == COMPACT_DEFERRED)
4813 * Looks like reclaim/compaction is worth trying, but
4814 * sync compaction could be very expensive, so keep
4815 * using async compaction.
4817 compact_priority = INIT_COMPACT_PRIORITY;
4822 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4823 if (alloc_flags & ALLOC_KSWAPD)
4824 wake_all_kswapds(order, gfp_mask, ac);
4826 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4828 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4831 * Reset the nodemask and zonelist iterators if memory policies can be
4832 * ignored. These allocations are high priority and system rather than
4835 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4836 ac->nodemask = NULL;
4837 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4838 ac->highest_zoneidx, ac->nodemask);
4841 /* Attempt with potentially adjusted zonelist and alloc_flags */
4842 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4846 /* Caller is not willing to reclaim, we can't balance anything */
4847 if (!can_direct_reclaim)
4850 /* Avoid recursion of direct reclaim */
4851 if (current->flags & PF_MEMALLOC)
4854 /* Try direct reclaim and then allocating */
4855 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4856 &did_some_progress);
4860 /* Try direct compaction and then allocating */
4861 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4862 compact_priority, &compact_result);
4866 /* Do not loop if specifically requested */
4867 if (gfp_mask & __GFP_NORETRY)
4871 * Do not retry costly high order allocations unless they are
4872 * __GFP_RETRY_MAYFAIL
4874 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4877 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4878 did_some_progress > 0, &no_progress_loops))
4882 * It doesn't make any sense to retry for the compaction if the order-0
4883 * reclaim is not able to make any progress because the current
4884 * implementation of the compaction depends on the sufficient amount
4885 * of free memory (see __compaction_suitable)
4887 if (did_some_progress > 0 &&
4888 should_compact_retry(ac, order, alloc_flags,
4889 compact_result, &compact_priority,
4890 &compaction_retries))
4894 /* Deal with possible cpuset update races before we start OOM killing */
4895 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4898 /* Reclaim has failed us, start killing things */
4899 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4903 /* Avoid allocations with no watermarks from looping endlessly */
4904 if (tsk_is_oom_victim(current) &&
4905 (alloc_flags & ALLOC_OOM ||
4906 (gfp_mask & __GFP_NOMEMALLOC)))
4909 /* Retry as long as the OOM killer is making progress */
4910 if (did_some_progress) {
4911 no_progress_loops = 0;
4916 /* Deal with possible cpuset update races before we fail */
4917 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4921 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4924 if (gfp_mask & __GFP_NOFAIL) {
4926 * All existing users of the __GFP_NOFAIL are blockable, so warn
4927 * of any new users that actually require GFP_NOWAIT
4929 if (WARN_ON_ONCE(!can_direct_reclaim))
4933 * PF_MEMALLOC request from this context is rather bizarre
4934 * because we cannot reclaim anything and only can loop waiting
4935 * for somebody to do a work for us
4937 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4940 * non failing costly orders are a hard requirement which we
4941 * are not prepared for much so let's warn about these users
4942 * so that we can identify them and convert them to something
4945 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4948 * Help non-failing allocations by giving them access to memory
4949 * reserves but do not use ALLOC_NO_WATERMARKS because this
4950 * could deplete whole memory reserves which would just make
4951 * the situation worse
4953 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4961 warn_alloc(gfp_mask, ac->nodemask,
4962 "page allocation failure: order:%u", order);
4967 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4968 int preferred_nid, nodemask_t *nodemask,
4969 struct alloc_context *ac, gfp_t *alloc_gfp,
4970 unsigned int *alloc_flags)
4972 ac->highest_zoneidx = gfp_zone(gfp_mask);
4973 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4974 ac->nodemask = nodemask;
4975 ac->migratetype = gfp_migratetype(gfp_mask);
4977 if (cpusets_enabled()) {
4978 *alloc_gfp |= __GFP_HARDWALL;
4980 * When we are in the interrupt context, it is irrelevant
4981 * to the current task context. It means that any node ok.
4983 if (!in_interrupt() && !ac->nodemask)
4984 ac->nodemask = &cpuset_current_mems_allowed;
4986 *alloc_flags |= ALLOC_CPUSET;
4989 fs_reclaim_acquire(gfp_mask);
4990 fs_reclaim_release(gfp_mask);
4992 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4994 if (should_fail_alloc_page(gfp_mask, order))
4997 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4999 /* Dirty zone balancing only done in the fast path */
5000 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5003 * The preferred zone is used for statistics but crucially it is
5004 * also used as the starting point for the zonelist iterator. It
5005 * may get reset for allocations that ignore memory policies.
5007 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5008 ac->highest_zoneidx, ac->nodemask);
5014 * This is the 'heart' of the zoned buddy allocator.
5017 __alloc_pages_nodemask(gfp_t gfp, unsigned int order, int preferred_nid,
5018 nodemask_t *nodemask)
5021 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5022 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5023 struct alloc_context ac = { };
5026 * There are several places where we assume that the order value is sane
5027 * so bail out early if the request is out of bound.
5029 if (unlikely(order >= MAX_ORDER)) {
5030 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5034 gfp &= gfp_allowed_mask;
5036 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5037 &alloc_gfp, &alloc_flags))
5041 * Forbid the first pass from falling back to types that fragment
5042 * memory until all local zones are considered.
5044 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5046 /* First allocation attempt */
5047 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5052 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5053 * resp. GFP_NOIO which has to be inherited for all allocation requests
5054 * from a particular context which has been marked by
5055 * memalloc_no{fs,io}_{save,restore}.
5057 alloc_gfp = current_gfp_context(gfp);
5058 ac.spread_dirty_pages = false;
5061 * Restore the original nodemask if it was potentially replaced with
5062 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5064 ac.nodemask = nodemask;
5066 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5069 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5070 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5071 __free_pages(page, order);
5075 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5079 EXPORT_SYMBOL(__alloc_pages_nodemask);
5082 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5083 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5084 * you need to access high mem.
5086 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5090 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5093 return (unsigned long) page_address(page);
5095 EXPORT_SYMBOL(__get_free_pages);
5097 unsigned long get_zeroed_page(gfp_t gfp_mask)
5099 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5101 EXPORT_SYMBOL(get_zeroed_page);
5103 static inline void free_the_page(struct page *page, unsigned int order)
5105 if (order == 0) /* Via pcp? */
5106 free_unref_page(page);
5108 __free_pages_ok(page, order, FPI_NONE);
5112 * __free_pages - Free pages allocated with alloc_pages().
5113 * @page: The page pointer returned from alloc_pages().
5114 * @order: The order of the allocation.
5116 * This function can free multi-page allocations that are not compound
5117 * pages. It does not check that the @order passed in matches that of
5118 * the allocation, so it is easy to leak memory. Freeing more memory
5119 * than was allocated will probably emit a warning.
5121 * If the last reference to this page is speculative, it will be released
5122 * by put_page() which only frees the first page of a non-compound
5123 * allocation. To prevent the remaining pages from being leaked, we free
5124 * the subsequent pages here. If you want to use the page's reference
5125 * count to decide when to free the allocation, you should allocate a
5126 * compound page, and use put_page() instead of __free_pages().
5128 * Context: May be called in interrupt context or while holding a normal
5129 * spinlock, but not in NMI context or while holding a raw spinlock.
5131 void __free_pages(struct page *page, unsigned int order)
5133 if (put_page_testzero(page))
5134 free_the_page(page, order);
5135 else if (!PageHead(page))
5137 free_the_page(page + (1 << order), order);
5139 EXPORT_SYMBOL(__free_pages);
5141 void free_pages(unsigned long addr, unsigned int order)
5144 VM_BUG_ON(!virt_addr_valid((void *)addr));
5145 __free_pages(virt_to_page((void *)addr), order);
5149 EXPORT_SYMBOL(free_pages);
5153 * An arbitrary-length arbitrary-offset area of memory which resides
5154 * within a 0 or higher order page. Multiple fragments within that page
5155 * are individually refcounted, in the page's reference counter.
5157 * The page_frag functions below provide a simple allocation framework for
5158 * page fragments. This is used by the network stack and network device
5159 * drivers to provide a backing region of memory for use as either an
5160 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5162 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5165 struct page *page = NULL;
5166 gfp_t gfp = gfp_mask;
5168 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5169 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5171 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5172 PAGE_FRAG_CACHE_MAX_ORDER);
5173 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5175 if (unlikely(!page))
5176 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5178 nc->va = page ? page_address(page) : NULL;
5183 void __page_frag_cache_drain(struct page *page, unsigned int count)
5185 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5187 if (page_ref_sub_and_test(page, count))
5188 free_the_page(page, compound_order(page));
5190 EXPORT_SYMBOL(__page_frag_cache_drain);
5192 void *page_frag_alloc_align(struct page_frag_cache *nc,
5193 unsigned int fragsz, gfp_t gfp_mask,
5194 unsigned int align_mask)
5196 unsigned int size = PAGE_SIZE;
5200 if (unlikely(!nc->va)) {
5202 page = __page_frag_cache_refill(nc, gfp_mask);
5206 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5207 /* if size can vary use size else just use PAGE_SIZE */
5210 /* Even if we own the page, we do not use atomic_set().
5211 * This would break get_page_unless_zero() users.
5213 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5215 /* reset page count bias and offset to start of new frag */
5216 nc->pfmemalloc = page_is_pfmemalloc(page);
5217 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5221 offset = nc->offset - fragsz;
5222 if (unlikely(offset < 0)) {
5223 page = virt_to_page(nc->va);
5225 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5228 if (unlikely(nc->pfmemalloc)) {
5229 free_the_page(page, compound_order(page));
5233 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5234 /* if size can vary use size else just use PAGE_SIZE */
5237 /* OK, page count is 0, we can safely set it */
5238 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5240 /* reset page count bias and offset to start of new frag */
5241 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5242 offset = size - fragsz;
5246 offset &= align_mask;
5247 nc->offset = offset;
5249 return nc->va + offset;
5251 EXPORT_SYMBOL(page_frag_alloc_align);
5254 * Frees a page fragment allocated out of either a compound or order 0 page.
5256 void page_frag_free(void *addr)
5258 struct page *page = virt_to_head_page(addr);
5260 if (unlikely(put_page_testzero(page)))
5261 free_the_page(page, compound_order(page));
5263 EXPORT_SYMBOL(page_frag_free);
5265 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5269 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5270 unsigned long used = addr + PAGE_ALIGN(size);
5272 split_page(virt_to_page((void *)addr), order);
5273 while (used < alloc_end) {
5278 return (void *)addr;
5282 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5283 * @size: the number of bytes to allocate
5284 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5286 * This function is similar to alloc_pages(), except that it allocates the
5287 * minimum number of pages to satisfy the request. alloc_pages() can only
5288 * allocate memory in power-of-two pages.
5290 * This function is also limited by MAX_ORDER.
5292 * Memory allocated by this function must be released by free_pages_exact().
5294 * Return: pointer to the allocated area or %NULL in case of error.
5296 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5298 unsigned int order = get_order(size);
5301 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5302 gfp_mask &= ~__GFP_COMP;
5304 addr = __get_free_pages(gfp_mask, order);
5305 return make_alloc_exact(addr, order, size);
5307 EXPORT_SYMBOL(alloc_pages_exact);
5310 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5312 * @nid: the preferred node ID where memory should be allocated
5313 * @size: the number of bytes to allocate
5314 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5316 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5319 * Return: pointer to the allocated area or %NULL in case of error.
5321 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5323 unsigned int order = get_order(size);
5326 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5327 gfp_mask &= ~__GFP_COMP;
5329 p = alloc_pages_node(nid, gfp_mask, order);
5332 return make_alloc_exact((unsigned long)page_address(p), order, size);
5336 * free_pages_exact - release memory allocated via alloc_pages_exact()
5337 * @virt: the value returned by alloc_pages_exact.
5338 * @size: size of allocation, same value as passed to alloc_pages_exact().
5340 * Release the memory allocated by a previous call to alloc_pages_exact.
5342 void free_pages_exact(void *virt, size_t size)
5344 unsigned long addr = (unsigned long)virt;
5345 unsigned long end = addr + PAGE_ALIGN(size);
5347 while (addr < end) {
5352 EXPORT_SYMBOL(free_pages_exact);
5355 * nr_free_zone_pages - count number of pages beyond high watermark
5356 * @offset: The zone index of the highest zone
5358 * nr_free_zone_pages() counts the number of pages which are beyond the
5359 * high watermark within all zones at or below a given zone index. For each
5360 * zone, the number of pages is calculated as:
5362 * nr_free_zone_pages = managed_pages - high_pages
5364 * Return: number of pages beyond high watermark.
5366 static unsigned long nr_free_zone_pages(int offset)
5371 /* Just pick one node, since fallback list is circular */
5372 unsigned long sum = 0;
5374 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5376 for_each_zone_zonelist(zone, z, zonelist, offset) {
5377 unsigned long size = zone_managed_pages(zone);
5378 unsigned long high = high_wmark_pages(zone);
5387 * nr_free_buffer_pages - count number of pages beyond high watermark
5389 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5390 * watermark within ZONE_DMA and ZONE_NORMAL.
5392 * Return: number of pages beyond high watermark within ZONE_DMA and
5395 unsigned long nr_free_buffer_pages(void)
5397 return nr_free_zone_pages(gfp_zone(GFP_USER));
5399 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5401 static inline void show_node(struct zone *zone)
5403 if (IS_ENABLED(CONFIG_NUMA))
5404 printk("Node %d ", zone_to_nid(zone));
5407 long si_mem_available(void)
5410 unsigned long pagecache;
5411 unsigned long wmark_low = 0;
5412 unsigned long pages[NR_LRU_LISTS];
5413 unsigned long reclaimable;
5417 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5418 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5421 wmark_low += low_wmark_pages(zone);
5424 * Estimate the amount of memory available for userspace allocations,
5425 * without causing swapping.
5427 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5430 * Not all the page cache can be freed, otherwise the system will
5431 * start swapping. Assume at least half of the page cache, or the
5432 * low watermark worth of cache, needs to stay.
5434 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5435 pagecache -= min(pagecache / 2, wmark_low);
5436 available += pagecache;
5439 * Part of the reclaimable slab and other kernel memory consists of
5440 * items that are in use, and cannot be freed. Cap this estimate at the
5443 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5444 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5445 available += reclaimable - min(reclaimable / 2, wmark_low);
5451 EXPORT_SYMBOL_GPL(si_mem_available);
5453 void si_meminfo(struct sysinfo *val)
5455 val->totalram = totalram_pages();
5456 val->sharedram = global_node_page_state(NR_SHMEM);
5457 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5458 val->bufferram = nr_blockdev_pages();
5459 val->totalhigh = totalhigh_pages();
5460 val->freehigh = nr_free_highpages();
5461 val->mem_unit = PAGE_SIZE;
5464 EXPORT_SYMBOL(si_meminfo);
5467 void si_meminfo_node(struct sysinfo *val, int nid)
5469 int zone_type; /* needs to be signed */
5470 unsigned long managed_pages = 0;
5471 unsigned long managed_highpages = 0;
5472 unsigned long free_highpages = 0;
5473 pg_data_t *pgdat = NODE_DATA(nid);
5475 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5476 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5477 val->totalram = managed_pages;
5478 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5479 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5480 #ifdef CONFIG_HIGHMEM
5481 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5482 struct zone *zone = &pgdat->node_zones[zone_type];
5484 if (is_highmem(zone)) {
5485 managed_highpages += zone_managed_pages(zone);
5486 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5489 val->totalhigh = managed_highpages;
5490 val->freehigh = free_highpages;
5492 val->totalhigh = managed_highpages;
5493 val->freehigh = free_highpages;
5495 val->mem_unit = PAGE_SIZE;
5500 * Determine whether the node should be displayed or not, depending on whether
5501 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5503 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5505 if (!(flags & SHOW_MEM_FILTER_NODES))
5509 * no node mask - aka implicit memory numa policy. Do not bother with
5510 * the synchronization - read_mems_allowed_begin - because we do not
5511 * have to be precise here.
5514 nodemask = &cpuset_current_mems_allowed;
5516 return !node_isset(nid, *nodemask);
5519 #define K(x) ((x) << (PAGE_SHIFT-10))
5521 static void show_migration_types(unsigned char type)
5523 static const char types[MIGRATE_TYPES] = {
5524 [MIGRATE_UNMOVABLE] = 'U',
5525 [MIGRATE_MOVABLE] = 'M',
5526 [MIGRATE_RECLAIMABLE] = 'E',
5527 [MIGRATE_HIGHATOMIC] = 'H',
5529 [MIGRATE_CMA] = 'C',
5531 #ifdef CONFIG_MEMORY_ISOLATION
5532 [MIGRATE_ISOLATE] = 'I',
5535 char tmp[MIGRATE_TYPES + 1];
5539 for (i = 0; i < MIGRATE_TYPES; i++) {
5540 if (type & (1 << i))
5545 printk(KERN_CONT "(%s) ", tmp);
5549 * Show free area list (used inside shift_scroll-lock stuff)
5550 * We also calculate the percentage fragmentation. We do this by counting the
5551 * memory on each free list with the exception of the first item on the list.
5554 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5557 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5559 unsigned long free_pcp = 0;
5564 for_each_populated_zone(zone) {
5565 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5568 for_each_online_cpu(cpu)
5569 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5572 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5573 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5574 " unevictable:%lu dirty:%lu writeback:%lu\n"
5575 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5576 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5577 " free:%lu free_pcp:%lu free_cma:%lu\n",
5578 global_node_page_state(NR_ACTIVE_ANON),
5579 global_node_page_state(NR_INACTIVE_ANON),
5580 global_node_page_state(NR_ISOLATED_ANON),
5581 global_node_page_state(NR_ACTIVE_FILE),
5582 global_node_page_state(NR_INACTIVE_FILE),
5583 global_node_page_state(NR_ISOLATED_FILE),
5584 global_node_page_state(NR_UNEVICTABLE),
5585 global_node_page_state(NR_FILE_DIRTY),
5586 global_node_page_state(NR_WRITEBACK),
5587 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5588 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5589 global_node_page_state(NR_FILE_MAPPED),
5590 global_node_page_state(NR_SHMEM),
5591 global_node_page_state(NR_PAGETABLE),
5592 global_zone_page_state(NR_BOUNCE),
5593 global_zone_page_state(NR_FREE_PAGES),
5595 global_zone_page_state(NR_FREE_CMA_PAGES));
5597 for_each_online_pgdat(pgdat) {
5598 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5602 " active_anon:%lukB"
5603 " inactive_anon:%lukB"
5604 " active_file:%lukB"
5605 " inactive_file:%lukB"
5606 " unevictable:%lukB"
5607 " isolated(anon):%lukB"
5608 " isolated(file):%lukB"
5613 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5615 " shmem_pmdmapped: %lukB"
5618 " writeback_tmp:%lukB"
5619 " kernel_stack:%lukB"
5620 #ifdef CONFIG_SHADOW_CALL_STACK
5621 " shadow_call_stack:%lukB"
5624 " all_unreclaimable? %s"
5627 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5628 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5629 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5630 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5631 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5632 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5633 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5634 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5635 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5636 K(node_page_state(pgdat, NR_WRITEBACK)),
5637 K(node_page_state(pgdat, NR_SHMEM)),
5638 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5639 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5640 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5641 K(node_page_state(pgdat, NR_ANON_THPS)),
5643 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5644 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5645 #ifdef CONFIG_SHADOW_CALL_STACK
5646 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5648 K(node_page_state(pgdat, NR_PAGETABLE)),
5649 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5653 for_each_populated_zone(zone) {
5656 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5660 for_each_online_cpu(cpu)
5661 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5670 " reserved_highatomic:%luKB"
5671 " active_anon:%lukB"
5672 " inactive_anon:%lukB"
5673 " active_file:%lukB"
5674 " inactive_file:%lukB"
5675 " unevictable:%lukB"
5676 " writepending:%lukB"
5686 K(zone_page_state(zone, NR_FREE_PAGES)),
5687 K(min_wmark_pages(zone)),
5688 K(low_wmark_pages(zone)),
5689 K(high_wmark_pages(zone)),
5690 K(zone->nr_reserved_highatomic),
5691 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5692 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5693 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5694 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5695 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5696 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5697 K(zone->present_pages),
5698 K(zone_managed_pages(zone)),
5699 K(zone_page_state(zone, NR_MLOCK)),
5700 K(zone_page_state(zone, NR_BOUNCE)),
5702 K(this_cpu_read(zone->pageset->pcp.count)),
5703 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5704 printk("lowmem_reserve[]:");
5705 for (i = 0; i < MAX_NR_ZONES; i++)
5706 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5707 printk(KERN_CONT "\n");
5710 for_each_populated_zone(zone) {
5712 unsigned long nr[MAX_ORDER], flags, total = 0;
5713 unsigned char types[MAX_ORDER];
5715 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5718 printk(KERN_CONT "%s: ", zone->name);
5720 spin_lock_irqsave(&zone->lock, flags);
5721 for (order = 0; order < MAX_ORDER; order++) {
5722 struct free_area *area = &zone->free_area[order];
5725 nr[order] = area->nr_free;
5726 total += nr[order] << order;
5729 for (type = 0; type < MIGRATE_TYPES; type++) {
5730 if (!free_area_empty(area, type))
5731 types[order] |= 1 << type;
5734 spin_unlock_irqrestore(&zone->lock, flags);
5735 for (order = 0; order < MAX_ORDER; order++) {
5736 printk(KERN_CONT "%lu*%lukB ",
5737 nr[order], K(1UL) << order);
5739 show_migration_types(types[order]);
5741 printk(KERN_CONT "= %lukB\n", K(total));
5744 hugetlb_show_meminfo();
5746 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5748 show_swap_cache_info();
5751 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5753 zoneref->zone = zone;
5754 zoneref->zone_idx = zone_idx(zone);
5758 * Builds allocation fallback zone lists.
5760 * Add all populated zones of a node to the zonelist.
5762 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5765 enum zone_type zone_type = MAX_NR_ZONES;
5770 zone = pgdat->node_zones + zone_type;
5771 if (managed_zone(zone)) {
5772 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5773 check_highest_zone(zone_type);
5775 } while (zone_type);
5782 static int __parse_numa_zonelist_order(char *s)
5785 * We used to support different zonlists modes but they turned
5786 * out to be just not useful. Let's keep the warning in place
5787 * if somebody still use the cmd line parameter so that we do
5788 * not fail it silently
5790 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5791 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5797 char numa_zonelist_order[] = "Node";
5800 * sysctl handler for numa_zonelist_order
5802 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5803 void *buffer, size_t *length, loff_t *ppos)
5806 return __parse_numa_zonelist_order(buffer);
5807 return proc_dostring(table, write, buffer, length, ppos);
5811 #define MAX_NODE_LOAD (nr_online_nodes)
5812 static int node_load[MAX_NUMNODES];
5815 * find_next_best_node - find the next node that should appear in a given node's fallback list
5816 * @node: node whose fallback list we're appending
5817 * @used_node_mask: nodemask_t of already used nodes
5819 * We use a number of factors to determine which is the next node that should
5820 * appear on a given node's fallback list. The node should not have appeared
5821 * already in @node's fallback list, and it should be the next closest node
5822 * according to the distance array (which contains arbitrary distance values
5823 * from each node to each node in the system), and should also prefer nodes
5824 * with no CPUs, since presumably they'll have very little allocation pressure
5825 * on them otherwise.
5827 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5829 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5832 int min_val = INT_MAX;
5833 int best_node = NUMA_NO_NODE;
5835 /* Use the local node if we haven't already */
5836 if (!node_isset(node, *used_node_mask)) {
5837 node_set(node, *used_node_mask);
5841 for_each_node_state(n, N_MEMORY) {
5843 /* Don't want a node to appear more than once */
5844 if (node_isset(n, *used_node_mask))
5847 /* Use the distance array to find the distance */
5848 val = node_distance(node, n);
5850 /* Penalize nodes under us ("prefer the next node") */
5853 /* Give preference to headless and unused nodes */
5854 if (!cpumask_empty(cpumask_of_node(n)))
5855 val += PENALTY_FOR_NODE_WITH_CPUS;
5857 /* Slight preference for less loaded node */
5858 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5859 val += node_load[n];
5861 if (val < min_val) {
5868 node_set(best_node, *used_node_mask);
5875 * Build zonelists ordered by node and zones within node.
5876 * This results in maximum locality--normal zone overflows into local
5877 * DMA zone, if any--but risks exhausting DMA zone.
5879 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5882 struct zoneref *zonerefs;
5885 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5887 for (i = 0; i < nr_nodes; i++) {
5890 pg_data_t *node = NODE_DATA(node_order[i]);
5892 nr_zones = build_zonerefs_node(node, zonerefs);
5893 zonerefs += nr_zones;
5895 zonerefs->zone = NULL;
5896 zonerefs->zone_idx = 0;
5900 * Build gfp_thisnode zonelists
5902 static void build_thisnode_zonelists(pg_data_t *pgdat)
5904 struct zoneref *zonerefs;
5907 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5908 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5909 zonerefs += nr_zones;
5910 zonerefs->zone = NULL;
5911 zonerefs->zone_idx = 0;
5915 * Build zonelists ordered by zone and nodes within zones.
5916 * This results in conserving DMA zone[s] until all Normal memory is
5917 * exhausted, but results in overflowing to remote node while memory
5918 * may still exist in local DMA zone.
5921 static void build_zonelists(pg_data_t *pgdat)
5923 static int node_order[MAX_NUMNODES];
5924 int node, load, nr_nodes = 0;
5925 nodemask_t used_mask = NODE_MASK_NONE;
5926 int local_node, prev_node;
5928 /* NUMA-aware ordering of nodes */
5929 local_node = pgdat->node_id;
5930 load = nr_online_nodes;
5931 prev_node = local_node;
5933 memset(node_order, 0, sizeof(node_order));
5934 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5936 * We don't want to pressure a particular node.
5937 * So adding penalty to the first node in same
5938 * distance group to make it round-robin.
5940 if (node_distance(local_node, node) !=
5941 node_distance(local_node, prev_node))
5942 node_load[node] = load;
5944 node_order[nr_nodes++] = node;
5949 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5950 build_thisnode_zonelists(pgdat);
5953 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5955 * Return node id of node used for "local" allocations.
5956 * I.e., first node id of first zone in arg node's generic zonelist.
5957 * Used for initializing percpu 'numa_mem', which is used primarily
5958 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5960 int local_memory_node(int node)
5964 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5965 gfp_zone(GFP_KERNEL),
5967 return zone_to_nid(z->zone);
5971 static void setup_min_unmapped_ratio(void);
5972 static void setup_min_slab_ratio(void);
5973 #else /* CONFIG_NUMA */
5975 static void build_zonelists(pg_data_t *pgdat)
5977 int node, local_node;
5978 struct zoneref *zonerefs;
5981 local_node = pgdat->node_id;
5983 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5984 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5985 zonerefs += nr_zones;
5988 * Now we build the zonelist so that it contains the zones
5989 * of all the other nodes.
5990 * We don't want to pressure a particular node, so when
5991 * building the zones for node N, we make sure that the
5992 * zones coming right after the local ones are those from
5993 * node N+1 (modulo N)
5995 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5996 if (!node_online(node))
5998 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5999 zonerefs += nr_zones;
6001 for (node = 0; node < local_node; node++) {
6002 if (!node_online(node))
6004 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6005 zonerefs += nr_zones;
6008 zonerefs->zone = NULL;
6009 zonerefs->zone_idx = 0;
6012 #endif /* CONFIG_NUMA */
6015 * Boot pageset table. One per cpu which is going to be used for all
6016 * zones and all nodes. The parameters will be set in such a way
6017 * that an item put on a list will immediately be handed over to
6018 * the buddy list. This is safe since pageset manipulation is done
6019 * with interrupts disabled.
6021 * The boot_pagesets must be kept even after bootup is complete for
6022 * unused processors and/or zones. They do play a role for bootstrapping
6023 * hotplugged processors.
6025 * zoneinfo_show() and maybe other functions do
6026 * not check if the processor is online before following the pageset pointer.
6027 * Other parts of the kernel may not check if the zone is available.
6029 static void pageset_init(struct per_cpu_pageset *p);
6030 /* These effectively disable the pcplists in the boot pageset completely */
6031 #define BOOT_PAGESET_HIGH 0
6032 #define BOOT_PAGESET_BATCH 1
6033 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
6034 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6036 static void __build_all_zonelists(void *data)
6039 int __maybe_unused cpu;
6040 pg_data_t *self = data;
6041 static DEFINE_SPINLOCK(lock);
6046 memset(node_load, 0, sizeof(node_load));
6050 * This node is hotadded and no memory is yet present. So just
6051 * building zonelists is fine - no need to touch other nodes.
6053 if (self && !node_online(self->node_id)) {
6054 build_zonelists(self);
6056 for_each_online_node(nid) {
6057 pg_data_t *pgdat = NODE_DATA(nid);
6059 build_zonelists(pgdat);
6062 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6064 * We now know the "local memory node" for each node--
6065 * i.e., the node of the first zone in the generic zonelist.
6066 * Set up numa_mem percpu variable for on-line cpus. During
6067 * boot, only the boot cpu should be on-line; we'll init the
6068 * secondary cpus' numa_mem as they come on-line. During
6069 * node/memory hotplug, we'll fixup all on-line cpus.
6071 for_each_online_cpu(cpu)
6072 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6079 static noinline void __init
6080 build_all_zonelists_init(void)
6084 __build_all_zonelists(NULL);
6087 * Initialize the boot_pagesets that are going to be used
6088 * for bootstrapping processors. The real pagesets for
6089 * each zone will be allocated later when the per cpu
6090 * allocator is available.
6092 * boot_pagesets are used also for bootstrapping offline
6093 * cpus if the system is already booted because the pagesets
6094 * are needed to initialize allocators on a specific cpu too.
6095 * F.e. the percpu allocator needs the page allocator which
6096 * needs the percpu allocator in order to allocate its pagesets
6097 * (a chicken-egg dilemma).
6099 for_each_possible_cpu(cpu)
6100 pageset_init(&per_cpu(boot_pageset, cpu));
6102 mminit_verify_zonelist();
6103 cpuset_init_current_mems_allowed();
6107 * unless system_state == SYSTEM_BOOTING.
6109 * __ref due to call of __init annotated helper build_all_zonelists_init
6110 * [protected by SYSTEM_BOOTING].
6112 void __ref build_all_zonelists(pg_data_t *pgdat)
6114 unsigned long vm_total_pages;
6116 if (system_state == SYSTEM_BOOTING) {
6117 build_all_zonelists_init();
6119 __build_all_zonelists(pgdat);
6120 /* cpuset refresh routine should be here */
6122 /* Get the number of free pages beyond high watermark in all zones. */
6123 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6125 * Disable grouping by mobility if the number of pages in the
6126 * system is too low to allow the mechanism to work. It would be
6127 * more accurate, but expensive to check per-zone. This check is
6128 * made on memory-hotadd so a system can start with mobility
6129 * disabled and enable it later
6131 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6132 page_group_by_mobility_disabled = 1;
6134 page_group_by_mobility_disabled = 0;
6136 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6138 page_group_by_mobility_disabled ? "off" : "on",
6141 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6145 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6146 static bool __meminit
6147 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6149 static struct memblock_region *r;
6151 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6152 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6153 for_each_mem_region(r) {
6154 if (*pfn < memblock_region_memory_end_pfn(r))
6158 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6159 memblock_is_mirror(r)) {
6160 *pfn = memblock_region_memory_end_pfn(r);
6168 * Initially all pages are reserved - free ones are freed
6169 * up by memblock_free_all() once the early boot process is
6170 * done. Non-atomic initialization, single-pass.
6172 * All aligned pageblocks are initialized to the specified migratetype
6173 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6174 * zone stats (e.g., nr_isolate_pageblock) are touched.
6176 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6177 unsigned long start_pfn, unsigned long zone_end_pfn,
6178 enum meminit_context context,
6179 struct vmem_altmap *altmap, int migratetype)
6181 unsigned long pfn, end_pfn = start_pfn + size;
6184 if (highest_memmap_pfn < end_pfn - 1)
6185 highest_memmap_pfn = end_pfn - 1;
6187 #ifdef CONFIG_ZONE_DEVICE
6189 * Honor reservation requested by the driver for this ZONE_DEVICE
6190 * memory. We limit the total number of pages to initialize to just
6191 * those that might contain the memory mapping. We will defer the
6192 * ZONE_DEVICE page initialization until after we have released
6195 if (zone == ZONE_DEVICE) {
6199 if (start_pfn == altmap->base_pfn)
6200 start_pfn += altmap->reserve;
6201 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6205 for (pfn = start_pfn; pfn < end_pfn; ) {
6207 * There can be holes in boot-time mem_map[]s handed to this
6208 * function. They do not exist on hotplugged memory.
6210 if (context == MEMINIT_EARLY) {
6211 if (overlap_memmap_init(zone, &pfn))
6213 if (defer_init(nid, pfn, zone_end_pfn))
6217 page = pfn_to_page(pfn);
6218 __init_single_page(page, pfn, zone, nid);
6219 if (context == MEMINIT_HOTPLUG)
6220 __SetPageReserved(page);
6223 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6224 * such that unmovable allocations won't be scattered all
6225 * over the place during system boot.
6227 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6228 set_pageblock_migratetype(page, migratetype);
6235 #ifdef CONFIG_ZONE_DEVICE
6236 void __ref memmap_init_zone_device(struct zone *zone,
6237 unsigned long start_pfn,
6238 unsigned long nr_pages,
6239 struct dev_pagemap *pgmap)
6241 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6242 struct pglist_data *pgdat = zone->zone_pgdat;
6243 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6244 unsigned long zone_idx = zone_idx(zone);
6245 unsigned long start = jiffies;
6246 int nid = pgdat->node_id;
6248 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6252 * The call to memmap_init_zone should have already taken care
6253 * of the pages reserved for the memmap, so we can just jump to
6254 * the end of that region and start processing the device pages.
6257 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6258 nr_pages = end_pfn - start_pfn;
6261 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6262 struct page *page = pfn_to_page(pfn);
6264 __init_single_page(page, pfn, zone_idx, nid);
6267 * Mark page reserved as it will need to wait for onlining
6268 * phase for it to be fully associated with a zone.
6270 * We can use the non-atomic __set_bit operation for setting
6271 * the flag as we are still initializing the pages.
6273 __SetPageReserved(page);
6276 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6277 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6278 * ever freed or placed on a driver-private list.
6280 page->pgmap = pgmap;
6281 page->zone_device_data = NULL;
6284 * Mark the block movable so that blocks are reserved for
6285 * movable at startup. This will force kernel allocations
6286 * to reserve their blocks rather than leaking throughout
6287 * the address space during boot when many long-lived
6288 * kernel allocations are made.
6290 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6291 * because this is done early in section_activate()
6293 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6294 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6299 pr_info("%s initialised %lu pages in %ums\n", __func__,
6300 nr_pages, jiffies_to_msecs(jiffies - start));
6304 static void __meminit zone_init_free_lists(struct zone *zone)
6306 unsigned int order, t;
6307 for_each_migratetype_order(order, t) {
6308 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6309 zone->free_area[order].nr_free = 0;
6313 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6315 * Only struct pages that correspond to ranges defined by memblock.memory
6316 * are zeroed and initialized by going through __init_single_page() during
6317 * memmap_init_zone().
6319 * But, there could be struct pages that correspond to holes in
6320 * memblock.memory. This can happen because of the following reasons:
6321 * - physical memory bank size is not necessarily the exact multiple of the
6322 * arbitrary section size
6323 * - early reserved memory may not be listed in memblock.memory
6324 * - memory layouts defined with memmap= kernel parameter may not align
6325 * nicely with memmap sections
6327 * Explicitly initialize those struct pages so that:
6328 * - PG_Reserved is set
6329 * - zone and node links point to zone and node that span the page if the
6330 * hole is in the middle of a zone
6331 * - zone and node links point to adjacent zone/node if the hole falls on
6332 * the zone boundary; the pages in such holes will be prepended to the
6333 * zone/node above the hole except for the trailing pages in the last
6334 * section that will be appended to the zone/node below.
6336 static u64 __meminit init_unavailable_range(unsigned long spfn,
6343 for (pfn = spfn; pfn < epfn; pfn++) {
6344 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6345 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6346 + pageblock_nr_pages - 1;
6349 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6350 __SetPageReserved(pfn_to_page(pfn));
6357 static inline u64 init_unavailable_range(unsigned long spfn, unsigned long epfn,
6364 void __meminit __weak memmap_init_zone(struct zone *zone)
6366 unsigned long zone_start_pfn = zone->zone_start_pfn;
6367 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6368 int i, nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6369 static unsigned long hole_pfn;
6370 unsigned long start_pfn, end_pfn;
6373 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6374 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6375 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6377 if (end_pfn > start_pfn)
6378 memmap_init_range(end_pfn - start_pfn, nid,
6379 zone_id, start_pfn, zone_end_pfn,
6380 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6382 if (hole_pfn < start_pfn)
6383 pgcnt += init_unavailable_range(hole_pfn, start_pfn,
6388 #ifdef CONFIG_SPARSEMEM
6390 * Initialize the hole in the range [zone_end_pfn, section_end].
6391 * If zone boundary falls in the middle of a section, this hole
6392 * will be re-initialized during the call to this function for the
6395 end_pfn = round_up(zone_end_pfn, PAGES_PER_SECTION);
6396 if (hole_pfn < end_pfn)
6397 pgcnt += init_unavailable_range(hole_pfn, end_pfn,
6402 pr_info(" %s zone: %llu pages in unavailable ranges\n",
6406 static int zone_batchsize(struct zone *zone)
6412 * The per-cpu-pages pools are set to around 1000th of the
6415 batch = zone_managed_pages(zone) / 1024;
6416 /* But no more than a meg. */
6417 if (batch * PAGE_SIZE > 1024 * 1024)
6418 batch = (1024 * 1024) / PAGE_SIZE;
6419 batch /= 4; /* We effectively *= 4 below */
6424 * Clamp the batch to a 2^n - 1 value. Having a power
6425 * of 2 value was found to be more likely to have
6426 * suboptimal cache aliasing properties in some cases.
6428 * For example if 2 tasks are alternately allocating
6429 * batches of pages, one task can end up with a lot
6430 * of pages of one half of the possible page colors
6431 * and the other with pages of the other colors.
6433 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6438 /* The deferral and batching of frees should be suppressed under NOMMU
6441 * The problem is that NOMMU needs to be able to allocate large chunks
6442 * of contiguous memory as there's no hardware page translation to
6443 * assemble apparent contiguous memory from discontiguous pages.
6445 * Queueing large contiguous runs of pages for batching, however,
6446 * causes the pages to actually be freed in smaller chunks. As there
6447 * can be a significant delay between the individual batches being
6448 * recycled, this leads to the once large chunks of space being
6449 * fragmented and becoming unavailable for high-order allocations.
6456 * pcp->high and pcp->batch values are related and generally batch is lower
6457 * than high. They are also related to pcp->count such that count is lower
6458 * than high, and as soon as it reaches high, the pcplist is flushed.
6460 * However, guaranteeing these relations at all times would require e.g. write
6461 * barriers here but also careful usage of read barriers at the read side, and
6462 * thus be prone to error and bad for performance. Thus the update only prevents
6463 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6464 * can cope with those fields changing asynchronously, and fully trust only the
6465 * pcp->count field on the local CPU with interrupts disabled.
6467 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6468 * outside of boot time (or some other assurance that no concurrent updaters
6471 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6472 unsigned long batch)
6474 WRITE_ONCE(pcp->batch, batch);
6475 WRITE_ONCE(pcp->high, high);
6478 static void pageset_init(struct per_cpu_pageset *p)
6480 struct per_cpu_pages *pcp;
6483 memset(p, 0, sizeof(*p));
6486 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6487 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6490 * Set batch and high values safe for a boot pageset. A true percpu
6491 * pageset's initialization will update them subsequently. Here we don't
6492 * need to be as careful as pageset_update() as nobody can access the
6495 pcp->high = BOOT_PAGESET_HIGH;
6496 pcp->batch = BOOT_PAGESET_BATCH;
6499 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6500 unsigned long batch)
6502 struct per_cpu_pageset *p;
6505 for_each_possible_cpu(cpu) {
6506 p = per_cpu_ptr(zone->pageset, cpu);
6507 pageset_update(&p->pcp, high, batch);
6512 * Calculate and set new high and batch values for all per-cpu pagesets of a
6513 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6515 static void zone_set_pageset_high_and_batch(struct zone *zone)
6517 unsigned long new_high, new_batch;
6519 if (percpu_pagelist_fraction) {
6520 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6521 new_batch = max(1UL, new_high / 4);
6522 if ((new_high / 4) > (PAGE_SHIFT * 8))
6523 new_batch = PAGE_SHIFT * 8;
6525 new_batch = zone_batchsize(zone);
6526 new_high = 6 * new_batch;
6527 new_batch = max(1UL, 1 * new_batch);
6530 if (zone->pageset_high == new_high &&
6531 zone->pageset_batch == new_batch)
6534 zone->pageset_high = new_high;
6535 zone->pageset_batch = new_batch;
6537 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6540 void __meminit setup_zone_pageset(struct zone *zone)
6542 struct per_cpu_pageset *p;
6545 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6546 for_each_possible_cpu(cpu) {
6547 p = per_cpu_ptr(zone->pageset, cpu);
6551 zone_set_pageset_high_and_batch(zone);
6555 * Allocate per cpu pagesets and initialize them.
6556 * Before this call only boot pagesets were available.
6558 void __init setup_per_cpu_pageset(void)
6560 struct pglist_data *pgdat;
6562 int __maybe_unused cpu;
6564 for_each_populated_zone(zone)
6565 setup_zone_pageset(zone);
6569 * Unpopulated zones continue using the boot pagesets.
6570 * The numa stats for these pagesets need to be reset.
6571 * Otherwise, they will end up skewing the stats of
6572 * the nodes these zones are associated with.
6574 for_each_possible_cpu(cpu) {
6575 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6576 memset(pcp->vm_numa_stat_diff, 0,
6577 sizeof(pcp->vm_numa_stat_diff));
6581 for_each_online_pgdat(pgdat)
6582 pgdat->per_cpu_nodestats =
6583 alloc_percpu(struct per_cpu_nodestat);
6586 static __meminit void zone_pcp_init(struct zone *zone)
6589 * per cpu subsystem is not up at this point. The following code
6590 * relies on the ability of the linker to provide the
6591 * offset of a (static) per cpu variable into the per cpu area.
6593 zone->pageset = &boot_pageset;
6594 zone->pageset_high = BOOT_PAGESET_HIGH;
6595 zone->pageset_batch = BOOT_PAGESET_BATCH;
6597 if (populated_zone(zone))
6598 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6599 zone->name, zone->present_pages,
6600 zone_batchsize(zone));
6603 void __meminit init_currently_empty_zone(struct zone *zone,
6604 unsigned long zone_start_pfn,
6607 struct pglist_data *pgdat = zone->zone_pgdat;
6608 int zone_idx = zone_idx(zone) + 1;
6610 if (zone_idx > pgdat->nr_zones)
6611 pgdat->nr_zones = zone_idx;
6613 zone->zone_start_pfn = zone_start_pfn;
6615 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6616 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6618 (unsigned long)zone_idx(zone),
6619 zone_start_pfn, (zone_start_pfn + size));
6621 zone_init_free_lists(zone);
6622 zone->initialized = 1;
6626 * get_pfn_range_for_nid - Return the start and end page frames for a node
6627 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6628 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6629 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6631 * It returns the start and end page frame of a node based on information
6632 * provided by memblock_set_node(). If called for a node
6633 * with no available memory, a warning is printed and the start and end
6636 void __init get_pfn_range_for_nid(unsigned int nid,
6637 unsigned long *start_pfn, unsigned long *end_pfn)
6639 unsigned long this_start_pfn, this_end_pfn;
6645 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6646 *start_pfn = min(*start_pfn, this_start_pfn);
6647 *end_pfn = max(*end_pfn, this_end_pfn);
6650 if (*start_pfn == -1UL)
6655 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6656 * assumption is made that zones within a node are ordered in monotonic
6657 * increasing memory addresses so that the "highest" populated zone is used
6659 static void __init find_usable_zone_for_movable(void)
6662 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6663 if (zone_index == ZONE_MOVABLE)
6666 if (arch_zone_highest_possible_pfn[zone_index] >
6667 arch_zone_lowest_possible_pfn[zone_index])
6671 VM_BUG_ON(zone_index == -1);
6672 movable_zone = zone_index;
6676 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6677 * because it is sized independent of architecture. Unlike the other zones,
6678 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6679 * in each node depending on the size of each node and how evenly kernelcore
6680 * is distributed. This helper function adjusts the zone ranges
6681 * provided by the architecture for a given node by using the end of the
6682 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6683 * zones within a node are in order of monotonic increases memory addresses
6685 static void __init adjust_zone_range_for_zone_movable(int nid,
6686 unsigned long zone_type,
6687 unsigned long node_start_pfn,
6688 unsigned long node_end_pfn,
6689 unsigned long *zone_start_pfn,
6690 unsigned long *zone_end_pfn)
6692 /* Only adjust if ZONE_MOVABLE is on this node */
6693 if (zone_movable_pfn[nid]) {
6694 /* Size ZONE_MOVABLE */
6695 if (zone_type == ZONE_MOVABLE) {
6696 *zone_start_pfn = zone_movable_pfn[nid];
6697 *zone_end_pfn = min(node_end_pfn,
6698 arch_zone_highest_possible_pfn[movable_zone]);
6700 /* Adjust for ZONE_MOVABLE starting within this range */
6701 } else if (!mirrored_kernelcore &&
6702 *zone_start_pfn < zone_movable_pfn[nid] &&
6703 *zone_end_pfn > zone_movable_pfn[nid]) {
6704 *zone_end_pfn = zone_movable_pfn[nid];
6706 /* Check if this whole range is within ZONE_MOVABLE */
6707 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6708 *zone_start_pfn = *zone_end_pfn;
6713 * Return the number of pages a zone spans in a node, including holes
6714 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6716 static unsigned long __init zone_spanned_pages_in_node(int nid,
6717 unsigned long zone_type,
6718 unsigned long node_start_pfn,
6719 unsigned long node_end_pfn,
6720 unsigned long *zone_start_pfn,
6721 unsigned long *zone_end_pfn)
6723 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6724 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6725 /* When hotadd a new node from cpu_up(), the node should be empty */
6726 if (!node_start_pfn && !node_end_pfn)
6729 /* Get the start and end of the zone */
6730 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6731 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6732 adjust_zone_range_for_zone_movable(nid, zone_type,
6733 node_start_pfn, node_end_pfn,
6734 zone_start_pfn, zone_end_pfn);
6736 /* Check that this node has pages within the zone's required range */
6737 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6740 /* Move the zone boundaries inside the node if necessary */
6741 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6742 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6744 /* Return the spanned pages */
6745 return *zone_end_pfn - *zone_start_pfn;
6749 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6750 * then all holes in the requested range will be accounted for.
6752 unsigned long __init __absent_pages_in_range(int nid,
6753 unsigned long range_start_pfn,
6754 unsigned long range_end_pfn)
6756 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6757 unsigned long start_pfn, end_pfn;
6760 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6761 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6762 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6763 nr_absent -= end_pfn - start_pfn;
6769 * absent_pages_in_range - Return number of page frames in holes within a range
6770 * @start_pfn: The start PFN to start searching for holes
6771 * @end_pfn: The end PFN to stop searching for holes
6773 * Return: the number of pages frames in memory holes within a range.
6775 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6776 unsigned long end_pfn)
6778 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6781 /* Return the number of page frames in holes in a zone on a node */
6782 static unsigned long __init zone_absent_pages_in_node(int nid,
6783 unsigned long zone_type,
6784 unsigned long node_start_pfn,
6785 unsigned long node_end_pfn)
6787 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6788 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6789 unsigned long zone_start_pfn, zone_end_pfn;
6790 unsigned long nr_absent;
6792 /* When hotadd a new node from cpu_up(), the node should be empty */
6793 if (!node_start_pfn && !node_end_pfn)
6796 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6797 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6799 adjust_zone_range_for_zone_movable(nid, zone_type,
6800 node_start_pfn, node_end_pfn,
6801 &zone_start_pfn, &zone_end_pfn);
6802 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6805 * ZONE_MOVABLE handling.
6806 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6809 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6810 unsigned long start_pfn, end_pfn;
6811 struct memblock_region *r;
6813 for_each_mem_region(r) {
6814 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6815 zone_start_pfn, zone_end_pfn);
6816 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6817 zone_start_pfn, zone_end_pfn);
6819 if (zone_type == ZONE_MOVABLE &&
6820 memblock_is_mirror(r))
6821 nr_absent += end_pfn - start_pfn;
6823 if (zone_type == ZONE_NORMAL &&
6824 !memblock_is_mirror(r))
6825 nr_absent += end_pfn - start_pfn;
6832 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6833 unsigned long node_start_pfn,
6834 unsigned long node_end_pfn)
6836 unsigned long realtotalpages = 0, totalpages = 0;
6839 for (i = 0; i < MAX_NR_ZONES; i++) {
6840 struct zone *zone = pgdat->node_zones + i;
6841 unsigned long zone_start_pfn, zone_end_pfn;
6842 unsigned long spanned, absent;
6843 unsigned long size, real_size;
6845 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6850 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6855 real_size = size - absent;
6858 zone->zone_start_pfn = zone_start_pfn;
6860 zone->zone_start_pfn = 0;
6861 zone->spanned_pages = size;
6862 zone->present_pages = real_size;
6865 realtotalpages += real_size;
6868 pgdat->node_spanned_pages = totalpages;
6869 pgdat->node_present_pages = realtotalpages;
6870 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6874 #ifndef CONFIG_SPARSEMEM
6876 * Calculate the size of the zone->blockflags rounded to an unsigned long
6877 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6878 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6879 * round what is now in bits to nearest long in bits, then return it in
6882 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6884 unsigned long usemapsize;
6886 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6887 usemapsize = roundup(zonesize, pageblock_nr_pages);
6888 usemapsize = usemapsize >> pageblock_order;
6889 usemapsize *= NR_PAGEBLOCK_BITS;
6890 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6892 return usemapsize / 8;
6895 static void __ref setup_usemap(struct zone *zone)
6897 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
6898 zone->spanned_pages);
6899 zone->pageblock_flags = NULL;
6901 zone->pageblock_flags =
6902 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6904 if (!zone->pageblock_flags)
6905 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6906 usemapsize, zone->name, zone_to_nid(zone));
6910 static inline void setup_usemap(struct zone *zone) {}
6911 #endif /* CONFIG_SPARSEMEM */
6913 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6915 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6916 void __init set_pageblock_order(void)
6920 /* Check that pageblock_nr_pages has not already been setup */
6921 if (pageblock_order)
6924 if (HPAGE_SHIFT > PAGE_SHIFT)
6925 order = HUGETLB_PAGE_ORDER;
6927 order = MAX_ORDER - 1;
6930 * Assume the largest contiguous order of interest is a huge page.
6931 * This value may be variable depending on boot parameters on IA64 and
6934 pageblock_order = order;
6936 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6939 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6940 * is unused as pageblock_order is set at compile-time. See
6941 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6944 void __init set_pageblock_order(void)
6948 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6950 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6951 unsigned long present_pages)
6953 unsigned long pages = spanned_pages;
6956 * Provide a more accurate estimation if there are holes within
6957 * the zone and SPARSEMEM is in use. If there are holes within the
6958 * zone, each populated memory region may cost us one or two extra
6959 * memmap pages due to alignment because memmap pages for each
6960 * populated regions may not be naturally aligned on page boundary.
6961 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6963 if (spanned_pages > present_pages + (present_pages >> 4) &&
6964 IS_ENABLED(CONFIG_SPARSEMEM))
6965 pages = present_pages;
6967 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6970 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6971 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6973 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6975 spin_lock_init(&ds_queue->split_queue_lock);
6976 INIT_LIST_HEAD(&ds_queue->split_queue);
6977 ds_queue->split_queue_len = 0;
6980 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6983 #ifdef CONFIG_COMPACTION
6984 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6986 init_waitqueue_head(&pgdat->kcompactd_wait);
6989 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6992 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6994 pgdat_resize_init(pgdat);
6996 pgdat_init_split_queue(pgdat);
6997 pgdat_init_kcompactd(pgdat);
6999 init_waitqueue_head(&pgdat->kswapd_wait);
7000 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7002 pgdat_page_ext_init(pgdat);
7003 lruvec_init(&pgdat->__lruvec);
7006 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7007 unsigned long remaining_pages)
7009 atomic_long_set(&zone->managed_pages, remaining_pages);
7010 zone_set_nid(zone, nid);
7011 zone->name = zone_names[idx];
7012 zone->zone_pgdat = NODE_DATA(nid);
7013 spin_lock_init(&zone->lock);
7014 zone_seqlock_init(zone);
7015 zone_pcp_init(zone);
7019 * Set up the zone data structures
7020 * - init pgdat internals
7021 * - init all zones belonging to this node
7023 * NOTE: this function is only called during memory hotplug
7025 #ifdef CONFIG_MEMORY_HOTPLUG
7026 void __ref free_area_init_core_hotplug(int nid)
7029 pg_data_t *pgdat = NODE_DATA(nid);
7031 pgdat_init_internals(pgdat);
7032 for (z = 0; z < MAX_NR_ZONES; z++)
7033 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7038 * Set up the zone data structures:
7039 * - mark all pages reserved
7040 * - mark all memory queues empty
7041 * - clear the memory bitmaps
7043 * NOTE: pgdat should get zeroed by caller.
7044 * NOTE: this function is only called during early init.
7046 static void __init free_area_init_core(struct pglist_data *pgdat)
7049 int nid = pgdat->node_id;
7051 pgdat_init_internals(pgdat);
7052 pgdat->per_cpu_nodestats = &boot_nodestats;
7054 for (j = 0; j < MAX_NR_ZONES; j++) {
7055 struct zone *zone = pgdat->node_zones + j;
7056 unsigned long size, freesize, memmap_pages;
7058 size = zone->spanned_pages;
7059 freesize = zone->present_pages;
7062 * Adjust freesize so that it accounts for how much memory
7063 * is used by this zone for memmap. This affects the watermark
7064 * and per-cpu initialisations
7066 memmap_pages = calc_memmap_size(size, freesize);
7067 if (!is_highmem_idx(j)) {
7068 if (freesize >= memmap_pages) {
7069 freesize -= memmap_pages;
7072 " %s zone: %lu pages used for memmap\n",
7073 zone_names[j], memmap_pages);
7075 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
7076 zone_names[j], memmap_pages, freesize);
7079 /* Account for reserved pages */
7080 if (j == 0 && freesize > dma_reserve) {
7081 freesize -= dma_reserve;
7082 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
7083 zone_names[0], dma_reserve);
7086 if (!is_highmem_idx(j))
7087 nr_kernel_pages += freesize;
7088 /* Charge for highmem memmap if there are enough kernel pages */
7089 else if (nr_kernel_pages > memmap_pages * 2)
7090 nr_kernel_pages -= memmap_pages;
7091 nr_all_pages += freesize;
7094 * Set an approximate value for lowmem here, it will be adjusted
7095 * when the bootmem allocator frees pages into the buddy system.
7096 * And all highmem pages will be managed by the buddy system.
7098 zone_init_internals(zone, j, nid, freesize);
7103 set_pageblock_order();
7105 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7106 memmap_init_zone(zone);
7110 #ifdef CONFIG_FLAT_NODE_MEM_MAP
7111 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7113 unsigned long __maybe_unused start = 0;
7114 unsigned long __maybe_unused offset = 0;
7116 /* Skip empty nodes */
7117 if (!pgdat->node_spanned_pages)
7120 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7121 offset = pgdat->node_start_pfn - start;
7122 /* ia64 gets its own node_mem_map, before this, without bootmem */
7123 if (!pgdat->node_mem_map) {
7124 unsigned long size, end;
7128 * The zone's endpoints aren't required to be MAX_ORDER
7129 * aligned but the node_mem_map endpoints must be in order
7130 * for the buddy allocator to function correctly.
7132 end = pgdat_end_pfn(pgdat);
7133 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7134 size = (end - start) * sizeof(struct page);
7135 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7138 panic("Failed to allocate %ld bytes for node %d memory map\n",
7139 size, pgdat->node_id);
7140 pgdat->node_mem_map = map + offset;
7142 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7143 __func__, pgdat->node_id, (unsigned long)pgdat,
7144 (unsigned long)pgdat->node_mem_map);
7145 #ifndef CONFIG_NEED_MULTIPLE_NODES
7147 * With no DISCONTIG, the global mem_map is just set as node 0's
7149 if (pgdat == NODE_DATA(0)) {
7150 mem_map = NODE_DATA(0)->node_mem_map;
7151 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7157 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7158 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7160 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7161 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7163 pgdat->first_deferred_pfn = ULONG_MAX;
7166 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7169 static void __init free_area_init_node(int nid)
7171 pg_data_t *pgdat = NODE_DATA(nid);
7172 unsigned long start_pfn = 0;
7173 unsigned long end_pfn = 0;
7175 /* pg_data_t should be reset to zero when it's allocated */
7176 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7178 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7180 pgdat->node_id = nid;
7181 pgdat->node_start_pfn = start_pfn;
7182 pgdat->per_cpu_nodestats = NULL;
7184 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7185 (u64)start_pfn << PAGE_SHIFT,
7186 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7187 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7189 alloc_node_mem_map(pgdat);
7190 pgdat_set_deferred_range(pgdat);
7192 free_area_init_core(pgdat);
7195 void __init free_area_init_memoryless_node(int nid)
7197 free_area_init_node(nid);
7200 #if MAX_NUMNODES > 1
7202 * Figure out the number of possible node ids.
7204 void __init setup_nr_node_ids(void)
7206 unsigned int highest;
7208 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7209 nr_node_ids = highest + 1;
7214 * node_map_pfn_alignment - determine the maximum internode alignment
7216 * This function should be called after node map is populated and sorted.
7217 * It calculates the maximum power of two alignment which can distinguish
7220 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7221 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7222 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7223 * shifted, 1GiB is enough and this function will indicate so.
7225 * This is used to test whether pfn -> nid mapping of the chosen memory
7226 * model has fine enough granularity to avoid incorrect mapping for the
7227 * populated node map.
7229 * Return: the determined alignment in pfn's. 0 if there is no alignment
7230 * requirement (single node).
7232 unsigned long __init node_map_pfn_alignment(void)
7234 unsigned long accl_mask = 0, last_end = 0;
7235 unsigned long start, end, mask;
7236 int last_nid = NUMA_NO_NODE;
7239 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7240 if (!start || last_nid < 0 || last_nid == nid) {
7247 * Start with a mask granular enough to pin-point to the
7248 * start pfn and tick off bits one-by-one until it becomes
7249 * too coarse to separate the current node from the last.
7251 mask = ~((1 << __ffs(start)) - 1);
7252 while (mask && last_end <= (start & (mask << 1)))
7255 /* accumulate all internode masks */
7259 /* convert mask to number of pages */
7260 return ~accl_mask + 1;
7264 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7266 * Return: the minimum PFN based on information provided via
7267 * memblock_set_node().
7269 unsigned long __init find_min_pfn_with_active_regions(void)
7271 return PHYS_PFN(memblock_start_of_DRAM());
7275 * early_calculate_totalpages()
7276 * Sum pages in active regions for movable zone.
7277 * Populate N_MEMORY for calculating usable_nodes.
7279 static unsigned long __init early_calculate_totalpages(void)
7281 unsigned long totalpages = 0;
7282 unsigned long start_pfn, end_pfn;
7285 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7286 unsigned long pages = end_pfn - start_pfn;
7288 totalpages += pages;
7290 node_set_state(nid, N_MEMORY);
7296 * Find the PFN the Movable zone begins in each node. Kernel memory
7297 * is spread evenly between nodes as long as the nodes have enough
7298 * memory. When they don't, some nodes will have more kernelcore than
7301 static void __init find_zone_movable_pfns_for_nodes(void)
7304 unsigned long usable_startpfn;
7305 unsigned long kernelcore_node, kernelcore_remaining;
7306 /* save the state before borrow the nodemask */
7307 nodemask_t saved_node_state = node_states[N_MEMORY];
7308 unsigned long totalpages = early_calculate_totalpages();
7309 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7310 struct memblock_region *r;
7312 /* Need to find movable_zone earlier when movable_node is specified. */
7313 find_usable_zone_for_movable();
7316 * If movable_node is specified, ignore kernelcore and movablecore
7319 if (movable_node_is_enabled()) {
7320 for_each_mem_region(r) {
7321 if (!memblock_is_hotpluggable(r))
7324 nid = memblock_get_region_node(r);
7326 usable_startpfn = PFN_DOWN(r->base);
7327 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7328 min(usable_startpfn, zone_movable_pfn[nid]) :
7336 * If kernelcore=mirror is specified, ignore movablecore option
7338 if (mirrored_kernelcore) {
7339 bool mem_below_4gb_not_mirrored = false;
7341 for_each_mem_region(r) {
7342 if (memblock_is_mirror(r))
7345 nid = memblock_get_region_node(r);
7347 usable_startpfn = memblock_region_memory_base_pfn(r);
7349 if (usable_startpfn < 0x100000) {
7350 mem_below_4gb_not_mirrored = true;
7354 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7355 min(usable_startpfn, zone_movable_pfn[nid]) :
7359 if (mem_below_4gb_not_mirrored)
7360 pr_warn("This configuration results in unmirrored kernel memory.\n");
7366 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7367 * amount of necessary memory.
7369 if (required_kernelcore_percent)
7370 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7372 if (required_movablecore_percent)
7373 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7377 * If movablecore= was specified, calculate what size of
7378 * kernelcore that corresponds so that memory usable for
7379 * any allocation type is evenly spread. If both kernelcore
7380 * and movablecore are specified, then the value of kernelcore
7381 * will be used for required_kernelcore if it's greater than
7382 * what movablecore would have allowed.
7384 if (required_movablecore) {
7385 unsigned long corepages;
7388 * Round-up so that ZONE_MOVABLE is at least as large as what
7389 * was requested by the user
7391 required_movablecore =
7392 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7393 required_movablecore = min(totalpages, required_movablecore);
7394 corepages = totalpages - required_movablecore;
7396 required_kernelcore = max(required_kernelcore, corepages);
7400 * If kernelcore was not specified or kernelcore size is larger
7401 * than totalpages, there is no ZONE_MOVABLE.
7403 if (!required_kernelcore || required_kernelcore >= totalpages)
7406 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7407 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7410 /* Spread kernelcore memory as evenly as possible throughout nodes */
7411 kernelcore_node = required_kernelcore / usable_nodes;
7412 for_each_node_state(nid, N_MEMORY) {
7413 unsigned long start_pfn, end_pfn;
7416 * Recalculate kernelcore_node if the division per node
7417 * now exceeds what is necessary to satisfy the requested
7418 * amount of memory for the kernel
7420 if (required_kernelcore < kernelcore_node)
7421 kernelcore_node = required_kernelcore / usable_nodes;
7424 * As the map is walked, we track how much memory is usable
7425 * by the kernel using kernelcore_remaining. When it is
7426 * 0, the rest of the node is usable by ZONE_MOVABLE
7428 kernelcore_remaining = kernelcore_node;
7430 /* Go through each range of PFNs within this node */
7431 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7432 unsigned long size_pages;
7434 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7435 if (start_pfn >= end_pfn)
7438 /* Account for what is only usable for kernelcore */
7439 if (start_pfn < usable_startpfn) {
7440 unsigned long kernel_pages;
7441 kernel_pages = min(end_pfn, usable_startpfn)
7444 kernelcore_remaining -= min(kernel_pages,
7445 kernelcore_remaining);
7446 required_kernelcore -= min(kernel_pages,
7447 required_kernelcore);
7449 /* Continue if range is now fully accounted */
7450 if (end_pfn <= usable_startpfn) {
7453 * Push zone_movable_pfn to the end so
7454 * that if we have to rebalance
7455 * kernelcore across nodes, we will
7456 * not double account here
7458 zone_movable_pfn[nid] = end_pfn;
7461 start_pfn = usable_startpfn;
7465 * The usable PFN range for ZONE_MOVABLE is from
7466 * start_pfn->end_pfn. Calculate size_pages as the
7467 * number of pages used as kernelcore
7469 size_pages = end_pfn - start_pfn;
7470 if (size_pages > kernelcore_remaining)
7471 size_pages = kernelcore_remaining;
7472 zone_movable_pfn[nid] = start_pfn + size_pages;
7475 * Some kernelcore has been met, update counts and
7476 * break if the kernelcore for this node has been
7479 required_kernelcore -= min(required_kernelcore,
7481 kernelcore_remaining -= size_pages;
7482 if (!kernelcore_remaining)
7488 * If there is still required_kernelcore, we do another pass with one
7489 * less node in the count. This will push zone_movable_pfn[nid] further
7490 * along on the nodes that still have memory until kernelcore is
7494 if (usable_nodes && required_kernelcore > usable_nodes)
7498 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7499 for (nid = 0; nid < MAX_NUMNODES; nid++)
7500 zone_movable_pfn[nid] =
7501 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7504 /* restore the node_state */
7505 node_states[N_MEMORY] = saved_node_state;
7508 /* Any regular or high memory on that node ? */
7509 static void check_for_memory(pg_data_t *pgdat, int nid)
7511 enum zone_type zone_type;
7513 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7514 struct zone *zone = &pgdat->node_zones[zone_type];
7515 if (populated_zone(zone)) {
7516 if (IS_ENABLED(CONFIG_HIGHMEM))
7517 node_set_state(nid, N_HIGH_MEMORY);
7518 if (zone_type <= ZONE_NORMAL)
7519 node_set_state(nid, N_NORMAL_MEMORY);
7526 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7527 * such cases we allow max_zone_pfn sorted in the descending order
7529 bool __weak arch_has_descending_max_zone_pfns(void)
7535 * free_area_init - Initialise all pg_data_t and zone data
7536 * @max_zone_pfn: an array of max PFNs for each zone
7538 * This will call free_area_init_node() for each active node in the system.
7539 * Using the page ranges provided by memblock_set_node(), the size of each
7540 * zone in each node and their holes is calculated. If the maximum PFN
7541 * between two adjacent zones match, it is assumed that the zone is empty.
7542 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7543 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7544 * starts where the previous one ended. For example, ZONE_DMA32 starts
7545 * at arch_max_dma_pfn.
7547 void __init free_area_init(unsigned long *max_zone_pfn)
7549 unsigned long start_pfn, end_pfn;
7553 /* Record where the zone boundaries are */
7554 memset(arch_zone_lowest_possible_pfn, 0,
7555 sizeof(arch_zone_lowest_possible_pfn));
7556 memset(arch_zone_highest_possible_pfn, 0,
7557 sizeof(arch_zone_highest_possible_pfn));
7559 start_pfn = find_min_pfn_with_active_regions();
7560 descending = arch_has_descending_max_zone_pfns();
7562 for (i = 0; i < MAX_NR_ZONES; i++) {
7564 zone = MAX_NR_ZONES - i - 1;
7568 if (zone == ZONE_MOVABLE)
7571 end_pfn = max(max_zone_pfn[zone], start_pfn);
7572 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7573 arch_zone_highest_possible_pfn[zone] = end_pfn;
7575 start_pfn = end_pfn;
7578 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7579 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7580 find_zone_movable_pfns_for_nodes();
7582 /* Print out the zone ranges */
7583 pr_info("Zone ranges:\n");
7584 for (i = 0; i < MAX_NR_ZONES; i++) {
7585 if (i == ZONE_MOVABLE)
7587 pr_info(" %-8s ", zone_names[i]);
7588 if (arch_zone_lowest_possible_pfn[i] ==
7589 arch_zone_highest_possible_pfn[i])
7592 pr_cont("[mem %#018Lx-%#018Lx]\n",
7593 (u64)arch_zone_lowest_possible_pfn[i]
7595 ((u64)arch_zone_highest_possible_pfn[i]
7596 << PAGE_SHIFT) - 1);
7599 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7600 pr_info("Movable zone start for each node\n");
7601 for (i = 0; i < MAX_NUMNODES; i++) {
7602 if (zone_movable_pfn[i])
7603 pr_info(" Node %d: %#018Lx\n", i,
7604 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7608 * Print out the early node map, and initialize the
7609 * subsection-map relative to active online memory ranges to
7610 * enable future "sub-section" extensions of the memory map.
7612 pr_info("Early memory node ranges\n");
7613 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7614 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7615 (u64)start_pfn << PAGE_SHIFT,
7616 ((u64)end_pfn << PAGE_SHIFT) - 1);
7617 subsection_map_init(start_pfn, end_pfn - start_pfn);
7620 /* Initialise every node */
7621 mminit_verify_pageflags_layout();
7622 setup_nr_node_ids();
7623 for_each_online_node(nid) {
7624 pg_data_t *pgdat = NODE_DATA(nid);
7625 free_area_init_node(nid);
7627 /* Any memory on that node */
7628 if (pgdat->node_present_pages)
7629 node_set_state(nid, N_MEMORY);
7630 check_for_memory(pgdat, nid);
7634 static int __init cmdline_parse_core(char *p, unsigned long *core,
7635 unsigned long *percent)
7637 unsigned long long coremem;
7643 /* Value may be a percentage of total memory, otherwise bytes */
7644 coremem = simple_strtoull(p, &endptr, 0);
7645 if (*endptr == '%') {
7646 /* Paranoid check for percent values greater than 100 */
7647 WARN_ON(coremem > 100);
7651 coremem = memparse(p, &p);
7652 /* Paranoid check that UL is enough for the coremem value */
7653 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7655 *core = coremem >> PAGE_SHIFT;
7662 * kernelcore=size sets the amount of memory for use for allocations that
7663 * cannot be reclaimed or migrated.
7665 static int __init cmdline_parse_kernelcore(char *p)
7667 /* parse kernelcore=mirror */
7668 if (parse_option_str(p, "mirror")) {
7669 mirrored_kernelcore = true;
7673 return cmdline_parse_core(p, &required_kernelcore,
7674 &required_kernelcore_percent);
7678 * movablecore=size sets the amount of memory for use for allocations that
7679 * can be reclaimed or migrated.
7681 static int __init cmdline_parse_movablecore(char *p)
7683 return cmdline_parse_core(p, &required_movablecore,
7684 &required_movablecore_percent);
7687 early_param("kernelcore", cmdline_parse_kernelcore);
7688 early_param("movablecore", cmdline_parse_movablecore);
7690 void adjust_managed_page_count(struct page *page, long count)
7692 atomic_long_add(count, &page_zone(page)->managed_pages);
7693 totalram_pages_add(count);
7694 #ifdef CONFIG_HIGHMEM
7695 if (PageHighMem(page))
7696 totalhigh_pages_add(count);
7699 EXPORT_SYMBOL(adjust_managed_page_count);
7701 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7704 unsigned long pages = 0;
7706 start = (void *)PAGE_ALIGN((unsigned long)start);
7707 end = (void *)((unsigned long)end & PAGE_MASK);
7708 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7709 struct page *page = virt_to_page(pos);
7710 void *direct_map_addr;
7713 * 'direct_map_addr' might be different from 'pos'
7714 * because some architectures' virt_to_page()
7715 * work with aliases. Getting the direct map
7716 * address ensures that we get a _writeable_
7717 * alias for the memset().
7719 direct_map_addr = page_address(page);
7721 * Perform a kasan-unchecked memset() since this memory
7722 * has not been initialized.
7724 direct_map_addr = kasan_reset_tag(direct_map_addr);
7725 if ((unsigned int)poison <= 0xFF)
7726 memset(direct_map_addr, poison, PAGE_SIZE);
7728 free_reserved_page(page);
7732 pr_info("Freeing %s memory: %ldK\n",
7733 s, pages << (PAGE_SHIFT - 10));
7738 void __init mem_init_print_info(void)
7740 unsigned long physpages, codesize, datasize, rosize, bss_size;
7741 unsigned long init_code_size, init_data_size;
7743 physpages = get_num_physpages();
7744 codesize = _etext - _stext;
7745 datasize = _edata - _sdata;
7746 rosize = __end_rodata - __start_rodata;
7747 bss_size = __bss_stop - __bss_start;
7748 init_data_size = __init_end - __init_begin;
7749 init_code_size = _einittext - _sinittext;
7752 * Detect special cases and adjust section sizes accordingly:
7753 * 1) .init.* may be embedded into .data sections
7754 * 2) .init.text.* may be out of [__init_begin, __init_end],
7755 * please refer to arch/tile/kernel/vmlinux.lds.S.
7756 * 3) .rodata.* may be embedded into .text or .data sections.
7758 #define adj_init_size(start, end, size, pos, adj) \
7760 if (start <= pos && pos < end && size > adj) \
7764 adj_init_size(__init_begin, __init_end, init_data_size,
7765 _sinittext, init_code_size);
7766 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7767 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7768 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7769 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7771 #undef adj_init_size
7773 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7774 #ifdef CONFIG_HIGHMEM
7778 nr_free_pages() << (PAGE_SHIFT - 10),
7779 physpages << (PAGE_SHIFT - 10),
7780 codesize >> 10, datasize >> 10, rosize >> 10,
7781 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7782 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7783 totalcma_pages << (PAGE_SHIFT - 10)
7784 #ifdef CONFIG_HIGHMEM
7785 , totalhigh_pages() << (PAGE_SHIFT - 10)
7791 * set_dma_reserve - set the specified number of pages reserved in the first zone
7792 * @new_dma_reserve: The number of pages to mark reserved
7794 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7795 * In the DMA zone, a significant percentage may be consumed by kernel image
7796 * and other unfreeable allocations which can skew the watermarks badly. This
7797 * function may optionally be used to account for unfreeable pages in the
7798 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7799 * smaller per-cpu batchsize.
7801 void __init set_dma_reserve(unsigned long new_dma_reserve)
7803 dma_reserve = new_dma_reserve;
7806 static int page_alloc_cpu_dead(unsigned int cpu)
7809 lru_add_drain_cpu(cpu);
7813 * Spill the event counters of the dead processor
7814 * into the current processors event counters.
7815 * This artificially elevates the count of the current
7818 vm_events_fold_cpu(cpu);
7821 * Zero the differential counters of the dead processor
7822 * so that the vm statistics are consistent.
7824 * This is only okay since the processor is dead and cannot
7825 * race with what we are doing.
7827 cpu_vm_stats_fold(cpu);
7832 int hashdist = HASHDIST_DEFAULT;
7834 static int __init set_hashdist(char *str)
7838 hashdist = simple_strtoul(str, &str, 0);
7841 __setup("hashdist=", set_hashdist);
7844 void __init page_alloc_init(void)
7849 if (num_node_state(N_MEMORY) == 1)
7853 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7854 "mm/page_alloc:dead", NULL,
7855 page_alloc_cpu_dead);
7860 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7861 * or min_free_kbytes changes.
7863 static void calculate_totalreserve_pages(void)
7865 struct pglist_data *pgdat;
7866 unsigned long reserve_pages = 0;
7867 enum zone_type i, j;
7869 for_each_online_pgdat(pgdat) {
7871 pgdat->totalreserve_pages = 0;
7873 for (i = 0; i < MAX_NR_ZONES; i++) {
7874 struct zone *zone = pgdat->node_zones + i;
7876 unsigned long managed_pages = zone_managed_pages(zone);
7878 /* Find valid and maximum lowmem_reserve in the zone */
7879 for (j = i; j < MAX_NR_ZONES; j++) {
7880 if (zone->lowmem_reserve[j] > max)
7881 max = zone->lowmem_reserve[j];
7884 /* we treat the high watermark as reserved pages. */
7885 max += high_wmark_pages(zone);
7887 if (max > managed_pages)
7888 max = managed_pages;
7890 pgdat->totalreserve_pages += max;
7892 reserve_pages += max;
7895 totalreserve_pages = reserve_pages;
7899 * setup_per_zone_lowmem_reserve - called whenever
7900 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7901 * has a correct pages reserved value, so an adequate number of
7902 * pages are left in the zone after a successful __alloc_pages().
7904 static void setup_per_zone_lowmem_reserve(void)
7906 struct pglist_data *pgdat;
7907 enum zone_type i, j;
7909 for_each_online_pgdat(pgdat) {
7910 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
7911 struct zone *zone = &pgdat->node_zones[i];
7912 int ratio = sysctl_lowmem_reserve_ratio[i];
7913 bool clear = !ratio || !zone_managed_pages(zone);
7914 unsigned long managed_pages = 0;
7916 for (j = i + 1; j < MAX_NR_ZONES; j++) {
7918 zone->lowmem_reserve[j] = 0;
7920 struct zone *upper_zone = &pgdat->node_zones[j];
7922 managed_pages += zone_managed_pages(upper_zone);
7923 zone->lowmem_reserve[j] = managed_pages / ratio;
7929 /* update totalreserve_pages */
7930 calculate_totalreserve_pages();
7933 static void __setup_per_zone_wmarks(void)
7935 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7936 unsigned long lowmem_pages = 0;
7938 unsigned long flags;
7940 /* Calculate total number of !ZONE_HIGHMEM pages */
7941 for_each_zone(zone) {
7942 if (!is_highmem(zone))
7943 lowmem_pages += zone_managed_pages(zone);
7946 for_each_zone(zone) {
7949 spin_lock_irqsave(&zone->lock, flags);
7950 tmp = (u64)pages_min * zone_managed_pages(zone);
7951 do_div(tmp, lowmem_pages);
7952 if (is_highmem(zone)) {
7954 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7955 * need highmem pages, so cap pages_min to a small
7958 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7959 * deltas control async page reclaim, and so should
7960 * not be capped for highmem.
7962 unsigned long min_pages;
7964 min_pages = zone_managed_pages(zone) / 1024;
7965 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7966 zone->_watermark[WMARK_MIN] = min_pages;
7969 * If it's a lowmem zone, reserve a number of pages
7970 * proportionate to the zone's size.
7972 zone->_watermark[WMARK_MIN] = tmp;
7976 * Set the kswapd watermarks distance according to the
7977 * scale factor in proportion to available memory, but
7978 * ensure a minimum size on small systems.
7980 tmp = max_t(u64, tmp >> 2,
7981 mult_frac(zone_managed_pages(zone),
7982 watermark_scale_factor, 10000));
7984 zone->watermark_boost = 0;
7985 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7986 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7988 spin_unlock_irqrestore(&zone->lock, flags);
7991 /* update totalreserve_pages */
7992 calculate_totalreserve_pages();
7996 * setup_per_zone_wmarks - called when min_free_kbytes changes
7997 * or when memory is hot-{added|removed}
7999 * Ensures that the watermark[min,low,high] values for each zone are set
8000 * correctly with respect to min_free_kbytes.
8002 void setup_per_zone_wmarks(void)
8004 static DEFINE_SPINLOCK(lock);
8007 __setup_per_zone_wmarks();
8012 * Initialise min_free_kbytes.
8014 * For small machines we want it small (128k min). For large machines
8015 * we want it large (256MB max). But it is not linear, because network
8016 * bandwidth does not increase linearly with machine size. We use
8018 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8019 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8035 int __meminit init_per_zone_wmark_min(void)
8037 unsigned long lowmem_kbytes;
8038 int new_min_free_kbytes;
8040 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8041 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8043 if (new_min_free_kbytes > user_min_free_kbytes) {
8044 min_free_kbytes = new_min_free_kbytes;
8045 if (min_free_kbytes < 128)
8046 min_free_kbytes = 128;
8047 if (min_free_kbytes > 262144)
8048 min_free_kbytes = 262144;
8050 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8051 new_min_free_kbytes, user_min_free_kbytes);
8053 setup_per_zone_wmarks();
8054 refresh_zone_stat_thresholds();
8055 setup_per_zone_lowmem_reserve();
8058 setup_min_unmapped_ratio();
8059 setup_min_slab_ratio();
8062 khugepaged_min_free_kbytes_update();
8066 postcore_initcall(init_per_zone_wmark_min)
8069 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8070 * that we can call two helper functions whenever min_free_kbytes
8073 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8074 void *buffer, size_t *length, loff_t *ppos)
8078 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8083 user_min_free_kbytes = min_free_kbytes;
8084 setup_per_zone_wmarks();
8089 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8090 void *buffer, size_t *length, loff_t *ppos)
8094 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8099 setup_per_zone_wmarks();
8105 static void setup_min_unmapped_ratio(void)
8110 for_each_online_pgdat(pgdat)
8111 pgdat->min_unmapped_pages = 0;
8114 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8115 sysctl_min_unmapped_ratio) / 100;
8119 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8120 void *buffer, size_t *length, loff_t *ppos)
8124 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8128 setup_min_unmapped_ratio();
8133 static void setup_min_slab_ratio(void)
8138 for_each_online_pgdat(pgdat)
8139 pgdat->min_slab_pages = 0;
8142 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8143 sysctl_min_slab_ratio) / 100;
8146 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8147 void *buffer, size_t *length, loff_t *ppos)
8151 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8155 setup_min_slab_ratio();
8162 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8163 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8164 * whenever sysctl_lowmem_reserve_ratio changes.
8166 * The reserve ratio obviously has absolutely no relation with the
8167 * minimum watermarks. The lowmem reserve ratio can only make sense
8168 * if in function of the boot time zone sizes.
8170 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8171 void *buffer, size_t *length, loff_t *ppos)
8175 proc_dointvec_minmax(table, write, buffer, length, ppos);
8177 for (i = 0; i < MAX_NR_ZONES; i++) {
8178 if (sysctl_lowmem_reserve_ratio[i] < 1)
8179 sysctl_lowmem_reserve_ratio[i] = 0;
8182 setup_per_zone_lowmem_reserve();
8187 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8188 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8189 * pagelist can have before it gets flushed back to buddy allocator.
8191 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8192 void *buffer, size_t *length, loff_t *ppos)
8195 int old_percpu_pagelist_fraction;
8198 mutex_lock(&pcp_batch_high_lock);
8199 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8201 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8202 if (!write || ret < 0)
8205 /* Sanity checking to avoid pcp imbalance */
8206 if (percpu_pagelist_fraction &&
8207 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8208 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8214 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8217 for_each_populated_zone(zone)
8218 zone_set_pageset_high_and_batch(zone);
8220 mutex_unlock(&pcp_batch_high_lock);
8224 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8226 * Returns the number of pages that arch has reserved but
8227 * is not known to alloc_large_system_hash().
8229 static unsigned long __init arch_reserved_kernel_pages(void)
8236 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8237 * machines. As memory size is increased the scale is also increased but at
8238 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8239 * quadruples the scale is increased by one, which means the size of hash table
8240 * only doubles, instead of quadrupling as well.
8241 * Because 32-bit systems cannot have large physical memory, where this scaling
8242 * makes sense, it is disabled on such platforms.
8244 #if __BITS_PER_LONG > 32
8245 #define ADAPT_SCALE_BASE (64ul << 30)
8246 #define ADAPT_SCALE_SHIFT 2
8247 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8251 * allocate a large system hash table from bootmem
8252 * - it is assumed that the hash table must contain an exact power-of-2
8253 * quantity of entries
8254 * - limit is the number of hash buckets, not the total allocation size
8256 void *__init alloc_large_system_hash(const char *tablename,
8257 unsigned long bucketsize,
8258 unsigned long numentries,
8261 unsigned int *_hash_shift,
8262 unsigned int *_hash_mask,
8263 unsigned long low_limit,
8264 unsigned long high_limit)
8266 unsigned long long max = high_limit;
8267 unsigned long log2qty, size;
8273 /* allow the kernel cmdline to have a say */
8275 /* round applicable memory size up to nearest megabyte */
8276 numentries = nr_kernel_pages;
8277 numentries -= arch_reserved_kernel_pages();
8279 /* It isn't necessary when PAGE_SIZE >= 1MB */
8280 if (PAGE_SHIFT < 20)
8281 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8283 #if __BITS_PER_LONG > 32
8285 unsigned long adapt;
8287 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8288 adapt <<= ADAPT_SCALE_SHIFT)
8293 /* limit to 1 bucket per 2^scale bytes of low memory */
8294 if (scale > PAGE_SHIFT)
8295 numentries >>= (scale - PAGE_SHIFT);
8297 numentries <<= (PAGE_SHIFT - scale);
8299 /* Make sure we've got at least a 0-order allocation.. */
8300 if (unlikely(flags & HASH_SMALL)) {
8301 /* Makes no sense without HASH_EARLY */
8302 WARN_ON(!(flags & HASH_EARLY));
8303 if (!(numentries >> *_hash_shift)) {
8304 numentries = 1UL << *_hash_shift;
8305 BUG_ON(!numentries);
8307 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8308 numentries = PAGE_SIZE / bucketsize;
8310 numentries = roundup_pow_of_two(numentries);
8312 /* limit allocation size to 1/16 total memory by default */
8314 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8315 do_div(max, bucketsize);
8317 max = min(max, 0x80000000ULL);
8319 if (numentries < low_limit)
8320 numentries = low_limit;
8321 if (numentries > max)
8324 log2qty = ilog2(numentries);
8326 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8329 size = bucketsize << log2qty;
8330 if (flags & HASH_EARLY) {
8331 if (flags & HASH_ZERO)
8332 table = memblock_alloc(size, SMP_CACHE_BYTES);
8334 table = memblock_alloc_raw(size,
8336 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8337 table = __vmalloc(size, gfp_flags);
8339 huge = is_vm_area_hugepages(table);
8342 * If bucketsize is not a power-of-two, we may free
8343 * some pages at the end of hash table which
8344 * alloc_pages_exact() automatically does
8346 table = alloc_pages_exact(size, gfp_flags);
8347 kmemleak_alloc(table, size, 1, gfp_flags);
8349 } while (!table && size > PAGE_SIZE && --log2qty);
8352 panic("Failed to allocate %s hash table\n", tablename);
8354 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8355 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8356 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8359 *_hash_shift = log2qty;
8361 *_hash_mask = (1 << log2qty) - 1;
8367 * This function checks whether pageblock includes unmovable pages or not.
8369 * PageLRU check without isolation or lru_lock could race so that
8370 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8371 * check without lock_page also may miss some movable non-lru pages at
8372 * race condition. So you can't expect this function should be exact.
8374 * Returns a page without holding a reference. If the caller wants to
8375 * dereference that page (e.g., dumping), it has to make sure that it
8376 * cannot get removed (e.g., via memory unplug) concurrently.
8379 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8380 int migratetype, int flags)
8382 unsigned long iter = 0;
8383 unsigned long pfn = page_to_pfn(page);
8384 unsigned long offset = pfn % pageblock_nr_pages;
8386 if (is_migrate_cma_page(page)) {
8388 * CMA allocations (alloc_contig_range) really need to mark
8389 * isolate CMA pageblocks even when they are not movable in fact
8390 * so consider them movable here.
8392 if (is_migrate_cma(migratetype))
8398 for (; iter < pageblock_nr_pages - offset; iter++) {
8399 if (!pfn_valid_within(pfn + iter))
8402 page = pfn_to_page(pfn + iter);
8405 * Both, bootmem allocations and memory holes are marked
8406 * PG_reserved and are unmovable. We can even have unmovable
8407 * allocations inside ZONE_MOVABLE, for example when
8408 * specifying "movablecore".
8410 if (PageReserved(page))
8414 * If the zone is movable and we have ruled out all reserved
8415 * pages then it should be reasonably safe to assume the rest
8418 if (zone_idx(zone) == ZONE_MOVABLE)
8422 * Hugepages are not in LRU lists, but they're movable.
8423 * THPs are on the LRU, but need to be counted as #small pages.
8424 * We need not scan over tail pages because we don't
8425 * handle each tail page individually in migration.
8427 if (PageHuge(page) || PageTransCompound(page)) {
8428 struct page *head = compound_head(page);
8429 unsigned int skip_pages;
8431 if (PageHuge(page)) {
8432 if (!hugepage_migration_supported(page_hstate(head)))
8434 } else if (!PageLRU(head) && !__PageMovable(head)) {
8438 skip_pages = compound_nr(head) - (page - head);
8439 iter += skip_pages - 1;
8444 * We can't use page_count without pin a page
8445 * because another CPU can free compound page.
8446 * This check already skips compound tails of THP
8447 * because their page->_refcount is zero at all time.
8449 if (!page_ref_count(page)) {
8450 if (PageBuddy(page))
8451 iter += (1 << buddy_order(page)) - 1;
8456 * The HWPoisoned page may be not in buddy system, and
8457 * page_count() is not 0.
8459 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8463 * We treat all PageOffline() pages as movable when offlining
8464 * to give drivers a chance to decrement their reference count
8465 * in MEM_GOING_OFFLINE in order to indicate that these pages
8466 * can be offlined as there are no direct references anymore.
8467 * For actually unmovable PageOffline() where the driver does
8468 * not support this, we will fail later when trying to actually
8469 * move these pages that still have a reference count > 0.
8470 * (false negatives in this function only)
8472 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8475 if (__PageMovable(page) || PageLRU(page))
8479 * If there are RECLAIMABLE pages, we need to check
8480 * it. But now, memory offline itself doesn't call
8481 * shrink_node_slabs() and it still to be fixed.
8488 #ifdef CONFIG_CONTIG_ALLOC
8489 static unsigned long pfn_max_align_down(unsigned long pfn)
8491 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8492 pageblock_nr_pages) - 1);
8495 static unsigned long pfn_max_align_up(unsigned long pfn)
8497 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8498 pageblock_nr_pages));
8501 /* [start, end) must belong to a single zone. */
8502 static int __alloc_contig_migrate_range(struct compact_control *cc,
8503 unsigned long start, unsigned long end)
8505 /* This function is based on compact_zone() from compaction.c. */
8506 unsigned int nr_reclaimed;
8507 unsigned long pfn = start;
8508 unsigned int tries = 0;
8510 struct migration_target_control mtc = {
8511 .nid = zone_to_nid(cc->zone),
8512 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8517 while (pfn < end || !list_empty(&cc->migratepages)) {
8518 if (fatal_signal_pending(current)) {
8523 if (list_empty(&cc->migratepages)) {
8524 cc->nr_migratepages = 0;
8525 pfn = isolate_migratepages_range(cc, pfn, end);
8531 } else if (++tries == 5) {
8532 ret = ret < 0 ? ret : -EBUSY;
8536 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8538 cc->nr_migratepages -= nr_reclaimed;
8540 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8541 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8544 putback_movable_pages(&cc->migratepages);
8551 * alloc_contig_range() -- tries to allocate given range of pages
8552 * @start: start PFN to allocate
8553 * @end: one-past-the-last PFN to allocate
8554 * @migratetype: migratetype of the underlaying pageblocks (either
8555 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8556 * in range must have the same migratetype and it must
8557 * be either of the two.
8558 * @gfp_mask: GFP mask to use during compaction
8560 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8561 * aligned. The PFN range must belong to a single zone.
8563 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8564 * pageblocks in the range. Once isolated, the pageblocks should not
8565 * be modified by others.
8567 * Return: zero on success or negative error code. On success all
8568 * pages which PFN is in [start, end) are allocated for the caller and
8569 * need to be freed with free_contig_range().
8571 int alloc_contig_range(unsigned long start, unsigned long end,
8572 unsigned migratetype, gfp_t gfp_mask)
8574 unsigned long outer_start, outer_end;
8578 struct compact_control cc = {
8579 .nr_migratepages = 0,
8581 .zone = page_zone(pfn_to_page(start)),
8582 .mode = MIGRATE_SYNC,
8583 .ignore_skip_hint = true,
8584 .no_set_skip_hint = true,
8585 .gfp_mask = current_gfp_context(gfp_mask),
8586 .alloc_contig = true,
8588 INIT_LIST_HEAD(&cc.migratepages);
8591 * What we do here is we mark all pageblocks in range as
8592 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8593 * have different sizes, and due to the way page allocator
8594 * work, we align the range to biggest of the two pages so
8595 * that page allocator won't try to merge buddies from
8596 * different pageblocks and change MIGRATE_ISOLATE to some
8597 * other migration type.
8599 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8600 * migrate the pages from an unaligned range (ie. pages that
8601 * we are interested in). This will put all the pages in
8602 * range back to page allocator as MIGRATE_ISOLATE.
8604 * When this is done, we take the pages in range from page
8605 * allocator removing them from the buddy system. This way
8606 * page allocator will never consider using them.
8608 * This lets us mark the pageblocks back as
8609 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8610 * aligned range but not in the unaligned, original range are
8611 * put back to page allocator so that buddy can use them.
8614 ret = start_isolate_page_range(pfn_max_align_down(start),
8615 pfn_max_align_up(end), migratetype, 0);
8619 drain_all_pages(cc.zone);
8622 * In case of -EBUSY, we'd like to know which page causes problem.
8623 * So, just fall through. test_pages_isolated() has a tracepoint
8624 * which will report the busy page.
8626 * It is possible that busy pages could become available before
8627 * the call to test_pages_isolated, and the range will actually be
8628 * allocated. So, if we fall through be sure to clear ret so that
8629 * -EBUSY is not accidentally used or returned to caller.
8631 ret = __alloc_contig_migrate_range(&cc, start, end);
8632 if (ret && ret != -EBUSY)
8637 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8638 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8639 * more, all pages in [start, end) are free in page allocator.
8640 * What we are going to do is to allocate all pages from
8641 * [start, end) (that is remove them from page allocator).
8643 * The only problem is that pages at the beginning and at the
8644 * end of interesting range may be not aligned with pages that
8645 * page allocator holds, ie. they can be part of higher order
8646 * pages. Because of this, we reserve the bigger range and
8647 * once this is done free the pages we are not interested in.
8649 * We don't have to hold zone->lock here because the pages are
8650 * isolated thus they won't get removed from buddy.
8654 outer_start = start;
8655 while (!PageBuddy(pfn_to_page(outer_start))) {
8656 if (++order >= MAX_ORDER) {
8657 outer_start = start;
8660 outer_start &= ~0UL << order;
8663 if (outer_start != start) {
8664 order = buddy_order(pfn_to_page(outer_start));
8667 * outer_start page could be small order buddy page and
8668 * it doesn't include start page. Adjust outer_start
8669 * in this case to report failed page properly
8670 * on tracepoint in test_pages_isolated()
8672 if (outer_start + (1UL << order) <= start)
8673 outer_start = start;
8676 /* Make sure the range is really isolated. */
8677 if (test_pages_isolated(outer_start, end, 0)) {
8682 /* Grab isolated pages from freelists. */
8683 outer_end = isolate_freepages_range(&cc, outer_start, end);
8689 /* Free head and tail (if any) */
8690 if (start != outer_start)
8691 free_contig_range(outer_start, start - outer_start);
8692 if (end != outer_end)
8693 free_contig_range(end, outer_end - end);
8696 undo_isolate_page_range(pfn_max_align_down(start),
8697 pfn_max_align_up(end), migratetype);
8700 EXPORT_SYMBOL(alloc_contig_range);
8702 static int __alloc_contig_pages(unsigned long start_pfn,
8703 unsigned long nr_pages, gfp_t gfp_mask)
8705 unsigned long end_pfn = start_pfn + nr_pages;
8707 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8711 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8712 unsigned long nr_pages)
8714 unsigned long i, end_pfn = start_pfn + nr_pages;
8717 for (i = start_pfn; i < end_pfn; i++) {
8718 page = pfn_to_online_page(i);
8722 if (page_zone(page) != z)
8725 if (PageReserved(page))
8728 if (page_count(page) > 0)
8737 static bool zone_spans_last_pfn(const struct zone *zone,
8738 unsigned long start_pfn, unsigned long nr_pages)
8740 unsigned long last_pfn = start_pfn + nr_pages - 1;
8742 return zone_spans_pfn(zone, last_pfn);
8746 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8747 * @nr_pages: Number of contiguous pages to allocate
8748 * @gfp_mask: GFP mask to limit search and used during compaction
8750 * @nodemask: Mask for other possible nodes
8752 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8753 * on an applicable zonelist to find a contiguous pfn range which can then be
8754 * tried for allocation with alloc_contig_range(). This routine is intended
8755 * for allocation requests which can not be fulfilled with the buddy allocator.
8757 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8758 * power of two then the alignment is guaranteed to be to the given nr_pages
8759 * (e.g. 1GB request would be aligned to 1GB).
8761 * Allocated pages can be freed with free_contig_range() or by manually calling
8762 * __free_page() on each allocated page.
8764 * Return: pointer to contiguous pages on success, or NULL if not successful.
8766 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8767 int nid, nodemask_t *nodemask)
8769 unsigned long ret, pfn, flags;
8770 struct zonelist *zonelist;
8774 zonelist = node_zonelist(nid, gfp_mask);
8775 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8776 gfp_zone(gfp_mask), nodemask) {
8777 spin_lock_irqsave(&zone->lock, flags);
8779 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8780 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8781 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8783 * We release the zone lock here because
8784 * alloc_contig_range() will also lock the zone
8785 * at some point. If there's an allocation
8786 * spinning on this lock, it may win the race
8787 * and cause alloc_contig_range() to fail...
8789 spin_unlock_irqrestore(&zone->lock, flags);
8790 ret = __alloc_contig_pages(pfn, nr_pages,
8793 return pfn_to_page(pfn);
8794 spin_lock_irqsave(&zone->lock, flags);
8798 spin_unlock_irqrestore(&zone->lock, flags);
8802 #endif /* CONFIG_CONTIG_ALLOC */
8804 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8806 unsigned int count = 0;
8808 for (; nr_pages--; pfn++) {
8809 struct page *page = pfn_to_page(pfn);
8811 count += page_count(page) != 1;
8814 WARN(count != 0, "%d pages are still in use!\n", count);
8816 EXPORT_SYMBOL(free_contig_range);
8819 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8820 * page high values need to be recalulated.
8822 void __meminit zone_pcp_update(struct zone *zone)
8824 mutex_lock(&pcp_batch_high_lock);
8825 zone_set_pageset_high_and_batch(zone);
8826 mutex_unlock(&pcp_batch_high_lock);
8830 * Effectively disable pcplists for the zone by setting the high limit to 0
8831 * and draining all cpus. A concurrent page freeing on another CPU that's about
8832 * to put the page on pcplist will either finish before the drain and the page
8833 * will be drained, or observe the new high limit and skip the pcplist.
8835 * Must be paired with a call to zone_pcp_enable().
8837 void zone_pcp_disable(struct zone *zone)
8839 mutex_lock(&pcp_batch_high_lock);
8840 __zone_set_pageset_high_and_batch(zone, 0, 1);
8841 __drain_all_pages(zone, true);
8844 void zone_pcp_enable(struct zone *zone)
8846 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
8847 mutex_unlock(&pcp_batch_high_lock);
8850 void zone_pcp_reset(struct zone *zone)
8852 unsigned long flags;
8854 struct per_cpu_pageset *pset;
8856 /* avoid races with drain_pages() */
8857 local_irq_save(flags);
8858 if (zone->pageset != &boot_pageset) {
8859 for_each_online_cpu(cpu) {
8860 pset = per_cpu_ptr(zone->pageset, cpu);
8861 drain_zonestat(zone, pset);
8863 free_percpu(zone->pageset);
8864 zone->pageset = &boot_pageset;
8866 local_irq_restore(flags);
8869 #ifdef CONFIG_MEMORY_HOTREMOVE
8871 * All pages in the range must be in a single zone, must not contain holes,
8872 * must span full sections, and must be isolated before calling this function.
8874 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8876 unsigned long pfn = start_pfn;
8880 unsigned long flags;
8882 offline_mem_sections(pfn, end_pfn);
8883 zone = page_zone(pfn_to_page(pfn));
8884 spin_lock_irqsave(&zone->lock, flags);
8885 while (pfn < end_pfn) {
8886 page = pfn_to_page(pfn);
8888 * The HWPoisoned page may be not in buddy system, and
8889 * page_count() is not 0.
8891 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8896 * At this point all remaining PageOffline() pages have a
8897 * reference count of 0 and can simply be skipped.
8899 if (PageOffline(page)) {
8900 BUG_ON(page_count(page));
8901 BUG_ON(PageBuddy(page));
8906 BUG_ON(page_count(page));
8907 BUG_ON(!PageBuddy(page));
8908 order = buddy_order(page);
8909 del_page_from_free_list(page, zone, order);
8910 pfn += (1 << order);
8912 spin_unlock_irqrestore(&zone->lock, flags);
8916 bool is_free_buddy_page(struct page *page)
8918 struct zone *zone = page_zone(page);
8919 unsigned long pfn = page_to_pfn(page);
8920 unsigned long flags;
8923 spin_lock_irqsave(&zone->lock, flags);
8924 for (order = 0; order < MAX_ORDER; order++) {
8925 struct page *page_head = page - (pfn & ((1 << order) - 1));
8927 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
8930 spin_unlock_irqrestore(&zone->lock, flags);
8932 return order < MAX_ORDER;
8935 #ifdef CONFIG_MEMORY_FAILURE
8937 * Break down a higher-order page in sub-pages, and keep our target out of
8940 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8941 struct page *target, int low, int high,
8944 unsigned long size = 1 << high;
8945 struct page *current_buddy, *next_page;
8947 while (high > low) {
8951 if (target >= &page[size]) {
8952 next_page = page + size;
8953 current_buddy = page;
8956 current_buddy = page + size;
8959 if (set_page_guard(zone, current_buddy, high, migratetype))
8962 if (current_buddy != target) {
8963 add_to_free_list(current_buddy, zone, high, migratetype);
8964 set_buddy_order(current_buddy, high);
8971 * Take a page that will be marked as poisoned off the buddy allocator.
8973 bool take_page_off_buddy(struct page *page)
8975 struct zone *zone = page_zone(page);
8976 unsigned long pfn = page_to_pfn(page);
8977 unsigned long flags;
8981 spin_lock_irqsave(&zone->lock, flags);
8982 for (order = 0; order < MAX_ORDER; order++) {
8983 struct page *page_head = page - (pfn & ((1 << order) - 1));
8984 int page_order = buddy_order(page_head);
8986 if (PageBuddy(page_head) && page_order >= order) {
8987 unsigned long pfn_head = page_to_pfn(page_head);
8988 int migratetype = get_pfnblock_migratetype(page_head,
8991 del_page_from_free_list(page_head, zone, page_order);
8992 break_down_buddy_pages(zone, page_head, page, 0,
8993 page_order, migratetype);
8997 if (page_count(page_head) > 0)
9000 spin_unlock_irqrestore(&zone->lock, flags);